Datasets:
kilian-group
/
arxiv-classifier

Dataset card Data Studio Files Community
1
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
1605.01784
1
1605
2016-05-05T22:45:29
Neural mechanisms underlying catastrophic failure in human-machine interaction during aerial navigation
[ "q-bio.NC" ]
Objective. We investigated the neural correlates of workload buildup in a fine visuomotor task called the boundary avoidance task (BAT). The BAT has been known to induce naturally occurring failures of human-machine coupling in high performance aircraft that can potentially lead to a crash; these failures are termed pilot induced oscillations (PIOs). Approach. We recorded EEG and pupillometry data from human subjects engaged in a flight BAT simulated within a virtual 3D environment. Main results. We find that workload buildup in a BAT can be successfully decoded from oscillatory features in the electroencephalogram (EEG). Information in delta, theta, alpha, beta, and gamma spectral bands of the EEG all contribute to successful decoding, however gamma band activity with a lateralized somatosensory topography has the highest contribution, while theta band activity with a frontocentral topography has the most robust contribution in terms of real world usability. We show that the output of the spectral decoder can be used to predict PIO susceptibility. We also find that workload buildup in the task induces pupil dilation, the magnitude of which is significantly correlated with the magnitude of the decoded EEG signals. These results suggest that PIOs may result from the dysregulation of cortical networks such as the locus coeruleus (LC) anterior cingulate cortex (ACC) circuit. Significance. Our findings may generalize to similar control failures in other cases of tight man machine coupling where gains and latencies in the control system must be inferred and compensated for by the human operators. A closed-loop intervention using neurophysiological decoding of workload buildup that targets the LC ACC circuit may positively impact operator performance in such situations.
q-bio.NC
q-bio
Neural mechanisms underlying catastrophic failure in human-machine interaction during aerial navigation Sameer Saproo1, Victor Shih1, David C. Jangraw2, and Paul Sajda1 1Department of Biomedical Engineering, Columbia University, New York 2National Institutes of Health, Bethesda, Maryland E-mail: [email protected], [email protected] Abstract. Objective. We investigated the neural correlates of workload buildup in a fine visuomotor task called the boundary avoidance task (BAT). The BAT has been known to induce naturally occurring failures of human-machine coupling in high performance aircraft that can potentially lead to a crash – these failures are termed pilot induced oscillations (PIOs). Approach. We recorded EEG and pupillometry data from human subjects engaged in a flight BAT simulated within a virtual 3D environment. Main results. We find that workload buildup in a BAT can be successfully decoded from oscillatory features in the electroencephalogram (EEG). Information in delta, theta, alpha, beta, and gamma spectral bands of the EEG all contribute to successful decoding, however gamma band activity with a lateralized somatosensory topography has the highest contribution, while theta band activity with a fronto-central topography has the most robust contribution in terms of real-world usability. We show that the output of the spectral decoder can be used to predict PIO susceptibility. We also find that workload buildup in the task induces pupil dilation, the magnitude of which is significantly correlated with the magnitude of the decoded EEG signals. These results suggest that PIOs may result from the dysregulation of cortical networks such as the locus coeruleus (LC) - anterior cingulate cortex (ACC) circuit. Significance. Our findings may generalize to similar control failures in other cases of tight man-machine coupling where gains and latencies in the control system must be inferred and compensated for by the human operators. A closed-loop intervention using neurophysiological decoding of workload buildup that targets the LC-ACC circuit may positively impact operator performance in such situations. 1. Introduction Superior human performance in complex tasks such as piloting a modern jet fighter or driving a Formula 1 car requires goal-directed navigation while operating within dynamic physical constraints or error margins. Such performance requires a careful balancing of cognitive resources, maximizing task engagement while keeping autonomic stress response in check. A failure to maintain this balance can result in catastrophic accidents. For instance, pilot-induced oscillations (PIO) are a dangerous flight characteristic that can spontaneously develop during periods of demanding task performance, e.g., when landing on the deck of a naval aircraft carrier, and can lead to loss of control and airframe damage (Hurt, 1965). Although the phenomenon of PIOs has been known in the flight community ever since the advent of manned flight, the underlying factors have not been completely understood. PIOs are defined as unstable oscillations in the longitudinal motion of an aircraft that are inadvertently caused by the pilot's own control input. Traditionally, PIOs have been attributed to non-optimal coupling between the pilot and the aircraft. Spontaneous dampened short-period oscillations are normal, but they can become dangerous if the pilot over-compensates small control errors in a way that increases the amplitude of aircraft oscillations to dangerous levels. Moreover, a pilot's unfamiliarity with the 'feel' of the aircraft (e.g., during training or test flights of prototype airplanes) can increase the likelihood of a PIO (Hurt, 1965). Previous investigations into PIOs have suggested that an aggressive mindset, high pilot workload, and tight error margins could be contributing factors (Gray, 2005, Gray, 2008). In particular, PIOs have been recreated in laboratory conditions using a 'boundary avoidance task' (BAT) paradigm, which entails gradually reducing the permitted margin of error while the pilot is attempting to closely track a complicated flight trajectory (Gray, 2005, Gray, 2008, Dotter, 2007, Warren, 2006). The BAT paradigm is thought to gradually increase a pilot's cognitive workload, arousal, and task engagement, until the cognitive conditions induce a catastrophic control failure, as in a PIO. However, what these cognitive conditions are, and how they mechanistically induce control failures, has not been determined so far. We present the first neurophysiological study of PIOs, where we investigated neural factors underlying PIOs using a naturalistic 3D BAT paradigm while simultaneously recording EEG and pupillary activity, which was then used to build a predictive classifier that tracked PIO susceptibility. Our data suggest that heightened error monitoring and error sensitivity in the decision-making circuitry of the brain, along with increased arousal, is coincident with a higher probability of PIOs. 2. Methods 2.1. Experimental Design and Stimuli We used NEDE (Jangraw et al., 2014), a 3D virtual environment created using Unity game development platform (Unity Technologies, CA), to create a realistic visuomotor task. This task required maneuvering a high-speed virtual aircraft in first-person perspective through a series of equidistant glide boxes that defined a complex undulating trajectory (Figure 1a). The environment was rendered on a 30 inch Apple Cinema HD display (1200 × 800, 60 Hz) that subtended 30 × 23 degrees of visual angle. The virtual aircraft could be maneuvered in the pitch axis using a flight joystick (Attack 3 Joystick, Logitech S.A.), but with yaw and roll controls disabled. Glide boxes were placed every 500 m in the virtual environment, while the aircraft moved forward with a constant velocity of 250 m/s. Thus, a subject had to navigate through a glide box approximately every 2 seconds. The trajectory formed by glide boxes was modeled as a weighted sum of 3 sinusoids of varying periods and amplitudes in the pitch axis (Figure 1b, Supplemental Figure 1). Glide box boundaries were task critical; failure to navigate through even a single glide box ended the flight abruptly. The task difficulty was manipulated by decreasing the size of the glide boxes at regular intervals (30s) during each trial. Therefore, the user-controlled flight during each trial (max 90s) could be divided into 3 distinct epochs of identical glide box trajectory but with steadily increasing difficulty; this is in consonance with previous 2D BAT investigations (Dotter, 2007). Supplementary Figure 1A shows a second glide path trajectory used in the experiment. Each trial started with 2 seconds of passive fixation (white cross in the middle of a blue screen) followed by 4 seconds of passive flight through the virtual environment towards the first glide box. The subject's joystick input had consequence only after the virtual aircraft had passed through the first glidebox in the trial. The specific characteristics of the aircraft's response to control input play a critical role in the causation of PIOs, with a more oscillatory pitch response as well as response lag leading to higher chance of a PIO (Dotter, 2007). Therefore, the pitch response of the virtual aircraft in our experiment incorporated both of these elements; an initial lag (0.2s or 0.3s) with a subsequent oscillatory movement of the aircraft in response to a step input of the joystick (Supplemental Figure 1). The maximum instantaneous and sustained pitch response was limited to 40 deg/s and 24 deg/s respectively. 2.2. Subjects A total of 12 healthy human subjects (ages 19-33, all right handed, 4 females) with normal or corrected-to-normal vision participated in the study. Informed consent was obtained from all subjects in accordance with the guidelines of the Institutional Review Board at Columbia University. Subjects were provided with a set of written instructions about the task. They were also familiarized with the virtual aircraft controls by providing least 60-120 minutes of flying practice in the NEDE environment prior to the experiment, either on the same day as the experiment (3 subjects) or the previous day (9 subjects). Data from 3 subjects were discarded from final analysis due to system malfunction during the experiment or poor EEG quality, i.e., inability to get a sufficient number (>1) of neural independent components after artifact rejection (see Data pre-processing section). 2.3. Data Collection During each trial, subjects' joystick input and the position of the aircraft were sampled at 60Hz. In addition to motor behavior, subjects' neurophysiological activity was measured at 2048Hz using an EEG system: Biosemi B.V. ActiveTwo AD-box, 64 Ag-AgCl active electrodes, 10-20 montage. All electrode offsets were below 40 mV at the beginning of the experiment. Gaze position and pupil size were recorded using an EyeLink 1000 eye tracker (SR Research, Ontario, Canada) at a 1000 Hz sampling frequency. The subject's head was stabilized during data collection using a chin and forehead rest. A 9-point calibration of the eye tracker was performed before each block of trials, and the subjects were instructed to not move their head between or during trials (they were provided regular breaks however). Note: stimulus delivery and behavior recording, EEG recording, and eye tracking were performed on 3 different computers simultaneously, with the respective data synced post-hoc (Jangraw et al., 2014). A total of 32, 40, or 48 flight runs were observed per subject, with the total experiment time not exceeding 1 hr. Note that given 2 different pitch response delays and 2 different glide- path trajectories, each unique combination of delay and trajectory was run an equal number of times (e.g., 10 per combination, for a total of 40 trials per subject). This was done to maximize the probability of sampling pilot behavior relevant to PIOs given inter-subject variability in innate performance on the task. Furthermore, slight changes in control parameters and flight trajectory across trials ensured that subjects had minimal learning or 'muscle memory' of the specific experimental parameters, and thus no steady increase in performance throughout the experiment. 2.4. Data Pre-Processing All EEG and pupillometry data were analyzed using the EEGLAB toolbox (Delorme and Makeig, 2004) in MATLAB (The MathWorks Inc., Natick, MA). The recorded EEG signals were first re- referenced to the average of all electrodes and then band-pass filtered to 0.5-100Hz using a Hamming windowed FIR filter. The result was then notch filtered at 60Hz to remove line noise, and finally down-sampled to 256 Hz. This pre-processing pipeline produced 'raw' datasets that contained signals from neural, ocular, and muscular sources, as well as non-physiological artifacts. To isolate the purely neural component of the EEG data, we used the following procedure: we first reduced the dimensionality of the EEG data by reconstituting the data using only the top 20 principal components derived from Principal Component Analysis (PCA). Thereafter, an Independent Component Analysis (ICA) decomposition of the data was performed using the Infomax algorithm (Bell and Sejnowski, 1995). We then used an ICA-based artifact removal algorithm called MARA (Winkler et al., 2011) to remove ICs attributed to blinks, horizontal eye movements (HEOG), muscular activity (EMG), and any loose or highly noisy electrodes. MARA performs automatic IC classification using a linear classifier trained on time- series, spectrum, and scalp map features of a large dataset of labeled IC artifacts. MARA assigns each IC a probability of being an artifact; we removed components with probabilities above 0.5. 2.5. EEG Data Classification The 64-channel EEG signals recorded during each trial were split into 1500 ms epochs that were centered at the onset of each stick movement. Thus, each epoch was construed as a unique data point for classification. We used spectral power as the classification feature; therefore, a spectrogram of the entire continuous data was computed using a sliding Short Time Fourier Transform (STFT) with a 128 sample Hamming window and 64 samples of overlap between windows, yielding five 500 ms windows of frequency data for each electrode. For further analysis, frequencies were separated into bands of interest: delta band consisted of frequencies 1- 3 Hz, theta band 4-7 Hz, alpha band 8-15 Hz, beta band 16-31 Hz, and gamma band 32-55 Hz. Spectral information from 56-128 Hz was not used in the classifier. Classification was performed either using the spectral power in different frequency bands as features (delta only, theta only, alpha only, beta only, gamma only) or using the power from all frequencies (all bands, Figure 3). Each data point was given a class label according to the nearest boundary size at the time of stick movement during the trial. In one classification regime (which we call LvsMS), the first class includes all stick movements made during navigation under large boundaries, and the second class includes all stick movements made during navigation under medium and small boundaries. In another classification regime (time-on-task or ToT), data points generated only within large boundaries were sub-divided according to their temporal occurrence into classes (first half or second half). Finally, the MvS classifier used the weight vector learned from the LvMS regime to classify Medium vs Small classes; this was done in order to show that the workload component (weight vector) was generalizable to different absolute boundary sizes. We used N-fold cross-validation to generate results, usually with both training and test data derived from the same subject, except in one case where leave-one-subject-out cross-validation was used to investigate whether the PIO classifier was generalizable to novel subjects. 2.5.1. Classification Algorithm The high dimensionality of data – there are 17,600 features per data point when using all frequency bands (1-55Hz) – required the use of a recently developed algorithm FaSTGLZ (Conroy et al., 2013), for efficient linear classification. FaSTGLZ classifies input data !∈ℝ! with binary class labels !∈{0,+1} by using logistic regression to create a separating hyperplane in the feature space that is parameterized by a normal vector != w!,…,w! ∈ℝ!. For the sake of simplicity, the classifier bias is estimated by incorporating a constant !!!!=1 and bias term !!!! into the classification. The posterior probability of the class label !! for each data point !! is modeled as a sigmoid 1+exp (−!!!!) (1) !!!=1!!,! = For logistic regression, denoting !!=1!,! =!!!! , the negative log-likelihood is given (2) !!log !!!!! + 1−!!log (1−!(!!!!)) !!!! !! = ℒ! + !!!!" (3) If the norm ! is any symmetric positive semi-definite matrix and ! a real-valued scalar, then !! A common problem with Maximum-likelihood estimators is the severe over-fitting of high dimensional training data. FaSTGLZ mitigates such over-fitting by using a penalized likelihood method based on L2-regularization that seeks to minimize: 1 function: by ℒ! =− would be a convex function, which is optimized by FaSTGLZ using the 'Alternating Direction Method of Multipliers' (ADMM) procedure. ADMM uses variable splitting to divide the main optimization into two simpler sub-procedures to minimize a differentiable objective and to solve a soft-thresholding operation. This allows the simultaneous training of high dimensional models across bootstraps, cross-validation folds, and permutation tests, thus considerably speeding up classifier learning. Note that all classifiers in our analyses were learned using 5-fold cross validation with 100 bootstraps for each fold. The optimal lambda values were chosen using a parameter sweep of 100 lambda values between 1e5 and 1e-5; the value that yielded the highest AUC was used for further analysis. The resulting classifier assigns a set of weights to the feature space used to train the model, such that each multidimensional data-point is projected onto a scalar dimension where the two classes are maximally separated. The classifier features – spectral power of EEG signals – were z-scored across epochs before classification, and therefore the learned classifier weights can be interpreted as the normalized contribution of each frequency at each electrode to the discriminating hyper plane. A positive weight would imply that the classification feature is more correlated with low pilot workload (Larger boundary size) and a negative weight would imply a stronger correlation with higher pilot workload (smaller boundary size), therefore describing the direction of the change in spectral magnitude across boundary size. Furthermore, the entire set of classifier weights (frequency band × time point × electrode) can be localized on the scalp, thus showing the spatial and temporal signature of neural correlates of workload (Figure 3). 3. Results The behavioral data show that our experimental paradigm elicited piloting behavior relevant to PIOs, i.e. there is an increase in the magnitude of PIO features with decreasing boundary size during the boundary avoidance task (Figure 2). Specifically, a reduction in boundary size led to quicker task failure (One-way repeated-measured ANOVA; F(2,16)=185.3, p<0.001), increased magnitude of joystick force (F(2,16)=63.59, p<0.001), increased frequency of joystick input (F(2,16)=22.24, p<0.001), and a rapid increase in the phase divergence between the input and the response (F(2,16)=25.37, p<0.001). Here, the phase divergence, between the input and the response, was computed by taking absolute the difference between the unwrapped phases of the Hilbert transform of accumulated joystick input and current aircraft heading. We used the spectral power of stick-locked EEG signals (1500ms around each stick movement) in different canonical frequency bands (delta, theta, alpha, beta, and gamma) as features to classify BAT-induced workload (Classes: Large vs Medium/Small boundaries), and we observed above-chance classification accuracy (chance AUC=0.5) for all subjects (Figure 3A, Table 1). The choice of the size of the epoch was dictated by the asymptote of classification accuracy across different epoch sizes (Supplemental Figure 2). In order to dissociate the contribution of different frequency bands to workload classification, we computed the classification accuracy for each band separately by filtering EEG signals to the respective band before classification (see Methods for specific frequencies for each band). The results show that the gamma band is the most informative band for classifying BAT-induced workload, as the gamma band classifier approximates the accuracy of a full spectrum classifier. This difference in classification accuracy does not seem to be a consequence of the higher dimensionality of the gamma band, as qualitatively similar results are produced when classifying using the average power in each band (Supplemental Figure 3). We also find that regularly sampling EEG signals (every 2s) for classification did not have a significant qualitative difference from stick-locked EEG signals (Supplemental Figure 4). The scalp topology of the classifier weights for a full-spectrum classifier suggests that the contribution of delta, theta, and gamma band activity to workload classification is spatially localized. Delta and theta band based classifiers had significant contributions from fronto-central sites, while gamma band modulation had a predominately lateralized somatosensory topography (Figure 3B). Furthermore, the scalp topology for all frequency bands was robust to the variation in the temporal overlap of windows used to compute spectral power using a Fast Fourier Transform (see Supplemental Figure 5 for a classifier with higher temporal overlap). We tested the robustness and the generalizability of classifiers from individual bands. We found that even though the theta band classifier does not produce high classification accuracy compared to the gamma band classifier, fronto-central theta activity is a more robust signature of workload since MARA-based artifact cleaning of raw EEG affected theta-band classifier the least (Figure 4). These data suggest that fronto-central theta activity might prove to be the best indicator of workload in an operational scenario (i.e., while flying a real fighter plane), due to the significant contamination of EEG signals with potentials from muscle activity. Similarly, training the classifier with data from non-test subjects (hold-one-subject-out cross-validation) impacted the accuracy of theta band classifier the least, further attesting to its generalizability (Figure 5). Workload can build up due to sustained focal attention required by the task, which would be unrelated to boundaries in the BAT scenario. However, we find that the contribution of time-on-task component to classifier performance for our data was not enough to explain the steadily increasing EEG signatures of workload in our BAT paradigm (Figure 6, Supplemental Figure 6). Furthermore, LvMS classifier (classes: large vs medium/small; data from medium and small boundaries were collapsed into a single class) could also reliably distinguish Medium from Small boundary conditions (MvS). This suggests that our assessed neural correlates (vector of weights normal to the classifying hyperplane) are independent of absolute boundaries and can reliably predict a continuum of workload states. More importantly, data suggest that neural correlates derived from laboratory BAT experiments can be effectively used to provide continuous feedback about pilot workload and PIO tendency in real-time (Figure 7). In addition to EEG, we collected pupillometry data while the subjects performed the Although, we trained classifiers to discriminate EEG signals based on boundary size, we demonstrate that the classifier output can also track PIO tendency. We estimated PIO tendency by creating a metric; the amplitude of Hilbert transform of band-passed stick movement (0.3-1.8Hz, range that is typical of PIOs; (TIAN, 2006, Rzucidło, 2007). We then separated the trials into 4 bands of increasing PIO tendency according to the magnitude of PIO measure in the last 5 seconds of a trial, and compared highest to lowest band (Q1: band lower than 1st quartile; Q3: band higher than 3rd quartile). We find that there is a significant difference between time- averaged (~10s) classifier output that leads up to a PIO event towards the end of the trial (Q3), compared to trials that did not end in a PIO (Q1), with the temporal trend showing steady divergence (Figure 8). experimental task. Data show that subject's pupils dilate with a decrease in boundary size (Figure 9A), directly implicating mental load and arousal (Murphy et al., 2011). We find a significant correlation between EEG-derived classifier output and pupil size for a full spectrum classifier (Figure 9B, one sample t-test; t(8) = -2.41, p = 0.043). More importantly, we find a significant increase in the correlation between the output of theta band classifier – with fronto-central topography – and the pupil size with a decrease in boundary size during BAT (Large vs. Medium boundaries), suggesting a close anterior cingulate cortex – locus coeruleus norepinephrine (ACC- LC-NE) interaction during induced workload buildup (One-way repeated-measured ANOVA; F(1,8)= 8.08, p=0.022). This change in correlation was not observed with the full-spectrum classifier (F(1,8)= 0.02, p=0.88). 4. Discussion We performed the first neurophysiological investigation into the phenomenon of pilot induced oscillations (PIOs), using a boundary avoidance task (BAT) in a gaming/virtual reality environment. We find that our task is able to gradually induce cognitive workload that in some cases causes PIO-like behavior. Furthermore, we find robust EEG signatures of workload in different spatio-spectral bands, with fronto-central theta band the most robust, in terms of being differentiable from potential artifacts. We also find a significant correlation between this EEG activity and pupil dilation due to BAT induced workload. Below we discuss these results within the context of specific circuits in the brain and also possible broader implications of our findings. The encoding and regulation of error monitoring has typically been associated with the anterior cingulate cortex (ACC), which is believed to be a key brain area for cognitive control and storing of predictive models of our environment (Tervo et al., 2014). The region ACC is believed to be at least partially modulated by the locus coeruleus, a tiny nucleus in the dorsal pons, regulating arousal levels in the brain via the neurotransmitter norepinephrine (Aston-Jones and Cohen, 2005). The link between arousal state and task performance has been shown to be non- linear (Aston-Jones and Cohen, 2005, Gompf et al., 2010). For example, the Yerkes-Dodson curve posits that a mid-level arousal state is the most optimal for task performance, though this "mid-level" is highly task and context dependent. Therefore, the LC-ACC circuit is of particular interest in decision-making under dynamic constraints (e.g., flying an aircraft or driving a vehicle) since dynamically adapting motor control strategies based on assessment of current performance and upcoming task constraints is often key to optimal performance. Recent work in animal models has shown a tight coupling between the LC- norepinephrine system (LC-NE) and the ACC when animals must dynamically switch between task-based models (Tervo et al., 2014). Specifically, the rats in the experiment faced a computer opponent in a competitive virtual task, where the computer was programmed to counter-predict a rat's behavior. When the LC-NE input to the ACC increased, the rats were less adept at incorporating environmental feedback into their internal model of choice and prediction. However, when the LC-NE input to ACC was suppressed, the rats were able to utilize feedback from the environment more effectively and therefore better model the computer's counter prediction to increase their performance and reward. Though we cannot directly measure LC-NE activity with scalp EEG, several studies have shown that pupil dilation can be used as a proxy for activity in the LC and thus provides some information of the state of arousal of an individual (Gompf et al., 2010, Joshi et al., 2016). ACC, on the other hand, is more accessible via EEG, with fronto-central theta activity having been identified as a correlate of ACC activation (Cavanagh and Frank, 2014). Thus, by linking EEG activity with pupillary measures, one can potentially, non-invasively, infer the dynamics of the LC-ACC circuit during a complex and dynamic task. Our results, therefore, can be interpreted within the context of the aforementioned study of LC-ACC interaction (Tervo et al., 2014), and may provide a mechanistic explanation for PIOs. An increase in observed fronto-central theta band power, which in turn is correlated with pupil dilation in our study, could be an indication of the subject switching into a behavioral model associated with high workload state. This might suggest that in a cognitive state associated with high workload, there is an increase in LC-NE input to ACC, which might lead to the subjects sticking with their current internal model of aircraft control, even when the boundaries have changed. This sub-optimal control model might lead to PIOs in certain instances. In contrast, a better strategy would be to incorporate environmental feedback and switch to a different internal model of aircraft control that is better adapted to steering within narrow boundaries. This interpretation of our results provides an interesting possibility for mitigating PIOs: since the LC-NE system is associated with arousal, using feedback from a hybrid BCI system (hBCI) to dynamically adjust arousal levels may regulate LC-NE input to the ACC, allowing updates to the internal model of the pilot based on the environmental feedback. For example, a hBCI that integrates pupillometry and EEG features could predict when a pilot is entering a state that will likely generate a PIO (as in Fig. 8, red curve), whereupon feedback in the form of a continuous auditory stimulus with calming influence, could be delivered to reduce arousal level and thus reduce LC-NE input to ACC. We hypothesize that an optimally calibrated feedback loop would help regulate LC-ACC interaction, resulting in piloting behavior improvements. Beyond cases of vehicular control, there is a large class of electronic games, such as the highly popular "Flappy Bird", that resemble a boundary avoidance task; the player controls a character moving at constant speed, avoiding obstacles and boundaries that become tighter as the game progresses. The objective in these games is to go as far along the course as possible. These games are known to be highly addictive, with the gamers repeatedly replaying the course from the beginning after a failure, trying to increase the distance they come along a course before failure. Although a recent attempt has been made to quantify the optimal parameters for such games such that they remain highly playable (Isaksen, 2015), the cognitive factors underlying their addictive nature remains unknown. Our results from the BAT investigation suggests that not only does arousal level (as evidenced by pupil dilation) increase progressively as the boundaries decrease and therefore difficulty increases, but there is a dramatic increase in the cognitive workload due to task monitoring (as evidenced by theta band activity over fronto-central sites). This presents a peculiar hypothesis: perhaps the addictive nature of such games comes from the ability to progressively achieve higher arousal levels, as the subjects improve their ability to increase the ACC-LC coupling to move to the more optimal position along the Yerkes-Dodson curve. Indeed, our data show that the correlation between fronto-central theta band activity (stand in for ACC) and pupil dilation (stand in for LC activity) increases during the course of the experiment. Finally, our results have relevance beyond the world of tracking physical boundaries as in gaming or vehicular navigation; humans frequently engage in sustained perception-decision- action loops that involve goal and error tracking under dynamic constraints. For example, a financial portfolio manager has to track changing market conditions and reallocate stocks so as to maximize portfolio value while managing risk within prescribed boundaries. Project managers must regularly track project progress and deal with exigencies so as to ensure high quality of work while avoiding unacceptable delays in completion. Viewed generally, these examples are illustrative of rapid decision-making that involves tracking optimal performance while avoiding frequently changing 'failure' boundaries. As with top-gun pilots, Formula 1 champions, top fund managers, and top project managers, the burning question is "what neural markers differentiate stellar performance (and performers) from catastrophic failures under challenging conditions"? Our results suggest that the key insight may lay in the interaction of neural circuitry that is engaged in error monitoring, decision-making, and regulating arousal. TABLES Stick Locked Classifier Regularly Sampled Classifier All Bands 0.75±0.02 Delta Theta Alpha Beta 0.56±0.01 0.57±0.02 0.59±0.02 0.65±0.03 Gamma 0.73±0.02 0.75±0.02 0.57+0.01 0.58+0.02 0.61±0.02 0.66±0.03 0.72±0.03 Table 1: Mean area-under-curve (AUC) values for workload classifiers (Classes: Large vs Medium/Small boundaries) using different spectral content (mean±SEM). Acknowledgments The authors would like to thank Eric Pohlmeyer and Mark Chevillet for their helpful comments and suggestions, and acknowledge Sona Roy and Wha-Yin Hsu for their assistance with data collection. The work was funded by the Defense Advanced Research Projects Agency (DARPA) and Army Research Office (ARO) under grant number W911NF-11-1-0219, the Army Research Laboratory under Cooperative agreement number W911NF-10-2-0022 and the U.K. Economic and Social Research Council under grant number ES/L012995/1. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the US Government. The US Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. References ASTON-JONES, G. & COHEN, J. D. 2005. An integrative theory of locus coeruleus- norepinephrine function: adaptive gain and optimal performance. Annu Rev Neurosci, 28, 403-50. BELL, A. J. & SEJNOWSKI, T. J. 1995. An information-maximization approach to blind separation and blind deconvolution. Neural Comput, 7, 1129-59. CAVANAGH, J. F. & FRANK, M. J. 2014. Frontal theta as a mechanism for cognitive control. Trends Cogn Sci, 18, 414-21. CONROY, B. R., WALZ, J. M., CHEUNG, B. & SAJDA, P. 2013. Fast Simultaneous Training of Generalized Linear Models (FaSTGLZ). eprint arXiv:1307.8430. DELORME, A. & MAKEIG, S. 2004. EEGLAB: an open source toolbox for analysis of single- trial EEG dynamics including independent component analysis. J Neurosci Methods, 134, 9-21. DOTTER, J. D. 2007. An analysis of aircraft handling quality data obtained from boundary avoidance tracking flight test techniques. GOMPF, H. S., MATHAI, C., FULLER, P. M., WOOD, D. A., PEDERSEN, N. P., SAPER, C. B. & LU, J. 2010. Locus ceruleus and anterior cingulate cortex sustain wakefulness in a novel environment. J Neurosci, 30, 14543-51. GRAY, W. 2005. Boundary Avoidance Tracking: A New Pilot Tracking Model. AIAA Atmospheric Flight Mechanics Conference and Exhibit. American Institute of Aeronautics and Astronautics. GRAY, W. 2008. A Generalized Handling Qualities Flight Test Technique Utilizing Boundary Avoidance Tracking. 2008 U.S. Air Force T&E Days. American Institute of Aeronautics and Astronautics. HURT, H. H. 1965. Aerodynamics for naval aviators, Aviation Supplies & Academics. ISAKSEN, A. G., DAN; NEALEN, ANDY 2015. Exploring Game Space Using Survival Analysis. Foundations of Digital Games. JANGRAW, D. C., JOHRI, A., GRIBETZ, M. & SAJDA, P. 2014. NEDE: an open-source scripting suite for developing experiments in 3D virtual environments. J Neurosci Methods, 235, 245-51. JOSHI, S., LI, Y., KALWANI, R. M. & GOLD, J. I. 2016. Relationships between Pupil Diameter and Neuronal Activity in the Locus Coeruleus, Colliculi, and Cingulate Cortex. Neuron, 89, 221-34. MURPHY, P. R., ROBERTSON, I. H., BALSTERS, J. H. & O'CONNELL R, G. 2011. Pupillometry and P3 index the locus coeruleus-noradrenergic arousal function in humans. Psychophysiology, 48, 1532-43. RZUCIDŁO, P. 2007. The detection of pilot(cid:1)induced oscillations. Aviation, 11, 15-22. TERVO, D. G., PROSKURIN, M., MANAKOV, M., KABRA, M., VOLLMER, A., BRANSON, K. & KARPOVA, A. Y. 2014. Behavioral variability through stochastic choice and its gating by anterior cingulate cortex. Cell, 159, 21-32. TIAN, F.-L. G., ZHENG-HONG YU,; ZHI-GANG 2006. Application of Gaussian Complex Wavelet in PIO Detection. International Congress of Aeronautical Sciences. WARREN, R. D. 2006. An Investigation of the Effects of Boundary Avoidance on Pilot Tracking. WINKLER, I., HAUFE, S. & TANGERMANN, M. 2011. Automatic classification of artifactual ICA-components for artifact removal in EEG signals. Behav Brain Funct, 7, 30. A B 2.5 x 104 2.5 x 104 2 2 view from above ) x ( h t p e d 50 0 −50 0 0.5 ) y ( n o i t i s o p l a c i t r e v 1200 1000 800 0 0.5 1 1 1.5 horizontal position (z) view from the side horizontal position (z) 1.5 Figure 1. A) Screenshot of subject's view during 3D Boundary Avoidance Task (BAT) experiment. Red squares depict waypoint boundaries, while the dotted horizontal green line in the center shows the current heading of the virtual aircraft. B) Full flight trajectory with the position and the size of the glide boxes; solid blue line shows the mean path through the center of glide boxes, while solid red lines denote glide boxes. All dimensions are in meters. Virtual aircraft moved steadily in z-axis @ 250 m/s and could be controlled in y (pitch) axis via joystick input. See supplemental figure 1 for the other glide path trajectory used in the experiment. ) s ( e m i t t h g i l F 30 20 10 0 m r o f s n a r T t r e b l i H f o e d u t i l p m A 0.5 0.4 0.3 0.2 0.1 0 Flight Length Comparison Max = 28s A Large Medium Small Stick Magnitude Comparison C Large Medium Small d n o c e s r e p s t u p n I ) s d a r ( φ ∆ 0.8 0.6 0.4 0.2 0 10 8 6 4 2 0 Stick Frequency Comparison B Large Medium Small Flightpath phase divergence D Large Medium Small Figure 2. Four measures that demonstrate that the BAT paradigm elicited control behavior typical of PIOs; A) a reduction in boundary size led to a decrease in flight length before failure (missing a glide box), B) more frequent and C) larger joystick inputs, and D) a quicker increase in the phase divergence between the Hilbert transform of cumulative control input and current aircraft trajectory. Error-bars reflect mean ± SEM across subjects (N=9). C U A 0.9 0.85 0.8 0.75 0.7 0.65 C U 0.6 A 0.55 0.5 0.45 0.4 1 (in descending order of AUC when using all bands) Subjects A B Workload classification All Bands Delta Theta Alpha Beta Gamma 2 3 4 5 6 7 8 9 Average weights for stick locked classifier -0.5 -0.25 0 FFT window center (s), 0.5s window size 0.25 0.5 Figure 3. A) Area under the receiver operating characteristic curve (AUC) for all subjects (in descending order of 'All Bands' classifier AUC), when using information from all spectral bands for classification ('All Bands'), as well as when using only individual bands (delta, theta, alpha, beta, or gamma). Classification was performed using 64-channel MARA-cleaned EEG signals -1.5s epochs around each joystick movement - that were labeled according to the size of the nearest glide path boundary at the time of their generation. B) Subject-averaged scalp distribution of normalized weights for the classifying hyperplane that best separated Large boundary from Medium and Small boundaries (scalp map corresponds to 'All Bands' classifier in panel A). Delta and theta band activity from fronto-central electrodes, as well as significant gamma band activity from a lateralized somatosensory topography, seems most indicative of higher workload induced by smaller boundaries. 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 C U A Comparison of MARA Cleaning Effect on Classification ** No MARA MARA ** ** * All Bands Delta Theta Alpha Beta Gamma Figure 4. Effect of artifact removal, using MARA algorithm, on classification accuracy (Large vs Medium/Small) when using information all bands or when using individual bands. MARA algorithm classifies ICA components in the data as artifact based on a pre- existing labeled set of artifactual ICs, including those for eye movements, muscle movements, and noisy electrodes. Error-bars reflect mean ± SEM across subjects. Paired t-test; * p<0.01, ** p<0.001. Figures 3, 5-9 show results from MARA cleaned data. C U A 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 Within Subject Classification Hold Subject Out Classification ** * * * All Bands Delta Theta Alpha Beta Gamma Average weights for stick locked 'Hold Subject Out' classifier A B FFT window center (s), 0.5s window size Figure 5. A) Comparison of classification accuracy (Large vs Medium/Small) when using 'Within subject' k-fold cross-validation or when using 'Hold Subject Out' cross-validation. Error-bars reflect mean ± SEM across subjects. Paired t-test; * p<0.05, ** p<0.01 B) Subject-averaged scalp distribution of normalized weights for the classifying hyperplane that best separated Large boundary from Medium and Small boundaries for 'Hold Subject Out' classification (scalp map corresponds to 'All Bands' classifier in panel A). Classifier Accuracy Comparison of Classifier Az LvMS MvS Time on Task 5 6 7 8 9 C z U A A 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 1 2 3 4 Subjects (Descending AllBands Az) Subjects (in descending order of LvMS AUC) Figure 6. Subject-wise Classification accuracy: LvMS, where data from Medium and Small boundaries was combined into a single training class, and the classifier learned to distinguish Large from Medium/Small boundary classes. MvS, used the weights learned from LvMS classifier and classified Medium vs Small boundary classes. Time on Task, when classifying data from the Large boundary class that is labeled according to temporal occurrence - first half or second half of flight within large boundaries. 0 0 5 10 15 20 25 30 35 40 5 10 15 20 25 30 35 40 −50 0 5 10 15 20 25 30 35 ) m ( n o ) d a r ( e s a h P a v e E i t l 1050 100 1000 50 950 900 0 ) . . U A ) d a r ( ( t u p n g a L e s a h P I k c i t s y o J 1 10 0 0 −1 −10 50 2 0 0 ) g e d ( ) z ( y h c t i 0 0 0 0 5 5 10 10 15 15 20 20 25 25 30 30 35 35 40 40 5 5 10 10 15 15 20 20 25 25 30 30 35 35 40 40 0 0 5 5 10 10 15 20 15 20 Time (s) 25 25 30 30 35 35 40 −50 −2 100 50 n o i t a v e E l 1000 950 900 1 0 −1 50 0 ) . U A . ( t u p n I k c i t s y o J ) g e d ( h c t i p δ p δ ) d a r ( e s a h P ) d a r ( g a L e s a h P −10 ) z ( y 2 0 −2 Subject Path: flight Rings Boundaries Joystick Input Input Onset Boundary Change Commanded δ pitch Actual δ pitch Boundary Change Subject Path: takeoff Phase: Commanded Subject Path: flight Phase: Actual Rings Boundary change Boundaries Joystick Input Phase Difference Input Onset Boundary Change Boundary Change Commanded δ pitch Classifier Output Actual δ pitch Potential Feedback Boundary Change Boundary Change Phase: Commanded Phase: Actual Boundary change 40 40 0 0 5 20 15 10 Figure 7: A representative trial flight showing the measured behavioral and neural markers across time. Top to bottom: Flight path, control stick movements, commanded and actual pitch of aircraft, 10 phase of pitch, phase lag of pitch, and z-scored classifier output y overlaid on joystick input onset. Potential feedback to the pilot is constructed by interpolating and filtering classifier output (cubic Phase Difference spline interpolation, then 3rd order Butterworth filter with 0.1 Hz cutoff to smooth output over 10s) Boundary Change 25 30 35 40 0 0 0 5 10 15 20 25 30 35 40 5 10 15 20 Time (s) 25 30 35 40 Classifier Output Potential Feedback Boundary Change ) . U A . ( t u p t u O r e i f i s s a C l 3 2 1 0 −1 −2 −30 Q3 PIO Q1 PIO −25 −20 −15 Seconds −10 −5 0 Figure 8. Figure shows that the output of workload classifier can be used to track PIO susceptibility in real-time. PIO susceptibility was estimated using a metric (0,1) based on joystick input (see Results). The last 5 seconds of each trial i that ended in the medium sized ring (40-60s after first ring is crossed) were analyzed for PIO susceptibility; maximum value Mi and time of maxima Ti were computed. 'Q3 PIO' reflects the interpolated classifier output in the 30 seconds leading up to Ti averaged over all trials where Mi > 0.75. 'Q1 PIO' reflects similar information for all trials where Mi < 0.25. This data suggests that even though the LvMS classifier only learns to differentiate EEG signals from different boundary conditions during trials, it can also differentiate piloting behavior. Note that data from -5s to 0s in the figure is guaranteed to reflect classifier output generated only within Medium glide path boundaries. Error-bars reflect mean ± SEM across subjects. ) z ( e z S i l i p u P 1 0.8 0.6 0.4 0.2 0 −0.2 −0.4 Pupil Size Comparison Large Medium Small A B ) z ( t u p t u o r e i f i s s a c l 2 1 0 −1 −2 −3 −4 Pupil−EEG relationship Large Medium Small −2 0 pupil size (z) 2 4 Figure 9: Pupillometry results. A) Data show an increase in pupil size (z-scored across each subject's data) with higher workload (smaller boundaries). B) The correlation between trial averaged classifier output (LvMS, full spectrum) and trial averaged pupil size, across all subjects (each datapoint represents a single trial). Overall correlation r=-0.1995, SEM=0.0764. 'Large' class correlation r = -0.0198, 'Small' class correlation r = -0.1210, difference n.s. A B ) s / g e d ( e t a r h c t i P 60 50 40 30 20 10 0 −10 0 0.5 1 ) x ( h t p e d 50 0 −50 0 0.5 ) y ( n o i t i s o p l a c i t r e v 1200 1000 800 0 0.5 1.5 2 2.5 Time from stick movement (s) view from above 1.5 horizontal position (z) view from the side 1 1 horizontal position (z) 1.5 3 3.5 4 2 2 2.5 x 104 2.5 x 104 Supplemental Figure 1. A) Sustained pitch (y) response of the virtual aircraft to full joystick input (at time 0s on x-axis). The initial delay between input onset and response onset could either be 0.2s or 0.3s (0.3s in the figure). B) Full flight trajectory represented by the position and the size of glide boxes. In the view from the side, solid line shows the mean path through the center of glide boxes, while dotted lines describe glide box boundaries. All dimensions are in meters. Virtual aircraft moved steadily in z-axis @ 250 m/s and could be controlled in y (pitch) axis via joystick input. Half of all trials for each subject has this glide path trajectory, and the other half at trajectory shown in figure 1b. Affect of Spectrogram Time Window on Classification The effect of spectrogram window on classification accuracy Even Sampled Stick Locked 0.8 C U A 0.75 0.7 0.65 0.6 0.55 0.5 500 700 1100 900 Window length (ms) 1300 1500 1700 1900 Supplemental Figure 2. The effect of increasing the length of EEG epochs, used to compute spectral features for each datapoint, on classification accuracy. AUC values asymptote around 1500ms, for both regularly sampled as well as stick locked classifier. Error-bars reflect SEM across 9 subjects. The effect of spectral averaging on classifier performance 0.9 Comparison of Averaging Across Frequencies Individual Freqs Averaged Across Freqs z A 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 All Bands Delta Theta Alpha Beta Gamma Supplemental Figure 3. Figure shows the comparison of AUC values for classifiers that used power of individual frequencies in each band as features versus classifiers that averaged power across frequencies in individual bands. The figure shows that the relatively higher contribution of Beta and Gamma band to overall classifier performance is not due to higher dimensionality of feature space when using individual frequencies as features. Error- bars reflect SEM across 9 subjects. C U A 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 1 A B Classifier Accuracy All Bands Delta Only Theta Only Alpha Only Beta Only Gamma Only 2 3 4 7 8 9 5 6 Subjects (in descending order of AUC) Average weights for regularly sampled -0.5 -0.25 0 0.25 0.5 FFT window center (s), 0.5s window size Supplemental Figure 4. A) AUC values for all subjects (in descending order of all band AUC), when using information from all spectral bands for regularly sampled data classification. Classification was performed using 64-channel MARA-cleaned EEG signals -1.5s epochs acquired every 2s - that were labeled according to the size of the nearest glide path boundary at the time of their generation. Also shown as AUC values for same subjects when using only individual bands (delta, theta, alpha, beta, or gamma). B) Subject- averaged scalp distribution of normalized weights for the classifying hyperplane that best separated Large boundary from Medium and Small boundaries (scalp map corresponds to 'All Bands' classifier in panel A). Average weights for stick locked classifier A B Window center (s) Average weights for regularly sampled classifier Window center (s) Supplemental Figure 5. A) Classification was performed using 64-channel 1.5s MARA-cleaned EEG epochs around each joystick movement, and were labeled according to the size of the nearest glide path boundary when they were generated. The Panel shows the scalp distribution of normalized weights for the classifying hyperplane that best separated Large boundary from Medium and Small boundaries. Delta and theta band activity from Fronto-central electrodes, as well as significant gamma band activity from a lateralized somatosensory topography, seems to contribute the most to classification accuracy. B) Figures show similar information as panel A, except that the classified EEG epochs that were acquired at regular intervals (2s). A B y1 Az(w+t) C1 Az(w) S M L θ Az(t) y2 C2 Supplemental Figure 6. Panel A shows schematic representation of hypothesized relationship of time-on-task (ToT) related signal to overall classification accuracy. Let C1 and C2 be two classifying features, e.g. power of a specific frequency at two electrodes. Let L, M, S be the data cloud generated while the subject operates under Large, Medium, and Small boundaries respectively. Dark colored points depict data sampled in the early half (first 15s) of flight within the respective boundary size, while light colored points depict the late half. Here, we assume that the classification accuracy of a workload classifier (L vs M/S) with AUC Az(w+t) has a time-on-task component with AUC - Az(t), i.e. a contribution of time since the start of the trial, that is unrelated to workload (or boundary change). Ideally, we would like to assess classification accuracy Az(w) of pure workload component. Let the vector normal to the classifying hyperplane for a workload classifier (L vs M/S) be y1, and the vector normal to the classifying hyperplane for a time-on-task classifier (L early vs L late) be y2. Therefore, a large angle between y1 and y2 would imply a small contribution of time-on-task component to the accuracy of boundary based classifier for workload. Indeed, the scatterplot in Panel B shows that the angle between vectors normal to the hyperplanes for workload classifier and ToT classifier is quite large for each subject. Consequently, correcting for ToT component (through projection) produces negligible effect on AUC (Az) values.
1501.01863
1
1501
2015-01-08T14:19:23
Estimating Information-Theoretic Quantities
[ "q-bio.NC" ]
Information theory is a practical and theoretical framework developed for the study of communication over noisy channels. Its probabilistic basis and capacity to relate statistical structure to function make it ideally suited for studying information flow in the nervous system. It has a number of useful properties: it is a general measure sensitive to any relationship, not only linear effects; it has meaningful units which in many cases allow direct comparison between different experiments; and it can be used to study how much information can be gained by observing neural responses in single trials, rather than in averages over multiple trials. A variety of information theoretic quantities are in common use in neuroscience - (see entry "Summary of Information-Theoretic Quantities"). Estimating these quantities in an accurate and unbiased way from real neurophysiological data frequently presents challenges, which are explained in this entry.
q-bio.NC
q-bio
Estimating Information-Theoretic Quantities   Robin  A.A.  Ince1,  Simon  R.  Schultz2  and  Stefano  Panzeri1,3   1  Institute  of  Neuroscience  and  Psychology,  58  Hillhead  Street,  University  of  Glasgow,  Glasgow  G12  8QB,  UK     2  Department  of  Bioengineering,  Imperial  College  London,  South  Kensington,  London  SW7  2AZ,  UK   3  Center  For  Neuroscience  and  Cognitive  Systems,  Italian  Institute  of  Technology,  Corso  Bettini  31   –  38068  Rovereto   (Tn)  Italy       16  pages  5724  words.   Definition Information  theory  is  a  practical  and  theoretical  framework  developed  for  the  study  of  communication  over   noisy  channels.  Its  probabilistic  basis  and  capacity  to  relate  statistical  structure  to  function  make  it  ideally   suited   for   studying   information   flow   in   the   nervous   system.   It   has   a   number   of   useful   properties:   it   is   a   general  measure  sensitive  to  any  relationship,  not  only  linear  effects;  it  has  meaningful  units  which  in  many   cases   allow   direct   comparison   between   different   experiments;   and   it   can   be   used   to   study   how   much   information   can   be   gained   by   observing   neural   responses   in   single   trials,   rather   than   in   averages   over   multiple  trials.  A  variety  of  information  theoretic  quantities  are  in  common  use  in  neuroscience  –  (see  entry   “Summary  of  Information-­‐Theoretic  Quantities”).  Estimating  these  quantities  in  an  accurate  and  unbiased   way  from  real  neurophysiological  data  frequently  presents  challenges,  which  are  explained  in  this  entry.   Detailed Description Information theoretic quantities Information   theory   provides   a   powerful   theoretical   framework   for   studying   communication   in   a   quantitative  way.  Within  the  framework  there  are  many  different  quantities  to  assess  different  properties   of   systems,   many   of   which   have   been   applied   fruitfully   to   analysis   of   the   nervous   system   and   neural   communication.  These  include  the  entropy  (H;  which  measures  the  uncertainty  associated  with  a  stochastic   variable),   mutual   information   (I;   which   measures   the   dependence   relationship   between   two   stochastic   variables),   and   several   more   general   divergence   measures   for   probability   distributions   (Kullbeck-­‐Liebler,   Jenson-­‐Shannon).   For   full   definitions   and   details   of   these   quantities,   see   the   "Summary   of   Information-­‐ Theoretic  Quantities"  entry.     The  Finite  Sampling  Problem     A   major   difficulty   when   applying   techniques   involving   information   theoretic   quantities   to   experimental   systems  is  that  they  require  measurement  of  the  full  probability  distributions  of  the  variables  involved.  If   we  had  an  infinite  amount  of  data,  we  could  measure  the  true  stimulus-­‐response  probabilities  precisely.   However,  any  real  experiment  only  yields  a  finite  number  of  trials  from  which  these  probabilities  must  be   estimated.  The  estimated  probabilities  are  subject  to  statistical  error  and  necessarily  fluctuate  around  their   true  values.  The  significance  of  these  finite  sampling  fluctuations  is  that  they  lead  to  both  statistical  error   (variance)   and   systematic   error   (called   limited   sampling   bias)   in   estimates   of   entropies   and   information.   This   bias   is   the   difference   between   the   expected   value   of   the   quantity   considered,   computed   from   probability   distributions   estimated   with   N   trials   or   samples,   and   its   value   computed   from   the   true   probability   distribution.   This   is   illustrated   in   Figure   1,   which   shows   histogram   estimates   of   the   response   distribution  for  two  different  stimuli  calculated  from  40  trials  per  stimulus.  While  the  true  probabilities  are   uniform  (dotted  black  lines)  the  calculated  maximum  likelihood  estimates  show  some  variation  around  the   true   values.   This   causes   spurious   differences   –   in   this   example   the   sampled   probabilities   indicate   that   obtaining   a   large   response   suggests   stimulus   1   is   presented   –   which   results   in   a   positive   value   of   the   information.   Figure   1C   shows   a   histogram   of   information   values   obtained   from   many   simulations   of   this   system.  The  bias  constitutes  a  significant  practical  problem,  because  its  magnitude  is  often  of  the  order  of   the  information  values  to  be  evaluated,  and  because  it  cannot  be  alleviated  simply  by  averaging  over  many   neurons  with  similar  characteristics.   Direct  estimation  of  information  theoretic  quantities   The  most  direct  way  to  compute  information  and  entropies  is  to  estimate  the  response  probabilities  as  the   experimental  histogram  of  the  frequency  of  each  response  across  the  available  trials,  and  then  plug  these   empirical  probability  estimates  into  Eqs.  (1-­‐3).  We  refer  to  this  as  the  “plug-­‐in”  method.   The   plug-­‐in   estimate   of   entropy   is   biased   downwards   (estimated   values   are   lower   than   the   true   value)   while  the  plug-­‐in  estimate  of  information  shows  an  upward  bias.  An  intuitive  explanation  for  why  entropy  is   biased   downwards   comes   from   the   fact   that   entropy   is   a   measure   of   variability.   As   we   saw   in   Figure   1,   variance   in   the   maximum   likelihood   probability   estimates   introduces   extra   structure   in   the   probability   distributions;   this   additional   variation   (which   is   dependent   on   the   number   of   trials   from   the   asymptotic   normality  of  the  ML  estimate)  always  serves  to  make  the  estimated  distribution  less  uniform  or  structured   than  the  true  distribution.  Consequently,  entropy  estimates  are  lower  than  their  true  values,  and  the  effect   of  finite  sampling  on  entropies  is  a  downward  bias.  For  information,  an  explanation  for  why  bias  is  typically   positive   is   that   finite   sampling   can   introduce   spurious   stimulus-­‐dependent   differences   in   the   response   probabilities,  which  make  the  stimuli  seem  more  discernible  and  hence  the  neuron  more  informative  than   it  really  is.  Alternatively,  viewing  the  information  as  a  difference  of  unconditional  and  conditional  entropies,   (first  two  expressions  of  Eq.  in  “Summary  of  Information  Theoretic  Quantities”)  the  conditional  entropy  for   each  stimulus  value  is  estimated  from  a  reduced  number  of  trials  compared  to  the  noise  entropy  (only  the   trials  corresponding  to  that  stimulus)  and  so  its  bias  is  considerably  larger.  Since  entropy  is  biased  down   and  the  conditional  entropy  term  appears  with  a  negative  sign,  this  causes  the  upward  bias  in  information.     In   the   above   explanations   for   the   source   of   the   bias,   the   source   is   the   variability   in   estimates   of   the   probability   distributions   considered.   Variance   in   the   maximum   likelihood   histogram   estimators   of   the   probabilities  induces  additional  “noise”  structure  affecting  the  entropy  and  information  values.     Asymptotic  estimates  of  the  limited  sampling  bias   To   understand   the   sampling   behaviour   of   information   and   entropy   better,   it   is   useful   to   find   analytical   approximations   to   the   bias.   This   can   be   done   in   the   so-­‐called   “asymptotic   sampling   regime”.   Roughly   speaking  this  is  when  the  number  of  trials  is  “large”.  More  rigorously,  the  asymptotic  sampling  regime  is   defined  as  N  being  large  enough  that  every  possible  response  occurs  many  times:  that  is,  Ns  P(rs)  >>  1  for   each   stimulus-­‐response   pair   s,r   such   that   P(rs)   >   0.   In   this   regime,   the   bias   of   the   entropies   and   information   can   be   expanded   in   inverse   powers   of   1/N   and   analytical   approximations   obtained   (Miller,   1955;  Panzeri  and  Treves,  1996).  The  leading  terms  in  the  biases  are  respectively:   BIAS H R ( [ ) ] =− BIAS H R S ( [ ) BIAS I S R ( ; ) [ ] ⎤ 1 ⎦ ∑ s R s ⎡ ⎣ ⎤ 1 ⎦ R ⎡ ⎣ =− 1 − N 2 ln(2) 1 − ] N 2 ln(2) 1 ⎧ −⎨ =− − N 2 ln(2) ⎩ ∑ R s ⎡ ⎣ ⎤ 1 ⎦ R ⎡ ⎣ ⎤ 1 ⎦ ⎫ ⎬ ⎭ s       (1)   sR denotes   the   number   of   relevant   responses   for   the   stimulus   conditional   response   probability   Here   distribution  P(rs)  –  i.e.  the  number  of  different  responses  r  with  non-­‐zero  probability  of  being  observed   when  stimulus  s  is  presented.   R  denotes  the  number  of  relevant  responses  for  P(r)  –  i.e.  the  number  of   different   responses   r   with     non-­‐zero   probability   of   being   observed   across   all   stimuli.   In   practice,   R  is   usually  going  to  be  equal  to  the  number  of  elements  constituting  the  response  space  R.  If  a  response  never   happens  across  all  stimuli,  it  can  be  removed  from  the  sum  over  r  in  the  calculation  of  each  entropy  term,   sR may  be  different  from R  if  some  responses  occur   and  thus  removed  from  the  response  set  R.  However,   only  with  particular  stimuli.   Although  valid  only  in  the  asymptotic  regime,  Eq.  (1)  sheds  valuable  light  on  the  key  factors  that  control  the   bias.  First,  Eq.  (1)  shows  that  the  bias  of   H(R S)  is  approximately  S  times  bigger  than  that  of   H(R) .  This   means  that,  in  the  presence  of  many  stimuli,  the  bias  of   I(S;R)  is  similar  to  that  of   H(R S) .  However,   I(S;R)  is  a  difference  of  entropies,  and  its  typical  values  are  much  smaller  than  those  of   H(R S).  This   implies  that  bias  correction  methods  must  be  validated  on  the  performance  of  information  and  not  only  on   entropies,   because,   in   many   cases,   the   bias   may   be   proportionally   negligible   for   entropies   but   not   the   information.  Second,  Eq.   (1)  shows  that  the  bias  is  small  when  the  ratio  Ns/ R  is  big,  i.e.  more  trials  per   stimulus   than   possible   responses.   This   is   because,   assuming   that   the   number   of   trials   per   stimulus   is   sR is   approximately   equal   to R ,   the   bias   of   H(R S)  is   approximately   – R approximately   constant   and   /[2Nsln(2)].  Thus,  Ns/ R  is  the  crucial  parameter  for  the  sampling  problem.  For  example,  in  the  simulations   of  Fig.  2,  with   R =28  the  bias  of   I(S;R)  became  negligible  for  Ns  ≥  213  (i.e.  Ns/ R  ≥  32).    Second,  Eq.   (1)   shows  that,  even  if  Ns  is  constant,  the  bias  increases  with  the  number  of  responses   R .  This  has  important   implications  for  comparing  neural  codes.  For  example,  a  response  consisting  of  a  given  number  of  spikes   can  arise  from  many  different  possible  temporal  patterns  of  spikes.  Thus,   R  is  typically  much  greater  for  a   spike  timing  code  than  for  a  spike  count  code,  and  it  follows  from  Eq.  (1)  that  the  estimate  of  information   conveyed  by  a  spike  timing  code  is  more  biased  than  that  measured  for  the  same  neurons  through  a  spike   count   code.     If   bias   is   not   eliminated,   there   is   therefore   a   danger   of   concluding   that   spike   timing   is   important  even  when  it  is  not.             A   further   important   feature   of   Eq.   (1)   is   that,   although   the   bias   is   not   the   same   for   all   probability   distributions,   in   the   asymptotic   sampling   regime   it   depends   on   some   remarkably   simple   details   of   the   response  distribution  (the  number  of  trials  and  the  number  of  relevant  responses).    Thus  Eq.  (1)  makes  it  is   possible  to  derive  simple  rules  of  thumb  for  estimating  the  bias  magnitude  and  compare  the  relative  bias  in   different  situations.  As  detailed  in  the  next  section,  the  simplicity  of  Eq.  (1)  can  also  be  exploited  in  order  to   correct  effectively  for  the  bias  (Panzeri  and  Treves,  1996).     Bias  Correction  Methods   A  number  of  techniques  have  been  developed  to  address  the  issue  of  bias,  and  allow  much  more  accurate   estimates  of  information  theoretic  quantities  than  the  “plug-­‐in”  method  described  above.  (Panzeri  et  al.,   2007)  and  (Victor,  2006)  provide  a  review  of  such  methods,  a  selection  of  which  are  briefly  outlined  here.     Panzeri–Treves  (PT)   In  the  so-­‐called  asymptotic  sampling  regime,  when  the  number  of  trials  is  large  enough  that  every  possible   response  occurs  many  times,  an  analytical  approximation  for  the  bias  (i.e.  the  difference  between  the  true   value  and  the  plug-­‐in  estimate)  of  entropies  and  information  can  be  obtained  (Miller,  1955;  Panzeri  and   Treves,  1996)  (Eq.  (1)).  The  value  of  the  bias  computed  from  the  above  expressions  is  then  subtracted  from   the   plug-­‐in   estimate   to   obtain   the   corrected   value.   This   requires   an   estimate   of   the   number   of   relevant   sR  by  the  count  of  responses   responses  for  each  stimulus,   that  are  observed  at  least  once  –  this  is  the  “naive”  count.  However  due  to  finite  sampling  this  will  be  an   underestimate  of  the  true  value.  A  Bayesian  procedure  (Panzeri  and  Treves,  1996)  can  be  used  to  obtain  a   more  accurate  value.   sR .  The  simplest  approach  is  to  approximate   Quadratic  Extrapolation  (QE)   In  the  asymptotic  sampling  regime,  the  bias  of  entropies  and  information  can  be  approximated  as  second   order  expansions  in  1/N,  where  N  is  the  number  of  trials  (Treves  and  Panzeri,  1995;  Strong  et  al.,  1998).  For   example,  for  information:     I R S ( ; ) = I true R S ( ; ) + plugin a b N N 2 +     (2)   This  property  can  be  exploited  by  calculating  the  estimates  with  subsets  of  the  original  data,  with  N/2  and   N/4  trials  and  fitting  the  resulting  values  to  the  polynomial  expression  above.  This  allows  an  estimate  of  the   parameters  a  and  b  and  hence  Itrue(S;  R).  To  use  all  available  data,  estimates  of  two  subsets  of  size  N/2  and   four   subsets   of   size   N/4   are   averaged   to   obtain   the   values   for   the   extrapolation.   Together   with   the   full   length  data  calculation,  this  requires  seven  different  evaluations  of  the  quantity  being  estimated.   Nemenman–Shafee–Bialek  (NSB)   The  NSB  method  (Nemenman  et  al.,  2002,  2004)  utilises  a  Bayesian  inference  approach  and  does  not  rely   on  the  assumption  of  the  asymptotic  sampling  regime.  It  is  based  on  the  principle  that  when  estimating  a   quantity,  the  least  bias  will  be  achieved  when  assuming  an  a  priori  uniform  distribution  over  the  quantity.   This   method   is   more   challenging   to   implement   than   the   other   methods,   involving   a   large   amount   of   function   inversion   and   numerical   integration.   However,   it   often   gives   a   significant   improvement   in   the   accuracy  of  the  bias  correction  (Montani  et  al.,  2007,  2007;  Montemurro  et  al.,  2007;  Nemenman  et  al.,   2008).   Shuffled  Information  Estimator  (Ish)   Recently,  an  alternative  method  of  estimating  the  mutual  information  has  been  proposed  (Montemurro  et   al.,  2007;  Panzeri  et  al.,  2007).  Unlike  the  methods  above,  this  is  a  method  for  calculating  the  information   only,  and  is  not  a  general  entropy  bias  correction.  However,  it  can  be  used  with  the  entropy  corrections   described  above  to  obtain  more  accurate  results.  In  brief,  the  method  uses  the  addition  and  subtraction  of   different  terms  (in  which  correlations  given  a  fixed  stimulus  are  removed  either  by  empirical  shuffling  or  by   marginalizing  and  then  multiplying  response  probabilities)  that  do  not  change  the  information  value  in  the   limit  of  an  infinite  number  of  trials,  but  do  remove  a  large  part  of  the  bias  for  finite  number  of  trials.  This   greatly  reduces  the  bias  of  the  information  estimate,  at  the  price  of  only  a  marginal  increase  of  variance.   Note   that   this   procedure   works   well   for   weakly   correlated   data,   which   is   frequently   the   case   for   simultaneously  recorded  neurons.       Examples  of  bias  correction  performance   Figure  2A  reports  the  results  of  the  performance  of  bias  correction  procedures,  both  in  terms  of  bias  and   Root  Mean  Square  (RMS)  error,  on  a  set  of  simulated  spike  trains  from  eight  simulated  neurons.  Each  of   these  neurons  emitted  spikes  with  a  probability  obtained  from  a  Bernoulli  process.  The  spiking  probabilities   were   set   equal   to   those   measured,   in   the   10–15   ms   post-­‐stimulus   interval,   from   eight   neurons   in   rat   somatosensory  cortex  responding  to  13  stimuli  consisting  of  whisker  vibrations  of  different  amplitude  and   frequency  (Arabzadeh  et  al.,  2004).  The  10–15  ms  interval  was  chosen  since  it  was  found  to  be  the  interval   containing   highest   information   values.   Figure   2A   shows   that   all   bias   correction   procedures   generally   improve  the  estimate  of  I(S;R)  with  respect  to  the  plug-­‐in  estimator,  and  the  NSB  correction  is  especially   effective.     Figure  2B  shows  that  the  bias-­‐corrected  estimation  of  information  (and  RMS  error;  lower  panel)  is  much   improved  by  using  Ish(R;S)  rather  than  I(R;S).  The  use  of  Ish(R;S)  makes  the  residual  errors  in  the  estimation   of  information  much  smaller  and  almost  independent  from  the  bias  correction  method  used.  Taking  into   account   both   bias   correction   performance   and   computation   time,   for   this   simulated   system   the   best   method   to   use   is   the   shuffled   information   estimator   combined   with   the   Panzeri–Treves   analytical   correction.   Using   this,   an   accurate   estimate   of   the   information   is   possible   even   when   the   number   of   samples  per  stimulus  is  R  where  R  is  the  dimension  of  the  response  space.   It   should   be   noted   that   while   bias   correction   techniques   such   as   those   discussed   above   are   essential   if   accurate  estimation  of  the  information  is  required,  whether  to  employ  them  depends  upon  the  question   being   addressed.   If   the   aim   is   quantitative   comparison   of   information   values   between   different   experiments,  stimuli  or  behaviours,  they  are  essential.  If  instead  the  goal  is  simply  to  detect  the  presence  of   a   significant   interaction,   the   statistical   power   of   information   as   a   test   of   independence   is   greater   if   the   uncorrected  plug-­‐in  estimate  is  used  (Ince  et  al.,  2012).  This  is  because  all  the  bias  correction  techniques   introduce   additional   variance;   for   a   hypothesis   test   which   is   unaffected   by   underlying   bias,   this   variance   reduces  the  statistical  power.   Information  Component  Analysis  Techniques     The   principal   use   of   Information   Theory   in   neuroscience   is   to   study   the   neural   code   –   and   one   of   the   commonly   studied   problems   in   neural   coding   is   to   take   a   given   neurophysiological   dataset,   and   to   determine   how   the   information   about   some   external   correlate   is   represented.   One   way   to   do   this   is   to   compute  the  mutual  information  that  the  neural  responses  convey  on  average  about  the  stimuli  (or  other   external  correlate),  under  several  different  assumptions  about  the  nature  of  the  underlying  code  (e.g.  firing   rate,  spike  latency,  etc).  The  assumption  is  that  if  one  code  yields  substantially  greater  mutual  information,   then  downstream  detectors  are  more  likely  to  be  using  that  code  (the  Efficient  Coding  Hypothesis).   A   conceptual   improvement   on   this   procedure,   where   it   can   be   done,   is   to   take   the   total   information   available  from  all  patterns  of  spikes  (and  silences),  and  to  break  it  down  (mathematically)  into  components   reflecting   different   encoding   mechanisms,   such   as   firing   rates,   pairwise   correlations,   etc   (Panzeri   et   al.,   1999;  Panzeri  and  Schultz,  2001).  One  way  to  do  this  is  to  perform  an  asymptotic  expansion  of  the  mutual   information,  grouping  terms  in  such  a  way  that  they  reflect  meaningful  coding  mechanisms.   Taylor  series  expansion  of  the  mutual  information   A  taylor  series  expansion  of  the  mutual  information  is  one  such  approach.  For  short  time  windows  (and  we   will  discuss  presently  what  “short”  means),  the  mutual  information  can  be  approximated  as     I(R;S) = TIt + T 2 2 Itt +...     (3)   where  the  subscript  t  indicates  the  derivative  with  respect  to  the  time  window  length  T.  For  time  windows   sufficiently  short  that  only  a  pair  of  spikes  are  contained  within  the  time  window  (either  from  a  population   of   cells   or   from   a   single   spike   train),   the   first   two   terms   are   all   that   is   needed.   If   the   number   of   spikes   exceeds  two,  it  may  still  be  a  good  approximation,  but  higher  order  terms  would  be  needed  to  capture  all   of   the   information.   One   possibility   is   that   Equation   (3)   could   be   extended   to   incorporate   higher   order   terms,   however   we   have   instead   found   it   better   in   practice   to   take   an   alternative   approach   (see   next   section).   With  only  a  few  spikes  to  deal  with,  it  is  possible  to  use  combinatorics  to  write  out  the  expressions  for  the   probabilities  of  observing  different  spike  patterns.  This  was  initially  done  for  the  information  contained  in   the   spikes   fired   by   a   small   population   of   cells   (Panzeri   et   al.,   1999),   and   then   later   extended   to   the   information  carried  by  the  spatiotemporal  dynamics  of  a  small  population  of  neurons  over  a  finite  temporal   wordlength   (Panzeri   and   Schultz,   2001).   In   the   former   formulation,   we   define ri(s)  to   be   the   mean   response  rate  (number  of  spikes  in  T  divided  by  T)  of  cell  i  (of  C  cells)  to  stimulus  s  over  all  trials  where  s   was  presented.  We  then  define  the  signal  cross-­‐correlation  density  to  be     νij = ri(s)rj(s) s ri(s) s rj(s) s −1  .   (4)   This  quantity  captures  the  correlation  in  the  mean  response  profiles  of  each  cell  to  different  stimuli  (e.g.   correlated  tuning  curves).  Analogously,  we  define  the  noise  cross-­‐correlation  density  to  be     γij(s) = ri(s)rj(s) ri(s)rj(s) −1  .   The  mutual  information  I(R;S)  can  then  be  written  as  the  sum  of  four  components,     with  the  first  component     I = Ilin + Isig-sim + Icor-ind + Icor-dep  ,   Ilin = T C ∑ i=1 ri(s)log2 ri(s) ri(s') s'     s (5)   (6)   (7)   capturing   the   rate   component   of   the   information,   i.e.   that   which   survives   when   there   is   no   correlation   between  cells  (even  of  their  tuning  curves)  present  at  all.  This  quantity  is  positive  semi-­‐definite,  and  adds   linearly  across  neurons.  It  is  identical  to  the  “information  per  spike”  approximation  calculated  by  a  number   of  authors  (Skaggs  et  al.,  1993;  Brenner  et  al.,  2000;  Sharpee  et  al.,  2004).  Isig-­‐sim  is  the  correction  to  the   information  required  to  take  account  of  correlation  in  the  tuning  of  individual  neurons,  or  signal  similarity:     Isig-sim = T 2 2ln2 C ∑ i=1 C ∑ j=1 ri(s) s rj(s) s ⎡ νij +(1+νij)ln ⎢ ⎢ ⎣ ⎛ ⎝⎜ 1 1+νij ⎞ ⎠⎟ ⎤ ⎥ ⎥ ⎦     (8)   This   is   negative   semi-­‐definite.   Icor-­‐ind   is   the   effect   on   the   transmitted   information   of   the   average   level   of   noise  correlation  (correlation  at  fixed  stimulus)  between  neurons:   1 1+νij ri(s)rj(s)γij(s) s Icor-ind = T 2 2 ∑ i=1 ∑ j=1 ⎞ ⎠⎟  .   (9)     C C log2 ⎛ ⎝⎜ Icor-­‐ind  can  take  either  positive  or  negative  values.  Icor-­‐dep  is  the  contribution  of  stimulus-­‐dependence  in  the   correlation   to   the   information   –   as   would   be   present,   for   instance,   if   synchrony   was   modulated   by   a   stimulus  parameter:     Icor-dep = T 2 2 C ∑ i=1 C ∑ j=1 ri(s)rj(s) 1+γij(s) ( )log2 ⎡ ⎢ ⎢ ⎣ ( ) ri(s')rj(s') s' 1+γij(s) ) s' ri(s')rj(s') 1+γij(s') ( ⎤ ⎥ ⎥ ⎦     s (10)   An   application   of   this   approach   to   break   out   components   reflecting   rate   and   correlational   coding   is   illustrated  in  Figure  3,  by  application  to  simulated  spike  trains  with  controlled  correlation.  This  approach   has  been  used  by  a  number  of  authors  to  dissect  out  contributions  to  neural  coding  (e.g.  Scaglione  et  al.,   2008).   This  approach  extends  naturally  to  the  consideration  of  temporal  correlations  between  spike  times  within  a   spike  train  (Panzeri  and  Schultz,  2001).  Note  that  the  Taylor  series  approach  can  be  applied  to  the  entropy   as  well  as  the  mutual  information  –  this  was  used  to  break  out  the  effect  of  spatiotemporal  correlations  on   the  entropy  of  a  small  population  of  neural  spike  trains  (Schultz  and  Panzeri,  2001).  While  the  Taylor  series   approach  can  be  extremely  useful  in  teasing  out  contributions  of  different  mechanisms  to  the  transmission   of  information,  it  is  not  recommended  as  a  method  for  estimation  of  the  total  information,  as  the  validity  of   the   approximation   can   break   down   quickly   as   the   time   window   grows   beyond   that   sufficient   to   contain   more  than  a  few  spikes  from  the  population.  It  is  however  useful  as  an  additional  inspection  tool  after  the   total   information   has   been   computed,   which   allows   the   information   to   be   broken   down   into   not   only   mechanistic  components  but  also  their  contributions  from  individual  cells  and  time  bins  (for  instance  by   visualizing  the  quantity  after  the  summations  in  Equation  (10)  as  a  matrix).   A  more  general  information  component  analysis  approach   A  limitation  of  the  Taylor  series  approach  is  that  it  is  restricted  to  the  analysis  of  quite  short  time  windows   of   neural   responses.   However,   an   exact   breakdown   approach   allowed   a   very   similar   information   component  approach  to  be  used  (Pola  et  al.,  2003).  In  this  approach,  we  consider  a  population  response   (spike  pattern)  r,  and  define  the  normalized  noise  cross-­‐correlation  to  be     γ(r s) = P(r s) Pind(r s) −1 if Pind(r s) ≠ 0 if Pind(r s) = 0 0  .   ⎧ ⎪ ⎨ ⎪ ⎩  where  the  marginal  distribution   Pind(r s) = P(rc s) . C ∏ c=1 Compare  the  expressions  in  Equation  (5)  and  (12).  Similarly,  the  signal  correlation  becomes   (11)   (12)       ν(r) = Using   these   definitions,   the   approximation  of  short  time  windows)  as   c∏ P(rc) ≠ 0  .   (13)   Pind(r) c∏ −1 if P(rc) if 0 ⎧ ⎪⎪ ⎨ ⎪ ⎪ ⎩ information   components   can   now   be   written   exactly   (without   the   P(rc) = 0 c∏   Ilin = ∑ c ∑ rc P(rc s)log2 P(rc s) P(rc)     s (14)         Isig-sim = 1 ln2 ∑ r Icor-ind = ⎡ ν(r) +(1+ν(r))ln ⎢ ⎣ ⎞ P(rc) ⎠⎟ c∏⎛ ⎝⎜ Pind(r s)γ(r s) s r∑ log2 ⎛ ⎝⎜ 1+ν(r) ⎞ ⎠⎟     1 1+ν(r) ⎞ ⎠⎟ ⎤ ⎥     ⎦ ⎛ ⎝⎜ 1 Icor-dep = r∑ Pind(r s)(1+γ(r s))log2 Pind(r s') s'(1+γ(r s)) Pind(r s')(1+γ(r s')) s'  .   s (15)   (16)   (17)   This  latter  component  is  identical  to  the  quantity  “ ΔI ”  introduced  by  Latham  and  colleagues  (Nirenberg  et   al.,  2001)  to  describe  the  information  lost  due  to  a  decoder  ignoring  correlations.  It  is  zero  if  and  only  if   P(s r) = Pind(s r)  for  every  s  and  r.     The   expressions   shown   above   are   perhaps   the   most   useful   description   for   obtaining   insight   into   the   behavior  of  the  information  components  under  different  statistical  assumptions  relating  to  the  neural  code,   however   they   are   not   necessarily   the   best   way   to   estimate   the   individual   components.   However,   the   components   can   also   be   written   as   a   sum   of   entropies   and   entropy-­‐like   quantities,   which   can   then   be   computed   using   the   entropy   estimation   algorithms   described   earlier   in   this   chapter   (Pola   et   al.,   2003;   Montani  et  al.,  2007;  Schultz  et  al.,  2009).  Note  that,  as  shown  by  Scaglione  and  colleagues  (Scaglione  et  al.,   2008,  2010)  the  components  in  Eq.  (14-­‐17)  can,  under  appropriate  conditions,  be  further  decomposed  to   tease  apart  the  relative  role  of  autocorrelations  (spikes  from  the  same  cells)  and  cross-­‐correlations  (spikes   from  different  cells).     Maximum  entropy  approach  to  information  component  analysis   The  conceptual  basis  of  the  information  component  approach  is  to  make  an  assumption  that  constrains  the   statistics  of  the  response  distribution  P(rs),  compute  the  mutual  information  subject  to  this  assumption,   and  by  evaluating  the  difference  between  this  and  the  “full”  information,  calculate  the  contribution  to  the   information   of   relaxing   this   constraint.   By   making   further   assumptions,   the   contributions   of   additional   mechanisms   can   in   many   cases   be   dissected   out   hierarchically.   As   an   example,   by   assuming   that   the   conditional  distribution  of  responses  given  stimuli  is  equal  to  the  marginal  distribution   P(r s) = Pind(r s) ,   and  substituting  in  to  the  mutual  information  equation,  one  can  define  an  information  component  Iind.  This   component  can  then  be  further  broken  up     Iind = Ilin + Isig-sim  ,   (18)   with   Ilin   and   Isig-­‐sim   as   defined   in   the   previous   section.   The   correlational   component,   Icor,   is   then   just   the   difference  between  Iind  and  the  total  mutual  information.  This  can  also  be  further  broken  up,  as     Icor = Icor-ind + Icor-dep  .   (19)   This  approach  can  be  extended  further.  For  instance,  the  assumption  of  a  Markov  approximation  can  be   made,  with  a  memory  extending  back  q  timesteps,  and  the  information  computed  under  this  assumption   (Pola   et   al.,   2005).   More   generally,   any   simplified   model   can   be   used,   although   the   maximum   entropy   models  are  of  special  interest  (Montemurro  et  al.,  2007;  Schaub  and  Schultz,  2012)     Psimp(r s) = 1 Z exp λ0 −1+ λigi(r) ∑ i=1 m ⎧ ⎨ ⎩     ⎫ ⎬ ⎭ (20)   with   parameters  λi  implementing   a   set   of   constraints   reflecting   assumptions   made   about   what   are   the   important   (non-­‐noise)   properties   of   the   empirical   distribution,   and   the   partition   function   Z   ensuring   normalisation.   An   example   of   this   is   the   Ising   model,   which   has   been   used   with   some   success   to   model   neural  population  response  distributions  (Schneidman  et  al.,  2006;  Shlens  et  al.,  2006;  Schaub  and  Schultz,   2012).   Estimation   of   the   Ising   model   for   large   neural   populations   can   be   extremely   computationally   intensive  if  brute  force  methods  for  computing  the  partition  function  are  employed,  however  mean  field   approximations   can   be   employed   in   practice   with   good   results   (Roudi   et   al.,   2009;   Schaub   and   Schultz,   2012).   Binless  methods  for  estimating  information   In   applications   in   neuroscience,   typically   at   least   one   of   the   random   variables   (stimuli   or   responses)   is   discrete,  and  thus  the  approach  of  discretizing  one  or  more  of  the  variables  is  often  taken.  However,  where   stimuli  and  responses  are  both  continuous  (an  example  might  be  local  field  potential  responses  to  a  white   noise   sensory   stimulus),   it   may   be   advantageous   to   take   advantages   of   techniques   better   suited   to   continuous   signals,   such   as   kernel   density   estimators   (Moon   et   al.,   1995),   nearest   neighbor   estimators   (Kraskov   et   al.,   2004)   or   binless   metric   space   methods   (Victor,   2002).   We   refer   to   the   entry   “Bin-­‐Less   Estimators  for  Information  Quantities”  for  an  in-­‐depth  discussion  of  these  techniques.     Calculation  of  information  theoretic  quantities  from  parametric  models   An  alternative  to  measuring  information-­‐theoretic  quantities  directly  from  observed  data  is  to  fit  the  data   to  a  model,  and  then  to  either  analytically  or  numerically  (depending  upon  the  complexity  of  the  model)   compute   the   information   quantity   from   the   model.   Examples   include   analytical   calculations   of   the   information  flow  through  models  of  neural  circuits  with  parameters  fit  to  anatomical  and  physiological  data   (Treves   and   Panzeri,   1995;   Schultz   and   Treves,   1998;   Schultz   and   Rolls,   1999),   and   parametric   fitting   of   probabilistic   models   of   neural   spike   firing   (Paninski,   2004;   Pillow   et   al.,   2005,   2008),   of   spike   count   distributions   (Gershon   et   al.,   1998;   Clifford   and   Ibbotson,   2000),   or   of   neural   mass   signals   (Magri   et   al.,   2009).  It  should  come  as  no  surprise  that  “there  is  no  free  lunch”,  and,  although  in  principle  information   quantities  can  be  computed  exactly  under  such  circumstances,  the  problem  is  moved  to  one  of  assessing   model  validity  –  an  incorrect  model  can  lead  to  a  bias  in  either  direction  in  the  information  computed.   Acknowledgements   Research   supported    by   the   SI-­‐CODE   (FET-­‐Open,   FP7-­‐284533)   project   and   by   the    ABC   and   NETT   (People   Programme   and   PITN-­‐GA-­‐2011-­‐289146)     projects   of    the    European    Union's  Seventh  Framework    Programme    FP7  2007-­‐2013.    Marie     Curie     Actions  PITN-­‐GA-­‐2011-­‐290011   References   Arabzadeh   E,   Panzeri   S,   Diamond   ME   (2004)   Whisker   Vibration   Information   Carried   by   Rat   Barrel   Cortex   Neurons.  J  Neurosci  24:6011–6020.   Brenner  N,  Strong  SP,  Koberle  R,  Bialek  W,  Steveninck  RRR  (2000)  Synergy  in  a  neural  code.  Neural  Comput   12:1531–1552.   Clifford  CW  g.,  Ibbotson  MR  (2000)  Response  variability  and  information  transfer  in  directional  neurons  of   the  mammalian  horizontal  optokinetic  system.  Vis  Neurosci  17:207–215.   Gershon   ED,   Wiener   MC,   Latham   PE,   Richmond   BJ   (1998)   Coding   Strategies   in   Monkey   V1   and   Inferior   Temporal  Cortices.  J  Neurophysiol  79:1135–1144.   Ince  RAA,  Mazzoni  A,  Bartels  A,  Logothetis  NK,  Panzeri  S  (2012)  A  novel  test  to  determine  the  significance  of   neural   selectivity   to   single   and   multiple   potentially   correlated   stimulus   features.   J   Neurosci   Methods  210:49–65.   Ince  RAA,  Mazzoni  A,  Petersen  RS,  Panzeri  S  (2010)  Open  source  tools  for  the  information  theoretic  analysis   of  neural  data.  Front  Neurosci  4:62–70.   Kennel   MB,   Shlens   J,   Abarbanel   HDI,   Chichilnisky   E   (2005)   Estimating   entropy   rates   with   Bayesian   confidence  intervals.  Neural  Comput  17:1531–1576.   Kraskov  A,  St\ögbauer  H,  Grassberger  P  (2004)  Estimating  mutual  information.  Phys  Rev  E  69:66138.   Magri  C,  Whittingstall  K,  Singh  V,  Logothetis  NK,  Panzeri  S  (2009)  A  toolbox  for  the  fast  information  analysis   of  multiple-­‐site  LFP,  EEG  and  spike  train  recordings.  BMC  Neurosci  10:81.   Miller  G  (1955)  Note  on  the  bias  of  information  estimates.  Inf  Theory  Psychol  Probl  Methods:95–100.   Montani  F,  Kohn  A,  Smith  MA,  Schultz  SR  (2007)  The  Role  of  Correlations  in  Direction  and  Contrast  Coding   in  the  Primary  Visual  Cortex.  J  Neurosci  27:2338.   Montemurro  MA,  Senatore  R,  Panzeri  S  (2007)  Tight  Data-­‐Robust  Bounds  to  Mutual  Information  Combining   Shuffling  and  Model  Selection  Techniques.  Neural  Comput  19:2913–2957.   Moon  YI,  Rajagopalan  B,  Lall  U  (1995)  Estimation  of  mutual  information  using  kernel  density  estimators.   Phys  Rev  E  52:2318.   Nemenman  I,  Bialek  W,  de  Ruyter  van  Steveninck  R  (2004)  Entropy  and  information  in  neural  spike  trains:   Progress  on  the  sampling  problem.  Phys  Rev  E  69:56111.   Nemenman  I,  Lewen  GD,  Bialek  W,  de  Ruyter  van  Steveninck  RR  (2008)  Neural  Coding  of  Natural  Stimuli:   Information  at  Sub-­‐Millisecond  Resolution.  PLoS  Comput  Biol  4:e1000025.   Nemenman  I,  Shafee  F,  Bialek  W  (2002)  Entropy  and  Inference,  Revisited.  Adv  Neural  Inf  Process  Syst  14   Proc  2002  Sic  Conf  14:95–100.   Nirenberg   S,   Carcieri   SM,   Jacobs   AL,   Latham   PE   (2001)   Retinal   ganglion   cells   act   largely   as   independent   encoders.  Nature  411:698–701.   Paninski  L  (2004)  Estimating  Entropy  on  Bins  Given  Fewer  Than  Samples.  IEEE  Trans  Inf  Theory  50:2201.   Panzeri  S,  Schultz  SR  (2001)  A  unified  approach  to  the  study  of  temporal,  correlational,  and  rate  coding.   Neural  Comput  13:1311–1349.   Panzeri  S,  Schultz  SR,  Treves  A,  Rolls  ET  (1999)  Correlations  and  the  encoding  of  information  in  the  nervous   system.  Proc  Biol  Sci  266:1001–1012.   Panzeri  S,  Senatore  R,  Montemurro  MA,  Petersen  RS  (2007)  Correcting  for  the  Sampling  Bias  Problem  in   Spike  Train  Information  Measures.  J  Neurophysiol  98:1064–1072.   Panzeri   S,   Treves   A   (1996)   Analytical   estimates   of   limited   sampling   biases   in   different   information   measures.  Netw  Comput  Neural  Syst  7:87–107.   Pillow   JW,   Paninski   L,   Uzzell   VJ,   Simoncelli   EP,   Chichilnisky   E   (2005)   Prediction   and   decoding   of   retinal   ganglion  cell  responses  with  a  probabilistic  spiking  model.  J  Neurosci  25:11003–11013.   Pillow   JW,   Shlens   J,   Paninski   L,   Sher   A,   Litke   AM,   Chichilnisky   E,   Simoncelli   EP   (2008)   Spatio-­‐temporal   correlations  and  visual  signalling  in  a  complete  neuronal  population.  Nature  454:995–999.   Pola   G,   Petersen   RS,   Thiele   A,   Young   MP,   Panzeri   S   (2005)   Data-­‐robust   tight   lower   bounds   to   the   information  carried  by  spike  times  of  a  neuronal  population.  Neural  Comput  17:1962–2005.   Pola  G,  Thiele  A,  Hoffmann  KP,  Panzeri  S  (2003)  An  exact  method  to  quantify  the  information  transmitted   by  different  mechanisms  of  correlational  coding.  Netw  Comput  Neural  Syst  14:35–60.   Roudi   Y,   Aurell   E,   Hertz   J   (2009)   Statistical   physics   of   pairwise   probability   models.   Arxiv   Prepr   ArXiv09051410.   Scaglione  A,  Foffani  G,  Scannella  G,  Cerutti  S,  Moxon  K  (2008)  Mutual  information  expansion  for  studying   the  role  of  correlations  in  population  codes:  How  important  are  autocorrelations?  Neural  Comput   20:2662–2695.   Scaglione  A,  Moxon  KA,  Foffani  G  (2010)  General  Poisson  Exact  Breakdown  of  the  Mutual  Information  to   Study  the  Role  of  Correlations  in  Populations  of  Neurons.  Neural  Comput  22:1445–1467.   Schaub   MT,   Schultz   SR   (2012)   The   Ising   decoder:   reading   out   the   activity   of   large   neural   ensembles.   J   Comput  Neurosci  32:101–118.   Schneidman  E,  Berry  II  MJ,  Segev  R,  Bialek  W  (2006)  Weak  pairwise  correlations  imply  strongly  correlated   network  states  in  a  neural  population.  Nature  440:1007–1012.   Schultz   S,   Treves   A   (1998)   Stability   of   the   replica-­‐symmetric   solution   for   the   information   conveyed   by   a   neural  network.  Phys  Rev  E  57:3302–3310.   Schultz   SR,   Kitamura   K,   Post-­‐Uiterweer   A,   Krupic   J,   Häusser   M   (2009)   Spatial   Pattern   Coding   of   Sensory   Information   by   Climbing   Fiber-­‐Evoked   Calcium   Signals   in   Networks   of   Neighboring   Cerebellar   Purkinje  Cells.  J  Neurosci  29:8005–8015.   Schultz  SR,  Panzeri  S  (2001)  Temporal  Correlations  and  Neural  Spike  Train  Entropy.  Phys  Rev  Lett  86:5823– 5826.   Schultz   SR,   Rolls   ET   (1999)   Analysis   of   information   transmission   in   the   Schaffer   collaterals.   Hippocampus   9:582–598.   Sharpee  T,  Rust  NC,  Bialek  W  (2004)  Analyzing  Neural  Responses  to  Natural  Signals:  Maximally  Informative   Dimensions.  Neural  Comput  16:223–250.   Shlens  J,  Field  GD,  Gauthier  JL,  Grivich  MI,  Petrusca  D,  Sher  A,  Litke  AM,  Chichilnisky  EJ  (2006)  The  Structure   of  Multi-­‐Neuron  Firing  Patterns  in  Primate  Retina.  J  Neurosci  26:8254–8266.   Shlens   J,   Kennel   MB,   Abarbanel   HDI,   Chichilnisky   E   (2007)   Estimating   information   rates   with   confidence   intervals  in  neural  spike  trains.  Neural  Comput  19:1683–1719.   Skaggs   WE,   McNaughton   BL,   Gothard   KM   (1993)   An   Information-­‐Theoretic   Approach   to   Deciphering   the   Hippocampal  Code.  In:  Advances  in  Neural  Information  Processing  Systems  5,  [NIPS  Conference],   pp   1030–1037.   San   Francisco,   CA,   USA:   Morgan   Kaufmann   Publishers   Inc.   Available   at:   http://dl.acm.org/citation.cfm?id=645753.668057  [Accessed  January  17,  2014].   Strong  SP,  Koberle  R,  de  Ruyter  van  Steveninck  RR,  Bialek  W  (1998)  Entropy  and  Information  in  Neural  Spike   Trains.  Phys  Rev  Lett  80:197–200.   Treves  A,  Panzeri  S  (1995)  The  Upward  Bias  in  Measures  of  Information  Derived  from  Limited  Data  Samples.   Neural  Comput  7:399–407.   Victor  J  (2006)  Approaches  to  information-­‐theoretic  analysis  of  neural  activity.  Biol  Theory  1:302–316.   Victor  JD  (2002)  Binless  strategies  for  estimation  of  information  from  neural  data.  Exp  Brain  Res  Phys  Rev  E   66:051903.         Figures  and  Figure  Captions       Figure   1:   The   origin   of   the   limited   sampling   bias   in   information   measures.     (A,   B)   Simulation   of   a   toy   uninformative  neuron,  responding  on  each  trial  with  a  uniform  distribution  of  spike  counts  ranging  from  0   to   9,   regardless   of   which   of   two   stimuli   (S   =   1   in   (A)   and   S   =   2   in   (B))   are   presented.   The   black   dotted   horizontal  line  is  the  true  response  distribution,  solid  red  lines  are  estimates  sampled  from  40  trials.  The   limited  sampling  causes  the  appearance  of  spurious  differences  in  the  two  estimated  conditional  response   distributions,  leading  to  an  artificial  positive  value  of  mutual  information.  (C)  The  distribution  (over  5000   simulations)  of  the  mutual  information  values  obtained  (without  using  any  bias  correction)  estimating  Eq.  1   from  the  stimulus–response  probabilities  computed  with  40  trials.  The  dashed  green  vertical  line  indicates   the   true   value   of   the   mutual   information   carried   by   the   simulated   system   (which   equals   0   bits);   the   difference   between   this   and   the   mean   observed   value   (dotted   green   line)   is   the   bias.   Redrawn   with   permission  from  (Ince  et  al.,  2010).       1.4 1.2 1 0.8 0.6 0.4 0.8 0.6 0.4 0.2 ) s t i b ( n o i t a m r o f n I r o r r e S M R 4 6 0 4 6 I 8 10 log2 trials/stim I 8 10 log2 trials/stim plugin nsb pt qe 12 14 1.4 1.2 1 0.8 0.6 0.4 0.8 0.6 0.4 0.2 ) s t i b ( n o i t a m r o f n I r o r r e S M R 4 6 12 14 0 4 6 Ish 8 10 log2 trials/stim Ish 8 10 log2 trials/stim 12 14 12 14   Figure  2:  Performance  of  various  bias  correction  methods.  Several  bias  correction  methods  were  applied  to   spike   trains   from   eight   simulated   somatosensory   cortical   neurons.   The   information   estimates   (upper   panels)  and  root  mean  square  (RMS)  error  (lower  panels)  are  plotted  as  a  function  of  the  simulated  number   of  trials  per  stimulus.  (A)  Mean  +/-­‐  SD  (upper  panel;  over  50  simulations)  and  RMS  error  (lower  panel)  of   I(S;R).  (B)  Mean  +/-­‐  SD  (upper  panel;  over  50  simulations)  and  RMS  error  (lower  panel)  of  Ish(S;R).       Figure   3.   Information   component   analysis   of   simulated   data   (for   a   five-­‐cell   ensemble).   (a)   Poisson   cells,   with   the   only   difference   between   the   stimuli   being   firing   rate.   (b)   Integrate-­‐and-­‐fire   cells,   with   common   input  due  to  sharing  of  one  third  of  connections,  resulting  in  a  cross-­‐correlogram  as  depicted,  leads  to  (c)   contribution  to  the  information  from  Icor-­‐ind.  (d)  Integrate-­‐and-­‐fire  cells  with  two  of  the  five  cells  increasing   their  correlation  due  to  increased  shared  input,  for  one  of  the  stimuli,  where  the  others  remain  randomly   correlated,   leading   to   cross-­‐correlograms   shown   in   panel   (e).   Firing   rates   are   approximately   balanced   between  the  two  stimuli.  This  results  in  a  substantial  contribution  to  the  information  from  the  stimulus-­‐ dependent  correlation  component,  Icor-­‐dep.  From  (Panzeri  et  al.,  1999)  with  permission.    
1701.08663
1
1701
2017-01-30T15:52:45
A computational study on synaptic and extrasynaptic effects of astrocyte glutamate uptake on orientation tuning in V1
[ "q-bio.NC", "q-bio.MN" ]
Astrocytes affect neural transmission by a tight control via glutamate transporters on glutamate concentrations in direct vicinity to the synaptic cleft and by extracellular glutamate. Their relevance for information representation has been supported by in-vivo studies in ferret and mouse primary visual cortex. In ferret blocking glutamate transport pharmacologically broadened tuning curves and enhanced the response at preferred orientation. In knock-out mice with reduced expression of glutamate transporters sharpened tuning was observed. It is however unclear how focal and ambient changes in glutamate concentration affect stimulus representation. Here we develop a computational framework, which allows the investigation of synaptic and extrasynaptic effects of glutamate uptake on orientation tuning in recurrently connected network models with pinwheel-domain (ferret) or salt-and-pepper (mouse) organization. This model proposed that glutamate uptake shapes information representation when it affects the contribution of excitatory and inhibitory neurons to the network activity. Namely, strengthening the contribution of excitatory neurons generally broadens tuning and elevates the response. In contrast, strengthening the contribution of inhibitory neurons can have a sharpening effect on tuning. In addition local representational topology also plays a role: In the pinwheel-domain model effects were strongest within domains - regions where neighboring neurons share preferred orientations. Around pinwheels but also within salt-and-pepper networks the effects were less strong. Our model proposes that the pharmacological intervention in ferret increases the contribution of excitatory cells, while the reduced expression in mouse increases the contribution of inhibitory cells to network activity.
q-bio.NC
q-bio
A computational study on synaptic and extrasynaptic effects of astrocyte glutamate uptake on orientation tuning in V1 Konstantin Mergenthaler1, Franziska Oschmann1, Jeremy Petravicz2, Dipanjan Roy1, Mriganka Sur 2, Klaus Obermayer1,* 1 Neural Information Processing Group, Fakultat IV and Bernstein Center for Computational Neuroscience, Technische Universitat Berlin, Berlin, Germany 2 Picower Institute for Learning and Memory, Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139, USA * [email protected] Abstract Astrocytes affect neural transmission by a tight control of the glutamate transporters which affect glutamate concentrations in direct vicinity to the synaptic cleft and in the extracellular space. The relevance of glutamate transporters for information representation has been supported by in-vivo studies in ferret and mouse primary visual cortex. A pharmacological block of glutamate transporters in ferrets broadened tuning curves and enhanced the response at preferred orientations. In knock-out mice with reduced expression glutamate transporters a sharpened tuning was observed. It is, however, unclear how focal and ambient changes in the glutamate concentration affect stimulus representation. Here, we developed a computational framework, which allows the investigation of synaptic and extrasynaptic effects of glutamate uptake on orientation tuning in recurrently connected network models with pinwheel-domain (ferret) or salt-and-pepper (mouse) organization. This model proposed that glutamate 1/37 uptake shapes information representation when it affects the contribution of excitatory and inhibitory neurons to the network activity. Namely, strengthening the contribution of excitatory neurons generally broadens tuning and elevates the response. In contrast, strengthening the contribution of inhibitory neurons can have a sharpening effect on tuning. In addition, local representational topology also plays a role: In the pinwheel-domain model effects were strongest within domains - regions where neighboring neurons share preferred orientations. Around pinwheels but also within salt-and-pepper networks the effects were less strong. Our model proposes that the pharmacological intervention in ferret increases the contribution of excitatory cells, while the reduced expression in mouse increases the contribution of inhibitory cells to network activity. Author Summary One of the key function of astrocytes is the clearance of neurotransmitters released during synaptic activity. Its importance for stimulus representation in the cortex was hypothesized following experiments that showed changes in selectivity when glutamate transport was blocked. Pharmacological and genetic interventions on glutamate transport considerably changed tuning width and strength of response in primary visual cortices of ferret and mouse. Here, we construct a modeling framework for visual cortices with pinwheel-domain and salt-and-pepper-organizations, which allows the detailed investigation of effects of altered glutamate uptake on orientation tuning. Our model proposes that changes in the representation of stimuli gets less selective if changes in glutamate uptake elicit stronger contribution of excitatory neurons to the network activity and selectivity is sharpened for a higher contribution of inhibitory neurons. Introduction Over the last years the view on astrocytes changed from mere supporting tissue providing metabolic support to active partners in information transmission and processing De Pitt`a et al. [2012], Alvarellos-Gonz´alez et al. [2012], Nadkarni et al. [2008], Reato et al. [2012], Perea et al. [2009]. Strongest drive to this shift of the perspective 2/37 was the development of calcium sensitive dyes Grynkiewicz et al. [1985] and the improvement of two-photon imaging Helmchen and Denk [2005]. These technique allowed the simultaneously observation of calcium transients in both astrocytes and neurons in-vivo Schummers et al. [2008]. Several pathways have been identified how neuronal and synaptic activity drive astrocyte activity Perea et al. [2014], Haydon and Nedergaard [2014], Benediktsson et al. [2012] or vice versa Perea et al. [2009], Araque et al. [2014], Chen et al. [2012], Nedergaard and Verkhratsky [2012]. Some of these pathways contain signaling cascades consisting of metabotropic receptors at the astrocyte membrane, internal second messenger signaling and vesicular release from astrocytes Panatier et al. [2011], Araque et al. [2014], De Pitt`a et al. [2011]. Other pathways contain transporters and pumps in the astrocyte plasma membrane, which directly link neuron and astrocyte activity via control of ion- and transmitter concentrations in a shared extracellular space Larsen et al. [2014], Rose and Karus [2013]. An in-vivo study in the ferret visual cortex (V1) revealed the relevance of astrocytes for stimulus representation in the cortex Schummers et al. [2008]. This study investigated the effects of a pharmacological block of the glutamate transport on the well-defined response to differently oriented gratings. While blocking the glutamate uptake in astrocytes leads to a stronger but less orientation selective response in neurons, the activity in astrocytes and the intrinsic optical signal were strongly attenuated. Another study revealed that a strongly reduced concentration of the primary astrocyte transporter (GLT-1) caused a sharpened orientation tuning Petravicz et al. [2014]. In a review Scimemi and Beato [2009] investigating how glutamate uptake might shape the synaptic glutamate concentration time course two key constraints were pointed out: geometry Rusakov and Kullmann [1998] and transporter efficiency Diamond [2001, 2005], Thomas et al. [2011], Zheng et al. [2008]. First, diffusion constraints, like a confined space Freche et al. [2011] and a clutter Min et al. [1998], shape the glutamate concentration after the release. Particularly, the size and the geometry of synapses play a role in glutamate clearance Tarczy-Hornoch et al. [1998], Meg´ıas et al. [2001], Guly´as et al. [1999]. Moreover, it has been observed that glutamatergic synapses to excitatory or to inhibitory cells differ in their geometry Koester and Johnston [2005]. Therefore, glutamatergic synapses are considered as a determining factor for these two types of 3/37 synapses Barbour et al. [1994]. In addition, Monte-Carlo modeling studies confirmed spatial constraints as a key determinant to glutamate clearance Freche et al. [2011], Rusakov [2001], Barbour [2001]. The second key constraint for the glutamate concentration time course are glutamate transporters, which shape the glutamate clearance from the synaptic cleft by buffering and complete uptake Danbolt [2001], Scimemi et al. [2009]. While some transporter subtypes are also found on pre- and postsynaptic neurons Danbolt [2001], Divito and Underhill [2014], the most abundant transporter (GLT-1) is highly concentrated on astrocyte processes ensheathing synapses Chaudhry et al. [1995], Benediktsson et al. [2012], Rusakov et al. [2014]. Dynamic changes in diffusion constraints occur primarily during maturation Thomas et al. [2011], Diamond [2005], but changes in neurotransmitter uptake can also be achieved by pharmacological blocking Schummers et al. [2008], or genetic ablation Petravicz et al. [2014]. Different effects of blocking glutamate transport on glutamate clearance have been found. One study proposes a shortening of glutamate clearance from the synaptic cleft when TBOA is applied, since less transporters are available to buffer glutamate within the cleft Scimemi et al. [2009]. Other studies propose a prolongation of the glutamate time course within the synapse during a block of the glutamate transport Murphy-royal et al. [2015], Barbour et al. [1994], Tong and Jahr [1994]. The modified glutamate concentration time course affects neurons via AMPA and NMDA receptors Tsukada et al. [2005], Bentzen et al. [2009]. During a blocked glutamate transport with TBOA a prolongation of AMPA-receptor mediated currents and a prevention of receptor desensitization were observed Mennerick et al. [1999]. Moreover, high concentrations of TBOA lead to a self-sustained pathologic rapid firing or to cell-death Tsukada et al. [2005], Rothstein et al. [1996]. Based on the studies named above we hypothesize that glutamate transporters shape physiological responses. The representation of stimulus specific features within the neo-cortex is largely considered to occur in networks which contain strong lateral connections with tightly calibrated excitatory and inhibitory contributions Stimberg et al. [2009], Shushruth et al. [2012], Marino et al. [2005]. This lead us to the question whether there are physiological properties of the glutamate uptake which could elicit changes in the proportion of the excitatory and inhibitory contribution. Moreover, the proportion of excitatory and inhibitory contribution crucially depends on the difference 4/37 in strength between excitatory and inhibitory neurons. Therefore, differences in extrasynaptic NMDA receptor-expression on excitatory and inhibitory neurons would determine the susceptibility of the network to ambient glutamate rise. To our knowledge studies which found differences in NMDA receptor properties on excitatory and inhibitory cells Martina et al. [2003, 2013] did not explicitly investigate extrasynaptic NMDA receptors. As such differences affect the proportion of excitatory and inhibitory contribution and no detailed experimental observations are available we incorporates differences in sensitivity into the model and explore its contribution to stimulus representation. In the following, we first investigate changes in the glutamate decay time within isolated synapses which comprise kinetic models for AMPA- and NMDA-receptors. Similar to Allam et al. [2012], David et al. [2009] we investigate changes in the glutamate decay time depending on the fraction of open AMPA and NMDA receptors. We particularly focus on fractions of open receptors when stimulations follow Poisson processes with different rates. In a next step these detailed synapses are integrated in a 2D- network for ferret visual cortex. When the glutamate transport is unchanged the network operates in a regime with strong lateral inhibitory and excitatory drive. For the integrated model we ask whether we can find combinations of glutamate decay times for synapses with excitatory and inhibitory connections which generate a similar loss in selectivity as in Schummers et al. Schummers et al. [2008]. As a second investigation we examine whether differences in the sensitivity to ambient glutamate between excitatory and inhibitory neurons shape orientation tuning in the network model. Motivated by the experiments which compare orientation tuning in GLT- wild type and knock-out mice we investigate the effects of different glutamate decay times and of different sensitivities to ambient glutamate. These experiments were also performed in a network with salt-and-pepper organization Runyan and Sur [2013]. 5/37 Results Glutamate uptake and its effect on the transmission properties of single excitatory synapses We studied the influence of astrocyte-mediated glutamate uptake on the transmission properties of glutaminergic synapses in a simplified setting. Here, the dynamics of synaptic AMPA receptors and NMDA receptors were described using kinetic models with 3 and 5 states (see Methods: Neurotransmitter concentration & receptor dynamics). The glutamate concentration in the synaptic cleft was quantified by bi-exponential pulses following every presynaptic spike. Different glutamate decay time constants accounted for changes in the efficacy of astrocytic glutamate uptake, where short (long) decay times corresponded to fast (slow) glutamate uptake. Fig. 1 shows the fraction of open NMDA and AMPA receptors in response to Poisson-distributed spike trains for three different decay time-constants of the glutamate pulses. The fraction of open NMDA receptors was mostly affected by different glutamate decay times when both the fraction of open NMDA receptors and the number of glutamate pulses were low (see Fig. 1A). As a consequence, different glutamate decay times had the biggest impact on the fraction of open NMDA receptors when the frequency of the presynaptic spike rates ranged between 10 and 15 Hz (see Fig. 1B). The fraction of open AMPA receptors was only marginally influenced by different glutamate decay times for large intervals between presynaptic spikes. However, an increase of the glutamate decay time prolonged the time to complete receptor closure (see Fig. 1A). Moreover, the effect of different glutamate decay time constants on the fraction of open AMPA receptors increased with the presynaptic spike rate (see Fig. 1B). Effects in a V1 with pinwheel-domain organization By asking whether affecting glutamate transport might have effects on representation of information in a recurrently connected networks, particular importance can be attributed to mechanisms weighting the contribution of excitatory and inhibitory populations. As the glutamate decay does not only depend on glutamate transporters 6/37 Figure 1. Simulation of a single synapse. A (upper) Time course of the glutamate concentration G for different decay times gf (fast: 0.6 ms (yellow); base line: 0.75 ms (black); slow: 0.975 ms (blue)). The glutamate pulses are generated by a Poisson-rate of 40 Hz, which drive the NMDA & AMPA receptors. (middle) Time course of the fraction of open NMDA receptors for different glutamate decay time constants following the glutamate pulses shown in the upper figure. (lower) Time course of the fraction of open AMPA receptors for different glutamate decay time constants following the glutamate pulses shown in the upper figure. B Fraction of open NMDA receptors after stimulation of 2 s with different Poisson-rates. Bold lines show the average proportion of open receptors and the shaded area its standard deviation. Largest differences in mean and strongest variation are found around 15 Hz. For high rates differences vanish. C Fraction of open AMPA receptors after stimulation of 2 s with different Poisson-rates. Bold lines show the average proportion of open receptors and the shaded area its standard deviation. Average proportion of AMPA-receptors increase with rate and decay-constants. Standard deviation is largest and less rate dependent for short decay times. but also on synapse geometries we independently varied the glutamate decay time for lateral synapses to either excitatory (EE-synapses) or inhibitory (IE-synapses) neurons and investigated changes in tuning. In our single layer model lateral synapses were synapses formed between neurons within the layer in contrast to afferent synapses, which originate from lower layers. Synaptic mechanism Starting from our reference point with the same decay constant (0.75 ms, red box in Fig. 2A) for EE-synapses and IE-synapses we observed that a prolongation of the glutamate decay time within EE-synapses broadens the firing rate tuning (Half-width-at-half-max: HWHM increases). The reference point (τf EE = τf IE = 0.75 ms) was chosen in accordance with values derived in Diamond 7/37 ABC0.100.20NMDApopenG [mM]0.00.8100 ms0.00.5AMPApopen050100Input Rate [Hz]NMDA popen0.000.10AMPA popen0501000.000.25gf = 0.75 msgf = 0.75 msgf = 0.975 msgf = 0.975 msgf = 0.60 msgf = 0.60 msgf = 0.75 msgf = 0.975 msgf = 0.60 ms [2005]. The broadening of tuning curves was even stronger with a simultaneous reduction of the decay constant in IE-synapses (exemplary point: blue box in Fig. 2A). Slight sharpening was observed when prolongation occurs mostly within IE-synapses (reference point: green box in Fig. 2A). This picture held within domain centers as well as close to pinwheels. However, close to pinwheels we observed markedly smaller effects for different decay constants. In addition to changes in firing rate tuning a very similar picture was found for excitatory and inhibitory conductances as well as the sub-threshold membrane potential. Interestingly, the membrane potential showed a prominent sharpening when the glutamate decay time within IE-synapses was prolonged (Fig. 2A lower-left panel). The broadening of the response for a prolonged decay in EE-synapses and a shortened decay in IE-synapses went hand in hand with an increase in firing rates (Fig. 2B). Therefore, a detailed prolongation of the glutamate decay time within EE-synapse and a simultaneous reduction of the glutamate decay time within IE-synapses provided a plausible condition for the experimentally observed change in tuning response during pharmacological block of the glutamate transport in ferret V1 Schummers et al. [2008]. In addition to changes in lateral connections, a simultaneous change in afferent excitatory connections (EA-synapses: to excitatory neurons, and IA-synapses: to inhibitory neurons) might occur. The exploration of the simultaneous prolongation in EE- and EA-synapses as well as in IE- and IA-synapses revealed, that prolongation and shortening became more effective and enhance the strengthening effect of one population above the other. However, no qualitative change were observed (data not shown). Extrasynaptic mechanism Another mechanism that weights the contribution of the excitatory and the inhibitory population differently and could originate from changes in the glutamate transport is a difference in sensitivity to the ambient glutamate level of excitatory and inhibitory neurons via extrasynaptic NMDARs. As a proxy for different NMDAR-densities we independently varied the ambient glutamate concentration affecting NMDARs on excitatory and inhibitory cells. With an increase of the ambient glutamate concentration effective on the excitatory neurons the orientation tuning broadened (higher HWHM values, cf. Fig. 3A). In addition, responses at preferred and non preferred orientations increased (Fig. 2B). Again this effect was much 8/37 Figure 2. Synaptic effect in pinwheel-domain network model. A: Glutamate decay within the IE- (horizontal axis) and EE-synapses (vertical axis) are independently varied. The reference condition point is 0.75/0.75 ms (red box). Values below 0.75 ms are shortened and values above 0.75 ms are prolonged glutamate clearance values. Half-width-at-half-max (HWHM) values (color coded) of the tuning curves are separately derived for neurons within orientation domains (left) and neurons close to pinwheels (right) for the firing rate, the received excitatory conductance, the received inhibitory conductance, and the membrane potential in excitatory neurons. All four investigated properties show a loss in selectivity and increased values if prolongation of glutamate decay preferentially occurs in connections to excitatory neurons. The effect is even stronger with a simultaneous reduction in decay time for connections to inhibitory neurons (exemplary: blue box). If prolongation would mostly occur in connections to inhibitory neurons responses are slightly sharpened (exemplary: green box). B Tuning curves for the exemplary points from A. 9/37 HWHMHWHMHWHMHWHMBA0.520.750.981.200.520.750.981.20Domain ge0.520.750.981.200.520.750.981.20Domain gi0.520.750.981.200.520.750.981.20Domain Rate0.520.750.981.20glutamate decay (IE)τfIE [ms]glutamate decay (IE)τfIE [ms]glutamate decay (IE)τfIE [ms]glutamate decay (IE)τfIE [ms]0.520.750.981.20glutamate decay (EE) τfEE [ms]Domain V0.520.750.981.20Pinwheel ge27.532.537.542.50.520.750.981.20Pinwheel gi27.532.537.542.50.520.750.981.20Pinwheel Rate27.532.537.542.50.520.750.981.20Pinwheel V42.046.551.0−50050Orientation [deg.]0204060Rate [Hz]Tuning Rate−50050Orientation [deg.]−60−55V [mV]Tuning V−50050Orientation [deg.]0.00.20.40.60.81.0Tuning gi−50050Orientation [deg.]0.00.10.20.30.40.5Conductance [pS]Conductance [pS]Tuning gecontrol DOcontrol PWIE prol DOIE prol PWEE prol DOEE prol PW more pronounced within domains and much weaker around pinwheels. For an even stronger effect on excitatory neurons the network entered a state of pathological self sustained activity. The effect of NMDAR-currents on inhibitory neurons was again small and only small changes in the tuning width were observed. Effects in a V1 with salt-and-pepper organization The smaller effects around pinwheels called for the investigation of effect of glutamate decay times in a network with a salt-and-pepper organization. We investigated the effect of differential changes of the glutamate decay time in a model, which was calibrated to reproduce the observed firing rate tuning in mouse V1 Runyan and Sur [2013]. Synaptic mechanism It turned out that changes in the glutamate decay constants only weakly changed the firing rate orientation tuning and had only negligible effects on the other quantities when considering half-width-at-half-max values (HWHM) Fig. 4A. While the shape of the tuning curves hardly changed and the tuning curves were mostly shifted upward for prolonged glutamate decay in EE-synapses and shifted downward for prolonged glutamate decay in IE-synapses Fig. 4B, responses at preferred and non-preferred orientations changed. The orientation-selectivity-index (OSI), however, merged shift- and shape-changes and OSI-distributions were either shifted to lower values when glutamate decay was prolonged in EE-synapses or were shifted to higher values when glutamate decay was prolonged in IE-synapses. For strongly prolonged glutamate decay times in EE-synapses with simultaneous shortening in IE-synapses the network reached self-sustained firing. Again during a simultaneous prolongation in EE- and EA-synapses as well as in IE- and IA-synapses the effect on HWHMs were more pronounced (Fig. 5A). Now, the selectivity loss for a prolonged decay in EE- and EA-synapses and a shortened decay in IE- and IA-synapses, and the selectivity increase for a prolonged decay in IE- and IA-synapses and a shortened decay in EE- and EA-synapses, were also visible in the HWHMs of the sub threshold properties. Nevertheless, the biggest change occurred as upward or downward shifts of the tuning curves independent of the preferred orientation (Fig. 5B). Both changes were reflected in changes in OSI-values and prolongation to excitatory neurons shifted OSI-distributions to lower values, and prolongation to 10/37 Figure 3. Extrasynaptic effect in pinwheel-domain network model. A Different elevated levels of ambient glutamate sensed by excitatory (vertical axis) and inhibitory (horizontal axis) neurons represent different efficiancies of NMDAR on excitatory and inhibitory neurons. Half-width-at-half-max (HWHM) values (color coded; gray = self sustained network activity) are derived for neurons within domains (left) and close to pinwheels (right) for firing rate, excitatory and inhibitory conductance and membrane potential. An increase in NMDAR-currents to excitatory neurons (exemplary: blue box) reduces orientation tuning selectivities and generally increases responses. An increase in NMDAR-currents on inhibitory neuron give rise to slightly sharpened but weaker responses (exemplary: green box). B Exemplary tuning curves from A. inhibitory neurons shifted OSI-distributions to higher values (Fig. 5C). 11/37 −50050Orientation [deg.]−60−55V [mV]Tuning VHWHMHWHMHWHMHWHMBA0.000.751.502.250.000.751.502.25Domain ge0.000.751.502.250.000.751.502.25Domain gi0.000.751.502.250.000.751.502.25Domain Rate0.000.751.502.250.000.751.502.25Domain V0.000.751.502.25Pinwheel ge27.532.537.542.50.000.751.502.25Pinwheel gi27.532.537.542.50.000.751.502.25Pinwheel Rate27.532.537.542.50.000.751.502.25Pinwheel V42.046.551.0−50050Orientation [deg.]0.00.10.20.30.40.5Tuning ge−50050Orientation [deg.]0.00.20.40.60.81.0Tuning gi−50050Orientation [deg.]020406080Rate [Hz]Conductance [pS]Conductance [pS]Tuning RateControl DOMControl PWI stronger DOMI stronger PWE stronger DOME stronger PWambient glutamate to inh. GambI [μM]ambient glutamate to inh. GambI [μM]ambient glutamate to inh. GambI [μM]ambient glutamate to inh. GambI [μM]ambient glutamate to exc. GambE[μM] Figure 4. Synaptic effect in salt-and-pepper network model. A: Glutamate decay time is separately varied in lateral synapses to excitatory neurons (vertical axis) and inhibitory neurons (horizontal axis) and HWHM of tuning curves for firing rate, excitatory and inhibitory conductance, and membrane potential are shown color-coded. Only HWHM for rate shows a prominent effect of changes in glutamate decay time. Boxes are for exemplary points (reference: red; prolongation in connections to excitatory neurons (EE): blue; prolongation to inhibitory neurons (IE): green) B Tuning curves for exemplary points show upward (prolonged EE) and downward (prolonged IE) shifts with little change in tuning width. C Orientation-Selectivity (OSI)-distributions for the exemplary points show higher OSI-values when IE-decay is prolonged (green), and lower OSI-values when EE-decay is prolonged (blue) for rate, excitatory conductance and membrane potential. Extrasynaptic mechanism We used different ambient glutamate concentrations as a proxy for different sensitivities of inhibitory and excitatory neurons to elevated ambient glutamate. In a salt-and-pepper network we observed that stronger sensitivity of excitatory neurons broadened the tuning (Fig. 6A blue box). For inhibitory cells more sensitive to ambient glutamate the HWHMs of the tuning curves of the firing 12/37 HWHMHWHMHWHMHWHMBAC−50050Orientation [deg.]−62−60−58−56V [mV]Conductance [pS]Conductance [pS]Tuning V−50050Orientation [deg.]0510152025Rate [Hz]Tuning RateNormalEE Prol.IE Prol.−50050Orientation [deg.]0.250.300.350.400.45Tuning gi−50050Orientation [deg.]0.050.100.150.20Tuning ge0.00.20.40.60.81.0OSI0246810Dist (OSI)OSI Distribution ge0.00.20.40.60.81.0OSI05101520Dist (OSI)OSI Distribution gi0.00.20.40.60.81.0OSI012345Dist (OSI)OSI Distribution RateControlEE Prol.IE Prol.0.00.20.40.60.81.0OSI051015Dist (OSI)OSI Distribution V0.520.750.981.200.520.750.981.20HWHM ge24.025.527.028.50.520.750.981.200.520.750.981.20HWHM gi24.025.527.028.50.520.750.981.20glutamate decay (IE) τfIE [ms]glutamate decay (IE) τfIE [ms]glutamate decay (IE) τfIE [ms]glutamate decay (IE) τfIE [ms]0.520.750.981.20glutamate decay (EE) τfEE [ms]HWHM Rate1720230.520.750.981.200.520.750.981.20HWHM V24.025.527.028.5 Figure 5. Synaptic effect in salt-and-pepper network model -- all synapses. A: In contrast to the exploration in Fig. 4 the decay time in afferent synapses is varied alongside the lateral ones. Again the HWHM for rate shows broadening for prolonging EE-synapses and EA-synapses and some sharpening for prolonged IE- and IA-synapses. In addition small difference could also be found in the sub threshold properties. Boxes are for exemplary points (reference: red; prolongation in connections to excitatory neurons (EE + EA): blue; prolongation to inhibitory neurons (IE + IA): green) B The tuning curves for exemplary points show that changes in HWHM are minor in comparison to the strong shifts (upward for EE + EA-synapses and downward for IE + IA-synapses) C The OSI-distribution combining baseline-shifts and width-changes show clearer separation of selected points. Generally, lower OSI-values are observed when glutamate decay in EE + EA-synapses is prolonged (blue) and higher OSI-values if prolongation occurs mostly in IE + IA-synapses (green). rates and the conductances were markedly reduced (Fig. 6A green box). The membrane potential showed almost unchanged HWHM-values. For the synaptic mechanism the strongest effect were orientation independent shifts (Fig. 6B), which elevated the baseline activity for ambient glutamate mostly affected excitatory cells, and pulled 13/37 0.00.20.40.60.81.0OSI0246Dist (OSI)−50050Orientation [deg.]0102030Rate [Hz]Tuning RateHWHMHWHMHWHMHWHMACConductance [pS]Conductance [pS]ControlEE + EAIE + IAOSI Distribution geOSI Distribution giOSI Distribution RateControlEE + EA Prol.IE + IA Prol.OSI Distribution V0.520.750.981.200.520.750.981.20HWHM ge24.025.527.028.50.520.750.981.200.520.750.981.20HWHM gi24.025.527.028.50.520.750.981.20glu. dec. (IE & IA) τfIE & τfIA [ms]glu. dec. (IE & IA) τfIE & τfIA [ms]glu. dec. (IE & IA) τfIE & τfIA [ms]glu. dec. (IE & IA) τfIE & τfIA [ms]0.520.750.981.20glutamate decay (EE & EA) τfEE & τfEA [ms]HWHM Rate1720230.520.750.981.200.520.750.981.20HWHM V24.025.527.028.5−50050Orientation [deg.]0.050.100.150.200.25Tuning ge−50050Orientation [deg.]0.250.300.350.400.45Tuning gi−50050Orientation [deg.]−62−60−58−56V [mV]Tuning VB0.00.20.40.60.81.0OSI0246810Dist (OSI)0.00.20.40.60.81.0OSI05101520Dist (OSI)0.00.20.40.60.81.0OSI051015Dist (OSI) down the baseline values for ambient glutamate mostly affected inhibitory cells. The prominent baseline shifts and the changes in tuning width combined to clear shifts in the OSI-distributions (Fig. 6C). For more sensitive inhibitory neurons OSI-distributions were shifted to higher values and for more sensitive excitatory neurons OSI-distributions were shifted to smaller values. Figure 6. Extrasynaptic-effect in salt-and-pepper network model. A: Again different elevated levels of ambient glutamate to excitatory (vertical axis) and inhibitory (horizontal axis) concentrations of ambient glutamate represent different NMDAR efficiancies. Tuning (HWHM) gets less selective if extrasynaptic NMDAR-currents have a stronger effect on excitatory neurons and more selective if NMDAR-currents are higher on interneurons in firing rates and again much weaker in the other variables. B The tuning curves for sub-threshold properties show strong stimulus-orientation independent changes. C OSI-distributions show higher values when the inhibitory population is primary target of ambient glutamate and lower values when the excitatory population is primary target. 14/37 HWHMHWHMHWHMHWHMC0.00.20.40.60.81.0OSI051015Dist (OSI)OSI DistributionV0.00.20.40.60.81.0OSI0246Dist (OSI)OSI Distribution RateControlE strongerI stronger0.00.20.40.60.81.0OSI05101520Dist (OSI)OSI Distribution gi0.00.20.40.60.81.0OSI051015Dist (OSI)OSI Distribution ge−50050Orientation [deg.]−64−62−60−58−56V [mV]Conductance [pS]Conductance [pS]Tuning V−50050Orientation [deg.]0102030Rate [Hz]Tuning RateControlE strongerI stronger−50050Orientation [deg.]0.250.300.350.400.45Tuning gi−50050Orientation [deg.]0.000.050.100.150.20Tuning geB0.000.751.502.250.000.751.502.25HWHM ge24.025.527.028.50.000.751.502.250.000.751.502.25HWHM gi24.025.527.028.50.000.751.502.250.000.751.502.25HWHM Rate1720230.000.751.502.25amb. glu. to inh. GambI[µM]amb. glu. to inh. GambI[µM]amb. glu. to inh. GambI[µM]amb. glu. to inh. GambI[µM]0.000.751.502.25ambient glutamate to exc. GambE [µM]HWHM V24.025.527.028.5A Discussion Schummers et al. Schummers et al. [2008] observed a loss in selectivity to oriented gratings when glutamate transport is blocked pharmacologically. Our presented model reproduces the loss in selectivity, but only if changes in the glutamate transport enhance the contribution of the excitatory population. Such a strengthening was achieved via two different pathways for glutamate uptake. One was the prolongation of glutamate decay within synapses to excitatory neurons. The second one was a postulated higher sensitivity of excitatory neurons to ambient glutamate. The importance of shifts in excitatory vs inhibitory contribution mediated by changes in glutamate transport can even be seen in a model with a salt-and-pepper organization of preferred orientations. Such a model shows sharpened tuning (higher OSI-values) as for GLT-1+/−-mice in Petravicz et al. [2014], but only if the contribution of inhibitory neurons is strengthened. Interestingly, the pinwheel-domain network showed an interaction between mapOSI as well as synaptic and extrasynaptic glutamate uptake effects. Changes in tuning were always stronger in domains and much less pronounced close to pinwheels. Following, the analogy of neurons within a salt-and-pepper network as neurons at pinwheels in a pinwheel-domain network it is not unexpected that effects on tuning width (HWHM) are small in such a network. Particularly the observation of lower effects of changes in lateral connections onto HWHM values but orientation unspecific shifts of tuning curves is in line with a suggested stronger contribution of weakly tuned inhibitory neurons as in Bopp et al. [2014]. Interestingly, the salt-and-pepper network with fewer lateral connections is more susceptible to pathologic self-sustained firing. For the pinwheel-domain as well as for the salt-and-pepper map we achieved to directly link effects of glutamate uptake to changes in information representation. To link these we, however, were forced to construct rather complex models with a lot of fixed parameters and a lot of detail. We took deliberate care in selecting fixed parameters to be in a physiological range, e.g., parameters describing the Hodgkin-Huxley dynamics of single neurons stem from models largely used for neurons in visual cortex Stimberg et al. [2009], Marino et al. [2005], Destexhe et al. [2001]. The necessity of 2D-network structures with local lateral connections follows arguments in Stimberg et al. [2009], Marino et al. [2005] and Roy 15/37 et al. [2013]. It allowed us to calibrate the model in reference condition to match experimentally observed orientation tuning in pinwheel-domain networks Marino et al. [2005] and salt-and-pepper networks Runyan and Sur [2013] and to investigate interactions of local heterogeneity in representing stimulus features with effects of glutamate uptake. With the low number of connections and neurons and the very high peak synaptic conductances we assume that each neuron and connection is representative for a subpopulation sharing exactly the same features as the representative single neuron. In contrast to our separate investigations of synaptic and extrasynaptic effect, we expect that glutamate uptake experiments show a combined effect of both mechanisms. To confine their exact contributions within the cortex, or answer whether changes in synaptic clearance or raising ambient glutamate can be ruled out -- due to no effective shift towards excitation or inhibition, required new careful experiments. For the synaptic effect, first synapses onto excitatory and inhibitory cells need to be separately investigated and separate assessments of synapses geometry, size, and transporter densities is required. Second, in single synapse studies -- either experimental or detailed modeling studies similar to Scimemi and Beato [2009], Rusakov and Kullmann [1998], Freche et al. [2011] -- effects on glutamate clearance and the susceptibility to altered glutamate uptake for the two types of synapses need to be investigated. For the extrasynaptic mechanism a separate estimation of only extrasynaptic NMDA receptor densities on excitatory and inhibitory cells would allow to estimate the effective impact of ambient glutamate. For the single synapse models we observed interactions between glutamate clearance decay time and firing rate in the contribution to average and fluctuations in open fractions of receptors. This leads to a range of medium frequencies (10-15 Hz) where NMDA-receptors show the highest sensitivity to changes in glutamate clearance. Similarly, AMPA-receptor fluctuations are most sensitive in a similar range. Considering that complex synapses will be present in networks which transit between fluctuation and mean driven phases Litwin-Kumar and Doiron [2012], Renart et al. [2007], we propose that changes in glutamate clearance interact with the cortical dynamical state. Finally, in the context of astrocytes as active partners, a short coming of our model is that both pathways of glutamate uptake were investigated without intrinsic dynamics. 16/37 Further investigations on the effects of glutamate uptake in networks would largely benefit from coupled dynamic models of neurons and astrocytes. In such models the dynamic intrinsic state of an astrocytes, e.g. Ca2+-concentration would interact with glutamate uptake and finally the neighboring neurons. Methods 0.1 Neuron model and postsynaptic currents Concentration of neurotransmitter in the synaptic cleft and channel kinetics. GY describes the time course of the neurotransmitter concentration in the synaptic cleft for the the excitatory neurotransmitter glutamate (GE) and the inhibitory neurotransmitter GABA (GI ). The time course of the neurotransmitter concentration in response to a presynaptic action potential follows a bi-exponential function: GY (t) = 1 τf Y − τrY (cid:18) (cid:88) tk<t (cid:18) exp − tk − t τf Y (cid:19) − exp (cid:19)(cid:19) (cid:18) − tk − t τrY . Here, the rise and decay constants τrY and τf Y (r: rise, f: decay) vary for different pairings of the post- (left letter) and presynaptic (right letter) cell type (Y ∈ {EE, EI, IE, II}, E: excitatory, I: inhibitory). tk denotes the arrival time of the action potential. Parameter values are summarized in Table 1. We chose τrEE = τrIE = τrE, τrEI = τrII = τrI , and τf EI = τf II = τf I . The rise constant τrE remained fixed, because of its small value. Variations in the decay constants τf EE and τf IE accounted for changes in the astrocytic glutamate uptake. GY was normalized, such that the peak concentrations of glutamate and GABA were set to 1 mM Clements et al. [1992], Vizi et al. [2010]. Fig. ?? shows the kinetic schemes used for the AMPA-, NMDA-, and GABAA-channels. The AMPA-channel is described by one closed, one desensitized and one open state Saftenku [2005]. The NMDA-channel passes through three closed, one desensitized and one open state Lester and Jahr [1992]. The GABA-A channel has three closed and two open states Destexhe et al. [1998]. Neuron model Neurons are described by conductance-based point neuron models, where changes of the membrane voltage VX for excitatory (X = E) and inhibitory 17/37 Table 1. Ligand gated receptor dynamics Description Source Glutamate concentration rise time Exc. to inh. concentration decay time Diamond [2005]* Exc. to exc. concentration decay time Diamond [2005]* GABA concentration rise time GABA concentration decay time Diamond [2005] † † AMPAR resensitization rate AMPAR desensitization rate AMPAR opening rate AMPAR closing rate AMPAR binding rate 0.16 ms 0.545 -- 1.275 ms 0.545 -- 1.275 ms 0.29 ms 0.291 ms 0.065 s−1 5.11 s−1 25.39 s−1 4. s−1 0.44 mM 1 × 106 M−1s−1 12.9 s−1 8.4 s−1 6.8 s−1 46.5 s−1 73.8 s−1 20 × 106 M−1s−1 GABAAR binding rate 1 10 × 106 M−1s−1 GABAAR binding rate 2 4.6 × 103 s−1 9.2 × 103 s−1 3.3 × 103 s−1 10.6 × 103 s−1 9.8 × 103 s−1 410 s−1 Parameter Value Synaptic -- Neurotransmitter τrE τf IE τf EE τrI τf I Synaptic -- AMPA Channel dynamics Rar Rad Rao Rac KB Synaptic -- NMDA Channel dynamics Rnb Rnu Rnd Rnr Rno Rnc Synaptic -- GABAA Channel dynamics Rgb1 Rgb2 Rgu1 Rgu2 Rgo1 Rgo2 Rgc1 Rgc2 *A range of values around 0.75ms (derived in ?) was explored. † Rise and decay constants were chosen such that the mean squared distance between the bi-exponential function and the concentration of GABA as a function of time calculated as in Destexhe et al. [1998] was minimal (particle swarm optimization). Destexhe et al. [1998] NMDAR binding rate NMDAR unbinding rate NMDAR desensitization rate NMDAR resensitization rate NMDAR opening rate NMDAR closing rate GABAAR unbinding rate 1 GABAAR unbinding rate 2 GABAAR opening rate 1 GABAAR opening rate 2 GABAAR closing rate 1 GABAAR closing rate 2 Saftenku [2005] Saftenku [2005] Saftenku [2005] Saftenku [2005] Saftenku [2005] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] (X = I) neurons are driven by a sum of transmembrane currents: Cm dVX dt = −IL,X − Iint,X − Isyn,X − Iamb,X − Ibg,X . (1) Cm denotes the membrane capacitance and t the time. We consider: (i) a leak current IL,X = −gL,X (VX − EL) with leak conductance gL,X and reversal potential EL, (ii) a sum Iint,X of three Hodgkin-Huxley type voltage-gated intrinsic currents (see below), (iii) the total synaptic ligand-gated current Isyn,X , (iv) a ligand-gated current Iamb,X driven by extrasynaptic glutamate, and (v) a background current Ibg,X for inducing a realistic level of spontaneous activity. Parameters are summarized in Table 2. 18/37 Table 2. Membrane capacitance and parameters for the leak, intrinsic, and background currents of the neuron model. Description Marino et al. [2005] Marino et al. [2005] Marino et al. [2005] Marino et al. [2005] Marino et al. [2005] Marino et al. [2005] Marino et al. [2005] Potassium current, no. of activation sites Potassium current, no. of inactivation sites Marino et al. [2005] M-channel, peak conductance, excit. neurons Marino et al. [2005] 0.35 nF Membrane capacitance 15.7 nS 31.4 nS −80 mV Reversal potential Leak conductance, excitatory neurons Leak conductance, inhibitory neurons Parameter Value Membrane capacitance and leak current Cm gL,E gL,I EL Intrinsic (voltage gated) currents 17.9 µS Sodium current, peak conductance gN a 50 mV Sodium current, reversal potential EN a 3 Sodium current, no. of activation sites lN a 1 Sodium current, no. of inactivation sites kN a 3.46 µS Potassium current, peak conductance gKd −90 mV Potassium current, reversal potential EKd 4 lKd 0 kKd 279 nS gM,E 27.9 nS M-channel, peak conductance, inhib. neurons gM,I −85 mV M-channel, reversal potential EM 1 lM 0 kM Background currents −5 mS EbgE Excitatory current, reversal potential −70 mV Inhibitory current, reversal potential EbgI 2.7 ms τbgE τbgI 10.7 ms 8.79 nS ¯gbgEE 28.8 nS ¯gbgEI 17.5 nS ¯gbgIE ¯gbgII 57.6 nS 0.157 nS Noise strength, excit. conductance σbgE σbgI 0.313 nS Noise strength, inhib. conductance Excitatory current, time constant Inhibitory current, time constant average excit. to excit. conductance average inhib. to excit. conductance average excit. to inhib. conductance average inhib. to inhib. conductance M-channel, no. of activation sites M-channel, no. of inactivation sites Source Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Marino et al. [2005] Marino et al. [2005] Marino et al. [2005] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Schummers et al. [2007] Intrinsic currents The Hodgkin-Huxley-type neuron model implements three intrinsic voltage-gated currents Iint,X = IN a + IKd + IM,X , a fast sodium current IN a, a delayed-rectified potassium current IKd, and a slow non-inactivating population specific potassium current IM,X . The intrinsic currents of each neuron follow: IZ = ¯gZmlZ Zact(V )mkZ Zinac(V )(V − EZ), Z ∈ {N a, Kd, M}, (2) with ¯gZ the peak conductance, mZact and mZinact the activation and inactivation variables, lZ and kZ the number of activation and inactivation sites, and EZ the reversal potential of the channel. The peak conductance for the M -current is population 19/37 selective ¯gM,X , to account for weaker adaptation in inhibitory neurons. The dynamics of activation (D = act) and inactivation (D = inac) are given by: dmZD dt = αZDo(1 − mZD) − αZDcmZD αZDd = v1v2(V ) exp(v3(V )) + v4 with αZDd opening (d = o) and closing (d = c) transition rates. The kinetics follow Destexhe and Par´e [1999] and are summarized in Table 3 Table 3. Channel dynamics Gating var. αN a act o αN a inac o αKd act o αM act o αN a act c αN a inact c αKd act c αM act c v1 [mV−1] (V + 51)/18 v3(V ) [mV] 0.32 0.128 0.032 v2(V ) [mV] v4 −(V + 45) −(V + 45)/4 −1 0 −(V + 40) −(V + 40)/5 −1 2.9529 × 10−4 −(V + 30) −(V + 30)/9 −1 −1 1 0 −1 (V + 18)/5 −(V + 28)/5 (V + 45)/40 (V + 30)/9 2.9529 × 10−4 0.28 4. 0.5 1 1 1 V + 18 V + 30 Expressions for channel dynamics as in Destexhe and Par´e [1999] Synaptic currents Each neuron receives a set of (lateral & afferent) glutamatergic and lateral GABA-ergic synaptic currents. (cid:88) Isyn,X = 1 NAf f j I j AM P Aaf f,X + 1 NXe AM P A,X + I k NM DA,X (cid:88) (cid:0)I k k (cid:1) + (cid:88) m 1 Nci I m GABAA with NAf f , NXe, Nci the number of received connections, and j, k, m the indices of projecting afferent, excitatory, and inhibitory neurons with X ∈ {E, I} the target population. The current through the specific receptor type at every synapse is governed by the introduced receptor dynamics. The post-synaptic current IR with R ∈ {AM P A, N M DA, GABAA} is given by: IR = ¯gRBlR (ΣOR) (V − ER), lR = (cid:124) (cid:123)(cid:122) gR (cid:125)  1 0 for R = NMDA , otherwise 20/37 with ¯gR the receptor specific peak conductance, ΣOR the sum of open states, V the membrane potential of the post-synaptic neuron, and ER the reversal potential. NMDA-receptors also express a voltage and magnesium dependent block B = (1 + exp (−0.062V + 1.2726) Mg) −1 and Mg the extracellular magnesium concentration Jahr and Stevens [1990]. For the considerations of tuning in the conductances the excitatory synaptic conductances gNM DA and gAM P A are aggregated (ge = gNM DA + gAM P A). For the inhibitory conductances the experimental limitations to distinguish between synaptic inhibitory conductances and adapting intrinsic conductances we combine the GABAA and the slow non-inactivating potassium current. (gi = gGABAA + gM,E) to provide compatible values Schummers et al. [2007]. Synaptic peak conductances A major difference to earlier models of V1 (cf. Stimberg et al. [2009] and Roy et al. [2013]) are the detailed synaptic kinetics. Therefore, afferent and lateral peak conductances had to be re-adjusted. To determine the peak-conductances for afferent (¯gAM P Aaf f,E and ¯gAM P Aaf f,I ) and inhibitory synapses (¯gGABAA) we stimulated simple exponential AMPA and GABAA synapses parametrized as in Stimberg et al. [2009] and the introduced detailed ones with the same 40 Hz Poisson spike-trains for 2 s. Then we determined the peak conductance for which the average conductances were equal. For the excitatory lateral peak conductances we assumed a 4:1 ratio for AMPA to NMDA receptors and we followed the paths described in Stimberg et al. [2009] and in Roy et al. [2013] to determine peak conductance values for connections to excitatory and inhibitory neurons. In detail we derived the peak conductances (¯gAM P A,E and ¯gNM DA,E, and ¯gAM P A,I , ¯gNM DA,I ) for the pinwheel-domain model by exactly following the procedure described in Stimberg et al. [2009] of matching orientation selectivity indices (OSI) for different mapOSIs to the data by Marino et al. [2005] (for definition of OSI and mapOSI see below). For the salt-and-pepper network we matched the peak conductances by using Kolmogorov-Smirnov-tests to compare OSI-distributions with the data by Runyan and Sur [2013], as described in Roy et al. [2013]. In both cases we used a grid search in the space spanned by peak conductances to excitatory and inhibitory neurons and selected the best matching point. Parameter used for peak-conductances are comprised in Tab. 4. 21/37 Table 4. Ligand gated receptors and currents Description Peak rec. GABAAR conductance AMPAR reversal potential NMDAR reversal potential GABAA reversal potential Unit-free magnesium concentration Value Parameter Synaptic -- Currents 549.51 nS Peak aff. AMPAR cond. to exc. neurons ¯gAM P Aaf f,E 0.73¯gAM P Aaf f,E Peak aff. AMPAR cond. to inh. neurons ¯gAM P Aaf f,I 281.8 nS ¯gGABAA 0 mV EAM P A 0 mV ENM DA −70 mV EGABAA 1 mM/M M g pinwheel-domain specific 879.40 nS ¯gAM P A,E 1538.61 nS ¯gAM P A,I 219.80 nS ¯gNM DA,E ¯gNM DA,I 384.65 nS salt-and-pepper specific ¯gAM P A,E ¯gAM P A,I ¯gNM DA,E ¯gNM DA,I Extra-synaptic ¯gamb α H Esom Gamb,E Gamb,E Peak rec. AMPAR conductance to exc. Peak rec. AMPAR conductance to inh. Peak rec. NMDAR conductance to exc. Peak rec. NMDAR conductance to inh. Peak rec. AMPAR conductance to exc. Peak rec. AMPAR conductance to inh. Peak rec. NMDAR conductance to exc. Peak rec. NMDAR conductance to inh. Source * Stimberg et al. [2009] * Destexhe et al. [1998] Destexhe et al. [1998] Destexhe et al. [1998] Jahr and Stevens [1990] * * * * * * * * 659.40 nS 879.20 nS 164.84 nS 219.80 nS 2.6 nS 0.54 M−H 1.5 55 mV 0 -- 2.5 µM 0 -- 2.5 µM Peak extra-synaptic NMDAR conductance Bentzen et al. [2009] Factor based on transition rates Bentzen et al. [2009] Bentzen et al. [2009] Hill-coefficient of extra-synaptic NMDAR Extra-synaptic NMDAR reversal potential Bentzen et al. [2009] Herman and Jahr [2007]◦ Amb. Glut. concentr. to exc. neurons Herman and Jahr [2007]◦ Amb. Glut. concentr. to inh. neurons * Peak synaptic conductances were determined as described in synaptic peak conductance paragraph. The ambient glutamate concentration are varied in a biological plausible range (cf. Herman and Jahr [2007]) Extrasynaptic ligand-gated currents Extrasynaptic NMDA-receptors are activated by ambient glutamate Gamb. Different densities of NMDA-receptors on excitatory and inhibitory neurons provide population specific currents Iamb,X . The currents follow the steady state descriptions of extrasynaptic NMDA-receptors (eNMDAR) in Bentzen et al. [2009]: Iamb,X = ¯gambB[eN M DAo] (V − Esom) with ¯gamb the peak conductance of eNMDARs, B the dynamics of its magnesium block , [eN M DAo] = αGH amb the fraction of open eNMDARs following a power law-dependence on ambient glutamate, and Esom the Nernst-potential of the eNMDARs (Parameters in Tab. 4). 22/37 Background currents We consider the network to be embedded in surrounding neuronal activity. Therefore, each neuron receives synaptic background activity: Ibg,X = gbgXE(t)(V − EbgE) − gbgXI (V − EbgI ) with specific reversal potentials (EbgE,EbgI ) and fluctuating conductances gbgXY , Y ∈ {E, I} the source population, following an Ornstein-Uhlenbeck process: dgbgXY = τbgY (¯gbgXY − gbgXY ) + σbgY dW , with τbgY the mean reversion speed, ¯gbgXY the average background conductance and σbgY the noise strength. V1 network The network layout is similar to Stimberg et al. [2009] and to Roy et al. [2013]. Two populations, an excitatory (size: NE) and an inhibitory (size: NI ) population of neurons represent layer 2-3 of the primary visual cortex, on the one hand for species with a pinwheel-domain organization, e.g., ferrets (Fig. 7C left part), on the other hand for species with a salt-and-pepper organization, e.g., mice (Fig. 7C right part). For both models the excitatory neurons are regularly placed on a 2d-grid of size √ NE. Inhibitory neurons are randomly placed on a third of all grid points. NE × √ Each neuron receives a number (Naf f ) of afferent Poisson inputs with stimulus specific rates. Additionally, excitatory and inhibitory neurons receive fixed specific numbers of recurrent excitatory (Nee, Nie) inputs and a number of recurrent inhibitory (Nci) inputs. The model for mouse features lower numbers of excitatory connections than the one for ferret (cf. Table 5). Independent of species all recurrent connections are randomly drawn, from the same radial symmetric 2d-Gaussian distance distribution, using the algorithm proposed in Efraimidis and Spirakis [2008],  0 P (r) = √ 1/ for r = 0 (no self-connections); 2πσ exp(−r2/2σ2 C) otherwise, with r the distance (in gridpoints) to the presynaptic neuron, and σC the width of the Gaussian. Thereby connections between neighboring neurons are more likely (see 23/37 Figure 7. V1 network model. A Synapses use explicit neurotransmitter descriptions GE and GI , which activate receptors described by extended kinetic schemes, with several closed C, desensitized D and open O stages and constant R and transmitter dependent Rf (G) transition rates. AMPA follows the description in Saftenku [2005], NMDA based on Lester and Jahr [1992], and GABAA on Destexhe et al. [1998]. B The model contains glutamatergic connections to inhibitory and to excitatory neurons, with NMDA & AMPA receptors. Effect of glutamate transporters on the glutamate time course in the two types of connection is separately varied due to the difference in synapse geometry. Extrasynaptic NMDA-receptors are activated by ambient glutamate. C The two one-layered V1 network models are composed of excitatory (brown) and inhibitory (gray) neurons which receive tuned excitatory affentent and lateral inhibitory and excitatory inputs. Neurons a the same location in the network share the preferred orientation, which is either organized in a pinwheel-domain (left) or salt-and-pepper map (right). Lateral connections are drawn from a 2d-Gaussian independent of preferred orientation. Afferent input already carries some tuning. For the pinwheel-domain model every neuron receives input from equally tuned neurons as in Stimberg et al. [2009]. For the salt-and-pepper map afferent tuning width is sampled from independent distributions for exc. and inhibitory neurons. Figure 7A). Each individual connection gets a transmission delay, which comprise synaptic and conduction delays, and is drawn from a gamma distribution Γ(kY , θY ) with shape kY and scale θY parameters specific for the source population Y ∈ {E, I}. Connections at the boundaries are generated using periodic boundary conditions. Parameters can be found in Table 5. 24/37 Orientation preference0.50.01.0connectionprobabilityastrocyteNMDA receptorAMPA receptorglutamate transporterglutamateGABAA receptorGABANMDACn0RnbRnuCn1RnbRnuCn2RndRnrDnRnoRncOnCg0Rgb1GIRgu1Cg1Rgb2GIRgu2Cg2Rgo1Rgc1Og1Rgo2Rgc2Og2GEGEDaRarRadf(GE)CaRaof(GE)RacOawith:f(GE)=G2E(GE+KB)2GABAAAMPAdomainmapOSI = 1.0mapOSI = 0.0Pinwheel-DomainSalt-and-PepperpinwheelAfferent input tuning widthCBAVisual Cortex (V1)excitatory cellsinhibitory cellsastrocyte Table 5. Geometry and stimulation parameters 2500 833 20 4 7 0.6 2.5 0.6 Parameter Value Geometry NE NI NAf f σC kE θE kI θI ferret specific Nee = Nie Nci mouse specific Nee Nie Nci Stimulation fA,max rbase ferret specific wA mouse specific wEA wIA σwEA σwIA 30 Hz 0.1 100 50 25 50 50 27.5 deg. 17.5 deg. 57.5 deg. 16 deg. 48 deg. Description Source Number of excit. neurons Number of inhib. neurons Number of afferent inputs Spread of recurrent conn. (std. dev.) Shape of Gamma distribution exc. conn. Scale of Gamma distribution exc. conn. Shape of Gamma distribution inh. conn. Scale of Gamma distribution inh. conn. Stimberg et al. [2009] Stimberg et al. [2009] Stimberg et al. [2009] Stimberg et al. [2009] Stimberg et al. [2009]* Stimberg et al. [2009]* Stimberg et al. [2009]* Stimberg et al. [2009]* Number of excit. recurrent inputs Number of inhib. recurrent inputs Stimberg et al. [2009] Stimberg et al. [2009] Number of excit. to excit. inputs Number of excit. to inhib. inputs Number of inhib. recurrent inputs Roy et al. [2013] Roy et al. [2013] Roy et al. [2013] max afferent firing rate fraction of stimulus indep. rate Stimberg et al. [2009] Stimberg et al. [2009] input tuning width Stimberg et al. [2009] input tuning width input tuning width input tuning width input tuning width Roy et al. [2013] Roy et al. [2013] Roy et al. [2013] Roy et al. [2013] * Mean and standard deviation of the gamma distribution are matched to the values in Stimberg et al. [2009] Organization of preferred orientations For species which express a pinwheel-domain organization we generate the preferred orientation θ(x, y) for each neuron based on its location (x, y) within a pinwheel-domain map representing 4 pinwheels (see Figure 7C left part). The map is constructed by, first producing a single pinwheel in the first quadrant 1q using equally spaced coordinates −1 ≤ x < 1 and −1 ≤ y < 1 and deriving every neurons preferred angle by: θ1q(x, y) = 90 π atan2(x, y), second the full map is generated by mirroring the first into the other three quadrants. For species without a distinct organization we generate a salt-and-pepper-map by uniformly distributing preferred orientations randomly (see Fig. 7C right part). 25/37 Stimulation All neurons receive a number NAf f of individual afferent Poisson-inputs generated by a full-field stimulation with a static fixed orientation θstim. The neuron specific rate fA (θstim, θ(x, y), wXA(x, y)) = fA,max (cid:32) rbase + (1 − rbase) exp (cid:32) − (θstim − θ(x, y))2 4σ2 XA(x, y) (cid:33)(cid:33) , depends on: selected stimulus orientation θstim, preferred orientation of the neuron θ(x, y), a base-line firing rate rbase, the maximal firing rate for optimal stimulation fA,max, and the population (X ∈ {E, I}) and neuron specific input tuning width wXA(x, y). For ferret all neurons independent of population receive input with identical tuning width wXA(x, y) = wA (Fig. 7C left part). For mouse individual neuron receive specific afferent inputs drawn from two truncated (0-90) Gaussian-distributions differently parametrized for excitatory (mean: wEA, standard deviation σwEA) and inhibitory (mean: wIA, standard deviation σwIA) neurons, cf. Fig. 7C right part, Roy et al. [2013], and Table 5 Blocking glutamate transport For the network we explore two ways how blocking glutamate transport affects tuning. First, at the synapses different decay times for τf EE and τf IE are considered. Second different NMDA-receptor current strengths to excitatory and inhibitory neurons are considered. We allow different Gamb,E and Gamb,I , as a proxy for different densities of NMDA-receptors. Numerical Simulations The synapses and the network model are implemented in Python 2.7 using Brian2 to generate C++ code. We used the Euler-integration scheme provided by the toolbox with an integration step of 0.01 ms. Every simulation is run first for 400 ms as initialization phase without recording data, then for 1600 ms data is recorded. Analyses Two different measures were used to analyze orientation tuning. First the orientation selectivity index (OSI; Swindale [1998]), given by R(θi), 26/37 (cid:118)(cid:117)(cid:117)(cid:116)(cid:32)(cid:88) (cid:33)2 (cid:32)(cid:88) (cid:33)2(cid:44)(cid:88) OSI = R(θi) cos(2θi) + R(θi) sin(2θi) i i i R(θi) is the investigated quantity (e.g. firing rate) observed as response to a stimulation with orientation θi. Stimulation orientations θi have to span the entire range of possible orientations and have to be equally spaced. Values for OSI range from 0 (unselective) to 1 (highly selective). When using the number of sites with a specific orientation preference of neighboring pixels in a radius of 8 pixels instead of R(θi) the OSI-measure can be used to derive the mapOSI, which quantifies the homogeneity of lateral inputs (cf. Fig 7C). As second measure we used the half-width at half max (HWHM) of the response tuning curves. Instead of deriving tuning curves for each individual neuron by stimulating with different orientations we generate pseudo-neurons from a single simulation with one fixed stimulus orientation θ = 43.8 deg. Pseudo-neurons are generated by splitting all excitatory neurons into batches of 50 neurons based on their mapOSI (in pinwheel-domain case) or afferent input tuning width (in salt-and-pepper case). E.g., the 50 neurons with the smallest mapOSI constitute a pseudo neuron. As the 50 neurons have different preferred orientations we can consider these as stimulations with different offsets to the preferred orientation of a pseudo-neuron. Therefore, neurons (of a range of mapOSIs or afferent tuning width) with a preferred orientation close to θ(x, y) = 43.8 will give the response of the pseudo-neuron (with the mapOSI or afferent tuning width) stimulated close to its preferred orientation. Equally spaced stimulations of the pseudo-neuron are obtained by, first fitting a flat-topped von-Mises distributions Swindale [1998] to the pseudo-neuron data, and second selecting points with a 10 deg. difference from the obtained distributions. For the separation in pseudo-neurons close to pinwheel and within domains we used mapOSI ≤ 0.4 and 0.6 < mapOSI ≤ 0.9, respectively. Acknowledgments The authors like to acknowledge the support team of Brian 2, which rapidly developed requested features which made it possible to run the simulations in standalone c++ code. 27/37 References Sushmita L Allam, Viviane S Ghaderi, Jean-Marie C Bouteiller, Arnaud Legendre, Nicolas Ambert, Renaud Greget, Serge Bischoff, Michel Baudry, and Theodore W Berger. A computational model to investigate astrocytic glutamate uptake influence on synaptic transmission and neuronal spiking. Front. Comput. Neurosci., 6(October): 70, jan 2012. ISSN 1662-5188. doi: 10.3389/fncom.2012.00070. URL http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid= 3461576&tool=pmcentrez&rendertype=abstract. Alberto Alvarellos-Gonz´alez, Alejandro Pazos, and Ana B Porto-Pazos. Computational models of neuron-astrocyte interactions lead to improved efficacy in the performance of neural networks. Comput. Math. Methods Med., 2012:476324, jan 2012. ISSN 1748-6718. doi: 10.1155/2012/476324. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=3357509&tool=pmcentrez&rendertype=abstract. Alfonso Araque, Giorgio Carmignoto, Philip G. Haydon, St´ephane H.R. Oliet, Richard Robitaille, and Andrea Volterra. Gliotransmitters Travel in Time and Space. Neuron, 81(4):728 -- 739, feb 2014. ISSN 08966273. doi: 10.1016/j.neuron.2014.02.007. URL http://linkinghub.elsevier.com/retrieve/pii/S0896627314001056. B Barbour. An evaluation of synapse independence. J. Neurosci., 21(20):7969 -- 7984, 2001. ISSN 1529-2401. doi: 21/20/7969[pii]. Boris Barbour, Bernhard U. Keller, Isabel Llano, and Alain Marty. Prolonged presence of glutamate during excitatory synaptic transmission to cerebellar Purkinje cells. Neuron, 12(6):1331 -- 1343, 1994. ISSN 08966273. doi: 10.1016/0896-6273(94)90448-0. Adrienne M Benediktsson, Glen S Marrs, Jian Cheng Tu, Paul F Worley, Jeffrey D Rothstein, Dwight E Bergles, and Michael E Dailey. Neuronal activity regulates glutamate transporter dynamics in developing astrocytes. Glia, 60(2):175 -- 88, feb 2012. ISSN 1098-1136. doi: 10.1002/glia.21249. URL http://www.ncbi.nlm.nih.gov/pubmed/22052455. N C K Bentzen, A M Zhabotinsky, and J L Laugesen. Modeling of glutamate-induced dynamical patterns. Int. J. Neural Syst., 19(6):395 -- 407, dec 2009. ISSN 0129-0657. 28/37 doi: 10.1142/S0129065709002105. URL http://www.ncbi.nlm.nih.gov/pubmed/20039463. Rita Bopp, Nuno Ma¸carico da Costa, Bjorn M Kampa, Kevan a C Martin, and Morgane M Roth. Pyramidal Cells Make Specific Connections onto Smooth (GABAergic) Neurons in Mouse Visual Cortex. PLoS Biol., 12(8):e1001932, 2014. ISSN 1545-7885. doi: 10.1371/journal.pbio.1001932. URL http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid= 4138028&tool=pmcentrez&rendertype=abstract. F a Chaudhry, K P Lehre, M van Lookeren Campagne, O P Ottersen, N C Danbolt, and J Storm-Mathisen. Glutamate transporters in glial plasma membranes: highly differentiated localizations revealed by quantitative ultrastructural immunocytochemistry. Neuron, 15(3):711 -- 720, 1995. ISSN 08966273. doi: 10.1016/0896-6273(95)90158-2. Naiyan Chen, Hiroki Sugihara, Jitendra Sharma, Gertrudis Perea, Jeremy Petravicz, Chuong Le, and Mriganka Sur. Nucleus basalis-enabled stimulus-specific plasticity in the visual cortex is mediated by astrocytes. Proc. Natl. Acad. Sci. U. S. A., 109(41): E2832 -- 41, oct 2012. ISSN 1091-6490. doi: 10.1073/pnas.1206557109. URL http://www.ncbi.nlm.nih.gov/pubmed/23012414. J D Clements, R a Lester, G Tong, Craig E Jahr, and G L Westbrook. The time course of glutamate in the synaptic cleft. Science, 258(5087):1498 -- 501, nov 1992. ISSN 0036-8075. URL http://www.ncbi.nlm.nih.gov/pubmed/1359647. N C Danbolt. Glutamate uptake. Prog. Neurobiol., 65(1):1 -- 105, sep 2001. ISSN 0301-0082. URL http://www.pubmedcentral.nih.gov/articlerender.fcgi? artid=2775085&tool=pmcentrez&rendertype=abstract. Yaron David, Luisa P Cacheaux, Sebastian Ivens, Ezequiel Lapilover, Uwe Heinemann, Daniela Kaufer, and Alon Friedman. Astrocytic dysfunction in epileptogenesis: consequence of altered potassium and glutamate homeostasis? J. Neurosci., 29(34): 10588 -- 99, aug 2009. ISSN 1529-2401. doi: 10.1523/JNEUROSCI.2323-09.2009. URL http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid= 2875068&tool=pmcentrez&rendertype=abstract. 29/37 Maurizio De Pitt`a, Vladislav Volman, Hugues Berry, and Eshel Ben-Jacob. A tale of two stories: astrocyte regulation of synaptic depression and facilitation. PLoS Comput. Biol., 7(12):e1002293, dec 2011. ISSN 1553-7358. doi: 10.1371/journal.pcbi.1002293. URL http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid= 3228793&tool=pmcentrez&rendertype=abstracthttp: //dx.plos.org/10.1371/journal.pcbi.1002293. Maurizio De Pitt`a, Vladislav Volman, Hugues Berry, Vladimir Parpura, Andrea Volterra, and Eshel Ben-Jacob. Computational quest for understanding the role of astrocyte signaling in synaptic transmission and plasticity. Front. Comput. Neurosci., 6(December):1 -- 25, 2012. ISSN 1662-5188. doi: 10.3389/fncom.2012.00098. URL http://www.frontiersin.org/Computational_Neuroscience/10.3389/fncom. 2012.00098/abstract. Alain Destexhe and D Par´e. Impact of network activity on the integrative properties of neocortical pyramidal neurons in vivo. J. Neurophysiol., 81(4):1531 -- 47, apr 1999. ISSN 0022-3077. URL http://www.ncbi.nlm.nih.gov/pubmed/10200189. Alain Destexhe, Z F Mainen, and Terrence J Sejnowski. Kinetic models of synaptic transmission. In I Koch, C. Segev, editor, Methods neuronal Model., pages 1 -- 25. MIT Press, Cambridge, 2nd editio edition, 1998. URL http://citeseerx.ist.psu.edu/ viewdoc/download?doi=10.1.1.164.9768&amp;rep=rep1&amp;type=pdf. Alain Destexhe, Michelle Rudolph, J M Fellous, and Terrence J Sejnowski. Fluctuating synaptic conductances recreate in vivo-like activity in neocortical neurons. Neuroscience, 107(1):13 -- 24, jan 2001. ISSN 0306-4522. URL http://www.ncbi.nlm.nih.gov/pubmed/11744242. Jeffrey S Diamond. Neuronal glutamate transporters limit activation of NMDA receptors by neurotransmitter spillover on CA1 pyramidal cells. J. Neurosci., 21(21): 8328 -- 38, nov 2001. ISSN 1529-2401. URL http://www.ncbi.nlm.nih.gov/pubmed/11606620. Jeffrey S Diamond. Deriving the glutamate clearance time course from transporter currents in CA1 hippocampal astrocytes: transmitter uptake gets faster during development. J. Neurosci., 25(11):2906 -- 16, mar 2005. ISSN 1529-2401. doi: 30/37 10.1523/JNEUROSCI.5125-04.2005. URL http://www.ncbi.nlm.nih.gov/pubmed/15772350. Christopher B. Divito and Suzanne M. Underhill. Excitatory amino acid transporters: Roles in glutamatergic neurotransmission. Neurochem. Int., 73(1):172 -- 180, 2014. ISSN 18729754. doi: 10.1016/j.neuint.2013.12.008. URL http://dx.doi.org/10.1016/j.neuint.2013.12.008. Pavlos Efraimidis and Paul Spirakis. Weighted Random Sampling. In Ming-Yang Kao, editor, Encycl. Algorithms, pages 1024 -- 1027. Springer US, 2008. ISBN 978-0-387-30770-1. doi: 10.1007/978-0-387-30162-4\ 478. URL http://dx.doi.org/10.1007/978-0-387-30162-4_478. Dominik Freche, Ulrike Pannasch, Nathalie Rouach, and David Holcman. Synapse geometry and receptor dynamics modulate synaptic strength. PLoS One, 6(10), 2011. ISSN 19326203. doi: 10.1371/journal.pone.0025122. G. Grynkiewicz, M. Poenie, and R. Y. Tsien. A new generation of Ca2+ indicators with greatly improved fluorescence properties. J. Biol. Chem., 260(6):3440 -- 3450, 1985. ISSN 00219258. doi: 3838314. a I Guly´as, M Meg´ıas, Z Emri, and T F Freund. Total number and ratio of excitatory and inhibitory synapses converging onto single interneurons of different types in the CA1 area of the rat hippocampus. J. Neurosci., 19(22):10082 -- 10097, 1999. ISSN 1529-2401. Philip G Haydon and Maiken Nedergaard. How Do Astrocytes Participate in Neural Plasticity? Cold Spring Harb. Perspect. Biol., pages 1 -- 16, 2014. ISSN 1943-0264. doi: 10.1101/cshperspect.a020438. URL http: //cshperspectives.cshlp.org/lookup/doi/10.1101/cshperspect.a020438. Fritjof Helmchen and Winfried Denk. Deep tissue two-photon microscopy. Nat. Methods, 2(12):932 -- 940, 2005. ISSN 1548-7091. doi: 10.1038/nmeth818. Melissa A Herman and Craig E Jahr. Extracellular glutamate concentration in hippocampal slice. J. Neurosci., 27(36):9736 -- 41, sep 2007. ISSN 1529-2401. doi: 10.1523/JNEUROSCI.3009-07.2007. URL 31/37 http://www.ncbi.nlm.nih.gov/pubmed/17804634http: //www.pubmedcentral.nih.gov/articlerender.fcgi?artid=2670936&tool= pmcentrez&rendertype=abstract. Craig E Jahr and C F Stevens. Voltage dependence of NMDA-activated macroscopic conductances predicted by single-channel kinetics. J. Neurosci., 10(9):3178 -- 3182, sep 1990. URL http://www.jneurosci.org/content/10/9/3178.abstract. Helmut J Koester and Daniel Johnston. Target cell-dependent normalization of transmitter release at neocortical synapses. Science, 308(5723):863 -- 866, 2005. ISSN 0036-8075. doi: 10.1126/science.1100815. Brian Roland Larsen, Mette Assentoft, Maria L. Cotrina, Susan Z. Hua, Maiken Nedergaard, Kai Kaila, Juha Voipio, and Nanna Macaulay. Contributions of the Na+/K+-ATPase, NKCC1, and Kir4.1 to hippocampal K+ clearance and volume responses. Glia, 62:608 -- 622, 2014. ISSN 08941491. doi: 10.1002/glia.22629. R a Lester and Craig E Jahr. NMDA channel behavior depends on agonist affinity. J. Neurosci., 12(2):635 -- 43, feb 1992. ISSN 0270-6474. URL http://www.ncbi.nlm.nih.gov/pubmed/1346806. Ashok Litwin-Kumar and Brent Doiron. Slow dynamics and high variability in balanced cortical networks with clustered connections. Nat. Neurosci., 15(11):1498 -- 1505, 2012. ISSN 1097-6256. doi: 10.1038/nn.3220. Jorge Marino, James Schummers, David C Lyon, Lars Schwabe, Oliver Beck, Peter Wiesing, Klaus Obermayer, and Mriganka Sur. Invariant computations in local cortical networks with balanced excitation and inhibition. Nat. Neurosci., 8(2): 194 -- 201, feb 2005. ISSN 1097-6256. doi: 10.1038/nn1391. URL http://www.ncbi.nlm.nih.gov/pubmed/15665876. Marzia Martina, Nicholas V Krasteniakov, and Richard Bergeron. D-Serine differently modulates NMDA receptor function in rat CA1 hippocampal pyramidal cells and interneurons. J. Physiol., 548(Pt 2):411 -- 423, 2003. ISSN 0022-3751. doi: 10.1113/jphysiol.2002.037127. 32/37 Marzia Martina, Tanya Comas, and Geoffrey a R Mealing. Selective pharmacological modulation of pyramidal neurons and interneurons in the CA1 region of the rat hippocampus. Front. Pharmacol., 4 MAR(March):1 -- 16, 2013. ISSN 16639812. doi: 10.3389/fphar.2013.00024. M Meg´ıas, Z Emri, T F Freund, and A I Guly´as. Total number and distribution of inhibitory and excitatory synapses on hippocampal CA1 pyramidal cells. Neuroscience, 102(3):527 -- 540, 2001. ISSN 0306-4522. doi: S0306-4522(00)00496-6[pii]. S Mennerick, W Shen, W Xu, a Benz, K Tanaka, K Shimamoto, K E Isenberg, J E Krause, and C F Zorumski. Substrate turnover by transporters curtails synaptic glutamate transients. J. Neurosci., 19(21):9242 -- 9251, 1999. ISSN 1529-2401. Ming Yuan Min, Dmitri A. Rusakov, and Dimitri M. Kullmann. Activation of AMPA, kainate, and metabotropic receptors at hippocampal mossy fiber synapses: Role of glutamate diffusion. Neuron, 21(3):561 -- 570, 1998. ISSN 08966273. doi: 10.1016/S0896-6273(00)80566-8. Ciaran Murphy-royal, Julien P Dupuis, Juan a Varela, Aude Panatier, Benoıt Pinson, J´erome Baufreton, Laurent Groc, and St´ephane H R Oliet. Surface diffusion of astrocytic glutamate transporters shapes synaptic transmission. Nat. Neurosci., 18(2), 2015. doi: 10.1038/nn.3901. Suhita Nadkarni, Peter Jung, and Herbert Levine. Astrocytes optimize the synaptic transmission of information. PLoS Comput. Biol., 4(5):e1000088, may 2008. ISSN 1553-7358. doi: 10.1371/journal.pcbi.1000088. URL http://www.ncbi.nlm.nih.gov/pubmed/18516277. Maiken Nedergaard and Alexei Verkhratsky. Artifact versus reality-How astrocytes contribute to synaptic events? Glia, 1023(January):1013 -- 1023, jan 2012. ISSN 1098-1136. doi: 10.1002/glia.22288. URL http://www.ncbi.nlm.nih.gov/pubmed/22228580. Aude Panatier, Joanne Vall´ee, Michael Haber, Keith K Murai, Jean-Claude Lacaille, and Richard Robitaille. Astrocytes are endogenous regulators of basal transmission at central synapses. Cell, 146(5):785 -- 98, sep 2011. ISSN 1097-4172. doi: 10.1016/j.cell.2011.07.022. URL http://www.ncbi.nlm.nih.gov/pubmed/21855979. 33/37 Gertrudis Perea, Marta Navarrete, and Alfonso Araque. Tripartite synapses: astrocytes process and control synaptic information. Trends Neurosci., 32(8):421 -- 31, aug 2009. ISSN 1878-108X. doi: 10.1016/j.tins.2009.05.001. URL http://www.ncbi.nlm.nih.gov/pubmed/19615761. Gertrudis Perea, Mriganka Sur, and Alfonso Araque. Neuron-glia networks: integral gear of brain function. Front. Cell. Neurosci., 8(November):1 -- 8, 2014. ISSN 1662-5102. doi: 10.3389/fncel.2014.00378. Jeremy Petravicz, N Mellios, Sami El Boustani, C Le, and Mriganka Sur. Role of astrocyte glutamate transporters in ocular dominance plasticity and response properties of visual cortex. In Soc. Neurosci., page 127.14, 2014. Davide Reato, Mario Cammarota, Lucas C Parra, and Giorgio Carmignoto. Computational model of neuron-astrocyte interactions during focal seizure generation. Front. Comput. Neurosci., 6(October):81, jan 2012. ISSN 1662-5188. doi: 10.3389/fncom.2012.00081. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=3467689&tool=pmcentrez&rendertype=abstract. Alfonso Renart, Rub´en Moreno-Bote, Xiao-Jing Wang, and N´estor Parga. Mean-driven and fluctuation-driven persistent activity in recurrent networks. Neural Comput., 19 (1):1 -- 46, 2007. ISSN 0899-7667. doi: 10.1162/neco.2007.19.1.1. Christine R Rose and Claudia Karus. Two sides of the same coin: Sodium homeostasis and signaling in astrocytes under physiological and pathophysiological conditions. Glia, pages 1 -- 15, apr 2013. ISSN 1098-1136. doi: 10.1002/glia.22492. URL http://www.ncbi.nlm.nih.gov/pubmed/23553639. Jeffrey D. Rothstein, Margaret Dykes-Hoberg, Carlos A. Pardo, Lynn A. Bristol, Lin Jin, Ralph W. Kuncl, Yoshikatsu Kanai, Matthias A. Hediger, Yanfeng Wang, Jerry P. Schielke, and Devin F. Welty. Knockout of glutamate transporters reveals a major role for astroglial transport in excitotoxicity and clearance of glutamate. Neuron, 16(3):675 -- 686, 1996. ISSN 08966273. doi: 10.1016/S0896-6273(00)80086-0. Dipanjan Roy, Yenni Tjandra, Konstantin Mergenthaler, Jeremy Petravicz, Caroline A Runyan, Nathan R Wilson, Mriganka Sur, and Klaus Obermayer. Afferent specificity, 34/37 feature specific connectivity influence orientation selectivity: A computational study in mouse primary visual cortex. arXiv Prepr. arXiv1301.0996, page 39, 2013. URL http://arxiv.org/abs/1301.0996. Caroline a Runyan and Mriganka Sur. Response selectivity is correlated to dendritic structure in parvalbumin-expressing inhibitory neurons in visual cortex. J. Neurosci., 33(28):11724 -- 33, jul 2013. ISSN 1529-2401. doi: 10.1523/JNEUROSCI.2196-12.2013. URL http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid= 3724550&tool=pmcentrez&rendertype=abstract. Dmitri A. Rusakov. The role of perisynaptic glial sheaths in glutamate spillover and extracellular Ca(2+) depletion. Biophys. J., 81(4):1947 -- 59, oct 2001. ISSN 0006-3495. doi: 10.1016/S0006-3495(01)75846-8. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=1301670&tool=pmcentrez&rendertype=abstract. Dmitri A. Rusakov and D M Kullmann. Extrasynaptic glutamate diffusion in the hippocampus: ultrastructural constraints, uptake, and receptor activation. J. Neurosci., 18(9):3158 -- 70, may 1998. ISSN 0270-6474. URL http://www.ncbi.nlm.nih.gov/pubmed/9547224. Dmitri A. Rusakov, Lucie Bard, Michael G. Stewart, and Christian Henneberger. Diversity of astroglial functions alludes to subcellular specialisation. Trends Neurosci., 37(4):228 -- 242, apr 2014. ISSN 01662236. doi: 10.1016/j.tins.2014.02.008. URL http://linkinghub.elsevier.com/retrieve/pii/S0166223614000228. E E Saftenku. Modeling of slow glutamate diffusion and AMPA receptor activation in the cerebellar glomerulus. J. Theor. Biol., 234(3):363 -- 82, jun 2005. ISSN 0022-5193. doi: 10.1016/j.jtbi.2004.11.036. URL http://www.ncbi.nlm.nih.gov/pubmed/15784271. James Schummers, Beau Cronin, Klaus Wimmer, Marcel Stimberg, Robert Martin, Klaus Obermayer, Konrad Koerding, and Mriganka Sur. Dynamics of orientation tuning in cat v1 neurons depend on location within layers and orientation maps. Front. Neurosci., 1(1):145 -- 59, nov 2007. ISSN 1662-453X. doi: 10.3389/neuro.01.1.1.011.2007. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=2570087&tool=pmcentrez&rendertype=abstract. 35/37 James Schummers, Hongbo Yu, and Mriganka Sur. Tuned responses of astrocytes and their influence on hemodynamic signals in the visual cortex. Science, 320(5883): 1638 -- 43, jun 2008. ISSN 1095-9203. doi: 10.1126/science.1156120. URL http://www.ncbi.nlm.nih.gov/pubmed/18566287. Annalisa Scimemi and Marco Beato. Determining the neurotransmitter concentration profile at active synapses. Mol. Neurobiol., 40(3):289 -- 306, dec 2009. ISSN 1559-1182. doi: 10.1007/s12035-009-8087-7. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=2777263&tool=pmcentrez&rendertype=abstract. Annalisa Scimemi, Hua Tian, and Jeffrey S Diamond. Neuronal transporters regulate glutamate clearance, NMDA receptor activation, and synaptic plasticity in the hippocampus. J. Neurosci., 29(46):14581 -- 95, nov 2009. ISSN 1529-2401. doi: 10.1523/JNEUROSCI.4845-09.2009. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=2853250&tool=pmcentrez&rendertype=abstract. S. Shushruth, P. Mangapathy, J. M. Ichida, P. C. Bressloff, L. Schwabe, and a. Angelucci. Strong Recurrent Networks Compute the Orientation Tuning of Surround Modulation in the Primate Primary Visual Cortex. J. Neurosci., 32(1): 308 -- 321, 2012. ISSN 0270-6474. doi: 10.1523/JNEUROSCI.3789-11.2012. Marcel Stimberg, Klaus Wimmer, Robert Martin, Lars Schwabe, Jorge Marino, James Schummers, David C Lyon, Mriganka Sur, and Klaus Obermayer. The Operating Regime of Local Computations in Primary Visual Cortex. Cereb. Cortex, 19(9): 2166 -- 2180, 2009. URL http: //www.culturacientifica.org/publicarch/cerebral_cortex_09_adv.pdf. N V Swindale. Orientation tuning curves : empirical description and estimation of parameters. Biol. Cybern., 78:45 -- 56, 1998. K Tarczy-Hornoch, K A C Martin, J J B Jack, and K J Stratford. Synaptic interactions bewteen smooth and spiny neurones in layer 4 of cat visual cortex \textit{in vitro}. J. Physiol., 508(2):351 -- 363, 1998. C. G. Thomas, Hua Tian, and Jeffrey S Diamond. The Relative Roles of Diffusion and Uptake in Clearing Synaptically Released Glutamate Change during Early Postnatal 36/37 Development. J. Neurosci., 31(12):4743 -- 4754, mar 2011. ISSN 0270-6474. doi: 10.1523/JNEUROSCI.5953-10.2011. URL http://www.jneurosci.org/cgi/doi/10.1523/JNEUROSCI.5953-10.2011. Gang Tong and Craig E. Jahr. Block of glutamate transporters potentiates postsynaptic excitation. Neuron, 13(5):1195 -- 1203, 1994. ISSN 08966273. doi: 10.1016/0896-6273(94)90057-4. Shota Tsukada, Masae Iino, Yukihiro Takayasu, Keiko Shimamoto, and Seiji Ozawa. Effects of a novel glutamate transporter blocker, (2S, 3S)-3-[3-[4-(trifluoromethyl)benzoylamino]benzyloxy]aspartate (TFB-TBOA), on activities of hippocampal neurons. Neuropharmacology, 48(4):479 -- 91, mar 2005. ISSN 0028-3908. doi: 10.1016/j.neuropharm.2004.11.006. URL http://www.ncbi.nlm.nih.gov/pubmed/15755476. E. S. Vizi, a. Fekete, R. Karoly, and a. Mike. Non-synaptic receptors and transporters involved in brain functions and targets of drug treatment. Br. J. Pharmacol., 160(4): 785 -- 809, 2010. ISSN 00071188. doi: 10.1111/j.1476-5381.2009.00624.x. Kaiyu Zheng, Annalisa Scimemi, and Dmitri A. Rusakov. Receptor actions of synaptically released glutamate: the role of transporters on the scale from nanometers to microns. Biophys. J., 95(10):4584 -- 96, nov 2008. ISSN 1542-0086. doi: 10.1529/biophysj.108.129874. URL http://www.pubmedcentral.nih.gov/ articlerender.fcgi?artid=2576387&tool=pmcentrez&rendertype=abstract. 37/37
1209.5029
1
1209
2012-09-23T01:17:02
Sparse Codes for Speech Predict Spectrotemporal Receptive Fields in the Inferior Colliculus
[ "q-bio.NC" ]
We have developed a sparse mathematical representation of speech that minimizes the number of active model neurons needed to represent typical speech sounds. The model learns several well-known acoustic features of speech such as harmonic stacks, formants, onsets and terminations, but we also find more exotic structures in the spectrogram representation of sound such as localized checkerboard patterns and frequency-modulated excitatory subregions flanked by suppressive sidebands. Moreover, several of these novel features resemble neuronal receptive fields reported in the Inferior Colliculus (IC), as well as auditory thalamus and cortex, and our model neurons exhibit the same tradeoff in spectrotemporal resolution as has been observed in IC. To our knowledge, this is the first demonstration that receptive fields of neurons in the ascending mammalian auditory pathway beyond the auditory nerve can be predicted based on coding principles and the statistical properties of recorded sounds.
q-bio.NC
q-bio
Sparse Codes for Speech Predict Spectrotemporal Receptive Fields in the Inferior Colliculus Nicole L. Carlson1,2, Vivienne L. Ming1, Michael Robert DeWeese1,2,3* 1 Redwood Center for Theoretical Neuroscience, University of California, Berkeley, California, United States of America, 2 Department of Physics, University of California, Berkeley, California, United States of America, 3 Helen Wills Neuroscience Institute, University of California, Berkeley, California, United States of America Abstract We have developed a sparse mathematical representation of speech that minimizes the number of active model neurons needed to represent typical speech sounds. The model learns several well-known acoustic features of speech such as harmonic stacks, formants, onsets and terminations, but we also find more exotic structures in the spectrogram representation of sound such as localized checkerboard patterns and frequency-modulated excitatory subregions flanked by suppressive sidebands. Moreover, several of these novel features resemble neuronal receptive fields reported in the Inferior Colliculus (IC), as well as auditory thalamus and cortex, and our model neurons exhibit the same tradeoff in spectrotemporal resolution as has been observed in IC. To our knowledge, this is the first demonstration that receptive fields of neurons in the ascending mammalian auditory pathway beyond the auditory nerve can be predicted based on coding principles and the statistical properties of recorded sounds. Citation: Carlson NL, Ming VL, DeWeese MR (2012) Sparse Codes for Speech Predict Spectrotemporal Receptive Fields in the Inferior Colliculus. PLoS Comput Biol 8(7): e1002594. doi:10.1371/journal.pcbi.1002594 Editor: Tim Behrens, University of Oxford, United Kingdom Received November 16, 2011; Accepted May 18, 2012; Published July 12, 2012 Copyright: ß 2012 Carlson et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Funding: NLC was supported by a National Science Foundation graduate fellowship; MRD gratefully acknowledges support from the National Science Foundation (www.nsf.gov), the McKnight Foundation (www.mcknight.org), the McDonnell Foundation (www.jsmf.org), and the Hellman Family Faculty Fund. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing Interests: The authors have declared that no competing interests exist. * E-mail: [email protected] Introduction Our remarkable ability to interpret the highly structured sounds in our everyday environment suggests that auditory processing in the brain is somehow specialized for natural sounds. Many authors have postulated that the brain tries to transmit and encode information efficiently, so as to minimize the energy expended [1], reduce redundancy [2–4], maximize information flow [5–8], or facilitate computations at later stages of processing [9], among other possible objectives. One way to create an efficient code is to enforce population sparseness, having only a few active neurons at a time. Sparse coding schemes pick out the statistically important features of a signal — those features that occur much more often than chance — which can then be used to efficiently represent a complex signal with few active neurons. The principle of sparse coding has led to important insights into the neural encoding of visual scenes within the primary visual cortex (V1). Sparse coding of natural images revealed local, oriented edge-detectors that qualitatively match the receptive fields of simple cells in V1 [10]. More recently, overcomplete sparse coding schemes have uncovered a greater diversity of features that more closely matches the full range of simple cell receptive field shapes found in V1 [11]. An encoding is called overcomplete if the number of neurons available to represent the stimulus is larger than the dimensionality of the input. This is a biologically realistic property for a model of sensory processing because information is encoded by increasing numbers of neurons as it travels from the optic nerve to higher stages in the visual path- way, just as auditory sensory information is encoded by increasing numbers of neurons as it travels from the auditory nerve to higher processing stages [12]. Despite experimental evidence for sparse coding in the auditory system [13,14], there have been fewer theoretical sparse coding studies in audition than in vision. However, there has been progress, particularly for the earliest stages of auditory processing. Sparse coding of raw sound pressure level waveforms of natural sounds produced a ‘‘dictionary’’ of acoustic filters closely resem- bling the impulse response functions of auditory nerve fibers [15,16]. Acoustic features learned by this model were best fit to the neural data for a particular combination of animal vocalizations and two subclasses of environmental sounds. Intriguingly, they found that training on speech alone produced features that were just as well-fit to the neural data as the optimal combination of natural sounds, suggesting that speech provides the right mixture of acoustic features for probing and predicting the properties of the mammalian auditory system. Another pioneering sparse coding study [17] took as its starting point speech that was first preprocessed using a model of the cochlea — one of several so-called cochleogram representations of sound. This group found relatively simple acoustic features that were fairly localized in time and frequency as well as some temporally localized harmonic stacks. These results were roughly consistent with some properties of receptive fields in primary auditory cortex (A1), but modeled responses did not capture the majority of spectrotemporal the specific shapes of neuronal receptive fields (STRFs; [18]) reported in the literature. That study only considered undercomplete dictionaries, and it focused solely on a ‘‘soft’’ sparse coding model that minimized the mean PLoS Computational Biology www.ploscompbiol.org 1 July 2012 Volume 8 Issue 7 e1002594 Author Summary The receptive field of a neuron can be thought of as the stimulus that most strongly causes it to be active. Scientists have long been interested in discovering the underlying principles that determine the structure of receptive fields of cells in the auditory pathway to better understand how our brains process sound. One possible way of predicting these receptive fields is by using a theoretical model such as a sparse coding model. In such a model, each sound is represented by the smallest possible number of active model neurons chosen from a much larger group. A primary question addressed in this study is whether the receptive fields of model neurons optimized for natural sounds will predict receptive fields of actual neurons. Here, we use a sparse coding model on speech data. We find that our model neurons do predict receptive fields of auditory neurons, specifically in the Inferior Colliculus (midbrain) as well as the thalamus and cortex. To our knowledge, this is the first time any theoretical model has been able to predict so many of the diverse receptive fields of the various cell-types in those areas. activity of the model’s neurons, as opposed to ‘‘hard’’ sparse models that minimize the number of active neurons. The same group also considered undercomplete, soft sparse coding of spectrograms of speech [19], which did yield some STRFs showing multiple subfields and temporally modulated harmonic stacks, but the range of STRF shapes they reported was still modest compared with what has been seen experimentally in auditory midbrain, thalamus, or cortex. Another recent study considered sparse coding of music [20] in order to develop auto- mated genre classifiers. To our knowledge, there are no published studies of complete or overcomplete, sparse coding of either spectrograms or cochleo- grams of speech or natural sounds. We note that one preliminary sparse coding study utilizing a complete dictionary trained on spectrograms did find STRFs resembling formants, onset-sensitive neurons, and harmonic stacks (J. Wang, B.A. Olshausen, and V.L. Ming, COSYNE 2008) but they did not obtain novel acoustic features, nor any that closely resembled STRFs from the auditory system. Our goal is two-fold. First, we test whether an overcomplete, hard sparse coding model trained on spectrograms of speech can more fully reveal the structure of natural sounds than previous models. Second, we ask whether our model can accurately predict receptive fields in the ascending auditory pathway beyond the auditory nerve. We have found that, when trained on spectro- grams of human speech, an overcomplete, hard sparse coding model does learn features resembling those of STRF shapes previously reported in the inferior colliculus (IC), as well as auditory thalamus and cortex. Moreover, our model exhibits a similar tradeoff in spectrotemporal resolution as previously reported in IC. Finally, our model has identified novel acoustic features for probing the response properties of neurons in the auditory pathway that have thus far resisted classification and meaningful analysis. Results Sparse Coding Model of Speech In order to uncover important acoustic features that can inform us about how the nervous system processes natural sounds, we Sparse Coding Predicts IC Receptive Fields have developed a sparse coding model of human speech (see Methods for details). As illustrated in Fig. 1, raw sound pressure level waveforms of recorded speech were first preprocessed by one of two simple models of the peripheral auditory system. The first of these preprocessing models was the spectrogram, which can be thought of as the power spectrum of short segments of the original waveform at each moment in time. We also explored an alter- native preprocessing step that was meant to more accurately model the cochlea [21,22]; the original waveform was sent through a filter bank with center frequencies based on the pro- perties of cochlear nerve fibers. Both models produced represen- tations of the waveform as power at different frequencies over time. The spectrograms (cochleograms) were then separated into segments of length 216 ms (250 ms). Because of the high dimensionality of these training examples, we performed prin- cipal components analysis (PCA) and retained only the first two hundred components to reduce the dimensionality (from 25625~6,400 values down to 200), as was done previously in some visual [23] and auditory [17] sparse coding studies; the latter group also performed the control of repeating their analysis without the PCA step and they found that their results did not change. We then trained a ‘‘dictionary’’ of model neurons that could encode this data using the Locally Competitive Algorithm (LCA), a recently developed sparse encoding algorithm [24]. This flexible algorithm allowed us to approximately enforce either the so-called ‘‘hard’’ sparseness (L0 sparseness; minimizing the number of simultaneously active model neurons) or ‘‘soft’’ sparseness (L1 sparseness; minimizing the sum of all simultaneous activity across all model neurons) during encoding by our choice of thresholding function. Additionally, we explored the effect of dictionary over- completeness (with respect to the number of principal components) by training dictionaries that were half-complete, complete, or overcomplete (two or four times). Following training, the various resulting dictionaries were analyzed for cell-types and compared to experimental receptive fields reported in the literature. Cochleogram-Trained Models In general, training our network on cochleogram representa- tions of speech resulted in smooth and simple shapes for the learned receptive fields of model neurons. Klein and colleagues [17] used a sparse coding algorithm that imposed an L1-like sparseness constraint to learn a half-complete dictionary of cochleograms. Their dictionary elements consisted of harmonic stacks at the lower frequencies and localized elements at the higher frequencies. To make contact with these results, we trained a half- complete L0-sparse dictionary on cochleograms and compared the response properties of our model neurons with those of the previous study. The resulting dictionary (Fig. 2) consists of similar shapes to this previous work with the exception of one ‘‘onset element’’ in the upper left (this is the least used of all of the elements from this dictionary). Subsequent simulations revealed that the form of the dictionary is strongly dependent on the degree of overcompleteness. Even a complete dictionary exhibits a greater diversity of shapes than this half-complete dictionary (Fig. S11). This was true for L1-sparse dictionaries trained with LCA [24] or with Sparsenet [10] (Figs. S15 and S19). The inability of the half-comp lete dictionary to produce the more comp lex receptive field shapes of the complete dictio- nary, such as those resembling STRFs measured in IC, or those in auditory thalamus or cortex, suggests that overcomp leteness in those regions is crucial to the flexibi lity of their auditory codes. PLoS Computational Biology www.ploscompbiol.org 2 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 1. Schematic illustration of our sparse coding model. (a) Stimuli used to train the model consisted of examples of recorded speech. The blue curve represents the raw sound pressure waveform of a woman saying, ‘‘The north wind and the sun were disputing which was the stronger, when a traveler came along wrapped in a warm cloak.’’ (b) The raw waveforms were first put through one of two preprocessing steps meant to model the earliest stages of auditory processing to produce either a spectrogram or a ‘‘cochleogram’’ (not shown; see Methods for details). In either case, the power spectrum across acoustic frequencies is displayed as a function of time, with warmer colors indicating high power content and cooler colors indicating low power. (c) The spectrograms were then divided into overlapping 216 ms segments. (d) Subsequently, principal components analysis (PCA) was used to project each segment onto the space of the first two hundred principal components (first ten shown), in order to reduce the dimensionality of the data to make it tractable for further analysis while retaining its basic structure [17]. (e) These projections PLoS Computational Biology www.ploscompbiol.org 3 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields were then input to a sparse coding network in order to learn a ‘‘dictionary’’ of basis elements analogous to neuronal receptive fields, which can then be used to form a representation of any given stimulus (i.e., to perform inference). We explored networks capable of learning either ‘‘hard’’ (L0) sparse dictionaries or ‘‘soft’’ (L1) sparse dictionaries (described in the text and Methods) that were undercomplete (fewer dictionary elements than PCA components), complete (equal number of dictionary elements), or over-complete (greater number of dictionary elements). doi:10.1371/journal.pcbi.1002594.g001 Spectrogram-Trained Models The spectrogram-trained dictionaries provide a much richer and more diverse set of dictionary elements than those trained on cochleograms. We display representative elements of the different categories of shapes found in a half-complete L0-sparse spectro- gram dictionary (Fig. 3a–f) along with a histogram of the usage of the elements (Fig. 3g) when used to represent individual sounds drawn from the training set (i.e., during inference). Interestingly, we find that the different qualitative types of neurons separate according to their usage into a series of rises and plateaus. The least used elements are the harmonic stacks (Fig. 3a), which is perhaps unsurprising since, in principle, only one of them needs to be active at many points in time for a typical epoch of a recording from a single human speaker. We note that, while such harmonic stack receptive fields are apparently rare in the colliculus, thalamus, and cortex, they are well represented in the dorsal cochlear nucleus (DCN) (e.g., see Fig. 5b in [25]). The neighboring flat region consists of onset elements (Fig. 3b), which contain broad frequency subfields that change abruptly at one moment in time. These neurons were all used approximately equally often across the training set since it is equally probable that a stimulus transient will occur any time during the 216 ms time window. The third region consists of more complex harmonic stacks that contain low-power subfields on the sides (Fig. 3c), a feature Figure 2. A half-complete sparse coding dictionary trained on cochleogram representations of speech. This dictionary exhibits a limited range of shapes. The full set of 100 elements from a half-complete, L0-sparse dictionary trained on cochleograms of human speech resemble those found in a previous study [17]. Nearly all elements are extremely smooth, with most consisting of a single frequency subfield or an unmodulated harmonic stack. Each rectangle can be thought of as representing the spectro-temporal receptive field (STRF) of a single element in the dictionary (see Methods for details); time is plotted along the horizontal axis (from 0 to 250 ms), and log frequency is plotted along the vertical axis, with frequencies ranging from 73 Hz to 7630 Hz. Color indicates the amount of power present at each frequency at each moment in time, with warm colors representing high power and cool colors representing low power. Each element has been normalized to have unit Euclidean length. Elements are arranged in order of their usage during inference (i.e., when used to represent individual sounds drawn from the training set) with usage increasing from left to right along each row, and all elements of lower rows used more than those of higher rows. doi:10.1371/journal.pcbi.1002594.g002 PLoS Computational Biology www.ploscompbiol.org 4 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 3. A half-complete, L0-sparse dictionary trained on spectrograms of speech. This dictionary exhibits a variety of distinct shapes that capture several classes of acoustic features present in speech and other natural sounds. (a–f) Selected elements from the dictionary that are representative of different types of receptive fields: (a) a harmonic stack; (b) an onset element; (c) a harmonic stack with flanking suppression; (d) a more localized onset/termination element; (e) a formant; (f) a tight checkerboard pattern (see Fig. S1 for the full dictionary). Each rectangle represents the spectro-temporal receptive field (STRF) of a single element in the dictionary; time is plotted along the horizontal axis (from 0 to 216 msec) and log frequency is plotted along the vertical axis, with frequencies ranging from 100 Hz to 4000 Hz. (g) A graph of the usage of the dictionary elements showing that the different types of receptive field shapes separate based on usage into a series of rises and plateaus; red symbols indicate where each of the examples from panels a–f fall on the graph. The vertical axis represents the number of stimuli that required a given dictionary element in order to be represented accurately during inference. doi:10.1371/journal.pcbi.1002594.g003 sometimes referred to as ‘‘temporal inhibition’’ or ‘‘band-passed inhibition’’ when observed in neural receptive fields; we will refer to this as ‘‘suppression’’ rather than inhibition to indicate that the model is agnostic as to whether these suppressed regions reflect direct synaptic inhibition to the neuron, rather than a decrease in excitatory synaptic input. The next flat region represents stimulus onsets, or ON-type cells, that tend to be more localized in frequency (Fig. 3d). The fifth group of elements is reminiscent of formants (Fig. 3e), which are resonances of the vocal tract that appear as characteristic frequency modulations common in speech. Formants are modulations ‘‘on top of’’ the underlying harmonic stack, often consisting of pairs of subfields that diverge or converge over time in a manner that is not consistent with a pair of harmonics rising or falling together due to fluctuations in the fundamental frequency of the speaker’s voice. The final region consists of the most active neurons, which are highly localized in time and frequency and exhibit tight checkerboard-like patterns of excitatory and suppressive subfields (Fig. 3f). These features are exciting because they are similar to experimentally measured receptive field shapes that to our knowledge have not previously been theoretically predicted, as discussed below. Overcompleteness Affects Learned Features Analogous to sparse coding studies in vision [11,26], we find that the degree of overcompleteness strongly influences the range and complexity of model STRF shapes. Fig. 4 presents representative examples of essentially all distinct cell types found in a four-times overcomplete L0-sparse dictionary trained on spectrograms. Features in the half-complete dictionary do appear as subsets of the larger dictionaries (Fig. 4a, c, e, g, l), but with increasing overcompleteness more complex features emerge, exhibiting richer patterns of excitatory and suppressive subfields. In general, optimized overcomplete representations can better capture structured data with fewer active elements, since the greater number of elements allows for important stimulus features to be explicitly represented by dedicated elements. In the limit of an infinite dictionary, for example, each element could be used as a so-called ‘‘grandmother cell’’ that perfectly represents a single, specific stimulus while all other elements are inactive. Novel features that were not observed in smaller dictionaries include: an excitatory harmonic stack flanked by a suppressive harmonic stack (Fig. 4b); a neuron excited by low frequencies (Fig. 4d); a neuron sensitive to two middle frequencies (Fig. 4f); a localized but complex excitatory subregion followed by a suppressive subregion that is strongest for high frequencies (Fig. 4h); a checkerboard pattern with roughly eight distinct subregions (Fig. 4i); a highly temporally localized OFF-type neuron (Fig. 4j); and a broadband checkerboard pattern that extends for many cycles in time (Fig. 4k). Several of these features resemble STRFs reported in IC and further up the auditory pathway (see the ‘‘Predicting acoustic features that drive neurons in IC and later stages in the ascending auditory pathway’’ section below). One interesting property of the checkerboard units is that they are largely separable in space and time [27], which has been studied for these and other types of neurons in ferret IC [28]. This is in contrast to some of the other model STRF shapes we have found, such as the example shown in Fig. 4e, which contains a strong diagonal subfield that is not well described by a product of independent functions of time and frequency. As in the case of the half-complete dictionary (Fig. 3), the different classes of receptive field shapes segregate as a function of usage even as more intermediary shapes appear (see Fig. S4 for the entire four-times overcomplete dictionary). However, the plateaus and rises evident in the usage plot for the half-complete dictionary (Fig. 3g) are far less distinct for the overcomplete representation (Fig. 4m). PLoS Computational Biology www.ploscompbiol.org 5 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 4. A four-times overcomplete, L0-sparse dictionary trained on speech spectrograms. This dictionary shows a greater diversity of shapes than the undercomplete dictionaries. (a–l) Representative elements a, c, e, g, j, and l resemble those of the half-complete dictionary (see Fig. 3). Other neurons display more complex shapes than those found in less overcomplete dictionaries: (b) a harmonic stack with flanking suppressive subregions; (d) a neuron sensitive to lower frequencies; (f) a short harmonic stack; (h) a localized but complex pattern of excitation with flanking suppression; (i) a localized checkerboard with larger excitatory and suppressive subregions than those in panel l; (k) a checkerboard pattern that extends for many cycles in time. Several of these patterns resemble neural spectro-temporal receptive fields (STRFs) reported in various stages of the auditory pathway that have not been predicted by previous theoretical models (see text and Figs. 6–8). (m) A graph of usage of the dictionary elements during inference. The different classes of dictionary elements still separate according to usage (see Fig. S4 for the full dictionary) although the notable rises and plateaus as seen in Fig. 3g are less apparent in this larger dictionary. doi:10.1371/journal.pcbi.1002594.g004 These same trends are present in the cochleogram-trained dictionaries. More types of STRFs appear when the degree of increased (Figs. S11, S12, S13). For overcompleteness is example, with more overcomplete dictionaries, some neurons have subfields spanning all frequencies or the full time-window within the cochleogram inputs. Additionally, we find neurons that exhibit both excitation and suppression in complex patterns, though the detailed shapes differ from what we find for the dictionaries trained on spectrograms. We wondered to what extent the specific form of sparseness we imposed on the representation was affecting the particular features learned by our network. To study this, we used the LCA algorithm [24] to find the soft sparse solution (i.e., one that minimizes the L1 norm), and obtained similar results to what we found for the hard sparse cases: increasing overcompleteness resulted in greater diversity and complexity of learned features (see Figs. S5, S6, S7, S8). We also trained some networks using a different algorithm, called Sparsenet [10], for producing soft sparse dictionaries, and we again obtained similar results as for our hard sparse dictionaries (Figs. S9, S10). It has been proven mathe- matically [29] that signals that are actually L0-sparse can be uncovered effectively by L1-sparse coding algorithms, which suggests that speech is an L0-sparse signal given that we find similar features using algorithms designed to achieve either L1 or L0 sparseness. Thus, preprocessing with spectrograms rather than a more nuanced cochlear model, and the degree of over- completeness, greatly influenced the learned dictionaries, unlike the different sparseness penalties we employed. The specific form of the sparseness penalty did, however, affect the performance of the various dictionaries. In particular, the level of sparseness achieved across the population of model neurons exhibited different relationships with the fidelity of their repre- sentations, suggesting that some model choices resulted in population codes that were more efficient at using small numbers of neurons to represent stimuli efficiently, while others were more effective at increasing their representational power when incorpo- rating more active neurons (Fig. S20). Modulation Power Spectra Our four-times overcomplete, spectrogram-trained dictionary exhibits a clear tradeoff in spectrotemporal resolution (red points, Fig. 5), similar to what has been found experimentally in IC [30]. IC is the lowest stage in the ascending auditory pathway to exhibit such a tradeoff, but it has yet to be determined for higher stages of processing, such as A1. This trend is not present in the half- complete cochleogram-trained dictionary (blue open circles, Fig. 5). Rather, these elements display a limited range of temporal modulations, but they span nearly the full range of possible PLoS Computational Biology www.ploscompbiol.org 6 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 5. Our overcomplete, spectrogram-trained model exhibits similar spectrotemporal tradeoff as Inferior Coliculus. Modulation spectra of half-complete cochleogram-trained dictionary and four-times overcomplete spectroram-trained dictionary are shown. The four-times overcomplete spectrogram-trained dictionary elements (red dots; same dictionary as in Fig. 4) display a clear tradeoff between spectral and temporal modulations, similar to what has been reported for Inferior Colliculus (IC) [30]. By contrast, the half-complete cochleogram-trained dictionary (blue circles; same dictionary as in Fig. 2) exhibits a much more limited range of temporal modulations, with no such tradeoff in spectrotemporal resolution. Each data point represents the centroid of the modulation spectrum of the corresponding element. The elements shown in Fig. 4 are indicated on the graph with the same symbols as before. doi:10.1371/journal.pcbi.1002594.g005 spectral modulations. Thus, by this measure the spectrogram- trained dictionary is a better model of IC than the cochleogram- trained model. In the next section, we compare the shapes of the various classes of model STRFs with individual neuronal STRFs from IC, and again find good agreement between our over- complete spectrogram-trained model and the neural data. Predicting Acoustic Features that Drive Neurons in IC and Later Stages in the Ascending Auditory Pathway Our model learns features that resemble STRFs reported in IC [30–33], as well as in the ventral side of the medial geninculate body (MGBv) [34] and A1 [34–36]. We are unaware of any previous theoretical work that has provided accurate predictions for receptive fields in these areas. Figs. 6, 7, and 8 present several examples of previously reported experimental receptive fields that qualitatively match some of our model’s dictionary elements. We believe the most important class of STRFs we have found are localized checker- board patterns of excitation and suppression, which qualitatively match receptive fields of neurons in IC and MGBv (Fig. 7). IC neurons often exhibit highly localized excitation and suppression patterns (Fig. 6), sometimes referred to as ‘‘ON’’ or ‘‘OFF’’ responses, depending on the temporal order of excitation and suppression. We show multiple examples drawn from the complete, two-times overcomplete, and four-times overcomplete dictionaries, trained on spectrograms, that exhibit these patterns. The receptive fields of two neurons recorded in gerbil IC exhibit suppression at a particular frequency followed by excitation at the same frequency (Fig. 6a). Such neurons are found in our model dictionaries (Fig. 6b). The reverse pattern is also found in which suppression follows excitation as shown in two cat IC STRFs (Fig. 6c) with matching examples from our model dictionaries (Fig. 6d). Note that the experimental receptive fields extend to higher frequencies because the studies were done in cats and gerbils, which are sensitive to higher frequencies than we were probing with our human speech training set. The difference in time-scales between our spectrogram representation and the experimental STRFs could reflect timescales of the different speech and behaviorally relevant sounds for cats and rodents. A common feature of thalamic and midbrain neural receptive fields is a localized checkerboard pattern of excitation and suppression (Fig. 7), typically containing between four to nine distinct subfields. We present experimental gerbil IC, cat IC and cat MGBv STRFs of this type in Fig. 7a beside similar examples (Fig. 7b). This pattern is displayed by many from our model elements in our sparse coding dictionaries, but to our knowledge it has not been predicted by previous theories. We also find some less localized receptive fields that strongly resemble experimental data. Some model neurons (Fig. 8b) consist of a suppression/excitation pattern that extends across most frequencies, reminiscent of broadband OFF and ON responses as reported in cat IC and rat A1 (Fig. 8a). Another shape seen in experimental STRFs of bat IC (top), and cat A1 (bottom; Fig. 8c) is a diagonal pattern of excitation flanked by suppression at the higher frequencies. This pattern of excitation flanked by suppression is present in our dictionaries (Fig. 8d), including at the highest frequencies probed. This type of STRF pattern is reminiscent of the two-dimentional Gabor-like patches seen in V1, which have been well captured by sparse coding models of natural scenes [10,11,26]. PLoS Computational Biology www.ploscompbiol.org 7 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 6. Model comparisons to receptive fields from auditory midbrain. Complete and overcomplete sparse coding models trained on spectrograms of speech predict Inferior Colliculus (IC) spectro-temporal receptive field (STRF) shapes with excitatory and suppressive subfields that are localized in frequency but separated in time. (a) Two examples of Gerbil IC neural STRFs [31] exhibiting ON-type response patterns with excitation following suppression; data courtesy of N.A. Lesica. (b) Representative model dictionary elements from each of three dictionaries that match this pattern of excitation and suppression. The three dictionaries were all trained on spectrogram representations of speech, using a hard sparseness (L0) penalty; the representations were complete (left column; Fig. S2), two-times overcomplete (middle column; Fig. S3), and four-times overcomplete (right column; Fig. 4 and Fig. S4). (c) Two example neuronal STRFs from cat IC [30] exhibiting OFF-type patterns with excitation preceding suppression; data courtesy of M.A. Escabı´. (d) Other model neurons from the same set of three dictionaries as in panel b also exhibit this OFF-type pattern. doi:10.1371/journal.pcbi.1002594.g006 PLoS Computational Biology www.ploscompbiol.org 8 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 7. Model comparisons to receptive fields from auditory midbrain and thalamus. An overcomplete sparse coding model trained on spectrograms of speech predicts Inferior Colliculus (IC) and auditory thalamus (ventral division of the medial geniculate body; MGBv) spectro- temporal receptive fields (STRFs) consisting of localized checkerboard patterns containing roughly four to nine distinct subfields. (a) Example STRFs of localized checkerboard patterns from two Gerbil IC neurons [31], one cat IC neuron [33], and one cat MGBv neuron [34] (top to bottom). Data courtesy of N.A. Lesica (top two cells) and M.A. Escabı´ (bottom two cells). (b) Elements from the four-times overcomplete, L0-sparse, spectrogram- trained dictionary with similar checkerboard patterns as the neurons in panel a. doi:10.1371/journal.pcbi.1002594.g007 Discussion We have applied the principle of sparse coding to spectrogram and cochleogram representations of human speech recordings in order to uncover some important features of natural sounds. Of the various models we considered, we have found that the specific form of preprocessing (i.e., cochleograms vs. spectrograms) and the degree of overcompleteness are the most significant factors in determining the complexity and diversity of receptive field shapes. Importantly, we have also found that features learned by our sparse coding model resemble a diverse set of receptive field shapes in IC, as well as MGBv and A1. Even though a spectrogram may PLoS Computational Biology www.ploscompbiol.org 9 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields Figure 8. Model comparisons to receptive fields from auditory midbrain and cortex. on spectrograms of speech predicts several classes of broadband spectro-temporal receptive field (STRF) shapes found in Inferior Colliculus (IC) and primary auditory cortex (A1). (a,b) An example broadband OFF-type STRF from cat IC [34] (top; data courtesy of M.A. Escabı´) and an example broadband ON-type subthreshold STRF from rat A1 [35] (bottom; data courtesy of M. Wehr) shown in panel a resemble example elements from a four-times overcomplete, L0-sparse, spectrogram-trained PLoS Computational Biology www.ploscompbiol.org 10 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields dictionary shown in panel b. (c) STRFs from a bat IC neuron [32] (top; data courtesy of S. Andoni) and a cat A1 neuron [34] (bottom; data courtesy of M.A. Escabı´) each consist of a primary excitatory subfield that is modulated in frequency over time, flanked by similarly angled suppressive subfields. (d) Example STRFs from four elements taken from the same dictionary as in panels b exhibit similar patterns as the neuronal STRFs in panel c. doi:10.1371/journal.pcbi.1002594.g008 not provide as accurate a representation of the output of the cochlea as a more explicit cochleogram model, such as the one we explored here, we have found that sparse coding of spectrograms yields closer agreement to experimentally measured receptive fields, demonstrating that we can infer important aspects of sensory processing in the brain by identifying the statistically important features of natural sounds without having to impose many constraints from biology into our models from the outset. Indeed, it is worth emphasizing that the agreement we have found did not result from fitting the neural physiology, per se; it emerged naturally from the statistics of the speech data we used to train our model. Specifically, the model parameters we explored — undercomplete vs. overcomplete representation, L0 vs. L1 spar- seness penalty, and cochleogram vs. spectrogram preprocessing — represent a low-dimensional space of essentially eight different choices compared with the rich, high-dimensional space of po- tential STRF shapes we could have obtained. Intriguingly, while we have emphasized the agreement between our model and IC, the receptive fields we have found resemble experimental data from multiple levels of the mammalian ascending auditory pathway. This may reflect the possibility that the auditory pathway is not strictly hierarchical, so that neurons in different anatomical locations may perform similar roles, and thus are represented by neurons from the same sparse coding dictionary. This view is consistent with the well-known observation that there is a great deal of feedback from higher to lower stages of processing in the sub-cortical auditory pathway [37], as compared with the visual pathway, for example. Some of our shapes have even been reported at lower levels. Harmonic stacks, including some with band-passed inhibition, have been reported in the dorsal cochlear nucleus [25,38] and they have been observed in presynaptic responses in IC (M.A. Escabı´, C. Chen, and H. Read, Society for Neuroscience Abstracts 2011), but these shapes have not yet been reported in IC spiking responses or further up the ascending auditory pathway. The tradeoff in spectrotemporal resolution we have found in our model resembles that of IC, which is the lowest stage of the ascending auditory pathway to exhibit a tradeoff that cannot be accounted for by the uncertainty principle, as is the case for auditory nerve fibers [30], but it remains to be seen if such a tradeoff also exists in MGBv or A1. A related issue is that an individual neuron might play different roles depending on the stimulus ensemble being presented to the nervous system. In fact, changing the contrast level of the acoustic stimuli used to probe individual IC neurons can affect the number of prominent subfields in the measured STRF of the neuron [31]. Our model does not specify which neuron should represent any given feature, it just predicts the STRFs that should be represented in the neural population in order to achieve a sparse encoding of the stimulus. Moreover, for even moderate levels of overcompleteness, our sparse coding dictionaries include categories of features that have not been reported in the experimental literature. For example, the STRF shown in Fig. 4k represents a well-defined class of elements in our sparse dictionaries, but we are unaware of reports of this type of STRF in the auditory pathway. Thus, our theoretical receptive fields could be used to develop acoustic stimuli that might drive auditory neurons that do not respond to traditional probe stimuli. In particular, our dictionaries contain many broadband STRFs with complex structures. These broadband neurons may not have been found experimentally since by necessity researchers often probe neurons extensively with stimuli that are concentrated around the neuron’s best frequency. It is important to recognize that STRFs do not fully capture the response properties of neurons in IC, just as most of the explainable variance is not captured by linear receptive fields of V1 simple cells [39]. We note, however, that while our sparse coding framework involves a linear generative model, the encoding is non-linear. Thus, one of the questions addressed by this study is the degree to which the competitive nonlinearity of a highly over-complete model can account for the rich assortment of STRFs in IC. We have found that this is a crucial factor in learning a sparse representation that captures the rich variety of STRF shapes observed in IC, as well as in thalamus and cortex. We have presented several classes of STRFs from our model that qualitatively match the shapes of neural receptive fields, but in many cases the neurons are sensitive to higher frequencies than the model neurons. This is likely due to the fact that we trained our network on human speech, which has its greatest power in the low kHz range, whereas the example neural data available in the literature come from animals with hearing that extends to much higher acoustic frequencies, and with much higher-pitched vocalizations, than humans. Our primary motivation for using speech came from the success of previous studies that yielded good qualitative [15] and quantitative [16] predictions of auditory nerve (AN) response properties based on sparse coding of speech. In fact, in order to obtain comparable results using environmental sounds and animal vocalizations, the relative proportion of training examples from each of three classes of natural sounds had to be adjusted to empirically match the results found using speech alone. Thus, speech provides a parameter-free stimulus set for matching AN properties, just as we have found for our model of IC. Moreover, good agreement between the model and AN physiology required selecting high SNR epochs within typically noisy recordings from the field; good results also required the selection of epochs containing isolated animal vocalizations rather than simultaneous calls from many individuals. By contrast, the speech databases used in those studies and the present study consist of clean, high SNR recordings of individual speakers. The issue of SNR is especially important for our study given that the dimensionality of our training examples is much higher (6,400 values for our spectrogram patches; 200 values after PCA) compared with typical vision studies (e.g., 64 pixel values [10]). Beyond the practical benefits of training on speech, the basic question of whether IC is best thought of as specialized for conspecific vocalizations or suited for more general auditory processing remains unanswered, but it seems reasonable to assume that it plays both roles. Questions such as this have inspired an important debate about the use of artificial and ecologically relevant stimuli [40,41] and what naturalistic stimuli can tell us about sensory coding [42–44]. The fact that several of the different STRFs we find have been observed in a variety of species, including rats, cats, and ferrets, suggests that there exist sufficiently universal features shared by the specific acoustic environments of these creatures to allow some understanding of IC function without having to narrowly tailor the stimulus set to each species. Even if sparse coding is, indeed, a central organizing principle throughout the nervous system, it could still be that the sparse PLoS Computational Biology www.ploscompbiol.org 11 July 2012 Volume 8 Issue 7 e1002594 representations we predict with our model correspond best to the subthreshold, postsynaptic responses of the membrane potentials of neurons, rather than their spiking outputs. In fact, we show an example of a subthreshold STRF (Fig. 8a bottom) that agrees well with one class of broadband model STRFs (Fig. 8b). The tuning properties of postsynaptic responses are typically broader than spiking responses, as one would expect, which could offer a clue as to which is more naturally associated with model dictionary elements. If our model elements are to be interpreted as sub- threshold responses, then the profoundly unresponsive regions surrounding the active subfields of the neuronal STRFs could be more accurately fit by our model STRFs after they are post- processed by being passed through a model of a spiking neuron with a finite spike threshold. It is encouraging that sparse encoding of speech can identify acoustic features that resemble neuronal STRFs from auditory midbrain, as well as those in thalamus and cortex, and it is notable that the majority of these features bear little resemblance to the Gabor-like shapes and elongated edge detectors that have been predicted by sparse coding representations of natural images. Clearly, our results are not an unavoidable consequence of the sparse coding procedure itself, but instead reflect the structure of the speech spectrograms and cochleograms we have used to train our model. Previous work to categorize receptive fields in A1 has often focused on oriented features that are localized in time and frequency [27,45], and some authors have suggested that such Gabor-like features are the primary cell types in A1 [46], but the emerging picture of the panoply of STRF shapes in IC, MGBv, and A1 is much more complex, with several distinct classes of features, just as we have found with our model. An important next step will be to develop parameterized functional forms for the various classes of STRFs we have found, which can assume the role that Gabor wavelets have played in visual studies. We hope that this approach will continue to yield insights into sensory processing in the ascending auditory pathway. Methods Sparse Coding In sparse coding, the input (spectrograms or cochleograms) y is encoded as a matrix A multiplied by a vector of weighting coefficients s: y~Asze where e is the error. Each column of A represents one dictionary element or receptive field, the stimulus that most strongly drives the neuron. If there are more columns in A than elements in y, this will be an overcomplete representation. We defined the degree of overcompleteness relative to the number of principle components. We learned the dictionary and inferred the coefficients by descending an energy function that minimizes the mean squared error of reconstruction under a sparsity constraint. E (t)~ 1 2 DDy(t){As(t)DD2zl X m C (sm (t)): ð1Þ Here l controls the relative weighting of the two terms and C represents the sparsity constraint. The sparsity constraint requires the column vector s to be sparse by some definition. In this paper, we focus on the L0-norm, minimizing the number of non-zero coefficients in s (or equi- valently the number of active neurons in a network). Another norm we have investigated is the L1-norm, minimizing the absolute activity of all of the neurons. Sparse Coding Predicts IC Receptive Fields Locally Competitive Algorithm We performed inference of the coefficients with a recently developed algorithm, a Locally Competitive Algorithm [24], which minimizes close approximations of either the L0- or L1- norms. Each basis function Ai is correlated with a computing unit defined by an internal variable ui as well as the output coefficient si . All of the neurons begin with the coefficients set to zero. These values change over time depending on the input. A neuron ui increases by an amount bi if the input overlaps with the receptive field of the neuron: bi (t)~SAi ,y(t)T. The neurons evolve as a group following dynamics in which the neurons compete with one another to represent the input. The neurons inhibit each other with the strength of the inhibition increasing as the overlap of their receptive fields and the output coefficient values increase. This internal variable is then put through a thresholding function Tl to produce the output value: si~Tl (ui ). In vector notation, the full dynamic equation of inference is: 1 t _uu(t)~f (u(t))~ s(t)~Tl (u(t)): ½b(t){u(t){(AT A{I )s(t)Š, ð2Þ The variable t sets the time-scale of the dynamics. The thresholding function Tl is determined by the sparsity constraint C . It is specified via: l dC (sm ) dsm ~um{sm~um{Tl (um ): ð3Þ Learning Learning is done via gradient descent on the energy function: r(t)~y(t){As(t), A~AzgA (r(t)sT (t))zh(A{AAT A): ð4Þ The h term is a device for increasing orthogonality between basis functions [47]. This is equivalent to adding in a prior that the basis functions are unique. Stimuli We used two corpora of speech recordings from the handbook of the International Phonetic Association (http://web.uvic.ca/ ling/resources/ipa/handbook_downloads.htm) and TIMIT [48]. These consist of people telling narratives in approximately 30 different languages. We resampled all waveforms to 16000 Hz, and then converted them into spectrograms by taking the squared Fourier Transform of the raw waveforms. We sampled at 256 frequencies logarithmically spaced between 100 and 4000 Hz. We monotonically transformed the output with the logarithm function, resulting in the log-power of the sound at specified frequencies over time. then divided into segments covering all The data was frequencies and 25 overlapping time points (16 ms each) representing 216 ms total. Subsequently, we performed principal components analysis on the samples to whiten the data as well as reduce the dimensionality. We retained the first 200 principal components as this captured over 93% of the variance in the spectrograms and lowered the simulation time. During analysis, the dictionaries were dewhitened back into spectrogram space. PLoS Computational Biology www.ploscompbiol.org 12 July 2012 Volume 8 Issue 7 e1002594 We also trained with another type of input, cochleograms [21,22]. These are similar to spectrograms, but the frequency filters mimic known properties of the cochlea via a cochlear model [21]. The cochlear model sampled at 86 frequencies between 73 and 7630 Hz. For this input, the total time for each sample was 250 ms (still 25 time points), and the first 200 principle com- ponents captured over 98% of the variance. Presentation of Dictionaries All dictionary neurons were scaled to be between 21 and 1 when displayed. The coefficients in the encoding can take on positive or negative values during encoding. To reflect this, we looked at the skewness of each dictionary element. If the skewness was negative, the colors of the dictionary element were inverted when being displayed to reflect the way that element was actually being used. Modulation Power Spectra To calculate the modulation power spectra, we took a 2D Fourier Transform of each basis function. For each element, we plotted the peak of the temporal and spectral modulation transfer functions (Fig. 5). For the cochleogram-trained basis functions, we approximated the cochleogram frequency spacing as being log-spaced to allow comparison with the spectrogram-trained dictionaries. Presentation of Experimental Data Data from [31] was given to us in raw STRF format. Each was interpolated by a factor of three, but no noise was removed. Data from [30,32–35] were given to us in the same format as they were originally published. Supporting Information Figure S1 The full set of elements from a half-complete, L0-sparse dictionary trained with LCA [24] on spectro- grams of speech. Each rectangle represents the spectrotemporal receptive field of a single element in the dictionary; time is plotted along the horizontal axis (from 0 to 216 msec), and log frequency is plotted along the vertical axis, with frequencies ranging from 100 Hz to 4000 Hz. Color indicates the amount of power present at each frequency at each moment in time, with warm colors representing high power and cool colors representing low power. Each element has been normalized to have unit Euclidean length. Elements are arranged in order of their usage during inference with usage increasing from left to right along each row, and all elements of lower rows used more than those of higher rows. (TIFF) Figure S2 The full set of elements from a complete, L0-sparse dictionary trained with LCA [24] on spectro- grams of speech. Same conventions as Fig. S1. (TIF) Figure S3 The full set of elements from a two times overcomplete, L0-sparse dictionary trained with LCA [24] on spectrograms of speech. Same conventions as Fig. S1. (TIF) Figure S4 The full set of elements from a four times overcomplete, L0-sparse dictionary trained with LCA [24] on spectrograms of speech. Same conventions as Fig. S1. (TIF) Sparse Coding Predicts IC Receptive Fields Figure S5 The full set of elements from a half-complete, L1-sparse dictionary trained with LCA [24] on spectro- grams of speech. Same conventions as Fig. S1. (TIF) Figures S6 The full set of elements from a complete, L1- sparse dictionary trained with LCA [24] on spectro- grams of speech. Same conventions as Fig. S1. (TIF) Figure S7 The full set of elements from a two times overcomplete, L1-sparse dictionary trained with LCA [24] on spectrograms of speech. Same conventions as Fig. S1. (TIF) Figure S8 The full set of elements from a four times overcomplete, L1-sparse dictionary trained with LCA [24] on spectrograms of speech. Same conventions as Fig. S1. (TIF) Figure S9 The full set of elements from a half-complete, L1-sparse dictionary trained with Sparsenet [10] on spectrograms of speech. Same conventions as Fig. S1. (TIF) Figure S10 The full set of elements from a complete, L1- sparse dictionary trained with Sparsenet [10] on spec- trograms of speech. Same conventions as Fig. S1. (TIF) Figure S11 The full set of elements from a complete, L0-sparse dictionary trained using LCA [24] on cochleo- grams of speech. Each rectangle represents the spectro- temporal receptive field of a single element in the dictionary; time is plotted along the horizontal axis (from 0 to 250 ms), and log frequency is plotted along the vertical axis, with frequencies ranging from 73 Hz to 7630 Hz. Color indicates the amount of power present at each frequency at each moment in time, with warm colors representing high power and cool colors represent- ing low power. Each element has been normalized to have unit Euclidean length. Elements are arranged in order of their usage during inference with usage increasing from left to right along each row, and all elements of lower rows used more than those of higher rows. (TIF) Figure S12 The full set of elements from a two times overcomplete, L0-sparse dictionary trained with LCA [24] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S13 The full set of elements from a four times overcomplete, L0-sparse dictionary trained with LCA [24] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S14 The full set of elements from a half- complete, L1-sparse dictionary trained with LCA [24] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S15 The full set of elements from a complete, L1- sparse dictionary trained with LCA [24] on cochleo- grams of speech. Same conventions as Fig. S11. (TIF) PLoS Computational Biology www.ploscompbiol.org 13 July 2012 Volume 8 Issue 7 e1002594 Figure S16 The full set of elements from a two times overcomplete, L1-sparse dictionary trained with LCA [24] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S17 The full set of elements from a four times overcomplete, L1-sparse dictionary trained with LCA [24] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S18 The full set of elements from a half- complete, L1-sparse dictionary trained with Sparsenet [10] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S19 The full set of elements from a complete, L1-sparse dictionary trained with Sparsenet [10] on cochleograms of speech. Same conventions as Fig. S11. (TIF) Figure S20 The signal to noise ratio (SNR) of sparse coding dictionaries increases with overcompleteness and with increasing numbers of active elements. Blue lines with triangles represent L0-sparse dictionaries, whereas green lines represent L1-sparse dictionaries. As expected, representations are more accurate with increasing numbers of active neurons and also when the level of overcompleteness is increased. Interestingly, the L0-sparse dictionaries typically have higher SNRs than the L1- sparse dictionaries. A few other general trends are evident as well. References 1. Laughlin SB (2001) Energy as a constraint on the coding and processing of sensory information. Curr Opin Neurobiol 11: 475–480. 2. Attneave F (1954) Some informational aspects of visual perception. Psychol Rev 61: 183–193. 3. Barlow HB (1961) Possible principles underlying the transformations of sensory messages. In: Rosenblith W, editor. Sensory Communication. Cambridge: MIT Press. pp. 217–234. 4. Atick JJ, Redlich AN (1992) What does the retina know about natural scenes? Neural Comput 4: 196–210. 5. Laughlin SB (1981) A simple coding procedure enhances a neuron’s information capacity. Z Naturforsch 36c: 910–912. 6. Rieke F, Warland D, de Ruyter van Steveninck R, Bialek W (1999) Spikes: Exploring the neural code. MIT Press. 7. DeWeese MR (1996) Optimization principles for the neural code. Network 7: 325–331. 8. Zhao L, Zhaoping L (2011) Understanding auditory spectro-temporal receptive fields and their changes with input statistics by efficient coding principles. PLoS Comp Bio 7: e1002123. 9. Fo ldia´ k P (1990) Forming sparse representations by local anti-hebbian learning. Biol Cybern 64: 165–170. 10. Olshausen BA, Field DJ (1996) Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature 381: 607–609. 11. Rehn M, Sommer FT (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. 12. DeWeese MR, Hromdka T, Zador AM (2005) Reliability and representational bandwidth in the auditory cortex. Neuron 48: 479–488. 13. DeWeese MR, Wehr M, Zador AM (2003) Binary spiking in auditory cortex. J Neurosci 23: 7940–7949. 14. Hromdka T, DeWeese MR, Zador AM (2008) Sparse representation of sounds in the unanesthetized auditory cortex. PLoS Biol 6: e16. 15. Lewicki MS (2002) Efficient coding of natural sounds. Nat Neurosci 5: 356– 1111. 16. Smith EC, Lewicki MS (2006) Efficient auditory coding. Nature 439: 978–982. 17. Klein D, Ko nig P, Ko rding KP (2003) Sparse spectrotemporal coding of sounds. J Appl Signal Proc 7: 659–667. 18. Aertsen AMHJ, Johannesma PIM (1981) A comparison of the spectro-temporal sensitivity of auditory neurons to tonal ad natural stimuli. Biol Cybern 42: 142–156. 19. Ko rding KP, Ko nig P, Klein D (2002) Learning of sparse auditory receptive fields. In: Proceedings of the International Joint Conference on Neural Networks 2002; 12–17 May 2002; Honolulu, Hawaii, United States. IJCNN ’02. Sparse Coding Predicts IC Receptive Fields Most notably, the L0-sparse dictionaries have higher SNRs than the L1-sparse dictionaries for similar levels of sparseness. Also, the more overcomplete dictionaries have higher SNRs than half- complete ones, even with the same absolute number of active neurons. The half-complete and complete dictionaries do not show much improvement in performance even as the number of active neurons increases. Interestingly, we find that the performance of the L0-sparse dictionaries tend to saturate as the fraction of active neurons approaches unity whereas the corresponding curves for the L1-sparse dictionaries tend to curve upwards. Note that we did not optimize the dictionaries at each data point, but instead used the same parameters used when training the network. (TIF) Acknowledgments The authors would like to thank J. Wang and E. Bumbacher for providing computer code and for helpful discussions. We are grateful to the following authors for sharing their neural data: F.A. Rodrı´guez, H.L. Read, M.A. Escabı´, S. Andoni, N. Li, G.D. Pollak, N.A. Lesica, B. Grohe, A. Qiu, C.E. Schreiner, C. Machens, M. Wehr, H. Asari, and A.M. Zador. We thank M.A. Escabı´, H.L. Read, A.M. Zador, and all members of the Redwood Center for useful discussions, and we thank K. Ko rding for helpful discussions and comments on the manuscript. Author Contributions Conceived and designed the experiments: NLC MRD. Performed the experiments: NLC. Analyzed the data: NLC VLM MRD. Wrote the paper: NLC MRD. 20. Henaff M, Jarrett K, Kavukcuoglu K, LeCun Y (2011) Unsupervised learning of sparse features for scalable audio classification. In: Proceedings of the 12th International Symposium on Music Information Retrieval; 24–28 October 2011; Miami, Florida, United States. ISMIR 2011. 21. Lyon RF (1982) A computational model of filtering, detection, and compression in the cochlea. In: Proc. IEEE Int. Conf. Acoust., Speech Signal Processing; Paris, France. pp. 1282–1285. 22. Slaney M (1998) Auditory toolbox version 2. Interval Research Corporation. Technical Report 1998-010. 23. van Hateren JH, van der Schaaf A (1998) Independent component filters of natural images compared with simple cells in primary visual cortex. Proc R Soc Lond B 265: 359–366. 24. Rozell CJ, Johnson DH, Baraniuk RG, Olshausen BA (2008) Sparse coding via thresholding and local competition in neural circuits. Neural Comput 20: 2526– 2563. 25. Backoff PM, Clopton BM (1991) A spectrotemporal analysis of dcn of single unit responses to wideband noise in guinea pig. Hearing Res 53: 28–40. 26. Olshausen BA, Cadieu CF, Warland DK (2009) Learning real and complex overcomplete representations from the statistics of natural images. In: Goyal VK, Papadakis M, van de Ville D, editors. Proc. SPIE, volume 7446. San Diego, California. 27. Depireux DA, Simon JZ, Klein DJ, Shamma SA (2001) Spectro-temporal response field characterization with dynamic ripples in ferret primary auditory cortex. J Neurophysiol 85: 1220–1111. 28. Shechtere B, Marvit P, Depireux DA (2010) Lagged cells in the inferior colliculus of the awake ferret. Eur J Neurosci 31: 42–48. 29. Donoho DL (2004) Compressed sensing. IEEE Trans Inform Theory 52: 1289– 1396. 30. Rodrı´guez FA, Read HL, Escabı´ MA (2010) Spectral and temporal modulation tradeoff in the inferior colliculus. J Neurophysiol 103: 887–903. 31. Lesica NA, Grothe B (2008) Dynamic spectrotemporal feature selectivity in the auditory midbrain. J Neurosci 28: 5412–5421. 32. Andoni S, Li N, Pollak GD (2007) Spectrotemporal receptive fields in the inferior colliculus revealing selectivity for spectral motion in conspecific vocalizations. J Neurosci 27: 4882–4893. 33. Qiu A, Schreiner CE, Escabı´ MA (2003) Gabor analysis of auditory midbrain receptive fields: spectro-temporal and binaural composition. J Neurophysiol 90: 456–476. 34. Escabı´ MA, Read HL (2005) Neural mechanisms for spectral analysis in the auditory midbrain, thalamus, and cortex. Int Rev Neurobiol 70: 207–252. PLoS Computational Biology www.ploscompbiol.org 14 July 2012 Volume 8 Issue 7 e1002594 Sparse Coding Predicts IC Receptive Fields 35. Machens CK, Wehr MS, Zador AM (2004) Linearity of cortical receptive fields measured with natural sounds. J Neurosci 24: 1089–1100. 36. Fritz J, Shamma S, Elhilali M, Klein D (2003) Rapid task-related plasticity of spectrotemporal receptive ı´elds in primary auditory cortex. Nat Neurosci 6: 1216–1223. 37. Read HL, Winer JA, Schreiner CE (2002) Functional architecture of auditory cortex. Curr Opin Neurobiol 12: 433–440. 38. Clopton BM, Backoff PM (1991) Spectrotemporal receptive fields of neurons in cochlear nucleus of guinea pig. Hearing Res 52: 329–44. 39. David SV, Gallant JL (2005) Predicting neuronal responses during natural vision. Network 16: 239–60. 40. Rust NC, Movshon JA (2005) In praise of artifice. Nat Neurosci 8: 1647–50. 41. Felsen G, Dan Y (2005) A natural approach to studying vision. Nat Neurosci 8: 1643–6. 42. Hsu A, Woolley SM, Fremouw TE, Theunissen FE (2004) Modulation power and phase spectrum of natural sounds enhance neural encoding performed by single auditory neurons. J Neurosci 24: 9201–11. 43. Rieke F, Bodnar DA, Bialek W (1995) Naturalistic stimuli increase the rate and efficiency of information transmission by primary auditory afferents. Proc Biol Sci 262: 259–65. 44. Vinje WE, Gallant JL (2000) Sparse coding and decorrelation in primary visual cortex during natural vision. Science 287: 1273–6. 45. Shamma SA (2001) On the role of space and time in auditory processing. TRENDS Cogn Sci 5: 340–348. 46. deCharms RC, Blake DT, Merzenich MM (1998) Optimizing sound features for cortical neurons. Science 280: 1439–1111. 47. Lee H, Pham P, Largman Y, Ng A (2009) Unsupervised feature learning for audio classification using convolutional deep belief networks. In: Bengio Y, Schuurmans D, Lafferty J, Williams CKI, Culotta A, editors. Advances in Neural Information Processing Systems 22. pp. 1096–1104. 48. Garofolo JS, et al (1993) Timit acoustic-phonetic continuous speech corpus. Philadelphia: Linguistic Data Consortium. PLoS Computational Biology www.ploscompbiol.org 15 July 2012 Volume 8 Issue 7 e1002594
1711.10814
2
1711
2017-12-10T08:43:26
A new fMRI data analysis method using cross validation: Negative BOLD responses may be the deactivations of interneurons
[ "q-bio.NC", "stat.AP" ]
Although functional magnetic resonance imaging (fMRI) is widely used for the study of brain functions, the blood oxygenation level dependent (BOLD) effect is incompletely understood. Particularly, negative BOLD responses(NBRs) is controversial. This paper presents a new fMRI data analysis method, which is more accurate than the typical conventional method. The authors conducted the experiments of simple repetition, and analyzed the data by the new method. The results strongly suggest that the deactivations(NBRs) detected by the new method are the deactivations of interneurons, because the deactivation ratios obtained by the new method approximately equals the deactivation ratios of interneurons obtained by the study of interneurons. The (de)activations detected by the new method are largely different from those detected by the conventional method. The new method is more accurate than the conventional method, and therefore the (de)activations detected by the new method may be correct and the (de)activations detected by the conventional method may be incorrect. A large portion of the deactivations of inhibitory interneurons is also considered to be activations. Therefore, the right-tailed t-test, which is usually performed in the conventional method, does not detect the whole activation, because the right-tailed t-test only detects the activations of excitatory neurons, and neglect the deactivations of inhibitory interneurons. A lot of fMRI studies so far by the conventional method should be re-examined by the new method, and many results obtained so far will be modified.
q-bio.NC
q-bio
A new fMRI data analysis method using cross validation: Negative BOLD responses may be the deactivations of interneurons. Hiroshi Tsukimoto and Takefumi Matsubara Tokyo Denki University [email protected] December 10, 2017 Abstract Although functional magnetic resonance imaging (fMRI) is widely used for the study of brain functions, the blood oxygenation level dependent (BOLD) effect is incompletely understood. Particularly, negative BOLD responses(NBRs) is controversial. This paper presents a new fMRI data analysis method, which is more accurate than the typical conventional method. The authors conducted the experiments of simple repetition, and analyzed the data by the new method. The results strongly suggest that the deactivations(NBRs) detected by the new method are the deactivations of interneurons, because the deactivation ratios obtained by the new method approximately equals the deactivation ratios of interneurons obtained by the study of interneurons. The (de)activations detected by the new method are largely different from those detected by the conventional method. The new method is more accurate than the conventional method, and therefore the (de)activations detected by the new method may be correct and the (de)activations detected by the conventional method may be incorrect. A large portion of the deactivations of inhibitory interneurons is also considered to be activations. Therefore, the right-tailed t-test, which is usually performed in the conventional method, does not detect the whole activation, because the right-tailed t-test only detects the activations of excitatory neurons, and neglect the deactivations of inhibitory interneurons. A lot of fMRI studies so far by the conventional method should be re-examined by the new method, and many results obtained so far will be modified. 1. Introduction Functional magnetic resonance imaging (fMRI) is widely used for the study of human brain functions. fMRI is based on the blood oxygenation level dependent (BOLD) effect(1). The 1 BOLD response is divided into positive BOLD response (PBR) and negative BOLD response (NBR). In particular, NBR remains incompletely understood(2,3,4). The cause of the problem of NBR may be an inappropriate analysis method. In the typical conventional fMRI data analysis method, the regression analysis of each voxel is performed. T-tests are performed on the coefficients of regression formulas. Multiple comparison adjustments are performed in many cases(5). In the multiple comparison adjustment, the threshold for p value is modified. The regression analysis of several adjacent voxels is expected to work better than the regression analysis of one voxel, because voxels are connected with each other in reality. This paper presents a new fMRI data analysis method. In the conventional method, tasks/rests are independent variables and fMRI values are dependent variables. In the new method, contrary to the conventional method, tasks/rests are dependent variables and fMRI values are independent variables. By this exchange, multiple regression analysis with several voxels, which properly models the situation that voxels are connected with each other, is possible. The regression analysis with several voxels is expected to be more accurate than the conventional method and is also expected to solve the problem of NBRs. In the new method, cross validations are performed to estimate the accuracies of regression formulas with several variables. Regression formulas with small mean squared prediction errors (MSPEs) are selected in an appropriate method, and t-tests (two-tailed) are performed on the coefficients of the regression formulas. Positive coefficients that are statistically significant show PBRs (activations), and negative coefficients that are statistically significant show NBRs (deactivations). The authors conducted the experiments of simple repetition using fMRI. It is well known that the temporal lobe and the motor cortex are activated in simple repetition(6). The new method was applied to the temporal lobe and the motor cortex. The new method with 8 voxels is more accurate than the conventional method. The deactivation ratio (= the number of deactivated voxels the number of deactivated voxels + the number of activated voxels ) of the temporal lobe is approximately 30%-33%. Neurons in the brain are broadly divided into excitatory neurons (e.g. pyramidal neurons) which release glutamate, and interneurons which are primarily inhibitory and release GABA. It was reported that the ratio of the population of interneurons in the human temporal lobe is approximately 37.7%(7). We have to consider the energy 2 consumption of interneurons and the ratio of deactivated interneurons in activation areas, the both of which are not well known. However, under an appropriate assumption, the ratio of deactivations of interneurons is approximately 25.4%-33.8%. As the ratio of deactivations obtained from the experiments is 30%-33%, the deactivations obtained from the experiments are probably the deactivations of interneurons. As described above, the new method found that NBRs are the deactivations of interneurons, which could not be found by the conventional method, because the conventional method is inaccurate. Metaphorically speaking, the conventional method is a microscope with magnification of 10 times and the new method is a microscope with magnification of 20 times. The deactivations of interneurons, which cannot be detected by a microscope with 10 times (the conventional method), can be detected by a microscope with magnification of 20 times (the new method). The (de)activations detected by the conventional method is largely different from those detected by the new method. The conventional method is less accurate than the new method, and therefore the (de)activations detected by the conventional method may be incorrect and the (de)activations detected by the new method may be correct. A large portion of the deactivations of inhibitory interneurons is also considered to be activations. Therefore, the right-tailed t-test, which is usually performed in the conventional method, does not detect the whole activation, because the right-tailed t-test only detects the activations of excitatory neurons, and neglect the deactivations of inhibitory interneurons. A lot of fMRI studies thus far by the conventional method should be re-examined by the new method, and many results obtained thus far will be modified. 2. A new fMRI data analysis method 2.1 The features of the new fMRI data analysis method 1. fMRI values are independent variables and tasks/rests are dependent variables. In the conventional method, tasks/rests(convolved with hemodynamic response function:HRF) are independent variables and fMRI values are dependent variables. In the new method, fMRI values are independent variables and tasks/rests (convolved with HRF) are dependent variables. By this exchange, multiple regression analysis with several voxels is possible. 2. Multiple regression analysis with several voxels 3 First, we explain the case of one voxel. The regression formula is as follows: 𝑦 = 𝑎𝑥 + 𝑏, where y stands for task(1)/rest(0) (As convolved with HRF, the range of 𝑦 is a little wider than the 0-1 range), 𝑥 stands for a fMRI value, 𝑎 stands for the coefficient and 𝑏 stands for the bias. The above formula is a prediction model, where a fMRI value predicts task/rest. This model does not show the physical causal relationship.This model shows the correlation between fMRI values and tasks/rests. It can be expected that the regression analysis with several voxels is more accurate than the regression analysis with one voxel, because voxels are connected with each other and voxels do not work independently. The typical simple set of several voxels is a cube with a side of 2 voxels. The total number of the voxels is 8(=2×2×2). The regression formula with 8 voxels is as follows: 𝑦 = 𝑎1𝑥1 + ⋯ + 𝑎8𝑥8 + 𝑏, where 𝑦 stands for task(1)/rest(0) and 𝑥𝑖 s stand for fMRI values. The above regression analyses are performed for all the cubes. In order to simplify the explanation, we explain the case of two dimensions, where squares with a side of two voxels are considered. Fig.1 shows 16 voxels. The first square is（1,2,5,6），the second square is (2,3,6,7)，the third square is (3,4,7,8)，the fourth square is（5,6,9,10）, and so on. Two examples of the regression formulas are as follows: 𝑦 = 𝑎1𝑥1 + 𝑎2𝑥2 + 𝑎3𝑥5 + 𝑎4𝑥6 + 𝑏1, 𝑦 = 𝑎5𝑥2 + 𝑎6𝑥3 + 𝑎7𝑥6 + 𝑎8𝑥7 + 𝑏2. The first formula corresponds to the first square (1,2,5,6), and the second formula corresponds to the second square (2,3,6,7). 1 5 9 2 6 3 7 10 11 13 14 15 4 8 12 16 Figure 1 16 voxels in the case of two dimensions 2.2 The procedures of the new method The procedures of the new method is explained with the two-dimensional example of Fig.1, for simplification. Consider the regression analysis of four(=2×2) voxels. There are nine 4 squares.((1,2,5,6),(2,3,6,7),(3,4,7,8),(5,6,9,10),(6,7,10,11),(7,8,11,12),(9,10,13,14),(10,11,14, 15), (11,12,15,16)) 1. Perform the cross-validations. Let us assume that the regression formulas and the MSPEs are obtained as shown in Table 1. Table 1 Regression formulas No. Regression formulas MSPEs 1 2 3 4 5 6 7 8 9 𝑦 = −0.01𝑥1 + 0.61𝑥2 + 0.82𝑥5 + 0.27𝑥6 + 0.41 𝑦 = 0.76𝑥2 + 0.06𝑥3 + 0.91𝑥6 − 0.02𝑥7 + 0.52 𝑦 = −0.01𝑥3 + 0.06𝑥4 − 0.07𝑥7 − 0.03𝑥8 + 0.45 𝑦 = 0.94𝑥5 + 0.64𝑥6 − 0.08𝑥9 − 0.72𝑥10 + 0.48 𝑦 = 0.70𝑥6 − 0.06𝑥7 − 0.88𝑥10 − 0.02𝑥11 + 0.46 𝑦 = 0.01𝑥7 − 0.06𝑥8 + 0.08𝑥11 − 0.05𝑥12 + 0.50 𝑦 = 0.04𝑥9 − 0.86𝑥10 + 0.09𝑥13 − 0.03𝑥14 + 0.48 0.11 0.08 0.20 0.09 0.15 0.18 0.19 𝑦 = −0.95𝑥10 + 0.04𝑥11 + 0.08𝑥14 − 0.01𝑥15 + 0.49 0.16 𝑦 = 0.01𝑥11 + 0.06𝑥12 − 0.08𝑥15 − 0.02𝑥16 + 0.51 0.17 2. Sort the regression formulas in ascending order of their MSPEs. The regression formulas are sorted in ascending order of their MSPEs. See Table 2. Table 2 Sorted formulas No. Regression formulas MSPEs 0.08 0.09 2 4 1 5 3 6 7 8 9 𝑦 = 0.76𝑥2 + 0.06𝑥3 + 0.91𝑥6 − 0.02𝑥7 + 0.52 𝑦 = 0.94𝑥5 + 0.64𝑥6 − 0.08𝑥9 − 0.72𝑥10 + 0.48 𝑦 = −0.01𝑥1 + 0.61𝑥2 + 0.82𝑥5 + 0.27𝑥6 + 0.41 0.11 𝑦 = 0.70𝑥6 − 0.06𝑥7 − 0.88𝑥10 − 0.02𝑥11 + 0.46 𝑦 = −0.01𝑥3 + 0.06𝑥4 − 0.07𝑥7 − 0.03𝑥8 + 0.45 𝑦 = 0.01𝑥7 − 0.06𝑥8 + 0.08𝑥11 − 0.05𝑥12 + 0.50 𝑦 = 0.04𝑥9 − 0.86𝑥10 + 0.09𝑥13 − 0.03𝑥14 + 0.48 0.15 0.16 0.17 0.18 𝑦 = −0.95𝑥10 + 0.04𝑥11 + 0.08𝑥14 − 0.01𝑥15 + 0.49 0.19 𝑦 = 0.01𝑥11 + 0.06𝑥12 − 0.08𝑥15 − 0.02𝑥16 + 0.51 0.20 5 3. Perform t-tests on the coefficients of the regression formulas in ascending order of their MSPEs. T-tests are performed on the coefficients of the regression formulas in ascending order of their MSPEs. Positive coefficients that are statistically significant show PBRs(activations), and negative coefficients that are statistically significant show NBRs(deactivations). Table 3 T-tests No. Regression formulas Significant Significant MSPEs Voxels I Voxels II 𝑦 = 0.76𝑥2 + 0.06𝑥3 + 0.91𝑥6 − 0.02𝑥7 + 0.52 𝑦 = −0.94𝑥5 + 0.64𝑥6 − 0.08𝑥9 − 0.11𝑥10 + 0.48 𝑦 = −0.01𝑥1 + 0.61𝑥2 − 0.82𝑥5 + 0.13𝑥6 + 0.41 𝑦 = 0.70𝑥6 − 0.06𝑥7 − 0.88𝑥10 − 0.02𝑥11 + 0.46 𝑦 = −0.01𝑥3 + 0.06𝑥4 − 0.07𝑥7 − 0.03𝑥8 + 0.45 𝑦 = 0.01𝑥7 − 0.06𝑥8 + 0.08𝑥11 − 0.05𝑥12 + 0.50 𝑦 = 0.04𝑥9 − 0.86𝑥10 + 0.09𝑥13 − 0.03𝑥14 + 0.48 𝑦 = −0.95𝑥10 + 0.04𝑥11 + 0.07𝑥14 − 0.01𝑥15 + 0.49 𝑦 = 0.01𝑥11 + 0.06𝑥12 − 0.08𝑥15 − 0.91𝑥16 + 0.51 𝑥2, 𝑥6 𝑥5, 𝑥6 𝑥2, 𝑥5 𝑥6, 𝑥10 𝑥10 𝑥10 𝑥16 𝑥2, 𝑥6 𝑥5 𝑥10 𝑥16 0.08 0.09 0.11 0.16 0.17 0.18 0.19 0.20 0.25 2 4 1 5 3 6 7 8 9 Let us assume that significant voxels are obtained as shown in "Significant Voxels I" in Table3. For example, 𝑥6(No.6 voxel) is statistically significant in No.2 formula, No.4 formula and No.5 formula. In such a case, 𝑥6 in No.2 formula is selected, and 𝑥6 in No.4 formula and No.5 formula are neglected. The same applies to other voxels(variables). Generally, a voxel that turned out to be significant by t-test in a regression formula, is neglected afterwards. The result of this processing is shown in "Significant Voxels II". "Significant Voxels I" stands for the results that do not consider double check. "Significant Voxels II" stands for the results that consider double check. 4. Select reliable significant voxels Regression formulas whose MSPEs are large, are unreliable, and the significant voxels in regression formulas whose MSPEs are large, are also unreliable. For example, in No.9 6 regression formula, 𝑥16 is significant, which is unreliable, because the MSPE of No.9 regression formula is large (0.25). See Table 3. Significant voxels in the regression formulas whose MSPEs are large (= unreliable significant voxels) should be excluded, and only significant voxels in the regression formulas whose MSPEs are small (=reliable significant voxels) should be selected. There are a few method to select reliable significant voxels. One method is to select significant voxels in the regression formulas whose MSPEs are smaller than a threshold. For example, let the threshold be 0.15 in Table 3, then 𝑥2, 𝑥6 and 𝑥5 are selected. Another method is to select significant voxels up to a certain number. For example, let a certain number be 4 in Table 3, then 𝑥2, 𝑥6, 𝑥5 and 𝑥10 are selected. As far as we have experienced, the latter method worked better than the former method, and therefore we adopt the latter method, in this paper. The actual algorithm is as follows: ①Perform t-tests on the coefficients of the regression formulas in ascending order of their MSPEs . ②Do not perform t-tests on the coefficients of the voxels that turned out to be significant before that. ③ Perform t-tests until the number of significant voxels reaches a certain number. In the above argument, a two dimensional example was explained for simplification. In reality, the regression analysis with 8(=2×2×2) voxels are performed. The regression analysis with 27(=3×3×3) voxels are also performed later. 3. Results 3.1 The new method is accurate. As the ratio of the population of interneurons in the human temporal lobe was reported(7), we conducted the block design experiments of simple repetition that activates the temporal lobe including the auditory cortex(6). In the tasks, the participants overtly spoke the sentences that they heard through a headphone. The temporal lobe (Brodmann Area 20,21,22,37,38,41, and 42) was analyzed by the conventional method, and the new method(1 voxel, 8 voxels, and 27 voxels). The regression formulas were sorted in ascending order of their MSPEs. Table 4 shows MSPEs. Values in the table are shown with three significant digits. The same applies to other tables. 7 Table 4 MSPEs of four methods No. 200th MSPEs 300th MSPEs 1V(cnv) 1V 8V 27V 1V(cnv) 1V 8V 27V 0.564 0.669 0.790 1.78 0.0974 0.0621 0.0652 0.626 0.111 0.0707 0.0729 0.107 0.0810 0.0817 0.727 0.113 0.0835 0.0855 0.0935 0.0646 0.0690 0.888 0.108 0.0684 0.0732 0.0958 0.0672 0.0684 1.91 0.105 0.0724 0.0724 0.635 0.132 0.0871 0.0844 0.678 0.143 0.0926 0.0908 1.03 0.571 0.450 0.988 0.0723 0.0547 0.0569 1.11 0.0801 0.0578 0.0595 0.116 0.0807 0.0769 0.615 0.130 0.0858 0.0846 0.122 0.0785 0.0789 0.477 0.128 0.0884 0.0857 0.154 0.110 0.114 1.10 0.165 0.118 0.124 1 2 3 4 5 6 7 8 9 10 1.01 0.158 0.108 0.111 1.10 0.170 0.117 0.118 Avg. 0.849 0.115 0.0794 0.0806 0.923 0.125 0.0855 0.0867 The left half of the table shows the MSPEs of the 200th regression formulas, and the right half of the table shows the MSPEs of the 300th regression formulas. "No." stands for the participant number. "1V (cnv)" stands for the conventional method. "1V", "8V" and "27V" stand for the new method with 1 voxels, 8 voxels and 27 voxels, respectively. The new method(8V or 27V) is more accurate than the new method(1V) and the conventional method. The MSPEs of the conventional method are large compared with those of the new method(1V, 8V or 27V), because the dependent variables in the conventional method are fMRI values and the dependent variables in the new method are tasks/rests convolved with HRF. Therefore, the comparison between the MSPEs of the conventional method and the MSPEs of the new method makes no sense. We want to check if the new method(8V or 27V) is more accurate than the conventional method. However, as explained above, the comparison between the conventional method and the new method(8V or 27V) makes no sense. The regression ability of the conventional method is almost the same as that of the new method(1V), because the latter is obtained from the former by exchanging the independent variable and the dependent variable. More detailed explanation is as follows. The coefficient 8 of determination is used as a measure of the regression ability of a regression formula(8). There are several definitions of the coefficient of determination, one of which is equal to the squared correlation coefficient of the dependent variable and the independent variable. As the new method(1V) is obtained from the conventional method by exchanging the independent variable and the dependent variable, the squared correlation coefficient of the dependent variable and the independent variable (=a coefficient of determination) of the conventional method is equal to that of the new method(1V). Therefore, the regression ability of the conventional method is almost the same as that of the new method(1V). Let us check if the new method(8V or 27V) is more accurate than the new method(1V) instead of checking if the new method(8V or 27V) is more accurate than the conventional method. Table 4 shows that the new method(8V) is more accurate than the new method(1V), and the accuracy of the new method(8V) is almost the same as that of the new method(27V). The above assertion is confirmed quantitatively by the analyses of variance (ANOVA) of MSPEs. Table 5 shows ANOVA of MSPEs in Table4. Table 5 The ANOVAs of MSPEs in Table 4 200/300 A B VR (TL) VR(MC) FCV DF1 DF2 200 200 200 300 300 300 1V 8V 11.5 27.4 8V 27V 0.0225 0.185 1V 27V 10.6 1V 8V 13.5 35.0 63.0 8V 27V 0.0178 0.183 1V 27V 12.5 45.8 4.41 4.41 4.41 4.41 4.41 4.41 1 1 1 1 1 1 18 18 18 18 18 18 "200/300" stands for the 200th regression formula or the 300th regression formula. "A" and "B" stand for the two groups of ANOVA. VR, TL, MC, FCV, DF1 and DF2 stand for variance ratio (=F value), temporal lobe, motor cortex, F-critical value at 0.05, degree of freedom between groups, and degree of freedom within groups, respectively. The variance ratios between the new method(1V) and the new method(8V), and those between the new method(1V) and the new method(27V) are greater than F-critical value. The variance ratios between the new method(8V) and the new method(27V) are less than F-critical value. We also analyzed the motor cortex (Brodmann Area 4 and 6), and obtained the similar results 9 to the temporal lobe. Table 5 also shows ANOVA of MSPEs in the motor cortex. The variance ratios between the new method(1V) and the new method(8V), and those between the new method(1V) and the new method(27V) are greater than F-critical value. The variance ratios between the new method(8V) and the new method(27V) are less than F-critical value. Table 4 and Table 5 show the results of the 200th(300th) regression formulas. At the nth regression formulas(n is neither 200 nor 300), we obtained similar results.We can conclude that the new method(8V or 27V) is more accurate than the new method(1V). As the regression ability of the conventional method(1V (cnv)) is almost the same as that of the new method(1V), we can also conclude that the new method(8V or 27V) is more accurate than the conventional method(1V (cnv)). The accuracy of the new method (8V) is almost the same as that of the new method (27V). However, the averages of the MSPEs of the new method (8V) are a little less than the averages of the MSPEs of the new method (27V) (See Table 4). Therefore, in this paper, we use the new method(8V). We conjecture that the new method with 23 or 33 voxels may work well, and the new method with 𝑛3(𝑛 is greater than 3) voxels may not work well, which is included in the future work. 3.2 The deactivations of interneurons We performed two-tailed t-tests (p ≦ 0.001,FWHM=8mm) on the coefficients of the regression formulas in ascending order of their MSPEs until the number of significant voxels reaches a certain number, which is denoted by C hereinafter. We counted the number of positive significant voxels and the number of negative significant voxels. The deactivation ratio is calculated as follows: Deactivation ratio = the number of deactivated voxels + the number of activated voxels the number of deactivated voxels = the number of negative significant voxels + the number of positive significant voxels . the number of negative significant voxels Table 6 shows the deactivation ratios (DRs) and MSPEs when C=200(300). The averages of the deactivation ratios in Table 6 are approximately 0.31-0.34. Notice that "200(300)" in Table 6 means the 200(300)th significant voxel, and "200(300)" in Table 4 means the 200(300)th regression formula. Generally, the 200(300)th significant voxels are included in the nth regression formulas (n>200(300)). When C is less than 200 (for example, C=100), the statistical stability is small, that is, the 10 deactivation ratios are unstable. When C is greater than 300 (for example, C=400), MSPEs are large, which means that the voxels included in the regression formulas are not related to the experimental task. Therefore, C=200(300) is appropriate. Generally, the appropriate C depends on the experimental task, that is, if the activation areas contain a lot of voxels, the appropriate C is large, and if the activation areas contain a small number of voxels, the appropriate C is small. It is another possible method to perform t-tests on the regression formulas whose MSPEs are less than a certain threshold (e.g. 0.15 or 0.20). Table 6 The deactivation ratios with FWHM=8mm and p≦0.001 No. Temporal lobe Motor cortex C=200 C=300 C=200 C=300 DR MSPE DR MSPE DR MSPE DR MSPE 0.343 0.105 0.357 0.120 0.365 0.127 0.357 0.146 0.280 0.119 0.283 0.139 0.345 0.118 0.327 0.145 0.290 0.110 0.307 0.128 0.350 0.148 0.360 0.168 0.310 0.107 0.280 0.132 0.390 0.150 0.363 0.171 0.340 0.135 0.302 0.167 0.340 0.153 0.329 0.175 0.330 0.0949 0.313 0.108 0.240 0.141 0.250 0.158 0.330 0.130 0.293 0.159 0.300 0.172 0.330 0.195 0.280 0.141 0.287 0.158 0.385 0.171 0.377 0.184 0.385 0.181 0.353 0.208 0.265 0.188 0.320 0.231 1 2 3 4 5 6 7 8 9 10 0.365 0.179 0.353 0.215 0.355 0.202 0.393 0.239 Avg. 0.325 0.313 0.334 0.341 "No." stands for the participant number. Notice that "200(300)" in Table 6 means the 200(300)th significant voxel, and "200(300)" in Table 4 means the 200(300)th regression formula. The averages of deactivation ratios are approximately 0.31-0.34. We also analyzed the data with FWHM=6mm and p≦0.001. The average of deactivation ratios of the temporal lobe was 0.344, and that of the motor cortex was 0.344. Therefore the averages of deactivation ratios with FWHM=6mm and p≦0.001 are a little greater than those of Table 6(FWHM=8mm, p≦0.001). We also analyzed the data with FWHM=8mm and p≦ 11 0.0027, which corresponds to 3σ of the normal distribution. The average of deactivation ratios of the temporal lobe was 0.357, and that of the motor cortex was 0.366. Therefore the averages of deactivation ratios with FWHM=8mm and p≦0.0027 are a little greater than those of Table 6 (FWHM=8mm, p≦0.001). Hereinafter, let FWHM=8mm and p≦0.001, because they both are commonly used. Let us investigate the deactivation ratios when C=300 in Table 6 in more detail. Table 7 shows the deactivation ratios (DRs) of the left hemisphere and the right hemisphere. The averages of deactivation ratios of the left hemisphere are almost the same as those of the right hemisphere. In the temporal lobe of No.5, left number=186 and right number=115, then 186+115=301, which is not equal to 300. The reason is that two voxels turned out to be significant by t-tests in the last regression formula. Table 7 The DRs of the left(right) hemisphere at the 300th significant voxel No. Temporal lobe Motor cortex Left Right Left Right Number DR Number DR Number DR Number DR 1 2 3 4 5 6 7 8 9 78 0.308 123 171 133 186 187 162 161 131 0.285 0.316 0.278 0.317 0.342 0.253 0.286 0.313 10 86 0.360 222 177 129 167 115 113 138 139 169 214 0.374 165 0.309 0.282 83 0.349 0.295 0.281 0.278 0.265 0.341 0.288 0.385 0.350 142 164 134 123 165 109 137 190 0.366 0.366 0.313 0.236 0.339 0.385 0.358 0.395 135 217 158 136 167 177 135 191 163 110 0.415 0.318 0.354 0.360 0.341 0.260 0.319 0.372 0.288 0.391 Average 141.8 0.306 158.3 0.314 141.2 0.342 158.9 0.342 "No." stands for the participant number. "Number" stands for the number of (positive and negative) significant voxels. The averages of deactivation ratios of the left hemisphere are almost the same as those of the right hemisphere. 12 We conducted the experiment of simple repetition five times with participant No.7. The deactivation ratios of the temporal lobe were 0.293, 0.363, 0.347, 0.310, and 0.347. The average was 0.332. The deactivation ratios of the motor cortex were 0.330, 0.387, 0.263, 0.317, and 0.307. The average was 0.321. From Table 6 and Table 7, the averages of the deactivation ratios of the temporal lobe are approximately 30%-33%. The averages of the deactivation ratios of the motor cortex are approximately 33%-35%. It was reported that the ratio of the population of interneurons in the human temporal lobe is approximately 37.7%(7). In order to find out the relation between the deactivation ratios and the ratio of the population of interneurons, we have to consider the energy consumption of interneurons and the ratio of the deactivated interneurons in an activation area. Let N stand for the ratio of the population of interneurons. Let A stand for the ratio of the average energy consumption of interneurons to the average energy consumption of excitatory neurons, which is called relative average energy consumption in this paper. Notice that the relative average energy consumption of excitatory neurons is 1. A = the average energy consumption of interneurons the average energy consumption of excitatory neurons . Let B stand for the ratio of the population of deactivated interneurons to the population of interneurons in an activation area. For simplification, "in an activated area" is omitted hereinafter. B = the popluation of deactivated interneurons . the population of interneurons Let F stand for the ratio of the energy consumption of deactivated interneurons to the energy consumption of the neurons. That is, F is described as follows: F = the energy consumption of deactivated interneurons the energy consumption of neurons . As the neurons are divided into excitatory neurons and interneurons, F is as follows: F = . the energy consumption of interneurons + the energy consumption of excitatory neurons the energy consumption of deactivated interneurons Let the numerator and the denominator of the above formula be divided by the energy consumption of excitatory neurons, then 13 F = . the relative energy consumption of interneurons + the relative energy consumption of excitatory neurons the relative energy consumption of deactivated interneurons The ratio of the population of deactivated interneurons is B × N. Let M stand for the total population of neurons in an activated area, then the population of deactivated interneurons is B × N × M. The relative energy consumption of deactivated interneurons is A × B × N × M. The relative energy consumption of interneurons is A × N × M. The relative energy consumption of excitatory neurons is 1 × (１ − N) × M. Therefore, F is as follows: F = A × B × N × M A × N × M＋(１ − N) × M = A × B × N A × N＋１ − N . Although the contributions of interneurons are very complicated at the micro level, the macro-level discussion is possible, because one voxel contains more than thousands of interneurons. Let us investigate A. The energy consumption of interneurons and those of excitatory neurons in humans are not well known. It was reported that the ratio of the population of interneurons in rats is approximately 15%(9,10,11). It was reported that the energy consumption ratio of interneurons in rats is 18%(12). Therefore, in rats, A is approximately 1.2. In humans, A may be higher than 1.2(13,14) . Let us investigate B. Interneurons are divided into Parvalbumin expressing (PV) interneurons, Somatostatin expressing (SOM) interneurons, Vasoactive intestinal peptide expressing (VIP) interneurons, and so on. When excitatory neurons in a certain area are activated, VIP interneurons are activated (excited) and inhibit other interneurons, and the interneurons inhibited by VIP interneurons are deactivated, which disinhibit excitatory neurons(15,16). Other inhibitions (for example, PV interneurons strongly inhibit one another) are also reported(17,18,19). Inhibition and disinhibition are controversial. For further argument, some assumptions on A and B are needed. Let us assume that A of humans is the same as A of rats (A=1.2). Let us assume that VIP interneurons in an activation area are activated and the majority of the other interneurons in the activation area are deactivated. The ratio of the population of VIP interneurons to the population of interneurons is 9.7±1.0%(20). Therefore, let us assume that B is approximately 1－0.097－α(=0.903－α), where α stands for the ratio of activated interneurons other than VIP interneurons. Thus, A =1.2, B=0.903－α, and N=0.377. Let us assume that 0.1 ≦ α ≦ 0.3. Substitute the above 14 values into F = A×B×N A×N＋１−N , then 0.254 ≦ F ≦0.338. The averages of the deactivation ratios of the temporal lobe in Table 6 and 7 are approximately 30%-33%, and therefore the deactivations detected in the experiments are probably the deactivations of interneurons. Moreover, as A of humans is probably higher than A of rats(=1.2)(13,14), let us assume that A of humans is 1.5, for example. And let us assume that F is 0.313(the average of deactivation ratios of the temporal lobe when C=300 in Table 6), then 0.313 = 1.5 × B × 0.377 1.5 × 0.377＋１ − 0.377 . holds, and B=0.658(α=0.245). Deactivated voxels and activated voxels are shown in Fig.2, which is explained in the next section. 3.3 The new method and the conventional method We applied two conventional methods to the fMRI data of simple repetition. One is t-test (p≦ 0.001). Another is t-test+ multiple comparison adjustment by random field theory, which is called "corrected" in Statistical Parametric Mapping (SPM)(5,21). Table 8 shows the deactivation ratios. The results of "1V" are similar to those of "0.05", because the regression formula of "1V" is obtained from that of "0.05" by exchanging the independent variables and the dependent variables. The deactivation ratios of "0.001" and "0.05" are approximately 0 with a few exceptions. The exceptions are No.1 temporal lobe, No.3 motor cortex, No.8 motor cortex and No.9 motor cortex. Fig. 2 shows the activated voxels and the deactivated voxels of "0.05" and "8V" of the above four cases. In Fig.2, the averages from the brain surface to the 30th voxel are displayed using MRIcro(22). The activated voxels and the deactivated voxels of "0.05" are largely different from those of "8V". In "0.05", the activated voxels and the deactivated voxels basically are not adjacent. On the other hand, in "8V", the activated voxels and the deactivated voxels are adjacent. The (de)activated voxels in "8V" are less than those in "0.05", because the number of (de)activated voxels are limited to 300 in "8V". We can increase the number to 400, for example, but if we do so, the regression formulas with large MSPEs will be considered to be involved in (de)activations, which will degrade the quality of the results. We can set the limit to the number of voxels in "0.05". Even if we do so, the (de)activated voxels only will be less 15 than those of Fig. 2, and the difference in the (de)activated voxels between "8V" and "0.05" will remain the same in essence. "Corrected" in the above argument is at the voxel-level, but at the cluster-level(5,21), the argument will be the same in essence. Table 8 Deactivation ratios of four methods No. Temporal lobe Motor cortex 0.001 0.05 1V 8V 0.001 0.05 1V 8V 1 2 3 4 5 6 7 8 9 0.190 0.158 0.000 0.357 0.0393 0.0278 0.000 0.357 0.0526 0.0362 0.000 0.283 0.000 0.000 0.000 0.327 0.0703 0.0693 0.000 0.307 0.386 0.307 0.020 0.360 0.00270 0.000 0.000 0.280 0.000 0.000 0.000 0.363 0.0181 0.000 0.000 0.302 0.0320 0.000 0.000 0.329 0.0183 0.00802 0.000 0.313 0.00896 0.000 0.000 0.250 0.00866 0.000621 0.000 0.293 0.0158 0.000 0.000 0.330 0.0285 0.0108 0.000 0.287 0.144 0.108 0.0233 0.377 0.0266 0.00483 0.000 0.353 0.260 0.148 0.000 0.320 10 0.0174 0.00184 0.000 0.353 0.000 0.000 0.000 0.393 "No." stands for the participant number. "0.001" stands for t-test (p≦ 0.001)(two-tailed). "0.05" stands for "corrected" (p≦ 0.05)(one-tailed). "1V" stands for the new method(1 voxel) when C=300. "8V" stands for the new method(8 voxels) when C=300. The results of "1V" are similar to those of "0.05", because the regression formula of "1V" is obtained from that of "0.05" by exchanging the independent variable and the dependent variable. The deactivation ratios of "0.001" and "0.05" are approximately 0 with a few exceptions. The exceptions are No.1 temporal lobe, No.3 motor cortex, No.8 motor cortex and No.9 motor cortex. The results of the new method(8V) are more reliable than those of the conventional method, because the new method(8V) is more accurate than the conventional method. Moreover, the deactivation ratios of the new method(8V) are stable, while the deactivation ratios of the conventional method are unstable, and the deactivations detected by the new method(8V) can be reasonably interpreted as the deactivations of interneurons, while the deactivations detected by the conventional method can be hardly interpreted. Many results on the deactivations 16 (NBRs)(23,24,25) were obtained by the conventional method. The results by the new method(8V) will be probably different from those by the conventional method. No. Method Area Left Left Right Right 1 1 3 3 8 Temporal 0.05 Lobe Temporal 8V Lobe Motor 0.05 cortex Motor 8V cortex Motor 0.05 cortex 8 Motor 8V cortex 9 Motor 0.05 cortex 9 Motor 8V cortex Activated voxels Deactivated voxels Activated voxels Deactivated voxels Figure 2 The activated voxels and the deactivated voxels of the conventional method 17 and the new method "No." stands for the participant number. "0.05" stands for "corrected" (p≦ 0.05). "8V" stands for the new method(8 voxels) when C=300. The (de)activated voxels of "0.05" are largely different from those of "8V". For example, in the left deactivated voxels of participant 1 and participant 9, and in the right deactivated voxels of participant 3 and participant 8, the results of the two methods are largely different. 4. Discussion We have presented a new fMRI data analysis method, which is more accurate than the conventional method. The deactivation ratios of the temporal lobe obtained by the new method are approximately 30%-33%, which probably mean the deactivations of interneurons by considering the ratio of the population of interneurons, the energy consumption of interneurons, and the inhibition and the disinhibition of interneurons. The deactivation ratios of the motor cortex are almost the same as those of the temporal lobe. The energy consumption of interneurons in humans are not well known. The inhibition and the disinhibition of interneurons are controversial. Therefore, in this paper, we have carried the argument on deactivation ratios under the obscure information. However, in the future, the argument will be more precise. Those who had the health problems such as high blood pressure, low blood pressure, lack of sleep, fatigue, and hunger, and those who take medicines (especially, the medicines that affect the cardiovascular system) were excluded from the experiments. However, as a reference, we also conducted the experiments with the above people. The deactivation ratios of the above people were higher (approximately 40%-60%) than Table 6. The reason may be local blood stealing, wherein blood is diverted to neurally active regions without a concomitant change of neural activity in the negative BOLD regions(26), because those who have the health problems may be unable to supply the sufficient blood to the activation areas, and the vascular systems of those who take medicines may not work well due to the medicines. The deactivation ratios of those who have the health problems or take medicines may equal the deactivations of interneurons plus local blood stealing. Participant No.9 had a little high blood pressure, and participant No.10 had a little low blood pressure. Therefore, the deactivation ratios of No.9 temporal lobe, No.10 temporal lobe and No.10 motor cortex are a little large among the deactivation ratios (See Table 6). Moreover, their MSPEs are a little greater than those of the others (See Table 6), perhaps because blood stealing does not occur synchronously with 18 tasks/rests, which increases their MSPEs. We conducted the experiments of simple repetition which activates the temporal lobe. We also conducted other experimental tasks which activate the temporal lobe. As far as we experienced, when the activation areas of an experimental task were large, the deactivation ratios tended to get larger. The reason may be that experimental tasks whose activation areas are large need a lot of blood, which easily causes local blood stealing. Therefore experimental tasks whose activation areas are large may be inappropriate for detecting the deactivations of interneurons. The (de)activated voxels detected by the conventional method, which are largely different from those by the new method(8V), may be incorrect, because the regression formulas of the conventional method are less accurate than those of the new method(8V). Moreover, in the conventional method, right-tailed t-tests are usually performed, which is based on the idea that only the excitations of excitatory neurons are the activations. That is, right-tailed t-tests neglect the deactivations of inhibitory interneurons. However, a large portion of the deactivations of inhibitory interneurons is also considered to be activations. Therefore, the right-tailed t-test does not detect the whole activation. A lot of fMRI studies thus far by the conventional method should be re-examined by the new method, and many results obtained thus far will be modified. Interneurons in the human cortex can hardly be investigated due to several problems (e.g. ethical problems). As the new method can detect the deactivations of interneurons, the new method will be a powerful tool for the study of the interneurons in the human cortex. The loss or dysfunction of interneurons is related to several diseases such as epilepsy, Alzheimer's disease, schizophrenia and so on(27,28,29). As the new method can detect the deactivations of interneurons, the new method will be a powerful tool for the image diagnosis of the above diseases. 5. Methods 5.1 Participants Ten participants in the experiments were drawn from right-handed healthy volunteers with normal blood pressures. Those who had the health problems such as lack of sleep, fatigue, and hunger and those who took medicines (especially, the medicines that affect the cardiovascular system) were excluded. Eight participants were male in their twenties. One participant was male in his sixties. One participant was female in her twenties. The study was approved by the 19 local ethics committee, and complied with the relevant ethical regulations. Written informed consents were obtained from all participants. 5.2 fMRI measurements fMRI measurements were performed with a 1.5T MRI system. The measurement parameters were as follows: slice thickness: 4mm, slice gap: 1mm, slice number: 24, matrix size: 64×64, echo time: 45msec, repetition time: 6000msec, flip angle: 90°, field of view: 240mm×240mm. The block design consisted of six task blocks and six rest blocks. Each block consisted of eight volumes (96 volumes in total). 5.3 Data analysis In Section 2 and Section 3, we explained the main points. We explain what we did not mention in Section 2 and Section3. 1.Preprocessing included realignment, normalization and smoothing (full width at half maximum (FWHM) = 8mm or 6mm) using SPM8(21) . Tasks/rests were convolved with HRF. 2. When cubes with a side of 2 voxels are generated in the temporal lobe, some cubes consist of the voxels inside the temporal lobe and the voxels outside the temporal lobe. We performed the regression analysis of the cubes consisting of the voxels inside the temporal lobe and the voxels outside the temporal lobe. We performed t-tests on the coefficients inside the temporal lobe, but we did not perform t-tests on the coefficients outside the temporal lobe. We did the same treatment in the motor cortex. 3. In a cross validation, regression analyses are performed with 95(=96－1) volumes 96 times, and 96 different regression formulas are generated. Which regression formula should be used for t-test? There is no special reason to select a certain regression formula among them for t- test. Therefore, regression analysis with 96 volumes was also performed to generate the regression formula for t-test. 4. T-tests were performed until the number of (de)activated voxels reaches a certain predetermined number (for example, 300). Sometimes, the number of (de)activated voxels did not reach the predetermined number. For example, when a participant did not perform an experimental task appropriately, the regression formula obtained was inaccurate, and the number of (de)activated voxels did not reach the predetermined number. 20 Acknowledgements We thank Hiroyuki Hioki for teaching interneurons. We thank the participants in the experiments. References 1. Ogawa S, Lee T-M, Kay A-R, Tank D-W (1990) Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc. Natl. Acad. Sci. U. S. A. 87:9868–9872. 2. Frankenstein U, Wennerberg A, Richter W, Bernstein C, Morden D, Remy F, Mcintyre M (2003) Activation and Deactivation in Blood Oxygenation Level Dependent Functional Magnetic Resonance Imaging. Concepts in Magnetic Resonance Part A. 16A:63–70. 3. Hayes D-J, Huxtable A-G (2012) Interpreting Deactivations in Neuroimaging. Front Psychol. 3: 27. 4. Mullinger K-J, Mayhewb S-D, Bagshawb A-P, Bowtell R, Francis S-T(2014) Evidence that the negative BOLD response is neuronal in origin:A simultaneous EEG–BOLD–CBF study in humans. NeuroImage 94:263–274. 5. Frackowiak R-S-J, Friston K-J, Frith C-D, Dolan R-J, Price C-J, Zeki S, John T. Ashburner J-T, Penny W-D (2004)Human Brain Function, Second Edition. Academic Press. London. 6. Perrone-Bertolotti M, Rapin L, Lachaux J-P, Baciu M, Loevenbruck H(2014) What is that little voice inside my head? Inner speech phenomenology, its role in cognitive performance, and its relation to self-monitoring. Behavioural Brain Research 261:220-239. 7. Del Río M-R, DeFelipe J(1996) Colocalization of calbindin D-28k, calretinin, and GABA immunoreactivities in neurons of the human temporal cortex. J. Comp. Neurol. 369:472- 482. 8. Draper N-R, Smith H(1998) Applied Regression Analysis, Third Edition. Wiley- Interscience NewYork. 9. Lin C-S, Lu S-M, Schmechel D-E(1986) Glutamic acid decarboxylase and somatostatin immunoreactivities in rat visual cortex. J. Comp. Neurol. 244:369-383. 10. Meinecke D-L, Peters A (1987) GABA immunoreactive neurons in rat visual cortex. J. Comp. Neurol. 261:388-404. 11. Beaulieu C. (1993) Numerical data on neocortical neurons in adult rats, with special reference to the GABAergic population. Brain Res. 609:284-292. 21 12. Patel A-B, de Graaf R-A, Mason G-F, Rothman D-L, Shulman R-G, Behar K-L(2005) The contribution of GABA to glutamate/glutamine cycling and energy metabolism in the rat cortex in vivo. Proc. Natl. Acad. Sci. USA 102:5588–5593. 13. Buzsáki G, Kaila K, Raichle M (2007) Inhibition and brain work. Neuron 56:771-783. 14. Kann O, Papageorgiou I-E, Draguhn A(2014) Highly energized inhibitory interneurons are a central element for information processing in cortical networks. J. Cereb. Blood Flow Metab. 34:1270–1282. 15. Pi H-Y, Hangya B, Kvitsiani, Sanders J-I, Huang Z-H, Kepecs A (2013) Cortical interneurons that specialize in disinhibitory control.Nature 503:521–524. 16. Karnani M-M. Jackson J, Ayzenshtat I, Sichani A-H, Manoocheri K, Kim S, Yuste R(2016) Opening Holes in the Blanket of Inhibition: Localized Lateral Disinhibition by VIP Interneurons. J. Neurosci. 36:3471-3480. 17. Pfeffer K-C, Xue M, He M, Huang Z-J, Scanziani M(2013) Inhibition of Inhibition in Visual Cortex: The Logic of Connections Between Molecularly Distinct Interneurons. Nat. Neurosci. 16:1068–1076. 18. Neske G-T, Connors B-W(2016) Distinct Roles of SOM and VIP Interneurons during Cortical Up States. Front Neural Circuits 10:52. 19. Garcia-Junco-Clemente P, Ikrar T, Tring E, Xu X, Ringach D-L, Trachtenberg J-T (2017)An inhibitory pull–push circuit in frontal cortex. Nature Neuroscience 20:389–392. 20. Hioki H, Okamoto S, Konno M, Kameda H, Sohn J, Kuramoto E, Fujiyama F, Kaneko T(2013) Cell type-specific inhibitory inputs to dendritic and somatic compartments of parvalbumin-expressing neocortical interneuron. J. Neurosci. 33:544-555. 21. University College London (2017) SPM - Statistical Parametric Mapping. http://www. fil.ion.ucl.ac.uk/spm/ . 22. University of South Carolina (2017) MRIcro. http://www.mccauslandcenter.sc.edu /crnl /mricro/. 23. Newton J-M, Sunderland A, Gowland P-A(2005) fMRI signal decreases in ipsilateral primary motor cortex during unilateral hand movements are related to duration and side of movement. NeuroImage 24:1080–1087. 24. Kastrup A, Baudewig J, Schnaudigel S, Huonker R, Becker L, Sohns J-M, Dechent P, Klingner C, Wittec O-W (2008) Behavioral correlates of negative BOLD signal changes in the primary somatosensory cortex. NeuroImage 41:1364–1371. 22 25. Schäfer K, Blankenburg F, Kupers R, Grüner J-M, Law I, Lauritzen M, Larsson H-B-W (2012) Negative BOLD signal changes in ipsilateral primary somatosensory cortex are associated with perfusion decreases and behavioral evidence for functional inhibition. NeuroImage 59:3119–3127. 26. Bressler D, Spotswood N, Whitney D (2007) Negative BOLD fMRI Response in the Visual Cortex Carries Precise Stimulus-Specific Information. PLoS ONE 2:e410. 27. Jacob J. (2016) Cortical interneuron dysfunction in epilepsy associated with autism spectrum disorders. Epilepsia 57:182-93. 28. Lewis D-A, Curley A-A, Glausier J, Volk D-W(2012) Cortical parvalbumin interneurons and cognitive dysfunction in schizophrenia. Trends in Neurosciences 35:57-67. 29. Schmid L-C, Mittag M, Poll S, Steffen J, Wagner J, Geis H-R, Schwarz I, Schmidt B, Schwarz M-K, Remy S, Fuhrmann M. (2016) Dysfunction of Somatostatin –Positive Interneurons Associated with Memory Deficits in an Alzheimer's Disease Model. Neuron 92:114–125. 23
1904.03399
1
1904
2019-04-06T09:37:27
Functional Geometry of Human Connectome and Robustness of Gender Differences
[ "q-bio.NC", "cond-mat.dis-nn", "math.AT" ]
Mapping the brain imaging data to networks, where each node represents a specific area of the brain, has enabled an objective graph-theoretic analysis of human connectome. However, the latent structure on higher-order connections remains unexplored, where many brain regions acting in synergy perform complex functions. Here we analyse this hidden structure using the simplicial complexes parametrisation where the shared faces of simplexes encode higher-order relationships between groups of nodes and emerging hyperbolic geometry. Based on data collected within the Human Connectome Project, we perform a systematic analysis of consensus networks of 100 female (F-connectome) and 100 male (M-connectome) subjects by varying the number of fibres launched. Our analysis reveals that the functional geometry of the common F\&M-connectome coincides with the M-connectome and is characterized by a complex architecture of simplexes to the 14th order, which is built in six anatomical communities, and short cycles among them. Furthermore, the F-connectome has additional connections that involve different brain regions, thereby increasing the size of simplexes and introducing new cycles. By providing new insights into the internal organisation of anatomical brain modules as well as into the links between them that are essential to dynamics, these results also highlight the functional gender-related differences
q-bio.NC
q-bio
Functional Geometry of Human Connectome and Robustness of Gender Differences Bosiljka Tadi´ca,b, Miroslav Andjelkovi´cc, Roderick Melnikd aDepartment of Theoretical Physics, Jožef Stefan Institute, Jamova 39, Ljubljana, Slovenia; bComplexity Science Hub, Josefstaedter Strasse 39, Vienna, Austria; c Institute for Nuclear Sciences Vinca, University of Belgrade, 11000 Belgrade, Serbia; dMS2Discovery Interdisciplinary Research Institute, M2NeT Laboratory and Department of Mathematics, Wilfrid Laurier University, Waterloo, ON, Canada Abstract Mapping the brain imaging data to networks, where each node represents a specific area of the brain, has enabled an ob- jective graph-theoretic analysis of human connectome. However, the latent structure on higher-order connections remains unexplored, where many brain regions acting in synergy perform complex functions. Here we analyse this hidden structure using the simplicial complexes parametrisation where the shared faces of simplexes encode higher-order relationships be- tween groups of nodes and emerging hyperbolic geometry. Based on data collected within the Human Connectome Project, we perform a systematic analysis of consensus networks of 100 female (F-connectome) and 100 male (M-connectome) subjects by varying the number of fibres launched. Our analysis reveals that the functional geometry of the common F&M- connectome coincides with the M-connectome and is characterized by a complex architecture of simplexes to the 14th order, which is built in six anatomical communities, and short cycles among them. Furthermore, the F-connectome has additional connections that involve different brain regions, thereby increasing the size of simplexes and introducing new cycles. By providing new insights into the internal organisation of anatomical brain modules as well as into the links between them that are essential to dynamics, these results also highlight the functional gender-related differences. Introduction Human psychology and behaviour are determined by functional brain connectivity among neurons, neural assemblies, or entire regions, making the patterns of circuitry that can be detected by brain imaging [1]. Recent large-scale research into the brain imaging data within the Human Connectome Project (HCP) [2, 3, 4] aims to uncover, describe and understand the functional structure of human connectome; the connectome is visualised as a network consisting of different brain regions (grey matter) and paths between them (white-matter fibre bundles) that can be determined by mapping the diffusion-MRI and tractography data. The network nodes are identified as distinct brain regions that are functionally similar and spatially close as well as equally connected to the other regions [5, 6, 7, 4]. The connections between these regions, which are determined from brain imaging data, can depend on a number of factors, and vary among different subjects, performed tasks and conditions. Therefore, the consensus between the pipelines in the structural connectome can be mapped from a large population tractography data [8] and depends on many parameters. Based on the data from HCP [2] and the brain mapper developed in [9], the Budapest connectome server [10] provides the possibilities to infer the consensus networks at a variety of the relevant parameters, as described in [11, 12]. The mapping of imaging data to the brain networks enables an objective analysis based on graph theory methods [13, 14]. Recently, different studies of brain imaging data revealed the strong evidences for gender differences in the structural connectome [15, 16, 17, 18, 19, 20, 21]. This subject was not well researched, but already it brought some controversial debates [22]. The exact origin of these gender differences and their potentials and impact on the level of individual and social behaviour are still to be investigated [23]. On the other hand, the current degree of reliability of the connectome data provides an opportunity for a mathematical analysis of structural differences at all levels. For example, a recent study [21] has shown that the consensus female connectome has superior connectivity than the consensus male connectome in many graph-theoretic measures. Recent investigations of geometrical properties of various complex systems [24, 25, 26, 27, 28, 29, 30, 31, 32, 33] show the relevance of the higher-order connectivity beyond standardly considered pairways interactions. Mathematically, the impact of these higher order interactions is adequately described by the simplicial complexes in the algebraic topology of graphs [34, 35, 36, 37]. In these complexes, elementary geometrical shapes (triangles, tetrahedra, and simplexes of higher order) are combined through shared substructures of various orders. These geometrical structures directly influence dy- namic processes that the complex system in question performs, such as transport, diffusion, or synchronisation among the involved nodes. In the case of brain networks, the main dynamic function pertains to maintaining an optimal balance be- tween the processes of integration and segregation where different regions of the brain can be simultaneously involved and the present modular structure of the brain plays an important role [38, 39, 40, 41]. Anatomical modules of the brain, which are recognized as different mesoscopic communities in the network [42, 43, 44, 45], are based on spatial topography and coexpression of genes in the brain cells [46]. It has been suggested that each module performs a discrete cognitive function while specific connector nodes take on communication between modules [40]. However, the fine functional organisation in- side these modules remains unexplored. Besides, the occurrence of simplicial complexes causes the emergent hyperbolicity or a negative curvature [47] in the structure of the graph, which affects its functional properties. In this sense, the complete graph and associated tree are ideally hyperbolic, characterised by the hyperbolicity parameter δ = 0. The graphs with small values of δ are subject to intensive investigations for their ubiquity in natural and social systems, as well as in technology applications [48, 24, 25, 30, 33]. Moreover, current theoretical studies reveal that Gromov hyperbolic graphs with a small hyperbolicity parameter have specific mathematical properties [48]. In particular, the bounds for the δ-parameter of the whole graph can be derived from subjacent simpler graphs, for example, induced cycles or clique separators of a given length [49, 50, 51, 52, 53, 54]. Therefore, the study of the hyperbolicity of brain graphs can reveal the presence of typical local structures that are potentially decomposable into some known forms, which underlie the brain's dynamic complexity. In this work, we considerably expand the analysis of human connectome beyond the simple pairwise connectivity. Using the mathematical techniques of algebraic topology of graphs, we identify hierarchically organised complexes that encode higher-order relationships between regions of the brain and explore the hyperbolic geometry of brain graphs. We consider the consensus connectomes mapped from 100 female (F-connectome) and 100 male (M-connectome) subjects, using the brain mapper and imaging data from the Human Connectome Project, which is provided by the Budapest server 3.0 [10]. The weighted edges are inferred according to the electrical connectivity criteria, which are most sensitive to the number of fibres observed in the tractography data. We analyse the connectomes that correspond to the significant variation in the number of fibres launched (see Methods). With the appropriate topology measures, our objectives are to determine the hidden structure of human connectome endowed with the relationships between groups of nodes and express the possible gender differences in this context. To this end, we construct and investigate a common F&M-connectome at different numbers of fibres and determine its structure, parametrised by simplicial complexes, and the graph's hyperbolicity parameter. Furthermore, by comparing edges in the F- and M-connectomes, we identify the excess edges that appear consistently in the F-connectome with an increased number of fibres. Our mathematical analysis reveals a rich structure of simplicial complexes that are common to the F&M-connectome and belong to different brain anatomical communities and cycles that connect them inside and across the two brain hemispheres. It further confirms the higher connectivity of the F-connectome and demonstrates that the excess edges have a well-organised structure that includes a particular set of paths and brain regions. Methods Input Data & Consensus Networks. We downloaded the data for male and female connectomes from the Budapest connectome server 3.0 [10]. Using the data from HCP [2] and the brain mapper provided in [9] the server produces the connectome corresponding to the settings of a variety of parameters, the meaning of which is specified in [11, 12]. For our study, we have selected the data that provide the consensus networks for female connectome and male connectome based on 100 subjects of each gender. The corresponding brain networks consist of N=1015 nodes (brain regions) and the weights of the connections between them determined according to the electrical connectivity criteria, i.e., the number of fibres between the considered pair of regions is divided by the average fibre length. We consider three different fibre counts, comprising of NF = 20K, 200K, and 1000K fibres, where for short K≡ 1000. For the additional parameters, we have set the minimum edge confidence as 100%, minimum edge weight as 4, and the median weight calculation. The resulting adjacency matrices of the weighted networks, herewith called F-connectome and M-connectome, respectively, are downloaded together with the node labels, i.e., standardly accepted names of the brain regions. Gromov hyperbolicity parameter of graphs. A generalization of the Gromov notion of hyperbolicity [47] is applied to graphs endowed with the shortest-path metric. Specifically, the 4-point Gromov criterion states that a graph G is δ- hyperbolic if f for any four vertices (A, B, C, D) there is a fixed small value δ(G) such that the following relation beween the sums of distances S ≡ d(A, B) + d(C, D) ≤ M ≡ d(A, C) + d(B, D) ≤ L ≡ d(A, D) + d(B, C) implies d(A, D) + d(B, C)− d(A, C)− d(B, D) ≤ 2δ(G). Thus, for a δ-hyperbolic graph, there is δ(G) such that any four nodes of the graph satisfy the condition ≤ δ(G) . (1) δ(A, B, C, D) ≡ L − M 2 From the triangle inequality, the value of (L−M)/2 is bounded brom above by the minimal distance dmin ≡ min{d(A, B), d(C, D)} in the smallest sum S. This relationship enables a direct computation of the hyperbolicity parameter of a graph, which is given by its adjacency matrix. In particular, by sampling a large number (109) 4-tuples of vertices we plot δ(A, B, C, D) against the corresponding dmin; the plot saturates at larger distances. We compute the average (cid:104)δ(cid:105) for all dmin as well as δmax = maxG{δ(A, B, C, D)}, which gives δ(G). We also determine the distribution P (d) of the shortest-path distances d on the graph. The largest distance defines the graph's diameter D, which gives the upper bound to the hyperbolicity parameter, δ(G) ≤ D/2. As mentioned above, the hyperbolic graphs with a small parameter δ have a specific structure of subgraphs, from which the upper bound of δ(G) can be derived [50, 51, 52, 53, 54]. In this context, the following definitions apply. A subgraph Γ of G is called isomet- ric if the distance between every pair of vertices (A, B) ∈ Γ is equal to the distance between them measured on G, i.e., dΓ(A, B) = dG(A, B). A cycle Cn is a sequence of n pairwise connected vertices with n + 1 → 1; an induced cycle does not contain a chord, an edge connecting nonconsecutive vertices. A clique of size s ≡ qmax + 1 is the full graph of s vertices and s(s − 1)/2 edges. Q-analysis of graphs: definition of structure vectors. Considering a connectome as an undirected and unweighted graph G, the higher-order connectivity of its vertices can be appropriately parametrised by the maximal complete subgraphs (or cliques) whose vertices belong to a clique complex C(G) in the graph G [36]. Two cliques σr and σq of the orders r, q can be interconnected by sharing some vertices; then the structure made by the shared vertices represents a common face of both cliques. For example, if for r < q all vertices of σr belong to σq, then the simplex σr represents a face of the order r in the simplex σq. The simplicial complex represents the aggregate of cliques that share the faces of different orders q = 0, 1, 2··· q(cid:48) max indicates the order of the largest clique in the complex. The order of a simplicial complex is the largest order of a simplex in it; we denote by qmax the order of the largest complex in the entire graph. max − 1, where q(cid:48) Applying the Bron-Kerbosch algorithm [55], the adjacency matrix of the graph G is converted into the incidence matrix Λ, which contains all cliques in the graph by identifying the vertices that belong to them; using this information, we then find how different cliques interconnect via shared nodes to make the higher-order structures. The overall hierarchical 2 organisation of the graph can be quantified [56, 57, 58, 59, 26, 27] by three structure vectors having the components along different topology levels q = 1, 2, 3,··· qmax. Specifically, for each considered graph, we determine: • FSV -- the first structure vector {Qq}, where each Qq represents the number of q-connected components; • SSV -- the second structure vector {nq}, where nq indicates the number of connected components from the level q upwards; • TSV -- the third structure vector { Qq} is introduced to quantify the degree of interconnectivity between cliques at each level q, and can be derived from the other two as { Qq} = 1 − Qq/nq. These structure vectors provide a measure of the graph's global architecture (see [60] for the application of Q-analysis for the vertex neighbourhood). For completeness, we also determine standard graph measures [61, 62], and community structure [43, 63, 64] of the typical connectome graph, see Results. Visualisation and standard graph parameters are made by using Gephi software [65]. Results Consensus Networks of Human Connectome According to the parameter settings (see Methods), the considered F-connectome consists of the edges that appear in all 100 female subjects, and similarly, the M-connectome contains the edges that are present in all 100 male subjects. For the illustration, the F-connectome at 1000K fibres is shown in Fig. 1 with the labelled brain regions as nodes. Here, we use the simplicial complexes parametrisation (see Methods) and the graph's hyperbolicity measures to uncover the hidden structure of human connectome, which is encoded in the higher-order connectivity between groups of nodes. Furthermore, using these mathematical measures, we analyse the variations of the brain connectivity patterns depending on the number of fibres launched and the gender of the subjects. As we will show in the following, these differ significantly depending on the number of launched fibres NF and between the genders. Figure 1: The female connectome at the highest resolution consisting of 1115 nodes (brain regions) and 11339 edges between them. The network is deduced from the HCP data provided at the server [10] with weighted edges as the median for 100 female subjects and NF =1000K fibres launched between each pair of nodes. To proceed, we first identify all edges that (although with different weights) are common for both F-connectome and M- connectome, here called C F &M -connectome at different NF . Table 1 and Fig. 2 summarise the number of edges and mutual relationships of different connectomes. Fig. 3 shows the corresponding graphs with the labelled brain regions, obtained for NF =200K and 1000K. Specifically, we find that: • The number of established edges in each considered connectome increases with the number of fibres launched NF ; 3 Figure 2: Schematic view of the number of edges E and their co-occurrence in the connectomes at the increasing number of fibres NF , see also Table 1. The common C F &M -connectome at a large number of fibres, NF , inherits all edges from the C F &M at a lower NF , black lines, and a fraction of the excess edges of the F-connectome, shown by pink lines. The top line (red) shows the number of robust excess edges in F-connectome which do not appear in any of the larger common C F &M -connectomes. Table 1: For the number of launched fibres NF , the corresponding number of edges are shown in the consensus male (M) and female (F) connectomes, the edges C F &M common to F&M connectomes, and the total number F e0 of excess edges in the F-connectome; the fractions of F e0 indicated as F ec+ and F ecc+ are the edges that appear in the common connectomes at the two higher NF , respectively, while F ex are the excess edges also at the higher NF . The difference between M and C F &M at 20K and 200K consists of 12 and 16 edges, which all appear in C F &M at 1000K. NF 20000 200000 1000000 M 776 4285 7110 F 1548 7634 11339 C F &M F e0 784 3365 4229 764 4269 7110 F ec+ F ecc+ 753 2170 27 - - - F ex 4 1195 ≥1195 • The common C F &M -connectome practically coincides with the M-connectome at each NF , whereas the F-connectome contains an increasing number of excess edges with the increasing NF ; • The common C F &M -connectome at a higher NF inherits all edges from the C F &M -connectome at a lower NF ; • A significant fraction of the excess edges found in the F-connectome at a lower NF appear in the common C F &M - connectome but at a higher NF ; • There is a large number of the excess edges in the F-connectome that are never found in the common C F &M - connectome at a higher NF ; the patterns of these edges make the fundamental difference between the human female and male connectomes. The Structure of Simplicial Complexes in Brain Graphs According to Table 1 and Fig. 2, at each NF , the common F &M-connectome practically coincides with the male con- nectome (apart from the exact weights of edges) while there are many excess edges in the female connectome. Here, by applying Q-analysis (see Methods) to the corresponding graphs at different numbers of fibres NF , we show that (i) the com- mon human connectome possesses a nontrivial hidden structure encoding multi-vertex connectivity; (ii) the excess edges of the F-connectome are not random but exhibit a highly organised structure, which thus implies a specific functionality, cf. Fig. 3. In Fig. 4 the results for the three structure vectors, defined in Methods, are presented for different NF . As Fig. 4 shows, the structure of connectomes becomes richer with the increased number of fibers NF . In particular, the cliques of a systematically larger order q appear and the degree of their inter-connectivity increases as measured by TSV. Moreover, the larger number of edges in the F-connectome leads to a much richer structure of the simplicial complexes, which is expressed by all structure vectors, cf. right panels of Fig. 4. We also notice that the difference between the M- and F-connectomes systematically increases with the increased NF . Representative quantitative properties are given in Table SI-I and Table 4 101102103NF102103104E(764)(4269)(753)(2170)(27)(4)(1195)CF&M20KCF&M200KCF&M1000K Figure 3: Networks of connections established among labelled brain regions at different numbers of launched fibres NF : (a) Common M&F connectome at NF = 200K and (c) common M&F connectome at NF = 1000K, the weights of M-connectome are shown. (b) and (d) The patterns of the additional edges appearing in the F-connectome (F-excess), which are not present in the M-connectome at NF =200K and NF = 1000K, respectively. The numbers of edges in the corresponding graph are indicated at each figure. The number of edges is inherited in the target graph at NF = 1000K from the graphs at NF = 200K. Explicitly, the graph (c) inherits all edges from the graph (a). The 2170 edges from the graph (b) appear in the common connectome (c), whereas 1195 edges of the graph (b) are inherited as the excess edges in the graph (d). SI-II in Supplementary Information. Noticeably, the Qq=0 component of the FSV, which gives the number of fragments of the graph, suggests that besides the largest component some vertices and small clusters remain disconnected. The number of fragments decreases and the connectivity increases with the increasing NF . The corresponding number of edges in the largest cluster is given in Table 1. The organisation of the present edges at each NF manifests in the presence of simplicial complexes with the largest order qmax. From Fig. 4 and Table SI-II, we see that the F-connectome possesses the cliques of a higher order; the difference increases from qM max = 20, at 1000K. The number of cliques of the highest order is different, as well as their connection to the other cliques at the level just below the qmax. Apart from the increased number of topology levels, the F-connectome also exhibits a significant degree of interconnections between the big cliques. For example, the TSV for the F-connectome at the level q = 13, which equals to qM max, is still very high, about 55%. Below, we identify the excess edges in the F-connectome and examine the patterns which they make. Table 2 shows a brief summary of different graphs' properties. max = 6, at 20K, to qM max = 5 and qF max = 13 and qF Hyperbolicity of the human connectome Neuroanatomy of the brain enclosed in a small volume of the skull was interpreted by the brain network which is embedded in a hyperbolic space [66]. Theoretically, the hyperbolicity of a path-connected geodesic metric space was proved [67, 68] to be equivalent to the hyperbolicity of the graph associated with it. In the brain graphs studied above, the hierarchical organisation of simplicial complexes reduces the distances between nodes in the graph's metric space, which implies their hyperbolicity. Here, using the 4-point Gromov criterion (see Methods), the hyperbolicity parameters are determined for F- and M-connectomes obtained by varying the number of fibres NF . In this context, we consider the corresponding adjacency matrix of the largest connected cluster as an unweighted symmetrical graph. Fig. 6 shows the results for the largest available NF = 1000K. In the bottom panels, the histograms of the distances between all pairs of vertices are plotted. Although the diameter D = 8 applies to both graphs, typical distances in the F-connectome appear to be smaller. In the top panels, we plot 5 Figure 4: The components of three structure vectors defined in Methods (FSV,SSV, TSV) plotted against the topology level q for the consensus connectomes determined from 100 male (left) and 100 female (right) subjects with the varied number of fibres NF , indicated in the legend. the values of the δ-parameter against the minimum distance dmin of a given 4-tuple, as described in Methods. Specifically, lower sets of curves represent the average value (cid:104)δ(cid:105) for a given dmin. Whereas the top lines contain the recorded maximum value δmax from all considered 4-tuples. We observe that the values of (cid:104)δ(cid:105) are very low, practically never exceed 0.25, which suggests the impact of the types of local structures populated by cliques. They are 0-hyperbolic subgraphs (atoms) [52] and induced cycles, whose hyperbol- icity depends on the length of the cycle and can be expressed as a multiple of 1/4 [48]. Moreover, δmax = 3/2 suggests that dominant isometric subgraphs, which determine the value of δmax for the whole graph [54] in both connectomes, can be cycles Cn that have n ≥ 6 but with the diameter D ≥ 3. While we regularly obtain δmax = 3/2 in the M-connectome, it was necessary to sample 109 different 4-tuples to find it in the F-connectome. Meanwhile, the value of δmax = 1 occurs often in the F-connectome. It suggests that the dominant subgraphs can be composed of cliques that are one-edge apart, which, according to the results in [52, 32], yields that δmax = δclique + 1 or they contain short cycles isomorphic to 4-cycle [48]. The situation is considerably different at the lower number of fibres where both F- and M-connectomes have gradually fewer edges (see Table 1). Consequently, the distances between vertices increase as well as the diameters of the graphs. The increased distances lead to the appearance of larger cycles and yield the distortion of the hyperbolicity parameter [51] while the graphs remain hyperbolic; we find the upper bound δmax ≤ 4 in both connectomes, as shown in Fig. 6. Figure 5: Top panles: Hyperbolicity parameters δmax (upper curves) and (cid:104)δ(cid:105) (lower curves) of the consensus connectome of female (right) and male (left) for NF =1000K fibres launched. Three lines are for 107, 108 and 109 sampled 4-tuples. Lower panels: The distribution P (d) of the shortest-path distances d for the corresponding female and male connectomes. The structure of common F&M-connectome and the excess edges in Female connectome By performing the edge-by-edge comparisons in the corresponding graphs, see Fig. 3, we identify every edge in terms of its source and destination vertex and the weight. For the highest NF , the common F &M-connectome consists of 7110 6 00.20.40.60.81TSV0100020003000SSV05101520q0200400600FSV M_1000KM_200KM_20K05101520qF_1000KF_200KF_20K02468dmin0.00.51.01.52.0<δ> δmax02468dmin02468d05×1041×1052×105P(d)Male02468dFemale Figure 6: Hyperbolicity parameters δmax and (cid:104)δ(cid:105) (top panels) and the shortest-path distances distribution (bottom panels) of the consensus female and male connectomes for the numbers of fibres NF = 20K (left) and 200K (right), shown at the same scale. The additional lines with triangle symbols in the right panels correspond to the excess edges in the female connectome at 200K, described in the text as F-excess1195. The number of sampled 4-tuples is 109. edges which coincide with the structure of the M-connectome, cf. Table 1 and Fig. 2. The corresponding network of the M-connectome, as shown in Fig. 7a, possesses a characteristic community structure related to different anatomical brain regions. Apart from the heterogeneity of the structure due to different degrees and weights of edges, this community structure is essential for the brain functional complexity [43, 42, 44, 45, 39, 40, 41] for both F- and M-connectomes. As mentioned above, the F-connectome possesses an extra structure on the top of the common F&M-connectome; it consists of many edges that connect different brain regions. The number of the extra edges varies with the number of launched fibres NF , as shown in Table 1. A subgraph of the identified excess edges in the F-connectome, here termed F-excess1195, consists of 1195 edges which systematically appear in the F-connectome, first at NF = 200K and then at NF = 1000K with increased weights; these edges are not present in the corresponding M-connectomes, and thus are not part of the universal F &M-connectome at the largest NF . A part of this graph, containing only the edges of a substantial weight, is shown in Fig. 7b. In the Supplementary Information list L-I, the names of source and target brain regions of these edges are given. The complete graph F-excess1195 is also shown in Fig-SI-3. Figure 7: (a) The common F&M-connectome at NF = 1000K with labelled brain regions belonging to the brain anatomical communities, indicated by different colours. Weights of the edges are from the M-connectome. (b) The robust structure of the excess connections among brain regions (labels) in the consensus female connectome that cannot be found in the consensus male connectome with up to 1000K fibres launched. Different colours indicate weighted communities. We show only the 490 edges with the significant weight in the tale of the weight distribution, cf. Fig. SI-1, and the involved 348 brain regions. It should be stressed that the excess edges observed in the F-connectome are attached to the central brain graph, the common F&M-connectome, at a large number of vertices. By considering F-excess as a separate graph, cf. Fig. 7b, we observe that these excess edges make nonrandom patterns and have a significant variation in weights (cf. Fig. SI-1); they involve 348 different brain areas in both hemispheres as well as the edges that connect the left and right hemispheres. The properties of the F-excess1195 subgraph are also summarised in Table 2, and the distribution of distances P (d), as well as the hyperbolicity parameters with δmax = 4 are shown in Fig. 6. Noticeably, the pattern of these extra connections 7 051015dmin0.01.02.03.04.0<δ> δmax051015dmin051015d05×1041×105P(d)20K200KMaleFemale051015dF-excess1195 in the F-connectome adds some larger cycles and 112 triangles. However, they are well embedded in the structure of the F&M-connectome, such that they do not appear as isomorphic cycles, and, consequently, do not increase the hyperbolicity parameter of the F-connectome. For comparison, we show the corresponding features of the randomised version of the F-excess1195 graph. Note that for this purpose we randomise the edges within each hemisphere separately while keeping the cross-hemisphere edges intact, so that the brain anatomical structure is observed. The parameters of the randomised graph are also shown in Table 2. Note that several other graph-theoretic properties, see the studies in reference [21], also differ in female and male connectomes. Table 2: Summary of graph parameters for the F-connectome and the M-connectome (which is equivalent to the common F&M-connectome) and the excess edges (F-excess) in the F-connectome at 1000K. The parameters of the F-excess1195 and its subgraph with large weights of edges F-ex1195w18, as well as its randomised version are shown. The quantities are computed for undirected graphs: the average degree < k >, path length < (cid:96) > and clustering coefficient < Cc >, the graph's density ρ, modularity mod and (the number of communities), diameter D, hyperbolicity parameter δmax, and the highest topology level qmax with the number (Qq) of the simplexes of that order. < k > < (cid:96) > < Cc > 12.07 7.01 4.17 1.77 1.41 0.94 0.69 0.67 0.13 0.064 0.031 0.006 3.45 3.97 4.36 5.91 6.54 9.95 ρ 0.025 0.014 0.008 0.005 0.008 0.003 mod 0.59 (6) 0.62 (6) 0.654 0.689 0.764 0.898 D δmax 3/2 8 3/2 8 11 5/2 4 17 4 19 30 5 qmax 20 (1) 13 (6) 3 (149) 2 (112) 2 (18) 2 (1) graph F-conn (Fig.1) M-conn (Fig.7a) F-excess (Fig.3d) F-excess1195 (Fig.SI-3) F-excess1195w18 (Fig.7b) randomised-F-excess1195 Discussion By analysing the HCP data provided at the Budapest connectome server, we acquired three sets of networks representing the consensus female and male connectomes at different numbers of launched fibres 20K, 200K, and 1000K. In addition to the standard graph parameters, by using algebraic topology methods we discovered a latent geometry that encodes higher-order connections in these brain graphs. Our main findings are: • Higher-order connectivity of the common F&M-connectome. We have shown that the human connectome, consisting of the edges that are common to both F&M connectomes, possesses a hidden structure beyond the node's pairwise connectivity. The higher-order connections between the groups of brain regions are suitably encoded by simplexes organised into larger complex structures and quantified by structure vectors, cf. Fig. 4. Remarkably, the complexity of the human connectome increases with the number of launched fibres, reaching the simplicial complexes of the order qmax + 1 = 14 at NF = 1000K. Specifically, there are six such cliques, which contain nodes in different brain modules (see Fig. SI-2 and the list L-I in Supplementary Information). We note that these simplicial complexes belong to different communities, which are anatomical mesoscopic structures of the brain graphs, cf. Fig. 7a. This architecture of connections in the brain graphs can be characterised by the tools of hyperbolic geometry. In particular, we find that they are Gromov hyperbolic graphs with small hyperbolicity constant δmax = 3/2, which characterises both F- and M-connectomes at 1000K launched fibres. Hyperbolicity varies with the network density, which is directly related to NF . In contrast, randomised (separately within each hemisphere) links exhibit much smaller simplexes (qrand max = 3) and increased hyperbolicity parameter that points to larger cycles. These findings indicate that the brain functional geometry consists of massive simplicial complexes as part of anatomical communities within each hemisphere as well as cycles that connect different regions inside and between the two hemispheres. • Structure of the excess edges in F-connectomes. F-connectome systematically appears to be better connected, i.e., has a more significant number of edges at every NF . Here, a more detailed inspection of the source-and-target brain region and the weight that identifies an edge indicates that two groups of excess edges occur: (1) The edges appearing in the F-connectome at a relatively low number of fibres which can appear in the M-connectome but only if a much larger number of fibres is launched; (2) The edges that robustly appear only in the F-connectome and have not been established in the M-connectome, including the highest available number 1000K of fibres. From the second group, the identity of 1195 edges that first appear at 200K in the F-excess subgraph and are not present in 8 the common F&M-connectome at 1000K are given in Supplementary Information. In particular, Fig.SI-3 shows the complete graph, while the list L-II contains only the edges with large weights. A comparison with the (inside the hemisphere) randomised graph has shown that these F-excess edges, considered as a separate graph, also have an organised structure involving a large number of brain regions, cf. Fig. 7b. Direct analysis and its hyperbolicity parameter suggest a geometry dominated by cycles and small simplexes. To summarise, our study reveals how the functional geometry of human connectome can be expressed by higher-order connectivity, described by simplicial complexes and induced cycles. This kind of structure is built into the anatomical communities of the brain at the mesoscopic scale in both brain hemispheres. However, the precise role of these simplicial complexes for the dynamical segregation in brain functional complexity remains to be better understood. In this context, the developed methodology provides new topological measures of the consensus brain networks and quantifies the robust gender differences. Specifically, a part of connections is more natural to invoke in the female than in the male brain, where much more fibres need to be launched to identify them. Whereas the other fraction of such connections consists of edges that appear exclusively in the consensus female connectome, they have not been identified in the consensus male connectome. It should be stressed that the considered consensus networks represent a kind of typical structures with the fixed number of vertices as 1015 brain regions while the edges are common for all 100 male and similarly for all 100 female, recorded within HCP in a representative set of (young and healthy) individuals. Note that, in each particular subject, the number of brain connections can deviate, e.g., being even considerably more abundant than in the respective consensus network. Moreover, the structure of possible connections is expected to vary with age, particular practice and with a development of diseases. Based on the brain imaging data, the methodology developed in this work would be suitable to reveal subtle differences between pairs of brains as well as changes in the brain of the same individual. Similar studies have been done with the patterns induced by the brain spontaneous fluctuations and content-related activity recorded by EEG [27, 30, 69], complementing the traditional methods. The application of our methodology to these issues warants a separate study which would include a more detailed investigation of the role of orientation and the weights of the edges. Conclusions Our analysis has revealed that the human connectome possesses a hyperbolic geometry and a complex structure on the scale between the node's edges and the mesoscopic anatomical communities within the cerebral hemispheres. This structure, composed of simplicial complexes of different sizes and cycles that connect them, accurately describes the higher-order connectivity among different regions of the brain, divided into anatomical modules. Therefore, it can provide a reliable basis for understanding the functional complexity of the brain. Moreover, the female connectome appears to have a structure different from the common F&M-connectome, not only in the number of edges but also in its organisation expressed by these higher-order connections. It might be conjectured that these excess connections imply additional functionality of the female connectome, which can have evolutionary, biological, biochemical, and even social origins. These issues go beyond our mathematical analysis of brain graphs. However, we believe that our findings can motivate further studies to better understand the origin and functional consequences of the apparent gender differences in the human connectome. 9 Bibliography [1] Babiloni, F., Cincotti, F., Babiloni, C., Carducci, F., Mattia, D., Astolfi, L., et al. Estimation of the cortical functional connectivity with the multimodal integration of high-resolution EEG and fMRI data by directed transfer function. NeuroImage, 24(1):118 -- 131 (2005). [2] McNab, BJ.A., Edlow, B.L., Witzel, T., Huang, S.Y., Bhat, H., Heberlein, K. et al. The human connectome project and beyond: Initial applications of 300mT/m gradients. NeuroImage, 80:234 -- 245 (2013). [3] Ganepola, T., Nagy, Z., Ghosh, A., Papadopoulo, T., Alexander, D.C. & Sereno, M.I. Using diffusion MRI to discrim- inate areas of cortical grey matter. NeuroImage, 182:456 -- 468 (2017). [4] Tittgemeyer, M., Rigoux,L., & Knösche, T.R. Cortical parcellation based on structural connectivity: A case for generative models. NeuroImage, 173:592 -- 603 (2018). [5] Rubinov, M. & Sporns, O. Complex network measures of brain connectivity: Uses and interpretations. NeuroImage, 52(3):1059 -- 1069 (2010). [6] Zalesky, A., Fornito, A., Harding, I.H., Cocchi, L., Yocel, M., Pantelis, C., & Bullmore, E.T. Whole-brain anatomical networks: Does the choice of nodes matter? NeuroImage, 50(3):970 -- 983 (2010). [7] Shen, X., Papademetris, X. & Constable, T.R. Graph-theory based parcellation of functional subunits in the brain from resting-state fmri data. NeuroImage, 50(3):1027 -- 1035 (2010). [8] Parker, C.S., Deligianni, F., Cardoso, M.J., Daga, P., Modat, M., Dayan, M., Clark, C.A., Ourselin, S. & Clayden, J.D. Consensus between pipelines in structural brain networks. PLoS ONE, 9(10):1 -- 10, (2014). [9] Zhang, Z., Descoteaux, M., Zhang, J., Girard, G., Chamberland, M., Dunson, D., Srivastava, A. & Zhu, H. Mapping population-based structural connectomes. NeuroImage, 172:130 -- 145 (2018). [10] Budapest reference connectome 3.0 https://pitgroup.org/connectome/ [11] Szalkai, B., Kerepesi, C., Varga, B. & Vince Grolmusz, V. The budapest reference connectome server v2.0. Neuro- science Letters, 595:60 -- 62 (2015). [12] Szalkai, B., Kerepesi, C., Varga, B. & Vince Grolmusz, V. Parameterizable consensus connectomes from the hu- man connectome project: the budapest reference connectome server v3.0. Cognitive Neurodynamics, 11(1):113 -- 116 (2017). [13] Sporns, O. Structure and function of complex brain networks. Dialogues Clin. Neurosci., 15(3):247 -- 262 (2013). [14] Pappo, D., Zanin, M., Pineda-Pardo, J.A., Boccaletti, S. & Buldu, J.M. Functional brain networks: great expectations, hard times, and the big leap forward. Philosophical transactions of the Royal Society of London. Series B, Biological sciences, 369(1653): 20130525 (2014). [15] Jahanshad, N. & Thompson, P.M. Multimodal neuroimaging of male and female brain structure in health and disease across the life span. Journal of Neuroscience Research, 95:371 -- 379 (2017). [16] Ruigrok, A. N. V., Salimi-Khorshidi, G., Lai, M.-C., Baron-Cohen, S., Lombardo, M. V., Tait, R. J., & Suckling, J. A meta-analysis of sex differences in human brain structure. Neuroscience and Behavioral Reviews, 39(100):34 -- 50 (2014). 10 BIBLIOGRAPHY BIBLIOGRAPHY [17] Miller, D.I. & Halpern, D.F. The new science of cognitive sex differences. Trends in Cognitive Sciences, 18(1):37 -- 45 (2014). [18] Taki, Y., Thyreau, B., Kinomura, S., Sato, K., Goto, R., Kawashima, R. & Fukuda, H. Correlations among brain gray matter volumes, age, gender, and hemisphere in healthy individuals. PLOS ONE, 6(7):1 -- 13, 07 (2011). [19] Zhang, C., Dougherty, C.C., Baum, S.A., White, T. & Michael, A.M. Functional connectivity predicts gender: Evi- dence for gender differences in resting brain connectivity. Human Brain Mapping, 39(4):1765 -- 1776 (2018). [20] Sun, Y., Lee, R., Chen, Y., Collinson, S., Thakor, N., Bezerianos, A. & Sim, K. Progressive gender differences of structural brain networks in healthy adults: A longitudinal, diffusion tensor imaging study. PLOS ONE, 10(3):1 -- 18, 03 (2015). [21] Szalkai, B., Varga, B. & Grolmusz, V. Graph theoretical analysis reveals: Women's brains are better connected than men's. PLOS ONE, 10(7):1 -- 30, 07 (2015). [22] Ingalhalikar, M., Smith, A., Parker, D., Satterthwaite, T.D., Elliott, M.A., Ruparel, K., Hakonarson, H., Gur, R.E., Gur, R.C. & Verma, R. Sex differences in the structural connectome of the human brain. Proceedings of the National Academy of Sciences, 111(2):823 -- 828 (2014). [23] Goyal, M.S., Blazey, T.M., Su, Y., Couture, L.E., Durbin, T.J., Bateman, R.J., Benzinger, T. L.-S., Morris, J.C., Raichle, M.E. & Vlassenko, A.G. Persistent metabolic youth in the aging female brain. Proceedings of the National Academy of Sciences, 201815917 (2019). [24] Albert, R., DasGupta, B. & Mobasheri, N. Topological implications of negative curvature for biological and social networks. Phys. Rev. E, 89:032811 (2014). [25] Narayan, O. & Saniee, I. Large-scale curvature of networks. Phys. Rev. E, 84:066108 (2011). [26] Andjelkovi´c, M., Tadi´c, B., Maleti´c, S. & Rajkovi´c, M. Hierarchical sequencing of online social graphs. Physica A: Statistical Mechanics and its Applications, 436:582 -- 595 (2015). [27] Tadi´c, B., Andjelkovi´c, M., Boshkoska, B.M. & Levnaji´c, Z. Algebraic topology of multi-brain connectivity networks reveals dissimilarity in functional patterns during spoken communications. PLOS ONE, 11(11):1 -- 25, 11 (2016). [28] Kleinberg, K.K., Boguña, M., Serrano, M.A. & Papadopoulos, F. Hidden geometric correlations in real multiplex networks. Nature Physics 12:1076 -- 1082 (2016). [29] Salnikov, V., Cassese, D. & Lambiotte, R. Simplicial complexes and complex systems. European Journal of Physics, 40:014001 (2018). [30] Tadi´c, B., Andjelkovi´c, M. & Šuvakov, M. Origin of hyperbolicity in brain-to-brain coordination networks. Frontiers in Physics, 6:7 (2018). [31] Bianconi, G. & Rahmede, C. Emergent hyperbolic network geometry. Sci. Rep., 7:41974 (2017). [32] Šuvakov, M., Andjelkovi´c, M. & Tadi´c, B. Hidden geometries arising in cooperative self-assembly. Sci. Rep., 8:1987 (2018). [33] B. Tadi´c. Self-organised criticality and emergent hyperbolic networks -- blueprint for complexity in social dynamics. European Journal of Physics, 40:024002 (2019). [34] Jonsson, J. Simplicial Complexes of Graphs. Lecture Notes in Mathematics, Springer-Verlag, Berlin, 2008. [35] Kozlov, D. Combinatorial Algebraic Topology. Springer Series "Algorithms and Computation in Mathematics", Vol. 21, Springer-Verlag Berlin Heidelberg, 2008. [36] Bandelt, H.-J. & Chepoi, V. Metric graph theory and geometry: a survey, in Goodman, J. E.; Pach, J.; Pollack, R., eds. "Surveys on discrete and computational geometry: Twenty years later". volume 453. Providence, RI: AMS, 2008. [37] Maleti´c, S. & Zhao, Y. Simplicial Complexes in Complex Systems: The Search for Alternatives. Harbin Institute of Technology, Harbin, Peoples Republic of China, first edition, 2017. 11 BIBLIOGRAPHY BIBLIOGRAPHY [38] Muñoz, M.A. Colloquium: Criticality and dynamical scaling in living systems. Rev. Mod. Phys. 90:031001 (2018). [39] Deco, G., Tononi, G., Boly, M. & Kringelbach, M.L.. Rethinking segregation and integration: contributions of whole- brain modelling. Nat Rev Neurosci, 16(7):430 -- 439 (2015). [40] Bertolero, M.A., Thomas Yeo, B.T. & D'Esposito, M. The modular and integrative functional architecture of the human brain. Proceedings of the National Academy of Sciences, 112(49):E6798 -- E6807 (2015). [41] Zamora-López G., Chen Y., Deco G., Kringelbach M. L. & Zhou C. Functional complexity emerging from anatomical constraints in the brain: the significance of network modularity and rich-club. Scientific Reports, 6:3842 (2016). [42] Meunier, D., Lambiotte, R. & Bullmore, E. Modular and hierarchically modular organization of brain networks. Frontiers in Neuroscience, 4:200 (2010). [43] Gronchi, G., Guazzini, A., Massaro, E. & Bagnoli, F. Mapping cortical functions with a local community detection algorithm. J. Complex Networks, 2(4):637 -- 653 (2014). [44] Mitra, P.P. The Circuit Architecture of Whole Brains at the Mesoscopic Scale. Neuron 83(6): 1273 -- 1283 (2014) [45] Zeng, H. Mesoscale connectomics. Current opinion in neurobiology 50:154-162 (2018). [46] Oldham, M.C., Konopka, G., Iwamoto, K., Langfelder, P., Kato, T., Horvath, S. & Geschwind, D.H. Functional organization of the transcriptome in human brain. Nature Neuroscience, 11:1271 (2008). [47] Gromov, M. Hyperbolic groups, pages 75 -- 263. Springer, New York, 1987. [48] Bermudo, S., Rodríguez, J.M., Rosario, O. & Sigarreta, J.M. Small values of the hyperbolicity constant in graphs. Discrete Mathematics, 339(12):3073 -- 3084 (2016). [49] Chepoi, V., Dragan, F.F., Estellon, B., Habib, M. & Vaxès, Y. Diameters, centers, and approximating trees of delta- In Proceedings of the 24th ACM Symposium on Computational Geometry, hyperbolicgeodesic spaces and graphs. College Park, MD, USA, June 9-11, 2008, pages 59 -- 68 (2008). [50] Bermudo, S., Rodríguez, J.M., Sigarreta, J.M. & Vilaire, J.-M.. Gromov hyperbolic graphs. Discrete Mathematics, 313(15):1575 -- 1585 (2013). [51] Carballosa, W., Pestana, D., Rodríguez, J.M. & Sigarreta, J.M. Distortion of the hyperbolicity constant in minor graphs. Electronic Notes in Discrete Mathematics, 46:57 -- 64 (2014). Jornadas de Matemática Discreta y Algorítmica. [52] Cohen, N., Coudert, D., Ducoffe, G. & Lancin, A. Applying clique-decomposition for computing gromov hyperbol- icity. Theoretical Computer Science, 690(Supplement C):114 -- 139 (2017). [53] Wu, Y. & Zhang, C. Hyperbolicity and chordality of a graph. The Electronic Journal of Combinatorics, 18:P43 (2011). [54] Martinez-Perez, A. Generalized chordality, vertex separators and hyperbolicity on graphs. arXiv:1708.06153v1, pages 1 -- 16 (2017). [55] Bron, C. & Kerbosch, J. Finding all cliques of an undirected graph. Comm. ACM, 16:575-577 (1973). [56] Freeman, C.L. Q-analysis and the structure of friendship networks. International Journal of Man-Machine Studies, 12(4):367 -- 378 (1980). [57] Gould, P. Q-analysis, or a language of structure: an introduction for social scientists, geographers and planners. International Journal of Man-Machine Studies, 13(2):169 -- 199 (1980). [58] Atkin, R.H. An algebra for patterns on a complex, ii. International Journal of Man-Machine Studies, 8(5):483 -- 498 (1976). [59] Maleti´c, S., Rajkovi´c, M. & Vasiljevi´c, D. Simplicial Complexes of Networks and Their Statistical Properties. Lecture Notes in Computer Science, 5102:568 -- 575 (2008). [60] Andjelkovi´c, M., Gupte, N. & Tadi´c, B. Hidden geometry of traffic jamming. Phys. Rev. E, 91:052817 (2015). 12 BIBLIOGRAPHY BIBLIOGRAPHY [61] Bollobas, B. Modern Graph Theory. Springer-Verlafg, Berlin Heidelberg, 1998. [62] Dorogovtsev, S.N. Lectures on Complex Networks. Oxford University Press, Inc., New York, NY, USA, 2010. [63] Mitrovi´c, M. & Tadi´c, B. Spectral and dynamical properties in classes of sparse networks with mesoscopic inhomo- geneities. Phys. Rev. E, 80(2):026123 (2009). [64] Lancichinetti, A., Kivela, M., Saramaki, J. & Fortunato, S. Characterizing the community structure of complex networks. PLoS ONE, 5(8):e11976 (2010). [65] Bastian, M., Heymann, S. & Jacomy, M. Gephi: An open source software for exploring and manipulating networks, 2009. [66] Cacciola, A., Muscoloni, A., Narula, V., Calamuneri, A., Nigro, S., Mayer, E.A. & et al. Coalescent embedding in the hyperbolic space unsupervisedly discloses the hidden geometry of the brain. arXiv:1705.04192, 2017. [67] Bowditch, B.H. Notes on gromov's hyperbolicity criterion for path-metric spaces. In E. Ghuy, A. Haeflinger, A. Verjovsky (Eds.) Group Theory from a Geometrical Viewpoint, Trieste, 1990, World Scientific, River Edge, NJ, pages 64 -- 167, 1991. [68] Rodriguez, J.M. & Touris, E. Gromov hyperbolicity through decomposition of metric spaces. Acta Math. Hungar., 103:53 -- 84 (2004). [69] Fox, M.D. & Raichle, M.E. Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nature Revw Neurosci, 8:700 -- 711 (2007). Acknowledgments Work supported by the Slovenian Research Agency (research code funding number P1-0044). MA received financial support from the Ministry of Education, Science and Technological Development of the Republic of Serbia, under the project OI 174014. RM is also grateful for the NSERC and CRC programs for their support. 13
1603.05897
1
1603
2016-03-18T15:53:07
Mapping multiplex hubs in human functional brain network
[ "q-bio.NC", "cond-mat.dis-nn", "physics.bio-ph", "physics.med-ph" ]
Typical brain networks consist of many peripheral regions and a few highly central ones, i.e. hubs, playing key functional roles in cerebral inter-regional interactions. Studies have shown that networks, obtained from the analysis of specific frequency components of brain activity, present peculiar architectures with unique profiles of region centrality. However, the identification of hubs in networks built from different frequency bands simultaneously is still a challenging problem, remaining largely unexplored. Here we identify each frequency component with one layer of a multiplex network and face this challenge by exploiting the recent advances in the analysis of multiplex topologies. First, we show that each frequency band carries unique topological information, fundamental to accurately model brain functional networks. We then demonstrate that hubs in the multiplex network, in general different from those ones obtained after discarding or aggregating the measured signals as usual, provide a more accurate map of brain's most important functional regions, allowing to distinguish between healthy and schizophrenic populations better than conventional network approaches.
q-bio.NC
q-bio
Mapping multiplex hubs in human functional brain network 1Departament d'Enginyeria Inform`atica i Matem`atiques, Universitat Rovira i Virgili, 43007 Tarragona, Spain, 2Department of Psychiatry, University of Wisconsin - Madison, Madison, WI, USA Manlio De Domenico1, Shuntaro Sasai2, Alex Arenas1 Typical brain networks consist of many peripheral regions and a few highly central ones, i.e. hubs, playing key functional roles in cerebral inter-regional interactions. Studies have shown that networks, obtained from the analysis of specific frequency components of brain activity, present pe- culiar architectures with unique profiles of region centrality. However, the identification of hubs in networks built from different frequency bands simultaneously is still a challenging problem, remain- ing largely unexplored. Here we identify each frequency component with one layer of a multiplex network and face this challenge by exploiting the recent advances in the analysis of multiplex topolo- gies. First, we show that each frequency band carries unique topological information, fundamental to accurately model brain functional networks. We then demonstrate that hubs in the multiplex network, in general different from those ones obtained after discarding or aggregating the measured signals as usual, provide a more accurate map of brain's most important functional regions, allow- ing to distinguish between healthy and schizophrenic populations better than conventional network approaches. The brain functional network is generally built by in- terconnecting brain regions according to some measure of functional connectivity [1–3]. Studies using functional magnetic resonance imaging [4, 5] (fMRI) provided con- vincing evidences supporting the existence of special re- gions, i.e. hubs, that play a fundamental role in brain functional connectivity [6, 7] by mediating interactions among other regions and favoring the brain's integrated operation. Generally, the strength of this connectivity is empirically estimated by inter-regional correlations cal- culated after post-processing and filtering fMRI signals with a conventional pass band, keeping components be- tween 0.01 and 0.1 Hz [8–10]. The importance of each region with respect to the overall connectivity, i.e. nodal centrality in the functional network, is of particular in- terest in many applications [2, 11–15]. However, it has been shown that networks with unique hub regions can be built from different frequency ranges [16] and that region centrality might largely fluctuate depending on frequency cuts [17], with components above 0.1 Hz also contributing to functional connectivity with unique topological infor- mation [18–23]. Such evidences impel the development of a novel framework to account for full information from all frequency bands separately and simultaneously, with- out discarding any particular component or aggregating some of them to build single networks. In this study, we tackle this challenging issue by employing the theoretical and computational tools re- cently developed for analyzing and modeling multiplex networks [24–27]. Multiplex architectures are special networks consisting of different layers, each encoding a different type of relationship or interaction between nodes [28, 29]. In this context, we identify each frequency component with a distinct layer of a multiplex network whose nodes represent the brain's regions of interest and edges represent their functional connectivity in a specific frequency range. This novel approach arises two fundamental questions, requiring to i) verify if and how brain regions playing the role of hubs in the new multiplex functional net- work differ from the ones obtained using standard net- work approaches; and ii) if and how we can exploit such differences to improve our understanding of brain disor- ders. In the following, we will provide extensive evidences demonstrating that hub regions in multiplex functional networks are different from hub regions in standard func- tional networks and that such differences in the nodal centrality profile allow us to identify patients affected by schizophrenia more accurately than conventional ap- proaches based on discarding or aggregating information about brain functional activity. RESULTS Building aggregated and multiplex functional connectivity networks. We use a publicly available COBRE data set of resting state fMRI, consisting of 71 patients affected by Schizophrenia and 74 healthy con- trols (age: 18–65). The set of 264 regions of interest (ROIs) introduced by Power et al. [30] is used to ex- tract the mean signal within each ROI, for each individ- ual separately. After estimating coherence between all pairs of ROIs, the frequency-specific connectivity matri- ces are obtained by averaging coherence within 12 fre- quency bands, defined by decomposing the frequency range from 0.01 to 0.25 Hz into intervals with equal widths of 0.02 Hz. The upper bound of this frequency range corresponds to the Nyquist frequency of fMRI sig- nals, while the lower bound is obtained by following con- ventional way to eliminate long term drift [9]. Weighted adjacency matrices, defining the functional network for each frequency component separately, are yielded by discarding from frequency-specific connec- tivity matrices those connections with non-significant amount of correlation (see Methods). The resulting multiplex network is obtained, for each individual sep- arately in control and patient groups, by interconnect- 2 the reduction process and its global maxima identify op- timal structural reduction strategies. We perform structural reducibility for each individual separately and calculate the corresponding quality func- tions, for control and patients groups (Fig. 2B). In both cases, we found that the maximum value of the qual- ity function is attained when no reduction is performed at all, providing evidence that the topological informa- tion carried by each functional network, corresponding to a different frequency component, should not be disre- garded from structural analyses. It is worth noting that the behavior of the quality function alone does not allow to distinguish between the two groups of individuals. To gain insights about (dis)similarities between differ- ent layers of the multiplex functional network in the two groups, we use the quantum Jensen-Shannon distance (see Methods) calculated during the structural reducibil- ity analysis. The distance matrix, whose entries provide the Jensen-Shannon distance between any pairs of layers, is first built for each individual separately, and group av- erage µ and standard deviation σ are calculated. The signal-to-noise ratio (SNR) defined by their ratio is suc- cessively calculated for each pair of layers and for each group, separately (Fig. 2C), as well as the relative dif- ference between the two group-averaged values. We ob- served differences of up to 30% in absolute value between the two groups, for specific pairs of layers. Dissimilari- ties between layers within the typical-band were higher in healthy individuals than in schizophrenic patients. On one hand, functional connectivity in healthy subjects is rather volatile and, in general, exhibits topological dif- ferences across individuals [16] that we did not observe in patients, suggesting the possibility that schizophrenia might alter brain's integrated operation to reduce such a functional diversity. On the other hand, an abnor- mal amount of dissimilarity between functional networks corresponding to other frequency bands (such as those within relatively higher ranges, e.g. 0.09-0.19 Hz) was ob- served in patients but not in healthy individuals. These results suggest that the dependence on frequency of pa- tients' functional connectivity is different from that of healthy individuals and we might use such dissimilarity patterns as a fingerprint of brain's functional organiza- tion for each group. Identifying schizophrenic patients by the central- ity profile of their brain functional connectivity. The importance of a region with respect to the overall brain functional connectivity can be quantified by cen- trality descriptors [6, 7, 11–13, 17, 31, 32]. Here, we propose to use PageRank [33, 34] centrality as a mea- sure of centrality, which is based on the rationale that nodes linked by influential nodes are more central than those linked by un-influential nodes. It has been used in several applications, from ranking relevant Web pages in the World Wide Web [33] to identifying important nodes in the human functional connectome [15] and, more re- cently, in a variety of multiplex networks [27]. Once cen- trality scores are calculated for each node, the set of all Figure 1. Schematic illustration of brain multiplex functional network construction. (A) We measure the brain activity with a set of 264 ROIs (here, we only draw five ROIs, for simplicity), and estimate the coherence spec- trum of signals between any pair of ROIs. (B) Averaged co- herence values are calculated in 12 frequency bands (here we only show four bands, for simplicity), to quantify the strength of frequency-specific functional connectivity. The statistical significance of each connection is calculated (see Methods) and connections with Z-score smaller than 3 are discarded. (C) The remaining connections are used to build adjacency matrices, weighted by Z-scores, that constitute the layers of the multiplex functional network once interconnected. (D) Resulting single-layer and multiplex networks obtained from this procedure. ing the layers encoding functional connectivity in each frequency band (Fig. 1A–C). We also define two single- layer networks, obtained by averaging coherence signals within 0.01-0.25 Hz and 0.01-0.1 Hz frequency ranges (Fig. 1D). We refer to such conventional networks as full-band and typical-band single-layer networks, respec- tively, both representing averaged and filtered versions of the full multiplex functional networks. Structural reducibility of the multiplex functional connectivity network. First, we verify if the multiplex network is a valid and suitable model of the underly- ing brain connectivity. For this purpose, we analyze the structural reducibility of a multiplex network [26], al- lowing to identify layers carrying redundant topological information. The method incorporates redundant layers into other ones to reduce the overall structure, while still maximizing the distinguishability between the multiplex network model and the corresponding fully aggregated graph, obtained by summing up the connectivity of all layers (Fig. 2A). The difference between connectivity in different layers is quantified by Jensen-Shannon diver- gence (see Methods), a powerful information-theoretical measure of (dis)similarity. A quality function controls 3 and 50 the number of top features used to discriminate between healthy and schizophrenic individuals. The com- parison between the results obtained from different cen- trality profiles are shown in Fig. 4A (see Supplementary Figs. 1–3 for further details). Remarkably, multiplex centrality profiles allow a more accurate discrimination of the two populations, confirming the hypothesis that multiplex functional networks provide a more suitable model of brain functional connectivity. This result is ro- bust against the selection of the number of features used to discriminate, with the multiplex approach significantly outperforming the other ones. Figure 2. Structural reducibility of the multiplex func- tional network. (A) Schematic illustration of how the anal- ysis structural reducibility of the network works: it allows to identify frequency bands providing redundant topological information and to verify the validity of the multiplex model with respect to conventional single-layer models. Global max- ima in the quality function identify optimal structural re- ductions. (B) The median quality function is shown for healthy control (solid) and schizophrenic patients (dashed), with shaded areas indicating the standard deviation around each value. (C) Signal-to-noise ratio (SNR; see text for fur- ther details) for Jensen-Shannon distance calculated for each pair of layers, color-coded for both groups, and corresponding relative difference between the two groups. their values constitutes the centrality profile of the under- lying functional network. The Spearman's correlation co- efficient between the centrality profiles corresponding to the multiplex, full-band and typical-band functional net- works is calculated (Fig. 3A and 3B). While very strong correlations are observed for centrality profiles calculated from single-layer networks, no significant correlation with multiplex centrality profiles were found. These results suggest the appealing possibility to use the multiplex centrality profiles to gain new insights about brain functional connectivity. To this aim, we interpret the centrality profiles as characteristic fea- tures of each individual (control or patient) and we use the well-known and robust random forest method [35] to train a classifier distinguishing between healthy and schizophrenic individuals (see Methods for further de- tails). At the very beginning, we trained the classifier by using all 264 centrality scores available for each in- dividual and found a classification accuracy of about 60- 65%, regardless for the type of centrality profile used (i.e. multiplex, full-band and typical-band). One of the main advantage of random forest classification is that it also ranks the features based on their classification power, i.e. on the degree of discrimination they have. We cap- italize on this precious information to perform a second round of classification, this time using only top-ranked features instead of the full set. We varied between 10 Figure 3. Comparing centrality profiles of multiplex and conventional functional networks. Spearman's cor- relation coefficient between the centrality profiles obtained from multiplex, full-band and typical-band functional net- works for (A) healthy and (B) patient groups. regions To gain further insights, we focus on the regions corre- sponding to the top 30 ROIs of the multiplex centrality profile, where we attained the maximum discrimination between control and patient groups. The spatial distri- bution of the corresponding brain's regions are shown in Fig. 4B. Anatomical information about these ROIs is summarized in Supplementary Table 1. of Characterizing distinctive schizophrenic brain functional activity. Capi- talizing on results from group discrimination using centrality profiles, we investigate more in detail the role of hub regions. In particular, our interest is twofold. On one hand, we wonder if the most central regions obtained from multiplex and conventional functional networks are the same (it is worth remarking that previous correlation analysis of centrality profiles does not provide this information, because differences might be due to low-ranked regions, for instance). On the other hand we want to clarify which hub regions are found only in healthy individuals, which ones are found only in schizophrenic individuals and which ones are found in both groups. 4 Figure 4. Discrimination performance of the multi- plex functional networks vs. conventional networks. (A) The statistical indicators of the discrimination between control and patient groups obtained from conventional net- work approaches (i.e. full-band and typical-band networks) are compared against the full multiplex functional network, which provides better overall discrimination. Note that the features are ROIs and their values are centrality scores. (B) Location of top 30 discriminating ROIs, obtained from mul- tiplex analysis. Hubs were identified as ROIs ranked in top 5% in terms of group-averaged region centrality. Figure 5 shows the spatial distributions in the brain of such hubs, for each group, while the corresponding anatomical information is reported in Supplementary Table 2. In all cases we found hub regions peculiar for each group and hubs re- gions that are common to both groups. While significant differences are not observed between networks built from conventional approaches, hubs from multiplex analysis constitute a distinct set. In both conventional networks, healthy-specific hubs are located in medial superior frontal, lateral frontal cor- tices and thalamus, whereas schizophrenic-specific hubs are localized in posterior parts of the brain, such as cuneus, precuneus, and superior temporal cortices. Hubs shared by both groups are identified along the midline of the brain, in particular the medial superior frontal, precuneus and cingulate cortices. In the multiplex net- work, healthy-specific hubs controls are located in ante- rior cingulate, superior frontal, insula and superior tem- poral cortices, whereas those pertaining to schizophrenic patients are distributed over frontal, parietal and occip- ital cortices. Hub regions shared by both groups groups are localized in frontal, occipital cortices and cerebellum. Notably, no hub region has been identified in the pre- cuneus cortex, a region well known to function as a hub in healthy individuals [31]. DISCUSSION Resting state functional connectivity has been widely investigated with fMRI in the past two decades. Since the first study conducted by Biswal et al. [36], functional Figure 5. Brain regions playing the role of hubs in functional connectivity. The most central regions, i.e. hubs, identified in multiplex and conventional functional net- works are shown (from top to bottom). Markers indicate their locations, whereas panels from left-hand to right-hand side show hubs found only in healthy controls (left), only in schizophrenic patients (center), or in both (right). connectivity has been defined as an inter-regional tempo- ral correlation of fMRI signals that are preprocessed with band-pass filters, removing frequency components below 0.01 and above 0.1 Hz. In fact, the power spectrum of spontaneous fluctuations of fMRI signals roughly follows a 1/f power-law scaling [37], where powers in the higher frequency range are relatively weaker than lower ones, suggesting the hypothesis that only the lower frequency range substantially contributes to brain's function. How- ever, recent studies have reported that conventionally ex- cluded frequency bands might provide additional insights on brain activity [16–18, 22, 23]. As a consequence, brain functional networks exhibit a peculiar architecture, con- sisting of a few regions acting as hubs, strongly dependent on the frequency components of brain activity that con- tribute to inter-regional interactions. However, a rigor- ous method to identify such hubs in networks built from different frequency bands simultaneously is a challenging problem remaining largely unexplored. Our results, based on multiplex modeling and anal- ysis of the brain activity, provide convincing evidence that characterization of brain functional networks can not prescind from considering the whole information ob- served from different frequency bands, simultaneously. This crucial finding allows to exploit new theoretical and computational tools for the analysis of brain activity and opens a new direction towards a deeper understanding of brain function and its operated integration. As a first hint of the power of the new methodology, we have shown that multiplex characterization of brain regions, in terms of network centrality, allows to find new areas of the brain that have never been classified as relevant in brain's functional integration (or the opposite). This is the case of ROIs in the precuneus cortex, a well-known region of highly central functional hubs [31], that are not found by our multiplex network analysis, reflecting the importance of considering the whole information si- multaneously, rather than aggregating or neglecting part of it [27]. We wondered if this result could be exploited for prac- tical applications, where the choice of specific frequency bands might play a crucial role. We focused our atten- tion on characterizing brain disorders in schizophrenic patients, a research topic of great interest that has been largely explored [12, 13, 38], although individual diagno- sis based on brain imaging remains still undeveloped [32]. With the aid of the MRI technique, it has been recently shown that regions affected by schizophrenia are dis- tributed across the brain [39, 40], impelling researchers to move from the conventional perspective where the causes of disorders are localized in specific areas, to a wider perspective with emphasis on abnormality in brain structural and functional connectivity [41]. Stud- ies on structural connectivity provided evidences that schizophrenic brains exhibit abnormal network architec- ture, characterized by reduced hierarchical organization, the loss of frontal hubs with emergence of non-frontal hubs [12, 13] and degraded rich-club organization [38]. Methods not based on networks were able to provide sat- isfactory performance in discriminating schizophrenic pa- tients from the analysis of their brain activity [42, 43], al- though they are often based on very complicated machine learning algorithms and make use of heterogenous data sources, thus not improving our understanding of brain function. Here, we have found that multiplex central- ity profile of brain regions allow to discriminate between control and schizophrenic groups of individuals more accurately than centrality profiles calculated from net- works obtained by using conventional approaches, such as aggregating and/or disregarding the measured activ- ity. Nevertheless, the discrimination accuracy is compa- rable to other methods, with the additional advantage of providing a framework facilitating the interpretation of results, without relying on external data sources or phenotypic information. In fact, we were able to identify many regions distinctive of schizophrenic brains, some of them localized where abnormality has been previously suggested [32, 44]. The analysis of dissimilarities between networks corresponding to different layers of the multi- plex functional network, confirmed significant differences between healthy and schizophrenic individuals in specific frequency ranges, including the higher ones. This find- ing demonstrates that brain activity in higher frequen- cies provides unique information about functional inter- action in the brain, even if their amplitudes are under- represented in the power spectrum. The proposed methodology suggests a guideline for fu- ture studies designed to consider brain's inter-regional 5 interactions at different frequencies, encouraging the ap- plication of other multiplex network measures to func- tional networks obtained, from variable brain states. for instance, ACKNOWLEDGMENTS M.D.D. acknowledges financial support from the Span- ish program Juan de la Cierva (IJCI-2014-20225). SS was supported by JSPS Postdoctoral Fellowships for Re- search Abroad. A.A. acknowledges financial support from ICREA Academia and James S. McDonnell Foun- dation and Spanish MINECO FIS2015-71582. METHODS Overview of the data set and fMRI preprocessing. The publicly available MR data set contributed by The Center for Biomedical Research Excellence (COBRE) was used in this study. The data set was downloaded from the following repository: http://fcon_1000. projects.nitrc.org/indi/retro/cobre.html. It in- cludes functional and anatomical MRI data acquired from 72 Schizophrenic patients and 75 healthy controls (age: 18–65 for both groups). One patient?s data was discarded from all analyses due to the shortness of the data length. The following pre-processing steps were applied to functional MR images by using the SPM8 package (Wellcome Department of Imaging Neu- roscience, London, UK): motion-correction, slice-timing- correction, spatial smoothing with Gaussian kernel (5- mm full-width-at-half-maximum) and spatial normaliza- tion. Signal fluctuations of fMRI are driven by not only neural but also physiological effects – such as respira- tion and cardiac pulsation – and environmental condi- tions – such as scanner instabilities and subject motion. These nuisance effects can be canceled out by discard- ing, for instance, the signal from the ROI centered in the white matter, the signal from the ventricular ROI, and the signal from the ROI located within the soft-tissue. We have linearly removed these components after tem- porally shifting them by optimal time-lags yielding the highest correlation with the averaged signal of all gray matter voxels [45]. Connectivity matrices. A set of 264 spherical ROIs (5 mm radii) was used to extract the mean signal within each ROI. For each individual, the coherence between all pairs of in-ROI averaged signals was estimated in spe- cific frequency bands, as described in the text. We kept the edges between pairs of ROIs whose weight was sig- nificantly different from a null model where observed sig- nals were replaced by surrogates. More specifically, we used the well-known iterative amplitude-adjusted Fourier transform (IAAFT) algorithm to build surrogate time se- ries preserving the power spectrum and the probability density of the original ones, while removing higher-order self-correlations. For each pair of ROIs i and j, we have verified that the distribution of the weights obtained from the null model corresponds to a Gaussian described by sample mean µij and variance σ2 ij. Let wij indicate the weight obtained from empirical data: we have calculated the absolute Z score as zij = wij−µij/σij and discarded all those edges for which zij < 3, corresponding to cross- coherence not statistically significant. We used the values zij as entries of the resulting connectivity matrix. Multiplex network model. A multilayer network al- lows to encode different types of interactions or relation- ships among a set of nodes. More specifically, in the case of our study we make use of multiplex networks to model functional connectivity. In a multiplex network, the links are of different type: one can assign a different "color" to each type, thus obtained an edge-colored representa- tion of the network. In this type of architectures, nodes exist in one or more layers, i.e. it is not required that all nodes exist in all layers. Correlation networks, as the ones used in this study, define edge-colored graphs where each layer encodes the correlations observed in a specific frequency band. However, it has been shown that by interconnecting nodes with their replicas across layers, the resulting interconnected multiplex network can be described by an adjacency tensor [25] with components M iα jβ , an object generalizing the well-known concept of adjacency matrix to higher orders, encoding connections between node i in layer α and node j in layer β. For in- jβ = 0 for i (cid:54)= j and terconnected multiplex networks, M iα simultaneously α (cid:54)= β. The presence of interconnections allows to exploit tensorial algebra to generalize many single-layer network descriptors, from centrality [27] to mesoscale structure [24, 46]. However, it is not always possible to assign a weight to inter-layer links by using the data, and it is common to parameterize the intensity of interconnections [47, 48], i.e. M iα jβ = D for i = j and simultaneously α (cid:54)= β, to study the resulting intercon- nected multiplex network as a function of this parameter D. This is exactly the case of the present study, where the choice of D depends on the analysis of interest. Structural reducibility of brain multiplex func- tional network. The analysis of structural reducibility of a multilayer network allows to find layers that pro- vide redundant topological information, suggesting how to merge some layers with other ones, to obtain an opti- mal multilayer network [26]. The whole procedure can be summarized as follows: i) compute the distance (based on quantum Jensen-Shannon divergence) between all pairs of layers; ii) perform hierarchical clustering of layers us- ing such distance matrix and use changes in the relative entropy q(•) as the quality function for the resulting par- tition; iii) finally, choose the partition which maximizes the quality function, i.e. the distinguishability from the fully aggregated graph obtained by summing up the ad- jacency matrices of all layers. It is worth remarking that this analysis is independent on the choice of interconnec- tions weight, i.e. it does not depend on D. Here, we do not enter into the details of the whole method; instead we focus on the Jensen-Shannon distance, that is a key measure for two analyses presented in this study. 6 The components A[α] ij (i, j = 1, 2, ..., N ; being N the number of ROIs in this study) of the adjacency matrix A[α] – encoding layer α – are obtained from the compo- nents of the multilayer adjacency tensor as A[α] ij = M iα jα. Here, A[α] ij > 0 if there is correlation between ROIs i and j in the frequency band represented by α. The Von Neumann entropy [49, 50] of the corresponding complex network is defined by hA[α] = −Tr (1) where L[α] = c×(S[α]−A[α]) is the combinatorial Lapla- , (cid:104)L[α] log2 L[α](cid:105) (cid:32) N(cid:80) (cid:33) A[α] ij i,j=1 cian rescaled by c = 1/ , and S is the diago- nal matrix of the strengths of the nodes. From the eigen- decomposition of the Laplacian, it is possible to show that the entropy can be calculated by hA[α] = − N(cid:88) i=1 λ[α] i log2(λ[α] i ), (2) where {λ[α] 1 , λ[α] 2 , . . . , λ[α] N } are the eigenvalues of L[α]. The similarity of two layers can be calculated in terms of differences in their entropy. Given two rescaled Lapla- cian matrices L[α] and L[β], it is possible to quantify to which extent layer α is different from layer β by their Kullback-Liebler divergence, defined by DKL(L[α]L[β]) = Tr[L[α](log2(L[α])−log2(L[β]))], (3) encoding the information gained about L[β] when the expectation is based only on L[α]. This divergence is not a metric and a more suitable dissimilarity measure is the Jensen-Shannon divergence, defined by DJS(L[α]L[β]) = DKL(L[α]L[µ]) + 1 2 2 (L[α] +L[β]). It can be shown that where L[µ] = 1 usually called Jensen–Shannon distance, takes values in [0, 1], satisfies all the properties of a metric distance and provides a very powerful measure of dissimilarity between layers. Random forest classification. Machine learning has been used to train a classifier to distinguish between con- trol and schizophrenic individuals. We used the random forest classifier [35], well-known for its robustness and for facilitating the interpretation of results. We have fixed to 5 the maximum number of terminal nodes trees the forest can have and to 2 the number of variables randomly sam- pled as candidates at each split. We have verified that forests consisting of 700 trees where enough to reach sta- ble results within this setup. Given the importance of interconnections weight for our analysis and, at the same time, the lack of knowledge 1 2 DKL(L[β]L[µ]), √DJS, (4) 7 dom walker from a node i in layer α to one of its neighbors j in layer β, L is the total number of layers and uiα jβ is the rank-4 tensor with all components equal to 1. The steady-state solution of the master equation πjβ(t + 1) = Riα jβπiα(t), (6) N(cid:88) L(cid:88) i=1 α=1 obtained in the limit t −→ ∞, provides the PageR- ank centrality for interconnected multiplex networks. To compute the overall PageRank of a node, accounting for the whole interconnected topology, we can safely sum up the stationary probabilities π(cid:63) jβ over the layers, to obtain the components of the centrality profile vector π(cid:63) j = π(cid:63) jβ used in our analysis. It is worth remarking L(cid:80) β=1 that the interconnection weight used for this purpose is D = 24.7708, the one yielding the highest classification accuracy. Supplementary Figure 1. Maximizing classification accuracy. Layers in the multiplex functional network are interconnected to allow the calculation of layer's and node's properties. However, the weight of such inter-layer links can not be deduced from the data. At the same time, not all ROIs have enough discriminant power and using sub-sets of them reduces the noise and generally improves the classifi- cation. We have chosen the value of the inter-layer strength (D = 24.7708) and the size of an appropriate sub-set of ROIs (30) where maximum classification accuracy is achieved (see Methods for details). about its value, we used random forest to learn also which value of D would be more suitable for calculations. We have performed a first exploratory classification us- ing a leave-one-out approach to maximize the amount of data used for training the classifier. The result of each classification, corresponding to exactly one different indi- vidual (without replacement) left out, was accompanied by the importance assigned by the classifier to each ROI in terms of mean decrease in its Gini index. Therefore, for each individual and each value of D, we have ranked the ROIs according to this measure and, eventually, summed up the ranks corresponding to all classifications. The result of the exploratory classification was an over- all ranking suggesting which ROIs, in general, have been more crucial than others in the classification process. Therefore, we performed a second classification round by using only the top ROIs according to the above rank- ing. We first varied the number of kept features and the value of D, to find the values with best classification performances in terms of accuracy (see Supplementary Fig. 1). The numerical analysis indicated that the best classification is achieved for interconnections weight close to 24.7708 and about 30 top ROIs: that value of D and that sub-set of ROIs have been used for analysis reported in the text. Using a similar approach, we have compared the best performance obtained from the full multiplex functional network (12 layers) against multiplex functional networks obtained by keeping layers in the typical band (Supple- mentary Fig. 2) and against classifier trained by includ- ing phenotypic data (Supplementary Fig. 3). In all cases, the classification obtained using the full multiplex func- tional network was equal or better than the other ones. ROIs PageRank centrality. PageRank is a mea- sure of node's centrality originally introduced by Google founders to rank Web pages according to their impor- tance in the Internet [33, 34]. The algorithm consider a random walker exploring the network with the follow- ing rules: 85% of times the walker jumps from the cur- rent node to one chosen with uniform probability from the neighborhood, whereas 15% of times the walker is allowed to jump to any node of the network, with uni- form probability. The stationary probability of finding the walker in a specific node is then used to rank the importance of nodes in the network, the rationale being that central nodes have high number of incoming links from other important nodes. The natural extension of the PageRank algorithm to the context of multiplex networks has been recently intro- duced [27] and proved to perform better than its single- layer counterpart in some applications. Let us indicate with Riα jβ the transition tensor, governing the dynamics of a random walker jumping to neighboring nodes with rate 0.85 and teleporting to any other node in the network with rate 0.15. This rank-4 tensor is given by jβ = 0.85 × T iα Riα jβ + 0.15 N L uiα jβ, (5) where T iα jβ governs the standard moves of a classical ran- 8 Supplementary Figure 2. Discrimination performance of two different multiplex functional networks. A new multiplex network, consisting only of the layers correspond- ing to the typical frequency range, has been built and used for discriminating between control and patient groups. The sta- tistical indicators of the discrimination are compared against the full multiplex functional network discussed in the text, providing better overall discrimination. Bars indicate stan- dard errors. Supplementary Figure 3. Discrimination performance of the multiplex functional networks with and with- out phenotypic information. The statistical indicators of the discrimination between control and patient groups ob- tained from the full multiplex functional network before and after including phenotypic information in the machine learn- ing process. The results are not significantly different, indi- cating that phenotypic data is redundant in this case and does not improve discrimination. Bars indicate standard errors. multiplex full-band common typical-band healthy Supplementary Table 1. 9 Network model Group MNI Talairach coordinate Anatomical label Transverse Temporal Gyrus healthy Superior Frontal Gyrus Superior Frontal Gyrus Superior Frontal Gyrus coordinate X Y Z X Y Z 53 -17 9 58 -16 7 29 -27 14 Claustrum 32 -26 13 -4 37 21 Anterior Cingulate -3 42 16 -1 25 30 Cingulate Gyrus 0 30 27 31 14 -2 Claustrum 34 16 -8 -27 46 26 -28 52 21 32 14 56 29 7 55 Middle Frontal Gyrus -11 40 14 Medial Frontal Gyrus -11 45 8 -14 -23 69 Precentral Gyrus -13 -17 75 -17 -11 67 schizophrenia -16 -5 71 21 -85 19 Cuneus 24 -87 24 16 -14 61 19 -8 64 -4 20 45 Medial Frontal Gyrus -3 26 44 46 -59 4 42 -57 3 Middle Temporal Gyrus -8 -26 60 Medial Frontal Gyrus -7 -21 65 34 38 -12 31 35 -4 Inferior Frontal Gyrus -47 -76 -10 -45 -72 -12 Fusiform Gyrus -37 -29 -26 -36 -27 -22 Parahippocampal Gyrus 11 -54 17 17 -80 -34 15 -74 -33 Uvula -50 -7 -39 -48 -5 -32 Inferior Temporal Gyrus -32 -1 54 -60 -25 14 -7 -52 61 58 -16 7 -72 24 6 30 27 0 -16 -77 34 -3 -81 21 10 -62 61 -31 -6 52 Precentral Gyrus -57 -26 14 -8 -55 54 Precuneus 53 -17 9 4 -1 25 30 Cingulate Gyrus -16 -76 28 Cuneus -4 -79 16 Cuneus 8 -11 -19 8 -20 57 25 23 43 32 Superior Parietal Lobule Thalamus Superior Frontal Gyrus Superior Frontal Gyrus -53 15 Posterior Cingulate schizophrenia -10 -18 7 -71 20 Cuneus common healthy -64 54 9 Superior Temporal Gyrus Transverse Temporal Gyrus Transverse Temporal Gyrus Superior Temporal Gyrus Insula Thalamus Inferior Parietal Lobule -50 46 Precuneus -20 64 19 26 50 27 27 -97 -13 24 -92 -15 Fusiform Gyrus -55 -9 12 36 22 3 9 -4 6 47 -30 49 -15 47 0 -3 2 53 -48 51 4 -3 26 44 -1 15 44 58 -16 7 -60 -25 14 6 -72 24 11 -39 50 10 -62 61 8 -72 11 32 -26 13 -53 -10 13 Precentral Gyrus 33 19 8 7 -6 9 42 -33 46 -1 -19 45 Paracentral Lobule -4 -3 51 Medial Frontal Gyrus 2 -4 20 45 Medial Frontal Gyrus -2 10 44 Medial Frontal Gyrus 53 -17 9 -57 -26 14 4 9 8 6 29 -27 14 Claustrum -11 -19 8 Thalamus -6 13 35 Cingulate Gyrus 23 43 32 5 2 3 17 39 Cingulate Gyrus -1 25 30 Cingulate Gyrus -53 -10 13 Precentral Gyrus -8 -55 54 Precuneus -1 -19 45 Paracentral Lobule -4 -3 51 Medial Frontal Gyrus 2 -4 20 45 Medial Frontal Gyrus -2 10 44 Medial Frontal Gyrus -5 18 34 26 50 27 8 7 51 5 23 37 0 30 27 -55 -9 12 -7 -52 61 -15 47 0 -3 2 53 4 -48 51 -3 26 44 -1 15 44 -71 20 Cuneus -42 46 Precuneus -64 54 -70 8 50 Medial Frontal Gyrus -50 46 Precuneus Superior Frontal Gyrus Superior Parietal Lobule Cuneus schizophrenia -10 -18 7 common Supplementary Table 2. 10 MNI -24 0 coordinate X Y Z 27 -97 -13 -10 -18 7 52 -34 -27 36 22 3 -25 -98 -12 65 -24 -19 -40 -19 54 6 33 -12 -34 24 32 -18 34 54 -13 55 -31 -17 0 -15 47 66 -8 25 -45 0 -56 -45 -24 -55 -9 12 25 -58 60 -21 41 -20 -31 -10 -36 9 13 55 38 -20 64 19 35 -67 -34 49 -3 -38 -37 -29 -26 24 45 -15 -1 15 44 -18 63 -9 53 33 1 54 3 9 Thalamus Insula -24 2 Thalamus Talairach coordinate Anatomical label X Y Z 24 -92 -15 Fusiform Gyrus -11 -19 8 48 -32 -23 Fusiform Gyrus 33 19 8 -25 -92 -15 Fusiform Gyrus 60 -23 -15 Middle Temporal Gyrus -39 -23 50 Postcentral Gyrus 5 30 -10 -27 Uncus 22 30 -10 Sub-Gyral 31 50 -4 Middle Frontal Gyrus 50 -30 -13 Inferior Temporal Gyrus -1 -19 45 Paracentral Lobule 61 -11 26 Precentral Gyrus -43 -2 11 -53 -42 -22 Fusiform Gyrus -53 -10 13 Precentral Gyrus 22 -60 53 -20 39 -11 Middle Frontal Gyrus -30 -8 -30 Uncus 8 11 47 42 -20 57 25 32 -62 -32 Cerebellar Tonsil 45 -2 -30 Middle Temporal Gyrus -36 -27 -22 Parahippocampal Gyrus 22 42 -6 Medial Frontal Gyrus -2 10 44 Medial Frontal Gyrus -18 59 0 Medial Frontal Gyrus 49 29 8 Inferior Frontal Gyrus Superior Frontal Gyrus Superior Frontal Gyrus 49 10 Medial Frontal Gyrus Superior Parietal Lobule Insula 11 [1] D. S. Bassett and E. Bullmore, The neuroscientist 12, Nature communications 6, 6864 (2015). 512 (2006). [2] E. Bullmore and O. Sporns, Nature Reviews Neuro- science 10, 186 (2009). [3] E. Bullmore and O. Sporns, Nature Reviews Neuro- science 13, 336 (2012). [4] M. P. Van Den Heuvel and H. E. H. Pol, European Neu- ropsychopharmacology 20, 519 (2010). [5] R. A. Poldrack and M. J. Farah, Nature 526, 371 (2015). [6] S. Achard, R. Salvador, B. Whitcher, J. Suckling, and E. Bullmore, The Journal of Neuroscience 26, 63 (2006). [7] J. D. Power, B. L. Schlaggar, C. N. Lessov-Schlaggar, and S. E. Petersen, Neuron 79, 798 (2013). [8] D. Cordes, V. M. Haughton, K. Arfanakis, J. D. Carew, P. A. Turski, C. H. Moritz, M. A. Quigley, and M. E. Meyerand, American Journal of Neuroradiology 22, 1326 (2001). [9] D. Cordes, V. Haughton, J. D. Carew, K. Arfanakis, and K. Maravilla, Magnetic resonance imaging 20, 305 (2002). [10] M. D. Fox and M. E. Raichle, Nature Reviews Neuro- [27] M. De Domenico, A. Sol´e-Ribalta, E. Omodei, S. G´omez, and A. Arenas, Nature Communications 6, 6868 (2015). [28] M. Kivela, A. Arenas, M. Barthelemy, J. P. Gleeson, Y. Moreno, and M. A. Porter, Journal of Complex Net- works 2, 203 (2014). [29] S. Boccaletti, G. Bianconi, R. Criado, C. Del Genio, J. G´omez-Gardenes, M. Romance, I. Sendina-Nadal, Z. Wang, and M. Zanin, Physics Reports 544, 1 (2014). [30] J. D. Power, A. L. Cohen, S. M. Nelson, G. S. Wig, K. A. Barnes, J. A. Church, A. C. Vogel, T. O. Laumann, F. M. Miezin, B. L. Schlaggar, et al., Neuron 72, 665 (2011). [31] M. P. van den Heuvel and O. Sporns, Trends in cognitive sciences 17, 683 (2013). [32] M. Rubinov et al., Dialogues in clinical neuroscience 15, 339 (2013). [33] S. Brin and L. Page, in Proceedings of the Seventh International Conference on World Wide Web 7 (Elsevier Science Publishers B. V., Amsterdam, The Netherlands, The Netherlands, 1998), WWW7, pp. 107–117. science 8, 700 (2007). [34] L. Ermann, K. M. Frahm, and D. L. Shepelyansky, Rev. [11] O. Sporns, C. J. Honey, and R. Kotter, PloS one 2, e1049 Mod. Phys. 87, 1261 (2015). (2007). [12] D. S. Bassett, E. Bullmore, B. A. Verchinski, V. S. Mat- tay, D. R. Weinberger, and A. Meyer-Lindenberg, The Journal of Neuroscience 28, 9239 (2008). [13] M.-E. Lynall, D. S. Bassett, R. Kerwin, P. J. McKenna, M. Kitzbichler, U. Muller, and E. Bullmore, The Journal of Neuroscience 30, 9477 (2010). [14] M. Rubinov and O. Sporns, Neuroimage 52, 1059 (2010). [15] X.-N. Zuo, R. Ehmke, M. Mennes, D. Imperati, F. X. Castellanos, O. Sporns, and M. P. Milham, Cerebral cor- tex 22, 1862 (2012). [16] S. Sasai, F. Homae, H. Watanabe, A. Sasaki, H. Tanabe, N. Sadato, and G. Taga, Frontiers in human neuroscience 8, 1022 (2014). [35] L. Breiman, Machine learning 45, 5 (2001). [36] B. Biswal, F. Z. Yetkin, V. M. Haughton, and J. S. Hyde, Magnetic resonance in medicine 34, 537 (1995). [37] B. J. He, The Journal of neuroscience 31, 13786 (2011). [38] M. P. van den Heuvel, O. Sporns, G. Collin, T. Scheewe, R. C. Mandl, W. Cahn, J. Goni, H. E. H. Pol, and R. S. Kahn, JAMA psychiatry 70, 783 (2013). [39] D. C. Glahn, A. R. Laird, I. Ellison-Wright, S. M. Thelen, J. L. Robinson, J. L. Lancaster, E. Bullmore, and P. T. Fox, Biological psychiatry 64, 774 (2008). [40] I. Ellison-Wright and E. Bullmore, Schizophrenia re- search 108, 3 (2009). [41] M. P. van den Heuvel and A. Fornito, Neuropsychology review 24, 32 (2014). [17] W. H. Thompson and P. Fransson, NeuroImage 121, 227 [42] H. Yang, J. Liu, J. Sui, G. Pearlson, and V. D. Calhoun, (2015). Frontiers in human neuroscience 4 (2010). [18] D. S. Bassett, A. Meyer-Lindenberg, S. Achard, T. Duke, [43] D. Chyzhyk, A. Savio, and M. Grana, Neural Networks and E. Bullmore, PNAS 103, 19518 (2006). 68, 23 (2015). [19] D. Mantini, M. G. Perrucci, C. Del Gratta, G. L. Romani, and M. Corbetta, Proceedings of the National Academy of Sciences 104, 13170 (2007). [20] K. Supekar, V. Menon, D. Rubin, M. Musen, M. D. Gre- icius, et al., PLoS Comput Biol 4, e1000100 (2008). [21] M. Chavez, M. Valencia, V. Navarro, V. Latora, and J. Martinerie, Physical Review Letters 104, 118701 (2010). [22] X.-H. Liao, M.-R. Xia, T. Xu, Z.-J. Dai, X.-Y. Cao, H.-J. Niu, X.-N. Zuo, Y.-F. Zang, and Y. He, Neuroimage 83, 969 (2013). [44] R. Honea, T. J. Crow, D. Passingham, and C. E. Mackay, American Journal of Psychiatry 162, 2233 (2005). [45] J. S. Anderson, T. J. Druzgal, M. Lopez-Larson, E.-K. Jeong, K. Desai, and D. Yurgelun-Todd, Human brain mapping 32, 919 (2011). [46] M. De Domenico, A. Lancichinetti, A. Arenas, and M. Rosvall, Physical Review X 5, 011027 (2015). [47] S. Gomez, A. Diaz-Guilera, J. Gomez-Gardenes, C. J. Perez-Vicente, Y. Moreno, and A. Arenas, Physical re- view letters 110, 028701 (2013). [48] M. De Domenico, A. Sol´e-Ribalta, S. G´omez, and A. Are- [23] J. E. Chen and G. H. Glover, NeuroImage 107, 207 nas, PNAS 111, 8351 (2014). (2015). [49] S. L. Braunstein, S. Ghosh, and S. Severini, Annals of [24] P. J. Mucha, T. Richardson, K. Macon, M. A. Porter, Combinatorics 10, 291 (2006). and J.-P. Onnela, Science 328, 876 (2010). [25] M. De Domenico, A. Sol`e-Ribalta, E. Cozzo, M. Kivela, Y. Moreno, M. A. Porter, S. G`omez, and A. Arenas, Phys. Rev. X 3, 041022 (2013). [26] M. De Domenico, V. Nicosia, A. Arenas, and V. Latora, [50] F. Passerini and S. Severini, Developments in Intelli- gent Agent Technologies and Multi-Agent Systems: Con- cepts and Applications: Concepts and Applications p. 66 (2010).
1505.00774
1
1505
2015-05-02T11:54:31
Modelling Microtubules in the Brain as n-qudit Quantum Hopfield Network and Beyond
[ "q-bio.NC", "quant-ph" ]
The scientific approach to understand the nature of consciousness revolves around the study of human brain. Neurobiological studies that compare the nervous system of different species have accorded highest place to the humans on account of various factors that include a highly developed cortical area comprising of approximately 100 billion neurons, that are intrinsically connected to form a highly complex network. Quantum theories of consciousness are based on mathematical abstraction and Penrose-Hameroff Orch-OR Theory is one of the most promising ones. Inspired by Penrose-Hameroff Orch-OR Theory, Behrman et. al. (Behrman, 2006) have simulated a quantum Hopfield neural network with the structure of a microtubule. They have used an extremely simplified model of the tubulin dimers with each dimer represented simply as a qubit, a single quantum two-state system. The extension of this model to n-dimensional quantum states, or n-qudits presented in this work holds considerable promise for even higher mathematical abstraction in modelling consciousness systems.
q-bio.NC
q-bio
Modelling Microtubules in the Brain as n-qudit Quantum Hopfield Network and Beyond Dayal Pyari Srivastava1, Vishal Sahni2, Prem Saran Satsangi3 ABSTRACT The scientific approach to understand the nature of consciousness revolves around the study of human brain. Neurobiological studies that compare the nervous system of different species have accorded highest place to the humans on account of various factors that include a highly developed cortical area comprising of approximately 100 billion neurons, that are intrinsically connected to form a highly complex network. Quantum theories of consciousness are based on mathematical abstraction and Penrose-Hameroff Orch-OR Theory is one of the most promising ones. Inspired by Penrose-Hameroff Orch-OR Theory, Behrman et. al. (Behrman, 2006) have simulated a quantum Hopfield neural network with the structure of a microtubule. They have used an extremely simplified model of the tubulin dimers with each dimer represented simply as a qubit, a single quantum two-state system. The extension of this model to n-dimensional quantum states, or n-qudits presented in this work holds considerable promise for even higher mathematical abstraction in modelling consciousness systems. Keywords : Quantum Hopfield Network, Qudits, Contextuality, Power Laws 1. Microtubules as Quantum Information / Computation Processing Devices : Penrose-Hameroff Orchestrated Objective Reduction (Orch. OR Theory) Penrose and Hameroff (2011) describe microtubules as self-assembling polymers of the peanut- shaped (4 nm x 8 nm x 5 nm) protein dimer tubulin, each tubulin hetero dimer molecule (110,000 atomic mass units) being composed of an alpha and beta monomer, and is a polar molecule with its positive end near the β-subunit. Typically, thirteen linear tubulin chains (“protofilaments”) align side-to-side to form hollow microtubule cylinders (25 nm diameter) with two types of hexagonal lattices A and B. The hollow microtubule cylinders of about 25 nm in diameter range from 200 nm to 25 micron in length. Structure of a microtubule and arrangement of its tubulins is given in Figure 1. Since there are at least 32 states, i.e., 5 bits (or 5/8 byte) of information per tubulin dumer molecule and 13 dimers (of 8 nm length) per each ring of microtubule-cylinder with 1250 rings per midsize (1-micron long) microtubule, the resulting information storage capacity is approximately 10 kilobytes per microtubule. Given 10,000 microtubules (or 100 million tubulin-dimers) per neuron, it represents 100 megabytes of processing power per neuron which translates into a total information storage capacity of human brain (with 100 billion neurons) of 10 exabytes (i.e. of the order of at least 1 exabyte = 1018 bytes). 1 Department of Physics and Computer Science & Quantum-Nano Systems Centre, Dayalbagh Educational Institute, Dayalbagh, Agra, India 2 Quantum-Nano Systems Centre & Centre for Consciousness Studies, Dayalbagh Educational Institute, Dayalbagh, Agra, India (email : [email protected]) 3 Chairman, Advisory Committee on Education, Dayalbagh Educational Institutions, Dayalbagh, Agra, India and Member, Editorial Board, International Journal of General Systems since inception. According to Penrose-Hameroff Orch. OR (Orchestrated Objective Reduction) theory, each tubulin molecule can exist as quantum superposition (i.e. quantum bit or qubit) of both states (black and white) coupled to London force dipole in hydrophobic pocket. Furthermore the A- lattice has multiple winding patterns which intersect on protofilaments at specific intervals matching the Fibonacci series found widely in nature and possessing a helical symmetry enabling topological quantum computing. Figure 1 Structure of a Microtubule and the arrangement of its tubulins Penrose and Hameroff (2011) estimate approximately 108 tubulins in each neuron which switch and oscillate in the range of 107 per second. This gives an information capacity as the single-cell value at the microtubule level of 1015 operations per second per neuron. The total brain capacity (1011 neurons, 103 synapses per neuron, 102 transmissions per synapse per second) would thus potentially translate at the microtubule level as 1026 operations per second in comparison to the earlier estimates of AI community for the information processing capacity of the entire brain of 1016operations per second at the level of neurons, synapses and their transmission-rate per second. If each tubulin dimer functions as a quantum bit and not a classical bit processor, the computational power becomes almost unimaginably vast. It has been claimed that as few as 300 quantum bits (qubits), have the same computational power as a hypothetical classical computer comprised of as many processing units as there are particles in the universe (Steane and Rieffel, 2000). Penrose-Hameroff Orchestrated Objective Reduction (Orch-OR) has three parts : the Gödel Part, the Gravity Part, and the Microtubule Part. The Gödel Part is an effort to use the famous Gödel’s Incompleteness Theorem [Any finite set of rules that encompass the rules of arithmetic is either inconsistent or incomplete; it entails either statements that can be proved to be both true and false, or statements that cannot be proved to be S E≈ G 2h π, such that the quantity S is given by : either true or false] to prove that human beings have intellectual powers of “non-computable” thought and understanding that they could not have if they functioned in accordance with the principles of classical physical theory. Proving this would reaffirm a conclusion of the von Neumann formulation of quantum theory, namely, that a conscious human being can behave in ways that a classical mechanical model cannot. [However, the Gödel Part cannot now be regarded as having been established successfully]. The Gravity Part addresses a key question pertaining to quantum dynamics : Exactly when do the sudden quantum jumps occur? The Diösi-Penrose (DP) expectation is that Objective Reduction (OR) occurs when the overall separation (the product of the temporal separation T with the spatial separation S) reaches a critical amount given by the Planck-Dirac constant =(cid:61) Planck constant (the gravitational self-energy) of the difference between the mass distributions of the two superposed state and T Z= (the life-time) GE= (cid:61) . Penrose-Hameroff (PH) theory proposes that this time interval is the duration of time for which Nature will endure this bifurcation of space-time structure into the two, incomputible parts, before jumping to one or the other of these two forms. This conjectured rule is based on two very general features of Nature : Planck’s universal constant of action h and Newton- Einstein universal law of gravitation. It invokes quantum gravity attempting to combine quantum theory with Einstein’s theory of gravity, namely General Relativity. Moreover, according to Orchestrated Objective Reduction (Orch OR), this is accompanied by an element of proto-consciousness. The best known temporal correlate for consciousness is gamma synchrony, EEG / MEG, 30-90 Hz, often referred to as coherent 40 Hz, representing a succession of 40 or so conscious moments per second (τ = 25 milliseconds). Global macroscopic states such as superconductivity ensue from quantum coherence (with attendant superposition) among only very small fractions (say 1%) of components (six tubulins per neuron). For τ = 25 msec, 20,000 such neurons would be required to elicit OR. In human brain, cognition and consciousness are, at any one time, thought to involve tens of thousands of neurons (10,000 to 1,00,000 neurons) which may be widely distributed throughout the brain. At 80 Hz or higher frequency, associated with Tibetan or other meditators, expanded awareness states of consciousness might be expected with more neuronal brain involvement. Does this rule have any empirical support? An affirmative answer can be provided by the Microtubule Part of Penrose-Hameroff (PH) theory by linking Diosi-Penrose objective reduction rule to Hameroff’s belief that consciousness is closely linked to the microtubular structure of the neuron. 2. Topological Quantum Computation Topological quantum computing in ‘Orch OR’ is shown in Figure 2 (Penrose and Hameroff, 2011). Figure 2 extending microtubule A-lattice hydrophobic channels results in helical winding patterns matching Fibonacci geometry (Fibonacci series : e.g. 1, 1, 2, 3, 5, 8 etc. in which each Fibonacci number is the sum of the preceding numbers). Bandyopadhyay (2011) has evidence for ballistic conductance and quantum interference along such helical pathways which may be involved in topological quantum computing. Quantum electronic states of London forces in hydrophobic channels result in slight superposition separation of atomic nuclei, sufficient EG for Orch OR. This image may be taken to represent superposition of four possible topological qubits (cid:61) , will undergo OR, and reduce to specific pathway(s) which then which, after time T = tau = GE implement function. B-lattice microtubules have a vertical seam dislocation. In the quantum theory, the quantum state of the n indistinguishable particles (e.g. quasiparticle, anyons) belongs to a Hilbert space that transforms as a unitary representation of the braid group nB of n strands (Sahni, Lakshminarayanan and Srivastava, 2011). nB can be presented as a set of generators that obey particular defining The braid group relations. To understand the defining relations, we may imagine that the n (say 3) particles occupy n (say 3) ordered positions (labeled 1, 2, 3, . . ., n) arranged on a line. Let 1σ denote a counterclockwise exchange of the particles that initially occupy positions 1 and 2, let 2σ denote a counterclockwise exchange of the particles that initially occupy positions 2 and 3, and so on. Any braid can be constructed as a succession of exchanges of neighboring particles; hence σ σ σ − are the group generators. 1 The second, slightly more subtle type of relation is σσ σ σ σσ j 1,2,..., 2 − ,..., n 1 j = j 1 + j j 1 + j , 1 + , 2 j = n 1 2 1 2 1 σσσ σσσ= 2 which is called the Yang-Baxter relation. We can verify the Yang-Baxter relation by drawing the two braids on a piece of paper (Fig. 3), and observing that both describe a process in which the particles initially in positions 1 and 3 are exchanged counterclockwise about the particle labeled 2, which stays fixed — i.e., these are topologically equivalent braids. Final time-slice (t=T) 3 2 1 3 2 1 1 2 3 Intermediate time slices Initial time slice (t=0) 1 2 3 1 2 1 2 1 with σσσ σσσ= 2 3n = strands each Figure 3 A simple diagram showing the two braids The most important issues facing topological quantum computation are twofold : (1) finding or identifying a suitable system with the appropriate topological properties, that is, non-abelian statistics, to enable quantum computation, and (2) figuring out a scheme to carry out the braiding operations necessary to achieve the required unitary transformations. The only topological system known to exist in nature is the quantum Hall regime. Thus, there is a strong need to find other systems satisfying the above two criteria for topological quantum computation. Topological quantum computation can then be carried out by moving quasi particles around one another in two space dimensions. The quasi particle world-lines form topologically nontrivial braids in three (= 2 + 1) dimensional space-time, and because these braids are topologically robust (i.e., they cannot be unbraided without cutting one of the strands) the resulting computation is protected against error. An Orch. OR qubit based on topological quantum computing specific to microtubule polymer geometry was suggested by Hameroff et. al. in 2002 (Penrose and Hameroff, 2011). Conductances along particular microtubule lattice geometry, e.g. Fibonacci helical pathways, were proposed to function as topological bits and qubits. Bandopadhyay (2011) has preliminary evidence for ballistic conductance along different, discrete helical pathways in single microtubules. 3. Modelling Mircotubules as Quantum Hopfield Networks Hopfield neural networks are a class of neural network models where non-linear graded response neurons organized into networks with effectively symmetric synaptic connections are able to implement interesting algorithms, thereby introducing the concept of information storage in the stable states of dynamical systems. The dynamics of the state-space trajectory as well as time domain evolution of sensitivities of the states with respect to circuit parameters using system dynamics simulation has been extensively studied (Kumar and Satsangi, 1993). Behrman et. al. (Behrman, 2006) have simulated a quantum Hopfield neural network with the structure of a microtubule. They have used an extremely simplified model of the tubulin dimers: each is represented simply as a qubit, a single quantum two-state system, the quantum analog of a classical bit. A qubit, unlike a bit, can exist in a superposition of the two states. They have included Coulombic interactions between qubits, and thermal effects for the qubits themselves. This is a kind of quantum Hopfield network. Each of the individual processing elements or neurons (here, the qubits) interacts with each of the others (a fully interconnected network); the elements (qubits) are initialized in some state, and allowed to evolve towards a local minimum. These networks store information and are also called “associative memory” or “content addressable memory” models. The final stable state is a pattern that is recalled by the network. Inspired by Penrose-Hameroff Orch. OR theory, Behrman et. al. (Behrman, 2006) use an extremely simplified model of tubulin dimers, each represented simply as a qubit, a single quantum two-state system. In order to obtain the Quantum Hopfield Net (QHN), they consider an array of N qubits. Hamiltonian or energy function operator for each qubit j is given by K Aσ σ z + x (1) = j H where xσ ⎡ = ⎢ ⎣ 0 1 1 0 ⎤ ⎥ ⎦ and Hamiltonian Operator H 0 ⎤ ⎥− 1 ⎦ are respectively the Pauli X and Z matrices; and the full zσ 1 0 = ∑ ⎡ = ⎢ ⎣ N H j j 1 = The first term in Eq. (1) represents the flipping of the qubit from one state to the other, called “tunneling”, with amplitude K. The second term in Eq. (1) represents the energy-difference 2A between the two states. This difference can be the result of external fields or interaction with other qubits. The Pauli Matrix zσ measures the state of the system; e.g. (-1) state. Thus each qubit j may be visualized as a loop having n discretization points, propagating in imaginary time from 0 to β. [A discretization point at any instant of imaginary time is the instantaneous state of the qubit at that corresponding value of the inverse temperature]. , telling us that the system is in the 0 ⎡ ⎤ ⎢ ⎥ 1 ⎣ ⎦ 0 ⎡ ⎤ ⎢ ⎥ 1 ⎣ ⎦ = − ( 1) zσ Let ψ δ γ + = 0 1 represent a qubit in terms of Quantum Terminal Graph (QTG) (Srivastava et. al. 2011, 2014) in Figure 4 where 0 is the (+1) state 1 ⎡ ⎤ ⎢ ⎥ 0 ⎣ ⎦ and 1 is the (-1) state 0 ⎡ ⎤ ⎢ ⎥ 1 ⎣ ⎦ . Notice that σ x δ γ ⎡ ⎤ ⎡ ⎤ ⎢ ⎥ ⎢ ⎥ γ δ ⎣ ⎦ ⎣ ⎦ = which flips the probability amplitudes. For example, the (+1) state 0 ' is flipped to the (-1) state 1 , and vice-versa. The Quantum Terminal Graph (QTG) in Fig. 4 (Srivastava et. al. 2011, 2014) is shown as the union of two separate parts such that 1 1a a denotes 1 1bb denotes the “ket 1”-part (which form an orthonormal basis for the the “ket 0”-part and Hilbert space). Here δ and γ are complex numbers such that 2δ is the probability of qubits of “information / computation data” flow-rate along unit directional (quantum across variable) 2γ is the probability of “information / computation data” flow-rate along unit vector 0 ; and directional (quantum across variable) vector 1 . Notice that the sum of probabilities of flows along 0 and 1 adds to 1, i.e. = . Accordingly, the probable qubit states can be said to be normalized to length 1. δ γ+ 1 2 2 ' 0 1a δ 1a ' 1 1b γ 1b ' ∪ Figure 4 Quantum Terminal Graph Srivastava et. al. (2014) have applied Graph Theoretic Quantum System Modelling (GTQSM) in continuum of protein heterodimer tubulin molecules of self-assembling polymers, viz., microtubules in the brain as a holistic system of interacting components representing hierarchical clustered quantum Hopfield network (hQHN) of networks. The quantum input-output ports of the constituent elemental interaction components (or processes) of tunneling interactions and Coulombic bidirectional interactions are in cascade and parallel interconnections with each other while the classical output ports of all elemental components are interconnected in parallel to accumulate micro-energy functions generated in the system as Hamiltonian (or Lyapunov) energy function. They have presented an insight, otherwise difficult to gain, for complex system of systems represented by clustered quantum Hopfield network (h-QHN), through application of graph theoretic quantum system modelling (GTQSM) construct presented in the next section. 4. Generalization of Quantum Hopfield Network Model for n-qudits Let − represent a generalized quantum odd prime based qudit (with dimension d=n) in comparison to the commonly used even-prime based quantum bit (quits with dimension 2) in terms of Quantum Terminal Graph (QTG) in Figure 5 ψ α β γ δ ... + + 1n ν = + + + 3 1 0 2 and 1 is the state and 2 is the state , and n is the state . 0 ⎡ ⎤ ⎢ ⎥ 1 ⎢ ⎥ 0 ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ 0 ⎢ ⎥⎣ ⎦ 0 ⎡ ⎤ ⎢ ⎥ 0 ⎢ ⎥ 1 ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ 0 ⎢ ⎥⎣ ⎦ 0 ⎡ ⎤ ⎢ ⎥ 0 ⎢ ⎥ 0 ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ 1 ⎢ ⎥⎣ ⎦ 1 b 2 c 3 d 1n − n ∪ γ δ ∪ . . . ∪ ∪ 1γ− 1 ⎡ ⎤ ⎢ ⎥ 0 ⎢ ⎥ 0 ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ 0 ⎢ ⎥⎣ ⎦ where 0 is the state 0 a α 'a ∪ β 'b 'c 'd 'n Figure 5 Generalized Qudit (dimension d=n) as union of n separate parts 'a a denote the first "ket-0" part, 'c c denotes the third "ket-2" part, The generalized quantum terminal graph (QTG) with dimension (d=n) is shown in Figure 5 as 'b b denotes the second the union of n separate parts such that 'd d denotes the fourth "ket-3" part and so on, "ket-1" part, 'n n which denotes the "ket-(n-1)" part (which form an orthonormal basis vector-set for the upto 2α is the probability of Hilbert space). Here qubits of "information / computation data" flow-rate along unit directional (quantum across 2β is the probability of qubits of "information / computation data" flow- variable) vector 0 ; 2γ is the probability of qubits rate along unit directional (quantum across variable) vector 1 ; of "information / computation data" flow-rate along unit directional (quantum across variable) 2δ is the probability of qubits of "information / computation data" flow-rate along vector 2 ; 2ε is the probability of qubits of unit directional (quantum across variable) vector 3 ; αβγδ ν− are complex numbers such that 1) ,( , , , 2 1 2 2 2 2 (2) 2 2 2 1 = 21ν− 2 α β+ ν + + − ... 1 α β γ + + α β γ δ + + + 1n − adds to 1, i.e. n = discretization points in a system of "information / computation data" flow-rate along unit directional (quantum across variable) is the probability of "information / computation data" flow-rate vector 4 and so on, and 1n − . Notice that the sum of along unit directional (quantum across variable) vector probabilities of flows along 0 , 1 , 2 , . . . and Accordingly, the probable qudit states can be said to be normalized to length 1. Eq. (2) is the equation of an n-dimensional sphere. Thus each qudit j may be visualized as a d- 2 1 = , dimensional hyper sphere (instead of a loop for qubit which has only 2 states, i.e. = , the equation of a sphere) having n a qudit which has only 3 states, i.e. discretization points, propagating in imaginary time from 0 to β (this is different from the index β for 1 ). [A discretization point at any instant of imaginary time is the instantaneous state of the qudits at that corresponding value of the inverse temperature]. Figure 6 shows a qubit with (Srivastava et. al., 2014). Figure 7 shows a qudit (d=3) with in different colours. Each qubit is now represented by a sphere with 8 discretization points. Figure 8 shows a qudit (d=5) with 5N = qudits. In this solar system model, each 3-dimensional sphere is rotating in a loop / circle and sample connections are shown for discretization points i = 9 to 16. Similar connections exist for i = 1 to 32 as shown by black lines. If radius of the sphere (planet) is 1r and radius of the orbit is 2r (r1 and r2 are labelled in qudit j = 3 for illustration), then n Total degrees of freedom Discretization Points used = = as the system equation for j which is the same as Eq. (2), i.e. We will have 1 2 2 α β γ δ ε This idea can be similarly extended for even higher dimensions of d. (equation of a sphere) (equation of a loop / circle) 3p = degrees of freedom q = degrees of freedom 8n = discretization points in a system of discretization points in a system of 1r α β γ 2 2 2r δ ε 2 + + + = = = for a 5-dimensional qudit. p q = + = 5 2 2 .2 q = p 3 × 3N = qubits (d=2) 3N = qudits shown = × = 8 4 32 r+ 2 2 2 + r 2 1 + n = 32 2 2 2 + 2 + 2 2 1 4 2 2 2 ζ ζ i=3 (3β/4) 31 3α 31 4α 12 3α i=4 (0,β) j=1 i=2 (β/2) ζ ζ i=1 (β/4) 31 1α i=3 (3β/4) ζ ζ i=4 (0,β) j=3 i=2 (β/2) 31 2α 12 2α 23 4α 23 2α i=3 (3β/4) 23 1α ζ i=1 (β/4) ζ 23 3α 12 4α ζ ζ 12 1α 'jj iα α= N ' 1, ∀ ≠ = and i=1,n j j ζ ζ i=4 (0,β) j=2 i=2 (β/2) i=1 (β/4) Figure 6 A qubit with n = discretization points in a system of 4 3N = qubits 31 1α 31 5α 31 8α 31 4α j=3 31 2α 31 6α 31 7α 31 3α 23 1α 23 7α 23 4α 'jj iα α= N ' 1, ∀ ≠ = and i=1,n j j j=1 12 1α 12 2α 12 6α 12 7α 12 8α 12 4α 12 5α 12 3α j=2 23 2α 23 6α 23 3α 23 5α 23 8α 23 3α Figure 7 A qudit (d=3) with 8n = discretization points in a system of 3N = qudits i = 1 to 8 i = 25 to 32 j = 1 i = 9 to 16 Similar connections i = 1 to 8 i = 17 to 24 i = 1 to 8 'jj iα α= N ' 1, ∀ ≠ = and i=1,n j j r2 j = 3 i = 9 to 16 i = 25 to 32 j = 2 i = 9 to 16 r1 i = 25 to 32 i = 17 to 24 Figure 8 A qudit (d=5) with 32 n = discretization points in a system of i = 19 to 24 3N = qudits Sample connections are shown for discretization points i = 9 to 16. Similar connections will exist for all i as shown by black lines 5. Clustered hierarchical level Model of the Brain Woolf N. (2006) has proposed five hierarchical levels of quantum entanglement-based neural interaction microtubules spread among 13-23 billion neurons in the human cortex of the estimated 100 billion neurons in brain. We can extends the graph-theoretic modelling framework to these, in general, l levels. An estimate of Nl and degree of entanglement lα for l = 1 to 5 by Nancy Woolf (Woolf, 2006) and adapted by Srivastava et. al. (2014) is summarized in Table 1. The cluster approach propounded by Srivastava et. al. (2014) views the quantum Hopfield network as Nn information processing subunits (each with states 1± ), corresponding to N qubits of tubulin molecule, which converge to )Nn -dimensional box. ( 2 Nn corners of an ( ) Table 1 Hierarchical levels of quantum entanglement Hierarchical Level of Quantum Entanglement Quantum entanglement among Estimate of number of qubits Nl by Woolf (2006) as adapted by Srivastava et. al. (2014) Estimate of Woolf lα by Tubulins of the same microtubule Pairs of microtubulins (qubits) in the same neuron pairs of microtubulins (qubits) belonging to different neurons highly interconnected cortical areas entanglement among cortical areas having few or no axonal connections 1 2 3 4 5 4 10 N → tubulins of the same 1 10 15 microtubule in any one of possible clusters m = 1 Srivastava et. al. (2014) 10 N → different microtubules in 2 the same neuron in any one of 4 possible clusters m = 2 10 15 Srivastava et. al. (2014) 3 10 3 N → different microtubules belonging to different neurons in same modules in any one of m = 3 10 16 possible clusters Srivastava et. al. (2014) N → different neurons in any 10 m = 4 10 12 one of possible clusters Srivastava et. al. (2014) N → different neurons in any 10 10 5 one of possible clusters m = 5 10 9 4 7 (Teleportation invoked by Srivastava et. al. (2014)) likely to be high α α<< 1 2 3 α α<< 2 3α α< 4 4α α< 5 2 4 10 4 10 m = 2 10 15 m = 1 10 15 α α<< 1 2 1j at the first clustered level. ' The first level of neural interaction would include quantum entanglement among approximately N → tubulins of the same microtubule in any one of possible clusters. Since this 1 level of neural interaction is between closely interacting particles, the degree or density of 1α of Coulombic bidirectional entanglement is likely to be high as reflected by the strength interaction process between the same discretization points i for each pair of different qubits 1j and The second level of neural interaction would involve entanglement between pairs of N → different microtubules in the microtubulins (qubits) each drawn from approximately possible clusters. Microtubule Associated Protein (MAP2) same neuron in any one of does link neighbouring microtubules together (or to actin filaments) and these linkages dynamically associate and disassociate. This degree of entangelemnt would be somewhat less of Coulombic than that of tubulins in the same microtubule as reflected by strength bidirectional interaction process at this second clustered level. The third degree of neural interaction would involve entanglement between pairs of N → different microtubules microtubulins (qubits) each drawn from approximately possible clusters. This belonging to different neurons in same modules in any one of degree of entanglement at the third clustered level would be expected to be less than that of tubulins in the different microtubules of the neuron at the second clustered level as reflected by α α<< 2 The fourth degree of neural interaction would involve entanglement within highly interconnected N → different cortical areas (e.g. cortical areas of the same sensory modality) involving neurons in any one of possible clusters. Entanglement can occur over long distances when there is a classical channel of information, for example, axonal connectivity between cortical areas or electromagnetic energy flow from one brain region to another, and a pre- existing entanglement. Qubits performing local operations can even send information to entangled qubits over long distances via classical channels as in quantum teleportation. This at degree of entanglement might be expected to be medium to strong as reflected by the fourth clustered level. The fifth level of neural interaction would involve entanglement among cortical areas having few possible or no axonal connections involving clusters. Brain-wide entanglement might occur in a manner predicted by Bell’s theorem stating that whenever a quantum measurement is made on one part of a holistic quantum system, this will produce an effect on the other part of the system. This degree of entanglement would be weakest, thereby allowing for greater flexibility in brain response, as reflected by the strengths of Coulombic bidirectional interaction processes at the various clustered levels of hierarchy α α α α α 1 N → different neurons in any one of 5 10 3 3 m = 3 10 16 m = 4 10 12 m = 5 10 9 10 10 3α α< 4 ( ) 7 10 < 5 4 < < 3 < 2 . 3 . 4 Quantum entanglement is considered in this section only in neuronal microtubules since non- neuronal cells divide repeatedly and recycle tubulins for mitotic spindle microtubules. On the other hand, neurons don’t divide, their microtubules remaining polymerized for the life of the cell, suitable for information encoding, memory and consciousness. At the first level, within a microtubule, direct quantum information transfer may take place and may be protected through topological or other error-correcting mechanism. At the fourth hierarchical level, where only a classical information channel exists, quantum information transfer and error correction may still take place via quantum teleportation. At the fifth level, quantum error correction may still be possible through weak classical channel and / or collateral learning and training as well as quantum teleportation because of quantum entanglement (Srivastava et. al., 2014). We model a microtubule as a quantum Hopfield network, an interconnected network consisting of tubulin dimers. Srivastava (2013) and Srivastava et. al. (2014) have represented each tubulin dimer as a qubit. It is essentially quantum mechanical (an Abelian Z2 lattice gauge theory) (Zee, 2010). We can generalize quantum information processing further to quantum dits or qudits that are d-level systems as an extension of qubits that could speed up computing tasks even further. A qubit is then a special case of a qudit with d = 2. In comparison to the qubit system, d- dimensional quantum states will be more efficient in quantum applications. In our ongoing work, we are considering tubulin dimers as higher order qubits called qudits, which are quantum systems of dimensionality d (an odd prime). So, the values of d, such as 3 and 5, this model is quantum field theoretic (a non-abelian lattice gauge theory). For larger values of d, our model may be string theoretic. Graph theory has helped us to address the problem of modelling continuum, which is essentially a field problem by converting it into a circuit problem by discretization, which is an approximation made by us using graph theory. Graph theory allows for multi-terminal representations, aggregation and disaggregation, thereby allowing us to model the brain at any of the five hierarchical levels. A system schematic for hierarchical clustered quantum Hopfield networks appears in Figure 9 (Srivastava 2013, Srivastava et. al., 2014). The Hamiltonian (or Lyapunov) Energy Function HhQHN for the simplest possible Quantum Hopfield Network, considering Nl qubits (at clustered hierarchy level l) each with n discretization points, propagating in imaginary time (inverse temperature) over 0 to β, (where each pair of qubits interacts only at equal imaginary discretized time i), is given by (Srivastava 2013, Srivastava et. al., 2014) : H hQHN ⎧ ⎪ = − ⎨ ⎪ ⎩ 1 β = I = H ⎧ ⎪ ⎨ ⎪ ⎩ ⎛ N ∑∑ ζ ⎜ ⎝ 1 = 1 = j 0 n i 0 n N 0 ∑∑ j 0 1 = i 1 = H j 0 1 + i i j 0 j 0 1 + S S i ⎧ ⎫ ⎞ ⎪ ⎪ + − ⎟ ⎬ ⎨ ⎪ ⎠ ⎪ ⎭ ⎩ 5 ≤ ∑∪ N N l l l 1 = ⎫ ⎧ ⎪ + ⎬ ⎨ ⎪ ⎩ ⎭ 1 β 5 l ∑ 1 = H II = n N l ∑∑ j 1 = i 1 = ⎡ α ⎢ l ⎢ ⎣ ⎛ ⎜ ⎜ ⎝ j n N l ∑ ∑ 1 ≠ = ' j l i 1 = ' j j l l H i ' j l j l S S i i ⎫ ⎬ ⎭ ⎞ ⎟ ⎟ ⎠ ⎤ ⎥ ⎥ ⎦ ⎫ ⎪ ⎬ ⎪ ⎭ (3) such that N = 0 l 1,5 = The first term in Eq. (3) is tunnelling energy and the second is Coulombic energy. Figure 9 Hierarchical Clustered Quantum Hopfield Network (hQHN) – System Schematic 6. Contextuality In a recent paper in Nature, Howard et. al. (2014) show that contextuality, an abstract generalization of nonlocality is a necessary condition for magic state distillation. They make an abstract graph called an exclusivity graph to obtain their results. Qudits of odd prime dimensionality are naturally differentiated from qubits in their analysis. In exclusivity graph in Figure 10(a), each of the 30 vertices corresponds to a two-qubit stabilizer state; connected vertices correspond to orthogonal states. A maximum independent set (representing mutually non-orthogonal states) of size 8 is highlighted in red. They have proved a remarkable equivalence between the onset of contextuality and the possibility of universal quantum computation via ‘magic state’ distillation, which is the leading model for experimentally realizing a fault-tolerant quantum computer. This is a conceptually satisfying link, because contextuality, which precludes a simple ‘hidden variable’ model of quantum mechanics, provides one of the fundamental characterizations of uniquely quantum phenomena. Furthermore, this connection suggests a unifying paradigm for the resources of quantum information: the non- locality of quantum theory is a particular kind of contextuality, and non-locality is already known to be a critical resource for achieving advantages with quantum communication. We are working on simulating a functional model of the brain based on quantum Hopfield networks with neurons selected from the various regions of the brain, based on functional layer diagram. Such models should be able to handle association and contextuality both as they will have neurons from portions of the brain tasked with association (association cortex) as well as contextuality (pre-frontal cortex). Table 2 gives a summary of the macrocolumn functional units mentioning number of columns, which is believed to be the fundamental computational unit of the cortex, and an estimate of the neurons in various lobes of the brain. (a) Macrocolumn functional units of the cortex Functional Unit Visual Cortex Auditory Cortex Somatosensory cortex Association cortex Motor cortex TOTAL Table 2 Columns 232 million columns 83 million columns 21 million columns 5 million columns 1 million columns 342 million columns Billion Neurons 68.3 24.0 6.0 1.4 0.3 100 billion (b) An estimate of the neurons in various lobes of the brain Frontal Lobe Motor cortex (precentral gyms) Temporal Lobe Parietal Lobe Occipital Lobe TOTAL 35 billion 6 billion 19 billion 16 billion 14 billion 90 billion (a) Exclusivity graph proposed by Howard et. al. applied to two qubits Figure 10 (b) QTG Representation Based on insights given from Srivastava (2013) and Srivastava et. al. (2014), it is proposed to derive exclusivity graph in Figure 10(b) by representing Quantum Terminal Graphs (Srivastava 2013, Srivastava et. al. 2014) as composite Quantum Terminal Graphs each represented by a single composite quantum terminal graph which may be further reduced to a representation in terms of a single composite quantum vertex or node as in Figure 10(b) by rendering the composite datum node as a floating unspecified datum node. This representation was also presented by a team from DEI before Department of Science and Technology, Govt. of India’s Cognitive Science Research Initiative (CSRI) Review Committee in March 2015 and revised project proposal prepared in April 2015. We believe that our graph theoretic approach based on physical systems theory, will give a simpler derivation of their results potentially allowing us to demonstrate that contextuality is not only necessary but also sufficient for magic state distillation. We are attempting to address the important question of how does the brain overcome decoherence using topological, fault tolerant computing and magic state distillation. 7. Power Laws To gain a picture of emergent patterns of brain activity, investigators need new sensing devices that can record from assemblages of thousands of neurons. Nanotechnology, with novel materials that sometimes measure less than the dimensions of individual molecules, may assist in making large-scale recordings. Prototype arrays have been built that incorporate more than 100,000 electrodes on a silicon base; such devices could record the electrical activity of tens of thousands of neurons in the retina. Electrodes are only one way to track the activity of neurons. Methods that move beyond electrical sensors are making their way into the lab. Biologists, borrowing from technologies developed by physicists, chemists and geneticists, are beginning to visualize living neurons in awake animals going about their daily paces using electro-encephalography (EEG), magento- encephalography (MEG) and functional Magentic Resonance Imaging (fMRI) and further analysis by Fourier transform methods. An interesting conclusion we can draw from a course on noise is that the 1/ρ behaviour of EEG and MEG (i.e. “one over of” power-spectrum) is a special noise (also called ‘pink’ noise) is the golden means between the disorder with high information content (white noise with 1/ρ0 flat (constant power spectrum) and the predictability with low information content (brown noise with “one over ρ2 power spectrum”). The cerebral cortex with its most complex architecture generates the most complex noise known to physics (Buzsaki, 2006). The anatomical-functional organization of the cerebral cortex should have consequences and limitations on cognitive behaviours as well. A well-known psychological law that comes to mind in connection with the 1/f nature of the cortical EEG / MEG is that of Weber and Fechnel : the magnitude of a subjective sensation (a cognitive unit) increases proportionally to the logarithm of the stimulus intensity (a physical unit). This collective behaviour of neurons, summed up crudely as the mean field (EEG / MEG) is a blend of rhythms. Neuronal network in the mammalian cortex generates several distinct oscillatory bands, covering frequencies from <0.05 hertz to >500 hertz. Brain evolution opted for a complex wiring pattern in the mammalian cortex. The resulting 1/fα temporal statistic of the mean field are the hallmark of the most complex dynamics and imply an inherently labile, self-organized state. Unlike so-called normal distribution (such as for human brain weight (e.g. 1.35 kg or 3 lbs.), in scale-free systems (governed by power law), things are different. There is no typical example in a scale-free system. A power law implies for the case of fracture, that if we plot the numbers of different size of pieces on a log-log scale, we will get an oblique line. In fractal geometry, fractals are usually defined in statistical or qualitative terms, loosely indicating anything that “looks like itself”. The scale invariance of fractals implies that knowledge of the properties of a model system at any scale can be used to predict the structure of the real system at larger or smaller scale. 8. Conclusions Price and Barrell (2012) show how a science of human experience can be developed through a strategy by integrating experiential paradigms with methods from the natural sciences. A science of human experience would largely be about human meanings and also would constitute a science of consciousness. Psychophysics includes a rudimentary form of introspection that is critical for developing a science of human meanings and consciousness. The qualitative aspects of human experience serve as a foundation for developing hypotheses to be used in experiments that have quantitative methods. This model serves as a useful example for correlating first person and third person approaches for explaining a conscious experience. Interesting conclusions can be drawn from a study of 1/f α behaviour of brain-scans of humans, particularly during Eastern Meditational Practice. The cerebral cortex with its most complicated architecture generates the most complex noise known to Physics. One over f (α ≡ 1) power spectrum (also called "pink noise") is a golden mean between the white noise (α ≡ 0, i.e. constant power spectrum) representing disorder with high information content, and brown noise (α ≡ 2) representing predictability with low information content. The anatomical functional organization of the cerebral cortex brings to mind a well-known psycho-physical law credited to Weber and Fechner. Accordingly, the magnitude of a subjective sensation such as meditational consciousness in cognitive and meta-cognitive domains increases proportionally to the logarithm of the stimulus intensity, "one over f (i.e. rhythmic frequency)", a physical unit. Notably, scale- free systems or fractals are governed by Power Laws (Satsangi and Sahni, 2015). The formulation of power law of meditational consciousness requires invoking a family of models based on Omni quantum theory and physical system theory (including fuzzy analytical hierarchy process), and requisite integration of first person inner experiences of meditationists as co-investigators with the third person scientific methodology of observing, reporting, understanding and hypothesis-testing (Satsangi and Sahni, 2015). REFERENCES Bandopadhyay A. (2011), Direct experimental evidence for quantum states in microtubules and topological invariance, Abstracts : Towards a Science of Consciousness, 2011, Stockholm, Sweden, http://www.consciousness.arizona.edu/TSC2011MicroTub.htm. Behrman E.C. (2006), Gaddam K, Steck JE and Skinner SR, MTs as a Quantum Hopfield Network, The Emerging Physics of Consciousness, ed. J.A. Tuszynski, pp.351-370, Springer, New York, 2006. Buzsaki G. (2006), Rhythms of the Brain, Oxford University Press, USA, 2006. Hameroff S, Hagan S and Tuszynski JA (2002), Quantum computation in brain MTs: Decoherence and biological feasibility, Physical Review E 2002;65:061901 1-11. Henry, Stapp, (2006), "Quantum Interactive Dualism: Libet and Einstein-Podolsky-Rosen Causal Anomalies." Invited contrib. to Erkenntnis, Feb. 2006. Howard M., Wallman J., Veitch V., Emerson J., “Contextuality supplies the ‘magic’ for quantum computation”, Nature, Vol. 510, pp. 351-355, June 19, 2014. Kumar and Satsangi (1993), “System Dynamics Simulation of Non-degenerate and Degenerate Circuits and Systems”, Paritantra (A Journal of Systems Science and Engineering, Systems Society of India, Wiley Eastern Limited, New Delhi) 1(1), 11-39. Penrose and Hameroff (2011), Consciousness in the Universe: Neuroscience, Quantum Space- Time Geometry and Orch OR Theory, Journal of Cosmology, Vol. 14, April-May 2011. Price D. and Barrell J. (2012), Inner Experience and Neuroscience : Merging both Perspectives, MIT Press, USA, 2012. Sahni V., Lakshminarayanan V., Srivastava D.P. (2011), Quantum Information Systems, ISBN 9780070707078, Tata McGraw Hill, New Delhi, April 2011. Satsangi P.S. (2006), “Generalizing Physical Systems Through Applied Systems Research from “Real” Physical Systems through “Conceptual” Socio-Economic-Environmental Systems to “Complete” (Physical-Mental-Spiritual) Creational Systems”, International Journal of General Systems, Vol. 35, No. 2, 2006, pp. 127-167. Satsangi P.S. (2010), Cosmology from the Twin Vantage Points of Radhasoami Faith and Systems Science, Vision Talk at International Seminar on Religion of Saints (‘Sants’) – Radhasoami Faith, Spiritual Consciousness Studies (SPIRCON 2010), Nov. 12-13, 2011. Satsangi P.S. (2011), Cosmology of Consciousness : Towards Quantum-Theoretic Systems Modelling; Spirit-Mind-Brain Interactions, Vision Talk at Inaugural Workshop of DEI Centre for Consciousness Studies (CONCENT 2011), Oct. 1-2, 2011, Abstract Book, pp. 50-67. Satsangi P.S., Sahni V. (2012), “Cosmic Consciousness Hierarchization : Analytic, Experimental and Experiential”, Concurrent Session on Science and Spirituality, Toward a Science of Consciousness (TSC 2012), Tucson, Arizona, USA, April 2012, Abstract Book, pp. 193. Satsangi P.S., Sahni V. (2015), “Power Law of Meditational Consciousness”, Concurrent Session on Meditation, Toward a Science of Consciousness (TSC 2015), Helsinki, Finland, June 2015. Srivastava D.P., Sahni V. and Satsangi P.S. (2011), “Graph Theoretic Quantum System Modelling for Information / Computation Processing Circuits”, International Journal of General Systems, Vol. 40, No. 8, ISSN 0308-1079, Taylor & Francis, UK, pp. 777-804. Srivastava D.P. (2013), "Graph Theoretic Quantum Field / System modelling for Quantum Information / Computation Circuits and Algorithms" Ph.D. Thesis, Dayalbagh Educational Institute, Dayalbagh, Agra, September 2013. Srivastava D.P., Sahni V., Satsangi P.S. (2014) “Graph Theoretic Quantum System Modelling for Neuronal Microtubles as Hierarchical Clustered Quantum Hopfield Networks”, International Journal of General Systems, Taylor & Francis, UK, Vol. 43, No. 6, pp. 633-648. Stapp (2007), Quantum Approaches to Consciousness, Chapter 31, Cambridge Handbook of Consciousness. Zee Anthony (2010), Quantum Field Theory in a Nutshell, Second Edition, Princeton University Press, USA, 2010.
1609.07656
1
1609
2016-09-24T19:06:35
Associative Memory Impairments arising from Neurodegenerative Diseases and Traumatic Brain Injuries in a Hopfield Network Model
[ "q-bio.NC", "q-bio.QM" ]
Neurodegenerative diseases and traumatic brain injuries (TBI) are among the main causes of cognitive dysfunction in humans. Both manifestations exhibit the extensive presence of focal axonal swellings (FAS). FAS compromises the information encoded in spike trains, thus leading to potentially severe functional deficits. Complicating our understanding of the impact of FAS is our inability to access small scale injuries with non-invasive methods, the overall complexity of neuronal pathologies, and our limited knowledge of how networks process biological signals. Building on Hopfield's pioneering work, we extend a model for associative memory to account for FAS and its impact on memory encoding. We calibrate all FAS parameters from biophysical observations of their statistical distribution and size, providing a framework to simulate the effects of brain disorders on memory recall performance. A face recognition example is used to demonstrate and validate the functionality of the novel model. Our results link memory recall ability to observed FAS statistics, allowing for a description of different stages of brain disorders within neuronal networks. This provides a first theoretical model to bridge experimental observations of FAS in neurodegeneration and TBI with compromised memory recall, thus closing the large gap between theory and experiment on how biological signals are processed in damaged, high-dimensional functional networks. The work further lends new insight into positing diagnostic tools to measure cognitive deficits.
q-bio.NC
q-bio
Associative Memory Impairments arising from Neurodegenerative Diseases and Traumatic Brain Injuries in a Hopfield Network Model Melanie Weber*, Pedro D. Maia and J. Nathan Kutz Department of Applied Mathematics, University of Washington, Seattle, WA 98195-3925, United States * Corresponding author. (Contact: [email protected], [email protected], [email protected]) Abstract Neurodegenerative diseases and traumatic brain injuries (TBI) are among the main causes of cognitive dysfunction in humans. Both mani- festations exhibit the extensive presence of focal axonal swellings (FAS). FAS compromises the information encoded in spike trains, thus leading to potentially severe functional deficits. Complicating our understanding of the impact of FAS is our inability to access small scale injuries with non- invasive methods, the overall complexity of neuronal pathologies, and our limited knowledge of how networks process biological signals. Building on Hopfield's pioneering work, we extend a model for associative memory to account for FAS and its impact on memory encoding. We calibrate all FAS parameters from biophysical observations of their statistical distribution and size, providing a framework to simulate the effects of brain disorders on memory recall performance. A face recognition example is used to demonstrate and validate the functionality of the novel model. Our re- sults link memory recall ability to observed FAS statistics, allowing for a description of different stages of brain disorders within neuronal networks. This provides a first theoretical model to bridge experimental observa- tions of FAS in neurodegeneration and TBI with compromised memory recall, thus closing the large gap between theory and experiment on how biological signals are processed in damaged, high-dimensional functional networks. The work further lends new insight into positing diagnostic tools to measure cognitive deficits. 1 Introduction Neurodegenerative diseases and traumatic brain injuries (TBI) are responsible for an overwhelming variety of functional deficits, both cognitive and behav- ioral, in animals and humans. Memory impairment, which is the focus of this work, is a particularly pernicious consequence for those affected. The patho- physiology of these induced brain disorders is usually complex, with key effects 1 of the injuries occurring at small spatial scales that are currently inaccessible by non-invasive diagnostic techniques. Indeed, a hallmark feature of the in- jury/disease pathology is the abundance of diffuse, focal axonal swellings (FAS) which compromise spike train encodings in neuronal networks. Thus, there is a broad need to understand how neuronal pathologies which develop at a cel- lular level compromise the functionality of a network of neurons responsible for cognitive function. In this article we extend well-established computational models of associative memory, i.e. the Hopfield network model, to incorporate FAS pathologies implicated in many brain disorders, providing novel metrics for quantifying memory impairments and functional deficits. The work is the first of its kind to integrate extensive experimental findings regarding FAS with mathematical models of spike-train encoding, neuronal networks, and memory association. Human memory is conjectured to work by forming associations [Gerstner et al(2014), Hopfield(1982), Squire et al(2007)]. Meaningful stimuli are usu- ally encoded, stored, and later recalled by the brain in response to some cue for use in a given activity. This cognitive process improves the interpreta- tion of subsequent stimuli that share common features with previously stored concepts [Hopfield(1982)]. What governs our memory recall abilities are coor- dinated exchanges of electrical signals between neurons in network structures. In fact, neurons constantly encode and transmit information in the form of spike trains along their axons [Gerstner et al(2014)]. Several pathological ef- fects, most notably FAS, can jeopardize this critical electrical activity, making memory performance particularly sensitive to common brain disorders. Given its ubiquity across many neurodegenerative diseases and traumatic brain injury, understanding the role of FAS in altering spike train encodings is of paramount importance, particularly on a network level where cognitive functionality oc- curs. In this work, we incorporate the biophysical effects of FAS observed in neurodegenerative diseases and/or following TBI to network models that exhibit memory retrieval capabilities. Specifically, we integrate cable theory models with FAS to quantify the impact of swellings on spike train propagation [Maia, Kutz(2014a), Maia, Kutz(2014b), Maia et al(2015)] and integrate these findings with a Hopfield associative memory model. Neuroscience has historically combined experimental observations with the- oretical models, from the early neuronal doctrine and single-cell analysis [Cajal, Estructura(1888),Bullock et al(2005)] to more recent neuronal-network paradigms [McCulloch, Pitts(1943), Hebb(1949)]. Experimental advances in physiological measurements of neuronal ensembles continuously challenge and improve ex- isting models, leading to larger and more complex simulations [Yuste(2015)]. Today, computational neuroscience is recognized as an important and acces- sible alternative to modeling the dynamical interplay between neurons, espe- cially when biophysical recordings are limited and/or inaccessible. Remarkably, John Hopfield's pioneer model for associative memory, and its later variations, are still widely used today to describe the recall of previously stored men- tal patterns through the evolution of neuronal dynamics [Hopfield(1982), Hop- field(1984),Hopfield, Tank(1985),Hopfield(2015),Hu et al(2015),Nowicki, Siegel- 2 mann(2010)]. Our aim is to modify the Hopfield model of associative memory by including the deleterious effects of FAS. Due to its simplicity and popularity, we found the Hopfield network model to ideally suited for first investigations of es- tablished single cell models for memory impairment [Maia, Kutz(2014a), Maia, Kutz(2014b), Maia et al(2015)] at the level of neuronal dynamics. The rich body of knowledge that has been developed for the Hopfield model serves as an ideal basis for this study and allows for an embedding of our results in existing theories [Anafi, Bates(2010), Menezes, Monteiro(2011), Ruppin, Reggia(2011)]. Overwhelming experimental evidence suggests that FAS is the hallmark manifestation of injuries on neuronal networks. Moreover, there now exists a wealth of experiments characterizing the statistical distribution of FAS as a func- tion of injury level, including size and frequency of swellings [Coleman(2005)]. Statistics can even be collected for specific neurodegenerative diseases such as Alzheimer's [Krstic, Knuesel(2012)], Parkinson [Galvin et al(1999)] or Multiple Sclerosis (MS) [Hauser et al(2006)], where the swellings occur as a consequence of complicated biophysical/biochemical deterioration of neurons. Thus, bio- physically plausible models can be posited that make use of the best available data regarding neurodegeneration and TBI. Understanding the role of FAS is of critical importance for comprehending cognitive deficits and memory impairment. Neurodegeneration via Alzheimer's, Parkinson's or MS is estimated to affect a large portion of adults. Indeed, Alzheimer's alone has recently be estimated to be the third leading cause of death, just behind heart disease and cancer, for older people. Likewise, TBI is responsible for millions of hospitalizations worldwide every year and is the leading cause of death among youngsters [Faul et al(2010), Jorge et al(2012), Xiong et al(2013)]. TBI is usually caused by external forces applied to the head as occurs often in contact sports. In fact, it has been reported [Fainaru, Fainaru(2013)] that 10 − 20% of professional football players suffer from TBI. The external forces damage neuronal axons in the brain resulting in FAS that modifies the shape of the axons, which in turn, distorts the transmitted pulses leading to a potentially significant loss of information [Maia, Kutz(2014b),Tang- Schomer et al(2010), Tang-Schomer et al(2012)]. 2 Results In this article, we introduce a biophysically realistic variation of Hopfield's model to incorporate a single-cell neuron model for FAS [Maia, Kutz(2014a)]. The FAS modeling is based upon state-of-the-art statistical estimates of size and distribution of FAS observed in experiments. We demonstrate the impact of FAS with intuitive examples from face recognition. From a translational viewpoint, our models give clear metrics for evaluating memory deficits as a function of injury level, thus suggesting how memory impairment metrics may allow one to infer network damage levels from memory recall tasks. 3 2.1 Face recognition on associative-memory networks We calibrate a Hopfield-like neuronal network [Hopfield(1982), Hopfield(1984)] to perform face recognition tasks (see Fig. 1). Our sample memory space consists of five human facial images (1044 × 1341 pixels), modified from the MIT faces database [Weyrauch et al(2004)] 1. For the second experiment (Fig. 2) we choose the three most correlated images to test for confusion between closely related memories. In our setting, neurons dynamically alternate between multiple firing-rate activity states as opposed to the binary formulation in the original standard Hopfield model [Hopfield(1982)]. This more realistic higher- dimensional feature space allows for the encoding and storing of more complex memories in the network. Fig 1. Face recognition task under the influence of neurodegenerative diseases and traumatic brain injury. A: For a noisy input with 80% overlap (i.e. 20% initial noise), the healthy system (left) achieves an overlap of 90% with the correct face (marked yellow). The chart presents the initial (blue) and final (green) Hamming distances between the current network state and the patterns corresponding to the respective facial images. When the recognition task is performed with an injured system (right), we observe a severe decrease in accuracy (overlap significantly smaller than in the healthy network) and confusion between two facial images. Both are characteristic injurious effects of memory impairments arising from FAS and are quantified with a recognition score in our study. B: Images of FAS -- highlighting their morphometric features -- adapted from experimental works used to calibrate our model [Wang et al(2011), Dikranian et al(2008)]. The weighting of neuronal connections captures key aspects of the network architecture, including how memories are encoded. They are initially calibrated to encode the desired set of memorized concepts as fixed points of the system. We add Brownian noise fluctuations to the dynamics as a proxy for natural stochastic fluctuations. As a result, when a partial or noisy version of one of 1Credit is hereby given to the Massachusetts Institute of Technology and to the Center for Biological and Computational Learning for providing the database of facial images. http: //cbcl.mit.edu/software-datasets/heisele/facerecognition-database.html 4 the memory concepts is presented, the system converges to the closest fixed point matching the input to the closest (least-square fit) memory. Our healthy network demonstrated in Fig. 1 achieved a recall accuracy of approximately 90% and distinguished perfectly between all images in our sample set. Injuries due to FAS deteriorate network performance by misclassifying given input stimulus, thus creating potential confusion in memory association. 2.2 Modeling dynamical impacts of FAS. The presence of FAS is known to distort neuronal spike-train dynamics, but precise electrophysiological recordings of pre- and post-FAS spike train dynamics are still unavailable. The recent theoretical models of Maia and Kutz [Maia, Kutz(2014a),Maia, Kutz(2014b),Maia et al(2015)], however, provide important estimates of spike train deterioration due to FAS. Using this model, we can characterize the effect of injury on the firing rate activity through a response function. The state variable of a Hopfield node, Si, is equivalent to the firing rate of that node. The collection of all nodes is denoted by the vector S. Injuries can then be characterized by the transfer function S = F (S, β). (1) where S is the effective firing rate (state) after the FAS, S is the firing rate (state) before the FAS, and β is a parameter vector indicating one of three in- jury types applied to individual nodes of the network [Maia, Kutz(2014a),Maia, Kutz(2014b)]. The function F (·) maps the pre-injury to post-injury firing rates using biophysically calibrated statistical distributions of injury in both frequency and size. If no FAS occurs to a given axon, then its state is unaffected (β0). This occurs with probability 1− p where p is the probability of injury and what we term injury level, i.e. larger p implies more injury. For those neurons with axonal swellings, the following manifestation of spike train deformations have been observed: unimpaired transmission (β1), filtered firing rates (β2), spike re- flection (β3), or blockage (β4). The injury type is dependent upon the geometry of the swelling, with blockage being the most severe injury type. From biophys- ical data collected on injury statistics [Dikranian et al(2008),Wang et al(2011)], both in swelling size and frequency, we assign a prescribed percentage of each type of injury (βj) to the network using calibrated simulations of spike propaga- tion dynamics [Maia, Kutz(2014a), Maia, Kutz(2014b)]. For a blockage injury (β4), no signal passes the swelling so the effective firing rate of the neuron goes to zero. Thus, S = 0 which prevents the neuron from adapting to the collective dynamics. Filtering injuries were taken to decrease the firing rate, with higher firing rates having a stronger chance of decreasing due to pile-up effects in the spike train [Maia, Kutz(2014b)]. Reflection of spike trains effectively filters the firing rate of an axon by a factor of two so that S = 0.5S. This is due to the fact that the reflected spike annihilates an oncoming spike [Maia, Kutz(2014a)]. Overall then, the method for producing the filtering function F (·) uses the most advanced experimental findings to date with recent computational studies of 5 Fig 2. Statistical evaluation of computational results. A: Effects of FAS on associative memory for a face recognition task. We measure recognition ability and accuracy for a sample set of three facial images [Weyrauch et al(2004)] over a given range of parameters (injury: 0 − 30%, initial noise: 0 − 30%). As the degree of injury increases, the noise handling ability of the system drops severely: The coloring of the heat maps change from yellow (significant recognition) in the upper left (small injury and initial noise) to dark orange and red (confusion) in the lower right corner (high injury and initial noise). We observe confusion and a decline in recall accuracy for every image in the set of samples. Regression over all N=1080 normalized data points indicates an exponentially declining relation between recognition scores and injury level as shown in the diagram. B: Mean recognition scores are displayed in red, standard deviations are shaded in gray. The dashed line indicates the recognition function obtained by linear regression. C,D: We tested our model in a biological realistic setting using experimental results from TBI studies by Wang et al. [Wang et al(2011)] and Dikranian et al. [Dikranian et al(2008)]. Our results show a modest, but unstable increase in memory performance for the adult brain during the first 24 hours after the injury (C, [Wang et al(2011)]), that drops again to the initial level during the following day. In the infant's mouse brain, we observe a steadily decreasing performance in the first 24 hours (D, [Dikranian et al(2008)]). The observations match the long recovering times that are commonly observed in patients suffering from TBI. 6 spike train propagation through FAS [Maia, Kutz(2014a), Maia, Kutz(2014b)]. See the Materials and Methods section for details on how we compute the sta- tistical distribution of injury types that are parametrized by the parameter βj. The presence of a significant number of blocked neurons in the Hopfield network creates blurring in the reconstructed concepts (faces) and therefore decreases the accuracy of the recalled information. Both filtering and reflection lead to confusion of correct states with their neighboring ones (in a Hamming distance sense). Ultimately, this decreases the ability of the network to perform denoising tasks. In most cases of FAS, all three mechanisms occur in simultaneously, depending on the type and time development of the injury [Maia, Kutz(2014a)]. 2.3 Memory impairments and confusion of concepts We examine the memory-recall performance of neuronal networks as a function of (a) distributions of FAS, (b) time-lapse from injury, and (c) percentage p of injured neurons. Our recognition score (see Fig. 6) quantifies the network's accuracy and ability to differentiate between previously stored concepts from noisy cues. Figure 2(a) shows recognition scores for the most similar triplet of faces in our set. In our first numerical experiment we considered FAS distributed according to results by Wang et al. [Wang et al(2011)] for TBI injuries after 12 hours. We varied the amount of noise (0−30%) and degree of injury (p = 0−30%) for each of the facial images in our triplet, allowing for cross-comparison of recognition scores and memory performance with respect to the choice of parameters (see Fig. 2). An interesting consequence of injury is that the network can produce, and often does for elevated injuries, erroneous associations of memories. Specif- ically, it confuses the concepts in the Hopfield network. Based on simulations of an uninjured network with noise, it could easily be conjectured that the fixed points associated with a given memory would disappear, with the system's dy- namics preventing it from converging to the correct pattern. On the contrary, we observe that the noise fluctuations cause the dynamics to converge to an erroneous fixed point under the influence of FAS in the injured system. For modest amounts of initial noise (15 − 30%), the affected system confuses the stored patterns and looses the ability to separate them properly. Confusion of concepts is especially pronounced when blockage and filtering account for the majority of FAS effects. Focal axonal swellings decrease the system's ability to handle noise and recall a previously stored pattern from partial information. Our simulations allowed us to quantify the rate at which the system's average recognition score R decrease as a function of injury strength p: R(p) = A − Bep (2) The functional form R(p) is produced from a linear regression over N = 1080 normalized data points obtained from computational experiments. We exam- ined the development of FAS from two different experimental data sets: in an 7 adult rat brain 12, 24 and 48 hours after TBI (Fig. 2c, [Wang et al(2011)]) and in an infant mice brain 30 min, 5, 16 and 24 hours after TBI (Fig. 2d, [Dikranian et al(2008)]). The values for the constants in eq. (2) are given by [A = 1, B = 0.24] and [A = 1.21, B = 0.22] respectively. We also evaluate recognition scores on injured networks as a function of time. Our results hint at a modest, but unstable increase in memory performance for the adult brain during the first 48 hours after the injury, and a performance decrease for the infant brain in the first 24 hours (see Fig. 2). This is consistent with the relatively long recovering times observed in patients suffering from TBI. 3 Discussion We consider memory impairments caused by Focal Axonal Swellings in a well- established Hopfield computational model for associate memory. The FAS are a hallmark feature of traumatic brain injuries and neurodegenerative diseases such as Alzheimer's, Parkinson's and MS. With this model, we can character- ize a neural network's ability to handle noise and perform recognition tasks at different levels of injury. Our computational model reproduces symptoms com- monly observed in patients suffering from brain disorders [Knutson(2012)]. The injured network-model for the task of face recognition loses accuracy in recalling facial images and confuses faces with similar features. From a Hopfield network viewpoint, less accuracy means that fixed points encoding facial images loose their stability. At higher injury levels, the network's dynamic is driven away from the correct face encoding, settling closer to an unstable fixed point asso- ciated with another facial pattern. Instability of the fixed points results from a large percentage of neurons manifesting FAS, leading ultimately to confusion between previously stored images. Confusion of concepts has a higher impact on the overall performance than the loss of accuracy. It severely affects the content and context of a stored concept. Such impairments are commonly observed in patients suffering from brain disorders [Coleman(2005)], especially in advanced stages of Alzheimer's diseases [Selkoe(2001),McKhann et al(1984)] and severe traumatic brain injuries [Cole, Bailie(2016), McGee(2016)]. To test our injury protocols in a biologically realistic setting, we calibrate our simulations with experimental FAS data from Wang et al. [Wang et al(2011)] and Dikranian et al. [Dikranian et al(2008)]. Both studies perform traumatic brain injuries to their model organism: [Dikranian et al(2008)] studied FAS in the cortex of infant mice, whereas [Wang et al(2011)] investigated swellings in the optic nerve of adult rat. Although [Dikranian et al(2008)] provides detailed morphometric data for injured axons in the infant mice, their functional assess- ments are still far from interpretable studies of memory development in human children [Mcallister(2004)]. TBI experiments in adult rats [Xiong et al(2013)] re- port functional impairments more analogous to human patients -- like deficits in tasks involving context memory [Schacter(1987)], conditional associative learn- 8 ing [Petrides(1985)], planning [Shallice, Evans(1978)] and other cognitive tasks [McDowell et al(1997)]. Adult rats, however, may exhibit memory deficits after mild TBI even without many signs of axonal injury [Lyeth(1990)]. In fact, recent studies demonstrate that catecholamines play a central role in the neurochem- ical activation and regulation of working memory [Mcallister(2004)] and such effects were not incorporated in the model. Thus, axonal structural damage may be sufficient but not necessary for the production of neurological and cognitive symptoms associated with TBI. This will be considered in future studies. There is much room for improvement in our injury model, especially given the variety of effects of different types of brain disorders. Although FAS are uni- versal pathological features, our results should be regarded as a foundational mathematical framework to study different disorders in neuronal networks. We hope to continue to develop computational models capable of describing a wide variety of cognitive deficits. Here, we use noise-handling as a proxy for cognition and describe functional deficits in terms of a few key model parameters. Our ongoing work includes the study of convergence and stability in the extended associate memory Hopfield model and the impact of memory impairments on these aspects. More importantly, this framework opens up opportunities for novel diagnostics: one could use a patient's inability to distinguish known visual impressions as an indirect proxy for cellular pathologies currently inaccessible to non-invasive techniques. There is a broad need in the neurological and biomedical commu- nities to understand the development, propagation, and functional implications of brain degeneration. As a consequence, medical imaging evolved dramatically with our ability to extract more information from highly under-sampled noisy data. Advances in Diffusion Tensor Imaging (DTI), functional MRI, acceler- ated MR imaging, and brain morphometry allow us, for the first time, to unveil network characteristics that distinguish healthy and pathological brains for a number of brain disorders. In fact, data-driven methods are key to graph-cut neuroimaging software, high angular resolution diffusion imaging (HARDI), and network-level analysis of cortical thickness from structural MRI. These methods, when combined, enable whole-brain network reconstruction and are revolutioniz- ing our understanding of incurable neurodegenerative diseases like Alzheimer's, Epilepsy, Autism, Schizophrenia, Stroke, and others. Such diseases, however, pose two fundamentally difficult challenges for functional diagnostics; changes in structural connectivity do not translate naively to dynamical neuroactivity. In fact, understanding information processing and network functionality is already a daunting task even in healthy brains. Secondly, there are pathological devel- opments in spatial scales inaccessible to non-invasive techniques. At a cellular level, for instance, the presence of Focal Axonal Swellings (FAS) is a critical biomarker for neurodegeneration and traumatic brain injuries. Differences in their geometrical ultrastructure can affect neuronal signaling and degrade the information encoded in spike trains. Our recent estimates highlight potential axonal trouble spots and provide a window of opportunity for theoretical and computational exploration of their collective effects in impaired neuronal net- works. 9 Theoretical and computational models may -- with all their limitations -- link injured neuronal networks to observable psychophysical deficits. Although much remains to be explored, the associate memory model based upon the classic work of Hopfield subject to FAS represents the first fundamental, biophysically- inspired theoretical model addressing cognitive deficits due to injury. 4 Methods and Materials 4.1 Hopfield network models for associative memory J. Hopfield made seminal contributions to the study of collective properties that emerge on systems of equivalent components (or neurons). He developed a model to describe content-addressable memory in an appropriate phase space for neuronal networks. The model incorporated aspects of neurobiology and its underlying neuronal circuitry. Within this framework, he was able to study properties such as familiarity recognition, categorization, error correction, and time sequence retention. Our computational model is based on J. Hopfield's original publications [Hopfield(1982), Hopfield(1984), Hopfield, Tank(1985)] and more recent extensions [Gerstner et al(2014), Benna, Fusi(2015)]. 4.1.1 Standard Hopfield Model In Hopfield's original model, a neuronal network is composed of N neurons that attend binary states Si ∈ {−1, 1}. The connections between neurons are responsible for information transference and processing in the network. They are represented by weights wij (linking neurons i and j), and stored in a connectivity matrix W = (wij). In this setting, the neuronal states evolve in time according to [Hopfield(1982), Hopfield(1984), Hopfield, Tank(1985)] where the gain function g is given by dSi(t) =Xj g(x) =(1, −1, wij · g(Sj(t))dt x ≥ 0 x < 0 . (3) (4) The most important property of the model is the ability to encode memories as fixed points of the system. When a noisy input is presented, it converges to the closest fixed point (closest known concept) in a process commonly referred as memory association. 4.1.2 Extended Hopfield Model In our extended Hopfield model, neurons may achieve multiple discrete states [Gerstner et al(2014), Benna, Fusi(2015)] Si ∈ {0, 1, ..., s − 1, s}. 10 The dynamical evolution of the system is also governed by a more sophisticated equation: 5) , ( (6) + µ · dBi noise {z } } with sigmoid gain function g given by g(x) = 0.5(1 + tanh(βx)). The constant τ gives the time-scale of the dynamics. Direct inputs for neuron i (e.g. external stimuli) are represented by Ii(t). The term Bi corresponds to a Wiener Process with intensity µ, and is a proxy for stochastic fluctuations in the firing rates. The (continuous) states are ultimately rounded to the nearest discrete state by a scaling function m(t) = max 1≤i≤N Si(t) m(t) (cid:21) Si(t) =(cid:20) s · Si(t) (7) (8) The resulting stochastic differential equation takes the following form when discretized: Si+1 = Si + ∆t f1(Si) + f2(Si) dBi (9) with f1(Si) = −τ−1Si +Xj wijg(Sj) + Ii , f2(Si) = µSi and Brownian increment dBi = Bi(t + ∆t) − Bi(t). We solved the system numerically using the Euler-Maruyama Method [Higham(2001)] and made all our codes available. As noted earlier, FAS may alter the activity state of an injured neuron. Thus, a broader class of states is required to study effects of more sophisticated mech- anisms. Whereas blockage of spike trains can be described in binary networks, more complex feature discriminations like reflection and filtering, that transfer partial or altered information, require a multi-state setting. For this, we con- verted our sample set of facial images [Weyrauch et al(2004)] from originally colored to nine distinct shades of gray (s = 8) in addition to the (standard) binary black-white states. With this extension, the computational network can also discriminate images of human faces that naturally share several features (high overlap). 4.2 Focal Axonal Swellings following brain injuries Focal Axonal Swellings are an ubiquitous pathological feature to several brain disorders [Xiong et al(2013), Coleman(2005), Tang-Schomer et al(2010), Tang- Schomer et al(2012), McArthur et al(2004), Bayly(2006b)]. Recent theoretical 11 dSi(t) = −τ−1 · Si(t)dt } {z self-dynamics + Ii(t)dt {z } +Xj wij · g(Sj(t))dt {z external input input from other neurons Fig 3. Schematics for injurious effects of Focal Axonal Swellings (FAS) to spike trains according to the theoretical framework of Maia and Kutz [Maia, Kutz(2014a), Maia, Kutz(2014b), Maia et al(2015)]. Spike trains may be distorted in four qualitatively distinct regimes depending on geometrical parameters associated to FAS: transmission (β1), filtering (β2), reflection (β3) or blockage (β4). See references and text for more details. Injuries can then be characterized by the transfer function S = F (S, β) where S is the effective firing rate (state) after the FAS, S is the firing rate (state) before. Such distortions are incorporated to injured neurons in our Hopfield network simulations in conjunction with experimental morphometric results from Wang et al. [Wang et al(2011)] and Dikranian et al. [Dikranian et al(2008)]. results suggest how morphological parameters affect the propagation of spike trains and thus, disrupt functional and coordinated network activity. In this section we provide details on how we introduce such effects in our Hopfield- networks and detail the experimental work used to calibrate the parameters of our model. 4.2.1 Theoretical framework for FAS effects Maia and Kutz developed in a series of papers a theoretical framework for char- acterizing the anomalous effects of FAS to spike propagation [Maia, Kutz(2014a), Maia, Kutz(2014b),Maia et al(2015)]. We review their main results and explain how to add such pathologies (or their proxies) into account for the firing-rate dynamics of neuronal networks (see the schematics in Fig. 3). The authors distinguish axonal enlargements that lead to minor changes in propagation (β1) from those that result in critical phenomena such as colli- sions, reflections or blockage of traveling spikes (β2, β3 and β4). They use three geometrical parameters (δB, δT , δA) to model a prototypical shaft enlargement and characterize all possible propagation regimes in an unmyelinated action potential model. The regimes can be distinguished by evaluating a (simple) function of the FAS geometrical parameters inferred through numerical simu- 12 lations. They suggest that evaluating this function along axon segments can help detect regions most susceptible to (i) transmission failure due to pertur- bations, (ii) structural plasticity, (iii) critical swellings caused by brain traumas and/or (iv) neurological disorders associated with the break down of spike train propagation. Swellings typically delete spikes by a mechanism called filtering (β2), when a first spike changes its profile at the axonal enlargement region and a close second spike interacts with its refractory period. As a consequence, the second spike is deleted in a mechanism of the so-called pile-up collision (see [Maia, Kutz(2014b)] for details). Distorted spike trains do not match their corresponding original firing rates (as illustrated in Fig. 3). Instead, they are confused with lower rates, which decrease the system's overall denoising abilities. We simulate the harmful effects of filtering by implementing a statistical version of the confusion matrix from the same source, that in simple terms, evaluates the probability that state i gets confused as state j due to the FAS. A less frequent mechanism of spike deletion is reflection (β3). There, a travel- ing pulse is divided into two pulses when it reaches the FAS: one propagating forwards and the other propagating backwards. The backward pulse will collide with the next spike and have them both deleted. Thus, only a fraction of the original encoded information is ultimately transmitted by the spike train. We add this effect in our neuronal network by halving the firing-rate of an injured neuron in this regime. Figure 3 finally illustrates the blockage of spikes (β4) that occur typically in regions of more dramatic axonal enlargement. In this scenario, no information is transmitted through the damaged axon and the neuron cannot adapt and play its role in the desired collective dynamics: it remains in its initial state. In our neuronal network model, a significant amount of neurons in the blockage regime causes blurs in the reconstructed concepts (memories) and therefore decreases the accuracy of the recalled information. We modeled this mechanism by introducing non-adapting neurons into the network, that keep their (possibly noise-affected) initial state over time. Traumatic Brain Injuries and neurodegenerative diseases induce FAS with tremen- dous variety of shapes and, consequently, with different functional deficits re- garding spike train propagation. Thus, we consider a distribution of different FAS mechanism where fractions of neurons are affected by blockage, reflection and filtering (confusion) respectively. 4.2.2 TBI/FAS data from adult rats (Wang et al. [Wang et al(2011)]) We use the data from Wang et al. [Wang et al(2011)] to calibrate the distri- butions of FAS following TBI in our simulations. In their study, the authors induce axonal injury in the optic nerve of a rat via central fluid percussion injury. They follow the progression of proximal and distal axonal injuries with biolog- ical markers for swellings. Transgenic animal models were used to better high- light damaged axonal segments with immunocytochemical markers. Through the use of qualitative and quantitative imaging approaches, they confirm and 13 Fig 4. Distribution of functional impairments in injured axons following TBI experiments from Wang et al. [Wang et al(2011)]. Panel A: we generate ovoid/spheroid FAS with areas compatible with the experimental distribution. The geometrical parameters of the FAS define the spike propagation regime. Panel B: we generate 12 FAS (column) for each injured axon (row) and order them from worst to best case scenario (upper "flags"). We assume that the worse FAS within an injured axon dominates the others, and classify the entire axon within that category (intermediary "flags"). This leads to the (bottom) pie-charts of impairments for an injured neuronal population. See text for more details. significantly extend previous data supporting the pathogenesis of traumatic ax- onal injury. Importantly, they provide unprecedented insight into this complex pathology, demonstrating post-TBI axonal swelling and disconnection. Within this setup, any potential axonal change induced by the traumatic event could be easily and consistently visualized via routine fluorescence and/or confocal microscopy. In this work we are especially interested in the distributions of shape, number and area of FAS from Wang et al. [Wang et al(2011)]. The number and area of FAS vary with time and spatial localization along the optic nerve. This allowed us to reconstruct distributions of total area per swelling and their functional deficits (blockage, reflection and filtering). See Fig. 4 for details. Wang et al. [Wang et al(2011)] describe a "spheroid" as a common type of FAS shape. We model this particular type of FAS with ovoids (see Fig. 4A) that include the three geometrical parameters (δB, δT , δA). The total area of each generated FAS is drawn from the distribution of area per swelling depicted in Fig. 4A (within one standard deviation). The authors divided the optic nerve into 12 spatial grids: distal (D1-D6) and proximal (P1-P6). We generate one swelling per grid for each injured axon and order them from worst to best case scenario: Fig. 4B 14 (upper "flag-charts") should be interpreted as a family of 20 injured axons (rows) with 12 FAS each (columns). Each FAS has a functional deficit (transmission, filtering, reflection or blockage) according to its geometrical parameters. Finally, we assume that the worst swelling within an axon (row) dominates and classifies the entire axon within that category (Fig. 4B, intermediate "flag-charts"). This lead to the pie-charts indicating the fraction of axons displaying transmission (green), filtering (yellow), reflection (red) and blockage (black). There are several drawbacks with this methodology, but we believe that our distributions are biophysically reasonable and compatible with available data. Better results could be obtained if the authors used the recently diagnostic tools developed in [Maia et al(2015)]. 4.2.3 TBI/FAS data from infant mice (Dikranian et al. [Dikranian et al(2008)]) Traumatic Brain Injuries (TBI) are responsible each year for millions of chil- dren hospitalizations [McArthur et al(2004)]. Infant brains (0-2 years) are highly susceptible to trauma [Schneier et al(2006)], and injuries sustained during this critical development period leads to profound neuronal and axonal degenera- tion [Gleckman et al(1999), Ewing-Cobbs(2000), Geddes et al(2001)]. K. Dikra- nian et al. [Dikranian et al(2008)] examined analogous injuries in the infant P7 mouse due to its physiological similarities with human children, inflicting con- trolled impact to the skull with an electromagnetic device based on a moving coil actuator. Animals were sacrificed at 30 min, 5h, 16h and 24h post-injury and well-established methodologies [Bayly(2006b)] were used to characterize the spatiotemporal pattern of axonal degeneration and cell death in the cin- gulum/external capsule region ipsilateral to the site of trauma. The Cingulate Cortex plays a critical role in memory performance [Vann et al(2012), Malin, McGaugh(2006), Weible et al(2012)], and they suggest that early degeneration there may disconnect cortical and thalamic neurons, leading to their apoptotic death. Dikranian's morphometric analysis [Dikranian et al(2008)] -- that included de- tailed swelling diameter measurements -- allowed us to realistically incorporate injury effects in our Hopfield network simulations. We are especially inter- ested in their distributions of shape, number and area of FAS. Notice that they provided distribution for the diameters of the spheroids, which allowed us to generate FAS in a much more direct way. The corresponding "flag-charts" and pie-charts for their distributions are depicted in Fig. 5. We believe these distri- butions to be more relevant than the ones obtained using the data from Wang et al. [Wang et al(2011)] since the cingulum is recognized as a region that plays a fundamental role in memory association. Thus, damage in this area can be also physiologically linked to memory impairments. 15 Fig 5. Distribution of functional impairments in injured axons following TBI experiments from Dikranian et al. [Dikranian et al(2008)]. We generate ovoid/spheroid FAS following the reported experimental distribution of FAS diameters. The geometrical parameters of the FAS define the spike propagation regime. We generate 5 FAS (column) for each one of the 40 injured axons (row) and order them from worst to best-case scenario (upper "flags"). We assume that the worse FAS within an injured axon dominates the others, and classify the entire axon within that category (intermediary "flags"). This leads to the (bottom) pie-charts of impairments for an injured neuronal population. See text for more details. 4.3 Implementation of memory storage. To simulate a face recognition task, the set of memories has to be learned by the network. For this, we encode them in the weights of the neuronal connections as specified by the weight matrix of the network: We consider a system of weighted neurons. The strength wij of the connection between neuron i and neuron j, described by the weight of the respective edge, characterizes the information transfer from i to j. Stored in the connectivity matrix W = (wij)1≤i,j≤N , they characterize the network's dynamics and encode the set of known concepts corresponding to the system's fixed points. The weight matrix is constructed from the training set of memories represented 16 → W = C T C (10) The theoretical storage capacity of a (Standard) Hopfield network of size N is 0.14N random patterns. In this study, we use a much smaller set of memo- ries, respectively five and three. This is due to the fact, that we store highly correlated facial images as opposed to random patterns. They have a pairwise overlap of 60% due to the structural similarity of faces. The high correlation pf the memories significantly decreases the storage capacity and therefore requires the choice of a small set of memories. We choose a setting with highly correlated memories to demonstrate the effects of memory confusion arising from FAS as described earlier. 4.4 Recognition score for network performance We developed a recognition score that measures recognition abilities with respect to significance and accuracy in recalling previously stored memory patterns (see Fig. 6). We assume the existence of an ideal observer (cf. Benna and Fusi [Benna, Fusi(2015)]), that knows the whole set of memories and the original pattern underlying the current noisy input. Our recognition score takes the place of this observer by evaluating the current network state against all memorized patterns. In what follows, we describe the computational steps of the recognition algorithm: δ j−j<1 X j 1 N mµ = (i) Overlap: We determine the overlap between the current network state and the set of stored memories by calculating the respective overlap mµ ∈ {0, 1} of individual neuronal states: (ii) Recognition and Significance: After a pre-defined number of time steps (system's parameter), the network's states are matched to the closest pat- tern, i.e., we determine the index i ∈ 1, ..., M , such that di = mµ orig − mµ i is minimal. If the output pattern matches the original one (i ≡ orig), we say that recognition occurs. Otherwise, we speak of confusion of the memories (concepts). The classification is considered significant only if di − dj < t ∀j = 1, ..., M ; (j 6= i), where t is a threshold parameter. With this scheme, we classify the mem- ory recall into four groups and assign (numerical) labels. 17 as network states: C := face 1 . . . . . . . . . . . . face 2 . . . . . . . . . . . . face M . . .        Fig 6. Calculation of recognition scores for measuring memory performance (see Hopfield Recognition Toolbox, current version available at GitHub: https://github.com/MelWe/hopf-recognition). We use the Hamming distance mµ i to measure the overlap between the current network state and the fixed points corresponding to known facial images. Confusion or recognition is characterized by mµ i : if the overlap with the correct facial image is highest, we speak of recognition, otherwise of confusion. A threshold for the difference between the highest and second highest overlap determines whether the recognition or confusion was significant. According to this classification, we assign color labels to each trial which can be displayed in a heat map. (iii) Evaluation: The recognition score was developed to evaluate the memory performance of our Hopfield neuronal network model over a broad range of injury and initial noise. For each pair of parameters (inj, noise) we calculate the score as value of the significance label scaled by the accuracy of the recognition (overlap mµ). The final result is a heat map (see Fig 2) that links recognition score, memory performance and noise handling to different levels of injury. Acknowledgments MW was supported by a scholarship of the Konrad-Adenauer-foundation. References Gerstner et al(2014). Gerstner W, Kistler WM, Naud R, Paninski L (2014) Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition. Memory and attractor dynamics, (Cambridge University Press) p.442-66 18 Hopfield(1982). Hopfield JJ (1982) Neural networks and physical systems with emergent collective computational properties, Proc. Nat. Acad. Sci. (USA) 79, p. 2554-2558. Hopfield(1984). Hopfield JJ (1984) Neurons with graded response have collec- tive computational properties like those of two-sate neurons, Proc. Nat. Acad. Sci. (USA) 81, p. 3088-3092. Hopfield, Tank(1985). Hopfield JJ, Tank DW (1985) Neural computation of decisions in optimization problems, Biological Cybernetics 55, p. 141-146. Hopfield(2015). Hopfield JJ (2015) Understanding Emergent Dynamics: Us- ing a Collective Activity Coordinate of a Neural Network to Recognize Time-Varying Patterns. Neural Comp., vol. 27, pp. 2011-2038 Higham(2001). Higham DJ (2001) An Algorithmic Introduction to Numerical Simulation of Stochastic Differential Equations, SIAM Review 43(3), p. 525-546. Maia, Kutz(2014a). Maia PD, Kutz JN (2014) Compromised axonal function- ality after neurodegeneration, concussion and/or traumatic brain injury, J Comput Neurosci 37, pp. 317-332. Maia, Kutz(2014b). Maia PD, Kutz JN (2014) Identifying critical regions for spike propagation in axon segments, Journal of Computational Neuro- science, vol. 36(2), pp. 141-155 Maia et al(2015). Maia PD et al. (2015) Diagnostic tools for evaluating the impact of Focal Axonal Swellings arising in neurodegenerative diseases and/or traumatic brain injury. J Neurosci Methods, vol. 253, pp. 233-43 Weyrauch et al(2004). Weyrauch B, Huang J, Heisele B, Blanz V (2004) Component-based Face Recognition with 3D Morphable Models, First IEEE Workshop on Face Processing in Video (Washington, D.C.) Dikranian et al(2008). Dikranian K et al. (2008) Mild traumatic brain injury to the infant mouse causes robust white matter axonal degeneration which precedes apoptotic death of cortical and thalamic neurons. Exp. Neuro., vol. 211, pp. 551-560 Wang et al(2011). Wang J, Hamm RJ, Povlishock JT. (2011) Traumatic ax- onal injury in the optic nerve: evidence for axonal swelling, disconnection, dieback and reorganization, Journal of Neurotrauma 28(7), p. 1185-1198. McEliece et al(1987). McEliece RJ, Posner EC, Rodemich ER, Venkatesh SS (1987) The Capacity of Hopfield Associative Memory, IEEE Trans. on Inf. Theory vol. IT-33(4), p. 461-482 Cerf et al(2010). Cerf M et. al. (2010) On-line, voluntary control of human temporal lobe neurons, Nature vol. 467, p. 1104-1108 19 Sekigawa et al(2013). Sekigawa A et. al. (2013) Diversity of Mitochon- drial Pathology in a Mouse Model of Axonal Degeneration in Synucle- inopathies, Hindawi Publishing Corporation Oxidative Medicine and Cel- lular Longevity vol. 2013, article ID 817807 Ballard et al(2011). Ballard C et. al. (2011) Alzheimer's disease, Lancet 2011 (377), p. 1019-31 Xiong et al(2013). Xiong Y, Mahmood A, Chopp M (2013) Animal models of traumatic brain injury, Nature Reviews Neuroscience vol. 14(22), pp. 128-142 Coleman(2005). Coleman M (2005) Axon degeneration mechanisms: com- monality amid diversity, Nature Reviews Neuroscience vol. 6(11), pp. 889-898 Fainaru, Fainaru(2013). Fainaru-Wada M, Fainaru S (2013) League of denial: The nfl, concussions, and the battle for truth, Crown Archetype Tang-Schomer et al(2010). Tang-Schomer MD et. al. (2010) Mechanical breaking of microtubules in axons during dynamic stretch injury under- lies delayed elasticity, microtubule disassembly, and axon degeneration, The FASEB Journal, vol. 24(5), pp. 1401-1410 Tang-Schomer et al(2012). Tang-Schomer MD et. al. (2012) Partial interrup- tion of axonal transport due to microtubule breakage accounts for the formation of periodic varicosities after traumatic axonal injury, Experi- mental Neurology, vol. 233, pp. 364-372 Krstic, Knuesel(2012). Krstic D, Knuesel I (2012) Deciphering the mech- anism underlying late-onset alzheimer disease, Nature Reviews Neuro- science, vol. 9(1), pp. 25-34 Galvin et al(1999). Galvin JE et. al. (1999) Axon pathology in parkinsons disease and lewy body dementia hippocampus contains a -, b -, and g - synuclein Proceedings of National Academy of Science, vol. 96, pp. 13450- 13455 Hauser et al(2006). Hauser SL, Oksenberg JR (2006) The neurobiology of multiple sclerosis: Genes, inflammation, and neurodegeneration. Neuron, vol. 52(1), pp. 61-76 Yuste(2015). Yuste R (2015) From the neuron doctrine to neural networks, Nature Reviews Neuroscience, vol. 16, pp. 487-497 McCulloch, Pitts(1943). McCulloch WS, Pitts W (1943) A logical calculus of the ideas immanent in nervous activity, Bull. Math. Biophys., vol. 5, pp. 115-133 Hebb(1949). Hebb DO (1949) The Organization Of Behaviour, Wiley 20 Cajal, Estructura(1888). Cajal R, Estructura S (1888) De los centros nerviosos de las aves, Rev. Trim. Histol. Norm. Part 1, pp. 1-10 (in Span- ish) Bullock et al(2005). Bullock TH et al. (2005) Neuroscience. The neuron doc- trine, redux. Science vol. 310, pp. 791-793 Squire et al(2007). Squire LR, Wixted JT, Clark RE (2007) Recognition memory and medial temporal lobe: A new perspective. Nat. Rev. Neuro. vol. 8, pp. 872-881 Hu et al(2015). Hu SG et al. (2015) Associative memory realized by a re- configurable memristive Hopfield neural network. Nat. Comm., DOI: 10.1038/ncomms8522 Nowicki, Siegelmann(2010). Nowicki D, Siegelmann H (2010) Flexible Kernel Memory. PlosOne, vol. 5(6) McArthur et al(2004). McArthur DL, Chute DJ, Villablanca JP (2004) Mod- erate and severe traumatic brain injury: epidemiologic, imaging and neu- ropathologic perspectives. Brain Pathol. vol. 14, pp. 185-194 Schneier et al(2006). Schneier AJ et al. (2006) Incidence of pediatric trau- matic brain injury and associated hospital resource utilization in the United States. Pediatrics, vol. 118, pp. 483-492 Gleckman et al(1999). Gleckman AM et al. (1999) Diffuse axonal injury in infants with nonaccidental craniocerebral trauma: enhanced detection by beta-amyloid precursor protein immunohistochemical staining. Arch. Pathol. Lab. Med., vol. 123, pp. 146-151 Ewing-Cobbs(2000). Ewing-Cobbs L (2000) Acute neuroradiologic findings in young children with inflicted or noninflicted traumatic brain injury. Childs Nerv. Syst., vol. 16, pp. 25-33; discussion 34 Geddes et al(2001). Geddes JF et al. (2001) Neuropathology of inflicted head injury in children. II. Microscopic brain injury in infants. Brain, vol. 124, pp. 1299-1306 Bayly(2006b). Bayly PV (2006b) Spatiotemporal evolution of apoptotic neu- rodegeneration following traumatic injury to the developing rat brain. Brain Res., vol. 1107, pp. 70-81 Hemphill(2015). Hemphill MA (2015) Traumatic Brain Injury and the Neu- ronal Microenvironment: A Potential Role for Neuropathological Mechan- otransduction. Neuron, vol. 85 Weible et al(2012). Weible AP et al. (2012) Neural Correlates of Long-Term Object Memory in the Mouse Anterior Cingulate Cortex. J. Neuro., vol. 32(16), pp. 5598-5608 21 Vann et al(2012). Vann SD, Aggleton PJ, Maguire EA (2012) What does the retrosplenial cortex do? Nat. Rev. Neuro., vol. 10, pp. 792-800 Malin, McGaugh(2006). Malin EL, McGaugh JL (2006) Differential involve- ment of the hippocampus, anterior cingulate cortex, and basolateral amygdala in memory for context and footshock. PNAS, vol. 103(6), pp. 1959-1963 McDowell et al(1997). McDowell S, Whyte J, D'Esposito (1997) Working memory impairments in traumatic brain injury: evidence from a dual- task paradigm. Neuropsychologia, vol. 35(10). pp. 1341 1353 Schacter(1987). Schacter DL (1987) Memory, amnesia, and frontal lobe dys- function. Psychobiology, vol.15(1), pp. 21-36 Petrides(1985). Petrides M (1985) Deficts on conditional associative learning after frontal- and temporal-lobe lesions in man. Neuropsychologia, vol. 23(5), 601-614 Shallice, Evans(1978). Shallice T, Evans ME (1978) The involvement of the frontal lobes in cognitive estimation. Cortex, vol. 14, pp. 294-303 Mcallister(2004). Mcallister TW (2004) Working memory deficits after trau- catecholaminergic mechanisms and prospects for matic brain injury: treatment - a review. Brain Injury, vol. 18(4), pp. 331-350 Lyeth(1990). Lyeth BG (1990) Prolonged memory impairment in the absence of hippocampal cell death following traumatic brain injury in the rat. Brain Research, vol. 526(2), pp. 249-258 Knutson(2012). Knutson AR (2012) Visual discrimination performance, memory, and medial temporal lobe function. Proc. Nat. Acad. Sci., vol. 109(32), pp. 13106-11 Selkoe(2001). Selkoe DJ (2001) Alzheimer's Disease: Genes, Proteins, and Therapy. Physiological Review, vol. 81(2) McKhann et al(1984). McKhann G et al. (1984) Clinical diagnosis of Alzheimer?s disease: Report of the NINCDS-ADRDA Work Group under the auspices of Department of Health and Human Services Task Force on Alzheimer's Disease. Neurology, vol. 34, pp. 939?944 Cole, Bailie(2016). Cole WR, Bailie JM (2016) Neurocognitive and Psychi- atric Symptoms following Mild Traumatic Brain Injury. Frontiers in Neu- roscience: Translational Research in Traumatic Brain Injury, Chapter 19 McGee(2016). McGee J (2016) Traumatic Brain Injury and Behavior: A Practical Approach. Neurol Clin., vol. 34(1, upcoming 2016 Feb), pp. 55-68 22 Benna, Fusi(2015). Benna MK, Fusi S (2015) Computational principles of biological memory. arXiv:1507.07580v1 Faul et al(2010). Faul M et al. (2010) Traumatic brain injury in the united states: emergency department visits, hospitalizations, and deaths. At- lanta (GA): Centers for Disease Control and Prevention, National Center for Injury Prevention and Control. Jorge et al(2012). Jorge RE et al. (2012) White matter abnormalities in vet- erans with mild traumatic brain injury. Amer. J. of Psychiatry, 169(12), pp. 1284 -- 1291. Anafi, Bates(2010). Anafi RC, Bates JHT (2010) Balancing Robustness against the Dangers of Multiple Attractors in a Hopfield-Type Model of Biological Attractors. PLoS ONE, vol. 5(12): e14413. Menezes, Monteiro(2011). Menezes RA, Monteiro LHA (2011) Synaptic com- implications for memory rehabilitation. pensation on Hopfield network: Neural Comput & Applic, vol. 20(753). Ruppin, Reggia(2011). Ruppin E, Reggia JA (1995) Patterns of Functional Damage in Neural Network Models of Associative Memory. Neural Com- putation, vol. 7(5), pp. 1105 -- 1127. 23
1111.6062
1
1111
2011-11-25T17:35:01
Simple, Fast and Accurate Implementation of the Diffusion Approximation Algorithm for Stochastic Ion Channels with Multiple States
[ "q-bio.NC" ]
The phenomena that emerge from the interaction of the stochastic opening and closing of ion channels (channel noise) with the non-linear neural dynamics are essential to our understanding of the operation of the nervous system. The effects that channel noise can have on neural dynamics are generally studied using numerical simulations of stochastic models. Algorithms based on discrete Markov Chains (MC) seem to be the most reliable and trustworthy, but even optimized algorithms come with a non-negligible computational cost. Diffusion Approximation (DA) methods use Stochastic Differential Equations (SDE) to approximate the behavior of a number of MCs, considerably speeding up simulation times. However, model comparisons have suggested that DA methods did not lead to the same results as in MC modeling in terms of channel noise statistics and effects on excitability. Recently, it was shown that the difference arose because MCs were modeled with coupled activation subunits, while the DA was modeled using uncoupled activation subunits. Implementations of DA with coupled subunits, in the context of a specific kinetic scheme, yielded similar results to MC. However, it remained unclear how to generalize these implementations to different kinetic schemes, or whether they were faster than MC algorithms. Additionally, a steady state approximation was used for the stochastic terms, which, as we show here, can introduce significant inaccuracies. We derived the SDE explicitly for any given ion channel kinetic scheme. The resulting generic equations were surprisingly simple and interpretable - allowing an easy and efficient DA implementation. The algorithm was tested in a voltage clamp simulation and in two different current clamp simulations, yielding the same results as MC modeling. Also, the simulation efficiency of this DA method demonstrated considerable superiority over MC methods.
q-bio.NC
q-bio
SIMPLE, FAST AND ACCURATE IMPLEMENTATION OF THE DIFFUSION APPROXIMATION ALGORITHM FOR STOCHASTIC ION CHANNELS WITH MULTIPLE STATES Patricio Orio1, Daniel Soudry2,3 (1) Centro Interdisciplinario de Neurociencia de Valparaíso, Facultad de Ciencias, Universidad de Valparaíso, 2360102 Valparaíso, Chile. (2) Department of Electrical Engineering, Technion, Haifa , Israel (3) Laboratory for Network Biology Research, Technion, Haifa , Israel Address for correspondence: Dr. Patricio Orio Centro Interdisciplinario de Neurociencia de Valparaíso Universidad de Valparaíso Gran Bretaña 1111 2360102 Valparaiso, Chile Tel: 56-32-2508185 Fax: 56-32-2508047 e-mail: [email protected] 1 Background ABSTRACT The phenomena that emerge from the interaction of the stochastic opening and closing of ion channels (channel noise) with the non-linear neural dynamics are essential to our understanding of the operation of the nervous system. The effects that channel noise can have on neural dynamics are generally studied using numerical simulations of stochastic models. Algorithms based on discrete Markov Chains (MC) seem to be the most reliable and trustworthy, but even optimized algorithms come with a non-negligible computational cost. Diffusion Approximation (DA) methods use Stochastic Differential Equations (SDE) to approximate the behavior of a number of MCs, considerably speeding up simulation times. However, model comparisons have suggested that DA methods did not lead to the same results as in MC modeling in terms of channel noise statistics and effects on excitability. Recently, it was shown that the difference arose because MCs were modeled with coupled activation subunits, while the DA was modeled using uncoupled activation subunits. Implementations of DA with coupled subunits, in the context of a specific kinetic scheme, yielded similar results to MC. However, it remained unclear how to generalize these implementations to different kinetic schemes, or whether they were faster than MC algorithms. Additionally, a steady state approximation was used for the stochastic terms, which, as we show here, can introduce significant inaccuracies. Main Contributions We derived the SDE explicitly for any given ion channel kinetic scheme. The resulting generic equations were surprisingly simple and interpretable – allowing an easy, transparent and efficient DA implementation, avoiding unnecessary approximations. The algorithm was tested in a voltage clamp simulation and in two different current clamp simulations, yielding the same results as MC modeling. Also, the simulation efficiency of this DA method demonstrated considerable superiority over MC methods, except when short time steps or low channel numbers were used. 2 INTRODUCTION Noise and variability are present throughout the nervous system, from sensory systems to the motor output and perhaps more importantly in the higher brain areas [1]. Far from being considered as a nuisance, noise is now understood as one of the key elements that shape the way the central nervous system (CNS) codes sensory inputs, builds internal representations and makes decisions [2]. Phenomena like stochastic resonance [3,4,5,6] enhance several aspects of sensory coding and signal detection [7,8]. Also, noise can be beneficial in various computational tasks [9,10,11,12]. One of the main sources of noise and variability is the stochastic opening and closing of ion channels, commonly called channel noise [13,14]. The effects of channel noise on neuronal excitability are to a large extent studied with the use of mathematical models, either by constructing and analyzing models with stochastic channels [e.g. 15,16,17,18] or by introducing a noisy conductances in dynamic clamp experiments [19,20]. It is of interest, then, to develop and analyze numerical models that faithfully reproduce the stochastic nature of ion channels. It is also of interest to develop fast algorithms that can be used in large scale simulations of neural networks or in real time simulation for dynamic clamp experiments. Ion channels are commonly modeled using the framework established by Hodgkin and Huxley [21, see also 22]. In this framework, ion channels contain one or more activation subunits that can be either in a resting or active state. The transition rates between states are voltage- dependent, and now we know that this is because these subunits contain a charged domain (the voltage sensor) that senses the membrane electrical potential [23]. In the pure Hodgkin and Huxley (HH) framework, the probability of a channel being open is equal to the probability of all its activation subunits being active. Usually the subunits are assumed to be independent and thus the probability of the open channel is the product of the probabilities of the active subunits. In the limit of infinitely many channels (deterministic HH model), probabilities are equivalent to the fraction of active subunits or open channels. The transition between resting and active states of subunits is described by ordinary differential equations of a deterministic nature, because the HH model fitted the behavior of a giant squid axon with such a large number of channels that individual stochastic contributions were completely neglected. When the stochastic behavior of ion channels is taken into account, it is best described by continuous-time, discrete state Markov jumping processes [24,25]. Several algorithms exist for the mathematical simulation of simultaneous and independent Markov Chains (MCs) representing a population of ion channels in a membrane patch or neuronal soma. Among these, the most efficient is a channel-number-tracking algorithm proposed by Gillespie [26] and first applied to ion channels in 1979 [27] (see [28] for a comparison with other MC algorithms). Nevertheless, all MC algorithms increase their computational complexity with the number of channels and the channel- number-tracking algorithm may be difficult to implement for complex kinetic schemes. Another approach for simulating stochastic ion channels relies on the fact that a large number of simultaneous and independent MCs can be approximated by a stochastic differential equation that describes the time evolution of the fraction of MCs that are in each possible state [29,30,31,32,33]. This algorithm, referred as Diffusion Approximation (DA), is dramatically more 3 efficient in terms of computational cost [28] and is the choice for dynamic clamp experiments where real-time simulation is required [19]. In the general form of DA [29], the time evolution of a variable vector containing the fraction of channels in each state is obtained by solving a Langevin equation (see eq. (1)) with both deterministic and stochastic transition matrices. The method, however, is less practical, since it requires the numerical calculation of a matrix square root at about each time step, making it a very time-consuming algorithm (each calculation usually requires ( )3 O M floating point operations [34], M being the number of channel states). To circumvent this, Fox and Lu [29] heuristically proposed to simulate the two-state activation subunits as separate stochastic processes and then calculate the conductance of each ion channel species as the product of subunit probabilities. This approach of uncoupled activation subunits requires a simple Stochastic Differential Equation (SDE) per subunit species without any matrix operation, easily constructed by adding simple noise terms to the deterministic differential equations of the mean channel kinetics. . This, in addition to its high computational efficiency, made the uncoupled subunits approach the main choice for DA implementations [18,19] However, the uncoupled subunits form of the DA does not approach the behavior of explicit MC appropriately. Mino and colleagues [28] found that this DA algorithm introduces less variability than MC modeling, evidenced as a shallower stimulus vs. action potential firing probability relationship. Later, Bruce [35] found that the DA algorithm, as it was being implemented, assumes that the stochastic term of the gating subunits is uncorrelated, while the MC modeling introduces correlated noise into the channel conductance behavior. Also, the variance of the conductance is higher for MCs than for the uncoupled subunits DA algorithm. Why was it assumed that activation subunit coupling is of minor importance when modeling stochastic channels? Mainly, because both approaches – coupled or uncoupled subunits – result in the same mean time evolution of the conductance. However, fluctuations introduced by both approaches are dramatically different, in terms of the variance of the conductances and their correlations at different times. This difference between approaches poses a serious problem since the purpose of any quantitative stochastic model is precisely to determine the effects of these fluctuations. The uncoupled subunits approach also has the disadvantage of not being applicable to kinetic schemes with non-independent activation subunits – such as channels with cooperative voltage sensors [36,37], or when the voltage sensors are not identical [38,39]. In recent works [33,40], it was further confirmed that considering coupled activation subunits produces more variability in the conductance and introduces noise with a particular covariance that cannot be reproduced by two-states models. Both works also proposed algorithms for the DA that better approached the results of MC modeling, in the context of the HH model. the square root of the stochastic diffusion matrix (an Goldwyn et al. [33] tested the general form of DA suggested by Fox [29], numerically computing ( )3 O M operation) at each time step, producing a very time-consuming algorithm. On the other hand, Linaro et al. [40] developed a set of SDEs that capture the statistical properties of the variations of conductance, adding it to the ion currents given by a deterministic model. 4 Here we present a different approach to derive the DA using basic probabilistic tools, for any given kinetic diagram of a channel. This derivations results in a practical, general and intuitive rules allowing for the accurate implementation of DA as a set of simple SDEs, with comparable (between simplicity to that of (inaccurate) uncoupled DA approach, allowing and efficient implementation ( )2 ( ) O M at each time step, depending on the number of kinetic O M and transitions). This makes the computational complexity of the stochastic algorithm comparable to that of the uncoupled DA approach and even the deterministic implementation that simply ignores the noise terms in the SDE. We thoroughly tested the proposed DA implementation, comparing its results to the behavior of explicit MC modeling in three different simulation tests: one under voltage clamp and two under current clamp. Notably, the methods previously suggested [33,40] displayed significant inaccuracies in two of these tests because they employ a steady-state approximation for the calculation of stochastic coefficients. Our method does not require such an approximation and therefore does not incur those errors. We also compare the computational efficiency and numerical stability of the algorithm for different numbers of channels and integration time steps, showing that in most cases DA will be algorithm of choice. Finally, we discuss how our method relates to other implementations previously published. 5 RESULTS We examine a specific population of N ion channels with M states, where the transition rate of a single channel from state j to state i is given by ijA . We define the rate matrix A to be = −∑ii A on the diagonal. In neuronal A ji j i ≠ ijA terms for all ≠i composed of these j , and also For brevity, we keep this voltage dependency implicit. We denote by models, these transition rates are usually voltage dependent (and so are also time-dependent). ix the fraction of channels in 1 = and it is common to use ix . Note that this normalization in order to reduce the number of variables [29,31,33,40]. However, here this each of the state, and by x a vector of ... + + x x M 1 substitution is not employed until the numerical implementation to make the algebraic operations easier. The DA proposed by Fox [29,31] for the stochastic dynamics of x leads to the following SDE x d dt where ξ is a vector of independent Gaussian white noise processes with zero mean and unit x A = + , ξ (1) S D , a square root of the diffusion matrix D (namely ). This matrix square root has been the main hindrance in the implementation of DA variance, A is the rate matrix, and =S D=⊤ SS [33]. If solved numerically in simulation time, it incurs a great computational cost, of order ( )3 O M at each time step. Interestingly, it is possible to obtain a direct analytical solution of =S D for certain kinetic schemes, such as the potassium channel scheme, prior to the simulation (we used Cholesky decomposition, see eq. (15) and below). However, it is not immediately clear how to do so for other schemes, such as the sodium channel scheme. We therefore explored a different derivation of the matrix S . Derivation of the Diffusion Approximation We denote X i Nx= i , the number of channels in state i , and X to be the corresponding vector. Recall that the channels are independent of each other and that transition rates are memoryless. j≠ Therefore, for all i ∆ ij ( t ) =     is a Random Variable (RV) composed of the sum of from state to state durin j i the number of channels switching ( ) + independent events (“trials”), in (2) , t t g d t ( ) = n X t j     which a channel either switched states, with probability of = p A dt ij , or did not switch states, with probability 1 − ijA dt (to first order in dt ). This entails that for all i j≠ , ( ) ij t∆ are independent and binomially distributed with ( ) = n X t j and = p A dt ij . Additionally, we define 6 ( ) ii t∆ 0 = . Assume the value of ( )tX is fixed. Denoting by ensemble) we use the properties of the binomial distribution and find the mean ⋅ the expectation (over the ∆ ij ( ) t ( ) = = p X t A dt j ij , ) ( ) t = n p ( 1 − p ) = ( ) X t A dt j ij ( 1 − A i j d t ) . ( ( ) ij t∆ are independent V ar ∆ ij (3) (4) (5) And the variance Since where 1 ij =δ N Cov ( j= , and 0 otherwise. ∆ ij ( ), t ∆ mk if i ) = ( ) t δ δ im jk V r a ( ∆ ij ( t ) ) , In the limit → ∞ → we get that n → ∞ and 0 , dt 0 p → for the binomial distribution of each ( ) ij t∆ . This allows us to approximate ( ) ij t∆ by a normal (Gaussian) distribution with both mean and variance equal to = np X A dt j ij (by the central limit theorem). In order to derive the SDE (eq. (1)), we need to assume that the Gaussian approximation is reasonable. Later, we confirm this numerically, as also did Linaro et al. [40] and Goldwyn et al. [33] (for example, this was numerically confirmed by [33] for channel numbers as low as N K 18, N= Na 60 = ). At each dt , iX changes according to the sum of channels entering and leaving state i ( ) ( ) ( ) ( ) ( ( ) ) ∑ X t X t dX t t t i i i ji j ( ) ( ) idX t ) is also normal, as a linear tX (the of vector of combination of independent normal RVs. Since the distribution of normal variables is entirely ( ) ij t∆ are all normal, then Assuming ∆ ij (6) ∆ . = + dt − − = d determined by their mean and covariance, we calculate them. We use eq. (3) to find the mean of eq. (6) ( ) ( ) dX t i i Xµ d = = ( ∑ j A ij X j − ( ) t A ji X i ) ( ) t d t . (7) Next, using eq. (5) we find the covariance 7 jm d X = − R ∆ ( ) t ( ), t ∆ ik ∆ ik Cov ( =Cov ) , =Cov i j  ( ∑   k  ∑   k  ∑ − Cov   k ( ∑ Cov k ( − Cov ∆ ( ∑ = k ( ), t ik ( ( ), t ( ) ( ( ) ( ) , dX t dX t j i ( ) ∑ ∆ ( ) , ( ) ∆ t t ki m  ∑   m  ∑ −   m ) + ( ), ( ) Cov t t ) ( − ( ) Cov t ) ( + ( ) Var t ji Var ∆ mj ∆ ik ∆ ik ∆ ij ∆ i ( ) t ( ) t δ ij δ ij ∆ ∆ = jm ji k  )   ∑ m ∑ m ( ) t ( ) t − ∆ mj ( ) t jm + ( ) t ( ) t ( ), t ( ), t ∆ ki ∆ ki ∆ ki Cov Cov ∆ mj ∆ ) )        ∑   k  ∑   k ( ∆ k i ( ), ∆ t ij ) ) − 2dt terms and dividing by dt we obtain ) ( ) , if t i , if i ) − ( ) Var t ( ), t ) Var + ( ) t A ki ∆ ki ∆ i j ( ) t ∆ . X X ( ( = j ji k i ) ( ) t Using the final line of eq. (5), neglecting (  ∑  =  ≠ k i −  A ji 1 dt , i j R X ( ) A ik X d Since we now know the mean of d components (eq. (8)), we can write j i X ( ) t − ( ) t A ij ( ) tX (eq. (7)) and the covariance between all of its ≠ j (8) d = X X µ d + R X d Z , (9) where Z is a vector of independent Gaussian RVs with mean zero and unit variance. To derive an 0 dt → , yielding =x X / N SDE for we divide eq. (9) by N and take the limit of x d dt x A = + , ξ S which is indeed eq. (1), with S D= D ij = 1 2 N dt ( , i j R X d , where     1 N = ) A x ik + ( ) t A ki ( ∑ ≠ k i − − ( ) ( ) A t A t x x j j i i ji x k i (10) ( ) t ) , if i , if i = ≠ j j . A Simpler Derivation of the Diffusion Approximation Now that we have the general expression for the diffusion matrix, and know its origin, we can procedures for matrix square root computation. The key idea behinds this is to use only devise a simple way to explicitly calculate S , which avoids the use of time consuming numerical ( ) ij t∆ and eqs. (3)- (6) to derive the SDE, and the Gaussian approximation. For simplicity, we demonstrate 3M = states this method step-by-step using a channel with 1 Α21 Α12 2 Α23 Α32 3 . 8 Using eq. (6) we write dX dX dX 1 2 3 W = ∆ − ∆ we notice that Denoting ji ij ij = ∆ − ∆ 12 21 = ∆ − ∆ + ∆ − ∆ 23 12 21 = ∆ − ∆ 32 23 ij∆ can be combined in opposing pairs 32 dX = dX W 12 1 = − + W W 12 23 = − W 23 dX 2 3 We now calculate the means, using ∆ ij ( ) t = ( ) X t A dt ij j (eq. (3) (7)), we obtain (11) (12) 2 1 = dX dX − X A dt X A dt 2 21 1 12 = − + − + X A dt X A dt X A dt X A dt 2 23 3 32 2 21 1 12 − dX X A dt X A dt 3 32 2 23 3 ( ) ( ) ( ) = − Y t W t W t Denoting ij ij ij dX , we obtain dX = = + + Y 23 , 1 dX d X 2 3 = = 1 X 2 d d X 3 Y 12 Y 12 Y 2 3 − − . ,Y Y are normal, independent, with zero mean and where 23 12 ( ( ) ) ( ) Y 12 ( ) ( ) ( ) Y 23 2dt terms. Now we can write where we used eq. (5), neglecting ∆ 12 ∆ 23 ∆ 21 ∆ = = Var Var Var Var Var Var + + = = 32 + X A dt X A dt 2 21 1 12 + X A dt X A dt 32 2 23 3 , dX dX 1 2 dX 3 = + − + Z X A dt X A dt X A dt X A dt 21 1 12 2 1 21 1 12 2 = − + − + X A dt X A dt X A dt X A dt 32 1 2 23 3 2 21 12 + + + Z X A dt X A dt Z X A dt X A dt 2 2 23 3 32 1 2 1 21 12 − − + X A dt X A dt Z X A dt X A dt 2 23 3 32 2 2 23 3 32 = − with ,Z Z are normal, independent, with zero mean and unit variance. 1 2 Dividng by N and taking the limit 0 dt → , we finally obtain the SDE = x A 2 12 − x A 1 21 + 1 N ξ 1 x A 2 12 + x A 1 21 = − x A 2 12 + x A 1 21 − x A 2 32 + x A 3 23 − 1 N ξ 1 x A 2 12 + x A 1 21 + 1 N ξ 2 x A 2 32 + x A 3 23 = x A 2 32 − x A 3 23 − 1 N ξ 2 x A 2 32 + x A 3 23 9 dx 1 dt dx 2 dt dx 3 dt Note that each component of ξ is associated with a transition pair i ) / ( signs. , and appears in the equations of + A x A x ji j ij dt and /idx N i j⇌ , multiplied by /jdx dt with opposite i Using a similar derivation we can now write S for a general channel with M states. To do this succinctly we must introduce several notations. We denote by T the set of all possible transitions ) pairs ( j⇌ that exist between states and then give each pair an index in k ( )1 / 2 M M − that T , the size of set T , can be any integer between 0 and ( )T i the index of the state connected by the k -th transition pair, excluding state i . to be the subset of all transitions pairs that connect to state i . Finally, we denote . Also, we denote . Note 1, ..., = T In that case, the matrix S is of size M T× , and S ik   =    sign ( − i m ik ) 1 N 0 A x im ik m ik + A x m i ik i ( ) ∈ k T i ( ) ∉ k T i , , Test Case – Potassium and Sodium Channels ikm to be (13) We have obtained the matrix S analytically, showing that it has a rather simple structure. It is necessary, however, to compare our result with previous definitions of the diffusion matrix as given by Fox [29,31] and used by Goldwyn [33]. For a simple comparison, we will use the case of the potassium channel (see the linear kinetic scheme for coupled subunits in Figure 1). Starting from eq. (1) and defining x  =  n  0 n 1 n 2 n 3 , the matrix KA is ⊤ n 4    0  − α 4     =       KS is defined such that β n 0 0 A K 0 n 4 α n 3 − α − β n n 2 β n 0 0 3 α n 2 − α − β 2 n n 3 β n 0 T S S K K D= [29], being 0 0 0 0 2 α n −α − β 3 n n 4 β n 0 α n 4 − β n            10 D = 1 N K n 1 n  α + β 4 n  0  − α − β 4 n  0  0       0 0 n 4 − α − β n 1 0 α + β + α + β 2 3 4 n n n 0 1 1 3 2 − α − β n n 1 2 1 0 0 n 2 0 3 − α − β 2 n 2 1 3 2 2 3 α + β + α + β n n n 1 2 2 2 3 − α − β n n 2 n 3 0 0 0 n 3 n 3 2 − α − β n 3 2 4 3 2 α + β + α + β n n n 2 3 3 −α − β 4 n n 3 4 n 4 0 0 0 4 −α − β n 3 α + β 4 n 3 n n 4 4            (14) (n sub indices in α and β were omitted for abbreviation). Using Cholesky decomposition, we can find KS : S K = 1 N K     −          4 n α 0 0 0 0 4 n α + β n 1 0 0 + β n 1 3 n α 1 + 2 β n 2 0 0 + 2 β n 2 2 n α 2 + 3 β n 3 0 0 0 − 3 n α 1 0 0 − 2 n α + 2 0 3 β n 3 n α 3 + 4 β n 4 − n α 3 + 4 β n 4 . 0 0 0 0 0              0 dt dn 1 dt dn 2 dt dn 3 dt dn 4 dt = = = = ( ( ( ( (15) Substituting in (1) and performing the matrix operations, the full system of SDE for the n variables can be now written as: ( ) = − α n α n β n β n dn ξ 1 + + + n n n n 1 4 4 0 1 0 1 N K 4 α n n 0 − β n n 1 − 3 α n n 1 + 2 β n n 2 ) − ξ 1 3 α n n 1 − 2 β n n 2 − 2 α n n 2 + 3 β n n 3 ) − ξ 2 2 α n n 2 − 3 β n n 3 − α n n 3 + 4 β n n 4 ) − ξ 3 1 N K 1 N K 1 N K 4 α n n 0 + β n n 1 + ξ 2 1 N K 3 α n n 1 + 2 β n n 2 (16) 3 α n n 1 + 2 β n n 2 + ξ 3 2 α n n 2 + 3 β n n 3 + ξ 4 1 N K 1 N K 2 α n n 2 + 3 β n n 3 α n n 3 + 4 β n n 4 α n n 3 − 4 β n n 4 ) − ξ 4 1 N α n n 3 + 4 β n n 4 , K where, again, ξ1, ξ2, ξ3 and ξ4 are independent Gaussian white noise terms with zero mean and unit variance. Note in (16) that although the length of the noise vector ξ is equal to the number of states, the number of noise terms actually employed is equal to the number of transition pairs j i + N A x ij A x ji multiplied by j⇌ . Also, as before, each component of ξ is associated with a transition pair i ( ) / /idx /jdx dt with opposite signs. Thus, the structure of equations we proposed is also obtained from the original definition of S . However, it is easy to see that Cholesky decomposition, which generates lower triangle matrices, ... M 2 will only work for “linear” kinetic schemes – 1 ⇌ ⇌ ⇌ . For a example, since a triangle , and then added to deterministic differential equations of j⇌ ; it is dt and i matrix must be square the Cholesky decomposition cannot work if M T< , as in the case of the 11 sodium channel, where 8M = and 10 T = . In that case, the S matrix we derive is different than that suggested by Fox [29,31] and used by Goldwyn et al. [33] – since in the latter approach the length of ξ was always equal to M , the number of states and not the number of transition pairs, as in our approach. With our approach, the SDE for sodium channels (see Supplemental Text T1) requires the use of 10 random terms instead of 8 (or 7, if the normalization of x is used). The use of more stochastic terms may appear computationally more expensive, but it comes with the benefit of simple stochastic equations that avoid complex matrix operations. Finally, it is noteworthy that the S matrix that we propose, with size M T× , also fulfills TSS D= , even if M T< . In the following sections we will prove that our equations faithfully reproduce the results that can be obtained in simulations with explicit MCs, with similar numerical stability and lower computational cost. Numerical Simulations To test the proposed DA algorithm, it was compared to MC modeling both in their uncoupled and coupled subunits approach. If properly implemented, a DA method considering coupled activation subunits should give the same results as multi-state MCs, while a DA method with uncoupled activation subunits should behave as independent, two-state MCs (see Figure 1). As we show next, this is indeed the case. Additionally, we examined a common “steady state” approximation employed when using DA methods. In this approximation the variable values in the expressions multiplying the noise terms are replaced by their steady state values [28,31,33,40]. Here we will show that the steady state approximation must be used with great caution depending on the kinetics of the channels simulated. The details of the specific models we used and the numerical implementation are described in Methods. Before we give the simulations results, we clarify a few important numerical issues. Numerical implementation issues An issue that is commonly debated in the implementation of DA is whether to manipulate the state variables to make them increase discretely or to bound them between 0 and 1. Mino et al. [28] did both, making the variables to represent an integer number of open channels by multiplying by the number of channels and then rounding them to the lowest integer. Later, Bruce [41] found that rounding to the lowest integer produced a shift of the Firing Efficiency curves to the left, and that it was more appropriate to make the rounding to the nearest integer. In both works the state variables were bounded between 0 and 1 (or between 0 and the number of channels), something that does not impose any mathematical difficulty when dealing with two- state gating subunits. However, when working with multi-state channels, bounding the variables by manually correcting an off-bound value causes the variable vectors to leave a bounded hyperplane that may 12 cause the diffusion matrix to be no longer positive semi-definite making it impossible to calculate its square root [33]. Therefore, Goldwyn and colleagues decided not to bind the variables and allowed values below 0 and above 1 and instead replaced the variable values in the random terms with their steady state values. We will show here that in some important cases this steady state approximation can introduce significant deviations compared to the exact equations. In the present work, neither the variables were converted to an integer number of channels nor were they bounded between 0 and 1. The only manipulation performed to ensure real valued random terms was to apply the square root to the absolute value of the argument. As evidenced by the simulations presented here, this did not introduce any noticeable deviation from the simulations with MCs. Voltage clamp simulations The behavior of the four simulation algorithms was first compared in voltage clamp simulations, using the potassium channel from the HH model alone. The initial condition was the steady state value at -90 mV and a 6 second simulation was performed with the kinetic constants fixed at +70 mV. The number of open channels was recorded at every time step of the simulation (Figure 2A, top, shows 8 simulated traces). 200 independent pulses were simulated and the mean and variance of open channels was calculated for every time step. Figure 2A, middle, shows mean and variance as a function of time and Figure 2A, bottom, shows the relationship between mean and variance of the number of open channels. It is well known that these two moments follow the relationship [42]: 2 where and 2 σ =< > − i I I < > I N 2 Iσ is the variance of the current at any given time, I< > is the mean of the current at the same time, i is the single channel current and N the number of channels. This relationship stems directly from the fact that the current in voltage clamp is the sum of independent binary < >= I channel current. In this case, if p is the probability of finding a channel open, then ) ( Comparison of Figures 2A and 2B show that the DA, implemented as coupled activation subunits, , which jointly give eq. (17). σ = 2 2 I Ni p 1 − p Nip (17) perfectly reproduces the behavior of MC simulations. In both simulations the fit of the data to Equation (17) yields the correct values of N and i. When the voltage clamp simulation is performed with uncoupled subunits models (Fig. 2C and 2D), the variance of the number of open channels increases with a longer delay than the mean, causing the mean vs. variance relationship not to be fit by the inverted parabola. The fit parameters fall very far from the real values regardless of the simulation algorithm. This stems directly from the fact that in the uncoupled subunit approximation the conductance is not the sum of the different channels conductance at a given time, but instead is the multiplication of such sums. Therefore, the derivation of eq. (17) is no longer accurate. However, it is noteworthy that the DA algorithm with uncoupled subunits behaves similar to modeling uncoupled Markov Chains. 13 The steady state approximation requires the kinetic constants to change slowly compared to the variables. As the kinetic constants are voltage-dependent, the voltage has to change slower than the variables. In a voltage clamp simulation, exactly the opposite happens as the voltage is changed instantaneously at time 0. As expected, simulations that use the steady state approximation performed very poorly, regardless of the activation subunits coupling (Figure 3). In the case of coupled activation subunits (Figure 3A), an almost constant variance of the number of open channels was obtained, and the maximum during the rising phase of the mean was lost. With uncoupled activation subunits (Figure 3B), the maximum in the variance trace was also lost but a longer delay was also observed. As a result, neither model recovered the correct parameters in the mean vs. variance fit. Thus, our proposed DA algorithm produces the same results as MC modeling. Significant differences appear when subunit coupling is not treated equally in the algorithms, and when steady state approximation is used. We will test it further with current clamp models also assessing the numerical stability and processor time cost. Mammalian Ranvier node model The performance of the different simulation algorithms in the mammalian Ranvier node (Rb) model [43] was tested using a 1 ms simulation in which a single current pulse of 0.1 ms duration and variable amplitude is given at the beginning (Figure 4A). 1000 simulations are performed at each current amplitude level and the measures of action potential variability (defined in Methods, Rb model) are presented in Figures 4B – 4D. There are clearly two pairs of overlapping curves (Rb2MC with Rb2DA and Rb8MC with Rb8DA), indicating that what makes a difference in the behavior of the models is the activation subunit coupling (Rb2 vs. Rb8) and not the numerical algorithm employed (MC vs. DA). While results in Figure 4 correspond to simulations performed with 1000 channels, simulations were also performed with 500, 5000 and 10000 channels. To present the data in a more concise way, the Firing Efficiency vs. Stimulus amplitude curves were fitted to a cumulative Gaussian distribution (Figure 5A). The mean of the distribution corresponds to the Threshold, the stimulus amplitude that has a probability 0.5 of firing an action potential, while the standard deviation (σ) is a measure of the spread or the input/output relationship. Figure 5B shows the fitting parameters obtained with different number of channels and the tested algorithms. The most relevant observation in these figures is that, like in Figure 3, simulations performed with the same state representation behave the same regardless of the numerical algorithm. In other words, DA reproduces the same behavior that is obtained with MC simulation. Also it is interesting to note that the threshold is almost independent of the number of channels, while σ is highly dependent on it. The latter fact is not surprising as fewer channels imply a noisier, more variable simulation and thus a flatter relationship between stimulus amplitude and Firing Efficiency. When more channels are present, noise is reduced and the curve gets steeper, becoming a step function in the deterministic limit (infinite number of channels). Figure 5C shows a comparison of the DA algorithms with and without the steady state approximation. In the case of uncoupled subunits there is no much difference introduced by this approximation, behaving almost exactly as the exact DA. However, the model with coupled 14 subunits deviates considerably from the exact algorithm, with less variability as evidenced in the lower spread of the activation curves (σ values). Therefore, it seems that the action potential in the Rb model is fast enough to make the steady state approximation not suitable for a model with coupled activation subunits. Numerical Stability To test and compare the numerical stability of the algorithms presented here, simulations were performed with increased time steps and the effect of time step on the Firing Efficiency curve was observed. Figure 6A shows that as the time step is increased the threshold also increases, indicating a shift to the right of the Firing Efficiency curve. At dt = 10 µs, there is a sudden drop in threshold, but this is probably a sign of a major instability occurring in the numerical integration. An important observation, however, is that all algorithms show the same behavior, reinforcing the idea that our DA algorithm reproduces the behavior of MC modeling. The spread of the Firing Efficiency curve (Figure 6B) remains to a great extent unchanged as dt is increased and once again the simulation algorithm (MC or DA) does not make any difference. In this case, however the state representation makes a difference as the Rb2 model (independent subunits) shows a steeper Firing Efficiency curve than the Rb8 model (coupled subunits). It should be mentioned that when using DA for the coupled subunits approach (Rb8 model) there was a significant number of simulations with dt=5 µs in which an out-of-range voltage value (NaN, ±Inf) was obtained, and all simulations ended out-of-range for dt≥10 µs. This is to some extent avoided if the variables are constrained to be between 0 and 1, but it comes with some computational cost. Normally, this constraint was not imposed in the simulations presented here (nor in the HH model) and for dt≤1 µs it was not necessary at all. Depending on the kinetics of model to be implemented a decision has to be made as to whether it is worth to add a couple of lines of code that will check and correct values out of boundaries. Computational cost Figure 6C-D plots the time it takes to run 16000 simulations (1000 simulations per stimulus amplitude) in the machine employed for this work, as a function of the integration time step (6C) or the number of channels simulated (6D). It is clear that MC modeling is slower than DA, with all state representations and all conditions tested. On the other hand, the 8-state representation that arises from an independent channel approach is always slower to calculate than its counterpart 2- state representation (independent subunits). This difference is bigger for MC modeling than for the DA algorithm, maybe because this model was tested in an environment mostly oriented to matrix operations. However, the most remarkable observation from Figure 6 is that MC modeling is highly affected by the number of channels in the simulation (more channels imply more transitions to calculate) while the DA method is only sensitive to the time step and completely unaffected by the number of channels. Squid axon model The original Hodgkin and Huxley [21] model for squid giant axon is deterministic and the channel activation functions are continuous variables. In the absence of a stimulus, no action potential is elicited and the system relaxes to a resting voltage very close to -65 mV. However, if discrete 15 stochastic channels are considered spontaneous action potentials arise due to sodium channels fluctuations [16]. Here, two types of stochastic HH models were simulated and the resulting spike frequency and intervals were analyzed. The HH2 model uses the independent subunits approach (2-state activation subunits), while HH58 model uses the coupled subunits approach, with 5-state potassium channels and 8-state sodium channels. As expected, the frequency of the spontaneous action potentials increases as the number of channels is decreased in all models and simulation algorithms (Figure 7). However, there is a striking difference between the behavior of the HH2 models and the behavior of the HH58 models at the same number of channels. While at NNa=1500 the HH2 models barely fires an action potential, the HH58 models fires about 30 action potentials per second. Importantly, our DA algorithm produces the same firing rates as the corresponding MC models. Figure 8A plots the mean action potential frequency observed in the 500 s simulation, as a function of the number of sodium channels (NNa) simulated (the number of potassium channels was always set to NNa*0.3). The pattern observed with the Rb model is repeated: models with different subunit coupling have a different behavior while the simulation algorithm makes no difference in the results. In order to go beyond the simple firing rate quantification, the Inter-Spike Intervals (ISIs) obtained in each case were plotted in histograms and fitted to an exponential decay function (Figure 8B, also see Eq. (22) in Methods). For all ISIs obtained, it was observed that the first two bins (marked with * in the histogram) did not follow the exponential trend so they were excluded when fitting the histograms. This was observed in all simulations and thus it is not caused by a specific simulation algorithm or subunit coupling. Indeed, it has been observed before [18] and is probably due to the resonant properties of the HH model [21,44,45] that, with a frequency of peak response of 67 Hz, will increase the probability of ISIs around 33 ms. Figures 8C and 8D show the fit parameters obtained as a function of the number of sodium channels, and it is evident that the simulation algorithm employed does not make any difference in the ISI distributions, while the subunit coupling does. As with the Rb model, a DA approximation algorithm was tested in which the variable values of the random term were replaced by their steady-state values. The results obtained with the coupled subunits model (HH58) is plotted in Figure 8 as well (gray triangles). Here the deviations from the exact DA (and MC as well) are minor, probably because the voltage dynamic in this model is slow enough to let the variables (at least the m variable) to be at its steady state value during almost all the simulation. Numerical Stability To check for numerical stability of the methods, simulations were repeated with increasing values of dt, the integration time step. As shown in Figure 9, increasing dt up to 100 µs has little or no effect in the mean rate of spikes (9A) or the parameters of the ISI distribution (9B and 9C). There are some deviations for dt > 10µs, but they are minor compared to what was observed with the Rb model. In this case, no out-of-range voltage values were produced throughout the 500 seconds simulated. Remarkably, the choice of the algorithm has no effect on the numerical stability within the dt values tested. 16 Computational cost Figure 9D-E plots the time it took to simulate 500 seconds as a function of the time step (9D) and the number of sodium channels (9E). As with the Rb model, MC modeling performance is severely affected by the number of channels while the DA algorithm is independent of it and only affected by the integration time step. However, in this case MC modeling turned out to be as efficient (in some cases more efficient) than DA at the lowest dt values. This is probably due to the longer time constants of the HH model (reproducing the behavior of squid axons at 6.3ºC) compared to the Rb model (mammalian ranvier node at 37ºC). In the HH model, there are fewer transitions per time step and probably when dt<1µs there are many steps in which no transition occurs, thus leaving all the computational weight to solving the membrane current equation. However as dt increases more transitions per step begin to occur and then the computational cost is dominated by the calculation of transitions rather than by the advancing of time steps. Accuracy of alternative DA implementations Two works recently proposed DA implementations that take into account subunit coupling [33,40]. Goldwyn and colleagues [33] tested the DA approach for coupled subunit originally developed by Fox [29], and solved the square root of the stochastic diffusion matrix numerically at each time step. Besides the computational cost of this approach, it demands the matrix D (eq. (14)) to be always positive semi-definite to compute real valued square roots. One simple solution for this, and the one they took, is to use the steady state approximation, replacing the values of the variables by their equilibrium values. On the other hand, Linaro et al. [40] deduced the covariance of the noise introduced by channel fluctuations and showed that it can be reproduced by a sum of Ornstein-Uhlenbeck processes (4 for potassium channels, 7 for sodium channels) with particular time constant and variance coefficients. This noise is then added to the sodium or potassium current, respectively, that are calculated by deterministic Hodgkin-Huxley equations. Importantly, they calculate the noise coefficients using steady-state approximation. As shown before, the use of a steady-state approximation can result in serious deviations from the explicit MC modeling because the fluctuations become independent on the actual value of the variables at the corresponding time. Figure 10 shows that indeed this is the case, with both algorithms falling short of reproducing the behavior of Markov Chains in the voltage-clamp simulations (note the resemblance of Figure 10A with Figure 3) as well as in the firing efficiency and firing time variance curves of the Ranvier Node model (Figure 10B). We managed to implement Fox’s equations without the steady-state approximation, just by extracting the absolute value of the variable vector prior to the matrix square root operation. In that case, the simulations give the same results as MC modeling and our DA implementation (not shown). Therefore, the matrix equations originally proposed by Fox and Lu are indeed a good numerical approximation to MC modeling although with a high computation cost – at least 20 times slower than our method in cases we examined. 17 DISCUSSION Accuracy of the Diffusion Approximation The original description of the Diffusion Approximation (DA), in its general form for a multiple (more than 2) state Markov Chain (MC), implies the calculation of the square root of a matrix [29,31]. As this is too time consuming an operation to be performed in real time, the uncoupled subunits approximation, consisting a stochastic form of the original Hodgkin and Huxley’s equations, seemed to be the right choice. Very recently is was described [35] and mathematically proven [32,33,40] that when the activation subunits are considered to be coupled or ‘tied’ in groups (as they really are in ion channels), the resulting conductance fluctuations have statistics that cannot be adequately reproduced in a model with uncoupled subunits, either with MC modeling or a DA algorithm. Previous reports, which suggested the inadequacy of DA methods [28,35,41], failed to notice that they were using an uncoupled subunits approach for the DA and a coupled subunits approach for the MC modeling. The simulations performed here confirm the results of Goldwyn et al. [33] in showing that the DA method was not being implemented properly for channels of more than two states. Furthermore, it is again confirmed here that the DA and MC algorithms give similar results – with two different models of neuronal excitability, and in both the coupled and uncoupled subunits approach; to our knowledge the most thorough testing that any DA algorithm has been subjected to. Relation to other Algorithms Goldwyn et al. [33], tested the DA approach for coupled subunit originally developed by Fox [29] and showed that a properly implemented DA can approach better the results of MC modeling, in the context of the HH model. However, they computed the square root of the diffusion matrix during execution, resulting in a slow computation speed. Another recent work [40] suggested an alternative DA implementation for the HH model, that uses similar equations to the uncoupled subunits approach (Eq. (25) for m, h and n) but with a noise term that is time-correlated in the way it should be when the subunits are considered to be coupled. The correlation of the noise terms requires solving 7 (Na) or 4 (K) additional differential equations, of a complexity comparable to those presented here. Importantly, both works, as well as many others, employed a steady-state approximation for the calculation of the stochastic term matrix introduced. As we showed here, this caused significant deviations in voltage clamp and the Rb model (but not in the HH model (Figure 10)). It is important to note that among the channels that work on the time scale of action potentials, the sodium channel of the Rb model has fast kinetics (resembling channels from mammalian Ranvier nodes), while the HH model possesses channels that are rather slow (giant squid axon at 6.3ºC). Most likely, this is the reason why the Rb model is more affected by the steady-state approximation than the HH model. As the time scale relevant for models based in the mammalian nervous system is precisely that of the Rb model, our conclusions about the steady- state approximation are of importance for such models. Both previous works [33,40], as well as the original derivation by Fox [31], give specific instructions on how to construct the SDE for sodium and potassium channels, in the context of the 18 HH model. However, generalizing these instructions to other kinetic schemes is not an easy task, since no explicit general expressions are given. In contrast, our alternative derivation gave simple and general closed form expressions for the both the diffusion matrix D (eq. (10)) and its matrix square root S (eq. (13)). In order to compare with previous DA formulations [31,33], we analytically found S for the potassium case using Cholesky decomposition. Surprisingly, but in tune with our proposed equations, the resulting matrix was simpler (compare eq. (14) with (15)) simulation speed considerably. Specifically, instead of the and sparse (containing many zero elements). The exact and simple expression for S (eq. (13)) allowed us to avoid the use of the inaccurate steady state approximation and to improve )3 ( O M computational complexity of the numerical matrix square root implementation (as done in [33]) our method has a complexity ( )2 ( ) O M and O M , depending on the number of kinetic transitions (see eq. (13)). Numerically testing this, we observed our method run at least 20 times faster, depending on the between software environment employed. Also our method only doubles the time required to solve the deterministic equations, that ignore the stochastic terms (not shown). Moreover, the equations that govern the dynamics of stochastic ion channels in our approach can be simply written as separate equations instead of matrix operations (e.g. eq. (16) for potassium and Supplemental Text T1 for sodium). This facilitates their implementation in non matrix-oriented computation software such as Neuron, and may also simplify future analytical analysis of the behavior of the stochastic neuron. We note a connection between the DA approach and another stochastic simulation method - the “binomial population” approach [46,47,48]. This approach employs eq. (11) directly, where each ij∆ is distributed binomially. So essentially, the only additional approximation we made was that ij∆ was a Gaussian RV. This can greatly reduce simulation speed since the generation of binomial RVs is much less efficient than Gaussian RVs, especially for large N [49]. As noted, our simulations (as well as Goldwyn’s [33] and Linaro’s [40]) indicate that this approximation is very good, as long as N is not too small. However, if N is small enough, so that the discrete nature of ion channel conductance becomes significant, then this approximation might break down. In that case, one ij∆ as a Poisson RV with parameter can speculate that it might be more accurate to approximate = np X A dt ij j (by the law of rare events). Note that also in this approximation it is possible to pair opposing transition pairs W = ∆ − ∆ (as in eq. (12) ) and generate ijW according to a ij ji ij Skellam distribution (the distribution of the difference between two Poisson RV). However, we have not investigated here whether or not the Poisson\Skellam approximations may actually improve the speed of binomial population algorithm or have any advantage over other methods (such as MC). Finally, we note that a similar approach to ours was previously introduced in the field of chemical physics. As in our case, this equation, named “the chemical Langevin equation” [50,51] sums the stochastic terms along transitions and not along states (compare our eq. (13) with eq. (23) in [50]). 19 The main difference between that approach and ours is that we sum together the noise contributions from both directions of each transition pair (done in the conversion from eq. (11) to eq. (12)). This approximately halves the computation time, when the generation of pseudo-RVs is the main computational bottleneck. Numerical efficiency – DA versus MC Following the practical approach of this work, we numerically evaluated the computational cost of both implementations (MC and DA). In almost all cases, our DA approach significantly outperforms the MC approach. In the Rb simulations (Figure 6C&D) the DA approach for coupled subunits is at least an order of magnitude faster than MC for all values of N and dt tested. In the HH simulations (Figure 9D&E) this remains true, except when low values of dt or N are used. Again we note that the results for Rb model are more significant to the mammalian nervous system, due to the similar kinetic timescales. Another issue to consider when comparing Figures 6 and 9 is that the Rb simulations presented here were performed in the Scilab numerical computation package while the HH simulations were implemented in NEURON. The latter will be always faster because it runs as compiled code; also variations in how each software implements numerical calculations at the processor level may cause further differences. In all cases, however, the speed of simulations performed with the DA algorithm was only affected by the size of the integration time step and completely independent of the number of channels to be simulated, because the number of channels is only a parameter in the equations. On the contrary, MC modeling was heavily affected by the number of channels and less affected by the integration time step. In this case a greater number of channels imply more transitions per time step, and for each transition two new calculations have to be made, each requiring a new random number. Thus, there will be situations where MC modeling may be numerically more efficient than DA. With a small number of channels there will be fewer transitions per time step and thus a MC simulation may run faster than a DA algorithm. This difference will be enhanced if the channels have slow kinetics, because this will reduce the probability of transitions. Also, if a small integration time step is required the DA algorithm can be as slow as MC modeling. In both these cases, it might be better to combine the MC and DA methods: use MC for channel with slow kinetics, while handle the faster channels using the DA approach. The waterline between “slow” and “fast” timescales here would be the time step duration. Also, note that in the simulations presented here, random numbers were generated in simulation time. Further speed-up of the DA algorithm can be achieved by the use of a pre-generated random number list. Conclusions This paper further confirms that the use of the Diffusion Approximation (DA), without any additional approximations, produce results that are indistinguishable from those of Markov Chain modeling (MC). Most importantly, we present the DA in a very simple, general and computationally efficient form, which will allow its easy implementation for any given kinetic scheme of a channel. We show that in the most common situations, the DA method proposed here has a numerical stability comparable to that of MC modeling (even with a simple Euler- 20 Maruyama integration scheme), while being much faster. The fast simulation speed achieved makes conceivable its use in dynamic clamp experiments. 21 METHODS Models To test the accuracy and efficiency of DA relative to MC modeling, both in their independent subunits and coupled subunits approaches, two models were employed in which different measures of simulation accuracy were calculated. Mammalian Ranvier node – Rb model The mammalian Ranvier node model [43] was the model employed previously to compare the performance of DA versus MC modeling [28,41]. This model consists only of a voltage-dependent sodium channel and a voltage-independent leak current. The membrane current equation is ( ) ( ) ( ) t V t ( ) t C + = − E app g Na Na I m ( ) ( ) dV t V t + dt R (18) with parameters Cm=18.9 pF; R=7.372MΩ; Nag that the leak reversal potential is 0. The α and β transition rates are given by the following voltage dependent functions: =6.808 µS; ENa=144 mV. The voltage is shifted so − = ) 1 − β m α m ( ) V ( ) V ( ) 1.872 25.41 V ( ) ( − / 6.06 exp 25.41 V ( ) − 3.973 21 V ( ) ( − 21 / 9.41 v ( ) + 0.549 27.74 V ( ( ) 27.74 / 9.06 V ) ( ( ) exp 56 Simulations of 1 ms were run in which a 100 µs current pulse was given at the beginning (Figure 1). 22.57 − − 1 exp (19) / 12.5 = − exp ( ) V ( ) V α h β h 1 + 1 − ) ) V = + = The pulse amplitude varied between 5 and 6.5 pA. 1,000 simulations were run and the following parameters were calculated: Firing efficiency, the fraction of simulations in which an action potential was evoked; and the mean and the variance of Firing time, time at which the voltage reached or surpassed 80 mV. Firing efficiency versus pulse amplitude curve was fit to the cumulative Gaussian distribution ( FiringEfficiency I 2 ) app = Φ − t 2 e dt 1 2 π x ∫ −∞ = 1 2  + 1   ( ) Φ x =     I app − Th σ    x 2 erf            erf(x) represents the error function. Th (threshold) gives the amplitude for a probability of firing of 0.5, while σ quantifies the spread of the input/output relationship. 22 Hodgkin and Huxley model of squid giant axon – HH model m C = − ( ) dV t dt The original Hodgkin and Huxley [21] model was simulated with the equation ( ( ( ) ) ( ) ( ) ( ) ( ) ( ) g V t g t V t t V t l K and parameters Cm=1, ENa=50 mV, EK=-77 mV, El=-54.4 mV, (20) Kg =36, gl=0.3 (Voltages Nag are shifted with respect to the original model to make the resting potential equal to -65 mV). The α and β functions employed are =120, − − − − − E E E ) g Na Na K l 0.1 ( V + α m ( ) V = ) 40 +  40    10  +  65    20  ) 55 +  55    10  1 = = − α h β m ( ) V ( ) V exp 4 exp  V  −    V  −    V  −   ( V  V  −   Simulations of 500 seconds were performed, and action potentials were recorded as the time at +  35    10  +   65 V   −     80  +  65    18   V  −   0.125 exp 0.07 exp (21) 0.01 exp exp ( ) V ( ) V ( ) V α n β n β h + = − = + = 1 1 1 which the voltage reached or surpassed 0 mV. The time of action potentials during the simulation were stored, and the Inter-Spike Intervals (ISIs) were calculated. The normalized ISI distribution was fitted to an exponential decay function with a refractory period [16]: ) ( ) ( ) ( The first two values of the ISI distribution histogram were not included in the fitting procedure. exp − r = − P τ τ r ref t (22) Uncoupled independent subunits N channels are simulated as 4N independent, 2-state subunits: a0 a1 αa βa (a = n,m,h) where αa is the transition probability from the 0 to the 1 state, and βa the transition probability from the 1 to the 0 state. NNa Sodium channels are simulated as 3NNa m subunits and NNa h subunits, and at each time step the sodium conductance is calculated as 3     Nm 1     3 N     Na Na NK Potassium channels are simulated as 4NK n subunits and the potassium conductance is calculated as Nh 1 N (23) = g g Na Na .   Nn 1   4 N   K Nm1, Nh1 and Nn1 are the number of m, h, and n subunits, respectively, that are in the ‘1’ state. = g g . K K (24) 4 23 a a − a = α Diffusion Approximation The DA in the case of independent subunits uses the variables m, h, n ∈[0,1] to keep track of the fraction of m, h, and n subunits, respectively, that are in the ‘1’ state. It follows immediately that the fraction of subunits in the ‘0’ state will be 1-m, 1-h, and 1-n, respectively. Fox and Lu [29] showed that the time evolution of the variables is given by the SDE da ( ) 1 dt where a represents either m, h or n. The stochastic term ξ(t) is a Gaussian white noise with zero mean and unit variance that is scaled by σa(t), being α − (1 a N a where Na is the number of a subunits (Nm=3NNa, Nh=NNa and Na=4NK). When the steady state approximation was used, the noise scaling factor was calculated as α β 2 a a α + β ( a − β + σ a a ξ ( ) ( ) t t ) + β (27) ( ) t = ( ) t = (25) (26) N a σ a σ a a a ) a a The conductance of sodium and potassium are calculated using the classical Hodgkin & Huxley expressions g Na = 3 g m h Na and g K = 4 g n K . In the voltage clamp simulations, the number of open potassium channels was calculated as 4 = N N n K . O Coupled independent subunits (independent channels) There are two possible ways of implementing a coupled subunits approach. The first consist of simulating 4 independent 2-state subunits per channel, as in the previous approach. However, each subunit is assumed to belong to a specific channel. A channel is considered open if and only if its four subunits are in the open state. Therefore, the state of each subunit (hence of each channel) must be tracked individually during the simulation [43]. In this paper a second approach is employed, that consists in building a multi-state MC per channel considering the possible combinations of active subunits. This allows for the faster number-tracking algorithm employed for simulations [16,26,27]. Given that subunits of a given kind are identical and independent, a Sodium channel has 8 possible states while a Potassium channel has 5 states: 24 m1h0 2αm 2βm m2h0 αm 3βm m3h0 αh βh αh βh αh βh Na channel m0h0 αh βh m0h1 3αm βm 3αm βm m1h1 2αm 2βm αm 3βm m3h1 m2h1 2αn 3βn n0 4αn βn n1 3αn 2βn n2 αn n3 n4 K channel 4βn In this approach, only one state of each MC represents the conducting or open channel, which is the state with all subunits active (m3h1 or n4). Then the conductance is calculated with the fraction of channels or MCs that are in the open state:     Nm h Nn 3 1 4     N N     Na K where Nm3h1 and Nn4 are the number of channels in the state m3h1 and n4, respectively. = = g g g g Na Na ; K K Diffusion Approximation The DA for channels with coupled activation subunits is detailed in the Results section. Numerical implementations Software implementation All models and algorithms were implemented in Scilab, a matrix-oriented numerical software (www.scilab.org), and NEURON, a simulation environment oriented to the modeling neurons and neural networks (www.neuron.yale.edu). Source files and scripts are available in ModelDB (http://senselab.med.yale.edu/ModelDB/). Both environments produced identical results but simulations in NEURON run faster because it runs in compiled mode. Results presented here (most importantly, processing time data) correspond to simulations in Scilab for the mammalian Ranvier node (Rb) model and simulations in NEURON for the squid giant axon (HH) model. Markov Chain modeling Independent MCs were modeled using a number-tracking algorithm [16,26,27,28]. Thoroughly τ = ( ) exp t λ described in [28], briefly this algorithm consist in keeping track of the number of MCs in each state, rather than keeping track of each MC individually. At any time t, the probability density function of the lifetime before the next transition (any transition) is ) ( ) ( tP where λ(t) is the effective transition rate given by S = ∑ i where S is the total number of states in the MC, Ni is the number of MCs in state i, and ζ(t) is the sum of transition rates escaping from state i. If there is more than one type of MC, they are all summed into λ. The time of the next transition tn is calculated by drawing a random number uniformly distributed within [0,1] and taking the inverse of the c.d.f. of the lifetime. If tn≤t, a ( ) ( ) N t t ζ i i − ( ) t λ τ ( ) t λ 25 transition has to be calculated before updating the current equation. Among all possible transitions, the probability of transition j to occur is = ( ) ( ) ( ) P t N t tα j i j where i is the state originating the transition j and αj its rate. A cumulative probability for all transitions is calculated and a transition is chosen by drawing a random number uniformly distributed within [0,1]. The number of MCs at each state is updated, and a new time for the next transition is calculated. When no more transitions are to occur in the current time step, the current equation is advanced one time step using an Euler integration scheme. Diffusion approximation Stochastic differential equations for DA were solved by an Euler-Maruyama integration method. For the coupled subunits approach, a better numerical stability is obtained if the fact that the sum of state variables for a given channel is 1 is taken into account, also reducing the number of SDEs + t dt = + t dt n 1 n n 2 n 2 , n α 3 n − β 2 n − α 2 to be solved. Thus, for potassium channels the equations used for advancing one time step are ( ) α − α + β α α + β = + − β − η + β + η 4 3 2 4 3 2 n n dt n n n n n n n n 1 2 1 1 1, 1 0 1 1 2 0 2 n n n ) ( α + η + β − η + β 3 2 2 n dt n 2 3 2 3 ) ( + η α − η n n 4 3 4 2 n ( = − β + β α + 4 4 n n 4 4 , n = 1 - n - n n - n + + + + 0 3 2 1 ,t dt ,t dt ,t dt ,t dt being η1, η2, η3, and η4 n n 3 3 n ) − η α 4 - n 4 n 1 + β 3 n 2 + β 4 n 2 + β 4 n − β 3 n + β 3 n α 3 independent Gaussian RVs with zero mean and standard − α α 2 n 3, n 2 n 4 n 4 n 2 n 4 n 3 2 α + dt n 3 n n 3 n n 3 n = n 3 n n 3 + t dt + t dt n n n n n n n n n n + d t n + ,t dt deviation dt N . n0 – n4 stand for n0,t – n4,t, i.e. the value of the variables at time t. A similar set K of equations was used for sodium channels. No rounding was performed on the variables, nor were they bound to lie between 0 and 1 (see discussion). To ensure real valued random terms, the square roots were applied to the absolute value of the operand. For the steady state approximation, the variables ni and mihj were replaced by their steady state values in all the noise terms:   4 =    i   − − 1 3 j i i α β α β m m h h + 3 ) ( α β h m   3 =    i   4 i α β n n + β n m h i ( α m ( α n (28) β h 4 ) ∞ i ∞ j + n ) . , − j i 26 ACKNOWLEDGMENTS We wish to thank Rolando Rebolledo, Mauricio Tejo and Rolando Biscay for helpful discussions and Joshua Goldwyn, Yuval Elhanati, Rolando Rebolledo and Ron Meir for critical reading of the manuscript. 27 REFERENCES 1. Faisal AA, Selen LP, Wolpert DM (2008) Noise in the nervous system. Nat Rev Neurosci 9: 292- 303. 2. Stein RB, Gossen ER, Jones KE (2005) Neuronal variability: noise or part of the signal? Nat Rev Neurosci 6: 389-397. 3. Longtin A, Bulsara A, Moss F (1991) Time-interval sequences in bistable systems and the noise- induced transmission of information by sensory neurons. Phys Rev Lett 67: 656-659. 4. Gluckman BJ, Netoff TI, Neel EJ, Ditto WL, Spano ML, et al. (1996) Stochastic Resonance in a Neuronal Network from Mammalian Brain. Phys Rev Lett 77: 4098-4101. 5. McDonnell MD, Abbott D, Friston KJ (2009) What Is Stochastic Resonance? Definitions, Misconceptions, Debates, and Its Relevance to Biology. PLoS Comput Biol 5: e1000348. 6. McDonnell MD, Ward LM (2011) The benefits of noise in neural systems: bridging theory and experiment. Nature reviews Neuroscience 12: 415-426. 7. Douglass JK, Wilkens L, Pantazelou E, Moss F (1993) Noise enhancement of information transfer in crayfish mechanoreceptors by stochastic resonance. Nature 365: 337-340. 8. Levin JE, Miller JP (1996) Broadband neural encoding in the cricket cercal sensory system enhanced by stochastic resonance. Nature 380: 165-168. 9. Molgedey L, Schuchhardt J, Schuster HG (1992) Suppressing chaos in neural networks by noise. Physical review letters 69: 3717-3719. 10. Fiete IR, Seung HS (2006) Gradient learning in spiking neural networks by dynamic perturbation of conductances. Physical review letters 97: 048104. 11. Kirkpatrick S, Gelatt CD, Jr., Vecchi MP (1983) Optimization by simulated annealing. Science 220: 671-680. 12. Motwani R, Raghavan P (1996) Randomized algorithms. ACM Comput Surv 28: 33-37. 13. Lecar H, Nossal R (1971) Theory of threshold fluctuations in nerves. II. Analysis of various sources of membrane noise. Biophys J 11: 1068-1084. 14. White JA, Rubinstein JT, Kay AR (2000) Channel noise in neurons. Trends Neurosci 23: 131-137. 15. White JA, Klink R, Alonso A, Kay AR (1998) Noise from voltage-gated ion channels may influence neuronal dynamics in the entorhinal cortex. J Neurophysiol 80: 262-269. 16. Chow CC, White JA (1996) Spontaneous action potentials due to channel fluctuations. Biophys J 71: 3013-3021. 17. Faisal AA, Laughlin SB (2007) Stochastic Simulations on the Reliability of Action Potential Propagation in Thin Axons. PLoS Comput Biol 3: e79. 18. Rowat PF, Elson RC (2004) State-dependent effects of Na channel noise on neuronal burst generation. J Comput Neurosci 16: 87-112. 19. Dorval AD, Jr., White JA (2005) Channel noise is essential for perithreshold oscillations in entorhinal stellate neurons. J Neurosci 25: 10025-10028. 20. Fernandez FR, White JA (2008) Artificial Synaptic Conductances Reduce Subthreshold Oscillations and Periodic Firing in Stellate Cells of the Entorhinal Cortex. J Neurosci 28: 3790-3803. 21. Hodgkin AL, Huxley AF (1952) A quantitative description of membrane current and its application to conduction and excitation in nerve. J Physiol 117: 500-544. 22. Hille B (2001) Ion Channels of Excitable Membranes. Sunderland, MA, USA: Sinauer Associates Inc. 23. Bezanilla F (2000) The voltage sensor in voltage-dependent ion channels. Physiol Rev 80: 555- 592. 28 24. Colquhoun D, Hawkes AG (1981) On the stochastic properties of single ion channels. Proc R Soc Lond B Biol Sci 211: 205-235. 25. Neher E, Stevens CF (1977) Conductance fluctuations and ionic pores in membranes. Annu Rev Biophys Bioeng 6: 345-381. 26. Gillespie DT (1977) Exact stochastic simulation of coupled chemical reactions. J Phys Chem 81: 2340-2361. 27. Skaugen E, Walloe L (1979) Firing behaviour in a stochastic nerve membrane model based upon the Hodgkin-Huxley equations. Acta Physiol Scand 107: 343-363. 28. Mino H, Rubinstein JT, White JA (2002) Comparison of algorithms for the simulation of action potentials with stochastic sodium channels. Ann Biomed Eng 30: 578-587. 29. Fox RF, Lu Y (1994) Emergent collective behavior in large numbers of globally coupled independently stochastic ion channels. Phys Rev E Stat Phys Plasmas Fluids Relat Interdiscip Topics 49: 3421-3431. 30. Goldwyn JH, Shea-Brown E (2011) The what and where of adding channel noise to the Hodgkin-Huxley equations. PLoS Comput Biol: (in press). 31. Fox RF (1997) Stochastic versions of the Hodgkin-Huxley equations. Biophys J 72: 2068-2074. 32. Pakdaman K, Thieullen M, Wainrib G (2010) Fluid Limit Theorems for Stochastic Hybrid Systems with Application to Neuron Models. Adv in Appl Probab 42: 761-794. 33. Goldwyn JH, Imennov NS, Famulare M, Shea-Brown E (2011) Stochastic differential equation models for ion channel noise in Hodgkin-Huxley neurons. Phys Rev E Stat Nonlin Soft Matter Phys 83: 041908. 34. Golub G, van Loan C (1996) Matrix Computations (Johns Hopkins Studies in Mathematical Sciences)(3rd Edition): The Johns Hopkins University Press. 35. Bruce IC (2009) Evaluation of stochastic differential equation approximation of ion channel gating models. Ann Biomed Eng 37: 824-838. 36. Schoppa NE, Sigworth FJ (1998) Activation of Shaker potassium channels. III. An activation gating model for wild-type and V2 mutant channels. J Gen Physiol 111: 313-342. 37. Bezanilla F, Perozo E, Stefani E (1994) Gating of Shaker K+ channels: II. The components of gating currents and a model of channel activation. Biophys J 66: 1011-1021. 38. Horn R, Ding S, Gruber HJ (2000) Immobilizing the moving parts of voltage-gated ion channels. J Gen Physiol 116: 461-476. 39. Vandenberg CA, Bezanilla F (1991) A sodium channel gating model based on single channel, macroscopic ionic, and gating currents in the squid giant axon. Biophys J 60: 1511-1533. 40. Linaro D, Storace M, Giugliano M (2011) Accurate and fast simulation of channel noise in conductance-based model neurons by diffusion approximation. PLoS Comput Biol 7: e1001102. 41. Bruce IC (2007) Implementation issues in approximate methods for stochastic Hodgkin-Huxley models. Ann Biomed Eng 35: 315-318; author reply 319. 42. Alvarez O, Gonzalez C, Latorre R (2002) Counting channels: a tutorial guide on ion channel fluctuation analysis. Adv Physiol Educ 26: 327-341. 43. Rubinstein JT (1995) Threshold fluctuations in an N sodium channel model of the node of Ranvier. Biophys J 68: 779-785. 44. 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. 45. Koch C (1999) Biophysics of computation : information processing in single neurons. Oxford: Oxford University Press. xxiii, 562 p. p. 46. Schneidman E, Freedman B, Segev I (1998) Ion channel stochasticity may be critical in determining the reliability and precision of spike timing. Neural Comput 10: 1679-1703. 29 47. Steinmetz PN, Manwani A, Koch C, London M, Segev I (2000) Subthreshold voltage noise due to channel fluctuations in active neuronal membranes. J Comput Neurosci 9: 133-148. 48. Faisal A (2010) Stochastic Simulation of Neurons, Axons, and Action Potentials. In: Laing C, Lord GJ, editors. Stochastic Methods in Neuroscience. New York: Oxford University Press. 49. Press WH (2007) Numerical recipes : the art of scientific computing. Cambridge, UK ; New York: Cambridge University Press. xxi, 1235 p. p. 50. Gillespie DT (2000) The chemical Langevin equation. J Chem Phys 113: 297. 51. Bhalla US, Wils S (2010) Reaction-Diffusion Modeling. In: De Schutter E, editor. Computational Modeling Methods for Neuroscientists. Cambridge, Massachusetts, Londond, England: The MIT Press. 30 FIGURE LEGENDS Figure 1. Comparison between models with uncoupled activation subunits and coupled activation subunits. Figure 2. Voltage clamp simulation and non-stationary noise analysis. 300 potassium channels from the HH model were simulated at a constant voltage of 70 mV. At t=0, they were in a steady state condition calculated at -90 mV. 200 independent simulations were performed with each simulation algorithm (indicated above each panel) and a non-stationary noise analysis was performed. Top row: 8 sample traces of the number of open channels against time. Middle: Mean (black) and variance (grey) of number of open channels as a function of time. Bottom: The variance of the number of open channels is plotted against the mean. The continuous line represents the best fit to equation (17), and the best fit parameters are indicated below. Figure 3. Effect of the steady-state approximation on the voltage clamp simulations. The same simulations as in Figure 1 were performed with the DA method, however the value of the variables in the random terms were replaced by their steady-state values (eq. (28)). The top, middle and bottom panels are as in Figure 1. Figure 4. Rb model simulations. A. 15 voltage traces (bottom) resulting from independent simulations with the Rb model, in which a 5.8 pA pulse of 100 µs duration (top) was applied. The simulations presented correspond to the Rb8 model (independent channels approach) using MC modeling, with 1000 Na channels and dt=1 µs. B-D. Firing efficiency (fraction of action potentials evoked in 1000 simulations), mean firing time, and variance of firing time as a function of stimulus amplitude for the different simulation methods. Rb2: independent subunits, Rb8: independent channels (tied subunits), MC: Markov chain modeling, DA: Diffusion approximation algorithm. N=1000, dt=0.1 µs Figure 5. Quantification of variability in the Rb model and its dependence on the number of channels simulated. A. Fitting of a firing efficiency curve to a sigmoid function (see Methods) that is characterized by a threshold (the stimulus amplitude that produces a firing efficiency of 0.5) and σ (the standard deviation of the threshold fluctuations). B. Dependence of the threshold and slope values on the number of channels simulated, for each of the simulation methods. dt=1 µs. C. Effect of the steady-state approximation (DA-ss) on the behavior of the model. Comparison of a firing efficiency curve for 1000 channels (left), and the fitting parameters of the firing efficiency (middle and right) for simulations performed with DA algorithm with and without steady state approximation. Figure 6. Numerical stability and computational cost of the simulation algorithms with the Rb model. A-B. Dependence of Rb model variability on the integration time step used in the simulation. Threshold (A) and σ (B) values calculated as in figure 4A, obtained at different values of integration time step (dt). N=1000. C. Dependence of computation time on integration time step 31 (dt) with N=1000 channels. D. Dependence of computation time on number of channels (N) with dt=0.5 µs. Computation time is the time, in seconds, needed to perform the 16000 simulations necessary for a single firing efficiency curve (1000 pulses at 16 current levels). This figure corresponds to simulations performed in the Scilab numerical computation software. 1000 simulations were performed as 10 batches of 100 simultaneous and independent simulations, in a Core i7 machine. Figure 7. Spontaneous firing in the Hodgkin and Huxley squid axon model. Sample voltage traces of 2 seconds of simulation of the stochastic HH model for all models and simulation algorithms tested. Figure 8. Firing rate and ISI distributions for the stochastic HH models. A. Mean firing rate of the stochastic HH models in a 500 seconds simulation with different number of channels. B. An inter-spike interval (ISI) was built for each simulation and the data was fitted to an exponential decay function with a refractory period (see Methods and ref. [16]). The histograms for only two simulations are shown here for illustration purposes. The first two points (marked with asterisks) were omitted in the fitting procedure (see text). The fit lines for the two histograms showed here overlap almost perfectly. C-D. Fit parameters of the ISI distributions at different number of channels. In all the simulations, NK=0.3*NNa. Data in this figure corresponds to dt=0.1 µs. Figure 9. Numerical stability and computational cost of the simulation algorithms with the HH model. A-C. Firing parameters of the stochastic HH models at different integration time steps. Mean firing rate (A) and fitting parameters of the ISI distributions (B-C) for the stochastic HH models tested as a function of the integration time step (dt). For the HH58 models, NNa=3000 and NK=900; for the HH2 models, NNa=300 and NK=90. D. Time to perform a 500 seconds simulation with NNa=6000 and NK=1800 as a function of dt. E. Time to perform a 500 seconds simulation with dt=5 µs as a function of NNa, the number of Na channels. NK=0.3*NNa. The segmented line indicates the 500 seconds limit; any simulation below this line run faster than real time. Figure 10. Inaccuracies introduced by previous DA algorithms. A. Performance of the of the Fox [29] algorithm for coupled subunits employed by Goldwyn et al. [33] and the Linaro et al. algorithm [40] in the voltage clamp simulation and non-stationary noise analysis. See legend of Figure 2 for further details. B. Performance of the algorithms in the Ranvier Node model simulations. Firing Efficiency, Firing Time Variance and Mean Firing Time versus Stimulus Amplitude are presented for simulations with NNa=1000. Standard Deviation for Threshold (σ) is plotted against number of channels (see Figure 5). dt=0.5 µs. 32 Figure 1 33 Figure 2 34 Figure 3 35 Figure 4 36 Figure 5 37 Figure 6 38 Figure 7 39 Figure 8 40 Figure 9 41 Figure 10 42 SUPPLEMENTAL TEXT Here we explain intuitively the method for building the stochastic differential equations (SDE) that will approximate any kinetic scheme for an ion channel, and derive the SDE for a sodium channel. We take as a working example the m1h0 state from the eight-state kinetic scheme for sodium channels (Supplemental Figure 1A) Deterministic terms For the deterministic (drift) terms of the stochastic differential equation, the six transitions that go from or come to the m1h0 state have to be considered. Each arrow represents a possible transition, its probability given by the product of the voltage-dependent kinetic constant times the value of the state that is at the beginning of the arrow. Terms given by arrows starting at m1h0 are negative (Supplemental Figure 1B) while the terms given by the arrows that end at m1h0 are positive (Supplemental Figure 1C). Thus, the six deterministic terms related to m1h0 are: Stochastic terms For the stochastic terms the transition arrows are to be considered in pairs. The pairs that are connected to m1h0 are (Supplemental Figure 1D): Each pair of arrows originates a random term with zero mean and standard deviation equal to the square root of the sum of the transition probabilities, divided by the square root of , the number of sodium channels. In this case, all terms are considered positive. For the m1h0 state, the stochastic terms are where 1, 2 and 3 are independent Gaussian white noise terms with zero mean and unit variance. Note that each pair of arrows connects two (and only two) states. In the SDE for the second of such states, the stochastic term has to be repeated exactly (the same Gaussian term) but with the opposite sign. For instance, the transition pair 2mm1h0/2mm2h0 connects m1h0 and m2h0; therefore the SDE for the m2h0 state must contain the term repeating the same Gaussian white noise term as in the m1h0 equation but with opposite sign. (being Gaussian terms with zero mean it doesn’t matter which one goes positive; the key point is to have the term positive in one equation and negative in the other). Because of this, the full set of 0010102010110020111322322mmmmhhmmhmmhmhmhmhmhmhmhmhmhmhmh0010102010113/2/2/mmmmhhmhmhmhmhmhmhNaN001010201011123322mmmmhhNaNaNamhmhmhmhmhmhNNN1020222mmNamhmhN equations for the sodium channels has 20 stochastic terms but only 10 random variables; analogously the equations for a 5-state potassium channels have 8 stochastic terms with 4 random variables. Following this procedure while keeping care of repeating stochastic term with opposite signs, the following set of equations for the sodium channel is obtained: 0000100001100104000110001010201011100102102051011203113322111322mmhhmmhhNaNammmmhhmmmmhhNaNaNadmhmhmhmhmhdtmhmhmhmhNNdmhmhmhmhmhmhmhdtmhmhmhmhmhmhNNNdmhdt1020203020212102032030620213020303031320307303101022311122331133mmmmhhmmmmhhNaNaNammhhmmhhNaNammhmhmhmhmhmhmhmhmhmhmhmhNNNdmhmhmhmhmhdtmhmhmhmhNNdmhmhdt11100018011140001110111112110118011191121510112111211332211132222mhhmmhhNaNammmmhhmmmmhhNaNaNammmhmhmhmhmhmhmhNNdmhmhmhmhmhmhmhdtmhmhmhmhmhmhNNNdmhmhmhdt121312021911211021316202131213130311020307303131112233113mmhhmmmmhhNaNaNammhhmmhhNaNamhmhmhmhmhmhmhmhmhmhNNNdmhmhmhmhmhdtmhmhmhmhNN Supplemental Figure 1
1310.1678
2
1310
2019-07-05T20:38:28
Cognitive Mechanisms Underlying the Creative Process
[ "q-bio.NC" ]
This paper proposes an explanation of the cognitive change that occurs as the creative process proceeds. During the initial, intuitive phase, each thought activates, and potentially retrieves information from, a large region containing many memory locations. Because of the distributed, content-addressable structure of memory, the diverse contents of these many locations merge to generate the next thought. Novel associations often result. As one focuses on an idea, the region searched and retrieved from narrows, such that the next thought is the product of fewer memory locations. This enables a shift from association-based to causation-based thinking, which facilitates the fine-tuning and manifestation of the creative work.
q-bio.NC
q-bio
Gabora, L. (2002). Cognitive mechanisms underlying the creative process. In T. Hewett & T. Kavanagh (Eds.), Proceedings of the Fourth International Conference on Creativity and Cognition (pp. 126-133). Cognitive Mechanisms Underlying the Creative Process Liane M. Gabora ABSTRACT This paper proposes an explanation of the cognitive change that occurs as the creative process proceeds. During the initial, intuitive phase, each thought activates, and potentially retrieves information from, a large region containing many memory locations. Because of the distributed, content-addressable structure of memory, the diverse contents of these many locations merge to generate the next thought. Novel associations often result. As one focuses on an idea, the region searched and retrieved from narrows, such that the next thought is the product of fewer memory locations. This enables a shift from association-based to causation-based thinking, which facilitates the fine-tuning and manifestation of the creative work. Keywords Associative hierarchy, bisociation, brainstorm, concepts, conjunction, context, creativity, defocused attention, distributed representation, emergent features, evaluation, focus, generativity, idea, impossibilist creativity, intuition, variable focus. INTRODUCTION What happens in the mind as a creative idea takes shape? This paper puts forth a theory of the cognitive mechanisms underlying creativity, elaborating on previous work [19, 20] to include the cognitive change that occurs as an idea transforms from inspiration to finished product. It assumes some form of homomorphism (rather than isomorphism) between the physical and the mental. STAGES OF THE CREATIVE PROCESS The creative process has a long history of being divided into stages [3, 23, 60]. It is assumed that prior to the onset of a particular creative act, the creator has acquired the tools of the trade. The first stage is known aspreparation, where the creator becomes obsessed with the problem, collects relevant data and traditional approaches to it, and perhaps attempts, unsuccessfully, to solve it. During the second stage, termed incubation, the creator does not actively attempt to solve the problem, but unconsciously continues to work on it. 1 In the third stage, illumination, a possible surfaces to consciousness in a vague and unpolished form. Subjective and theoretical accounts of this phase of the creative process speak of discovering a previously unknown 'bisociation' [33], or underlying order. For example, Poincaré [50] claims that creative ideas "reveal unsuspected kinships between other facts well known but wrongly believed to be strangers to one another" (p. 115). The classic example is Kekule's discovery of the ring-shaped structure of the benzene molecule via a dream about a serpent biting its tail. In the final stage, verification, the idea is worked into a form that can be proven and communicated to others. Some argue that an incubation period may not be necessary [61], and that the creative process can be boiled down to a generative brainstorming stage followed by an evaluative focusing stage [9, 10]. Dennett suggests that the generative-evaluative process is cyclic; a new product is generated, evaluated, new goals are set, and the cycle is repeated. COGNITIVE MODES: ASSOCIATIVE AND ANALYTIC The existence of two stages of the creative process is consistent with the widely-held view that there are two distinct forms of thought [9, 10, 30, 31, 46, 49, 54, 56]. The first is a suggestive, intuitive associative mode that reveals remote or subtle connections between items that are correlated but not necessarily causally related. This could yield a potential solution to a problem, though it may still be in a vague, unpolished form. The second form of thought is a focused, evaluative analytic mode, conducive to analyzing relationships of cause and effect. In this mode, one could work out the logistics of the solution and turn it into a form that is presentable to the world, and compatible with related knowledge or artifacts. This suggests that creativity requires not just the capacity for both associative and analytic modes of thought, but also the ability to adjust the mode of thought to match the demands of the problem, and how far along one is in solving it. What cognitive mechanisms might underlie this? ARCHITECTURE OF THE MIND Before we can piece together the cognitive mechanisms underlying creativity, we must briefly look at how episodes of experience, as well abstract items such as concepts, attitudes, and stories, are stored in memory. Memory is Sparse The human mind would have to have more memory locations than the number of particles in the universe to store all the permutations of sound, colour, and so forth that the senses are capable of detecting. Thus, the number of memory locations is much smaller than the number of possible experiences. 2 This is illustrated schematically in Figure 1. Every vertex represents a possible constellation of stimulus properties that could be present in some experience we might have, and stored as an episode in memory. (A property might be something concrete such as 'blue' or more abstract such as 'honorable', or it may be something one would be unlikely to ever think of or come up with a word for.) Only a fraction of these constellations of properties is realized as actual memory locations (those with circles on them). The memory is therefore sparse. And in fact, only a fraction of those actually has some previous episode stored in them (those with black dots in the center). Figure 1. Each vertex represents a possible memory location. Each black ring represents an actual location in a particular memory architecture or mind. The three rings with circles inside represent actual locations where an item has been stored. Degree of whiteness indicates degree of activation by current thought or experience. Activation is greatest for location k and falls with distance from k. In this case, only one location in the activated region has had something stored to it, and it is only marginally activated, so a reminding or retrieval event may or may not take place. If many memories had been stored in locations near k, they would blend to generate the next experience. Memory is Distributed but Distributions are Constrained If the mind stored each item in just one memory location as a computer does, then in order for an experience to evoke a reminding of a previous experience, it would have to be identical to that previous experience. And since the space of possible experiences is so vast that no two ever are exactly identical, this kind of organization would be somewhat useless. In neural network models of cognitive processes, this problem can be solved by distributing the storage of an item across many memory locations [26, 32, 47, 62]. Likewise, each location participates in the storage of many items. However, in a fully distributed memory, where each item is stored in every memory location, the stored items interfere with one another. (This phenomenon goes by many names: 'crosstalk', 'superposition catastrophe', 'false memories', 'spurious memories' or 'ghosts' [17, 28, 29, 59] ). The problem can be solved by constraining the distribution region. This is illustrated Figure 1; only a portion of the memory region (indicated by the degree of whiteness) gets activated. One way of constraining distributions in neural networks is to use a radial basis function, or RBF [24, 27, 36, 63]. Each input activates a hypersphere of 3 memory locations, such that activation is maximal at the center k of the RBF and tapers off in all directions according to a (usually) Gaussian distribution of width s . Both k and s are determined in a training phase. The result is that one part of the network can be modified without affecting the capacity of other parts to store other patterns. A spiky activation function means that s is small. Therefore only those locations closest to k get activated, but they get activated a lot. A flat activation function means that s is large. Therefore locations relatively far from k still get activated, but no location gets very activated. The mind is similarly constructed such that items in memory are distributed across assemblies of nerve cells [25, 37] but the distributions are constrained because assemblies are limited in size. Thus a given instant of experience activates not just one location in memory, nor does it activate every location to an equal degree, but activation is distributed across many memory locations, with degree of activation decreasing with distance from the most activated one, which we call k. The further a memory location is from k, the less activation it not only receives from the current stimulus experience but in turn contributes to the next instant of experience, and the more likely its contribution is cancelled out by that of other simultaneously activated locations. Following the neural network terminology, we say that the degree to which episodes and concepts are distributed is determined by the spikiness or flatness of the activation function. The flatter the activation function, the more distributed the memory. The choice of k does not have to be determined in advance as it is in the neural network if memory locations differ in their capacity to detect and respond to different features. There is much evidence that this works through temporal coding [1, 6, 7, 8, 14, 35, 44, 45, 48, 53, 58, for reviews see 11, 51]. As Cariani points out, temporal coding drastically simplifies the problem of how the brain coordinates, binds, and integrates information. Different features or stimulus dimensions are carried by different frequencies like a radio broadcast system, and each memory location is attuned to respond to a slightly different frequency, or set of frequencies. Thus, k arises naturally due to the differential effect of the stimulus on the various memory locations. The greater the number of stimulus frequencies impacting the memory architecture, the greater the number of memory locations that respond. Memory is Content Addressable Content addressability refers to the fact that there is a systematic relationship between the content of an experience (not just as the subject matter, but the qualitative feel of it) and the memory locations where it gets stored (and from which material for the next instant of experience is evoked). In a computer, this one-to-one correspondence arises naturally because each possible input a unique address in memory. Retrieval is thus simply a matter of looking at the address in the address register and fetching the item at the specified location. The distributed nature of human memory prohibits this, but content addressability is still achievable as follows. Memory locations are contained in nerve cells called neurons. A given experience induces a chain reaction such that some neurons are inhibited and others excited. For an experience to be engraved to a certain memory 4 location, a particular pattern of activation must occur. The 'address' of a memory location is thus the pattern(s) of activation that lead it to be affected. Associative Richness We have looked at three aspects of the architecture of memory: it is sparse, distributed (yet distributions are constrained), and content-addressable. In such a memory architecture, if the regions where two stored episodes or abstract concepts overlap, it means they share one or more common features or properties. The relationship between memories and concepts stored in overlapping regions of memory is therefore one of correlationrather than causation . Now let us look at why this turns out to be important for creativity. Martindale [39] has identified a cluster of psychological attributes associated with creativity which includes defocused attention [12, 13, 43], and high sensitivity [39, 40], including sensitivity to subliminal impressions; that is, stimuli that are perceived but of which we are not conscious of having perceived [57]. Another characteristic of creative individuals is that they have flat associative hierarchies [42]. The steepness of an individual's associative hierarchy is measured experimentally by comparing the number of words that individual generates in response to stimulus words on a word association test. Those who generate only a few words in response to the stimulus have steep associative hierarchies. Those who generate many have flatassociative hierarchies. One can also refer to this as associative richness . 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 totable) come to mind. The experimental evidence that creativity is associated with not just flat associative hierarchies but also defocused attention and heightened sensitivity suggests that associative richness stems from a tendency to perceive more of the detail of a stimulus or situation. One includes in ones' internal representation of the stimulus situation features that are less central to the concept that best categorizes it, features that may in fact make it defy straightforward classification as strictly an instance of one concept or another. This could be accomplished through a tendency toward a flat activation function. Experiences get more widely etched into memory, thus the storage regions for episodes and concept overlap more, resulting in greater potential for associations to be found amongst them. A Stream of Thought Since content addressability ensures that items with related meanings get stored in overlapping locations, one naturally retrieves items that are similar or relevant to the current experience. As a result, the entire memory does not have to be searched in order for, for example, one person to remind you of another. It is because the size of the region 5 of activated memory locations must falls midway between the two possible extremes--not distributed and fully distributed--that one can generate a stream of coherent yet potentially creative thought. The current thought or experience activates a certain region of memory. Episodes or concepts stored in the locations in this region provide the 'ingredients' for the next thought. This next thought is slightly different, so it activates and retrieves from a slightly different region, and so forth, recursively. In a state of defocused attention or heightened sensitivity to detail, stimulus properties that are less directly relevant to the current goal get encoded in memory. Since more features of attended stimuli participate in the process of storing to and evoking from memory, more memory locations are activated and participate in the encoding of an instant of experience and release of 'ingredients' for the next instant. The more memory locations activated, the more they in turn activate, and so on; thus streams of thought tend to last longer. So if a stimulus does manage to attract attention, it will tend to be more thoroughly assimilated into the matrix of associations that constitutes the worldview, and more time is taken to settle into any particular interpretation of it. We refer to this as a state of conceptual fluidity. In such a state, new stimuli are less able to compete with what has been set in motion by previous stimuli, i.e. the memory network plays a larger role in conscious experience. Another interesting consequence of a flat activation function is that an episode or concept that lies relatively far from the one that best captures the properties of the current thought, but that at least lies within the hypersphere of activated memory locations, can 'pull' the next thought quite far from the one that preceded and evoked it. Thus there is an increased probability that one thought will lead in a short period of time to a seemingly unrelated thought; consecutive instants are less correlated. So the worldview is penetrated more deeply, but also traversed more quickly. VARIABLE FOCUS AS THE KEY TO CREATIVITY There is in fact a considerable body of research suggesting that creativity is associated with, not just high conceptual fluidity, nor just extraordinary control, but both [2, 15, 16, 18, 52, 55]. As Feist [16] 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 [2] said over 30 years ago: "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). There is also evidence of an association between creativity and high variability in physiological measures of arousal such as heart rate [4], spontaneous galvanic skin response [38], and EEG alpha amplitude [39, 41]. Knowing that creativity is associated with both conceptual fluidity on the one hand, and focus or control on the other, puts us in a good position to posit an underlying mechanism: the capacity to spontaneously adjust the spikiness of the activation function 6 in response to the situation. Each new instant of thought can touches more or fewer memory locations, depending on the nature of the problem, and how far along one is in the process of solving it. We can refer to this as the capacity for variable focus. Let us now look in more detail at how it could explain what happens during the creative process from initial brainstorming stage to fine-tuning of the finished product. Brainstorming and insight Let us now consider what happens in the mind of an artist or scientist who does something particularly creative. The first response to a problem, perceived inconsistency, or desire to express oneself or generate something asthetically pleasing, may be a rational or deductive approach. When this doesn't work, it seems likely that there is a tendency to brainstorm; temporarily 'loosen' one's internal model of reality, weaken inter-concept relationships so as to allow new insights to more readily percolate through and exert the needed revolutionary impact. One becomes receptive to new ways of perceiving the world. How might this happen? Dennett [10] believes that the generative component of the creative process operates randomly. Campbell [5] claims it operates through a process of blind variation. He specifically states that by 'blind' he does not mean random, and is emphatic that it is not causal, though how it does work he does not say. Nevertheless, the notion that the generative stage is not random yet not causal is readily explainable given the attributes of memory described above. In a state of defocused attention and heightened sensitivity, more features of the stimulus situation or concept under consideration get processed. Thus the more memory locations the current instant of experience gets stored to; the activation function is flat, as in Figure 2. Figure 2. Schematic representation of region of memory activated and retrieved from during the associative stage of creative process. Stored items in all the memory locations in whitened region will blend to generate the next instant of thought. The flat activation function results in a greater likelihood of 'catching' a stored episode or concept isn't usually associated with the experience that evoked it. The new idea arises in an unpolished form; it exists in a state of potentiality, in the sense identified relationship could be resolved different ways depending on the contexts one encounters, both immediately, and down the road. For example, consider the cognitive state of the person who thought up the idea of building a snowman. It seems reasonable that this involved thinking of snow not just in 7 the newly that that terms of its most typical features such as `cold' and `white', but also the less typical feature `moldable'. At the instant of inventing snowman there were many ways of resolving how to give it a nose. However, perhaps because the inventor happened to have a carrot handy, the concept snowman has come to acquire the feature `carrot nose'. Thus new ideas arise through a sort of 'conceptual meltdown', in that the details of episodes or the meanings of concepts merge or blend into one another more than usual, such that they are more readily recombined to give something unique. Focusing the Creative Idea In the short run, a wide activation function is conducive to creativity because it provides a high probability of 'catching' new combinations of properties by reconstructing unusual blends of stored items. But maintaining it indefinitely would be untenable since the relationship between one thought and the next can be so remote that a stream of thought lacks continuity. Thus, once the overall framework of a unique idea has been painted in the broad strokes, one goes from a state of mind that is more likely to simultaneously evoke items that are correlated and therefore stored in overlapping memory locations, to a state of mind that is more conducive to establishing relationships of causation. This may happen by gradually narrowing the region of memory that gets activated, such that fewer memory locations are activated, as in Figure 3. Fewer locations release their contents to 'participate in' the formation of the next though, thus affording one finer control over what concepts gets evoked. Thought becomes focused and logical; access to remote associations, and the ensuing generation of strange new combinations, would at this point be a distraction. Figure 3. Schematic representation of region of memory activated and retrieved from during the more analytic stage of the creative process. The activation function is spiky. Through such focusing, one slowly settles on a 'draft' of the idea that incorporates aspects that are relevant, pleasing, or useful, while weeding out aspects that are irrelevant, distasteful, or misleading. In the process, the idea becomes grounded more firmly in consensus reality, so that when is more widely understandable and less vulnerable to attack. is born it it MODELING THE CREATIVE PROCESS 8 Is it possible to mathematically model the creative process? One big stumbling block is that a creative idea often possesses features which are said to be emergent: not true of the constituent ideas of which it was composed. For example, the concept snowman has as a feature or property 'carrot nose', though neither snow nor man does). Most mathematical formalisms are not only incapable of predicting what sorts of features will emerge (or disappear) in the conjunctive concept, but they do not even provide a place in the formalism for the gain (or loss) of features. This problem stems back to a limitation of the mathematics underlying not only representational theories of concepts (as well as compositional theories of language) but all classical physical theories. The mathematics of classical physics only allows one to describe a composite or joint entity by means of the product state space of the state spaces of the two subentities. Thus if X1 is the state space of the first subentity, and X2 the state space of the second, the state space of the joint entity is the Cartesian product space X1 xX 2. For this reason, classical physical theories cannot describe the situation wherein two entities generate a new entity with properties not strictly inherited from its constituents. One could try to solve the problem ad hoc by starting all over again with a new state space each time there appears a state that was not possible given the previous state space; for instance, every time a conjunction comes into existence. However, this happens every time one generates a sentence that has not been used before, or even uses the same sentence in a slightly different context. Another possibility would be to make the state space infinitely large to begin with. However, since we hold only a small number of items in mind at any one time, this is not a viable solution to the problem of describing what happens in cognition. This problem is hinted at by Boden [3], who uses the term impossibilist creativity to refer to creative acts that not only explore the existing state space but transform that state space; in other words, it involves the spontaneous generation of new states with new properties. However, this sort of problem is addressed by mathematical formalisms originally developed for quantum mechanics, as follows. When quantum entities combine, they do not stay separate as classical physical entities tend to do, but enter a state of entanglement. If H 1 is the Hilbert space describing first subentity and H 2 the Hilbert space describing second, the state space of the joint entity is described by the tensor product of the two Hilbert spaces. The tensor product always allows for the emergence of new states-- specifically the entangled states--with new properties. Because of the linearity of Hilbert space, the mathematics of pure quantum mechanics is too limited to begin to describe how concepts combine in the mind to form new ideas. However, this (and other) limitations are overcome using a generalization of the pure quantum formalism known as the state context property system, or SCOP formalism. This formalism is being used to articulate a theory of creativity by treating complex ideas as conjunctions as concepts in the context of one another [21, 22]. Whereas the role of context is generally neglected, as we see it, ideas require a context to actualize them in a thought or experience. The SCOP formalism enables us to explicitly incorporate the context that elicits a reminding of a concept, and the change of state this induces in the concept (possibly transforming it into something new and creative) into the 9 formal description of the concept itself. A concept is viewed not as a fixed mental representation, but as a source of potentiality which predisposes it to dynamically attract context-specific cognitive states (both concrete stimulus experiences and imagined or counterfactual situations) into a certain subspace of conceptual space. Interaction with the context (the stimulus or situation) causes a concept to `collapse' to an instantiated form of it. Thus a concept cannot be described in a context-independent manner (except as a superposition of every possible context-driven instantiation of it). In this view, not only does a concept give meaning to a stimulus or situation, but the situation evokes meaning in the concept, and when more than one is active they evoke meaning in each other. Each of the two concepts in a conjunction constitutes a context for the other that `slices through' it at a particular angle, thereby mutually actualizing one another's potentiality in a specific way. The stimulus situation plays the role of the measurement in physics, acting as context that induces a change of the cognitive state from superposition state to collapsed state. The collapsed state is more likely to consist of a conjunction of concepts for associative than analytic thought because more stimulus or concept properties take part in the collapse. As a metaphorical explanatory aid, if concepts were apples, and the stimulus a knife, then the qualities of the knife would determine not just which apple to slice, but which direction to slice through it. Changing the knife (the context) would expose a different face of the apple (elicit a different version of the concept). And if the knife were to slash through several apples (concepts) at once, we might end up with a new kind of apple (a conjunction). SUMMARY We have looked at a cognitive mechanism to explain what happens in the mind during the course of the creative process. Until a creative insight has been obtained, one is in an intuitive, brainstorming state of mind, and the activation function is wide. The creative insight may take many forms: for example, an invention or scientific theory, story or myth, a way or moving, acting, or accomplishing something, or a way of portraying relationship and emotion artistically. At this point the new idea is still vague and needs to come into focus. One reflects on an idea by reflecting it back into the memory with an increasingly spiky activation function, seeing what it evokes back into awareness, and repeating the process until it comes into focus. By taking what is retrieved from memory (which may consist of many items blended together) and feeding the most promising properties of this construction back at the memory, seeing what is then retrieved, and so forth, certain properties get abstracted. The initially unfocused idea eventually turns into one that can solve the problem at hand, account for the inconsistency, or convey the desired relations and emotions. The mathematical modeling of the creative process is a difficult endeavor, partly because the new idea often has properties that were not present in the constituent ideas or concepts that went into the making of it. It is possible to use a mathematical formalism that was originally devised partly to cope with the problems of context and emergent properties in the quantum world. In this model, interaction with a context (the stimulus or situation) causes a concept to `collapse' to a possibly new instantiated form of it. The collapsed state is more likely to consist of a conjunction of concepts for associative than 10 analytic thought because, due to the flat activation function, more stimulus or concept properties take part in the collapse. ACKNOWLEDGMENTS I would like to acknowledge the support of Grant FWOAL230 of the Flemish Fund for Scientific Research. The part at the end about modeling the contextual aspects of cognition using formalisms originally developed for description of contextuality in physics was developed with Diederik Aerts, and is more fully described in our joint papers. REFERENCES Abeles, M. & Bergman, H. Spatiotemporal firing patterns in the frontal cortex of behaving monkeys. Journal of Neurophysiology 70 , 4 (1993), 1629-1638. Barron, F. Creativity and Psychological Health, Van Nostrand, 1963. Boden, M. The Creative Mind: Myths and Mechanisms. Weidenfeld & Nicolson. Revised edition, Cardinal, 1990/1992. Bowers, K.S. & Keeling, K.R. Heart-rate variability in creative functioning. PsychologicalReports 29 (1971), 160-162. Campbell, D. Evolutionary Epistomology. In Evolutionary Epistomology, Rationality, and the Sociology of Knowledge, eds. G. Radnitzky &. W.W. Bartley III, Open Court, LaSalle IL, 1987. Campbell, F.W. & Robson, J.G. Application of Fourier analysis to the visibility of gratings. Journal of Physiology, 197 (1968), 551-566. Cariani, P. As if time really mattered: temporal strategies for neural coding of sensory information. In Origins: Brain and self-organization , ed. K. Pribram. Erlbaum, Hillsdale NJ, 1995, 161-229. Cariani, P. Temporal coding of sensory information. In Computational neuroscience: Trends in research 1997, ed. J.M. Bower. Plenum, Dordrecht, Netherlands, 1997, 591-598. Dartnell, T. Artificial intelligence and creativity: An introduction. Artificial Intelligence and the Simulation of Intelligence Quarterly 85 (1993). Dennett, D. Brainstorms: Philosophical Essays on Mind and Psychology , Harvester Press, 1978. 11 De Valois, R.L. & De Valois, K.K. Spatial Vision. Oxford University Press, Oxford UK, 1988. Dewing, K. & Battye, G. Attentional deployment and non-verbal fluency. Journal of Personality and Social Psychology 17 (1971), 214-218. Dykes, M. & McGhie, A. A comparative study of attentional strategies in schizophrenics and highly creative normal subjects. British Journal of Psychiatry 128 (1976), 50- 56. Emmers, R. Pain: A Spike-Interval Coded Message in the Brain. Raven Press, Philadelphia PA, 1981. Eysenck, H.J. Genius: The Natural History of Creativity, Cambridge University Press, Cambridge UK, 1995. Feist, G.J. The influence of personality on artistic and scientific creativity. In: Handbook of Creativity, ed. R. J. Sternberg, Cambridge University Press, Cambridge UK (1999), 273-296. Feldman, J.A. & Ballard, D.H. Connectionist models and their properties. Cognitive Science 6 (1982), 205-254. Fodor, E.M. Subclinical manifestations of psychosis-proneness, ego-strength, and creativity. Personality and Individual Differences 18 (1995), 635-642. Gabora, L. The beer can theory of creativity. In Creative Evolutionary Systems, eds. P. Bentley & D. Corne, Morgan Kauffman (2002), 147-161. Available athttp://cogprints.soton.ac.uk/documents/disk0/00/00/09/76/ Gabora, L. Toward a theory of creative inklings. In (R. Ascott, Ed.) Art, Technology, and Consciousness, Intellect Press (2000), 159-164. Available at http://cogprints.soton.ac.uk/documents/disk0/00/00/08/56/ Gabora, L. & Aerts, D. Contextualizing concepts. Proceedings of the 15th International FLAIRS Conference (Pensacola Florida, May 2002), American Association for Artificial Intelligence, 148-152. Gabora, L. & Aerts, D. Contextualizing concepts using a mathematical generalization of the quantum formalism. Journal of Experimental and Theoretical Artificial Intelliegence, to appear in special issue on concepts and categories (2002). http://www.vub.ac.be/CLEA/liane/papers/jetai.pdf Hadamard, J. The Psychology of Invention in the Mathematical Field . Princeton University Press, Princeton NJ, 1949. 12 Hancock, P.J.B., Smith, L.S. & Phillips, W.A. A biologically supported error-correcting learning rule. Neural Computation 3, 2 (1991), 201-212. Hebb, D.O. The Organization of Behavior. Wiley, 1949. Hinton, G., McClelland, J.L. & Rummelhart, D.E. Distributed representations. In Parallel distributed processing: Explorations in the microstructure of cognition, eds. Rummelhart, D. E. & J. L. McClelland. MIT Press, Cambridge MA, 1986. Holden, S.B. & Niranjan, M. Average-case learning curves for radial basis function networks. Neural Computation 9, 2 (1997), 441-460. Hopfield, J.J. Neural networks and physical systems with emergent collective computational abilities', Proceedings of the National Academy of Sciences 79 (1982), 2554-2558. Hopfield, J.J., Feinstein, D.L. & Palmer, R.G. "Unlearning" has a stabilizing effect in collective memories. Nature 304 (1983), 158-159. James, W. The Principles of Psychology. Dover, New York, 1890/1950. Johnson-Laird, P.N. Mental Models. Harvard University Press, Cambridge MA, 1983. Kanerva, P. Sparse Distributed Memory, MIT Press, Cambridge MA, 1988. Koestler, A. The Act of Creation. Macmillan, 1964. Lestienne, R. Determination of the precision of spike timing in the visual cortex of anesthetized cats. Biology and Cybernetics 74 (1996), 55-61. Lestienne, R. amd Strehler, B.L. Time structure and stimulus dependence of precise replicating patterns present in monkey cortical neuron spike trains. Neuroscience (April 1987). Lu, Y.W., Sundararajan, N. & Saratchandran, P. A sequential learning scheme for function approximation using minimal radial basis function neural networks. Neural Computation 9, 2 (1997) 461-478. Marr, D. A theory of the cerebellar cortex. Journal of Physiology , 202 (1969), 437-470. Martindale, C. Creativity, consciousness, and cortical arousal. Journal of Altered States of Consciousness 3 (1977), 69-87. Martindale, C. Biological bases of creativity. In Handbook of Creativity , ed. R. J. Sternberg, Cambridge University Press, Cambridge UK (1999), 137-152. 13 Martindale, C. & Armstrong, J. The relationship of creativity to cortical activation and its operant control. Journal of Genetic Psychology 124 (1974), 311-320. Martindale, C. & Hasenfus, N. EEG differences as a function of creativity, stage of the creative process, and effort to be original. Biological Psychology 6 (1978), 157-167. Mednick, S.A. The associative basis of the creative process. Psychological Review 69 (1962), 220-232. Mendelsohn, G.A. Associative and attentional processes in creative performance. Journal of Personality 44 (1976), 341-369. Metzinger, T. Faster than thought: Holism, homogeneity, and temporal coding. In Conscious Experience, ed. T. Metzinger. Schoningh/Academic Imprint, Thorverton U.K., 1995. Mountcastle, V. Temporal order determinants in a somatosthetic frequency discrimination: Sequential order coding. Annals of the New York Academy of Science 682 (1993), 151-170. Neisser, U. The multiplicity of thought. British Journal of Psychology 54 (1963), 1-14. Palm, G. On associative memory. Biological Cybernetics 36 (1980), 19-31. Perkell, D.H. & Bullock, T.H. Neural coding. Neurosciences Research Program Bulletin 6, 3 (1968), 221-348. Piaget, J. The Language and Thought of the Child. Routledge & Kegan Paul, London, 1926. Poincare, H. The Foundations of Science. Science Press, Lancaster PA, 1913. Pribram, K.H. Brain and Perception: Holonomy and Structure in Figureal Processing. Erlbaum, Hillsdale NJ, 1991. Richards, R.L., Kinney, D.K., Lunde, I., Benet, M., & Merzel, A. Creativity in manic depressives, cyclothymes, their normal relatives, and control subjects. Journal of Abnormal Psychology 97 (1988), 281-289. Riecke, F. & Warland, D. Spikes: Exploring the Neural Code . MIT Press, Cambridge, MA, 1997. Rips, L.J. Necessity and natural categories. Psychological Bulletin 127, 6 (2001) 827-852. Russ, S.W. Affect and Creativity. Erlbaum, Hillsdale NJ, 1993. 14 Sloman, S. The empirical case for two systems of Reasoning. Psychological Bulletin 9, 1 (1996), 3-22. Smith, G.J.W. & Van de Meer, G. Creativity through psychosomatics. Creativity Research Journal 7 (1994), 159-170. Stumpf, C. Drug action on the electrical activity of the hippocampus. International Review of Neurobiology 8 (1965), 77-138. Von der Malsburg, C. Am I thinking assemblies? In Proceedings of the 1984 Trieste Meeting on Brain Theory, ed. G. Palm & A. Aertsen. Springer-Verlag, Berlin, 1986. Wallas, G. The Art of Thought. Harcourt, Brace & World, New York, 1926. Weisberg, R. Creativity: Genius and Other Myths. Freeman Press, New York, 1986. Willshaw, D.J. Holography, associative memory, and inductive generalization. In Parallel Models of Associative Memory, eds. G.E. Hinton & J.A. Anderson, Lawrence Earlbaum Associates (1981), 83-104. Willshaw, D. J. & Dayan, P. Optimal plasticity from matrix memory: What goes up must come down. Journal of Neural Computation 2 (1990), 85-93. 15
1711.08533
1
1711
2017-11-22T22:59:52
Universality of macroscopic neuronal dynamics in Caenorhabditis elegans
[ "q-bio.NC" ]
Recordings of whole brain activity with single neuron resolution are now feasible in simple organisms. Yet, it is still challenging to appropriately simplify such complex, noisy, and multivariate data in order to reveal general principles of nervous system function. Here, we develop a method that allows us to extract global brain dynamics from pan-neuronal imaging. Success of this method is rooted in a surprising mathematical connection between dimensionality reduction and a general class of thermodynamic systems. Application of this theoretical framework to the nervous system of C. elegans reveals the manifold that sculpts global brain dynamics. This manifold allows us to predict switches between worm behaviors across individuals, implying that macroscopic dynamics embodied by the manifold are universal. In contrast, activation of individual neurons differs consistently between worms. These findings suggest that brains of genetically identical individuals express distinct microscopic neuronal configurations which nonetheless yield equivalent macroscopic dynamics.
q-bio.NC
q-bio
Universality of macroscopic neuronal dynamics in Caenorhabditis elegans. Connor Brennan1 & Alex Proekt2 1Department of Neuroscience, University of Pennsylvania 2Department of Anesthesiology and Critical Care, University of Pennsylvania ∗To whom correspondence should be addressed; E-mail: [email protected]. Recordings of whole brain activity with single neuron resolution are now fea- sible in simple organisms. Yet, it is still challenging to appropriately simplify such complex, noisy, and multivariate data in order to reveal general princi- ples of nervous system function. Here, we develop a method that allows us to extract global brain dynamics from pan-neuronal imaging. Success of this method is rooted in a surprising mathematical connection between dimen- sionality reduction and a general class of thermodynamic systems. Applica- tion of this theoretical framework to the nervous system of C. elegans reveals the manifold that sculpts global brain dynamics. This manifold allows us to predict switches between worm behaviors across individuals, implying that macroscopic dynamics embodied by the manifold are universal. In contrast, activation of individual neurons differs consistently between worms. These findings suggest that brains of genetically identical individuals express dis- tinct microscopic neuronal configurations which nonetheless yield equivalent macroscopic dynamics. 1 Both deterministic (1, 2) and stochastic (3–5) models have been successfully used to un- derstand brain function. Detailed biophysical models can be used to faithfully model voltage dynamics in single neurons (1, 2, 6, 7) while pairwise correlations between firing of individual neurons can predict the likelihood of global firing patterns (3–5). Yet, detailed biophysical mod- els of even simple brains are experimentally intractable (8), computationally costly (9, 10), and not necessarily conceptually revealing (8, 11, 12). On the other hand, stochastic models are not sufficient to reproduce the observed macroscopic dynamics (13, 14). Significant simplification is necessary in order to arrive at a model of the nervous system at the behaviorally-relevant macroscopic level. Yet, it is not clear how the variability observed at the microscopic level in the nervous system can be appropriately simplified. Collective macroscopic properties are commonly understood as emergent epiphenomena that arise as a consequence of microscopic interactions but do not themselves affect the sys- tem (15). Yet, from an evolutionary standpoint, selection operates on the level of organismal behavior (16) mediated by macroscopic brain dynamics rather than individual ion channels or neurons. Thus, here we hypothesized that this evolutionary constraint will manifest as invari- ance of macroscopic brain dynamics. However, because multiple microscopic configurations can give rise to equivalent macroscopic dynamics it is possible for individual components of the system to vary across individuals. We test this hypothesis in nematode C. elegans because of its unique advantages as a model organism. All 302 neurons (17) and all of their connections are known (18–20). Simultaneous recordings of the majority of the neurons in the brain (head ganglia) of the worm have been performed in vivo (21–24) (Figure 1A, Supplement S1) and made publicly available. Motor commands of the worm can be inferred from activation of well-characterized neurons (21, 25– 27). For instance activation of the AVA neuron can be used to infer the direction of locomotion (Figure 1B). 2 Changes in behavior are associated with coherent changes in the activity of multiple neurons (28–31) even in the simple nervous system of C. elegans (21) (Figure 1). Correlations among neurons can be used to reduce the dimensionality of neuronal activity using principal component analysis (PCA) (21) (Figures S6 and S21). Indeed, in the nervous system of C. elegans only two principal components (PCs) are required to extract approximately 60% of the fluctuations in neuronal activity (21). There is, however, a fundamental distinction between neuronal activity and neuronal dynam- ics (31, 32). Neuronal activity is an output of the dynamical system governed by the biophysics of individual neurons and their connections (33–35). These biophysical processes influence neuronal activity and are in turn influenced by it. Yet they are not directly apparent in record- ings of neuronal activity. Neuronal dynamics can be thought of as a trajectory through the space spanned by the rele- vant variables. This trajectory can never cross itself in a deterministic dynamical system (36). In real systems, however, measurement noise and stochastic processes lead to some tangling of the trajectories. Thus, a combination of deterministic and stochastic approaches is necessary. Here we successfully combine deterministic and stochastic approaches to discover trajectory bundles (37) that together form a manifold. This manifold governs macroscopic dynamics of the C. elegans nervous system. Our key finding is that the shape of this manifold is conserved among individuals. This universality allows us to make predictions concerning future behaviors of the worm. Remarkably these predictions are valid across individuals despite the fact that activations of some identified neurons are consistently and qualitatively different among worms within a clonal population. 3 Figure 1: behavioral assignment of C. elegans pan-neuronal imaging A) Heat plot of fluores- cence (∆F/F) for 129 neurons in C. elegans head ganglia recorded by Kato et al (21). Neurons 1 to 15 (above horizontal red line) were identified in all five individual worms (21). For the purposes of com- paring activity between worms this measure was normalized and expressed as z-score. B) Trace of AVA neuron colored according to the inferred behavioral state (top) (21). Solid black line shows the duration of a sample backing bout, defined as repeated backing behavior uninterrupted by forward locomotion. Vertical red lines show beginning of a backing bout. Neuronal activity in A projected onto the first two principal components (PC1 and PC2) (bottom). Turns and reversals are not easily assigned on the basis of PC1 alone. Thus, only forward and backward locomotion are labeled (blue and green respectively). 4 Stochastic processes alone are insufficient to model single worm dynamics. Assuming only that the processes giving rise to the observed neuronal activity do not change appreciably during the time course of the experiment, and given the constraints of observed neu- ronal statistics (3,4), the simplest model of neuronal dynamics is Brownian motion (Supplement S2) (38). To construct this model, we estimate the probability density function P (Figure 2A, middle panel), by binning neuronal activity (Figure 2A, bottom) in the plane spanned by the first two principal components (PC1 and PC2). Brownian motion given by dX dt = D ∇P (X) P (X) + , (1) is determined entirely by P . Diffusion constant D and noise  set the temporal scale but do not influence the dynamics. In this case X is the position vector in the plane spanned by PC1 and PC2. Dashed arrows (Figure 2A top panel) show the flux of neuronal activity predicted by Brownian motion – preferred direction always tends towards the local energy minimum. While simulations of Brownian motion recapitulate the statistics, they fail to qualitatively or quantitatively recapitulate the observed neuronal or behavioral dynamics even for a single worm (Figures 2B and 2C, respectively). Thus, stochastic processes alone are unable to give rise to the observed neuronal dynamics. Local flux of neuronal activity is sufficient to model a single worm but fails to generalize across worms. The flux of neuronal activity computed as the observed transition probability between neighbor- ing bins rather than the gradient of P (Eq. 1) is shown in Figure 2D. We refer to the dynamics given by these experimentally observed transition probabilities as the Empirical Markov Model (EMM). EMM simulations not only recapitulate the observed statistics (Figure 2D, top) but 5 also the qualitative and quantitative aspects of neuronal dynamics (Figure 2E-F) within a sin- gle worm. Near energy minima both models have similar fluxes (red arrows, Figures 2A and 2D). EMM is distinguished, however, by the presence of an additional flux component (blue arrows, Figure 2D) (Figure S7). Because this additional component cannot be inferred from P alone, we refer to it as deterministic. Note that the deterministic component of the flux is cyclic (Figures 2D and S7). Surprisingly, attempts to model neuronal dynamics by directly projecting all worms onto a common PCA-based coordinate system followed by application of EMM fail catastrophically (Figure 2G) despite being successful in each worm individually (Figure S8). One reason for this failure is that only 15 neurons were consistently identified across worms (Figure S8). The pri- mary problem, however, is that projecting all worms onto a common coordinate system leads to tangling of trajectories (Figure S10 and S11). Consequently, the flux computed from the cross- worm EMM model is similar to that predicted by Brownian motion (Figure S12 and S13). While including more principal components could in principle improve generalizability of EMM, the amount of data needed to constrain P grows exponentially with the number of dimensions and quickly becomes experimentally intractable (39). Thus, in order to be experimentally tractable, dynamics have to lie in a low dimensional manifold. Yet, lack of generalizability of EMM illus- trates that there is no simple linear relationship between neuronal activity and the macroscopic variables that sculpt neuronal dynamics. 6 7 Figure 2: Neuronal dynamics as diffusion in neuronal activity space A-C) Brownian motion is not sufficient to capture neuronal dynamics in a single worm. A) Neuronal activity projected onto PC1 and 2 and colored according to the behavioral state (bottom). Empirically derived distribution of points (middle). Blue shows high probability (low energy). Red shows low probability (high energy). Brownian motion predicts that neuronal activity will flow down the gradient of the distribution (arrows top). B) Simulated neuronal activity (only PC1 shown) is qualitatively different from observed neu- ronal dynamics (Figure 1B) C) Simulated dwell times in different behavioral states differ from those observed experimentally. D-F) Empirical Markov model (EMM) captures dynamics in a single worm. D) Neuronal activity projected as in A-C. Unlike in A-C, individual snapshots are now connected accord- ing to the order in which they were empirically observed. Empirically observed transition probabilities are shown as flow (middle). The distribution computed by simulating empirically observed transition probabilities (D, top) is identical to the empirically observed distribution (A, middle). E) Simulation of neuronal activity projected onto PC1 qualitatively recapitulates observed dynamics. F) Dwell time distributions of behavioral states match those observed empirically. G) Empirical Markov model fails to generalize across worms. Projection of recordings from the 15 neurons identified across all worms onto common PCs fails to give rise to a meaningful EMM. Simulated PC1 (left) and predicted dwell time distributions of behavioral states (right) are qualitatively and quantitatively different from the observed neuronal activity. 8 Macroscopic dynamics are conserved among worms. While worms may differ in terms of specific microscopic details of neuronal activation, we hy- pothesized that at the collective macroscopic level dynamics will be conserved between worms. This hypothesis asserts that when projected onto a suitable coordinate system the trajectories of neuronal activity, i.e. the manifolds, have similar shapes in different individuals (Figure 3). This coordinate system, known as the phase space, contains all of the relevant dynamical variables. Consistent with the hypothesis, we identified a phase space in which the manifold con- structed by projecting activity of 107 neurons in a single worm (Figure 3A) is essentially identi- cal to the manifold constructed across individual worms using just the common set of 15 neurons (Figure 3B) (Supplement S4). To construct the manifolds in Figure 3 we averaged neuronal tra- jectory with respect to the phase of the cyclic flux rather than with respect to neuronal activity (Figures 2G, S10-S13,). Thus, phase of the cyclic flux (arrow) is the first relevant macroscopic dynamical variable. In addition to the phase, the only other relevant variable is identity of the flux – specifying which loop the system is currently in. Therefore, the phase space of the C. elegans nervous system is only two dimensional (Figure S16). The separation of trajectories in phase space allows for near perfect prediction of behavioral statistics (Figure 3C-D) as well as efficient simulations of whole-brain dynamics (Figure S21) valid for each worm individually and collectively across worms. To further strengthen the argument for manifold invariance across worms, we constructed a manifold based on the data from four out of the five worms and used it to predict the behavior of the fifth worm not used in the construction of the manifold (Figure 3E, S17, S18). Remarkably, the manifold provides meaningful information not just about statistics of be- haviors but also about behavioral transitions on a cycle-by-cycle basis. In (Figures 3F-I) we focus on predicting the termination of backwards locomotion (transition from green to yel- 9 low/orange). Figures S19 and S20 show other behavioral transitions. In Figure 3G, we compute the time until termination as a function of phase along the manifold (Supplement S3). As ex- pected, the time to behavioral transition decreases as the phase evolves towards the transition zone between behaviors (orange trace). This cannot be predicted based on dwell time statistics alone (blue trace). As the system nears the point of transition, the influence of the stochastic- ity decreases and uncertainty (entropy) of the prediction decreases concomitantly (Figure 3H orange). This information is also not contained in dwell time statistics (Figures 3H blue trace, S19 and S20). There is a key distinction between neuronal activity and neuronal dynamics. Neuronal ac- tivity changes abruptly (Figure 1) and transitions appear stochastic. In contrast, the manifold space is smooth. This is akin to smooth Hodgkin Huxley variables giving rise to abrupt changes in membrane voltage such as action and plateau potentials (6). The smoothness of the manifold exposes the deterministic dynamics that depend predominantly on the phase of the cyclic flux. Yet this phase is not readily apparent in observations of neuronal activity. 10 11 Figure 3: Universality of manifold allows cross worm predictions. A) Manifold of coherent trajectory bundles for a single worm computed on the basis of 100+ neurons. Color shows behavioral state, width shows point density, arrows show the preferred direction of motion along each loop. Neu- ronal trajectory averaged with respect to phase of the flux is projected onto the first 3 PCs. PCA is not required for the simulations or quantitative predictions and is solely used for illustration B) Manifold of neuronal dynamics computed for all worms using only the 15 neurons shared across worms. The shape of the manifold in B is essentially identical to A. C) AVA neuron simulated using the all-worm manifold (Figure 3B) is qualitatively similar to observed AVA activity (Figure 1B). D) Statistics of simulated be- havioral dwell times are essentially identical to observed behavioral statistics (forward blue; backward green; backing bouts black). E) Manifold constructed on the basis of 4 worms. Raw data from the left out fifth worm is superimposed onto the manifold. Manifold in the same orientation as A and B (left). Man- ifold rotated to highlight the coherence in trajectories (right). F-I) Predictions of behavioral transitions on a cycle-by-cycle basis. F) Evolution of the system from the circle to the diamond marker is tracked by phase along the cyclic flux. G) Each point on the manifold is identified by the time since the initiation of backward locomotion and independently by the phase of the cyclic flux. The former gives rise to a prediction for termination of backwards locomotion based on behavioral statistics alone (mean shown by blue line, shading shows 95% confidence interval). The latter gives rise to the prediction based on the manifold (mean shown by orange line, shading shows 95% confidence interval). Time to termination is progressively shortened as the phase proceeds from the circle to the diamond marker. This information is not contained in the behavioral statistics alone. H) Uncertainty (entropy) about expected time until termination of backward locomotion for manifold-based model (orange) is lower than for dwell time- based model (blue) for all phases of the manifold. Uncertainty decreases progressively as phase evolves from circle to diamond. I) Empirical distributions of time until termination at the start (circle) and end (diamond) of the subsection. 12 Construction of an asymmetrical transition probability matrix to identify the phase of the flux of neuronal activity. Altogether, findings in Figure 3 strongly argue that the phase of the cyclic flux is a key macro- scopic dynamical variable of C. elegans nervous system. Cyclicity of neuronal dynamics (Fig- ures 3, 2D and S7) is not accidental – it naturally arises in stochastic differential equations. To show this, we start with the law of probability conservation dP (X) dt = −∇J(X, t). (2) Where J(X, t) is the flux at state X and time t. We first assume steady state dP (X) dt = 0 and then use the Fokker-Planck equation (38) – which describes the time-evolution of driven stochastic systems – to arrive (Supplement S2) at the stochastic partial differential equation that governs the temporal evolution of the broad class of systems with both stochastic and deterministic processes dX dt = D ∇P (X) P (X) + J(X) P (X) + . (3) The first term in Eq. 3 is Brownian motion. The second term contains the deterministic flux J. Brownian motion is the trivial solution (J = 0) often assumed in stochastic models (3). The non-trivial solution (J (cid:54)= 0) satisfies the steady state assumption so long as the flux is cyclic – that is, if the net flux along any loop is zero (40, 41). Thus, divergence-free cyclic flux is the most parsimonious solution that at once satisfies the steady state assumption and allows for deterministic dynamics. Since Eq. 3 naturally arises in reaction-diffusion systems, we adapt diffusion mapping (42– 44) to express distances between two different states of the nervous system (X and Y) as the probability of transition between them PX→Y = exp (cid:32) −(cid:107)X − Y(cid:107)2 2 4ε (cid:33) , (4) 13 where ε is a measure of the size of the local neighborhood and (cid:107)·(cid:107)2 is the Euclidean distance. Eq. S22 describes a Gaussian distribution which peaks at X and decays as a function of distance from it. Transition probabilities between all pairs of states can be summarized in a transition probability matrix M commonly referred to as a diffusion map (DM) because of the analogy between diffusion and transition probability (Supplement S4). Yet, by construction M is symmetric PX→Y = PY →X and therefore compatible only with Brownian motion (J = 0). This symmetry arises because Eq. S22 does not take into account the order in which states X and Y are observed. To illustrate how DMs can be modified to include dynamics we use a simple system (Supplement S5) governed by Eq. 3 (Figure 4A). We denote the state of the system at time t as Xt. Because of the cyclic flux, the system can readily transition from Xt to Xt+1 but not backwards (Figure 4C). We use this simple intuition to modify DMs as PXt→Y = exp (cid:32) −(cid:107)Xt+1 − Y(cid:107)2 2 (cid:33) . (5) 2σ2 The key difference from Eq. S22 is that the transition probability peaks at the next empirically observed state of the system (Figure 4C). σ in Eq 5 is a data-based estimate of the size of the local neighborhood (Supplement S4). Thus, the system in state Xt will most likely transition to the next empirically observed state Xt+1, but may alternatively transition to a state near Xt+1, corresponding to a different nearby trajectory (Figure 4D). The transition probability after N steps, given by MN (Figure 4E), reveals robustly recurrent behavior (red shows nearby points). Diagonal bands in this recurrence plot occur when the trajectory has parallel segments separated in time (45) and thus reveal the existence of cyclic flux along the trough of the energy landscape. Figure 4F shows the time evolution of the distribution of points – a kymograph – driven by both stochastic and deterministic processes in the infinite data limit. Using asymmetrical M we reconstruct the kymograph in Figure 4F, using a simulation of just four revolutions around the trough (Figure 4G, red line). Despite undersampling, the estimated kymograph closely 14 approximates the results obtained in the infinite data limit. This strength of the asymmetric transition matrix is indispensable for the reconstruction of the manifold on the basis of short recordings of neuronal activity. The key mathematical insight is that, after appropriate normalization, eigenvalues and eigen- vectors of M approximate those of the discrete version of Eq. 3 (42). Asymmetry of M admits complex eigenvalues that give rise to cyclic fluxes (second term of Eq. 3). M can therefore be directly implemented to simulate Eq. 3. At steady state, all fluxes except for the one with the largest associated complex eigenvalue dampen out and may be safely ignored. Thus, eigenmode decomposition of M naturally leads to dimensionality reduction that preserves the most salient dynamics. The fundamental advan- tage of asymmetrical DMs is that the relationship between the observed states of the system (elements of M) and the relevant macroscopic dynamical variable – phase – is given by the eigenvectors of M. In the case of the system in Figure 4 phase is the only relevant variable, while in C. elegans (Figure 3) one additionally needs to specify the which loop the system is currently in (flux ID). 15 16 Figure 4: Asymmetric diffusion maps recover phase of the cyclic flux. A) A simulated sys- tem consists of a potential (mesh) with associated Brownian motion (white arrows), and cyclic flux (red arrows). B) Simulation of the system without cyclic flux has no persistent phase velocity (top). With cyclic flux the direction of the rotation is consistent but, because the magnitude of the flux is not uni- form, the velocity is variable (bottom). C) Schematic of asymmetric transition probability. The arrow indicates the direction of preferred flux. Transition probabilities from the state of the system at time t (Xt) are computed as in Eq. 5 (red → high probability; blue → low probability). Transition to the next observed state of the system Xt+1 is most likely but transitions to parallel trajectories are also possible. D) Asymmetrical transition probability matrix M computed for all pairs of states. M is sparse because for each point, transition probability is nonzero just for k = 12 nearest neighbors. The parallel diagonal bands highlight the high level of recurrence in the system. E) Exponentiation of M further highlights the asymmetry. The dashed line shows the main diagonal illustrating that the parallel bands are off center. F) Simulated time evolution of distribution of points (kymograph) created by averaging and smoothing 1000 individual simulations of the system. G) Reconstruction of the kymograph using M estimated from a single simulation of four revolutions around the trough (red line). 17 Relevant macroscopic variables are revealed by delay embed- ding observed neuronal activity. It is typically assumed that there is one-to-one correspondence between a snapshot of neuronal activity and position in phase space of the brain (3). Yet, at any given moment, neuronal activity is influenced by myriad unobserved biophysical variables. Thus, trajectories originating from identical instantaneous neuronal activity patterns can evolve in different directions depending on the state of these unobserved variables. Indeed, trajectories in the space spanned by neuronal activity are tangled (Figures 2G, S10, S11). While it is impossible to simultaneously observe all of the microscopic variables that define even the simplest nervous system, one can nonetheless reconstruct system dynamics from the observed neuronal activity based on Takens delay embedding theorem (37,45–47) (Figure 5C). To illustrate how delay embedding works we use a simple system (Supplement S5) of two coupled differential equations (Figure 5A) summarized in a single function f dX dt = f (O, H). (6) The position X in phase space is given by the pair of dynamical variables (O, H) and only O is experimentally observed (Eq. S27 and S28). However, given a suitably chosen delay time τ and the number of embeddings n, one can construct a function φ only on the basis of O such that the dynamics given by f and φ are essentially equivalent (46). = φ(O(t), O(t − τ ), O(t − 2τ ), ..., O(t − nτ )), (7) dX dt For the system given by Eq. 6 this approach is shown in 5A-D. Delay embedding is a key aspect in the manifold construction for the C. elegans nervous system. The eigenvectors of M map delay embedded neuronal activity onto the state space spanned by phase and flux ID. Without delay embedding, M fails to produce meaningful pre- 18 dictions (Figure 5E). Thus, there does not appear to be a simple one-to-one mapping between snapshots of neuronal activity and states of the brain. This conclusion does not rest only on the manifold construction method (Supplement S3). In Figure 5F-I we attempt to directly predict the onset of a backing bout (Figure 1 vertical dashed red line) given an antecedent snapshot of neuronal activity. This prediction works well within a single worm with either all recorded neurons (Figure 5F) or the 15 neurons identified in all worms (Figure 5G). Thus, within each worm instantaneous neuronal activity and behavior are closely related. Yet, the prediction degrades dramatically across worms (Figure 5H-I blue) suggesting that the relationship between neuronal activity and behavior is inconsistent among individuals. This is consistent with failure of EMM and diffusion maps constructed on the basis of raw neuronal activity and argues strongly that these failures are not due to dimensionality reduction or binning. Remarkably, delay embedding restores the ability to predict across worms (Figure 5F-I or- ange). Thus, data in Figure 5E-I strongly imply that the instantaneous neuronal activity varies among individuals and is not sufficient to specify the state of the brain. Yet, the slow dynamics extracted by the manifold are universal. To illustrate this further, we plot normalized activation of representative neurons as a func- tion of behavioral phase in different worms. This reveals that even within a clonal population, individual worms exhibit qualitative differences between activation of some neurons (Figure 5J, Figures S22) while other neurons (Figure S23) are consistent among worms. Furthermore, the same identified neuron can be consistent among worms in one behavior but not in others (Figure S24). This makes it unlikely that the observed differences in neuronal activations among worms can be attributed to an artifact of imaging or neuron identification. 19 20 Figure 5: Necessity of Delay Embedding for Extraction of Universal Neuronal Dynamics. A) Time evolution of the system given by Eq. 6 is described by two variables (O → blue; H → orange). Only O is experimentally observed B) Phase space plot of the system in A. C) Phase space reconstructed using approach in Eq 7. For visualization purposes only the dimensionality is reduced by projecting onto the first two PCs. D) The hidden variable H is reconstructed from the delay-embedded manifold. E) Manifold constructed as in 3 on the basis of the 15 shared neurons across worms, fails to qualitatively (left) or quantitatively (right) predict behavioral statistics (forward blue; backward green; backing bouts black) without delay embedding. F-I) Comparison of predictive power on initiations of backing bouts (red dashed line, Figure 1) using different combinations of worms and delay embedding. Leave one out analysis calculates the prediction model based on the data of 4 worms and predicts events for the excluded worm. Blue bars are computed using raw activity while orange bars use delay-embedded data. p-values using two-sample t-test: single worm (all neurons) → 0.85, single worm (shared neurons) → 10−5, leave one out (shared neurons) → 10−38 and all worms (shared neurons) → 10−84 J) Average normalized traces of specific neuronal activations during specific behaviors.Dotted lines show 95% con- fidence intervals. Traces are colored according to the identify of the worm.To account for differences in duration of individual behaviors, time is converted to phase which runs from the initiation to termination of each behavior. While in each invidual worm activation is consistent, there are qualitative consistent differences between worms with respect to certain neurons. p-values of traces coming from the same distribution (1-way ANOVA): AVAL → 2.1 × 10−7, RMED → 3.7 × 10−17, SMDVR → 1.0 × 10−9 21 Discussion. Here we developed a method for extracting salient dynamical features from complex, multivari- ate, nonlinear, and noisy time series which arise in a variety of experimental settings such as gene expression, signal transduction, financial fluctuations, and others. We apply this method to whole brain imaging in C. elegans to demonstrate its use in simulating activity of the ner- vous system and predicting in a model-free fashion switches between different behaviors across individuals. The manifold in C. elegans nervous system is composed of two cyclic fluxes corre- sponding to forward and backward locomotion. While the system is in either one of the fluxes, its fate is entirely predictable. Yet, in the neighborhood where fluxes merge, the behavior cannot be clearly predicted and stochastic forces play a stronger role. This implies that the region where the two cyclic fluxes merge is a decision point where the nervous system is most susceptible to noise and/or sensory inputs. The manifold shape is conserved among individuals. Yet, this universality is not observed at the microscopic level of individual neurons. While some neurons exhibit consistent activation patterns across different worms, many others do not. This striking observation is not without precedent. Hodgkin-Huxley models of conductances measured in individual AB neurons in crustacean stomatogastric ganglion exhibit bursting akin to the biological neuron. However, averaging these measurements across different AB neurons yields models that fail to burst (7). This suggests that differences between individual AB neurons (48, 49) or individual worms are not simply random deviations from a common template that can be averaged away at the microscopic level. This is because from the evolutionary perspective the relevant variables are macroscopic. While hereditary variability between individuals is indeed implemented at the microscopic level, selection operates at the macroscopic level of organismal behavior (16) embodied by the 22 global dynamics of the brain. Thus, there is no explicit selective pressure for each worm to produce identical neuronal activation during behavior. While undoubtedly there are important constraints imposed by the biomechanics of the animal, the connectome, and other variables, any degenerate solution that gives rise to the appropriate macroscopic dynamics is equally valid from an evolutionary standpoint. Even in the simplest nervous systems, it is not possible to appropriately experimentally constrain a detailed biophysical model of the brain. Yet, asym- metrical diffusion maps can be used to extract the relevant macroscopic variables that allow for predictions of brain activity valid across individuals. References 1. A. L. Hodgkin, A. F. Huxley, The Journal of physiology 116, 449 (1952). 2. A. L. Hodgkin, A. F. Huxley, B. Katz, The Journal of physiology 116, 424 (1952). 3. E. Schneidman, M. J. Berry, R. S. II, W. Bialek, Nature 440, 1007 (2006). 4. E. Schneidman, S. Still, M. J. Berry, W. Bialek, et al., Physical review letters 91, 238701 (2003). 5. G. Tkacik, et al., Journal of Statistical Mechanics: Theory and Experiment 2013, P03011 (2013). 6. R. Fitzhugh, The Journal of general physiology 43, 867 (1960). 7. J. Golowasch, M. S. Goldman, L. Abbott, E. Marder, Journal of Neurophysiology 87, 1129 (2002). 8. J. Gjorgjieva, G. Drion, E. Marder, Current opinion in neurobiology 37, 44 (2016). 9. E. M. Izhikevich, IEEE Transactions on neural networks 14, 1569 (2003). 23 10. H. Markram, et al., Cell 163, 456 (2015). 11. A. I. Selverston, Behavioral and Brain Sciences 3, 535 (1980). 12. J. Barral, A. D. Reyes, Nature 201, 6 (2016). 13. A. Tang, et al., Journal of Neuroscience 28, 505 (2008). 14. I. E. Ohiorhenuan, et al., Nature 466, 617 (2010). 15. D. T. Campbell, Studies in the Philosophy of Biology (Springer, 1974), pp. 179–186. 16. M. Lassig, A. Valleriani, Biological Evolution and Statistical Physics, vol. 585 (Springer, 2008). 17. J. G. White, E. Southgate, J. N. Thomson, S. Brenner, Phil. Trans. R. Soc. Lond 314, 1 (1986). 18. E. J. Izquierdo, R. D. Beer, PLoS computational biology 9, e1002890 (2013). 19. C. I. Bargmann, E. Marder, Nature methods 10, 483 (2013). 20. L. R. Varshney, B. L. Chen, E. Paniagua, D. H. Hall, D. B. Chklovskii, PLoS computational biology 7, e1001066 (2011). 21. S. Kato, et al., Cell 163, 656 (2015). 22. J. P. Nguyen, et al., Proceedings of the National Academy of Sciences 113, E1074 (2016). 23. R. Prevedel, et al., Nature methods 11, 727 (2014). 24. L. Tian, et al., Nature methods 6, 875 (2009). 25. Z. Li, J. Liu, M. Zheng, X. S. Xu, Cell 159, 751 (2014). 24 26. L. Luo, et al., Neuron 82, 1115 (2014). 27. J. Larsch, D. Ventimiglia, C. I. Bargmann, D. R. Albrecht, Proceedings of the National Academy of Sciences 110, E4266 (2013). 28. J. A. Michaels, B. Dann, H. Scherberger, PLoS computational biology 12, e1005175 (2016). 29. E. E. Fetz, Behavioral and brain sciences 15, 679 (1992). 30. A. P. Georgopoulos, A. B. Schwartz, R. E. Kettner, Science pp. 1416–1419 (1986). 31. E. Salinas, T. J. Sejnowski, Nature reviews. Neuroscience 2, 539 (2001). 32. M. M. Churchland, et al., Nature 487, 51 (2012). 33. H. S. Seung, Proceedings of the National Academy of Sciences 93, 13339 (1996). 34. R. D. Beer, Artificial intelligence 72, 173 (1995). 35. J. P. Miller, A. I. Selverston, Journal of neurophysiology 48, 1416 (1982). 36. S. H. Strogatz, Nonlinear dynamics and chaos: with applications to physics, biology, chem- istry, and engineering (Westview press, 2014). 37. G. Sugihara, et al., science 338, 496 (2012). 38. R. K. Pathria, Statistical mechanics (Butterworth-Heinemann, 1996). 39. C. C. Aggarwal, A. Hinneburg, D. A. Keim, ICDT (Springer, 2001), vol. 1, pp. 420–434. 40. J. Wang, L. Xu, E. Wang, Proceedings of the National Academy of Sciences 105, 12271 (2008). 25 41. H. Yan, et al., Proceedings of the National Academy of Sciences 110, E4185 (2013). 42. B. Nadler, S. Lafon, R. R. Coifman, I. G. Kevrekidis, Applied and Computational Har- monic Analysis 21, 113 (2006). 43. R. R. Coifman, S. Lafon, Applied and computational harmonic analysis 21, 5 (2006). 44. W. Lian, R. Talmon, H. Zaveri, L. Carin, R. Coifman, Signal Processing 116, 13 (2015). 45. J.-P. Eckmann, S. O. Kamphorst, D. Ruelle, EPL (Europhysics Letters) 4, 973 (1987). 46. F. Takens, et al., Lecture notes in mathematics 898, 366 (1981). 47. N. H. Packard, J. P. Crutchfield, J. D. Farmer, R. S. Shaw, Physical review letters 45, 712 (1980). 48. M. S. Goldman, J. Golowasch, E. Marder, L. Abbott, Journal of Neuroscience 21, 5229 (2001). 49. A. A. Prinz, D. Bucher, E. Marder, Nature neuroscience 7, 1345 (2004). 50. T. Schrodel, R. Prevedel, K. Aumayr, M. Zimmer, A. Vaziri, Nature methods 10, 1013 (2013). 51. M. Rubinov, O. Sporns, Neuroimage 52, 1059 (2010). 52. M. E. Newman, Proceedings of the national academy of sciences 103, 8577 (2006). Acknowledgments We thank Sarah Friedensen, Guillermo Cecchi, Marcelo Magnasco, Drew Hudson, Tom Joseph, Manuel Zimmer, and Max Kelz for critically reading the manuscript. We also thank Manuel Zimmer and his lab for sharing their recordings of neuronal activity. 26 Supplementary S1. Neuronal Imaging Calcium imaging. Our analyses use whole-brain single-cell-resolution Ca2+ imaging data published by Kato et al., 2015. The deviation of fluorescence from baseline (∆F/F ) is considered as a proxy for neuronal activity. Each of the five worms were immobilized in a microfluidic device (50) under environmentally constant conditions. The 107 to 131 neurons detected in each worm span all head ganglia, all head motor neurons and most of the sensory neurons and interneurons along with most of the anterior ventral cord motor neurons (17). Of the identified neurons for each worm there is a subset of 15 neurons which were successfully identified across all five worms which are used when generalizing models across multiple worms. The 15 unambigu- ously identified neurons are: AIBL, AIBR, ALA, AVAL, AVAR, AVBL, AVER, RID, RIML, RIMR, RMED, RMEL, RMER, VB01, VB02. This set of neurons is used in all cross worm analyses. There are 3 more neurons that are very likely (though not unambiguously) identified in all 5 worms – SMDVR, RIVR and RIVL. These neurons are considered as candidates for representative neurons for Figure 5J and Figures S22 and S23. We adopt the same behavior states defined by Kato et al., 2015. The four primary behavioral states are forward locomotion, turns (or FALL), RISE and backwards locomotion (Figure 1). More details of the experiment can be found in Kato et al., 2015. Preprocessing. No preprocessing is applied to the data for the Brownian motion and EMM model formu- lations (Figure 2). For the Fokker-Planck diffusion map method the data are smoothed with a Gaussian filter σ = 1. Time derivatives are calculated using the diff function. Further, the data ∆F/F is z-scored in each channel using the zscore function of MATLAB. These values 27 were chosen to maximize the results of each method, but we find that the results are robust to the details of this preprocessing procedure. Cross worm analyses pose two unique concerns. First, there are only 15 neurons that were identified for in all five animals. These 15 neurons account for the majority of the variance in the data and serve as our cross worm basis (Figure S6). Second, the five time series data sets are disjoint. The switches from one set to another are padded with a set of 50 zero values. This length was chosen so that the delay embedding of one worm would never interfere with the embedding from another worm. Autocorrelation of each channel was computed separately. We are concerned with the au- tocorrelation of all channels simultaneously and so use the minimum autocorrelation across all channels, which was found to be τ ≈ 10 frames. S2. Stochastic Dynamical Systems (Fokker-Planck) Background. Because neuronal systems are constantly exposed to internal and external noise, the most sensible approach is to model the system probabilistically (41), dX dt = F(X) + , (S8) where F(X) is the driving force, X is the position in state space and  is noise. Since we are concerned with the overall statistics of the system we will consider not the trace of a single trajectory, but the flow of probability about the neuronal space (38). The law of probability conservation, dP (X) dt = −∇J(X, t), (S9) states that the change in probability P is due to the local flux, J(X, t), in that region. In systems 28 with homogeneous (constant in space) noise the probability flux is defined by the equation J(X, t) = F(X)P − ∇P, (S10) where F(X) is the driving force of the system at state X. Our first model, Brownian motion, assumes that the only dynamics acting on the system are diffusion. The steady state distribution of this model much match the true probability distribution observed in the data. At steady state the probability distribution no longer changes with time and so, dP (X) dt = 0 = ∇J(X, t), (S11) where P is the steady state probability distribution function of the model. In the purely diffusive model we assume that flux of the system vanishes at all points, J(X, t) = 0, which satisfies the constraint that the divergence of the flux be 0. The relationship between driving force and flux further implies that F(X) = D ∇P (X) P (X) . (S12) The above uses the trivial solution to Eq. S11 (J = 0), but another class of solutions exist when the flux does not vanish at steady state (41). These solutions have flux that is purely cyclic, J(X, t) = ∇ × A, (S13) where A is an arbitrary vector field. Such fluxes are divergence free, and will always form complete loops. In this case the driving force is, F(X) = D ∇P (X) P (X) + J(X) P (X) , (S14) Where J(X) is the flux at steady state. Now the driving force is made of two distinct terms. The first term corresponds to diffusion, while the second corresponds to a deterministic cyclic flux. It is in this way that both stochastic and deterministic elements enter into our models. 29 Brownian motion model. The Brownian motion model is defined by a right stochastic matrix having the property that each row in the matrix sums to one. This kind of matrix is also known as a Markov matrix and can be interpreted as a transition probability matrix from each neuronal state, i, to all other neuronal states, j. We limit the diffusion of the system to adjacent pairs of states (bins in a plane spanned by PC1 and PC2), and so the probability of transitioning from state i to state j is proportional to the driving force at that state. This gives us a formula for each off diagonal element of the transition matrix, T(i, j) ∼ pi − pj pi , (S15) where T(i, j) are the elements of the transition matrix and pi is the probability of a given state in the observed probability distribution. The equation for the transition probability is the discrete equivalent of the driving force (Eq. S12). Finally, the matrix is scaled by some value so that all rows have a sum less than one and the diagonals are filled to bring the row sum to one. This scaling step represents the level of noise in the system and has no influence on the long term statistics of the model or our results. Eq. S12 ensures that the Markov matrix for the Brownian motion model satisfies the detailed balance condition, πiP (Xt+1 = jXt = i) = πjP (Xt+1 = iXt = j), where P is the probability of transitioning from one state to another. This means that the Markov chain produced by this matrix will be reversible, P (Xt = iXt−1 = j) = P (Xt+1 = iXt = j), which is equivalent to the purely diffusive (J = 0) case of solutions to Eq. S11. Empirical Markov Model. In the EMM the transition probability matrix is defined by the observed transitions in the data, (S16) Where (cid:107)si → sj(cid:107) is the number of times the system transitions from state i to state j and (cid:107)si(cid:107) is the total number of times the system is found in state i. The resulting Markov chain (cid:107)si(cid:107) Tij = (cid:107)si → sj(cid:107) , 30 can be irreversible. This is because the data are not required to have detailed balance, and so the Markov matrix is able to have complex eigenvalues. In a Markov matrix the eigenvalues can be thought of as the modes along which the system decays to its steady state distribution. This is because the a Markov matrix is the time evolution operator of the system given by the differential equation, ∆Xt = XtL, where L is the Markov matrix. The solution to this differential equation is, Xt = XoeLt, (S17) (S18) where Xo are the initial condition of the system. Alternatively this equation can be rewritten in terms of the eigenmodes of L, (cid:88) i Xt = cieλitφi, (S19) where λi are the eigenvalues, φi are the eigenvectors and ci are the initial conditions of the system projected onto the i-th eigenvector. L, under a broad range of conditions, has a single eigenvalue with λ = 1. This corresponds to an assertion that such systems always come to a single steady state. The associated eigenvector corresponds to the steady state distribution of the system. When eigenmodes are complex Eq. S19 becomes an equation of a decaying wave in the plane spanned by a pair of complex conjugate eigenvectors. As discussed in the main text, these decaying spirals correspond to the cyclic flux of Eq. S14. In the long time limit all eigenmodes with eigenvalues much less than 1 damp out, but complex modes with eigenvalues near 1 heavily shape the dynamics of the system even in the long time limit. These eigenmodes are used to identify the cyclic fluxes of neuronal activity. S3. Model performance measures Dwell Time Statistics. 31 Dwell time statistics are calculated by setting a threshold on the simulated AVA neuron. The threshold is chosen by inspection, and any time point above the threshold is assumed to belong to backing locomotion, while any time point below the threshold belongs to forward locomotion. Backing bout times are periods in which the animal stops backing up and immediately backs up again. These events are defined as periods in which the forward locomotion state fails to last for more than 30 frames ( 10 seconds). Dwell time distributions are shown in Figure 1-3, 5. Statistics for single worm models using EMM are shown in Figure S8. Manifold-Free behavioral prediction. We selected a neuronal event that is likely to be highly correlated with behavior, and thus informative of the relevant neuronal dynamics (28–31). behaviorally, this event is the onset of a backing bout. An average activation pattern was constructed by aligning the time series of neuronal activity to each initiation of a backing bout and averaging across them (akin to spike triggered average). In the simplest case, this average activation pattern is a single snapshot of neuronal activity immediately preceding the initiation of the backing bout. Delay embedding (see below and main text) involves taking several snapshots of neuronal activity separated by a fixed time interval. Average activation patterns of multiple lengths (number of snapshots) were evaluated for their ability to predict (see below) the initiation of the bout. The snapshots were separated by τ ≈ 10 ( 3 seconds). This value was chosen based on the autocorrelation of neuronal activity. To evaluate the ability of the average activation pattern to predict the onset of backing bout, the average activation pattern was convolved with neuronal activity and smoothed. The out- put of this convolution (score) was then used to calculate the prediction probability shown in Figure 5F-I (see below). In order to predict the onset of the behavioral event an appropriate threshold value of the score has to be chosen. From an information-theoretic perspective, an optimal threshold is a 32 threshold that gives the highest separation between the distribution of scores associated with the initiation of backing bout and those that are not. To construct these distributions only local maxima of the score were considered. This min- imizes the spurious effects of noise and compensates for the fact that the behavioral events of interest are very rare in the data. To construct the distribution of scores associated with the be- havioral event, we found peaks closest to the behavioral event within tpeak − tevent ≤ 10 ≈ 3s. The amplitude of these peaks is used to construct the distribution of true events θ = 1. All other peaks are considered to be false events, θ = 0. The probability of correctly identifying a behavioral event given a specific Xthres is p(θ = 1 X ≥ Xthres) = (cid:107)Xθ=1 ≥ Xthres(cid:107) (cid:107)Xθ=1(cid:107) + (cid:107)Xθ=0 ≥ Xthres(cid:107), (S20) where Xθ=1 are the scores of true events, Xθ=0 are scores of false events and Xthres is the score cutoff threshold, and (cid:107)·(cid:107) denotes the number of elements in the set. Optimal threshold is found as argmax of Eq. S20 with respect toXthres. The average activation pattern is constructed using 50% of the events in the data. This training set is used to find Xthres and the remaining 50% are used to determine the accuracy of the prediction based on the scores exceeding Xthres. To obtained standard deviation of the prediction probability, the training set is bootstrapped. Time to transition analysis. Phase along the cyclic flux in the manifold is binned. Points within each bin are identified. Each point is characterized by two independent values: time since the onset of the behavior and bin ID. The null hypothesis here is the expected time to behavioral transition based solely the dwell time distribution. This corresponds to finding the survival function given by the right tail of the integral of the dwell time distribution (to infinity). Pnull(t) = P (t + ti) N . (S21) N(cid:88) i 33 Where ti is the time since onset of the behavior, and P (t) is the probability of the transition occurring at time t. For the manifold-based prediction the distribution of times until behavioral switch is explicitly found in the data. This analysis applied to backward locomotion is shown in Figure 3 and for forward locomotion in Figure S19 and S20. S4. Manifold Finding Delay embedding and true phase space. In the case of C. elegans benefits of delay embedding are not infinite – i.e. more data does not always yield better predictions (Figure S14). This conclusion is independent of the manifold construction method. This observation suggests the existence of a true phase space, where trajectories are maximally coherent, and minimal trajectory overlap occurs. While several algorithms for finding a good delay embedding parameters exist (47), they have significant limitations for highly noisy and short (3000 to 4000 frames per animal). To improve our chances of finding a reasonable embedding, rather than embedding just the raw neuronal activity, we embedded the adjoint space formed by the raw neuronal activity and its derivative (akin to position and velocity). While there are some differences depending on the specifics of the parameters of the delay embedding, the results are fairly robust to changes in the parameters. For the figures in the main manuscript we used: ∆τ = 10 (Number of frames to delay in each delay embedding) Number of τ = 5 (Total number of delay embeddings) For the toy model in Figure 4 we used: ∆τ = 1 (Number of frames to delay in each delay embedding) Number of τ = 1 (Total number of delay embeddings) 34 For the toy model in Figure 5A we used: ∆τ = 149 (Number of frames to delay in each delay embedding) Number of τ = 20 (Total number of delay embeddings) Diffusion mapping. Diffusion maps refer to a class of methods in which distances between points (in state space) are cast as transition probabilities (42, 43) – nearby points have high transition probability. The transition probability between nearby points can be approximated by a Gaussian diffusion kernel, kε(X, Y) = exp −(cid:107)X − Y(cid:107)2 2 4ε (cid:32) (cid:33) , (S22) where ε is a parameter of the algorithm, and (cid:107)·(cid:107)2 is the standard euclidean norm. It turns out that, diffusion maps converge to the normalized graph Laplacian, the Fokker- Planck (42) operator or the Laplace-Beltrami operator depending on the normalization of kε, kα ε (X, Y) = kε(X, Y) qα ε (X)qα ε (Y) , (S23) where qε(X) is the probability of point X at steady state, and kε is the pre-normalized kernel. The Fokker-Planck operator corresponds to the case where α = 1 2. This normalization is used for the rest of this discussion. While diffusion mapping is an extremely powerful tool it is not directly applicable to neu- ronal activity because the temporal component of the time series data is not utilized in the analysis (44). As a consequence, the diffusion kernel and resulting transition matrix is symmet- ric. Therefore, we will seek to use the temporal component of the data to form an asymmetric transition probability matrix. We start by redefining the diffusion kernel in order to make use of the time series data. We accomplish this simply by centering the kernel on the next empirically observed data point as 35 follows kF P (Xt, Y) = exp (cid:32) −(cid:107)Xt+1 − Y(cid:107)2 2 2σ2 (cid:33) . (S24) Where Xt is the position of the system in state space at time t, Y is a nearby position in state space, and σ is a normalization term (see below). There are two other differences in our kernel compared to the original version (43). First, the normalization term, σ, is defined by using the amount of local noise at the two points Y and Xt+1. The local noise, σl(x) is approximated by taking the standard deviation of the velocity of the trace in some small time window around the point of interest. This means that σ2 is now the product of two standard deviations, which is a natural symmetric scaling factor for the diffusion-based kernel. The method is robust to the exact value of time window used, and smoothing techniques can also be employed as the noise level is assumed to be a function of position, and the time series data moves smoothly through space. Second, in order to remove the dependence of kε on the ε parameter we include in σ the mean value from all of the local neighbors found throughout the full data set. Again, this normalization is quite robust and only serves to automatically pick a ε that is known to be in a reasonable range. Putting these together yields the full kernel, kF P (Xt, Yt) = exp (cid:32) − (cid:107)Xt+1 − Yt(cid:107)2 2σl(Xt+1)σl(Yt)(cid:104)kF P(cid:105)XY 2 (cid:33) , (S25) where (cid:104)kF P(cid:105)XY is the mean value of kF P over all data points and associated nearest neighbors. The time series data is assumed to have coherent trajectories, however, in the limit of finite data with progressively finer time steps the above kernel will fail to appropriately model dif- fusion between coherent trajectories. This is because there will be many points from the same trajectory (due to the small time steps) that are very close to Xt+1 while points from separate yet coherent trajectories may be relatively far away. In order to overcome this problem, the notion of nearest neighbors is adjusted to instead be nearest trajectories. The kernel for Xt 36 to all other points is calculated and then the best point is added to the nearest trajectories list (this will always be point Xt+1 since it has a kernel value of 1). Any points that are within some minimum number of time steps of Xt+1 are removed from consideration. This assures that the next nearest point will come from a trajectory other than the original trajectory. Then the closest remaining point is added to the nearest trajectories list and the process repeats for k nearest trajectories. Both this minimum time frame and the number of nearest trajectories to be considered are important parameters of the method and can heavily influence the results. The minimum time frame must be chosen to be long enough that neighboring trajectories are indeed independent trajectories and so a good estimate can be found using the autocorrelation or the medium recurrence time of the time series data. The number of nearest trajectories is a trade off between needing enough data to be able to cluster the coherent trajectory bundles appropriately and not forcing the method to group trajectories that are not in fact in the same coherent bundle. However, both of these issues are resolved in the large data limit. For the reconstruction of worm manifold we used: Local trace size = 6 (Number of points to approximate local noise) Number of neighbors = 12 (k nearest trajectories to link in the transition probability ma- trix) Minimum time frame = 50 (Minimum number of frames to wait for the nearest trajecto- ries) For the toy model in Figure 4 we used: Local trace size = 4 (Number of points to approximate local noise) Number of neighbors = 10 (k nearest trajectories to link in the transition probability ma- trix) 37 Minimum time frame = 50 (Minimum number of frames to wait for the nearest trajecto- ries) Phase Identification. The key objective of the method is to identify a small set of parameters by which to char- acterize the state of the brain across individual worms. In terms of the Fokker-Planck system given by the diffusion map, the key parameter is phase. Phase is given by the eigenvector asso- ciated with the largest complex eigenvalue as all other eigenmodes do not strongly influence the dynamics in the long term. This phase can be directly computed from the transition probability matrix formed by exponentiating the matrix constructed as in Eq. S25. If there are multiple cyclic fluxes as in C. elegans CNS, then in addition to the phase one needs to also know the identify of the flux. This latter problem is addressed in clustering section (see below). Trajectory clustering. Any standard clustering algorithm will suffice, and this section will only detail one of many possible choices (51) that can be used. We did not explore the effects of the choice of clustering and suspect that, as is the case with many clustering applications, the best choice will depend on the specifics of the dataset. We use a maximum modularity algorithm (52) on the transition probability matrix defined in Eq. S25. By construction, the transition probability matrix is sparse (only transitions in local neighborhoods are considered). Therefore, in its raw form the system given by this matrix will not explore the manifold sufficiently as it will be trapped in each individual isolated neighborhood. To overcome this problem the matrix is exponentiated N times until a minimum fraction of elements of each row are non-zero (25% in the worm data). Conceptually this corresponds to finding the evolution of the system after N time steps. Specific choice of N does not have a strong influence on the results, so long as the resultant matrix is not too sparse. 38 Two major features are found in the transition probability matrix (Figure S15): patches and diagonals. Square patches identify locations where the system exhibits Brownian motion near a point attractor. Diagonal traces identify coherent trajectories. The square patches are already suitable for community detection. If two elements of the ma- trix belong to the same point attractor, they will be found in the same square patch. The situation is slightly more complex for coherent trajectories identified by diagonal bands of high transition probability. To determine whether two elements of state space belong to the same coherent tra- jectory, we compute the maximum correlation of each row (distances from each element of state space) and time lagged copies all the other rows max(corr(rowi, shif t(rowj, t))). Where rowi is the ith row, shift moves all elements in the row t steps to the right and the maximum is taken over all t. This newly formed matrix has the same dimensions as the original transition probabil- ity matrix. We apply standard maximum modularity clustering using the community louvain function from the Brain Connectivity Toolbox to this matrix (51). Manifold simulation. Our ability to construct EMM (Figure 2) in the PCA space hinges on the relative simplicity of the C. elegans CNS allowing for significant dimensionality reduction. In high dimensional spaces, direct binning is not possible as each bin would be too sparsely occupied to give mean- ingful information concerning the probability distribution or the transition probabilities between neighboring bins. Yet, as we show in Figure 2 the PCA space is not universal across animals and is thus limiting in terms of our ability to simulate brain dynamics. In contrast, the phase and cluster identity of each point defines a natural basis on which to build a low dimensional representation of the time series data that is densely populated. In fact, our simulations of brain dynamics are essentially EMM model in the proper space of flux ID and phase. The density of points in this space can be seen in the plot of log probability density function over the parame- ters of phase and community identity (Figure S16 top). The space also retains information about 39 the behavioral state of the animal, as shown by the alignment of behavior to each community identity (Figure S16 bottom). The simulated AVA traces in Fig 3C are created by simulating the system with the same procedure as the EMM using this parameterized space. A map from parameterized state space, (ID, φ) ∈ Θ, to the true neuronal phase space, X, is defined by averaging the neuronal space position of all points found in each parameter space bin, Θ → X (ID, φ)i (cid:55)→ (cid:104)Xt ∈ (ID, φ)i(cid:105) . (S26) Manifold reconstruction. In order to reconstruct the shape of the manifold in neuronal space (Figure 3) we need some way in which to determine the mean trajectory followed by each trajectory bundle. We can find this mean by using a combination of the two parameters discussed in the previous sections. The phase provides a continuous parameter that runs along the flow of the trajectory bundle, and the cluster identity provides an insurance that the trajectory bundles with overlapping phases are treated separately. In order to ensure that our phase parameter continuously covers all space, we construct a Gaussian smoothed phase density histogram in order to find the bounds of that trajectory by including only those phases that have density larger than some cutoff. The method is robust to the exact value used. These averaged trajectory start and end points are then used to segment the time series data and align it by the phase over this particular trajectory. The mean trajectory is then found by averaging the neuronal position of points using weights from a sliding Gaussian window over the phase values. The method is robust to the exact sigma used. The resulting mean trajectories form disjoint bundles in the phase space. In order to con- struct the smooth manifolds as shown in Figure 3 the bundles are joined together by interpolat- 40 ing a spline (over both position and direction) from the end of one bundle to the beginning of the next bundle. The qualitative accuracy of the interpolation is checked vs. the raw time series data. Nonetheless, these extensions are not necessary for quantitative analyses – which are all done in (ID, φ) ∈ Θ Average behavioral trajectories. To compare the differences in patterns of neuronal activation within a single worm and across multiple worms we synchronize activity time to behavioral outputs and average over this synchronized time. For each point in a particular instance of a behavior we set its synchronized time to be tsync(ti) = (ti − tstart)/(tend − tstart). Where ti is the time of the point, tstart is the time of the start of the instance of behavior and tend is the time of the end of the instance of behavior. This time warping insures that tsync ranges from 0 to 1 over all instances of behavior, and thus can be used to average trajectories from multiple instances of the same behavior. p-values were calculated using only the first PC of the traces. Time series data of each behavioral trace was projected onto the first PC calculated from the concatenated matrix of all traces for each neuron/behavior pair. Each individual trace was mean subtracted before performing PCA. p-values were then computed using 1-way ANOVA. S5. Toy Systems Delay Embedding. The system used to highlight delay embedding is given by the two-parameter (O and H) system, O(cid:48) = H · O + , H(cid:48) = a(b − O2) + , (S27) (S28) where a, b and noise  are selected to keep the system near trajectory bundles. We use a = 41 0.01, b = 1.2 and noise is Gaussian with mean zero and σ = 0.1. These parameters were chosen such that the system takes roughly 5000 timesteps to recur. The two equations above completely define the phase space of the system, along with future and past trajectories (up to noise) (Figure 5). Model Stochastic Differential Equation. The system used in Figure 4 was created by constructing an energy landscape and flux term and then simulating using the stochastic differential equation, ∆X = λU∇U (X) + λJ J(X, t) + , (S29) where λU and λJ are strength coefficients for the diffusion and flux terms respectively, and  is noise. The energy landscape is defined by single ring formed by tracing a Gaussian around the circle and taking the maximum value at each increment. The flux follows the equation, F(X) ∼ (cid:98)θ 1 + αsin(θ) (r − µr)β , (S30) where (cid:98)θ indicates flux in the polar direction, and µr is the radial coordinate with the highest flux. µris set to be equal to the radius of the trough formed by the energy landscape, and the coefficients λU, λJ, α and β are set such that the system makes 1 complete cycle every 1000 frames. 42 Figure S6: PCA projections of pan-neuronal imaging data Projection of neuronal activity onto the first 2 PCs of each individual worm. All neurons recorded in each worm are included in the analysis. Data are colored by behavioral states as in the main text. behaviors localize into mostly coherent patterns. Note, however, that the shape of the data differs significantly between worms. Furthermore, the PCs themselves are different between worms. Plot of the explained variance for the first 10 PCs for each worm (bottom right). Note that the first 2 PCs explain approximately 60% of the variance in each worm. 43 Figure S7: Comparison of Brownian motion and EMM Element-wise quotient of the tran- sition matrices (EMM / Brownian motion) is computed. If the value at position (i, j) of the resultant matrix exceeds threshold, an arrow is drawn from point i to point j. The arrow is col- ored from blue (EMM ≈ Brownian motion) to red (EMM >> Brownian motion). Large values of the quotient are mostly concentrated along the transitions between the two energy minima. This flux is, on average, confined to a clockwise cycle. 44 Figure S8: Dwell time statistics predicted by EMM constructed for each worm individually Dwell time statistics of forward locomotion (blue), backwards locomotion (green) and backing bouts (black) observed experimentally (solid line) and predicted by EMM (dashed lines) in each worm. EMM was constructed as in the main manuscript on the basis of the first two PCs of neuronal activity. Either the full neural set (dashed) or the common 15 neuron subset (dot dash) were used. In each case, PCs were calculated individually for each worm. The main manuscript shows data from Worm B (all neurons). This worm is most closely predicted by EMM and makes the best possible case for this approach. Forward locomotion is well approximated in all worms. Backward locomotion is generally less well approximated by EMM and appears more stochastic than is experimentally observed. Backing bouts are well approximated in all worms with the exception of worm C. Which is likely due to the high level of tangling of trajectories found in its projection onto PC1 and PC2 (Figure S6). 45 46 Figure S9: Projections of 15 shared neurons on the first two PCs computed individually for each worm Projection of the shared set of 15 neurons from each individual worm onto the first 2 PCs calculated from its data. Contrast this with Figure S6 which shows the same analysis for all neurons observed in each worm. Data are colored by the same behavioral states as assigned in the main text. Note that as in Figure S6 the PCs are computed individually for each worm and thus will not generally be the same between worms. Plot of the explained variance for the first 10 PCs (bottom right). Note that the first 2 PCs explain approximately 85% of the variance for each worm. 47 Figure S10: 15 shared neuronal activity of all worms projected onto same PCA basis Ac- tivity of each of the 15 neurons identified in all worms was transformed to z-scores. Thus normalized data from each worm was concatenated and subjected to PCA. Projection onto first 3PCs is shown. Data are colored by the same behavioral states as assigned in the main text. When data from all worms is plotted on the same axes, trajectories form a tangle, and behav- ioral states are not clearly separated. This tangling of the trajectories explains why EMM is not directly generalizable across worms even if the number of dimensions is increased from 2 to 3. 48 Figure S11: Derivative of 15 shared neuronal activity of all worms projected onto same PCA basis It was noted by Kato et al (21) that derivatives of neuronal activity separated behav- ioral states better than raw activity. While this is true, projecting the appropriately normalized derivatives onto a common set of PCs still gives rise to significant trajectory tangling. Note that consistent with Kato et al (21) the trajectories are separated better than in Figure S10. This is because projecting the derivative of neuronal activity in discrete time is closely related to the idea of delay embedding as the calculation of the first derivative requires information about the present and future state of the system. 49 Figure S12: Quotient between Brownian motion and EMM using space spanned by 15 shared neuronal activity Quotient between flux predicted by Brownian motion and experi- mentally estimated as in EMM for the 15 neurons in all worms projected onto common set of PCs. Unlike Figure S7 no clear cyclic flux is observed. This is because trajectories in Fig- ure S10 are tangled. Note that the order of magnitude scale difference for the quotient. This is consistent with the observation that EMM across worms does not perform any better than the stochastic model Figure 2G of the main manuscript 50 Figure S13: Quotient when using space spanned by derivative of 15 shared neuronal ac- tivity Quotient between flux predicted by Brownian motion and experimentally estimated as in EMM for the 15 neurons in all worms projected onto common set of PCs. In this case normal- ized derivatives of neuronal activity were used as the basis of the PCs. Looking at the derivatives does not restore the cyclic flux clearly present in Figure 2D 51 Figure S14: Benefits of delay embedding are not infinite Plot of predictive power (using manifold-free behavioral prediction) as in Figure 5. Varying the amount of delay embedding increases the predictive power to a point ( 40 frames). After this point, the added dimensionality begins to have adverse effects and the predictive power begins to fall. 52 Figure S15: Asymmetric diffusion matrix modified for clustering of trajectories Asymmet- ric diffusion matrix constructed using diffusion kernel as in Eq. S25. This matrix has been modified for clustering by convolving each row of the matrix with all other rows and taking the maximum resulting value. In this way, the diagonal bands that make up trajectories tend to cluster together more tightly. 53 Figure S16: Manifold space density and behavioral localization The matrix of the kind shown in Figure S15 was subjected to clustering. This identified 5 clusters (communities) each of which is shown by a row in the plot. X-axis is the phase computed from the asymmetric dif- fusion matrix constructed according to Eq. S25. Top plot shows density of points. Red arrows show the flux of the system corresponding to a backwards locomotive behavior, and blue arrows show the flux of the system corresponding to a forward locomotive behavior. Bottom plot shows behavioral states (colored as in the main text). 54 Figure S17: Manifold simulation of individual worms Simulations using the manifold found on the basis of all observed neurons in each individual worm. Most worms recapitulate the observed statistics qualitatively, and well approximate the statistics quantitatively. Note that in contrast to EMM (Figure S8), worm C is well approximated by the manifold. Deviations from the observed dwell times are likely due in part to under sampling (Number of observed behaviors is about 20 for all worms but D, which has about 10 behaviors). 55 Figure S18: Manifold simulation of left out worms Simulations using the manifold found on the basis of 15 shared neurons by leaving each one of the 5 worms in turn. Simulations of neuronal activity given by this manifold are used to predict the behavioral statistics from the left out worm (the data from which was not used in manifold construction). All simulations show structure in their dwell time statistics that qualitatively mimic the statistics found in the left out worm. Worm D is a bit of an outlier in terms of dwell time statistics, which is why the other 4 worms have trouble predicting it. Deviations from the observed dwell times are likely due in part to under sampling (Number of observed behaviors is about 20 for all worms but D, which has about 10 behaviors). 56 Figure S19: Manifold dynamics of forward locomotion Similar analysis to Figure 3F-I of the main text. This subsection of the manifold is taken from the forward locomotion loop and predicts the time at which the worm transitions from forward locomotion to a turn. Figure S20: Entropy of forward locomotion Similar to the main text the entropy of the man- ifold based prediction is always lower than the null hypothesis. However, the entropy takes longer to drop off, and does so in a much less linear fashion, due to the high degree of stochas- ticity found in the forward locomotive trajectory. 57 Figure S21: Reconstruction of pan-neuronal dynamics Original data from Figure 1B after z-scoring (top). Reconstructed of the z-scored data based on the projection onto the first two PCs (middle). Note that the main features of the original data are preserved, but the noisier elements are lost. Manifold simulation of the same worm as the above plots using all recorded neurons (bottom). Note that more details of the structure is preserved in this simulation that was preserved in the reconstruction based on the first 2 PCs. 58 Figure S22: Representative neurons which are different across individuals Average traces of specific neurons during specific behaviors in all 5 worms demonstrate individual variance in the clonal population. Each color trace is a different worm. Dotted lines show 95% confidence intervals. p-values of traces coming from the same distribution (1-way ANOVA): AVBL → 0.0056, RIVR → 0.0018, AVAR → 0.0027, RMED → 3.2 × 10−22, RMEL → 1.6 × 10−8, SMDVR → 7.4 × 10−10 59 Figure S23: Representative neurons which are similar across individuals Average traces of specific neurons during specific behaviors in all 5 worms show that some neurons do have invariant stereotyped dynamics. Each color trace is a different worm. Dotted lines show 95% confidence intervals. p-values of traces coming from the same distribution (1-way ANOVA): ALA → 0.63, AVBL → 0.77, AVER → 0.38, VB02 → 0.42, AVAL (Forward locomotion) → 0.48, AVAL (Dorsal turn) → 0.70 60 Figure S24: p-values for all neuron/behavior combinations Each column is a specific neuron over the 6 identified behaviors. p-values are calculated as in Figures S22 and S23. Note that some neurons tend to be highly conserved across all behaviors (ALA, RID), while most are only highly similar in a few behaviors. RMED varies consistently in all behaviors. 61
1603.01880
3
1603
2017-09-22T17:22:28
Optimal sequence memory in driven random networks
[ "q-bio.NC", "nlin.CD" ]
Autonomous randomly coupled neural networks display a transition to chaos at a critical coupling strength. We here investigate the effect of a time-varying input on the onset of chaos and the resulting consequences for information processing. Dynamic mean-field theory yields the statistics of the activity, the maximum Lyapunov exponent, and the memory capacity of the network. We find an exact condition that determines the transition from stable to chaotic dynamics and the sequential memory capacity in closed form. The input suppresses chaos by a dynamic mechanism, shifting the transition to significantly larger coupling strengths than predicted by local stability analysis. Beyond linear stability, a regime of coexistent locally expansive, but non-chaotic dynamics emerges that optimizes the capacity of the network to store sequential input.
q-bio.NC
q-bio
Optimal sequence memory in driven random networks Jannis Schuecker∗,1 Sven Goedeke∗,1 and Moritz Helias1, 2 1Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany 2Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany (Dated: September 25, 2017) Autonomous randomly coupled neural networks display a transition to chaos at a critical coupling strength. We here investigate the effect of a time-varying input on the onset of chaos and the resulting consequences for information processing. Dynamic mean-field theory yields the statistics of the activity, the maximum Lyapunov exponent, and the memory capacity of the network. We find an exact condition that determines the transition from stable to chaotic dynamics and the sequential memory capacity in closed form. The input suppresses chaos by a dynamic mechanism, shifting the transition to significantly larger coupling strengths than predicted by local stability analysis. Beyond linear stability, a regime of coexistent locally expansive, but non-chaotic dynamics emerges that optimizes the capacity of the network to store sequential input. ∗ These authors contributed equally PACS numbers: 87.19.lj, 87.85.Ng, 05.45.-a, 05.40.-a Large random networks of neuron-like units can ex- hibit collective chaotic dynamics [1–4]. Their informa- tion processing capabilities have been a focus in neuro- science [5] and in machine learning [6] and show opti- mal performance close to the transition to chaos [7–9]. Due to its rich chaotic dynamics, the seminal network [1] until today serves as a model by Sompolinsky et al. model for various activity patterns observed in working memory tasks [10–13], motor control [14], and percep- tual decision making [15]. The interplay between a time- dependent input signal and the dynamical state of the network, however, is poorly understood; notwithstand- ing consequences for information processing. In the absence of a signal the network dynamics is au- tonomous. Networks of randomly coupled rate neurons display a transition from a fixed point to chaotic fluc- tuations at a critical coupling strength [1], illustrated in Figure 1a. The transition is well understood by dynamic mean-field theory, originally developed for spin glasses [16, 17]. The onset of chaos is equivalent to the emer- gence of a non-zero, decaying autocorrelation function, whose decay time diverges at the transition. This equiv- alence has been used in several subsequent studies [18– 20]. Furthermore a tight relationship to random matrix theory exists: the transition happens precisely when the fixed point becomes linearly unstable, which identifies the spectral radius of the random connectivity matrix [21, 22] as the parameter controlling the transition. These relations, however, lose their validity in the pres- ence of fluctuating input: stochasticity per se decorre- lates the network activity even if the dynamics is sta- ble (Figure 1b), so that a decaying autocorrelation func- tion does not necessarily indicate chaos. The stochastic drive, furthermore, causes perpetual fluctuations also in the regular regime. Therefore, a transition to chaos, if existent at all, must be of qualitatively different kind than the transition from the silent fixed point in the au- tonomous case. Time-dependent driving has indeed been found to stabilize network dynamics [18, 23]. However, the mechanism is only understood for low-dimensional systems in the context of chaos synchronization by noise [24], in networks driven by deterministic signals [18], and in systems with time-discrete dynamics [23]. In the latter model, the effect of the fluctuating input on the transi- tion to chaos is completely captured by its influence on the spectral radius of the Jacobian. Its single neuron dy- namics, moreover, does not possess non-trivial temporal correlations. But these temporal correlations are indeed essential for the transition to chaos and for information processing in time-continuous systems, as we will show here. Realistic continuous-time network models can gener- ate complex but controlled responses to input [8] that resemble activity patterns observed in motor cortex. In particular, the dynamical state of the network plays a crucial role during the involved learning process. How- ever, the effect of the input on the dynamical state has remained obscure. To investigate the generic influence of external input on the network state, we include additive white noise in the seminal model by Sompolinsky et al. [1] and develop the dynamic mean-field theory for the resulting stochas- tic continuous-time dynamics. In contrast to the original work, we here reformulate the problem in terms of the functional formalism for stochastic differential equations [25–29]. The application of the auxiliary field formula- tion known from large N field theory [30] then allows us to derive the mean-field equations by a saddle point ap- proximation. We find that the autocorrelation function is formally identical to the motion of a classical parti- cle in a potential, where the noise amounts to an ini- tial kinetic energy. We then determine the maximum Lyapunov exponent [31] by considering two copies of the system with different initial conditions [32] in a replica calculation. Our main result is a closed-form condition for the transition from stable to chaotic dynamics. We find that the input suppresses chaos significantly more strongly than expected from time-local linear stability, the criterion valid in time-discrete systems. This obser- vation is explained by a dynamic effect: the decrease of the maximum Lyapunov exponent is related to the sharp- ening of the autocorrelation function by the fluctuating drive. The regime in the phase diagram between local instability, as indicated by the spectral radius of the Ja- cobian, and transition to chaos, corresponding to a pos- itive maximum Lypunov exponent, constitutes an as yet unreported dynamical regime that combines locally ex- pansive dynamics with asymptotic stability. Moreover, in contrast to the autonomous case, the decay time of the autocorrelation function does not diverge at the transi- tion. Its peak is strongly reduced by the input and occurs slightly above the critical coupling strength. To study information processing capabilities we evalu- ate the capacity to reconstruct a past input signal by a linear readout of the present state, the so-called memory curve [33]. Dynamic mean-field theory and a replica cal- culation lead to a closed form expression for the memory curve. We find that the memory capacity peaks within the expansive, non-chaotic regime, indicating that locally expansive while asymptotically stable dynamics is ben- eficial to store input sequences in the dynamics of the neural network. I. DYNAMIC MEAN-FIELD EQUATION We study the continuous-time dynamics of a random network of N neurons, whose states xi(t) ∈ R, i = 1, . . . , N, evolve according to the system of stochastic dif- ferential equations dxi dt = −xi + N Xj=1 Jij φ(xj ) + ξi(t) . (1) The Jij are independent and identically Gaussian dis- tributed random coupling weights with zero mean and variance g2/N , where the intensive gain parameter g controls the recurrent coupling strength or, equivalently, the weight heterogeneity of the network. We further ex- clude self-coupling, setting Jii = 0. The time-varying in- puts ξi(t) are pairwise independent Gaussian white-noise processes with autocorrelation function hξi(t)ξj (s)i = 2σ2δij δ(t − s). We choose the sigmoidal transfer func- tion φ(x) = tanh(x), so that without input, for σ = 0, the model agrees with the autonomous one studied in [1]. The dynamical system (1) contains two sources of ran- domness: the quenched disorder due to the random cou- pling weights and temporally fluctuating drive. A par- 2 ticular realization of the random couplings Jij defines a fixed network configuration and its dynamical prop- erties usually vary between different realizations. For large network size N , however, certain quantities are self- averaging, meaning that their values for a typical realiza- tion can be obtained by an average over network config- urations [34]. An important example is the population- averaged autocorrelation function. We here derive a dynamic mean-field theory that de- scribes the statistical properties of the system under the joint distribution of disorder, noise, and possibly ran- dom initial conditions in the limit of large network size N → ∞. The theory can be derived via a heuristic "local chaos" assumption [35] or using a generating functional formulation [17, 36]. We here follow the latter approach, because it casts the problem into the established language of statistical field theory for which a wealth of approx- imation techniques is available [37]. A mathematically rigorous proof uses large deviation techniques [38]. The general idea is that for large network size N the local j=1 Jij φ(xj ) in (1) approaches a Gaus- sian process with self-consistently determined statistics. We interpret the stochastic differential equations in the Ito-convention [39] and formulate the problem (1) in terms of a moment-generating functional Z. Using the Martin-Siggia-Rose-De Dominicis-Janssen path integral formalism [25, 28, 40] we obtain recurrent inputPN Z[l](J) =Z DxZ Dx exp(cid:16)S0[x, x] − xTJφ (x) + lTx(cid:17) (2) with S0[x, x] = xT (∂t + 1) x + σ2 xT x, (3) k=1ΠM l=1R ∞ k=1ΠM (2πi)−1dxl where xTy =PiR xi(t)yi(t) dt denotes the scalar prod- sures are defined asR Dx = limM→∞ ΠN and R Dx = limM→∞ ΠN l=1R i∞ uct in time and in neuron space and x and l represent a response field and a source field, respectively. The mea- dxl k −∞ k with the subscript k denoting the k-th unit and the super- script l denoting the l-th time slice. The action S0 in (3) contains all single unit properties, therefore excluding the coupling term −xTJφ (x), which is written explicitly in (2). Assuming that the dynamics is self-averaging, we aver- age over the quenched disorder in the connectivity J and perform a saddle-point approximation (Appendix A). The resulting functional factorizes into N terms −i∞ g2 2 ¯Z∗ ∝Z DxZ Dx exp(cid:16)S0[x, x] + xTCφ(x)φ(x) x(cid:17), with Cφ(x)φ(x)(t, s) := hφ(x(t))φ(x(s))i denoting the av- erage autocorrelation function of the non-linearly trans- formed activity of the units (A9) and xTCφ(x)φ(x) x := RR dt ds x(t) Cφ(x)φ(x)(t, s) x(s). The factorization re- duces the network to N non-interacting units, each on (4) a background of an independent Gaussian noise with identical self-consistently determined statistics. At this level of approximation, the problem is hence equivalent to a single unit system. The effective equation of mo- tion corresponding to this system can be read from (4) (Appendix A) (a) x2 x1 x2 x1 g = 0.5 (b) 3 g = 0.5 g = 1.7 g = 1.7 dx dt = −x + η(t) + ξ(t). (5) (c) 0 150 t 190 150 t 190 Here, ξ(t) is a Gaussian white-noise process as in (1), independent of η(t). The centralized Gaussian process η(t) is fully specified by its autocorrelation function 4 0 1 · ) c ( V hη(t)η(s)i = g2Cφ(x)φ(x)(t, s). (6) −20 0.0 (e) 0.5 c (d) 0 3 0 1 · ) c ( V 0.5 c −15 1.0 (f) 0.0 g = 1.2 g = 1.48 c0 0.8 ) τ ( c Ekin c0 1.0 g = 0.0 g = 1.2 g = 1.48 g = 1.7 II. EFFECTIVE EQUATION OF MOTION OF THE AUTOCORRELATION Our goal is to determine the mean-field autocorrela- tion function hx(t)x(s)i, which, by self-averaging, also de- scribes the population-averaged autocorrelation function. Assuming that x(t) is a stationary process, c(τ ) = hx(t + τ )x(t)i obeys the differential equation (Appendix B) c = d2c dτ 2 = c − g2fφ(c, c0) − 2σ2δ(τ ) (7) with c0 = c(0). The Dirac-δ inhomogeneity originates from the white-noise autocorrelation function of the time- varying input and is absent in [1]. The same inhomogene- ity arises from Poisson spiking noise with 2σ2 = g2r [41], where r is the population-averaged firing rate. In (7) we write fφ(c(τ ), c0) = Cφ(x)φ(x)(t + τ, t), introducing the notation fu(c, c0) =ZZ u(cid:18)qc0 − c2 c0 (8) z1 + c√c0 z2(cid:19) u (√c0 z2) Dz1Dz2 i /2)/√2π dzi, for an arbitrary function u(x) and the Gaussian inte- gration measures Dzi = exp(−z2 i = 1, 2. This representation holds since x(t) is itself a Gaussian process. Note that (8) reduces to a one-dimensional integral for fu(c0, c0) = hu(√c0z1)2i and fu(0, c0) = hu(√c0z1)i2. We formulate (7) as the one-dimensional motion of a classical particle in a potential: c = −V ′(c) − 2σ2δ(τ ) , (9) where we define 1 2 c2 + g2fΦ(c, c0) − g2fΦ(0, c0) , (10) V (c) = V (c; c0) = − with Φ(x) = R x 0 φ(y) dy and ∂/∂c fΦ(c, c0) = fφ(c, c0) following from Price's theorem [42, 43]. The autocor- relation c(τ ) here plays the role of the position of the ) τ ( c 0.5 0.0 0 20 τ 0 0 20 τ Figure 1. Activity statistics of autonomous and driven network. Autonomous case σ = 0 (left column) and driven case σ = √0.125 (right column). Upper row: Simulated tra- jectories of two example neurons for sub-critical g = 0.5 (up- per part of vertical axis) and super-critical coupling g = 1.7 (lower part of vertical axis). Middle row: Classical potential (10) with self-consistently determined variance c0 following from energy conservation (11) for different coupling strengths g (corresponding legends in lower row); dashed horizontal line at minus initial kinetic energy Ekin = σ4/2. In the driven case the critical coupling gc = 1.48 from eq. (20) is shown in red. Lower row: Self-consistent autocorrelation function (solid line) compared to simulations (crosses). The variance (peak height) c0 corresponds to the largest value of c at which the potential (middle row) is defined, indicated for g = 1.7 with gray dotted lines in (d) and (f). Network size in simulations is N = 10000. particle and the time lag τ the role of time. The poten- tial (10) depends on the initial value c0, which has to be determined self-consistently. We obtain c0 from classical energy conservation c2/2 + V (c) = constant. Considering τ ≥ 0 and the symmetry of c(τ ), the fluctuating drive in (9) amounts to an initial velocity c(0+) = −σ2 and thus to the kinetic energy c2(0+)/2 = σ4/2. Since c(τ ) ≤ c0, the solution c(τ ) and its first derivative must approach zero as τ → ∞. Thus we obtain the self-consistency con- dition for c0 as 1 2 σ4 + V (c0; c0) = V (0; c0) = 0. (11) For the autonomous case, Figure 1c,e shows the resulting potential and the corresponding self-consistent autocor- relation function c(τ ) in the chaotic regime. Approaching the transition from above, g → gc = 1, the amplitude c0 vanishes and the decay time of c(τ ) diverges [1]. This pic- ture breaks down in the driven case (Figure 1d,f), where c0 is always nonzero, c(τ ) decays with finite time scale and has a kink at zero. The mean-field prediction is in excellent agreement with the population-averaged auto- correlation function obtained from numerical simulations of one network instance showing that the self-averaging property is fulfilled. In the following we derive a condi- tion for the transition from stable to chaotic dynamics in the presence of the time-varying input. III. EFFECT OF INPUT ON THE TRANSITION TO CHAOS The maximum Lyapunov exponent quantifies how sen- sitive the dynamics depends on the initial conditions [31]. It measures the asymptotic growth rate of infinitesimal perturbations. For stochastic dynamics the stability of the solution for a fixed realization of the noise or equiv- alently the stochastic input is also characterized by the maximum Lyapunov exponent [44]: If it is negative, tra- jectories with different initial conditions converge to the same time-dependent solution; the dynamics is stable. If it is positive, the distance between two initially arbi- trary close trajectories grows exponentially in time; the dynamics exhibits sensitive dependence on initial condi- tions and is hence chaotic. We derive the maximum Lyapunov exponent by us- ing dynamic mean-field theory. To this end, we consider two copies of the network distinguished by superscripts α ∈ {1, 2}. These copies, or replicas, have identical cou- pling matrix J and, for σ > 0, are subject to the same realization of the stochastic input ξi(t). The maximum 4 Lyapunov exponent can be defined as the asymptotic growth rate of the Euclidean distance between trajec- tories of the two copies: λmax = lim t→∞ lim x1(0)−x2(0)→0 1 2t x1(0) − x2(0)2(cid:19) . ln(cid:18) x1(t) − x2(t)2 We now follow an idea by Derrida and Pomeau [32] and exploit the self-averaging property of population- averaged correlation functions, i.e, 1 i (s) ≈ cαβ(t, s), where cαβ denote the correlation functions av- eraged over the realization of the couplings. We express the mean squared Euclidean distance as N PN i=1 xα i (t)xβ 1 N N Xi=1(cid:0)x1 i (t) − x2 i (t)(cid:1)2 ≈ c11(t, t) + c22(t, t) − 2c12(t, t) ≡ d(t) , where we defined the mean-field squared distance d(t). Thus the asymptotic growth rate of d(t) provides us with a mean-field description of the maximum Lyapunov ex- ponent. To obtain this growth rate we first consider d(t, s) = c11(t, s) + c22(t, s) − c12(t, s) − c21(t, s) (12) with the obvious property d(t) = d(t, t). We then deter- mine the temporal evolution of d(t, s) for infinitesimally perturbed initial conditions x1(0)− x2(0) = ǫ. To this end it is again convenient to use a generating functional that captures the joint statistics of the two systems and in addition allows averaging over the quenched disorder [see also 37, Appendix 23, last remark]. The generating func- tional describing the two copies is defined analogously to the single system (2) as Z[{lα}α∈{1,2}](J) = Π2 α=1nZ DxαZ Dxα exp(cid:16)S0[xα, xα] − xαTJφ (xα) + lαTxα(cid:17)o exp(cid:0)2σ2x1T x2(cid:1) (13) with the single system "free action" S0[x, x] defined in (3). The factor in the last line results from the identical external input in the two copies and effectively couples the two systems. We also note that the coupling matrix J is the same in both copies. Averaging (13) over the quenched disorder of the ran- dom coupling matrix J and performing a saddle-point approximation we obtain a pair of effective dynamical equations (Appendix C), (∂t + 1) xα(t) = ξ(t) + ηα(t) , α ∈ {1, 2} , (14) together with a set of self-consistency equations for the statistics of the noises ηα hηα(t) ηβ(s)i = g2 hφ(xα(t))φ(xβ (s))i. (15) Now, there are two terms which introduce correlations between the two copies. First the common temporal fluctuations ξ(t) injected into both systems. Second the effective noises ηα and ηβ are correlated between repli- cas (15), arising from the two systems having the same coupling J in each realization. The origin of the latter coupling is hence of static nature. The distance (12) between the two copies is given by the auto-correlations of the single systems and the cross- correlations between them. We consider the case where both copies are prepared with identical initial conditions and thus are fully synchronized: the cross-correlation c12 initially equals the auto-correlations c11, c22. The latter are identical to the single-system autocorrelation func- tion c, because the marginal statistics of each subsystem (a) 0 ) τ ( W -0.2 0.15 ) τ ( 2 ψ 0 (b) 0 -0.3 0.15 0 g = 1.2 g = 1.48 g = 1.2 g = 1.48 g = 1.7 −30 0 τ 30 −30 0 τ 30 Figure 2. Ground state of Schrödinger equation deter- mines Lyapunov exponent. Upper part of vertical axis: Quantum potential W (solid curve) and ground state energy E0 (dashed line) for autonomous case (a) and driven case (b) for σ = √0.125. Lower part of vertical axis: Correspond- ing squared ground state wave function. Parameters as in Figure 1 (driven case for g = 0 left out). is not affected by the mere presence of the respective other system. An increase of the distance d(t), by (12), amounts to a decline of c12 away from its initial value c. Here c is the stationary autocorrelation as we are interested in the Lyapunov exponent averaged over ini- tial conditions drawn from the stationary distribution. To determine the growth rate in the limit of small dis- tances d(0) ∝ ǫ between the two copies we therefore ex- pand the cross-correlation around its stationary solution c12(t, s) = c(t − s) + ǫ k(1)(t, s) , ǫ ≪ 1, which leads to an equation of motion for the first order deflection (Appendix C 1) (∂t + 1) (∂s + 1) k(1)(t, s) = g2fφ′(c(t − s), c0) k(1)(t, s) (16) with d(t) = −2ǫ k(1)(t, t). A separation ansatz in the coordinates τ = t − s and T = t + s then yields an eigenvalue problem in the form of a time-independent Schrödinger equation [1, 41] (Appendix C 2) (cid:2)−∂2 τ + W (τ )(cid:3) ψ(τ ) = E ψ(τ ), (17) where now τ plays the role of a spatial coordinate. Here, the quantum potential W (τ ) = −V ′′(c(τ )) = 1− g2fφ′ (c(τ ), c0) is given by the negative second deriva- tive of the classical potential V (c) evaluated along the self-consistent autocorrelation function c(τ ). The ground state energy E0 of (17) determines the asymptotic growth rate of k(1)(t, t) as t → ∞ and, hence, the maximum Lyapunov exponent via λmax = −1 + √1 − E0 (C11). Therefore, the dynamics is predicted to become chaotic if E0 < 0. The quantum potential together with the so- lution for the ground state energy and wave function is 5 shown in Figure 2. The latter are obtained as solutions of a finite difference discretization of (17). In the autonomous case, a decaying autocorrelation function corresponds to a positive maximum Lyapunov exponent [1]. This follows from the observation that for g > 1 the derivative of the self-consistent autocorrela- tion function c(τ ) solves the Schrödinger equation with E = 0. But as c(τ ) is an eigenfunction with a single node it cannot be the ground state, which has zero nodes. The ground state energy, which is necessarily lower, must therefore be negative, E0 < 0. So the dynamics is chaotic and λmax crosses zero at g = 1 (Figure 3a). In the presence of fluctuating drive, the maximum Lyapunov exponent becomes positive at a critical cou- pling strength gc > 1; with increasing input amplitude the transition shifts to larger values (Figure 3a). The mean-field prediction λmax = −1 + √1 − E0 shows ex- cellent agreement with the maximum Lyapunov expo- nent obtained in simulations using a standard algorithm [31]. Since the ground state energy E0 must be larger than the minimum W (0) = 1 − g2h[φ′(x)]2i of the quan- tum potential, an upper bound for λmax is provided by −1 + gph[φ′(x)]2i leading to a necessary condition (18) gph[φ′(x)]2i ≥ 1 for chaotic dynamics. However, close to the transition λmax is clearly smaller than the upper bound, which is a good approximation only for small g (Figure 3a, in- set): the actual transition occurs at substantially larger coupling strengths. In contrast, for memoryless discrete- time dynamics the necessary condition found here is also sufficient for the transition to chaos [23, eq. 13]. The local linear stability of the dynamical system (1) is analyzed via the variational equation d dt yi(t) = −yi(t) + N Xj=1 Jijφ′(xj (t)) yj(t) , (19) i = 1, . . . , N, describing the temporal evolution of an in- finitesimal deviation yi(t) about a reference trajectory xi(t). Interestingly, ρ = gph[φ′(x)]2i (cf. (18)) is also the radius of the disk formed by the eigenvalues of the Ja- cobian matrix in the variational equation (19) estimated by random matrix theory [21, 22]. Therefore, the dynam- ics is expected to become locally unstable if this radius exceeds unity, as shown in the inset in Figure 3b display- ing ρ and the eigenvalues at an arbitrary point in time. But even for the case with ρ > 1 the system is not neces- sarily chaotic. Hence, contrary to the autonomous case [1, 21], the transition to chaos is not predicted by random matrix theory. To derive an exact condition for the transition we de- termine a ground state with vanishing energy E0 = 0. As in the autonomous case, c(τ ) solves (17) for E = 0, ex- cept at τ = 0 where it exhibits a jump, because c(τ ) has a kink due to the input (7). By linearity c(τ ) is a contin- uous and symmetric solution with zero nodes. Therefore, if its derivative is continuous as well, requiring c(0+) = 0, it constitutes the searched for ground state. This is in contrast to the autonomous case, where c(τ ) corresponds to the first excited state. Consequently, with (7) we find the condition for the transition g2 c fφ(c0, c0) − c0 = 0 , (20) in which c0 is determined by the self-consistency condi- tion (11) resulting in the transition curve (gc, σc) in pa- rameter space (Figure 3b). This reveals the relationship between the onset of chaos, the statistics of the random coupling matrix, and the input amplitude. From (20) follows that the system becomes chaotic pre- cisely when the variance c0 of a typical single unit equals the variance of its recurrent input from the network g2 chφ2i. At the transition the classical self-consistent po- tential V (c; c0) has a horizontal tangent at c0, while in the chaotic regime a minimum emerges (Figure 1d). This implies with (7) that the curvature c(0+) of the autocor- relation function at zero changes sign from positive to negative (Figure 1f). Close to the transition a standard perturbative approach shows that λmax is proportional to g2hφ2(x)i − c0, indicating a self-stabilizing effect: since both terms grow with g, the growth of their difference is attenuated, explaining why λmax(g) bends down as the transition is approached (Figure 3a). The condition (20) predicts the transition at signifi- cantly larger coupling strengths compared to the neces- sary condition (Figure 3b), which is explained as follows. For continuous-time dynamics the effect of fluctuating in- put is twofold: First, because the slope φ′(x) is maximal at the origin, fluctuations reduce the averaged squared slope in g2h[φ′(x)]2i, thereby stabilizing the dynamics. This is an essentially static effect as it can be fully at- tributed to the increase of the variance c0 caused by the additional input; static heterogeneous inputs would have the same effect. Second, the input sharpens the auto- correlation function (Figure 1e,f) and hence the quan- tum potential (Figure 2). This shifts the ground-state energy to larger values, further decreasing the maximum Lyapunov exponent. Because this effect depends on the temporal correlations, the input suppresses chaos by a dynamic mechanism yielding stable dynamics even in the presence of local linear instability. To understand this dynamic mechanism we return to the variational equation (19): its fundamental solution can be regarded as a product of short-time propagator matrices, where each factor has the same stability prop- erties with unstable directions given by the local Jacobian matrix at the respective time. Even though the fraction of eigenvalues with positive real part stays approximately constant, the corresponding unstable directions vary in time. The sharpening of the autocorrelation function suggests that fluctuating input causes a faster variation, (a) 0 x a m λ −1 1 0 0 E σ = 0 σ = 0.11 σ = 0.35 σ = 0.5 (b) 2 chaos 0 -0.3 1 g 2 1 (c) 0 10 ∞ τ 0 0 0 1 2 3 g σ 1 2 g 6 1 3 Figure 3. Transition to chaos. (a) Upper part of verti- cal axis: Maximum Lyapunov exponent λmax as a function of the coupling strength g for different input amplitude lev- els. Mean-field prediction (solid curve) and simulation (dia- monds). Comparison to the upper bound −1 + gph[φ′(x)]2i (dashed) for σ = 0.5 in inset. Zero crossings marked with dots. Lower part of vertical axis: Ground state energy E0 as a function of g. (b) Phase diagram with transition curve (solid red curve) obtained from (20) and necessary condition ((18) with equal sign, gray dashed curve). Dots correspond to zero crossings in inset in (a). Disk of eigenvalues of the Jacobian matrix in (19) for σ = 0.8 and g = 1.25 (lower) and g = 2.0 (upper) centered at −1 in the complex plane (gray). Radius ρ = gph[φ′(x)]2i from random matrix theory (black). Vertical line at zero. (c) Asymptotic decay time τ∞ of auto- correlation function. Vertical dashed lines mark the transition to chaos. Color code as in (a). Network size of simulations N = 5000. such that perturbations cannot grow in the direction of unstable modes, but rather decay asymptotically. In low-dimensional systems the suppression of chaos by external fluctuations is understood: Noise forces the system to visit regions of the phase space with locally contracting dynamics more frequently [24] so that con- traction dominates expansion, in total yielding stable asymptotic behavior. This mechanism is similar to the static stabilization effect described above, where fluctuat- ing drive causes the system to sample regions of the phase space with smaller eigenvalues of the Jacobian. The self- averaging high-dimensional system, however, has a con- stant spectral radius over time and hence the dynamics is either locally contracting or locally expanding for all times. While the previous effects are explained by local stability, the dynamic suppression of chaos found here is a genuinely time-dependent mechanism, explained by the time evolution of the Jacobian. In the autonomous case the time-scale of fluctuations diverges at the transition to chaos [1]. We here con- sider the effect of the input on the asymptotic decay time τ∞ = 1/p1 − g2hφ′(x)i2 of the autocorrelation function (Figure 3c). For weak input amplitude, the decay time peaks at the transition, corresponding to the diverging time scale in the autonomous case. For larger input am- plitudes, the peak is strongly reduced and the maximum decay time is attained above the transition. IV. INFORMATION PROCESSING CAPABILITIES We expect the expansive, non-chaotic regime to be beneficial for information processing: The local instabil- ity of the network ensures sufficient initial amplification of the impinging external signal. The asymptotic stabil- ity is required for the driving signal not to be corrupted by the unbounded amplification of small variations of the input; it is hence necessary to ensure generalization. In the following we investigate these ideas quantitatively by considering the sequential memory capacity of the net- work. We focus on the component z(t) = 1√N PN i=1 ξi(t) of the input that is received by all neurons with equal strength. In other words, the total input to each neuron is decomposed into the signal z(t) and the remaining in- puts ξi(t)− z(t) which act as noise. We then consider the dynamical short-term memory defined as the capacity to reconstruct the input z(t) from the state at time t + τ using a linear readout,PK i=1 wixi(t+ τ ), where K ≤ N is the number of readout neurons. The reconstruction ca- pacity as a function of the delay time τ yields the mem- ory curve m(τ ) = 1− ǫ(τ ) [33], where ǫ(τ ) is the minimal relative mean-squared error between readout and signal. Alternatively this measure quantifies the fidelity by which a sequence of past inputs can be reconstructed from the current network activity. For optimal readout weights w that minimize ǫ(τ ), the memory curve is given by [33, 45] m(τ ) = hx(t + τ )z(t)iThx(t)x(t)Ti−1hx(t + τ )z(t)i . hz(t)2i (21) We follow the approach by Toyoizumi and Abbott [9] and neglect the off-diagonal terms in hxxTi, which is justi- fiable for a sparse readout with K ≪ N . Additionally, i(cid:11) are given by their for large N the diagonal terms (cid:10)x2 mean-field value c0, identical for all units. Determining the memory curve (21) then amounts to computing the sum of squared correlation functions Pihxi(t + τ )z(t)i2 between the signal and the network activity, which we obtain by a replica calculation (Appendix D). The key idea is to express the correlation functions hxizi as a sum of response functions hxi xji; this is possible due to the Gaussian statistics of z. The calculation is similar to the derivation of the Schrödinger equation (Appendix C) with the difference, however, that the two replicas re- ceive independent realizations of the inputs. The memory 7 curve follows from a differential equation for the correla- tion between the two systems and is measured in units of the readout ratio K/N (D22): m(τ ) = 2σ2 c0 e−2τ I0 [2ghφ′(x)i τ ] Θ(τ ) dτ (22) with the modified Bessel function of the first kind I0. The memory curve has two contributions (D22): memory due to the collective network dynamics and local memory due to the leaky integration of the single units. The latter ef- fect is trivial and is reflected in the initial steep falloff ∝ e−2τ of the memory curves with time lag τ , indepen- dent of the coupling strength (Figure 4a). Its decay time is half the time constant of the neurons, which is set to unity here (1). With increasing coupling strength the variance c0 increases, so that the memory curve (22) at zero time lag τ = 0 reduces. For time lags that are large compared to the single unit time constant, the network contribution to the memory dominates. A non-vanishing memory capacity for longer time lags is therefore only the result of the reverberation of the input through the net- work interaction. The analytical results are in excellent agreement with direct simulations. We isolate the interesting network memory by sub- tracting the single unit contribution (first term in (D22)) mnet(τ ) = m(τ ) − 2σ2 c0 e−2τ Θ(τ ) dτ. (23) This quantity is particularly important in situations where the readout does not have access to the neurons receiving the signal. The network memory curve consis- tently vanishes for the uncoupled case. We compare the performance of two different couplings strengths: gnec, following from (18) with equal sign and corresponding to the onset of the local instability, and gc, marking the on- set of chaos (cf. Figure 3b). For short time lags, τ < 5, the network memory curve is larger for gnec, while for longer time lags it is larger for gc due to a slower de- cay of the memory curve. This behavior is confirmed by the memory curve as a function of g, shown for differ- ent time lags (Figure 4c, upper panel). For τ ≥ 4, the memory capacity m is entirely given by mnet, which is in line with the fast decay of the single-unit contribution. Moreover, while for small time lags the memory is max- imal around gnec, for larger time lags it peaks nearby gc, indicating that the intermediate, expansive, non-chaotic regime supports storage of the input. The memory capacity is defined as the integral over the memory curve M =Z ∞ 0 m(τ ) = σ2 c0s 1 1 − g2hφ′(x)i2 , (24) which follows directly from the Laplace transform of the Bessel function. Typically the memory capacity is bounded by the num- ber of neurons N [45]. The signal z in our situation, however, can be seen as one out of N independent in- puts and m(τ ) its corresponding memory curve. The expressions are therefore independent of N [46] and the memory capacity satisfies M ≤ 1. The network memory capacity is defined as Mnet = R mnet. While the mem- ory capacity decreases in the chaotic regime, the network memory peaks within the expansive, non-chaotic regime (Figure 4b, lower panel). So far we have considered the memory capacity at a fixed amplitude σ of the input. In the following we in- vestigate the memory capacity over the whole phase di- agram. The total memory capacity shows a steep falloff directly above the onset of chaos (Figure 5a). This is ex- pected because the information about the input is lost in the chaotic network dynamics. The contour lines of the memory capacity are nearly parallel to the transition criterion, the curve with vanishing Lyapunov exponent. This observation closely links the transition to chaos to the memory capacity: The direction in the phase dia- gram in which the system most quickly enters the chaotic regime is accompanied by the steepest decline of memory. The contour lines of the total network memory capacity show a ridge running through the expansive, non-chaotic regime (Figure 5b); it confirms the results found above: The memory is optimal in the dynamical regime of local instability and asymptotic stability. Moreover, the net- work capacity has a substantial contribution to the total memory capacity of about 50%. The optimal network memory in the hitherto unre- ported regime between gnec and gc can be understood in an intuitive manner. Two conditions must be met for good memory. First, individual units must be susceptible to the signal; the susceptibility equals the noise-averaged slope hφ′i of the gain function. Second, the signal must propagate effectively through the network, requiring a sufficiently strong coupling g. These two requirements are reflected in the monotonic increase of the memory curve (22) with the effective slope g hφ′i, independent of the time lag τ . An increase in the coupling strength, however, elevates the intrinsically generated fluctuations as well. These have a twofold effect on the memory ca- pacity. First they decrease hφ′i, so that g hφ′i assumes a maximum. Second, the intrinsic fluctuations propagate through the network as well; they hence reduce the signal to noise ratio of the readout, as they enter the denomina- tor in Mnet through c0. The interplay of these two mecha- nisms leads to optimal memory located in the expansive, non-chaotic regime, explaining why the combination of time-local expansive dynamics and stable long-term be- havior maximizes the memory capacity of the network. V. DISCUSSION We here present a completely solvable network model that allows us to investigate the effect of time-varying (a) 10−1 m 10−3 (b) 10−2 t e n m 10−4 0 0 10 τ 10 τ 8 τ = 2.0 τ = 4.0 τ = 6.0 τ = 8.0 τ = 10.0 (c) 0.01 ) . t s n o c = τ ( m 0.00 1 M 20 g = 0.0 g = 1.7 g = 2.4 20 0 0 2 g 4 Figure 4. Sequential memory. Mean-field prediction (solid curves) and simulation (crosses). (a) Memory (22) as a func- tion of time lag τ between signal and readout for different cou- pling strengths encoded in color (legend in b). (b) Network contribution (23) to memory for different coupling strengths g. (c) Upper part of vertical axis: Memory at different time lags τ over coupling strength. Network contribution to mem- ory shown as dashed thick light-gray curves, which coincide with total memory curves for τ ≥ 4. Vertical gray line marks local instability (18) and vertical red line marks transition to chaos (cf. Figure 3b). Lower part of vertical axis: Memory capacity M (24) (black) and network contribution to memory capacity Mnet (red). (a) 2 1 g (b) 1 M g 2 1 0.6 t e n M 0 0.01 σ 0 1.00 0 0.01 0.0 1.00 σ Figure 5. Memory capacity in different phases of the network dynamics. (a) Total memory capacity (24) en- coded in color. Phase boundary (20) between regular and chaotic dynamics (red) and necessary condition (18) of local instability (gray) as in Figure 3. Contour lines of memory shown in black. (b) Same as (a) for network contribution to memory capacity. input on the transition to chaos and information process- ing capabilities. Adding time-varying stochastic forcing to the seminal model by Sompolinsky et al. [1] yields a stochastic continuous-time dynamical system. Contrary to the original model [1], we here reformulate the stochas- tic differential equations as a field theory [25–28]. This formal step allows us to develop the dynamic mean-field theory by standard tools: a saddle point approximation of the auxiliary field generating functional [17, 30, 36]. As in the original model, this procedure reduces the in- teracting system to the dynamics of a single unit. The self-consistent solution of the effective equation yields a standard physics problem: the autocorrelation function of a typical unit is given by the motion of a classical particle in a potential. We find that the amplitude of the input corresponds to the initial kinetic energy of the particle. The field theoretical formulation then allows us to per- form a replica calculation to determine the maximum Lyapunov exponent; the problem formally reduces to finding the ground-state energy of a single-particle quan- tum mechanical problem. The transition to chaos ap- pears at the point where the ground state energy changes sign, which allows us to obtain a closed form condition relating the coupling strength and the input amplitude at the transition. We find a simple hallmark of the tran- sition in the single unit activity: at the transition point the variance of the recurrent input to a single unit equals the variance of its own activity. Correspondingly, the autocorrelation function at zero time lag changes its cur- vature from convex to concave at the transition point. These features can readily be measured in most physical systems. The assessment of chaos by these passive ob- servations in particular does not require a perturbation of the system. The transition criterion allows us to map out the phase diagram spanned by the coupling strength and the input amplitude. It shows that external drive shifts the tran- sition to chaos to significantly larger coupling strengths than predicted by time-local linear stability analysis. The transition in the stochastic system is thus qualitatively different from the transition in the autonomous system, where loss of local stability and transition to chaos are equivalent [1, 21]. The discrepancy of these two measures in the driven system is explained by a dynamic effect: the decrease of the maximum Lyapunov exponent is related to the sharpening of the autocorrelation function by the fluctuating drive. The displacement between local in- stability and transition to chaos leads to an intermediate regime which is absent in time-discrete networks [23] and only exists in their more realistic time-continuous coun- terparts studied here. This hitherto unreported dynam- ical regime combines locally expansive dynamics with asymptotic stability. The seminal works [1, 21] have established a tight link between the fields of random matrix theory and au- tonomous neural networks with random topology: De- terministic chaos emerges if the spectral radius of the coupling matrix exceeds unity. In contrast, we find in stochastically driven networks that the spectral radius only yields a necessary condition for a positive Lyapunov exponent; it determines the minimum of the quantum mechanical potential whose ground state energy relates to the Lyapunov exponent. The presented closed-form relation between input strength, the statistics of the ran- dom matrix, and the onset of chaos (20) generalizes the 9 well known link to non-autonomous stochastic dynamics. It is controversially discussed whether the instability of deterministic rate dynamics explains a transition to chaos in networks of spiking neurons [47–49]. It was ar- gued that such a transition is absent in spiking models because the correlation time does not peak at the point where the corresponding deterministic rate dynamics be- comes unstable [48–50]. For the analysis of oscillations [51] and correlations [52, 53] in these networks, the irreg- ular spiking activity of the neurons can be approximated by effective stochastic rate equations, whereby the real- ization of the spikes is represented by an explicit source In this setting, the input ξ in (1) can be in- of noise. terpreted as such spiking noise, explicitly investigated in [41]. For weak noise one may neglect its impact on the location of the transition. In this limit, noise suppresses the divergence of the correlation time [41]. Our work is not bound to small noise amplitudes and suggests that a diverging time scale in spiking networks does not occur at the instability for two reasons. First, we have shown that in a stochastic system the transition to chaos is not predicted by local instability. If a diverging time-scale at the transition existed it would not occur at the instabil- ity, but as a larger coupling strength. But the presented analysis shows that the decay time of the autocorrelation function does not even peak at the transition to chaos, but rather in the chaotic regime. While these results strictly only hold for the rate dynamics considered here, they still strongly suggest the absence of a diverging time scale in networks of spiking neurons. Indeed, the absence of a diverging time-scale has been observed in simula- tion of spiking neurons [48] as well as in an iterative approach solving for the self-consistent autocorrelation function [54, 55]. To assess whether this richer dynamics found in the driven network has functional consequences, we investi- gate sequential memory [33]. The obtained closed-form expression for the network memory capacity exhibits a peak within the expansive, non-chaotic regime. We iden- tify two mechanisms whose partly antagonistic interplay causes optimal memory: Local amplification of the stim- ulus and intrinsically-generated noise. Local instability of the network ensures sensitivity to the external input, so that on short time scales the incoming signal is am- plified and can therefore more reliably be read out. But larger coupling also increases network-intrinsic fluctua- tions, which, in turn, reduce the susceptibility as well as the signal to noise ratio of the readout. Therefore it is plausible that the optimal memory appears at a point where local amplification of the external input is large enough, but intrinsic chaoticity is still limited. Sequential memory has been studied in a time-discrete neural network model [9], which receives a single weak ex- ternal input. In contrast, we here investigate the memory of a single signal in the presence of multiple simultane- ous inputs with arbitrary amplitude σ. Without addi- tional observation noise, Toyoizumi and Abbott [9] find that sequential memory does not posses a maximum; it is constant and optimal for sub-critical coupling values 0 < g < 1 and falls off in the chaotic regime g > 1 due to intrinsically-generated fluctuations. Perfect reconstruc- tion in the non-chaotic regime is possible, because the single-step delayed activity is a direct linear function of the input. In our setting, the single neuron memory has a similar effect (Figure 5 a). In the discrete system, opti- mal memory close to the transition only arises in presence of observation noise [9]. Memory falloff in the chaotic phase is much more shallow than in the direction of reg- ular dynamics, so that a fine-tuning is not needed if the network dynamic is slightly chaotic. For continuous dynamics the situation is qualitatively different. The network component of the memory or, equivalently, the memory for longer delay times τ , shows non-monotonic behavior even without observation noise. For small signal amplitude σ, memory is optimal right below the transition to chaos and steeply falls off above (Figure 5 b). For large σ, the falloff is weaker in the chaotic regime, qualitatively more similar to Toyoizumi and Abbott [9]. A negative maximum Lyapunov exponent for nonau- tonomous dynamics indicates the echo state property, the reliability of the network response to input [56]. We could indeed show for the analytically tractable model here that memory capacity quickly declines in the chaotic regime due to intrinsically generated chaotic fluctuations. Echo state networks show long temporal memory near the edge of chaos [57–60]. Typically these networks are time- discrete and thus the onset of chaos is directly linked to the spectral radius of the Jacobian. This relation is used in the design of these systems, exploiting that a spec- tral radius close to local instability ensures long memory times. We here show for the driven and time-continuous system, that the edge of chaos and local instability are two different concepts and that memory capacity is a third, distinct measure: Memory is optimal at neither of the two other criteria, but rather in between. In partic- ular, our analytical results for the memory capacity can be used to determine the optimal coupling strength for a given input amplitude. Recently, an algorithm was proposed to train a random network as given by (1) to produce a wide range of activ- ity patterns [8]. Such learning shows best performance if initially the random network without input is in the chaotic state. It has been argued that such networks have a large dynamic range and are able to produce a wide va- riety of outputs. In the training phase the input to the network needs to suppress chaos so that learning con- verges. The procedure therefore requires the choice of an initial coupling that is large enough to ensure chaos, but not too large so that the input can suppress chaos. Our quantitative criterion for the transition easily enables a proper choice of parameters and facilitates the design, 10 control, and understanding of functional networks. In this work we have considered memory of the input signal. An important task of the brain, however, is not only to maintain the input but also to perform non-linear transformations on it. We expect the locally unstable but globally stable dynamics to be beneficial for such a task: The expansive behavior can project the input into a high-dimensional space, which is crucial for non-linear computations or discrimination tasks [58]. Thus, this dynamical regime not only provides memory but might serve as a basis for more complex computations. To show the generic effect of input we added a time- varying input to the seminal model by Sompolinsky et al. [1]. Even though the original model makes some simplify- ing assumptions, such as the all-to-all Gaussian connec- tivity and a sigmoidal symmetric gain function, the tran- sition to chaos is qualitatively the same as in networks with more biologically realistic parameters [20, 41, 50]. To focus on the new physics arising in non-autonomous systems, we have here chosen to present the simplest but non-trivial, and yet application-relevant extension. The reformulation of the derivation of the dynamic mean-field theory by help of established methods from field theory [17, 30, 36, 61, 62] here allowed us to find the explicit form of network memory by a replica calculation. In general, the presented formulation opens the study of recurrent random networks to the rich and powerful set of field the- oretical methods developed in other branches of physics. This language allows a straight-forward extension of our results in various directions. Among them more biolog- ically realistic settings, such as sparse connectivity re- specting Dale's law [63], threshold like-activation func- tions, non-negative activity variables or multiple popula- tions. The latter extension would allow the study of the interesting case in which the population receiving the sig- nal is separated from the readout population. Such a sit- uation would most likely emerge in the cortex where the input population of local microcircuits typically differs from the output population. More generally, the stability of complex dynamical sys- tems plays an important role in various other field of physics, biology, and technology. Examples include os- cillator networks [64], disordered soft-spin models [16], power grids [65], food webs [66], and gene-regularity net- works [67]. Presenting exact results for a prototypical and solvable model this work contributes to the under- standing of chaos and signal propagation in such high dimensional systems. VI. ACKNOWLEDGEMENTS This work was partially supported by Helmholtz young investigator's group VH-NG-1028, Helmholtz portfolio theme SMHB, Jülich Aachen Research Alliance (JARA). This project received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No. 720270. J.S. and S.G. contributed equally to this work. [1] H. Sompolinsky, A. Crisanti, and H. J. Sommers, Phys. Rev. Lett. 61, 259 (1988). [2] C. van Vreeswijk and H. Sompolinsky, Science 274, 1724 (1996). [3] M. Monteforte and F. Wolf, Phys. Rev. Lett. 105, 268104 (2010). [4] G. Lajoie, K. K. Lin, and E. Shea-Brown, Phys. Rev. E 87, 052901 (2013). [5] W. Maass, T. Natschläger, and H. Markram, Neural Comput. 14, 2531 (2002). [6] H. Jaeger and H. Haas, Science 304, 78 (2004). [7] R. Legenstein and W. Maass, Neural Networks 20, 323 (2007). [8] D. Sussillo and L. F. Abbott, Neuron 63, 544 (2009). [9] T. Toyoizumi and L. F. Abbott, Phys. Rev. E 84, 051908 (2011). [10] O. Barak, D. Sussillo, R. Romo, M. Tsodyks, and L. Ab- bott, 103, 214 (2013). [11] K. Rajan, C. D. Harvey, and D. W. Tank, Neuron 90, 128 (2016). [12] N. Li, K. Daie, K. Svoboda, and S. Druckmann, Nature 532, 459 (2016). [13] J. D. Murray, A. Bernacchia, N. A. Roy, C. Constan- tinidis, R. Romo, and X.-J. Wang, Proc. Nat. Acad. Sci. USA p. 201619449 (2016). [14] R. Laje and D. V. Buonomano, Nat. Neurosci. 16, 925 (2013). [15] V. Mante, D. Sussillo, K. V. Shenoy, and W. T. Newsome, Nature 503, 78 (2013). [16] H. Sompolinsky and A. Zippelius, Phys. Rev. Lett. 47, 359 (1981). [17] H. Sompolinsky and A. Zippelius, Phys. Rev. B 25, 6860 (1982). [18] K. Rajan, L. Abbott, and H. Sompolinsky, Phys. Rev. E 82, 011903 (2010). [19] J. Aljadeff, M. Stern, and T. Sharpee, Phys. Rev. Lett. 114, 088101 (2015). [20] O. Harish and D. Hansel, PLoS Comput Biol 11, e1004266 (2015). [21] H. Sommers, A. Crisanti, H. Sompolinsky, and Y. Stein, Phys. Rev. Lett. 60, 1895 (1988). [22] K. Rajan and L. F. Abbott, Phys. Rev. Lett. 97, 188104 (2006). [23] L. Molgedey, J. Schuchhardt, and H. Schuster, Phys. Rev. Lett. 69, 3717 (1992). [24] C. Zhou and J. Kurths, Phys. Rev. Lett. 88, 230602 (2002). [25] P. Martin, E. Siggia, and H. Rose, Phys. Rev. A 8, 423 (1973). [26] H.-K. Janssen, Zeitschrift für Physik B Condensed Mat- ter 23, 377 (1976). [27] C. De Dominicis, J. Phys. Colloques 37, C1 (1976). [28] C. De Dominicis and L. Peliti, Phys. Rev. B 18, 353 (1978). [29] J. Schuecker, S. Goedeke, D. Dahmen, and M. Helias, arXiv (2016), 1605.06758 [cond-mat.dis-nn]. 11 [30] M. Moshe and J. Zinn-Justin, Physics Reports 385, 69 (2003), ISSN 0370-1573. [31] J.-P. Eckmann and D. Ruelle, Reviews of modern physics 57, 617 (1985). [32] B. Derrida and Y. Pomeau, EPL (Europhysics Letters) 1, 45 (1986). [33] H. Jaeger, Short term memory in echo state networks, vol. 5 (GMD-Forschungszentrum Informationstechnik, 2001). [34] K. Fischer and J. Hertz, Spin glasses (Cambridge Uni- versity Press, 1991). [35] S.-I. Amari, Systems, Man and Cybernetics, IEEE Trans- actions on pp. 643–657 (1972). [36] A. Crisanti and H. Sompolinsky, Phys. Rev. A 36, 4922 (1987). [37] J. Zinn-Justin, Quantum field theory and critical phe- nomena (Clarendon Press, Oxford, 1996). [38] T. Cabana and J. Touboul, J. statist. Phys. 153, 211 (2013). [39] C. W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences (Springer- Verlag, Berlin, 1985), 2nd ed., ISBN 3-540-61634-9, 3- 540-15607-0. [40] A. Altland and S. B., Concepts of Theoretical Solid State Physics (Cambridge university press, 2010). [41] J. Kadmon and H. Sompolinsky, Phys. Rev. X 5, 041030 (2015). [42] R. Price, IRE Transactions on Information Theory 4, 69 (1958). [43] A. Papoulis, Probability, Random Variables, and Stochas- (McGraw-Hill, Boston, Massachusetts, tic Processes 1991), 3rd ed. [44] Note1, the theory of random dynamical systems makes this more precise; a brief overview is given in [4]. [45] J. Dambre, D. Verstraeten, B. Schrauwen, and S. Massar, Scientific reports 2, 514 (2012). [46] M. Hermans and B. Schrauwen, in Neural Networks (IJCNN), The 2010 International Joint Conference on (IEEE, 2010), pp. 1–7. [47] S. Ostojic, Nat. Neurosci. 17, 594 (2014). [48] R. Engelken, F. Farkhooi, D. Hansel, C. van Vreeswijk, and F. Wolf, bioRxiv p. 017798 (2015). [49] S. Ostojic, bioRxiv p. 020354 (2015). [50] F. Mastroguiseppe and S. Ostojic, arXiv p. 1605.04221 (2016). [51] N. Brunel and V. Hakim, Neural Comput. 11, 1621 (1999). [52] J. Trousdale, Y. Hu, E. Shea-Brown, and K. Josic, PLOS Comput. Biol. 8, e1002408 (2012). [53] M. Helias, T. Tetzlaff, and M. Diesmann, New J. Phys. 15, 023002 (2013). [54] B. Dummer, S. Wieland, and B. Lindner, Front. Comput. Neurosci. 8, 104 (2014). [55] S. Wieland, D. Bernardi, T. Schwalger, and B. Lindner, Phys. Rev. E 92, 040901 (2015). [56] G. Wainrib and M. N. Galtier, Neural Networks 76, 39 (2016), ISSN 0893-6080. [57] N. Bertschinger and T. Natschläger, Neural Comput. 16, 1413 (2004). [58] R. Legenstein and W. Maass, What makes a dynamical (MIT Press, 2007), system computationally powerful? pp. 127–154. [59] L. Büsing, B. Schrauwen, and R. Legenstein, Neural Comput. 22, 1272 (2010). [60] J. Boedecker, O. Obst, J. T. Lizier, N. M. Mayer, and M. Asada, Theory in Biosciences 131, 205 (2012). [61] C. Chow and M. Buice, The Journal of Mathematical Neuroscience 5 (2015). [62] J. A. Hertz, Y. Roudi, and P. Sollich, Journal of Physics A: Mathematical and Theoretical 50, 033001 (2017). [63] J. Eccles, F. P, and K. Koketsu, J Physiol (Lond) 126, 524 (1954). [64] F. A. Rodrigues, T. K. D. Peron, P. Ji, and J. Kurths, Phys. Rep. 610, 1 (2016). [65] T. Nishikawa and A. E. Motter, New J. Phys. 17, 015012 (2015). [66] S. Allesina and S. Tang, Population Ecology 57, 63 (2015). [67] A. Pomerance, E. Ott, M. Girvan, and W. Losert, Proc. Nat. Acad. Sci. USA 106, 8209 (2009). [68] J. W. Negele and H. Orland, Quantum Many-Particle Systems (Perseus Books, 1998). [69] M. Abramowitz and I. A. Stegun, Handbook of Mathe- matical Functions: with Formulas, Graphs, and Mathe- matical Tables (Dover Publications, New York, 1974). Appendix A: Derivation of mean-field equation The generating functional Z[l](J) in (2) is properly normalized independent of the realization of J. This property allows us to follow [28] and to introduce the disorder-averaged generating functional ¯Z[l] := hZ[l](J)iJ =Z Πij dJij N (0, g2 N , Jij ) Z[l](J). (A1) The coupling term exp(cid:16)−Pi6=j Jij xT torizes over unit indices i, j and the random weights Jij appear linear in the exponent. Thus we can sepa- rately integrate over the independently and identically distributed Jij, i 6= j, by completing the square and ob- tain i φ(xj )(cid:17) in (2) fac- g2 N Z dJijN (0, = exp(cid:18) g2 2N (cid:0)xT , Jij) exp(cid:0)−Jij xT i φ(xj )(cid:1)2(cid:19) . i φ(xj )(cid:1) (A2) We reorganize the resulting sum in the exponent of the coupling term as 12 1 = = g2 g2 g2 2N Xi6=j(cid:18)Z xi(t)φ(xj (t)) dt(cid:19)2 2N Xi6=jZ Z xi(t)φ(xj (t)) xi(t′)φ(xj (t′)) dt dt′ φ(xj (t))φ(xj (t′)) 2Xi Z Z xi(t)xi(t′)   dt dt′  2Xi Z Z xi(t)xi(t′) − where we used (cid:0)R f (t)dt(cid:1)2 = RR f (t)f (t′) dt dt′ in the first step andPij xiyj =Pi xiPj yj in the second. The last line is the diagonal (self-coupling) to be skipped in It is a correction of order N−1 and the double sum. will be neglected in the following. The disorder-averaged generating functional (A1) therefore takes the form φ(xi(t))φ(xi(t′)) dt dt′, N Xj g2 N (A3) 1 ¯Z[l] =Z DxZ Dx exp(cid:16)S0[x, x] + lTx(cid:17) × exp(cid:16) 1 2 xT Q1x(cid:17) , (A4) where we extended our notation with xTAy RR x(t) A(t, t′) y(t′) dt dt′ to bi-linear forms and defined := φ(xj (t))φ(xj (t′)) . (A5) Q1(t, t′) := g2 N Xj The field Q1 is an empirical average over N contributions, which, by the law of large numbers and in the case of weak correlations, will converge to its expectation value for large N . This heuristic argument is shown in the following more formally: A saddle-point approximation leads to the replacement of Q1 by its (self-consistent) expectation value. To this end we first decouple the in- teraction term by inserting the Fourier representation of the Dirac-δ functional: (A6) N g2 QT N g2 Q1(t, s) + φ(x(t))T φ(x(s))] 1 Q2 + φ(x)TQ2φ(x)(cid:19) , δ[− =Z DQ2 exp(cid:18)− 1 Q2 := RR Q1(t, s) Q2(t, s) dt ds = PN i=1 φ(xi(t))φ(xi(s)). We note that the conjugate field Q2 ∈ iR is purely imaginary. We hence rewrite (A4) as where we further extended our notation with QT φ(x(t))Tφ(x(s)) and ¯Z[j, j] =Z DQ1Z DQ2 exp(cid:18)− N g2 QT Ω[Q1, Q2] =Z DxZ Dx exp(cid:16)S0[x, x] + 1 2 1 Q2 + N ln Ω[Q1, Q2] + jTQ1 + jTQ2(cid:19) xTQ1 x + φ(x)TQ2φ(x)(cid:17), 13 (A7) where we introduced source terms j, j for the auxiliary fields and dropped the original source terms lTx. The integral measures DQ1,2 must be defined suitably. In writing N ln Ω[Q1, Q2] we have used that the auxiliary fields couple only to sums of fields Pi φ2(xi) andPi x2 i , so that the generating functional for the fields x and x factorizes into a product of N identical factors Ω[Q1, Q2]. The remaining problem can be considered a field the- ory for the auxiliary fields Q1 and Q2. The form (A7) clearly exposes the N dependence of the action for these latter fields: It is of the form R dQ exp(N f (Q)), which, for large N , suggests a saddle point approximation, which neglects fluctuations in the auxiliary fields and hence sets them equal to their expectation value; this point is the dominant contribution to the probability mass. To ob- tain the saddle point equations we consider the Legendre- Fenchel transform of ln ¯Z as Γ[q1, q2] := sup j,j jTq1 + jTq2 − ln ¯Z[j, j], δΓ δq1 It holds that = j and δΓ δq2 called the vertex generating functional or effective ac- = j, tion [37, 68]. called equations of state. The leading order mean-field or tree-level approximation amounts to the approximation Γ[q1, q2] ≃ −S[q1, q2], where S[Q1, Q2] = − N 1 Q2 + N ln Ω[Q1, Q2] is the action for the auxiliary fields Q1 and Q2. We insert this tree-level approximation into the equations of state and further set j = j = 0 since the source fields have no physical meaning and thus must vanish. We get the saddle point equations g2 QT 0 = = δS[Q1, Q2] δQ{1,2} δ δQ{1,2}(cid:18)− N g2 QT 0 = − N g2 Q∗1(t, s) + from which we obtain a pair of equations (A8) (A9) N Ω δΩ[Q1, Q2] 1 Q2 + N ln Ω[Q1, Q2](cid:19) δQ2(t, s) (cid:12)(cid:12)(cid:12)(cid:12)Q∗ δQ1(t, s) (cid:12)(cid:12)(cid:12)(cid:12)Q∗ δΩ[Q1, Q2] ↔ Q∗1(t, s) = g2 hφ(x(t))φ(x(s))iQ∗ =: g2Cφ(x)φ(x)(t, s) N Ω N g2 Q∗2(t, s) + 0 = − g2 2 hx(t)x(s)iQ∗ = 0, ↔ Q∗2(t, s) = where we defined the average autocorrelation function Cφ(x)φ(x)(t, s) of the non-linearly transformed activity of the units. The second saddle point Q∗2 = 0 vanishes, because the field was introduced to represent a Dirac δ constraint in Fourier domain. One can show that conse- quently R DQ exp(S[Q1, Q2])Q2 = 0, which is the true mean value Q∗2 = hQ2i = 0, known as the Deker-Haake theorem. Here hiQ∗ denotes the expectation value with respect to realizations of x evaluated at the saddle point Q∗. The expectation value must be computed self-consistently, since the values of the saddle points, by (A7), influence the statistics of the fields x, which in turn determines the function Q∗1 by (A9). Inserting the saddle point solution into the generating functional (A7) we get (4) ¯Z∗ ∝Z DxZ Dx exp(cid:16)S0[x, x] + g2 2 xTCφ(x)φ(x) x(cid:17). The action has the important property that it de- composes into a sum of actions for individual, non- interacting units that each feel a field with a common, self-consistently determined statistics, characterized by its second cumulant Cφ(x)φ(x). Prior to the saddle point approximation (A7) the fluctuations in the field Q1 are common to all the single units, which effectively couples them. The saddle-point approximation replaces the fluc- tuating field Q1 by its mean (A9), which reduces the network to N non-interacting units, or, equivalently, a single unit system. The second term in (4) is a Gaussian noise with a two point correlation function Cφ(x)φ(x)(t, s). The physical interpretation is the noisy signal each unit receives due to the input from the other N units. Its au- tocorrelation function is given by the summed autocorre- lation functions of the output activities φ(xi(t)) weighted by g2N−1, which incorporates the Gaussian statistics of the couplings. The interpretation of the noise can be appreciated by explicitly considering the moment generating functional of a Gaussian noise with a given autocorrelation func- tion C(t, s), which leads to the cumulant generating func- tional ln Zη[x] that appears in the exponent of (4) and has the form ln Zη[−x] = lnhexp(cid:0)−xT η(cid:1)i xT C x. = 1 2 Note that the only non-vanishing cumulant of the effec- tive noise is the second cumulant; the cumulant generat- ing functional is quadratic in x. This means the effective noise is Gaussian and only couples pairs of time points in proportion to the correlation function. Appendix B: Stationary process We rewrite equation (5) as (∂t + 1) x(t) = η(t), (B1) where we combined the two independent Gaussian pro- cesses η and ξ appearing in (5) into η(t). We then multi- ply (B1) for time points t and s and take the expectation value over realizations of the noise η on both sides, which leads to (∂t + 1) (∂s + 1) Cxx(t, s) = g2 Cφ(x)φ(x)(t, s) + 2σ2δ(t − s), (B2) where we defined the covariance function of the activi- ties Cxx(t, s) := hx(t)x(s)i. We are now interested in the stationary statistics Cxx(t, s) =: c(t − s) of the system. The inhomogeneity in (B2) is then also time-translation invariant; Cφ(x)φ(x)(t+τ, t) is only a function of τ . There- fore the differential operator (∂t + 1) (∂s + 1) c(t − s), with τ = t − s, simplifies to (−∂2 τ + 1) c(τ ) so we get τ + 1) c(τ ) = g2 Cφ(x)φ(x)(t + τ, t) + 2σ2 δ(τ ), (−∂2 (B3) given as (7) in the main text. of the connectivity J, as in (A2). We therefore need to evaluate the Gaussian integral 14 xαT i φ(xα j )! 2 Xα=1 Z dJijN (0, = exp g2 × exp(cid:18) g2 2N 2 g2 N , Jij ) exp −Jij j )(cid:3)2! Xα=1(cid:2)xαT j )(cid:3)(cid:2)x2T i φ(xα i φ(x1 i φ(x2 N (cid:2)x1T j )(cid:3)(cid:19) . The first exponential factor only includes variables of a single subsystem and is identical to the term appearing in (A4). The second exponential factor is a coupling term between the two systems arising from the identical ma- trix J in the two replicas in each realization that enters the expectation value. We treat the former terms as be- fore and here concentrate on the mixed coupling term. Analogous to (A3), the exponent of the mixed coupling term can be rewritten as (C1) (C2) g2 N Xi6=j (cid:2)x1T =ZZ Xi i φ(x1 i φ(x2 j )(cid:3) j )(cid:3)(cid:2)x2T N Xj g2 i (s) x1 i (t)x2 φ(x1 j (t)) φ(x2 j (s)) dt ds + O(N−1), where we included the self coupling term i = j, which is again a subleading correction of order N−1. We now follow the steps in Appendix A and introduce three pairs of auxiliary variables. The pairs Qα 2 are defined as before in (A5) and (A6), but for each subsys- tem, while the pair T1, T2 decouples the mixed term (C2) by defining 1 , Qα Appendix C: Replica calculation to assess the Lyapunov exponent T1(t, s) := g2 N Xj φ(x1 j (t)) φ(x2 j (s)). We start from the generating functional for the pair of systems (13) and perform the average over realizations Taken together, we can therefore rewrite the generating functional (13) averaged over the couplings as ¯Z := hZ(J)iJ = Π2 S[{Qα 1 , Qα 2}α∈{1,2}, T1, T2] := − Ω12[{Qα 1 , Qα 2}α∈{1,2}, T1, T2] = Π2 2 2(cid:27)Z DT1Z DT2 exp(cid:16)S[{Qα 1 Z DQα N 1 T2 + N ln Ω12[{Qα g2 T T 2 − α=1(cid:26)Z DQα N Xα=1 g2 QαT α=1nZ DxαZ Dxα exp(cid:16)S0[xα, xα] + xαTQα 1 Qα 1 , Qα 1 2 × exp(cid:16)x1T(cid:0)T1 + 2σ2(cid:1) x2 + φ(x1)TT2φ(x2)(cid:17)(cid:17) , 1 , Qα 2}α∈{1,2}, T1, T2](cid:17) (C3) 2}α∈{1,2}, T1, T2] 1 xα + φ(xα)TQα 2 φ(xα)(cid:17)o where we used that the generating functional factorizes into a product of 2N identical factors Z 12. 1 , Qα Analogously to Appendix A we could introduce sources for the auxiliary fields Qα 2 , T1, T2. Then the equa- tions of state are obtained from the vertex-generating functional Γ as before, which, in the tree-level approx- imation is given by Γ = −S and for vanishing sources != 0. leads to the saddle-point equations From the latter we obtain the set of equations = δS δT1,2 δS δQα 1,2 Qα∗1 (t, s) = g2 1 Ω12 δΩ12 2 (t, s) δQα = g2 hφ(xα(t))φ(xα(s))iQ∗ ,T ∗ (C4) Qα∗2 (t, s) = 0 T ∗1 (t, s) = g2 1 Ω12 T ∗2 (t, s) = 0. δΩ12 δT2(t, s) = g2 hφ(x1(t))φ(x2(s))iQ∗ ,T ∗ 15 for α, β ∈ {1, 2} 2σ2δ(t − s) + g2Fφ(cid:0)cαβ(t, s), cαα(t, t), cββ(s, s)(cid:1) , (C6) where the function Fφ is defined as the expectation value (∂t + 1) (∂s + 1) cαβ(t, s) = for the centered bi-variate Gaussian distribution Fφ(c12, c11, c22) :=(cid:10)φ(x1)φ(x2)(cid:11) (cid:18)x1 x2(cid:19) ∼ N2(cid:18)0,(cid:18)c11 c12 c12 c22(cid:19)(cid:19) . First, we observe that the equations for the autocorrela- tion functions cαα(t, s) decouple and can each be solved separately, leading to the same equation (9) as before. This formal result could have been anticipated, because the marginal statistics of each subsystem cannot be af- fected by the mere presence of the respective other sys- tem. Their solutions The generating functional at the saddle point therefore is c11(t, s) =c22(t, s) = c(t − s) xαTQα∗1 xα(cid:17) (C5) then provide the "background" for the equation for the cross-correlation function between the two copies; they fix the second and third argument of the function Fφ on the right-hand side of (C6). It remains to determine the equation of motion for c12(t, s). α=1DxαDxα exp(cid:16) 2 ¯Z∗ =ZZ Π2 Xα=1 × exp(cid:0)x1T(cid:0)T ∗1 + 2σ2(cid:1) x2(cid:1) . We make the following observations: S0[xα, xα] + 1 2 1. The two subsystems α = 1, 2 in the first line of (C5) have the same form as in (4). This has been expected, because there is no physical cou- pling between the two systems. This implies that the marginal statistics of the activity in one sys- tem cannot be affected by the mere presence of the second. Hence in particular the saddle points Qα∗1,2 must be the same as in (4). 2. If the term in the second line of (C5) was absent, the statistics in the two systems would be indepen- dent. Two sources, however, contribute to corre- lations between the systems: The common Gaus- sian white noise that gives rise to the term ∝ 2σ2 and the non-white Gaussian noise due to a non-zero value of the auxiliary field T ∗1 (t, s). 3. Only products of pairs of fields appear in (C5), so that the statistics of the xα is Gaussian. From (C5) and (C4) we can read off the pair of effective dynamical equations (14) with self-consistent statistics (15). 1. Derivation of the variational equation We first determine the stationary solution c12(t, s) = k(t − s). We see immediately from (C6) that k(τ ) obeys the same equation of motion as c(τ ), so k(τ ) = c(τ ) is a solution. The distance (12) between replicas for this solution therefore vanishes; the dynamics in both repli- cas follows identical realizations. Let us now study the stability of this solution. We hence need to expand c12 around the stationary solution c12(t, s) = c(t − s) + ǫ k(1)(t, s) , ǫ ≪ 1 . We expand the right hand side of (C6) into a Taylor series using Price's theorem and (8) Fφ(cid:0)c12(t, s), c0, c0(cid:1) = fφ(cid:0)c12(t, s), c0(cid:1) = fφ (c(t − s), c0) + ǫ fφ′ (c(t − s), c0) k(1)(t, s) + O(ǫ2) . Inserted into (C6) and using that c solves the lowest or- der equation, we get the linear equation of motion for the first order deflection (16). By (12) the first order deflec- tion k(1)(t, s) determines the distance between the two subsystems as d(t) = c11(t, t) + c22(s, s) c0 {z } = −2ǫ k(1)(t, t). {z } c0 −c12(t, t) − c21(t, t) } −2c0−2ǫ k(1) (t,t) {z (C7) We multiply the equation (14) for α = 1 and α = 2 and take the expectation value on both sides, so we get The negative sign makes sense, since we expect in the t,s→∞= 0, so k(1) must be of chaotic state that c12(t, s) opposite sign than c > 0. 2. Schrödinger equation for the maximum Lyapunov exponent We here want to reformulate the equation for the variation of the cross-system correlation (16) into a Schrödinger equation, as in the original work [1, eq. 10]. First, noting that Cφ′φ′(t, s) = fφ′ (c(t − s), c0) is time translation invariant, it is advantageous to introduce the coordinates T = t + s and τ = t− s and write the covari- ance k(1)(t, s) as k(T, τ ) with k(1)(t, s) = k(t + s, t − s). The differential operator (∂t + 1) (∂s + 1) with the chain rule ∂t → ∂T + ∂τ and ∂s → ∂T − ∂τ in the new coordinates is (∂T + 1)2 − ∂2 τ . A separation ansatz 2 κT ψ(τ ) then yields the eigenvalue equation k(T, τ ) = e 1 16 (D1) hxi(t)z(t0)i2 , K Xi=1 with t0 = t − τ and K denoting the number of neurons connected to the readout, which we initially leave as a free parameter. Here hi denotes the average over real- izations of the inputs ξi (or alternatively over time) and the overbar the average over realizations of the connec- tivity J as in (A1). Moreover, we here examine a more j=1 vjξj (t) where v denotes general input signal z(t) =PN We pick two points in time t, s ≥ t0 and define the input weights. ( κ 2 + 1)2ψ(τ ) − ∂2 τ ψ(τ ) = g2fφ′ (c(τ ), c0) ψ(τ ) hK(t, s) := hxi(t)z(t0)ihxi(s)z(t0)i . (D2) K Xi=1 for the growth rates κ of d(t) = −2ǫk(1)(t, t) = −2ǫk(2t, 0). We can express the right hand side by the second derivative of the potential (10) so that with V ′′(c(τ ); c0) = −1 + g2fφ′ (c(τ ), c0) (C8) we get the time-independent Schrödinger equation + 1(cid:17)2(cid:19) τ − V ′′(c(τ ); c0)(cid:1) ψ(τ ) =(cid:18)1 −(cid:16) κ (cid:0)−∂2 } {z =:E 2 ψ(τ ), (C9) where the time lag τ plays the role of a spatial coordi- nate for the Schrödinger equation. The eigenvalues ("en- ergies") En determine the exponential growth rates κn of the solutions k(2t, 0) = eκnt ψn(0) at τ = 0 with κ±n = 2(cid:16)−1 ±p1 − En(cid:17) . (C10) We can therefore determine the growth rate of the mean- square distance of the two subsystems by (C7). The fastest growing mode of the distance is hence given by the ground state energy E0 and the plus in (C10). The deflection between the two subsystems therefore growth with the rate λmax = κ+ 0 (C11) 1 2 = −1 +p1 − E0, where the factor 1/2 in the first line is due to d being the squared distance, hence the length √d growth with half the exponent as d. The measure of interest, (D1), then follows for t = s. The key idea is to express the correlator hxi(t)z(t0)i as a weighted sum of response functions hxi(t)x(t0)i, which we show in the following. We introduce a scalar source term k(t) for the signal z(t) and average over the noise ξ. This yields the generating functional Z[l, k] =Z Dx Z D x× × exp(Xi S0[xi, xi] −Xj (D3) Jij φ(xj ) − 2σ2vi kT xi + lT i xi). Evaluating the correlator leads to hxi(t)z(t0)i = δ2Z δli(t)δk(t0)(cid:12)(cid:12)(cid:12)ji=k=0 N = −2σ2 vjhxi(t)xj (t0)i . Xj=1 (D4) We now consider a pair of systems (replicas) similarly as in Appendix C with the difference, however, that the two systems receive independent realizations of the inputs ξ. We need two independent systems to express the product of the two correlators in (D1). By independence, the corresponding average factorizes, hxi(t)z(t0)ihxi(s)z(t0)i = (cid:10)x1 =: hi(t, s, t0), i (t)z1(t0) x2 i (s)z2(t0)(cid:11) (D5) where the superscript denotes the replicon index as be- fore. To obtain h, it is sufficient to introduce a single source term 4σ4Xj,l vjvlZ dtǫ(t)x1 j (t)x2 l (t) (D6) Appendix D: Memory curve with source ǫ(t) to the corresponding generating func- tional, which allows us to obtain hxi(t)z(t0)i, with (D4) and (D5) as To evaluate (21) we need to determine the disorder- averaged sum of squared response functions hxi(t)z(t0)ihxi(s)z(t0)i = δ δl1 i (t) δ δl2 i (s) δ δǫ(t0) Z(cid:12)(cid:12)l=ǫ=0. The additional source term (D6) has the physical inter- pretation of a common input with time-dependent vari- ance ǫ(t) injected into a pair of units between the two replicas. The absence of quadratic terms ∝ (xα)2 shows that this common input does not affect the marginal statistics of the two systems in isolation. This interpreta- tion is here only mentioned for illustrative purposes; the derivation does not rely on it.Due to the weight vjvl for different unit pairs j, l we keep the single neuron index in the following. The goal now is to derive a differential equation for the disorder averaged hi(t, s, t0) similar to Appendix C 1, needed to compute (D1). First, after averaging over the disorder, completely analogous to Appendix C, we can read off effective equa- tions for the single units (∂t + 1) xα i (t) = ξα i (t) + ηα i (t) + ρα i (t) (D7) α ∈ {1, 2}, i ∈ {1, ..., N}, together with a set of self- consistency equations for the statistics of the noises i (t) ξβ hξα i (t) ηβ hηα i (t) ρβ hρα j (s)i = 2σ2δαβδij δ(t − s) hφ(xα i (s))i j (s)i = j (s)i = 4σ4(1 − δαβ) νiνjǫ(t) δ(t − s) . δij Xi i (t))φ(xβ g2 N (D8) The first line in (D8) represents the independent noise between the systems, the second line the common con- nectivity and the third line the common noise component we introduced in (D6) to express the squared response function (D2). i (t), l2 Second, we obtain hi(t, s, t0) by a functional deriva- tive with respect to l1 i (s) and ǫ(t0), which can be seen from its representation as the four-point correlator in (D5). Writing the functional derivative with respect to ǫ explicitly as a limit, we can express h by the correlation between the pair of systems hi(t, s, t0) = lim ι→0 1 ι hx1 i (t)x2 i (s)i(cid:12)(cid:12)(cid:12)ǫ=ι δ(◦−t0)] , (D9) where we used that for ǫ = 0 the two systems are un- correlated. We now combine the effective equation (D7) and (D9) to obtain a partial differential equation for h : (∂t + 1) (∂s + 1) hi(t, s, t0) (D10) 1 = g2 N ιh N Xi=1 hφ(x1 lim ι→0 +4σ4ν2 i δ(t − s) δ(t − t0). i (t))φ(x2 i (s))ii(cid:12)(cid:12)(cid:12)ǫ=ι δ(◦−t0) Since we are interested in the limit ι → 0, we expand the first term to linear order around the uncorrelated state 17 hφ(x1 i (t))φ(x2 i (s))i = fφ (0, c0) + ∂1fφ (0, c0) c12 i (t, s) 2 = hφ′(xi)i hi(t, s, t0) , (D11) 2 = 0. Inserting (D11) into (D10) we arrive at where the first term vanishes as it factorizes into hφ(x)i (∂t + 1) (∂s + 1) hi(t, s, t0) =Xi 2 hi(t, s, t0) hφ′(xi)i i δ(t − t0) δ(s − t0) hi(t, s, t0) N 2 + 4σ4 ν2 = hφ′(xj)i + 4σ4 ν2 Xi i δ(t − t0) δ(s − t0) . In the latter step we used that hφ′(xi)i is independent of i because the expectation value is taken with respect to the disorder averaged unperturbed system and thus we use a representative unit j as the index. Taking the sum with respect to i = 1, ..., K yields (∂t + 1) (∂s + 1) hK(t, s) = g2hφ′ (xj )i 2 K N hN (t, s) (D12) + 4σ4kvKk2δ(t − t0)δ(s − t0) i . For the complete sum of squared response functions, hN (t, s), the following closed linear partial differential equation holds: with kvKk2 =PK i=1 v2 (∂t + 1) (∂s + 1) hN (t, s) = g2hφ′ (xj)i 2 hN (t, s) + 4σ4δ(t − t0)δ(s − t0) , (D13) where we set kvNk2 = kvk2 = 1 without loss of general- ity. The solution to this equation describes the shape of the memory curve if the readout has access to the states of all neurons. To determine hK(t, s) we note that the difference hK(t, s) − K N hN (t, s) is proportional to the so- lution of (∂t + 1) (∂s + 1) h(0)(t, s) = δ(t − t0)δ(s − t0) , which by direct integration yields h(0)(t, s) = e−(t−t0)Θ(t − t0)e−(s−t0)Θ(s − t0) . Thus, hK(t, s) is given by (D14) (D15) hK(t, s) = K N K hN (t, s) + 4σ4(cid:18)kvKk2 − h(1)(t, s) + 4σ4(cid:18)kvKk2 − = 4σ4 K N N(cid:19) h(0)(t, s) K N(cid:19) h(0)(t, s) , (D16) where h(1)(t, s) solves (∂t + 1) (∂s + 1) h(1)(t, s) = a2h(1)(t, s) + δ(t − t0)δ(s − t0) (D17) For the memory curve we only need the solution u(T, τ ) for τ = 0, the diagonal s = t in the original coordinates. Setting τ = 0 in the Fourier representation gives the Laplace transform of u(T, τ = 0): 18 with parameter a2 = g2hφ′(xj )i ∞ . Here τ∞ is the time scale of the asymptotic decay of the autocor- relation function. = 1 − 1/τ 2 2 As in Appendix C 2 it is useful to change coordinates to T = t + s − 2t0 and τ = t − s. In these coordinates (D17) takes the form u(p) = u(p, τ = 0) = = 1 k2 + p2 − a2 dk 1 π +∞Z−∞ pp2 − a2 1 (D20) τ h(1)(T, τ ) = a2h(1)(T, τ ) + 2δ(T )δ(τ ) (∂T + 1)2h(1)(T, τ ) − ∂2 and setting h(1)(T, τ ) = e−T u(T, τ ) simplifies the PDE further to with p ∈ C such that Re(cid:0)p2(cid:1) > a2. The function on the right is the Laplace transform of the modified Bessel function of the first kind I0(aT ) [69]. Together with h(1)(T, τ ) = e−T u(T, τ ) we therefore obtain the shape of the memory curve as ∂2 T u(T, τ ) − ∂2 τ u(T, τ ) = a2u(T, τ ) + 2δ(T )δ(τ ) , (D18) a Klein-Gordon wave equation with temporal coordi- nate T and spatial coordinate τ (and negative squared mass −a2). We are looking for the solution u(T, τ ) in T ≥ 0, τ ∈ R. To this end we consider the tempo- ral Laplace and the spatial Fourier transform of (D18). Fourier transformation in τ yields ∂2 T u(T, k) +(cid:0)k2 − a2(cid:1)u(T, k) = 2δ(T ) with k ∈ R and the Fourier representation (D19) u(T, τ ) = 1 2π eikτ u(T, k) dk . +∞Z−∞ For each k ∈ R the Laplace transformation in T , u(p, k) = ∞Z0 e−pT u(T, k) dT , of (D19) reads p2 u(p, k) − p u(0, k) {z }=0 − ∂T u(0, k) {z } =0 Hence, in the Fourier-Laplace domain we obtain +(cid:0)k2 − a2(cid:1)u(p, k) = 2 . u(p, k) = 2 p2 + k2 − a2 . h(1)(T ) = h(1)(T, τ = 0) = e−T I0(aT ) Θ(T ) . (D21) Finally, using (D15) and (D21) in (D16) gives the fol- lowing explicit expression (setting t0 = 0) for the sum of squared response functions hK(t) = hK(t, t) = 4σ4 K N e−2t I0(a2t) Θ(t) N(cid:19) e−2t Θ(t) K + 4σ4(cid:18)kvKk2 − = 4σ4 K N + 4σ4kvKk2e−2t Θ(t) . e−2t (I0(a2t) − 1) Θ(t) (D22) In (D22) we split hK(t) into two contributions: a network contribution hnet K (t) proportional to K/N with shape e−2t (I0(a2t) − 1) Θ(t) and a local contribution propor- tional to kvKk2 with shape e−2t Θ(t). The latter is just the memory of the signal due to the leaky integration of the single units, while the former describes the memory due to the collective network dynamics; only this contri- bution is affected by the network parameters. We can evaluate (21) using (D22) with the choice vi = 1/√N ∀i which leads to (22) and to the network memory (23).
1310.0479
1
1310
2013-10-01T20:20:13
Nonlinear Observer Design and Synchronization Analysis for Classical Models of Neural Oscillators
[ "q-bio.NC" ]
This work explores four nonlinear classical models of neural oscillators, the Hodgkin-Huxley model, the Fitzhugh-Nagumo model, the Morris-Lecar model, and the Hindmarsh-Rose model. Nonlinear contraction theory is used to develop observers and perform synchronization analysis on these systems. Neural oscillation and signaling models are based on the biological function of the neuron, with behavior mediated through the channeling of ions across the cell membrane. The variable assumed to be measured is the membrane potential, which may be obtained empirically through the use of a neuronal force-clamp system, or may be transmitted through the axon to other neurons. All other variables are estimated by using partial state or full state observers. Basic observer rate convergence analysis is performed for the Fitzhugh Nagumo system, partial state observer design is performed for the Morris-Lecar system, and basic synchronization analysis is performed for both the Fitzhugh-Nagumo and the Hodgkin-Huxley systems.
q-bio.NC
q-bio
Nonlinear Observer Design and Synchronization Analysis for Classical Models of Neural Oscillators Ranjeetha Bharath and Jean-Jacques Slotine Massachusetts Institute of Technology ABSTRACT This work explores four nonlinear classical models of neural oscillators, the Hodgkin- Huxley model, the Fitzhugh-Nagumo model, the Morris-Lecar model, and the Hindmarsh-Rose model. Nonlinear contraction theory is used to develop observers and perform synchronization analysis on these systems. Neural oscillation and signaling models are based on the biological function of the neuron, with behavior mediated through the channeling of ions across the cell membrane. The variable assumed to be measured is the membrane potential, which may be obtained empirically through the use of a neuronal force-clamp system, or may be transmitted through the axon to other neurons. All other variables are estimated by using partial state or full state observers. Basic observer rate convergence analysis is performed for the Fitzhugh Nagumo system, partial state observer design is performed for the Morris-Lecar system, and basic synchronization analysis is performed for both the Fitzhugh-Nagumo and the Hodgkin- Huxley systems. 1. INTRODUCTION The goal of this work is to analyze four different biological models from a control systems perspective by designing nonlinear observers for them and studying their synchronization properties. The four models are the Hodgkin-Huxley model, the Fitzhugh-Nagumo model, the Morris-Lecar model, and the Hindmarsh-Rose model. These four models describe neuronal activity by modeling neurons in an oscillatory fashion. Neuronal networks involve “networks of interactions between cells” and consist of “the network of synaptic connections between neurons.” 1 The time scale for neuronal activity is milliseconds and the length scale is cellular.1 Analyzing the mathematical behavior of neural networks can lead to interesting insights about cellular behavior and interaction. The first section describes a set of mathematical notions underlying the theory presented in this work, such as contraction analysis. The next section discusses the biological motivation and mathematical analysis of the Hodgkin-Huxley model developed as a four- state description of signaling. The third section describes the Fitzhugh-Nagumo model, which is a two-state simplification of the Hodgkin-Huxley model. The fourth section describes the Morris-Lecar model, which is a three-state description. Finally, the fifth section describes the Hindmarsh-Rose model, which is another three state model description. In all cases, for observer design, the membrane potential or voltage is taken to be the measured variable. These sections are followed by a conclusion and references. 2. MATHEMATICAL NOTIONS: CONTRACTION 12 The concept of contraction can be made rigorous by studying the system mathematically. The deterministic system, which can be multivariable and nonlinear in any and all of the state variables, is given as: ( ) It is valuable to define what is meant, in matrix algebra, by uniformly negative definite ( ) is uniformly negative before proceeding with the system analysis. To say that ( ) definite is to say that ( ) However, a simple application of this idea can be used in the following manner instead14: a virtual system constructed as a copy of the real system which accounts for the coupling “is contracting if the maximum eigenvalue of the symmetric part of F, where F is the generalized Jacobian, is uniformly negative.” This is the notion used in the remainder of this work in order to study the synchronization behavior of coupled neurons. 2 5-8 3. HODGKIN-HUXLEY 3.1 BIOLOGICAL MOTIVATION By studying the electrical behavior of the giant axon in squid, A. L. Hodgkin and A. F. Huxley were able to create a mathematical model describing neuronal activity in 1955. Naturally, this model is a simplification of the significantly more complicated real neural potential dynamics. The three current components, which together make up the basis for ion transfer across the membrane, are INa, IK, and Il, referring to sodium, potassium, and other ions, respectively. The sodium and potassium conductances, gNa and gK, are functions of time, but other factors such as ionic driving gradients (e.g. ENa), are viewed as constant. It is proposed that the membrane potential controls permeability. ( ) ( ) ( ) The paper finds equations to describe conductances, and then tries to provide a physical basis for the equations, since it claims that “there is little hope of calculating the time course of the sodium and potassium conductances from first principles.”4 The variable n is used as a dimensionless variable representing the proportion of particles inside versus outside the membrane. It varies between 0 and 1. In a like manner, m and h are variables used to represent proportions and range from 0 to 1 as well. The Hodgkin-Huxley variable ranges are given in Table 1. Table 1: Range for variables V, m, n, and h. Variable V (mV) m n h Range ~-110-110 0 to 1 0 to 1 0 to 1 3.2 MATHEMATICAL ANALYSIS The Hodgkin-Huxley model is mathematically described by the following differential equations: ( ) ( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( ) ( ) Using the model description above and differential calculus for partial differential equations, it thus follows that the Jacobian of the entire system is: 3 3.3 OBSERVER DESIGN In reality, not all these variables might be measurable with physical system. Assuming that only the membrane potential, V, can be measured, it is valuable to construct an observer in order to fill in for the state variables that cannot be measured but are important to describe the system. Replace the system with a partial state observer, copying the m, n, and h dynamics. Then, assume that the membrane potential is the only measured variable. This line of analysis yields the following results: ( ) ( ) ( ) ( ) ( ) Taking the derivatives of the observer dynamics with respect to the estimated variables yields: ( ) ( ) The diagonal Jacobian of the observer is thus: 4 ( ) [ ( ) ] ( ) It is not difficult to show the range of the diagonal terms for the domain of V. Plotting the (1,1) and (2,2) terms as functions of V reveal their ranges. The (1,1) and (2,2) terms are negative for all values of V, since the negative exponential value offsets the cases when V is less than -10 or -25. The (3,3) entry is obviously negative for any value of V, since the range of the exponential function (ex) is strictly positive. 3.4 COUPLING AND SYNCHRONIZATION ANALYSIS The next step is to study the possible coupling of multiple Hodgkin-Huxley neurons. ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( ) ( ) In the case of two neurons, there are eight state variables and equations, displayed as: ∑ ( ) ( ) ( ) ( ) ( ) ∑ ( ) ∑ ( ) ( ) ( ) ( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) This leads to a very complicated Jacobian of size 8 by 8. However, this is a very similar Jacobian to the original 4 by 4 Jacobian of the uncoupled system. The matrix will still have similar diagonal and first column and row terms as the form of the equations has not changed, except for the addition of a ( ) and a ( ) term. Of course, if 5 one were to make an observer for this expanded system, this coupling does not affect the reduced order observer, since these terms do not include m, n, and h. The (1,1), (1,2), (2,1), and (2,2) terms are thus: [ ] , , , and This matrix just concerning eigenvalues to be negative, the (1,1) and (2,2) diagonal terms must be larger in absolute value than the off diagonal terms, and yet more negative, through diagonal dominance. Since contains k21 and m, n, and h are always nonnegative ranging from 0 to 1, this will always be the case except for m, n, and h=0. The matrix will in this case reduce to: is symmetric. In order for both [ ] The eigenvalues will then be -2k21 and 0, meaning that there is a loss in stability for this portion of the Jacobian. Similar analyses regarding the nature of submatrices of the rest of the 8x8 Jacobian can be used to better characterize the synchronization properties and see whether or not V1 converges to V2. In particular, it is suggested that by studying the four 4x4 matrices which compose the Jacobian, and seeing if they can be rearranged in combination (e.g. hierarchical combination12), further analysis can be performed. 4. FITZHUGH-NAGUMO MODEL 2 4.1 BIOLOGICAL MOTIVATION The Hodgkin-Huxley model proposed in 1952 can be approximated to two state variables instead of four, presented in the Fitzhugh-Nagumo model analysis. The Fitzhugh- Nagumo model attempts to characterize the threshold phenomenon in neuronal firing and signaling behavior. Note that a special case of this biological neural oscillator model is a simple van der Pol oscillator. Biologically, v represents the membrane potential just as it does in the parent Hodgkin-Huxley model. The variable w represents a recovery variable for the system behavior, and Iext refers to the external current applied. 4.2 MATHEMATICAL ANALYSIS The model is described by the following set of two differential equations: ( ) ( It is easy to do a quick contraction analysis on this system after building an observer and using a transformation matrix to form a generalized Jacobian. ) ∑ ( ) 6 ( ( ) ⁄ ) 4.3 OBSERVER DESIGN Replace the system with a copy. ( ) ( ) ⁄ By using the transformation matrix and the generalized Jacobian arrive at a transformed Jacobian. ( ) ⁄ The symmetric part of the Generalized Jacobian is simply: ( ) ( ⁄ Constructing an observer, assuming that v is measured yields: ( ) ( ) ) ( ( ) ( ) ) we Feeding in the observer gain k gives: ) ( ) ( ⁄ For simplicity in mathematical analysis, take a reasonable choice for c, c=1. Then, it is clear that k must be greater than or equal to 1, since the maximum value for the (1,1) coordinate is 1, since x2 is nonnegative for all real x. Because v is real, 1-v2 will always be less than or equal to 1. Also, -b/c is always negative since both b and c are positive constants. Since the system is contracting, will converge to v. 4.4 OBSERVER RATE CONTROL APPROACH It is desirable to control both rows of the system in order to affect the rate of convergence for the observer. The first option to study is feeding back the v error to the second row of the symmetric part of the generalized Jacobian, in the form ( ). This yields: ( ) ( ⁄ ) Sylvester’s criterion (this is not a diagonal matrix) yields that it is symmetric negative definite because the 1x1 determinant is negative for proper choice in k, and the full determinant is negative is: ( ( ) ( ) Of course, choosing k2 appropriately yields a negative 2x2 determinant. The next step is to consider how to choose k1 and k2 to place both the eigenvalues of the system and thus control the rate of convergence of the observer, not just the fact that it does indeed converge. Unfortunately, the two eigenvalues are still the same as in the case without this feedback, so it is not valuable—k2 still does not play a role. 7 Another option is to choose other theta transformation matrices. This was tried for many cases but it made the generalized Jacobian very complicated and difficult to analyze, without yielding much physical gain. The generalized Jacobian loses its anti-symmetric properties for other choices of θ than what is presented above. 4.5 COUPLING AND SYNCHRONIZATION ANALYSIS It is then valuable to analyze the generalized Jacobian and dynamics for a coupled system. The uncoupled Jacobian for both neurons is given as: ) ( ⁄ ( ) ( ⁄ ⁄ ⁄ ) By using the transformation matrix and the generalized Jacobian arrive at a transformed Jacobian. we ( ) ) ( ( ) ⁄ ( ⁄ ) This transformed Jacobian is anti-symmetric or skew-symmetric, and the diagonal terms can be analyzed for contraction purposes. Now, by appropriately adding controller gains k1 and k2: ) ( ( ) ( appropriately chosen to be positive and greater than 1, the coupled system is contracting. (1-v2) cannot be greater than 1 because v is a real number so v2 is positive. If the coupling is added in: ( ) ∑ ( ) ⁄ ⁄ ) ( ) ∑ ( ) ( ) ( ) 8 ( ) ∑ ( ( ) ∑ ⁄ ⁄ ) This concludes a preliminary analysis of the Fitzhugh-Nagumo neural oscillator model from a nonlinear control systems perspective. 5. MORRIS-LECAR MODEL 10 5.1 BIOLOGICAL MOTIVATION In 1981 at the National Institutes of Health in Bethesda, Maryand, Catherine Morris and Harold Lecar developed a model related to the Hodgkin-Huxley model, but using barnacle muscle fibers subjected to stimulation and a different dynamical description. Their model, called the Morris-Lecar model, allows for IK and and ICa to dominate the potential dynamics. The model is conducted in three variables: V, which is the same voltage from the Hodgkin-Huxley model, m, which is the fraction of open Ca2+ models, and n, which is the fraction of open K+ channels. These variables are analogous to the m and n in the Hodgkin-Huxley model, which were ratios of these ions inside and outside the cell membrane. In all the experiments that this model is based off of (at least those presented in the paper first describing the model), the voltage ranges from -50mV to 50 mV. In addition, m and n are defined to (as ratios) range from 0 to 1. Various other physical parameters are defined in the paper. Table 2 gives the values of particular constants used in the Morris-Lecar model which are pertinent to this preliminary control systems analysis. Table 2: Values for constants in the Morris-Lecar model. Constant Value V1 V2 V3 V4 gk gCa C 10 mV 15 mV -1 mV 14.5 mV 0 to 20 0 to 20 20 Figure 1 depicts plots from the original Morris-Lecar paper showing how the variables gk and gCa vary with one another for different values of I. 9 Figure 1. Variation of gCa and gk with one another for varying values of I10. 5.2 MATHEMATICAL ANALYSIS The Morris-Lecar model can be described by the following set of three differential equations in three state variables: ( ) ( ) ( ( ( ( ( ( The Jacobian of this system is thus: ) ) [ ( ( )] ) ( ( ( ) [ ( ) ( ( ) [ ) ( ( ) ( ) ( ) ( ( ) )) ) )) ) ] ) )] An analysis of individual Jacobian matrix entries yields interesting results for observer design. The (2,1) and (3,1) entries can be treated as: [ ( ) ( ( ) ) ( ) ( )] ( ) It is useful to characterize the greatest value of these off-diagonal terms in order to study it from a matrix norm perspective of diagonal dominance to assess contraction properties of the system. Experimentally, from values given in the original paper, the range of the x parameter given above can be determined. Using these given values, it is shown that the value of the x parameter varies approximately from In addition, the m and n parameters, taken from the Hodgkin-Huxley model, are defined to vary from 0 to 1. Table 3 displays the values for the (2,1) and (3,1) terms for various x (column) and m (row) values. 10 Table 3: Values for the (2,1) and (3,1) terms, respectively, with maximum values given in bold. x m 0 -2 -1 0 1 2 x n -2 -1 0 1 2 0.06 0.24 0.50 0.24 0.06 0.00 0.06 0.24 0.50 0.24 0.06 0.50 0.06 0.24 0.50 0.24 0.06 0.50 0.06 0.24 0.50 0.24 0.06 1 0.05 0.24 0.50 0.24 0.05 1.00 0.05 0.24 0.50 0.24 0.05 The maximum values for the (2,1) and the (3,1) entries are given in bold. Recall that the diagonal dominance norm as it is used for contraction analysis purposes is given as ∑ ) ) The (1,2) and (1,3) terms are bounded by constants, and can easily be off-set by a gain added to the diagonal term. Finally, the (2,3) and (3,2) terms are both zero. Bounding the off-diagonal terms reduces the Jacobian analysis to simply: ( ( ] [ Finally, looking at the diagonals, it is easy to see that the (1,1) term can be 0 if m and n are 0. The (2,2) and (3,3) entries can be studied by looking at the ranges of the cosh functions. The range of cosh is always positive. and are both positive constants (1/15 and 1/10, respectively), and thus (2,2) and (3,3) are always negative. For the ranges of V for which this study and model is concerned with, (2,2) ranges from approximately -0.106 to -0.1045, and (3,3) ranges from the values -0.002267 to -0.00459. When choosing the gains, it is valuable to add a buffer for diagonal dominance. As is obvious, this diagonal dominance analysis for the full system is very complicated and also very conditional on the particular circumstances and current state values, which would need to be fed back to cancel out the varying terms. Thus, it is simpler to use a partial state observer when moving on to the next step of designing an observer for the system. 11 )) ) )) ) ( ( ( ( ( 5.3 PARTIAL STATE OBSERVER DESIGN It is useful to analyze a partial state observer, assuming once again that V (membrane potential) is the measured variable. Assuming that V is measured, the dynamics for the observer and estimated variables become: ( ( Then, the partial state observer Jacobian is simply: ( ) ] [ Again, the range of the ( )function for is always positive. This means that the eigenvalues of this diagonal matrix are always negative, because they are multiplied by and , where both are positive constants, and then multiplied by -1. In order to affect the rate of convergence, on the other hand, it is necessary to feed back some kind of gain. Because only V is assumed to be measured, the easiest way is to feed back ( ) 5.4 POSSIBLE APPROACH TO A FULL-STATE OBSERVER Of course, this is only possible through the use of a full state observer, since V must be estimated along with m and n to use this form of feedback. This can be used to control the second and third lines of the equation in order to guarantee convergence, based on the bounds shown above for the off-diagonal terms. Namely, once specific feedback is used, contraction can be guaranteed for the observer through diagonal dominance, and the observer has all the correct pieces as solutions (e.g. will converge to V). ( ) ( ) ( ) ) ( ) ( The next step is to vary the gains k1-3 in order to try to change the eigenvalues of the system in order to control the rate of convergence. 6. HINDMARSH-ROSE MODEL )) ) ( ) )) ) ( ) ( ( ( ( ( ( ) ( ) ( ) 3 6.1 BIOLOGICAL MOTIVATION In 1983, J.L. Hindmarsh and R. M. Rose developed a model composed of three coupled differential equations to describe neural activity. The work was motivated by the discovery of a cell in the pond snail Lymnaea which generated a burst after being depolarized by a short current pulse. This model describes the phenomenon of bursting in two-dimensional space. However, the model also incorporates a third dimension, which is 12 a slow current that hyperpolarizes the cell. The variable x here refers to the membrane potential of the cell, while y and z represent the transport of ions across the cell membrane. The variable y represents the fast transport of sodium and potassium, while z represents the slower transport of other ions and is correlated to the bursting phenomenon. 6.2 MATHEMATICAL ANALYSIS The Hindmarsh-Rose model is described by the following set of differential equations with three state variables: [ ( ) ] The constant parameters are chosen to be at s=4, a=1, b=3, c=1, d=5, and r=10-3. This simplifies the Jacobian to the following form: ⁄ ⁄ After trying various methods (e.g. diagonal dominance, negative definiteness, anti- symmetric negative diagonal) to assess the contraction properties of this system, the best (and simplest) method was determined to be the anti-symmetric with negative diagonal. Start with θ as the identity matrix. By varying the diagonal terms in θ, it is possible to affect the off-diagonal terms in the generalized Jacobian, . By testing out various values of the transformation matrix, and seeing how they affect the generalized Jacobian, the transformation matrix can then be chosen appropriately. For example, from simple matrix algebra: [ ] [ ] [ ⁄ ⁄ ] [ ] [ ⁄ ⁄ ] [ ] 13 [ ⁄ ⁄ ] By using a linear combination of the second two choices of transformation matrix, the simple choice of θ is: This yields a generalized Jacobian of: [ ⁄ √ ] [ √ √ ⁄ ⁄ ⁄ ] This matrix is anti-symmetric, meaning that M=-MT. Now, all that remains is to ensure that the diagonals are negative for all values of x, y, and z. It is obvious that the significant choice is the variable (1,1) term, which depends only on x. The domain of concern is x=[0,2], since in this range of x values, the (1,1) term of the generalized Jacobian is positive. To ensure that it is always negative, choose appropriate k gain values, namely, . [ ] Here, the gain k is arbitrarily chosen as -10, and as long as it is more negative than -3, can be varied depending on control power available to set this gain in the system. Thus, for all x, y, and z, this system as defined above is contracting. 6.3 FULL STATE OBSERVER DESIGN AND ANALYSIS It is now possible to create an observer for this system. Assume that only the membrane potential (the x variable) can be measured, and it is necessary to reconstruct the state through the use of this measured variable. Recall the form: ( ) ( ) Here, refers to the measured variable x. For this particular system, only x the membrane potential is assumed to be measured, while y and z are recreated through the use of a full-state observer. ( ) ( ) ( ) [ ( ) ] ( ) ( ) 14 Then, the generalized Jacobian (of the form aforementioned gain for k1, becomes ) incorporating the observer and using the √ [ √ ⁄ ⁄ ⁄ ] Thus, for all x, y, and z, this observer as defined above is contracting. Because this is a full state observer, feedback using the estimate error for x can be used for all three lines of the observer, yielding the result of having control over the rate of convergence for the system, not just whether it converges. 6.4 SIMULATION A simulation of the Hindmarsh Rose model in the following form yields the graph in Figure 2. The terms used are I=4, xr=0, a=1, b=3, c=1, d=5, r=0.001, and s=4. The form of the model used is reproduced here for convenience. The difference between the model above and the one here is that I is not 0 (current applied). It does not affect the Jacobian and analyses above since it is a constant. ( ( ) ) [ ] Figure 2a depicts the neuron bursting phenomenon derived from the Hindmarsh-Rose model definition and its simulation. Figure 2b plots two state variables against one another. 15 Figure 2a: MATLAB simulation of Hindmarsh-Rose model. Figure 2b: MATLAB plot of x and y variables against each other. Figure 2b displays a limit cycle in the behavior of the system. Compare this simulation’s results to a similar set of plots from the original Hindmarsh-Rose paper, to see the parallels, shown in Figure 3a and 3b. 16 Figure 3a: Figure from original paper showing neuron spikes3. Figure 3b: Figure from original paper showing x and y coordinate plotted against one another. 7. Conclusions and Further Work In conclusion, the four models analyzed using controls systems perspective provide some insights into the function of the various components in neuronal networks and their roles in signaling. The four-state Hodgkin-Huxley model can be simplified to smaller models such as the two-state Fitzhugh-Nagumo model to be studied in greater detail. Further work should address the possibility of a full-state observer for the Hodgkin-Huxley model, and also provide a better analysis of rate control for the Fitzhugh-Nagumo model. In addition, the possibility of a full-state observer for the Morris-Lecar model is open as well. This work provides a preliminary analysis and description of these four neural oscillator models from a nonlinear control systems approach, which can be expanded further to include improved detail, accuracy, and variety of methods and analysis techniques. 8. Acknowledgements and References Many thanks to Professor Jean-Jacques Slotine at MIT, the advisor for this entire project who provided invaluable instruction and help to learn this material on nonlinear controls and neurobiology. 17 REFERENCES 1 Alon, Uri. An Introduction to Systems Biology: Design Principles of Biological Circuits. Boca Raton, FL: Chapman & Hall/CRC, 2007. Print. 2 Fitzhugh, Richard. "Mathematical Models of Threshold Phenomena in the Nerve Membrane." Bulletin of Mathematical Biophysics 17 (1955): 257-78. 3 Hindmarsh, J. L., and R. M. Rose. "A Model of Neuronal Bursting Using Three Coupled First Order Differential Equations." Proceedings of the Royal Society B: Biological Sciences 221.1222 (1984): 87-102. 4 Hodgkin, A., and A. Huxley. "A Quantitative Description of Membrane Current and Its Application to Conduction and Excitation in Nerve." J. Physiology 117 (1952): 500-544. 5 Hodgkin, A., and A. Huxley. "The dual effect of membrane potential on sodium conduction." J. Physiology 116 (1952): 483-296. 6 Hodgkin, A., and A. Huxley. "The components of membrane conductance in the giant axon of loligo." J. Physiology 116 (1952): 483-296. 7 Hodgkin, A., and A. Huxley. "Currents carried by sodium and potassium ions through the membrane of the giant axon of loligo." J. Physiology 116 (1952): 497-506. 8 Hodgkin, A., and A. Huxley. "Measurement of current-voltage relations in the membrane of the giant axon of loligo." J. Physiology 116 (1952): 424-448. 9 Horn, Roger A., and Charles R. Johnson. Matrix Analysis. New York: Cambridge UP, 2013. Print. 10 Morris, C. "Voltage Oscillations in the Barnacle Giant Muscle Fiber." Biophysical Journal 35.1 (1981): 193-213. 11 Purves, Dale. Voltage-Gated Ion Channels. U.S. National Library of Medicine 12 Slotine, Jean-Jacques, and Winfried Lohmiller. "On Contraction Analysis for Nonlinear Systems." Automatica 34.6 (1998): 683-96. 13 Slotine, Jean-Jacques, and Wei Wang. "A Study of Synchronization and Group Cooperation Using Partial Contraction Theory." 14 Slotine, Jean-Jacques, and Giovanni Russo. "Global Convergence of Quorum-sensing Networks."Physical Review E 82.4 (2010). 15 Slotine, Jean-Jacques. "2.14: Analysis and Design of Feedback Control Systems." Ed. Sze Zheng Yong. 18 16 "The Mind of a Worm: The Nervous System." Worm Atlas. Web. 17 Wang, Wei, and Jean-Jacques E. Slotine. "On Partial Contraction Analysis for Coupled Nonlinear Oscillators." Biological Cybernetics 92.1 (2005): 38-53. Print. 19
1912.10489
1
1912
2019-12-22T17:40:19
Recurrent Feedback Improves Feedforward Representations in Deep Neural Networks
[ "q-bio.NC", "cs.LG" ]
The abundant recurrent horizontal and feedback connections in the primate visual cortex are thought to play an important role in bringing global and semantic contextual information to early visual areas during perceptual inference, helping to resolve local ambiguity and fill in missing details. In this study, we find that introducing feedback loops and horizontal recurrent connections to a deep convolution neural network (VGG16) allows the network to become more robust against noise and occlusion during inference, even in the initial feedforward pass. This suggests that recurrent feedback and contextual modulation transform the feedforward representations of the network in a meaningful and interesting way. We study the population codes of neurons in the network, before and after learning with feedback, and find that learning with feedback yielded an increase in discriminability (measured by d-prime) between the different object classes in the population codes of the neurons in the feedforward path, even at the earliest layer that receives feedback. We find that recurrent feedback, by injecting top-down semantic meaning to the population activities, helps the network learn better feedforward paths to robustly map noisy image patches to the latent representations corresponding to important visual concepts of each object class, resulting in greater robustness of the network against noises and occlusion as well as better fine-grained recognition.
q-bio.NC
q-bio
RECURRENT FEEDBACK IMPROVES FEEDFORWARD REPRESENTATIONS IN DEEP NEURAL NETWORKS Siming Yan Computer Science Department Peking University [email protected] Bowen Xiao Computer Science Department Peking University [email protected] Yimeng Zhang Computer Science Department and Neuroscience Institute Carnegie Mellon University [email protected] Xuyang Fang Computer Science Department Carnegie Mellon University [email protected] Harold Rockwell Computer Science Department Carnegie Mellon University [email protected] Tai Sing Lee Computer Science Department and Neuroscience Institute Carnegie Mellon University [email protected] December 24, 2019 ABSTRACT The abundant recurrent horizontal and feedback connections in the primate visual cortex are thought to play an important role in bringing global and semantic contextual information to early visual areas during perceptual inference, helping to resolve local ambiguity and fill in missing details. In this study, we find that introducing feedback loops and horizontal recurrent connections to a deep convolution neural network (VGG16) allows the network to become more robust against noise and occlusion during inference, even in the initial feedforward pass. This suggests that recurrent feedback and contextual modulation transform the feedforward representations of the network in a meaningful and interesting way. We study the population codes of neurons in the network, before and after learning with feedback, and find that learning with feedback yielded an increase in discriminability (measured by d-prime) between the different object classes in the population codes of the neurons in the feedforward path, even at the earliest layer that receives feedback. We find that recurrent feedback, by injecting top-down semantic meaning to the population activities, helps the network learn better feedforward paths to robustly map noisy image patches to the latent representations corresponding to important visual concepts of each object class, resulting in greater robustness of the network against noises and occlusion as well as better fine-grained recognition. 1 Introduction The primate visual system is organized as a hierarchy of many visual areas with massive recurrent connections, both within each area and between different areas [1]. These recurrent connections are thought to encode statistical priors of natural scenes, as well as to bring in higher order semantic and contextual information to help resolve local ambiguity, filling in missing details in lower visual areas during inference. There has been considerable recent interest in exploring and exploiting the use of recurrent feedback in deep convolutional neural networks [2, 3, 4]. This work shows promising results suggesting that recurrent connections allow networks achieve comparable performance in image classification with less parameters and fewer layers than a deep feedforward network. A PREPRINT - DECEMBER 24, 2019 Figure 1: The schematic of a VGG Contextual Modulation(VGG-CM) model. (a) 3 loops between adjacent stages of VGG16. (b) microcircuit of recurrent interaction inside one loop. See Section 3.1, 3.2 for details. These earlier studies focused on the contribution of recurrent connections for inference. Little is known about the impact of recurrent connections on refining the feedforward computation. In this paper, we provide evidence that the recurrent feedback of semantic and contextual information might highlight important visual concepts (clusters of latent representations, [5]) at each level during training, steering the feedforward circuits towards these beacons, resulting in a feedforward network that is more robust against noise, occlusion and adversarial attacks, and performs better in fine-grained recognition. 2 Related Work Using global contextual information to disambiguate local early processing is a classic idea in computer vision [6, 7]. In this early work, recurrent connections were used to encode hand-crafted, generic statistical priors of natural scenes, such as smoothness and uniqueness constraints. This work was concerned primarily with early vision tasks, such as contour completion [8] and image reconstruction [9, 10]. Recently, there has been considerable interest in the computer vision community in exploring the functional advantages of top-down feedback and local recurrent connections in deep neural networks for object recognition [2, 4, 3, 11]. [2, 4] used feedback essentially to introduce top-down information from the fully semantic layers, enhancing relevant features in early layers in an attention-like mechanism. The large semantic gap between the semantic layer and early feature layers make feature alignment difficult, and such methods work mostly for low resolution CIFAR datasets [4]. Other work has explored the use of local recurrence [3, 11] and/or long range loops [3], and was successful in demonstrating that local recurrent connections can achieve ImageNet object recognition performance that is comparable to a very deep Residual Network, but with much fewer layers and parameters. A striking feature of the primate visual system is that the most massive recurrent feedback connections are between adjacent visual areas, and feedback can propagate all the way down to LGN. This ubiquitous feature has not been utilized in the above studies. The recurrent feedback in predictive coding network models [11] updates the activities of neurons to reconstruct their input and is analogous to local recurrence within a visual area, rather than between adjacent visual areas. In this paper, we explore networks with multiple loops between adjacent stages of a deep convolution network, which is more consistent with the anatomical literature. We study how networks with local horizontal and top-down feedback can bring in contextual information to help resolve local ambiguity in challenging scenarios. Our first contribution is in showing that VGG16, fine-tuned under recurrent feedback, becomes more robust against noises, adversarial "attacks," and occlusion, and also performs better in fine-grained object recognition. Most interestingly, the lion's share of this improvement is due to the refinement in feedforward connections, not the effects of feedback during 2 A PREPRINT - DECEMBER 24, 2019 inference. Our second contribution is in characterizing the changes in latent representations of the networks underlying the improvement in performance. We are able to show that there is an increase in semantic clustering in the population activities subsequent to learning under feedback, and that feedback might have steered the network to learn more robust feedforward information processing paths toward important clusters of latent representations, resulting in improvement in performance in challenging situations. 3 Methods Our first objective is to characterize how contextual modulation created by local recurrent information from the current visual area and top-down feedback from the adjacent higher visual area can be useful for improving network performance in object recognition. Anatomical and neurophysiological evidence suggest that the surround contextual modulation a neuron experiences can be divided into a local component that is mediated by local circuits, and a global component that is mediated by top-down feedback [12]. The near-surround includes surround with spatial extent that is 2-3 times the size of the receptive field, while the far-surround includes a space that is 3-6 times the size of the receptive field. 3.1 Feedback loops -- contextual modulation from far surround We approximate the far-surround feedback architecture by introducing feedback loops to adjacent stages in a standard VGG16 network [13]. We consider each stage in VGG, which contains multiple convolutional layers followed by a pooling layer as output, to be roughly equivalent to a visual area. Thus, we model the feedback loop between adjacent visual areas by adding feedback loops between the last convolution layers before the pooling layer in each stage. So, the three feedback loops added are from conv3.3 to conv2.2, from conv4.3 to conv3.3 and from conv5.3 to conv4.3 respectively (as shown in Figure 1a). Figure 1b illustrates the details of feedback loop 2 from conv4.3 (stage 4) to conv3.3 (stage 3). The feedback generates a tensor FB(3.3) with the same spatial resolution and feature dimensions as its target layer (i.e. conv 3.3) by performing an expansion deconvolution using a 3 X 3 convolution filter on conv 4.3. Since there are three 3 x 3 convolutions and a pooling layer between conv 3.3 and conv 4.3, each unit in the FB(3.3) layer has effectively integrated spatially global information from over 19 x 19 columns in conv 3.3, thus containing information coarser in spatial resolution but with higher order semantics, corresponding to the far-surround effect in neurophysiology [12]. The feedforward R and the feedback FB representations are specified by the following equations Rt FF t FB t s(k) + CM s(k) = FF t s(k) = Conv(Conv(Pooling(Rt s(k) = Sigmoid(Deconv(Rt s(k−1)))) s(k+1))) (1) (2) (3) where Rs(k) are the unit responses at time t of a particular VGG16 layer in the visual stage k, e.g. s(2) stands for VGG16 conv 2.2, s(3) for VGG16 conv 3.3, s(4) for VGG16 conv 4.3, s(5) for VGG16 conv 5.3. These are the layers that provide as well as receive "inter-areal" feedback modulation. Rt s(k) is the sum of the feedforward input signal FF s(k) and a contextual modulation signal CM to be described below. FB t s(k) is a tensor that contains the feedback information from the next stage Rs(k+1), at the same spatial and feature dimension as Rt s(k) and is derived from expansion (up-sampling), followed by a 3 x 3 convolution. 3.2 Lateral connections - contextual modulation from near surround While top-down feedback provides semantic information of a more global nature, more detailed fine-grained feature constraints and priors are better encoded by local lateral or horizontal connections. For example, in V1, local connectivity has been shown to encode a contour continuation constraint [8, 15, 16] and surface smoothness constraint [17, 18], providing a facilitatory associative field for contour and surface completion. Inhibitory local connectivity can also implement a uniqueness constraint, mediating competitive surround suppression for redundancy reduction or sparse coding. Here, rather than specifying these local priors explicitly by hand, as in traditional Markov random field models [7] or sparse coding models [19, 20], we used a convolutional layer (H(3, 3)) to learn and model the local contextual priors implicitly using 3 x 3 convolution on the feedforward representational layer R(3.3) as specified as follows. H t s(k) = Conv(Rt s(k)) 3 (4) A PREPRINT - DECEMBER 24, 2019 Figure 2: Examples of images used in three different challenging scenarios. Each pair contains the original image (left) and the corrupted image (right). (a): Corruption with 50% Gaussian noise. (b) Corruption with 30% adversarial attack noise. (c) Corruption with occlusion as in the VehicleOcclusion dataset [14]. 3.3 Contextual Modulation and Information Flow The feedback tensor F Bs(k) and the horizontal context tensor Hs(k) are concatenated and then subject to 3 x 3 convolution to derive the contextual modulation tensor CONTEXT s(k) that is of the same dimension as as Rs(k), which it modulates by point-wise addition. The convolution is passed through a tanh function to generate an output that can be positive or negative, as the contextual modulation effect can be inhibitory or excitatory, mediated by inhibitory and excitatory interneurons. CONTEXT t Rt s(k) = Tanh(Conv(Concat(H t s(k) + CONTEXT t−D s(k) = FF t s(k) s(k), FB t s(k)))) (5) (6) where D indicates a delay. For simplicity, we used a delay of D = 1. Given an input image, the network computes a first pass as a feedforward DCNN (unroll 0). At time t = 0, there is no contextual modulation, i.e. CM = 0. For simplicity, we assumed there is no delay in computation through the entire feedforward hierarchy. The context modulation tensors (CONTEXT (2.2), CONTEXT (3.3),CONTEXT (4.3)) are then computed based on feedback and horizontal integration and then added to the bottom-up activities of corresponding layers in the second pass (unroll 1), and continuing through the subsequent passes (unrolls). Because of the loops, the bottom-up feedforward input FF s(k) activities of every layer will continue to be modified, and the activities of the higher layer will propagate down to modulate the activities in the lower layers, all the way down to conv2.2 in our current implementation. 3.4 Training and Testing We train and test the networks on four benchmark datasets: CIFAR-10 (10 classes, low resolution 32 x 32 pixels images), CIFAR-100 (100 classes, low resolution 64 x 64 pixels images), ImageNet (1000 classes, high-resolution 224 ×224 pixels images) and CUB-200 (for fine-grained recognition). Our model is implemented in PyTorch. Training on the CIFAR and CUB-200 datasets is completed in 4 hours by 1 Nvidia GeForce GTX GPU, and on ImageNet in 2 days by 4 such GPUs. We then test the 3-loop network trained on ImageNet in three challenging scenarios: (1) Gaussian noise up to 50% (meaning that the half-height bandwidth of the Gaussian is 50% of the total range of image pixel values) added to the ImageNet images in the validation set; (2) Adversarial attack noise up to 30% added to the ImageNet images; (3) Occlusion as in the VehicleOcclusion dataset [14]. This occlusion dataset contains 4549 training images and 4507 testing images covering six types of vehicles, e.g., airplane, bicycle, bus, car, motorbike and train. For each test image in this dataset, some randomly-positioned, irrelevant occluders were placed onto the target object. Example images of each scenario are shown in Figure 2. In training the network, the hyperparameters were selected by grid search. For nonlinearity imposed after the convolution operation, we used ReLU for both integrating the feedback and the horizontal contextual information, as empirically we found them to work better than tanh and sigmoid. We used tanh after the convolution on their concatenation C to generate the contextual modulation (CONTEXT) that can exert either positive (add) or negative (subtract) modulation to the feed-forward path. Contextual modulation potentially can exert gain modulation (i.e. multiplication) on the feedforward path, rather than addition or subtraction. However, empirically we found that multiplication in this particular architecture actually produces inferior performance. We did not tried division, which might potentially model the divisive normalization mechanism in the cortex. 4 A PREPRINT - DECEMBER 24, 2019 Models VGG16 VGG-ATT VGG-LR-2 VGG-CM-0 VGG-CM-4 Experiments CIFAR-10 CIFAR-100 CUB-200 ImageNet Gaussian Noise 50% Adversarial Noise 30% Occlusion 91.20* 91.77 91.49 92.10 91.52 67.06* 69.48 68.99 71.51 71.43 64.880* 73.200 72.990 74.824 74.910 71.076 71.213 71.551 71.645 71.741 12.983 13.002 14.041 17.970 18.202 42.541 42.759 43.623 47.249 47.924 34.500 37.619 38.721 50.012 50.720 Table 1: Top-1 image classification accuracy on different experiments. Notice that in the CIFAR and CUB-200 experiments, VGG16 models have only one FC layer. VGG-ATT is the model proposed in [4], VGG-LR-n is the "rethinking" one-FC-layer VGG model with 2 unrolling times proposed in [2] and tested with n unrolling times. VGG-CM-n is our 3-loop model trained with 4 unrolling times and tested with n unrolling times. Noise Level 0 10 20 30 40 50 Models VGG16 Unroll 0 Unroll 1 Unroll 2 Unroll 3 Unroll 4 71.076 65.366 53.988 39.045 23.971 12.983 71.645 67.493 56.711 43.331 28.581 17.970 71.655 67.576 56.776 43.376 28.919 18.052 71.681 67.591 56.843 43.438 28.981 18.071 71.702 67.603 56.854 43.441 29.012 18.120 71.741 67.620 56.988 43.686 29.120 18.202 Table 2: Noisy image classification top-1 accuracy for different unrolling times of our proposed model. VGG16 is the standard feedforward VGG16 model and Unroll n indicates n unroll times during the test process. 4 Experimental Results 4.1 Performance improvement in feedforward computation in noisy and occluded situations Table 1 provides a summary comparison between our VGG-CM with three loops against VGG16, as well as two other VGG-based models with recurrent feedback (VGG-ATT [4] ; VGG-LR-2 [2]) in their performance on four benchmark datasets (CIFAR10, CIFAR100, CUB200, and ImageNet), as well as three challenging test sets: ImageNet images corrupted with Gaussian noise and adversarial noise, as well as occlusion. Our VGG-CM model's performance is comparable to VGG16 in CIFAR10 and ImageNet, but exhibits a significant improvement relative to VGG16 in fine-grained object recognition (15% improvement), in noisy (30% improvement for high level of Gaussian noises, and 13% for adversary noise) and occlusion situations (47% for VehicleOcclusion dataset). Our VGG-CM (with 3 loops and 4 unrolls) also outperformed VGG-ATT and VGG-LR in these challenging noisy and occluded test sets. The adversarial noise was generated by the standard Fast Gradient Signed Method (FGSM) attack. Table 2 shows the performance of VGG-CM at the different unrolls for different levels of Gaussian noise. Note that VGG-CM-0 is the VGG-CM at unroll 0, meaning that its performance is based purely on the first pass of the feed-forward computation, without any benefit of recurrent feedback contextual modulation. It is equivalent to removing all the loops and contextual modulation in our network in the testing stage, reducing to a network with exactly the same number of parameters as the original VGG16. Interestingly, VGG-CM-0's performance is very close to VGG-CM-4 in the four challenging tasks (fine-grained recognition, Gaussian noise, adversarial noise and occlusion), revealing that the majority of the performance improvement in challenging situations happens in the feedforward computation, with only small incremental benefits from the recurrent iteration during inference. More specifically, at 40% noise level, unroll 0 achieves 89% of the improvement attained by unroll 4; and at 50% noise level, unroll 0 already achieves 96% of the improvement attained by unroll 4. Thus, it seems that recurrent feedback and contextual modulation is very important during training, significantly modifying the feedforward connections of the system and accounting for the lion's share of the improvement, but actually plays a lesser role in our setup during inference, contributing only small incremental improvements. Given the network showed most of the performance improvement in the feedforward connections and only minor improvement with additional recurrent iterations, i.e. unroll 4 is not significant better than unroll 0 or unroll 1, one wonder how many unrolls are necessary during training. Table 3 shows the test performance of the full network trained with different number of unrolls. The data suggests maybe one recurrent iteration during training is sufficient to produce most of the performance benefits. 5 A PREPRINT - DECEMBER 24, 2019 Noise Level 0 10 20 30 40 50 Models VGG16 Unroll 1 Unroll 2 Unroll 3 Unroll 4 71.076 65.366 53.988 39.045 23.971 12.983 71.132 66.468 55.980 42.392 28.110 17.941 71.313 66.484 55.938 42.516 28.544 17.954 71.316 67.481 56.894 43.551 29.031 17.982 71.741 67.620 56.988 43.686 29.120 18.202 Table 3: Object recognition accuracy at different noise levels for our full networks that are trained with different unrolls. Unroll 1 here means the network only has one iteration (unroll) during training. Figure 3: (a) Average d-prime between all possible pairs of object classes for VGG-CM at unroll 0 as well as unroll 4 relative to VGG16. (b) Percentage changes in average d-prime relative to that of VGG for VGG-CM-0 and VGG-CM-4. It should be noted that since VGG-CM-0, our network on the first pass, during inference or testing, has essentially the same number of parameters as VGG16, hence the improvement in performance is not simply due to the network being effectively deeper or having more parameters. The improvement in the robustness of the network thus primarily involves the refinement of the feedforward representations. As a side note, we found that having three loops is better than two loops and having two loops is better than one loop, even for unroll 0. 4.2 Increase in semantic and categorical discriminability of the feedforward latent representations What has happened to the feedforward latent representations that lead to the increase in the robustness of the network in these challenging situations? Our conjecture is that the cascade of feedback from global and semantic representations of the higher layers has steered the representations at each level to become more semantically distinct. To evaluate this conjecture, we compute the average d-prime (or sensitivity index; distance divided by standard deviation) between all possible pairs of the 1000 object classes at each layer of latent representation along the feedforward hierarchy of VGG16 and VGG-CM-0. Figures 3a and 3b show the average d-prime as well as the change in average d-prime between object classes for VGG-CM-0 and VGG-CM-4 relative to VGG16 at some selected layers of the network. They show that starting at conv2.2, the first target layer that receives the feedback, VGG-CM's d-primes have substantially increased relative to that of the VGG16, supporting the idea that feedback has made the early representations more semantically distinct at every layer. Like the robustness to noise, most of this improvement is in VGG-CM-0, showing that while this is learned with the help of feedback, the improvement itself no longer depends on feedback. 4.3 Impact of feedback and horizontal modulation: ablation studies We performed a series of ablation experiments to find out whether top-down feedback alone, or horizontal recurrence alone is sufficient to generate the performance improvement. We found that combining both top-down feedback and horizontal connections yielded the best result in robustness against noise. An intuitive rationale for this is that top-down feedback introduces spatially more global surround but coarse contextual modulation, while horizontal recurrent connections bring in more precise local interaction, implementing more specific and precise local priors. Table 3 compares the performance of four models (standard VGG16, VGG-CM without feedback (no FB), VGG-CM without horizontal modulation (no H), VGG-CM full model) in different noise levels. For each version of the VGG-CM, we also compare the model's performance at unroll 0 (feedforward only) and at 4 unrolls (after 4 iterations of recurrent modulation). All the VGG-CM network models were trained with 4 unrolls. The results show that horizontal modulation 6 A PREPRINT - DECEMBER 24, 2019 Noise Level 0 10 20 30 40 50 Models VGG16 No FB U0 No FB U4 No H U0 No H U4 Full U0 Full U4 71.076 65.366 53.988 39.045 23.971 12.983 71.428 65.680 54.808 40.368 25.980 15.106 71.244 66.098 55.440 41.104 27.136 16.012 71.546 65.768 53.110 36.584 21.494 11.294 71.410 66.556 56.188 41.914 27.522 16.094 71.645 65.366 53.988 43.331 28.581 17.970 71.741 67.620 56.988 43.686 29.120 18.202 Table 4: Object recognition accuracy at different noise levels, comparing four models -- standard VGG16, horizontal modulation only (No FB), top-down modulation only (No H), and both top-down and horizontal contextual modulation (Full). The models are compared at unroll 0 (U0, the feedforward case) and unroll 4 (U4, fully unrolled). alone can bring about some improvement in the feedforward circuit, achieving 17% of the 38% increase achievable by the full model at high noise level. On the other hand, top-down modulation alone somehow cannot improve the feedforward circuit, though its' contribution is important for the network to realize its full potential. It is possible that the top-down feedback is needed to gate the detailed computation by the horizontal connections that encode precise geometric and spatial priors as in the high-resolution buffer theory [21], predictive encoding model [22], the gated Boltzmann machine [9], or gated Markov random field [10]. 4.4 Representational changes along the feedforward path? What might have happened to the feature representations to bring about this untangling of object representations along the hierarchy, even at the earlier layers? Given that the number of feature detectors at each layer is rather limited in the VGG, the changes in the receptive fields or feature tuning of the individual neurons are not obvious. We also looked for an increase in tuning selectivity, but found the changes in the population sparsity and life-time sparsity (tuning sharpness) of the neuronal populations to be negligible. The increase in semantic discriminability in the latent representations can arise from a increase in the distance in the population code or neural representational space between the concepts or the reduction of the variability of the samples belong to a concept. Next, we study the changes in the population codes by examining the changes in the visual concepts they encode. Visual concepts are clusters of population activities associated with each class of visual objects that correspond to parts of objects at different levels of scales and abstraction. For example, the face category will have visual concepts that correspond to the eyes and lips in lower layers, and correspond to the a portion of the faces in the higher layers (see Figures 4a-c). Yuille and colleagues [5] have shown that k-means clusters of hypercolumn population activities for each class of objects, with suitable pruning, give rise to semantically meaningful subparts and parts of objects, and that these parts can be used in a voting scheme to detect the existence of semantic parts, allowing the system to recognize an object based on its parts, under severe occlusion. For example, seeing an eye and a nose would be sufficient to recognize it is the face of a human. They call these clusters "visual concepts". Figure 4a-c shows the clusters of visual concepts of face category (each row is a cluster, contains a number of examples) at three different layers for illustration. Given recurrent feedback has made our VGG-CM more robust against noises and occlusion, we speculate that the feedforward path might have learned better visual concept representations. We applied the method described in [5] to extract visual concepts from the max pooling layers of the third stage of the VGG network, for 99 of the 1000 ImageNet categories. This is accomplished by collecting the hypercolumn population response vectors across space and across images, and then perform K-means++ to cluster these population code, starting with a number of clusters matching the number of channels, which are then condensed via a greedy pruning method to yield the visual concepts. To assess whether there is any improvement in the visual concepts, we first measure the importance of each visual concept by the relative drop in object recognition performance the network experiences for that class of objects when that visual concept is removed from the intermediate representation at that layer. This is accomplished to air-brushing away the image patch in the original image associated with that visual. We use the following score S(V k m) to measure the importance of a visual concept V k m for category m, (cid:80) S(V k m) = k∈Mk,Poriginal(I i I i M k)>0.3( Poriginal(I i Poriginal(I i k) k)−Poccluded(I i k) ) (7) where Mk ⊂ M are the images containing visual concept V k m in the classification layer of the neural network, and Poccluded(I i concept V k m is air-brushed away. The threshold of 0.3 for Poriginal(I i m, Poriginal(I i k) is image I i k) is image I i k 's output probability for class k 's output probability after the visual k) is chosen to ensure that the difference is only 7 A PREPRINT - DECEMBER 24, 2019 Figure 4: Some examples of visual concepts for faces at different stages, and plots of their within-class frequency and between-class uniqueness against their importance. measured for images that are reasonably well classified by the network in the first place. Specifically: (cid:80) Poccluded(I i k) = Px,y∈P i k,l F (f air−brushed P i k,l l,x,y ) l,x,y ∈ RWl×Hl×Nl is generated by where F (x) is the neural network's computation starting at layer x, and f air−brushed air-brushing away a visual concept feature vector px,y ∈ RNl from Ik,i's intermediate layer response fl. P i k,l is the set m. The airbrushing process point-multiplies the 5 × 5 spatial of population responses in fl belonging to visual concept V k area around the visual concept with 1 − G(0, 1), where G(0, 1) is a 2D Gaussian template with standard deviation 1, normalized to equal 1 in the center. In this way, the visual concept itself is entirely erased, and the nearby points that share its corresponding image patch are attenuated. The modified representation is then allowed to propagate through the remaining layers of the deep network to compute the probability of the different targeted classes. Note that the concepts are removed one at a time, so those that appear multiple times in an image (e.g. eyes) may have reduced importance by this measure. Figure 5 shows the ranked importance score thus computed of the 100 concept percentiles represented by the population codes of pool 3 layer for the VGG16 network and the VGG-CM-0 network. The importance curve for 99 randomly selected classes of objects is rather sharp, indicating that a few visual concepts are significantly more important than the others for each class. What makes a visual concept important for object recognition? Intuitively, concepts (and their corresponding semantic parts) that appear frequently within one image class, but rarely within others, should be more useful for discriminating between those classes. We find this to be the case: within a class, the occurrence frequency of a concept is positively correlated with its importance (Figure 4d), and the number of similar concepts in different classes is negatively correlated with importance (Figure 4e). However, the ranked importance score curves of VGG16 and VGG-CM-0, as shown in Figure 5, are not different. This suggests that fine-tuning with recurrent feedback has not changed the visual concepts embedded in the population codes or, equivalently, that the robustness against noise and occlusion is not due to the visual concepts becoming different, better or more important. We next analyze what happened to the representations of these population codes when the images corresponding to the visual concept are corrupted with 50% Gaussian noise. For each of the 99 chosen classes, we corrupt the images, pass them through the two networks, and calculate how many of the patches were still mapped into the same visual concept. We find that VGG-CM-0 performs better by this measure, retaining more of the noisy patches (Figure 5). Since the visual concepts have not fundamentally changed, mapping a noisy patch to the same population code as an unaffected patch should help the downstream neurons classify the concepts or object parts. We also compute the average distance of all the representations of the noisy samples relative to their original concepts' center, and found that indeed it is lower for the VGG-GM-0, meaning that the noisy population codes are not just more likely to be closest to, but are on average less deviated from their original visual concepts (Figure 5). Note that we only showed results from pool 3 concepts because these concepts tend to be semantically more meaningful than lower layer concepts and can be more reliably estimated than higher layer concepts [14]. These observations suggest that the recurrent feedback during training might have allowed higher level semantic and contextual information to highlight and select important visual concepts at each level, using these "beacons" to help the network to learn better paths to project more reliably to the same correct representations along the hierarchy. 8 A PREPRINT - DECEMBER 24, 2019 Figure 5: Plots of visual concept performance in VGG16 (red) and VGG-GCM-0 (blue), averaged over all 99 classes, ranked by percentile of the concepts. (left): The importance of the visual concepts, as calculated by Equation (9). (center): The probability (or frequency) of reassignment of a patch to its original concept after the addition of 50% Gaussian noise. (right): The average distance of feature responses from the cluster centers after the addition of 50% Gaussian noise. 5 Conclusion In this paper, we investigate a biologically-inspired neural architecture, with feedback loops between adjacent stages of processing as well as horizontal contextual modulation in a deep convolutional neural network. We find that fine-tuning such a network with recurrent feedback and contextual modulation allows it to become more robust against noise and occlusion, as well as perform better at fine-grained discrimination. We develop an approach to evaluate the latent representations of deep neural networks based on the notion of visual concepts. We find that while the visual concepts in the feedforward latent representations have not been changed by feedback, the domain of images mapped to these visual concepts have been increased, allowing corrupted versions of the image patches to be mapped to the representation of the semantically meaningful visual concepts. Our finding suggests that higher order semantic feedback might have highlighted the visual concepts at the lower level during training, providing beacons to help the network find its way to the appropriate target concepts, thus improving its robustness. References [1] Daniel J. Felleman and David C. Van Essen. Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex, 1(1):1 -- 47, 1991. [2] Xin Li, Zequn Jie, Jiashi Feng, Changsong Liu, and Shuicheng Yan. Learning with rethinking: Recurrently improving convolutional neural networks through feedback. Pattern Recognition, 79:183 -- 194, 2018. [3] Aran Nayebi, Daniel Bear, Jonas Kubilius, Kohitij Kar, Surya Ganguli, David Sussillo, James J. DiCarlo, and Daniel L. K. Yamins. Task-driven convolutional recurrent models of the visual system. CoRR, abs/1807.00053, 2018. [4] Saumya Jetley, Nicholas A. Lord, Namhoon Lee, and Philip H. S. Torr. Learn to pay attention. CoRR, abs/1804.02391, 2018. [5] Jianyu Wang, Zhishuai Zhang, Cihang Xie, Vittal Premachandran, and Alan Yuille. Unsupervised learning of object semantic parts from internal states of cnns by population encoding. Computer Science, 2015. [6] David Marr and Tomaso Poggio. Cooperative computation of stereo disparity. Science, 194(4262):283 -- 287, 1976. [7] Stuart Geman and Donald Geman. Stochastic relaxation, gibbs distributions, and the bayesian restoration of images. In Readings in computer vision, pages 564 -- 584. Elsevier, 1987. [8] James H. Elder, Amnon Krupnik, and Leigh A. Johnston. Contour grouping with prior models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(6):661 -- 674, 2003. [9] Roland Memisevic and Geoffrey E Hinton. Learning to represent spatial transformations with factored higher-order boltzmann machines. Neural computation, 22(6):1473 -- 1492, 2010. [10] Volodymyr Mnih, Geoffrey E Hinton, et al. Generating more realistic images using gated mrf's. In Advances in Neural Information Processing Systems, pages 2002 -- 2010, 2010. [11] Kuan Han, Haiguang Wen, Yizhen Zhang, Di Fu, Eugenio Culurciello, and Zhongming Liu. Deep predictive coding network with local recurrent processing for object recognition. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31, pages 9201 -- 9213. Curran Associates, Inc., 2018. 9 A PREPRINT - DECEMBER 24, 2019 [12] Alessandra Angelucci and Paul C Bressloff. Contribution of feedforward, lateral and feedback connections to the classical receptive field center and extra-classical receptive field surround of primate v1 neurons. Progress in brain research, 154:93 -- 120, 2006. [13] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. [14] Jianyu Wang, Cihang Xie, Zhishuai Zhang, Jun Zhu, Lingxi Xie, and Alan L. Yuille. Detecting semantic parts on partially occluded objects. CoRR, abs/1707.07819, 2017. [15] Wu Li, Valentin Piëch, and Charles D Gilbert. Contour saliency in primary visual cortex. Neuron, 50(6):951 -- 962, 2006. [16] Ramakrishnan Iyer and Stefan Mihalas. Cortical circuits implement optimal context integration. bioRxiv, page 158360, 2017. [17] Tai Sing Lee, Tom Stepleton, Brian Potetz, and Jason M Samonds. Neural encoding of scene statistics for surface and object inference. 2008. [18] Christof Koch, Jose Marroquin, and Alan Yuille. Analog" neuronal" networks in early vision. Proceedings of the National Academy of Sciences, 83(12):4263 -- 4267, 1986. [19] Bruno A Olshausen and David J Field. Sparse coding of sensory inputs. Current opinion in neurobiology, 14(4):481 -- 487, 2004. [20] Christopher J Rozell, Don H Johnson, Richard G Baraniuk, and Bruno A Olshausen. Sparse coding via thresholding and local competition in neural circuits. Neural computation, 20(10):2526 -- 2563, 2008. [21] Tai Sing Lee and David Mumford. Hierarchical bayesian inference in the visual cortex. JOSA A, 20(7):1434 -- 1448, 2003. [22] Mingmin Zhao, Chengxu Zhuang, Yizhou Wang, and Tai Sing Lee. Predictive encoding of contextual relationships for perceptual inference, interpolation and prediction, 2014. 10
1610.07181
2
1610
2017-03-12T15:12:08
Death and rebirth of neural activity in sparse inhibitory networks
[ "q-bio.NC", "cond-mat.dis-nn", "nlin.CD" ]
In this paper, we clarify the mechanisms underlying a general phenomenon present in pulse-coupled heterogeneous inhibitory networks: inhibition can induce not only suppression of the neural activity, as expected, but it can also promote neural reactivation. In particular, for globally coupled systems, the number of firing neurons monotonically reduces upon increasing the strength of inhibition (neurons' death). However, the random pruning of the connections is able to reverse the action of inhibition, i.e. in a sparse network a sufficiently strong synaptic strength can surprisingly promote, rather than depress, the activity of the neurons (neurons' rebirth). Thus the number of firing neurons reveals a minimum at some intermediate synaptic strength. We show that this minimum signals a transition from a regime dominated by the neurons with higher firing activity to a phase where all neurons are effectively sub-threshold and their irregular firing is driven by current fluctuations. We explain the origin of the transition by deriving an analytic mean field formulation of the problem able to provide the fraction of active neurons as well as the first two moments of their firing statistics. The introduction of a synaptic time scale does not modify the main aspects of the reported phenomenon. However, for sufficiently slow synapses the transition becomes dramatic, the system passes from a perfectly regular evolution to an irregular bursting dynamics. In this latter regime the model provides predictions consistent with experimental findings for a specific class of neurons, namely the medium spiny neurons in the striatum.
q-bio.NC
q-bio
Death and rebirth of neural activity in sparse inhibitory networks Aix Marseille Univ, INSERM, INMED and INS, Inst Neurosci Syst, Marseille, France and Aix Marseille Univ, Universit´e de Toulon, CNRS, CPT, UMR 7332, 13288 Marseille, France David Angulo-Garcia∗ CNR - Consiglio Nazionale delle Ricerche - Istituto dei Sistemi Complessi, 50019 Sesto Fiorentino, Italy and INFN - Istituto Nazionale di Fisica Nucleare - Sezione di Firenze, 50019 Sesto Fiorentino, Italy Stefano Luccioli† CNR - Consiglio Nazionale delle Ricerche - Istituto dei Sistemi Complessi, 50019 Sesto Fiorentino, Italy and Aix Marseille Univ, INSERM, INS, Inst Neurosci Syst, Marseille, France Weierstrass Institute, Mohrenstrasse 39, 10117 Berlin, Germany Simona Olmi‡ Alessandro Torcini§ Laboratoire de Physique Th´eorique et Mod´elisation, Universit´e de Cergy-Pontoise, CNRS, UMR 8089, 95302 Cergy-Pontoise cedex, France Aix Marseille Univ, INSERM, INMED and INS, Inst Neurosci Syst, Marseille, France Aix Marseille Univ, Universit´e de Toulon, CNRS, CPT, UMR 7332, 13288 Marseille, France CNR - Consiglio Nazionale delle Ricerche - Istituto dei Sistemi Complessi, 50019 Sesto Fiorentino, Italy and Max-Planck-Institut fur Physik komplexer Systeme, Nothnitzer Strasse 38, 01187 Dresden, Germany (Dated: April 3, 2018) Inhibition is a key aspect of neural dynamics playing a fundamental role for the emergence of neural rhythms and the implementation of various information coding strategies. Inhibitory populations are present in several brain structures and the comprehension of their dynamics is strategical for the understanding of neural processing. In this paper, we clarify the mechanisms underlying a general phenomenon present in pulse-coupled heterogeneous inhibitory networks: inhibition can induce not only suppression of the neural activity, as expected, but it can also promote neural reactivation. In particular, for globally coupled systems, the number of firing neurons monotonically reduces upon increasing the strength of inhibition (neurons' death). However, the random pruning of the connections is able to reverse the action of inhibition, i.e. in a random sparse network a sufficiently strong synaptic strength can surprisingly promote, rather than depress, the activity of the neurons (neurons' rebirth). Thus the number of firing neurons reveals a minimum at some intermediate synaptic strength. We show that this minimum signals a transition from a regime dominated by the neurons with higher firing activity to a phase where all neurons are effectively sub-threshold and their irregular firing is driven by current fluctuations. We explain the origin of the transition by deriving a mean field formulation of the problem able to provide the fraction of active neurons as well as the first two moments of their firing statistics. The introduction of a synaptic time scale does not modify the main aspects of the reported phenomenon. However, for sufficiently slow synapses the transition becomes dramatic, the system passes from a perfectly regular evolution to an irregular bursting dynamics. In this latter regime the model provides predictions consistent with experimental findings for a specific class of neurons, namely the medium spiny neurons in the striatum. PACS numbers: 87.19.lj,05.45.Xt,87.19.lm I. INTRODUCTION The presence of inhibition in excitable systems in- duces a rich dynamical repertoire, which is extremely relevant for biological [13], physical [33] and chemical systems [84]. In particular, inhibitory coupling has been invoked to explain cell navigation [87], morphogenesis in ∗ [email protected] † [email protected] ‡ [email protected] § [email protected] animal coat pattern formation [46], and the rhythmic ac- tivity of central pattern generators in many biological systems [28, 45]. In brain circuits the role of inhibi- tion is fundamental to balance massive recurrent exci- tation [73] in order to generate physiologically relevant cortical rhythms [12, 72]. Inhibitory networks are important not only for the emergence of rhythms in the brain, but also for the fun- damental role they play in information encoding in the olfactory system [40] as well as in controlling and reg- ulating motor and learning activity in the basal gan- glia [5, 15, 47]. Furthermore, stimulus dependent sequen- tial activation of neurons or group of neurons, reported for asymmetrically connected inhibitory cells [34, 52], has 1 a) nA 0.8 0.6 0.4 0.2 0 0.1 1 10 g 100 1000 2 1 b) 0.9 nA 0.8 4 P(ν) g=1 2 0 0 0.7 0.6 0.1 0.2 0.4 ν 0.6 1 g 10 5 0 0 P(ν) g=10 0.2 ν 0.4 0.6 10 FIG. 1. a-b) Fraction of active neurons nA as a function of the inhibitory synaptic strength g for a globally coupled system (a), where K = N − 1, and a randomly connected (sparse) network with K = 20 (b). In a) is reported the asymptotic value nA calculated after a time tS = 1 × 106. Conversely in b), nA is reported at successive times: namely, tS = 985 (red squares), tS = 1.1 × 104 (brown stars), tS = 5 × 105 (blue diamonds) and tS = 1 × 106 (green triangles). An estimation of the times needed to reach nA = 1 can be obtained by employing Eq. (13) these values range from ts = 5 × 109 for g = 0.1 to 5 × 105 for g = 50. Insets in b) depict the probability distributions P (ν) of the single neuron firing rate ν for the sparse network for a given g indicated in the inset at two different times: tS = 985 (red filled histograms) and tS = 1 × 106 (thick empty green histograms). The histograms are calculated by considering only active neurons. The reported data refer to instantaneous synapses, to a system size N = 400 and to an uniform distribution P (I) with [l1, l2] = [1.0, 1.5] and θ = 1, the values reported in a) and b) have been also averaged over 10 random realizations of the network. been suggested as a possible mechanism to explain se- quential memory storage and feature binding [67]. These explain the long term interest for numerical and theoretical investigations of the dynamics of inhibitory networks. Already the study of globally coupled homo- geneous systems revealed interesting dynamical features, ranging from full synchronization to clustering appear- ance [24, 83, 85], from the emergence of splay states [90] to oscillator death [6]. The introduction of disorder, e.g. random dilution, noise or other form of heterogeneity in these systems leads to more complex dynamics, ranging from fast global oscillations [9] in neural networks and self-sustained activity in excitable systems [38], to irreg- ular dynamics [3, 30–32, 42, 49, 56, 82, 90]. In particular, inhibitory spiking networks, due to stable chaos [63], can display extremely long erratic transients even in linearly stable regimes [3, 30, 31, 42, 49, 82, 89, 90]. One of the most studied inhibitory neural population is represented by medium spiny neurons (MSNs) in the striatum (which is the main input structure of the basal ganglia) [37, 57]. In a series of papers, Ponzi and Wickens have shown that the main features of the MSN dynamics can be reproduced by considering a randomly connected inhibitory network of conductance based neurons sub- ject to external stochastic excitatory inputs [64–66]. Our study has been motivated by an interesting phenomenon reported for this model in [66]: namely, upon increasing the synaptic strength the system passes from a regularly firing regime, characterized by a large part of quiescent neurons, to a biologically relevant regime where almost all cells exhibit a bursting activity, characterized by an alternation of periods of silence and of high firing. The same phenomenology has been recently reproduced by employing a much simpler neural model [1]. Thus sug- gesting that this behaviour is not related to the specific model employed, but it is indeed a quite general prop- erty of inhibitory networks. However, it is still unclear the origin of the phenomenon and the minimal ingredi- ents required to observe the emergence of this effect. In order to exemplify the problem addressed in this pa- per we report in Fig. 1 the fraction of active neurons nA (i.e. the ones emitting at least one spike during the simu- lation time) as a function of the strength of the synaptic inhibition g in an heterogeneous network. For a fully cou- pled network, nA has a monotonic decrease with g (Fig. 1 (a)), while for a random sparse network nA has a non monotonic behaviour, displaying a minimum at an inter- mediate strength gm (Fig. 1 (b)). In fully coupled net- works the effect of inhibition is simply to reduce the num- ber of active neurons (neurons' death). However, quite counter-intuitively, in presence of dilution by increasing the synaptic strength the previously silenced neurons can return to fire (neurons' rebirth). Our aim is to clarify the physical mechanisms underlying neuron's death and re- birth, which are at the origin of the behaviour reported in [1, 66]. In particular, we consider a deterministic network of purely inhibitory pulse-coupled Leaky Integrate-and-Fire (LIF) neurons with an heterogeneous distribution of ex- citatory DC currents, accounting for the different level of excitability of the neurons. The evolution of this model is studied for fully coupled and for random sparse topology, as well as for synapses with different time courses. For the fully coupled case, it is possible to derive, within a self-consistent mean field approach, the analytic expres- sions for the fraction of active neurons and for the average firing frequency ¯ν as a function of the coupling strength g. In this case the monotonic decrease of nA with g can be interpreted as a Winner Takes All (WTA) mecha- nism [16, 21, 88], where only the most excitable neurons survive to the inhibition increase. For random sparse networks, the neurons' rebirth can be interpreted as a re- activation process induced by erratic fluctuations in the synaptic currents. Within this framework it is possible to obtain semi-analytically, for instantaneous synapses, a closed set of equations for nA as well as for the average firing rate and coefficient of variation as a function of the coupling strength. In particular, the firing statistics of the network can be obtained via a mean-field approach by extending the formulation derived in [70] to account for synaptic shot noise with constant amplitude. The in- troduction of a finite synaptic time scale do not modify the overall scenario as far as this is shorter than the mem- brane time constant. As soon as the synaptic dynamics becomes slower, the phenomenology of the transition is modified. At g < gm we have a frozen phase where nA does not evolve in time on the explored time scales, since the current fluctuations are negligible. Above gm we have a bursting regime, which can be related to the emergence of correlated fluctuations induced by slow synaptic times, as discussed in the framework of the adiabatic approach in [50, 51]. The paper is organized as follows: In Sect. II we present the models that will be considered in the pa- per as well as the methods adopted to characterize its dynamics. In Sect. III we consider the globally cou- pled network where we provide analytic self-consistent expressions accounting for the fraction of active neurons and the average firing rate. Section IV is devoted to the study of sparsely connected networks with instantaneous synapses and to the derivation of the set of semi-analytic self-consistent equations providing nA, the average fir- ing rate and the coefficient of variation. In section V we discuss the effect of synaptic filtering with a particular attention on slow synapses. Finally in Sect. VI we briefly discuss the obtained results with a focus on the biological relevance of our model. II. MODEL AND METHODS We examine the dynamical properties of an heteroge- neous inhibitory sparse network made of N LIF neurons. The time evolution of the membrane potential vi of the i-th neuron is ruled by the following first order ordinary differential equation: vi(t) = Ii − vi(t) − gEi(t) ; (1) where g > 0 is the inhibitory synaptic strength, Ii is the neuronal excitability of the i-th neuron encompass- 3 ing both the intrinsic neuronal properties and the ex- citatory stimuli originating from areas outside the con- sidered neural circuit and Ei(t) represents the synaptic current due to the recurrent interactions within the con- sidered network. The membrane potential vi of neuron i evolves accordingly to Eq. (1) until it overcomes a con- stant threshold θ = 1, this leads to the emission of a spike (action potential) transmitted to all the connected post-synaptic neurons, while vi is reset to its rest value vr = 0. The model in (1) is expressed in adimensional units, this amounts to assume a membrane time con- stant τm = 1, for the conversion to dimensional variables see Appendix A. The heterogeneity is introduced in the model by assigning to each neuron a different value of input excitability Ii drawn from a flat distribution P (I), whose support is I ∈ [l1, l2] with l1 ≥ θ, therefore all the neurons are supra-threshold. The synaptic current Ei(t) is given by the linear super- position of all the inhibitory post-synaptic potentials (IP- SPs) η(t) emitted at previous times tj n < t by the pre- synaptic neurons connected to neuron i, namely Ei(t) = 1 K Xj6=i Cij Xntn<t η(t − tj n) ; (2) where K is the number of pre-synaptic neurons. Cij rep- resent the elements of the N × N connectivity matrix associated to an undirected random network, whose en- tries are 1 if there is a synaptic connection from neu- ron j to neuron i, and 0 otherwise. For the sparse net- work, we select randomly the matrix entries, however to reduce the sources of variability in the network, we as- sume that the number of pre-synaptic neurons is fixed, namely Pj6=i Cij = K << N for each neuron i, where autaptic connections are not allowed. We have verified that the results do not change if we choose randomly the links accordingly to an Erdos-Renyi distribution with a probability K/N . For a fully coupled network we have K = N − 1. The shape of the IPSP characterize the type of fil- tering performed by the synapses on the received action potentials. We have considered two kind of synapses, in- stantaneous ones, where η(t) = δ(t), and synapses where the PSP is an α-pulse, namely η(t) = H(t)α2te−tα ; (3) with H denoting the Heaviside step function. In this latter case the rise and decay time of the pulse are the same, namely τα = 1/α, and therefore the pulse duration τP can be assumed to be twice the characteristic time τα. The equations of the model Eqs. (1) and (2) are integrated exactly in terms of the associated event driven maps for different synaptic filtering, these correspond to Poincar´e maps performed at the firing times (for details see Appendix A) [53, 91]. For instantaneous synapses, we have usually consid- ered system sizes N = 400 and N = 1, 400 and for the sparse case in-degrees 20 ≤ K ≤ 80 for N = 400 and 20 ≤ K ≤ 600 for N = 1400 with integration times up to tS = 1 × 106. For synapses with a finite decay time we limit the analysis to N = 400 and K = 20 and to maximal integration times tS = 1 × 105. Finite size dependences on N are negligible with these parameter choices, as we have verified. In order to characterize the network dynamics we mea- sure the fraction of active neurons nA(tS) at time tS, i.e. the fraction of neurons emitting at least one spike in the time interval [0, tS]. Therefore a neuron will be consid- ered silent if it has a frequency smaller than 1/tS, with our choices of tS = 105−106, this corresponds to neurons with frequencies smaller than 10−3− 10−4 Hz, by assum- ing as timescale a membrane time constant τm = 10 ms. The estimation of the number of active neurons is always started after a sufficiently long transient time has been discarded, usually corresponding to the time needed to deliver 106 spikes in the network. Furthermore, for each neuron we estimate the time av- eraged inter-spike interval (ISI) TISI , the associated fir- ing frequency ν = 1/TISI, as well as the coefficient of variation CV , which is the ratio of the standard devia- tion of the ISI distribution divided by TISI. For a regu- lar spike train CV = 0, for a Poissonian distributed one CV = 1, while CV > 1 is an indication of bursting activ- ity. The indicators usually reported in the following to characterize the network activity are ensemble averages over all the active neurons, which we denote as ¯a for a generic observable a. To analyze the linear stability of the dynamical evo- lution we measure the maximal Lyapunov exponent λ, which is positive for chaotic evolution, and negative (zero) for stable (marginally stable) dynamics [4]. In par- ticular, by following [2, 55] λ is estimated by linearizing the corresponding event driven map. III. FULLY COUPLED NETWORKS: WINNER TAKES ALL In the fully coupled case we observe that the number of active neurons nA saturates, after a short transient, to a value which remains constant in time. In this case, it is possible to derive a self-consistent mean field approach to obtain analytic expressions for the fraction of active neurons nA and for the average firing frequency ¯ν of the neurons in the network. In a fully coupled network each neuron receives the spikes emitted by the other K = N−1 neurons, therefore each neuron is essentially subject to the same effective input µ, apart corrections O(1/N ). citability I, is given by The effective input current, for a neuron with an ex- µ = I − g ¯νnA ; (4) where nA(N − 1) is the number of active pre-synaptic neurons assumed to fire with the same average frequency ¯ν. 4 In a mean field approach, each neuron can be seen as isolated from the network and driven by the effective input current µ. Taking into account the distribution of the excitabilities P (I) one obtains the following self- consistent expression for the average firing frequency ¯ν = 1 ∆Z{IA} dI P (I)(cid:20)ln(cid:18) I − g ¯νnA − vr I − g ¯νnA − θ (cid:19)(cid:21)−1 (5) where the integral is restricted only to active neurons, i.e. to I ∈ {IA} values for which the logarithm is defined, while ∆ =R{IA} dI P (I) is the measure of their support. In (5) we have used the fact that for an isolated LIF neuron with constant excitability C, the ISI is simply given by TISI = ln[(C − vr)/(C − θ)] [11]. An implicit expression for nA can be obtained by esti- mating the neurons with effective input µ > θ, in partic- ular the number of silent neurons is given by 1 − nA =Z l∗ l1 dIP (I) , (6) where l1 is the lower limit of the support of the distri- bution, while l∗ = g ¯νnA + θ. By solving self-consistently Eqs.(5) and (6) one can obtain the analytic expression for nA and ¯ν for any distribution P (I). In particular, for excitabilities distributed uniformly in the interval [l1, l2], the expression for the average fre- quency Eq. (5) becomes ¯ν = 1 nA(l2 − l1)Z{IA} dI(cid:20)ln(cid:18) I − g ¯νnA − vr I − g ¯νnA − θ (cid:19)(cid:21)−1 ;(7) while the number of active neurons is given by the fol- lowing expression nA = l2 − θ l2 − l1 + g ¯ν ; (8) with the constraint that nA cannot be larger than one. The analytic results for these quantities compare quite well with the numerical findings estimated for different distribution intervals [l1, l2], different coupling strengths and system sizes, as shown in Fig. 2. Apart for definitely large coupling g > 10 where some discrepancies among the mean field estimations and the simulation results for ¯ν are observable (see Fig. 2 (b)). These differences are probably due to the discreteness of the pulses, which can- not be neglected for very large synaptic strengths. As a general feature we observe that nA is steadily de- creasing with g, thus indicating that a group of neurons with higher effective inputs (winners) silence the other neurons (losers) and that the number of winners eventu- ally vanishes for sufficiently large coupling in the limit of large system sizes. Furthermore, the average excitability of the active neurons (the winners) ¯IA increases with g, as shown in the inset of Fig. 2 (a), thus revealing that only the neurons with higher excitabilities survive to the silencing action exerted by the other neurons. At the a) 10 100 b) IA g 1 10 100 0.1 1 g 1 0.8 0.6 0.4 0.2 nA 1.8 1.7 1.6 1.5 0 0.01 1 ν 0.75 0.5 0.25 0 0.01 0.1 1 g 10 100 FIG. 2. Fraction of active neurons nA (a) and average net- work's frequency ¯ν (b) as a function of the synaptic strength g for uniform distributions P (I) with different supports. Inset: average neuronal excitability of the active neurons ¯IA versus g. Empty (filled) symbols refer to numerical simulation with N = 400 (N = 1400) and dashed lines to the corresponding analytic solution. Symbols and lines correspond from bottom to top to [l1, l2] = [1.0, 1.5] (black); [l1, l2] = [1.0, 1.8] (red) and [l1, l2] = [1.2, 2.0] (blue). The data have been averaged over a time interval tS = 1 × 106 after discarding a transient of 106 spikes. same time, as an effect of the growing inhibition the av- erage firing rate of the winners dramatically slows down. Therefore despite the increase of ¯IA the average effective input ¯µ indeed decreases for increasing inhibition. This represents a clear example of the winner takes all (WTA) mechanism obtained via (lateral) inhibition, which has been shown to have biological relevance for neural sys- tems [20, 22, 60, 88]. It is important to understand which is the minimal coupling value gc for which the firing neurons start to die. In order to estimate gc it is sufficient to set nA = 1 in Eqs. (7) and (8). In particular, one gets gc = (l1 − θ)/¯ν , (9) thus for l1 = θ even an infinitesimally small coupling is in principle sufficient to silence some neurons. Furthermore, from Fig. 3 (a) it is evident that whenever the excitabil- 5 ities become homogeneous, i.e. for l1 → l2, the critical synaptic coupling gc diverges towards infinity. Thus het- erogeneity in the excitability distribution is a necessary condition in order to observe a gradual neurons' death, as shown in Fig. 2 (a). This is in agreement with the results reported in [7], where homogeneous fully coupled networks of inhibitory LIF neurons have been examined. In particular, for finite systems and slow synapses the authors in [7] reveal the existence of a sub-critical Hopf bifurcation from a fully synchronized state to a regime characterized by oscilla- tor death occurring at some critical gc. However, in the thermodynamic limit gc → ∞ for fast as well as slow synapses, in agreement with our mean field result for in- stantaneous synapses. We also proceed to investigate the isolines correspond- ing to the same critical gc in the (l1, l2)-plane, the results are reported in Fig. 3 (b) for three selected values of gc. It is evident that the l1 and l2-values associated to the isolines display a direct proportionality among them. However, despite lying on the same gc-isoline, different parameter values induce a completely different behaviour of nA as a function of the synaptic strength, as shown in the inset of Fig. 3 (b). Direct simulations of the network at finite sizes, namely for N = 400 and N = 1400, show that for sufficiently large coupling neurons with similar excitabilities tend to form clusters, similarly to what reported in [42], for the same model here studied, but with a delayed pulse transmission. However, at variance with [42], the overall macroscopic dynamics is asynchronous and no collective oscillations can be detected for the whole range of con- sidered synaptic strengths. IV. SPARSE NETWORKS : NEURONS' REBIRTH In this Section we will consider a network with sparse connectivity, namely each neuron is supra-threshold and it receives instantaneous IPSPs from K << N randomly chosen neurons in the network. Due to the sparseness, the input spike trains can be considered as uncorrelated and at a first approximation it can be assumed that each spike train is Poissonian with a frequency ¯ν correspon- dent to the average firing rate of the neurons in the net- work [8, 9]. Usually, the mean activity of a LIF neural network has been estimated in the context of the diffu- sion approximation [69, 80]. This approximation is valid whenever the arrival frequency of the IPSPs is high with respect to the firing emission, while the amplitude of each IPSPs (namely, G = g/K) is small with respect to the firing threshold θ. This latter hypothesis in our case is not valid for sufficiently large (small) synaptic strength g (in-degree K), as it can be appreciated by the com- parison shown in Fig. 13 in Appendix B. Therefore the synaptic inputs should be treated as shot noise. In par- ticular, here we apply an extended version of the analytic a) l2/θ=1.5 l2/θ=2 l2/θ=2.5 l2/θ=3 10 1 0.1 gc 0.01 1 1.5 3 b) 2.5 l2/θ 2 1.5 1 1 2.5 3 2 l1/θ 1 nA 0.9 0.8 0 1 1.5 2 l1/θ 3 2 g 2.5 3 FIG. 3. a) Critical value gc as a function of the lower value of the excitability l1 for several choices of the upper limit l2. b) Isolines corresponding to constant values of gc in the (l1, l2)-plane: namely, gc = 0.5 (black solid line), gc = 1.0 (red dashed line), gc = 2.0 (blue dotted line). Inset: Dependence of nA on g for three couples of values (l1, l2) chosen along each of the isolines reported in the main figure. approach derived by Richardson and Swabrick in [70] to estimate the average firing rate and the average coef- ficient of variation for LIF neurons with instantaneous synapses subject to inhibitory shot noise of constant am- plitude (for more details see Appendices B and C). At variance with the fully coupled case, the fraction of active neurons nA does not saturate to a constant value for sufficiently short times. Instead, nA increases in time, due to the rebirth of losers which have been previously silenced by the firing activity of the winners, as shown in in Fig. 1(b). This effect is clearly illustrated by con- sidering the probability distributions P (ν) of the firing rates of the neurons at successive integration times tS. These are reported in the insets of Fig 1(b) for two cou- pling strengths and two times: namely, tS = 985 (red lines) and tS = 1 × 106 (green lines). From these data is evident that the fraction of neurons with low firing rate (the losers) increases in time, while the fraction of high firing neurons remains almost unchanged. More- 6 over, the variation of nA slows down for increasing tS and nA approaches some apparently asymptotic profile for sufficiently long integration times. Furthermore, nA has a non monotonic behaviour with g, as opposite to the fully coupled case. In particular, nA reveals a minimum nAm at some intermediate synaptic strength gm followed by an increase towards nA = 1 at large g. As we have verified, as far as 1 < K << N finite size effects are neg- ligible and the actual value of nA depends only on the in-degree K and the considered simulation time tS. In the following we will try to explain the origin of such a behaviour. Despite the model is fully deterministic, due to the random connectivity the rebirth of silent neurons can be interpreted in the framework of activation processes in- duced by random fluctuations. In particular, we can as- sume that each neuron in the network will receive nAK independent Poissonian trains of inhibitory kicks of con- stant amplitude G characterized by an average frequency ¯ν , thus each synaptic input can be regarded as a single Poissonian train with total frequency R = nAK ¯ν. There- fore, each neuron, characterized by its own excitability I, will be subject to an average effective input µ(I) (as reported in Eq. (4)) plus fluctuations in the synaptic current of intensity σ = gr nA¯ν K . (10) Indeed, we have verified that (10) gives a quantitatively correct estimation of the synaptic current fluctuations over the whole range of synaptic coupling here consid- ered (as shown in Fig. 4). A closer analysis of the prob- ability distributions P (IAT ) of the inter-arrival times (IATs) indicates that these are essentially exponentially distributed, as expected for Poissonian processes, with a decay rate given by R, as evident from Fig. 5 for two different synaptic strenghts. For instantaneous IPSP, the current fluctuations are due to stable chaos [63], since the maximal Lyapunov exponent is negative for the whole range of coupling, as we have verified. Therefore, as reported by many au- thors, erratic fluctuations in inhibitory neural networks with instantaneous synapses are due to finite amplitude instabilities, while at the infinitesimal level the system is stable [3, 30, 31, 42, 49, 82, 90]. In this picture, the silent neurons stay in a quiescent state corresponding to the minimum of the effective po- tential U(v) = v2/2 − µv and in order to fire they should overcome a barrier ∆U = (θ − µ)2/2. The average time tA required to overcome such barrier can be estimated accordingly to the Kramers' theory for activation pro- cesses [26, 80], namely tA ≃ τ0 exp (θ − µ(I))2 σ2 ! ; (11) where τ0 is an effective time scale taking in account the intrinsic non stationarity of the process, i.e. the fact that 1.5 1 σ , A µ 0.5 0 0 10 20 g 30 40 50 FIG. 4. Effective average input of the active neurons ¯µA (black circles) and average fluctuations of the synaptic cur- rents ¯σ (red squares) as a function of the inhibitory coupling g. The threshold potential θ = 1 is marked by the (blue) horizontal dotted line and gm by the (green) vertical dash- dotted line. The dashed black (red) line refer to the theo- retical estimation for µA (σ) reported in Eq. (4) (Eq. (10)) and averaged only over the active neurons. The data refer to N = 1400, K = 140, [l1, l2] = [1.0 : 1.5] and to a simulation time tS = 1 × 106. 100 gc 10 1 0.1 1 7 10 ts/τ0 100 1000 FIG. 6. Critical values gc1 (black) and gc2 (red) as calculated from Eq. (13) for l1 = 1.2 (dash-dotted), l1 = 1.15 (contin- uous), l1 = 1.1 (dashed) and l1 = 1.0 (dotted line). All the values entering in Eq. (13) are taken from simulation. All other parameters used for the simulation as in Fig. 1 (b). acterized by the following excitability I = l2 − nA(tS)(l2 − l1) ; (12) a) 2 1.5 1 0.5 ) T A I ( P b) for excitabilities I uniformly distributed in the interval [l1, l2]. In order to obtain an explicit expression for the fraction of active neurons at time tS, one should solve the equation (11) for the neuron with excitability I by setting tS = tA, thus obtaining the following solution ) T A I ( P 5 4 3 2 1 0 0 0.25 0.5 IAT 0.75 1 0 0 0.5 1 1.5 IAT 2 2.5 3 FIG. 5. Probability distributions of the inter-arrival times (IATs) for a generic neuron in the network for a) g = 1.3 and b) g = 10. In both panels, the red continuous line indicate the exponential distribution corresponding to a purely Poissonian process with arrival rate given by R = nAK ¯ν and the dashed blue vertical lines refer to the average IAT for the Poissonian distribution, namely 1/R. The distributions have been evalu- ated for the arrival of 5 × 105 IPSPs. Other parameters used for the simulation as in Fig. 1 (b). the number of active neurons increases during the time evolution. It is important to stress that the expression (11) will remain valid also in the limit of large synaptic cou- plings, because not only σ2, but also the barrier height will increase with g. Furthermore, both these quantities grow quadratically with g at sufficiently large synaptic strength, as it can be inferred from Eqs. (4) and (10). It is reasonable to assume that at a given time tS all neurons with tA < tS will have fired at least once and that the more excitable will fire first. Therefore by as- suming that the fraction of active neurons at time tS is nA(tS), the last neuron which has fired should be char- nA(tS) =  where φ − 2βγ +pφ2 − 4φβγ 2γ2 1 if nA < 1 otherwise (13) g2 K ¯ν ln(tS/τ0) β = θ − l2 . γ = (l2 − l1) + g ¯ν φ = Equation (13) gives the dependence of nA on the cou- pling strength g for a fixed integration time tS and time scale τ0, whenever we can provide the value of the av- erage frequency ¯ν. A quick inspection to Eqs. (11) and (13) shows that, setting nA = 1, we obtain two solu- tions for the critical couplings gc1 (gc2) below (above) which all neurons will fire at least once in the cosidered time interval. The solutions are reported in Fig 6, in particular we observe that whenever l1 → vth the crit- ical coupling gc1 will vanish, analogously to the fully coupled situation. These results clearly indicate that nA should display a minimum for some finite coupling strenght gm ∈ (gc1, gc2). Furthermore, as shown in Fig 6 the two critical couplings approach one another for in- creasing tS and finally merge, indicating that at suffi- ciently long times all the neurons will be active at any synaptic coupling strength g. The average frequency ¯ν can be obtained analytically by following the approach described in Appendix B for LIF neurons with instantaneous synapses subject to in- hibitory Poissonian spike trains. In particular, the self- consistent expression for the average frequency reads as ¯ν =Z{IA} dIP (I)ν0(I, G, nA, ¯ν) ; (14) where the explicit expression of ν0 is given by Eq. (31) in Appendix B. The simultaneous solution of Eqs. (13) and (14) pro- vides a theoretical estimation of nA and ¯ν for the whole considered range of synaptic strength, once the unknown time scale τ0 is fixed. This time scale has been deter- mined via an optimal fitting procedure for sparse net- works with N = 400 and K = 20, 40 and 80 at a fixed integration time tS = 1 × 106. The results for nA are reported in Fig. 7 (a), the estimated curves reproduce reasonably well the numerical data for K = 20 and 40, while for K = 80 the agreement worsens at large cou- pling strengths. This should be due to the fact that by increasing g and K the spike trains stimulating each neu- ron cannot be assumed to be completely independent, as done for the derivation of Eqs. (13) and (14). Neverthe- less, the average frequency ¯ν is quantitatively well repro- duced for the considered K values over the entire range of the synaptic strengths, as it is evident from Figs. 7 (b-d). A more detailed comparison between the theoret- ical estimations and the numerical data can be obtained by considering the distributions P (ν) of the single neu- ron firing rate for different coupling strengths reported in Figs. 7 (k-m) for K = 40. The overall agreement can be considered as more than satisfactory, the observable discrepancies are probably due to the fact that our ap- proach neglect a further source of disorder present in the system and related to the heterogeneity in the number of active pre-synaptic neurons [9]. We have also estimated analytically the average coeffi- cient of variation of the firing neurons CV by extending the method derived in [70] to obtain the response of a neuron receiving synaptic shot noise inputs. The analytic expressions of the coefficient of variation for LIF neurons subject to inhibitory shot noise with fixed post-synaptic amplitude are obtained by estimating the second moment of the associated first-passage-time distribution, the de- tails are reported in Appendix C. The coefficient of vari- ation can be estimated, once the self consistent values for nA and ¯ν have been obtained via Eqs. (13) and (14). The comparison with the numerical data, reported in Figs 7 (e-g), reveals a good agreement over the whole range of synaptic strengths for all the considered in-degrees. At sufficiently small synaptic coupling the neurons fire tonically and almost independently, as it emerges clearly from the raster plot in Fig 7 (h) and by the fact that ¯ν approaches the average value for the uncoupled system (namely, 0.605) and CV → 0. Furthermore, the neu- ronal firing rates are distributed towards finite values in- dicating that the inhibition as a minor role in this case, as shown in Fig 7 (k). By increasing the coupling, nA decreases, as an effect of the inhibition more and more 8 neurons are silenced (as evident from Fig. 7 (l)) and the average firing rate decrease, at the same time the dynamics becomes slightly more irregular as shown in Fig 7 (i). At large coupling g > gm, a new regime ap- pears, where almost all neurons become active but with an extremely slow dynamics which is essentially stochas- tic with CV ≃ 1, as testified also by the raster plot re- ported in Fig 7 (j) and by by the firing rate distribution shown in Fig 7 (m). Furthermore, from Fig. 7(a) it is clear that the value of the minimum of the fraction of active neurons nAm de- creases by increasing the network in-degree K, while gm increases with K. This behaviour is further investigated in a larger network, namely N = 1400, and reported in the inset of Fig. 8. It is evident that nA stays close to the globally coupled solutions over larger and larger intervals for increasing K. This can be qualitatively understood by the fact that the current fluctuations Eq. (10), respon- sible for the rebirth of silent neurons, are proportional to g and scales as 1/√K, therefore at larger in-degree the fluctuations have similar intensities only for larger synap- tic coupling. The general mechanism behind neurons' rebirth can be understood by considering the value of the effective neuronal input and of the current fluctuations as a func- tion of g. As shown in Fig. 4, the effective input current ¯µA, averaged over the active neurons, essentially coincide with the average of the excitability ¯IA for g → 0, where the neurons can be considered as independent one from the others. The inhibition leads to a decrease of ¯µA, and to a crossing of the threshold θ exactly at g = gm. This indicates that at g < gm the active neurons, being on average supra-threshold, fire almost tonically inhibiting the losers via a WTA mechanism. In this case the firing neurons are essentially mean-driven and the current fluc- tuations play a role on the rebirth of silent neurons only on extremely long time scales, this is confirmed by the low values of ¯σ in such a range, as evident from Fig. 4. On the other hand, for g > gm, the active neurons are now on average below threshold while the fluctuations dominate the dynamics. In particular, the firing is now extremely irregular being due mainly to reactivation pro- cesses. Therefore the origin of the minimum in nA can be understood as a transition from a mean-driven to a fluctuation-driven regime [68]. A quantitative definition of gm can be given by requir- ing that the average input current of the active neurons ¯µA crosses the threshold θ at g = gm, namely ¯µA(gm) = ¯IA − gm¯νmnAm = θ ; where ¯IA is the average excitability of the firing neurons, while nAm and ¯νm are the number of active neurons and the average frequency at the minimum. For an uniform distribution P (I), a simpler expression for gm can be derived, namely gm = ¯ν−1 m (cid:20) l2 − θ nAm + 1 2 (l1 − l2)(cid:21) . (15) 1 a) 0.9 nA 0.8 0.7 0.6 0.1 400 300 x e d n i n o r u e n 200 100 0 0 2 h) k) 1.5 P(ν) 1 0.5 0 0 1 g 10 20 t ν 30 10 400 300 200 100 i) 40 0 0 l) 5 4 3 2 1 9 1 CV 0.5 0 1 CV 0.5 0 1 CV 0.5 0 0.6 0.4 0.2 0.6 0.4 0.2 0.6 0.4 0.2 ν ν ν b) e) c) f) d) g) 100 0 0.1 1 g 0.1 1 10 g 10 400 300 200 100 10 20 t ν 30 j) 40 0 0 10 20 t ν 30 40 m) 20 15 10 5 0 0 0.1 0.2 0.3 ν 0.4 0.5 0.2 0.4 ν 0.6 0.8 0 0 0.1 0.2 0.3 0.4 ν 0.5 0.6 FIG. 7. a) Fraction of active neurons nA as a function of inhibition for several values of K. b-d) Average network's firing rate ¯ν for the same cases depicted in a), and the corresponding CV (e-f). In all panels, filled symbols correspond to numerical data and dashed lines to semi-analytic values: black circles correspond to K = 20 (ts/τ0 = 11), red squares to K = 40 (ts/τ0 = 19) and blue diamonds to K = 80 (ts/τ0 = 26.6). The data are averaged over a time interval tS = 1 × 106 and 10 different realizations of the random network. h-j) Raster plots for three different synaptic strengths for N = 400 and K = 40: namely, h) g = 0.1; i) g = 1 and j) g = 8. The corresponding value for the fraction of active neurons, average frequency and average coefficient of variation are nA = (0.94, 0.76, 0.88), ¯ν = (0.55, 0.34, 0.10) and CV = (0.04, 0.27, 0.76), respectively. The neurons are ordered in terms of their intrinsic excitability and the time is rescaled by the average frequency ¯ν. k-l) Probability distributions P (ν) of the the single neuron firing rate ν, for the same values of g in the panels above. Empty-discontinuous bars correspond to the theoretical prediction while the filled bars indicate the histogram calculated with the simulation. The remaining parameters as in Fig. 1 (b). We have compared the numerical measurements of gm with the estimations obtained by employing Eq. (15), where nAm and ¯ν are obtained from the simulations. As shown in Fig. 8 for a network of size N = 1, 400, the overall agreement is more than satisfactory for in- degrees ranging over almost two decades (namely, for 20 ≤ K ≤ 600). This confirms that our guess (that the minimum nAm occurs exactly at the transition from mean-driven to fluctuation-driven dynamics) is consis- tent with the numerical data for a wide range of in- degrees. It should be stressed that, as we have verified for var- ious system sizes (namely, N = 700,1400 and 2800) and for a constant average in-degree K = 140, for instanta- neous synapses the network is in an heterogeneous asyn- chronous state for all the considered values of the synap- tic coupling. This is demonstrated by the fact that the intensity of the fluctuations of the average firing activ- ity, measured by considering the low-pass filtered linear super-position of all the spikes emitted in the network, vanishes as 1/√N [85]. Therefore, the observed transi- tion at g = gm is not associated to the emergence of irreg- nA 1 0.8 0.6 0.4 0.2 0 0.1 1 g 10 gm 8 6 4 2 10 100 K 1000 FIG. 8. gm as a function of the in-degree K. The symbols refer to numerical data, while the dashed line to the expression (15). Inset: nA versus g for the fully coupled case (solid black line) and for diluted networks (dashed lines), from top to bottom K = 20, 40, 80, 140. A network of size N = 1400 is evolved during a period tS = 1 × 105 after discarding a transient of 106 spikes, the data are averaged over 5 different random realizations of the network. Other parameters as in Fig. 1. ular collective behaviours as reported for globally coupled heterogeneous inhibitory networks of LIF neurons with delay [42] and of pulse-coupled phase oscillators [82]. V. EFFECT OF SYNAPTIC FILTERING In this Section we will investigate how synaptic fil- tering can influence the previously reported results. In particular, we will consider non instantaneous IPSP with α-function profile (3), whose evolution is ruled by a single time scale τα. A. Fully Coupled Networks Let us first examine the fully coupled topology, in this case we observe analogously to the δ-pulse coupling that by increasing the inhibition, the number of active neu- rons steadily decreases towards a limit where only few neurons (or eventually only one) will survive. At the same time the average frequency also decreases mono- tonically, as shown in Fig. 9 for two different τα differ- ing by almost two orders of magnitude. Furthermore, the mean field estimations (7) and (8) obtained for nA and ¯ν represent a very good approximation also for α- pulses (as shown in Fig. 9). In particular, the mean field estimation essentially coincides with the numeri- cal values for slow synapses, as evident from the data reported in Fig. 9 for τα = 10 (black filled circles). This can be explained by the fact that for sufficiently slow synapses, with τP > ¯TISI, the neurons feel the synaptic input current as continuous, because each in- 10 put pulse has essentially no time to decay between two firing events. Therefore the mean field approximation for the input current (4) works definitely well in this case. This is particularly true for τα = 10, where τP = 20 and ¯TISI ≃ 2 − 6 in the range of the considered cou- pling. While for τα = 0.125, we observe some deviation from the mean field results (red squares in Fig. 9) and the reason for these discrepancies reside in the fact that τP < ¯TISI for any coupling strength, therefore the dis- creteness of the pulses cannot be completely neglected in particular for large amplitudes (large synaptic couplings) analogously to what observed for instantaneous synapses. B. Sparse Networks For the sparse networks nA has the same qualitative behaviour as a function of the synaptic inhibition ob- served for instantaneous IPSPs, as shown in Fig. 9 (b) and Fig. 10 (a). The value of nA decreases with g and reaches a minimal value at gm, afterwards it increases towards nA = 1 at larger coupling. The origin of the minimum of nA as a function of g is the same as for in- stantaneous synapses, for g < gm the active neurons are subject on average to a supra-threshold effective input ¯µA, while at larger coupling ¯µA < θ, as shown in the inset of Fig. 10 (b). This is true for any value of τα, however, this transition from mean- to fluctuation-driven becomes dramatic for slow synapses . As evidenced from the data for the average output firing rate ¯ν and the average co- efficient of variation CV , these quantities have almost discontinuous jumps at g = gm, as shown in Fig. 11 . Therefore, let us first concentrate on slow synapses with τα larger than the membrane time constant, which is one for adimensional units. For g < gm the fraction of active neurons is frozen in time, at least on the con- sidered time scales, as revealed by the data in Fig. 9 (b). Furthermore, for g < gm, the mean field approximation obtained for the fully coupled case works almost perfectly both for nA and ¯ν, as reported in Fig. 10 (a). The frozen phase is characterized by extremely small values of the current fluctuations ¯σ (as shown Fig. 10 (b)) and a quite high firing rate ¯ν ≃ 0.4 − 0.5 with an associated average coefficient of variation CV almost zero (see black circles and red squares in Fig. 11). Instead, for g > gm the num- ber of active neurons increases in time similarly to what observed for the instantaneous synapses, while the aver- age frequency becomes extremely small ¯ν ≃ 0.04 − 0.09 and the value of the coefficient of variation becomes def- initely larger than one. These effects can be explained by the fact that, below gm the active neurons (the winners) are subject to an effective input ¯µA > θ that induces a quite regular firing, as testified by the raster plot displayed in Fig. 10 (c). The supra-threshold activity of the winners joined to- gether with the filtering action of the synapses guarantee that on average each neuron in the network receive an al- most continuous current, with small fluctuations in time. 1 a) nA 0.8 0.6 0.4 0.2 0 0.6 0.4 0.2 ν 0 0.1 1 g 10 11 1 b) nA 0.8 0.6 0.4 0.2 0 0.1 1 g 10 0.1 1 g 10 FIG. 9. Fraction of active neurons nA as a function of the inhibition with IPSPs with α-profile for a fully coupled topology (a) and a sparse network (b) with K = 20. a) Black (red) symbols correspond to τα = 10 (τα = 0.125), while the dashed lines are the theoretical predictions (7) and (8) previously reported for instantaneous synapses. The data are averaged over a time window tS = 1 × 105. Inset: average frequency ¯ν as a function of g. b) nA is measured at successive times : from lower to upper curve the considered times are tS = {1000, 5000, 10000, 50000, 100000}, while τα = 10. The system size is N = 400 in both cases, the distribution of excitabilities is uniform with [l1, l2] = [1.0, 1.5] and θ = 1. These results explain why the mean field approximation still works in the frozen phase, where the fluctuations in the synaptic currents are essentially negligible and un- able to induce any neuron's rebirth, at least on realistic time scales. In this regime the only mechanism in action is the WTA, fluctuations begin to have a role for slow synapses only for g > gm. Indeed, as shown in Fig. 10 (b), the synaptic fluctuations ¯σ for τα = 10 (black circles) are almost negligible for g < gm and show an enormous increase at g = gm of almost two orders of magnitude. Similarly at τα = 2 (red square) a noticeable increase of ¯σ is observable at the transition. In order to better understand the abrupt changes in ¯ν and in CV observable for slow synapses at g = gm, let us consider the case τα = 10. As shown in Fig. 11 (c), τP > ¯TISI ≃ 2 − 3 for g < gm, therefore for these couplings the IPSPs have no time to decay between a firing emission and the next one, thus the synaptic fluc- tuations ¯σ are definitely small in this case, as already shown. At gm an abrupt jump is observable to large val- ues ¯TISI > τP , this is due to the fact that now the neu- rons display bursting activities, as evident from the raster plot shown in Fig. 10 (d). The bursting is due to the fact that, for g > gm, the active neurons are subject to an effective input which is on average sub-threshold, there- fore the neurons preferentially tend to be silent. How- ever, due to current fluctuations a neuron can pass the threshold and the silent periods can be interrupted by bursting phases where the neuron fires almost regularly. As a matter of fact, the silent (inter-burst) periods are very long ≃ 700 − 900, if compared to the duration of the bursting periods, namely ≃ 25 − 50, as shown in Fig. 11 (c). This explains the abrupt decrease of the average firing rate reported in Fig. 11 (a). Furthermore, the inter-burst periods are exponentially distributed with an associated coefficient of variation ≃ 0.8 − 1.0, which clearly indicates the stochastic nature of the switching from the silent to the bursting phase. The firing periods within the bursting phase are instead quite regular, with an associated coefficient of variation ≃ 0.2, and with a duration similar to ¯TISI measured in the frozen phase (shaded gray circles in Fig. 11 (c)). Therefore, above gm the distribution of the ISI exhibits a long exponential tail associated to the bursting activity and this explains the very large values of the measured coefficient of vari- ation. By increasing the coupling the fluctuations in the input current becomes larger thus the fraction of neu- rons that fires at least once within a certain time interval increases. At the same time, ¯ν, the average inter-burst periods and the firing periods within the bursting phase remain almost constant at g > 10, as shown in Fig. 11 (a). This indicates that the decrease of ¯µA and the increase of ¯σ due to the increased inhibitory coupling essentially compensate each other in this range. Indeed, we have verified that for τα = 10 and τα = 2 ¯µA (¯σ) decreases (increases) linearly with g with similar slopes, namely ¯µA ≃ 0.88 − 0.029g while ¯σ ≃ 0.05 + 0.023g. For faster synapses, the frozen phase is no more present. Furthermore, due to rebirths induced by current flutuations, nA is always larger than the fully coupled mean field result (8), even at g < gm. It is interesting to notice that by decreasing τα, we are now approaching the instantaneous limit, as indicated by the results reported for nA in Fig. 10 (a) and CV in Fig. 11 (b). In particular, for τα = 0.125 (green triangles) the data almost collapse on the corresponding values measured for instantaneous synapses in a sparse networks with the same characteris- tics and over a similar time interval (dashed line). Fur- thermore, for fast synapses with τα < 1 the bursting ac- tivity is no more present as it can be appreciated by the fact that at most CV approaches one in the very large coupling limit. For sufficiently slow synapses, the average firing rate ¯ν can be estimated analytically by applying the so-called adiabatic approach developed by Moreno-Bote and Parga in [50, 51]. This method applies to LIF neurons with a synaptic time scale definitely longer than the membrane time constant. In these conditions, the output firing rate can be reproduced by assuming that the neuron is sub- ject to an input current with time correlated fluctuations, which can be represented as colored noise with a correla- tion time given by the pulse duration τP = 2τα (for more details see Appendix D). In this case we are unable to develop a self-consistent approach to obtain at the same time nA and the average frequency. However, once nA is provided by simulations the analytic estimated (45) obtained with the adiabatic approach gives very good agreement with the numerical data for sufficiently slow synapses, namely for τP ≥ 1, as shown in Fig. 11(a) for τα = 10, 2 and 0.5. The theoretical expression (45) is even able to reproduce the jump in the average frequen- cies observable at gm and therefore to capture the burst- ing phenomenon. By considering τP < 1, as expected, the theoretical expression fails to reproduce the numer- ical data in particular at large coupling (see the dashed green line in Fig. 11(a) corresponding to τα = 0.125). By following the arguments reported in [50], the burst- ing phenomenon observed for τα > 1 and g > gm can be interpreted at a mean field level as the response of a sub-threshold LIF neuron subject to colored noise with correlation τP . In this case, the neuron is definitely sub- threshold, but in presence of a large fluctuation it can be lead to fire and due to the finite correlation time it can remain supra-threshold regularly firing for a period ≃ τP . The validity of this interpretation is confirmed by the fact that the measured average bursting periods are of the order of the correlation time τP = 2τα, namely, ≃ 27 − 50 (≃ 7 − 14) for τα = 10 (τα = 2). As a final point, to better understand the dynamical origin of the measured fluctuations in this deterministic model, we have estimated the maximal Lyapunov expo- nent λ. As expected from previous analysis, for non- instantaneous synapses we can observe the emergence of regular chaos in purely inhibitory networks [1, 31, 89]. In particular, for sufficiently fast synapses, we typically note a transition from a chaotic state at low coupling to a linearly stable regime (with λ < 0) at large synap- tic strengths, as shown in Fig. 12 (a) for τα = 0.125. Despite the fact that current fluctuations are monoton- ically increasing with the synaptic strength. Therefore, fluctuations are due to chaos, at small coupling, while at larger g they are due to finite amplitude instabilities, as expected for stable chaotic systems [3]. However, the passage from positive to negative values of the maximal Lyapunov exponent is not related to the transition oc- curring at gm from a mean-driven to a fluctuation-driven dynamics in the network. For slow synapses, λ is essentially zero at small cou- pling in the frozen phase characterized by tonic spiking 12 of the neurons, while it becomes positive by approach- ing gm. For larger synaptic strengths λ, after reaching a maximal value, decreases and it can become eventually negative at g >> gm, as reported in Fig. 12 (b-c). Only for extremely slow synapses, as shown in Fig. 12 (c) for τα = 10, the chaos onset seems to coincide with the tran- sition occurring at gm. These findings are consistent with recent results concerning the emergence of asynchronous rate chaos in homogeneous inhibitory LIF networks with deterministic [27] and stochastic [32] evolution. However, a detailed analysis of this aspect goes beyond the scope of the present paper. VI. DISCUSSION In this paper we have shown that the effect reported in [1, 66] is observable whenever two sources of quenched disorder are present in the network: namely, a random distribution of the neural properties and a random topol- ogy. In particular, we have shown that neuron's death due to synaptic inhibition is observable only for hetero- geneous distributions of the neural excitabilities. Fur- thermore, in a globally coupled network the less excitable neurons are silenced for increasing synaptic strength un- til only one or few neurons remain active. This scenario corresponds to the winner-takes-all mechanism via lateral inhibition, which has been often invoked in neuroscience to explain several brain functions [88]. WTA mechanisms have been proposed to model hippocampal CA1 activ- ity [16], as well as to be at the basis of visual velocity estimate [25], and to be essential for controlling visual attention [29]. However, most brain circuits are characterized by sparse connectivity [10, 39, 60], in these networks we have shown that an increase in inhibition can lead from a phase dominated by neuronal death to a regime where neuronal rebirths take place. Therefore the growth of in- hibition can have the counter-intuitive effect to activate silent neurons due to the enhancement of current fluctu- ations. The reported transition is characterized by a pas- sage from a regime dominated by the almost tonic activ- ity of a group of neurons, to a phase where sub-threshold fluctuations are at the origin of the irregular firing of large part of the neurons in the network. For instan- taneous synapses, the average first and second moment of the firing distributions have been obtained together with the fraction of active neurons within a mean-field approach, where the neuronal rebirth is interpreted as an activation process driven by synaptic shot noise [70]. For a finite synaptic time smaller than the character- istic membrane time constant, the scenario is similar to the one observed for instantaneous synapses. However, the transition from mean-driven to fluctuation-driven dy- namics becomes dramatic for sufficiently slow synapses. In this situation one observes for low synaptic strength a frozen phase, where the synaptic filtering washes out the current fluctuations leading to an extremely regular dy- 1 0.8 0.6 0.4 0.2 nA 0 0.1 400 300 x e d n τα = 10 τα = 2 τα = 0.5 τα = 0.125 1 g 10 a) 1 σ 0.01 0.0001 0.1 400 300 x e d n 13 b) 1.2 µ A 1 0.8 0.1 1 g 10 1 g 10 100 i n o r u e n 200 100 i n o r u e n 200 100 0 0 20 40 c) 60 80 100 t ν 0 0 20 40 60 80 t ν d) 100 FIG. 10. a) Fraction of active neurons for a network of α-pulse coupled neurons as a function of g for various τα: namely, τα = 10 (black circles), τα = 2 (red squares), τα = 0.5 (blue diamonds) and τα = 0.125 (green triangles). For instantaneous synapses, the fully coupled analytic solution is reported (solid line), as well as the measured nA for the sparse network with same level of dilution and estimated over the same time interval (dashed line). b) Average fluctuations of the synaptic current ¯σ versus g for ISPS with α-profile the symbols refer to the same τα as in panel (a). Inset: Average input current ¯µA of the active neurons vs g, the dashed line is the threshold value θ = 1. The simulation time has been fixed to tS = 1 × 105. c-d) Raster plots for two different synaptic strengths for τα = 10: namely, c) g = 1 corresponds to nA ≃ 0.52, ¯ν ≃ 0.45 and CV ≃ 3 × 10−4; while i) g = 10 to nA ≃ 0.99, ¯ν ≃ 0.06 and CV ≃ 4.1. The neurons are ordered according to their intrinsic excitability and the time is rescaled by the average frequency ¯ν. The data have been obtained for a system size N = 400 and K = 20, other parameters as in Fig. 9. a) 0.6 0.5 0.4 ν 0.3 0.2 0.1 0 0.1 τα = 10 τα = 2 τα = 0.5 τα = 0.125 1 g 10 CV b) 5 4 3 2 1 0 0.1 τ P c) 1000 100 10 1 0.1 1 g 10 1 g 10 FIG. 11. a) Average firing rate ¯ν vs g for a network of α-pulse coupled neurons, for four values of τα. Theoretical estimations for ¯ν calculated with the adiabatic approach (45) are reported as dashed lines of colors corresponding to the relative symbols. b) Average coefficient of variation CV for four values of τα as a function of the inhibition. The dashed line refers to the values obtained for instantaneous synapses and a sparse network with the same value of dilution. c): Average inter-spike interval ¯TISI (filled black circles) as a function of g for τα = 10. For g > gm the average inter-burst interval (empty circles) and the average ISI measured within bursts (gray circles) are also shown, together with the position of gm (green veritical line). The symbols and colors denote the same τα values as in Fig. 10. All the reported data were calculated for a system size N = 400 and K = 20 and for a fixed simulation time of tS = 1 × 105. 14 0.1 a) λ 0 -0.1 -0.2 b) 0.04 λ 0.02 0 c) 0.006 λ 0.004 0.002 0.1 1 10 g -0.02 0.1 100 1 10 g 100 0 0.1 1 g 10 100 FIG. 12. Maximal Lyapunov exponent λ versus g for a network of α-pulse coupled neurons, for τα = 0.125 (a), τα = 2 (b) and τα = 10 (c). The blue dashed vertical line denote the gm value. All the reported data were calculated for a system size N = 400 and K = 20 and for simulation times 5 × 104 ≤ tS ≤ 7 × 105 ensuring a good convergence of λ to its asymptotic value. The other parameters are as in Fig. 9. namics controlled only by a WTA mechanism. As soon as the inhibition is sufficiently strong to lead the active neurons below threshold, the neuronal activity becomes extremely irregular exhibiting long silent phases inter- rupted by bursting events. The origin of these bursting periods can be understood in terms of the emergence of correlations in the current fluctuations induced by the slow synaptic timescale, as explained in [50]. In our model, the random dilution of the network con- nectivity is a fundamental ingredient to generate current fluctuations, whose intensity is controlled by the aver- age network in-degree K. A natural question is whether the reported scenario will be still observable in the ther- modynamic limit. On the basis of previous studies we can affirm that this depends on how K scales with the system size [23, 41, 75]. In particular, if K stays finite for N → ∞ the transition will still be observable. For K diverging with N , the fluctuations become negligible for sufficiently large system sizes, impeding neuronal re- births and the dynamics will be controlled only by the WTA mechanism. An additional source of randomness present in the net- work is related to the variability in the number of ac- tive pre-synaptic neurons. In our mean-field approach we have assumed that each neuron is subject to nAK spike trains, however this is true only on average. The num- ber of active pre-synaptic neurons is a random variable binomially distributed with average nAK and variance nA(1 − nA)K. Future developments of the theoretical approach here reported should include also such variabil- ity in the modeling of the network dynamics [9]. As a further aspect, we show that the considered model is not chaotic for instantaneous synapses, in such a case we observe irregular asynchronous states due to stable chaos [63]. The system can become truly chaotic only for finite synaptic times [3, 31]. However, we report clear indications that for synapses faster than the mem- brane time constant τm the passage from mean-driven to fluctuation-driven dynamics is not related to the on- set of chaos. Only for extremely slow synapses we have numerical evidences that the appearance of the bursting regime could be related to a passage from a zero Lya- punov exponent to a positive one, in agreement with the results reported in [27, 32] for homogeneous inhibitory networks. These preliminary indications demand for fu- ture, more detailed investigations of deterministic spiking networks in order to relate fluctuation-driven regime and chaos onsets. Moreover, we expect that it will be hard to distinguish whether the erratic current fluctuations are due to regular chaos or stable chaos on the basis of the analysis of the network activity, as also pointed out in [31]. For what concerns the biological relevance of the pre- sented model, we can attempt a comparison with exper- imental data obtained for MSNs in the striatum. This population of neurons is fully inhibitory with sparse lat- eral connections (connection probability ≃ 10− 20% [77, 81]), unidirectional and relatively weak [78]. Further- more, for MSNs within the same collateral network the axonal propagation delays are quite small ≃ 1− 2 ms [76] and they can be safely neglected. The dynamics of these neurons in behaving mice, reveals a low average firing rate with irregular firing activity (bursting) with asso- ciated large coefficient of variation [47]. As we have shown, these features can be reproduced by sparse net- works of LIF neurons with sufficiently slow synapses at g > gm and τα > τm. For values of the membrane time constant which are comparable to the ones mea- sured for MSNs [59, 61] (namely, τm ≃ 10 − 20 msec), the model is able to capture even quantitatively some of the main aspects of the MSNs dynamics, as shown in Table I. We obtain a reasonable agreement with the experiments for sufficiently slow synapses, where the in- teraction among MSNs is mainly mediated by GABAA receptors, which are characterized by IPSP durations of the order of ≃ 5 − 20 ms [35, 81]. However, apart the burst duration, which is definitely shorter, all the other aspects of the MSN dynamics can be already captured for τα = 2τm (with τm = 10 ms) as shown in Table I. Therefore, we can safely affirm, as also suggested in [66], that the fluctuation driven regime emerging at g > gm is the most appropriate in order to reproduce the dynamical 15 τα/τm τm (msec) Spike Rate (Hz) CV Burst Duration (msec) Spike Rate within Bursts (Hz) 2 10 10 20 10 20 4-6 2-3 4-6 2-3 ≃ 1.8 ≃ 1.8 ≃ 4.2 ≃ 4.2 100 ± 40 200 ± 80 400 ± 150 800 ± 300 42 ± 2 21 ± 1 41 ± 2 20 ± 1 Experimental data 2-3 ≃ 1.5 − 3 500 − 1100 31 ± 15 TABLE I. Comparison between the results obtained for slow α- synapses and experimental data for MSNs.The numerical data refer to results obtained in the bursting phase, namely for synaptic strenght g in the range [10 : 50], for simulation times tS = 1 × 105, N = 400 and K = 20. The experimental data refer to MSNs population in striatum for free behaving wild type mice, data taken from [47]. evolution of this population of neurons. Other inhibitory populations are present in the basal ganglia. In particular two coexisting inhibitory popula- tions, arkypallidal (Arkys) and prototypical (Protos) neu- rons, have been recently discovered in the external globus pallidus [43]. These populations have distinct physiolog- ical and dynamical characteristics and have been shown to be fundamental for action suppression during the per- formance of behavioural tasks in rodents [44]. Protos are characterized by a high firing rate ≃ 47 Hz and a not too large coefficient of variation (namely, CV ≃ 0.58) both in awake and slow wave sleep (SWS) states, while Arkys have a clear bursting dynamics with CV ≃ 1.9 [18, 44]. Furthermore, the firing rate of Arkys is definitely larger in the awake state (namely, ≃ 9 Hz) with respect to the SWS state, where the firing rates are ≃ 3 − 5 Hz [44]. On the basis of our results, on the one hand Protos can be modeled as LIF neurons with reasonable fast synapses in a mean driven regime, namely with a synaptic cou- pling g < gm. On the other hand, Arkys should be char- acterized by IPSP with definitely longer duration and they should be in a fluctuation driven phase as suggested from the results reported in Fig. 11. Since, as shown in Fig. 11 (a), the firing rate of inhibitory neurons de- crease by increasing the synaptic strenght g we expect that the passage from awake to slow wave sleep should be characterized by a reinforcement of Arkys synapses. Our conjectures about Arkys and Protos synaptic properties based on their dynamical behaviours ask for for experi- mental verification, which we hope will happen shortly. Besides the straightforward applicability of our find- ings to networks of pulse-coupled oscillators [48], it has been recently shown that LIF networks with instan- taneous and non-instantaneous synapses can be trans- formed into the Kuramoto-Daido model [17, 36, 62]. Therefore, we expect that our findings should extend to phase oscillator arrays with repulsive coupling [79]. This will allow for a wider applicability of our results, due to the relevance of limit-cycle oscillators not only for model- ing biological systems [86], but also for the many scientific and technological applications [19, 58, 71, 74]. ACKNOWLEDGMENTS Some preliminary analysis on the instantaneous synapses has been performed in collaboration with A. Imparato, the complete results will be reported else- where [54]. We thank for useful discussions J. Berke, B. Lindner, G. Mato, G. Giacomelli, S. Gupta, A. Politi, MJE Richardson, R. Schmidt, M.Timme. This work has been partially supported by the European Commis- sion under the program "Marie Curie Network for Initial Training", through the project N. 289146, "Neural Engi- neering Transformative Technologies (NETT)" (D.A.-G., S.O., and A.T), by the A∗MIDEX grant (No. ANR-11- IDEX-0001-02) funded by the French Government "pro- gram Investissements d'Avenir" (D.A.-G. and A.T.), and by "Departamento Adminsitrativo de Ciencia Tecnologia e Innovacion - Colciencias" through the program "Doc- torados en el exterior - 2013" (D.A.-G.). This work has been completed at Max Planck Institute for the Physics of Complex Systems in Dresden (Germany) as part of the activity of the Advanced Study Group 2016 entitled "From Microscopic to Collective Dynamics in Neural Cir- cuits". APPENDIX A: EVENT DRIVEN MAPS By following [53, 91] the ordinary differential equations (1) and (2) describing the evolution of the membrane po- tential of the neurons can be rewritten exactly as discrete time maps connecting successive firing events occurring in the network. In the following we will report explicitly such event driven maps for the case of instantaneous and α synapses. For instantaneous PSPs, the event-driven map for neu- ron i takes the following expression: vi(n + 1) = vi(n)e−Tδ + Ii(1 − e−Tδ ) − g K Cmi , (16) where the sequence of firing times {tn} in the network is denoted by the integer indices {n}, m is the index of the neuron firing at time tn+1 and Tδ ≡ tn+1 − tn is the inter-spike interval associated with two successive neuronal firing. This latter quantity is calculated from the following expression: Tδ = log(cid:20) Im − vm Im − 1 (cid:21) . (17) For α-pulses, the evolution of the synaptic current Ei(t), stimulating the i-th neuron can be expressed in terms of a second order differential equation, namely Ei(t)+2α Ei(t)+α2Ei(t) = α2 K Xj6=i Xntn<t Cij δ(t−tn) . (18) Eq. (18) can be rewritten as two first order differential equations by introducing the auxiliary variable Q ≡ Ei− αEi, namely α2 K Xntn<t Qi = −αQi + Ei = Qi − αEi, Cij δ(t − tn) (19) Finally, the equations (1) and (19) can be exactly in- tegrated from the time tn, just after the deliver of the n-th pulse, to time tn+1 corresponding to the emission of the (n + 1)-th spike, to obtain the following event driven map: Qi(n + 1) = Qi(n)e−αTα + α2 K Cmi (20a) Ei(n + 1) = Ei(n)e−αTα + Qi(n)Tαe−αTα (20b) vi(n + 1) = vi(n)e−Tα + Ii(1 − e−Tα) − gHi(n) , (20c) In this case, the inter-spike interval Tα ≡ tn+1 − tn should be estimated by solving self-consistently the fol- lowing expression Tα = ln(cid:20) Im − gHm(n) − 1(cid:21) , Im − vm(n) (21) where the explicit expression for Hi(n) appearing in equations (20c) and (21) is Hi(n) = − e−Tα − e−αTα α − 1 Tαe−αTα α − 1 Qi(n) . (cid:18)Ei(n) + Qi(n) α − 1(cid:19) 16 membrane potential (1) can be easily re-expressed in terms of dimensional variables as follows τm Vi(t) = Ii − Vj(t) − τmg Ei(t) i = 1,··· , N ; (23) where we have chosen τm = 10 ms as the membrane time constant, Ii represents the neural excitability and the external stimulations. Furthermore, t = t · τm, the field Ei = Ei/τm has the dimensionality of a frequency and g of a voltage. The currents { Ii} have also the dimen- sionality of a voltage, since they include the membrane resistance. For the other parameters/variables the transformation to physical units is simply given by (24) Vi = Vr + (Vth − Vr)vi Ii = Vr + (Vth − Vr)Ii g = (Vth − Vr)g (25) (26) where Vr = −60 mV and Vth = −50 mV are realistic values of the membrane reset and threshold potential. The isolated i-th LIF neuron is supra-threshold whenever Ii > Vth. APPENDIX B: AVERAGE FIRING RATE FOR INSTANTANEOUS SYNAPSES In this Appendix, by following the approach in [70] we derive the average firing rate of a supra-threshold LIF neuron subject to inhibitory synaptic shot noise of con- stant amplitude G, namely v(t) = I − v(t) − GX{tk} δ(t − tk) ; (27) where I > 1.The post-synaptic pulses reaching the neu- ron are instantaneous and their arrival times are Poisson- distributed and characterized by a rate R. In order to find the firing rate response of the LIF neuron we in- troduce the probability density P (v) and the flux J(v) associated to the membrane potentials, these satisfy the continuity equation: ∂P ∂t + ∂J ∂v = ρ(t)[δ(v − vr) − δ(v − θ)] ; (28) where ρ(t) is the instantaneous firing rate of the neuron. The flux can be decomposed in an average drift term plus the inhibitory part, namely J = (I − v)P + Jinh ∂Jinh(v, t) ∂v = R[P (v, t) − P (v − G, t)] ; (30) (29) The model so far introduced contains only adimen- sional units, however, the evolution equation for the J(θ, t) = ρ(t) Jinh(θ, t) = 0 P (θ, t) = 0 ; (22) The set of equations (28) to (30) is complemented by the boundary conditions: and by the requirement that membrane potential distri- bution should be normalized, i.e Z θ −∞ P (v, t)dv = 1 . By introducing bilateral Laplace transforms f (s) = −∞ dvesvf (v) and by performing some algebra along the lines described in [70] it is possible to derive the analytic expression for the average firing rate R ∞ 1 ν0 =Z ∞ 0 ds s esθ − esvr Z0(s) . (31) where Z0(s) is the Laplace transform of the sub-threshold probability density. Namely, it reads as Z0(s) = Ehs−ResI+RE(Gs)i ; (32) where E(y) = −R ∞ −y dte−t/t is the exponential integral, E = e−R(Γ+ln G) is the normalization constant ensuring that the distribution Z0(v) is properly normalized, and Γ is the Euler-Mascheroni constant. 17 In order to validate the method here outlined to obtain the average firing frequency ¯ν, we compare the theoreti- cal estimates given by (14) with numerical data obtained for sparse networks with in-degree K and instantaneous inhibitory synapses. The agreement is quite remarkable as shown in Fig. 13. In the same figure the solid magenta line refers to the results obtained by employing the diffu- sion approximation [8, 9, 69, 80]: clear discrepancies are evident already for g ≥ 1. In particular, for the diffusion approximation and the evaluation of (14) we assume that each neuron receives a Poissonian spike train with an ar- rival rate given by R = nA¯νK. Furthermore, it should be stressed that in this case we limit to test the quality of the approach described in this Appendix versus the diffusion approximation, therefore nA, required to esti- mate ¯ν, is obtained from the simulation and not derived self-consistently as done in Sect. IV. APPENDIX C: COEFFICIENT OF VARIATION FOR INSTANTANEOUS SYNAPSES In order to derive the coefficient of variation for the shot noise case it is necessary to obtain the first two mo- ments of the first-passage-time density q(t). By following the same approach as in Appendix B, the time-dependent continuity equation with initial condition P (v, 0) = vr is written as ∂P ∂t + ∂J ∂v = ρ(t)[δ(v − vr) − δ(v − θ)] + δ(t)δ(v − vr) . As suggested in [70], Eq. (33) can be solved by perform- ing a Fourier transform in time and a bilateral Laplace transform in membrane potential. This allows to obtain the Fourier transform of the spike-triggered rate, namely obtaining ρ(ω) ≃ n0 + n1ω + n2ω2 d0 + d1ω + d2ω2 ; where n0 = −1, d0 = 0, d1 = −in0/ν0 and (33) (37) (38) 0 ds siωA′(s) Z ∞ ds siω [B′(s) − A′(s)] ρ(ω) = Z ∞ 0 ; (34) where A(s) = esvr / Z0(s) and B(s) = esθ/ Z0(s). The Fourier transform of the first-passage-time density is q(ω) = ρ(ω) 1+ ρ(ω) and the first and second moment of the distribution are given by ∂ q ∂ωω=0 = −ihti ∂2 q ∂ω2ω=0 = −ht2i (35) (36) The integrals appearing in Eq. (34) cannot be exactly solved, therefore we have expanded it to the second order 0 n1 = iZ ∞ d2 =Z ∞ 0 log sA′(s)ds log s [B(s) − A(s)]ds . (39) s From the expression (34) we can finally obtain the first and second moment of q(t), namely hti = 1 ν0 , ht2i = 2[d2 1 + d2 + d1n1] . (40) Once these quantities are known the coefficient of vari- ation can be easily estimated for each neuron with ex- citability I. The results obtained for the average coefficient of vari- ation CV for a sparse network are compared with nu- merical data and with the diffusion approximation in the 2 1.5 V C 1 0.5 0 0.1 18 current z with an arbitrary distribution P (z) is given by ν0 ≃Z dzP (z)ν(z) (41) 1 g 10 where ν(z) is the input to rate transfer function of the neuron under a stationary input which for the LIF neuron is simply: 0.5 0.4 ν 0.3 0.2 0.1 0 0.1 1 g 10 FIG. 13. Average network's frequency ¯ν as a function of the synaptic strength g for instantaneous synapses and uniform distributions P (I) with support [l1, l2] = [1.0, 1.5]. Inset: av- erage coefficient of variation CV versus g. Filled symbols re- fer to numerical simulation for N = 400 and K = 20, dashed lines to the corresponding analytic solutions reported in Ap- pendices B and C and the solid (magenta) lines to the dif- fusion approximation. The data have been averaged over a time interval tS = 1 × 106 after discarding a transient of 106 spikes. inset of Fig. 13. It is evident that the approximation here derived is definitely more accurate than the diffusion ap- proximation for synaptic strengths larger than g ≃ 1. APPENDIX D: AVERAGE FIRING RATE FOR SLOW SYNAPSES In Section V we have examined the average activity of the network for non instantaneous IPSPs with α-function profiles. In presence of synaptic filtering, whenever the synaptic time constant is larger than the membrane time constant one can apply the so-called adiabatic approach to derive the firing rate ν0 of a single neuron, as described in [50, 51]. In this approximation, the output firing rate ν0 of the single neuron driven by a slowly varying stochastic input ν(z) =(cid:20)ln(cid:18) z − vr z − θ (cid:19)(cid:21)−1 . (42) The synaptic filtering induces temporal correlations in the input current z, which can be written as: h(z(t) − µ)(z(t′) − µ)i = σ2 2τs exp(cid:20)−t − t′ τs (cid:21) ; (43) here τs is the synaptic correlation time. In the case of α- pulses, where the rise and decay synaptic times coincide, we can assume that the correlation time is given by τs = τP = 2τα. Analogously to the diffusion approximation [8, 9, 14], the input currents are approximated as a Gaussian noise with mean µ and variance σ2 z = σ/2τs. In our network model, a single neuron receives an average current µ given by Eq. (4) with a standard deviation σ given by Eq. (10). In particular, the fraction of active neurons nA entering in the expressions of µ and σ is in this case obtained by the numerical simulations. Therefore, the single neuron output firing rate reads as ν0(I) =Z dz √2πσz e − (z−µ(I))2 2σ2 z (cid:20)ln(cid:18) z − vr z − θ (cid:19)(cid:21)−1 ;(44) where I is the neuronal excitability. The average firing rate of the LIF neurons in the net- work, characterized by an excitability distribution P (I), can be estimated as − (z−µ(I))2 2σ2 z dz √2πσz e (cid:20)ln(cid:18) z − vr z − θ (cid:19)(cid:21)−1 ¯ν =Z{IA} dIP (I)Zθ (45) where we impose the self-consistent condition that the average output frequency is equal to the average input one. [1] D Angulo-Garcia, Joshua D. Berke, and A Torcini. Cell assembly dynamics of sparsely-connected inhibitory for the collective activity networks: of striatal projection neurons. PLoS Comput Biol, 12(2):e1004778, 2015. a simple model [2] D Angulo-Garcia and A Torcini. Stochastic mean-field formulation of the dynamics of diluted neural networks. Physical Review E, 91(2):022928, 2015. [3] David Angulo-Garcia and Alessandro Torcini. Stable chaos in fluctuation driven neural circuits. Chaos, Soli- tons & Fractals, 69(0):233 – 245, 2014. [4] Giancarlo Benettin, Luigi Galgani, Antonio Giorgilli, and Jean-Marie Strelcyn. Lyapunov characteristic exponents for smooth dynamical systems and for hamiltonian sys- tems; a method for computing all of them. part 1: The- ory. Meccanica, 15(1):9–20, 1980. [5] Joshua D Berke, Murat Okatan, Jennifer Skurski, and Howard B Eichenbaum. Oscillatory entrainment of stri- atal neurons in freely moving rats. Neuron, 43(6):883– 896, 2004. [6] Paul C Bressloff and Stephen Coombes. Dynamics of strongly coupled spiking neurons. Neural computation, 12(1):91–129, 2000. [7] PC Bressloff and S Coombes. A dynamical theory of spike train transitions in networks of integrate-and-fire oscilla- tors. SIAM Journal on Applied Mathematics, 60(3):820– 841, 2000. [8] Nicolas Brunel. Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. J. Comput. Neurosci., 8(3):183–208, 2000. [9] Nicolas Brunel and Vincent Hakim. Fast global oscil- lations in networks of integrate-and-fire neurons with low firing rates. Neural. Comput., 11(1):1621–1671, mar 1999. [10] Ed Bullmore and Olaf Sporns. Complex brain networks: graph theoretical analysis of structural and functional systems. Nature Reviews Neuroscience, 10(3):186–198, 2009. [11] Anthony N Burkitt. A review of the integrate-and-fire neuron model: I. homogeneous synaptic input. Biological cybernetics, 95(1):1–19, 2006. [12] Gyorgy Buzs´aki and Andreas Draguhn. Neuronal oscilla- tions in cortical networks. Science, 304(5679):1926–1929, 2004. [13] Scott Camazine. Self-organization in biological systems. Princeton University Press, 2003. [14] RM Capocelli and LM Ricciardi. Diffusion approxima- tion and first passage time problem for a model neuron. Kybernetik, 8(6):214–223, 1971. [15] Luis Carrillo-Reid, Fatuel Tecuapetla, Dagoberto Tapia, Arturo Hern´andez-Cruz, Elvira Galarraga, Ren´e Drucker-Colin, and Jos´e Bargas. Encoding network states by striatal cell assemblies. Journal of neurophysi- ology, 99(3):1435–1450, 2008. [16] Robert Coultrip, Richard Granger, and Gary Lynch. A cortical model of winner-take-all competition via lateral inhibition. Neural networks, 5(1):47–54, 1992. [17] Hiroaki Daido. Onset of cooperative entrainment in limit- cycle oscillators with uniform all-to-all interactions: bi- furcation of the order function. Physica D: Nonlinear Phenomena, 91(1):24–66, 1996. [18] Paul D Dodson, Joseph T Larvin, James M Duffell, Farid N Garas, Natalie M Doig, Nicoletta Kessaris, Ian C Duguid, Rafal Bogacz, Simon JB Butt, and Peter J Mag- ill. Distinct developmental origins manifest in the spe- cialized encoding of movement by adult neurons of the external globus pallidus. Neuron, 86(2):501–513, 2015. [19] Florian Dorfler and Francesco Bullo. Synchronization in complex networks of phase oscillators: A survey. Auto- matica, 50(6):1539–1564, 2014. [20] Bard Ermentrout. Complex dynamics in winner-take- all neural nets with slow inhibition. Neural networks, 5(3):415–431, 1992. [21] Tomoki Fukai and Shigeru Tanaka. A simple neural net- work exhibiting selective activation of neuronal ensem- bles: from winner-take-all to winners-share-all. Neural computation, 9(1):77–97, 1997. [22] Tomoki Fukai and Shigeru Tanaka. A simple neural net- work exhibiting selective activation of neuronal ensem- bles: from winner-take-all to winners-share-all. Neural computation, 9(1):77–97, 1997. [23] D Golomb, D Hansel, and G Mato. Mechanisms of syn- chrony of neural activity in large networks. Handbook of biological physics, 4:887–968, 2001. 19 [24] David Golomb and John Rinzel. Clustering in globally coupled inhibitory neurons. Physica D: Nonlinear Phe- nomena, 72(3):259–282, 1994. [25] Norberto M Grzywacz and AL Yuille. A model for the es- timate of local image velocity by cells in the visual cortex. Proceedings of the Royal Society of London B: Biological Sciences, 239(1295):129–161, 1990. [26] Peter Hanggi, Peter Talkner, and Michal Borkovec. Reaction-rate theory: fifty years after kramers. Reviews of modern physics, 62(2):251, 1990. [27] Omri Harish and David Hansel. Asynchronous rate chaos in spiking neuronal circuits. PLoS Comput Biol, 11(7):e1004266, 2015. [28] Ronald M Harris-Warrick. Dynamic biological networks: the stomatogastric nervous system. MIT press, 1992. [29] Laurent Itti and Christof Koch. Computational mod- elling of visual attention. Nature reviews neuroscience, 2(3):194–203, 2001. [30] Sven Jahnke, Raoul-Martin Memmesheimer, and Marc Timme. Stable irregular dynamics in complex neural net- works. Phys. Rev. Lett., 100:048102, Jan 2008. [31] Sven Jahnke, Raoul-Martin Memmesheimer, and Marc Timme. How chaotic is the balanced state? Front. Comp. Neurosci., 3(13), Nov 2009. [32] Jonathan Kadmon and Haim Sompolinsky. Transition to chaos in random neuronal networks. Physical Review X, 5(4):041030, 2015. [33] BS Kerner and Vyacheslav Vladimirovich Osipov. Self- organization in active distributed media: scenarios for the spontaneous formation and evolution of dissipative structures. Physics-Uspekhi, 33(9):679–719, 1990. [34] MA Komarov, GV Osipov, and JAK Suykens. Sequen- tially activated groups in neural networks. EPL (Euro- physics Letters), 86(6):60006, 2009. [35] Tibor Koos, James M Tepper, and Charles J Wilson. Comparison of ipscs evoked by spiny and fast-spiking neurons in the neostriatum. The Journal of neuroscience, 24(36):7916–7922, 2004. [36] Yoshiki Kuramoto. Chemical oscillations, waves, and turbulence, volume 19. Springer Science & Business Me- dia, 2012. [37] Jos´e L Lanciego, Natasha Luquin, and Jos´e A Obeso. Functional neuroanatomy of the basal ganglia. Cold Spring Harbor perspectives in medicine, 2(12):a009621, 2012. [38] Daniel B Larremore, Woodrow L Shew, Edward Ott, Francesco Sorrentino, and Juan G Restrepo. Inhibition causes ceaseless dynamics in networks of excitable nodes. Physical review letters, 112(13):138103, 2014. [39] Simon B Laughlin and Terrence J Sejnowski. Commu- nication in neuronal networks. Science, 301(5641):1870– 1874, 2003. [40] Gilles Laurent. Olfactory network dynamics and the cod- ing of multidimensional signals. Nature Reviews Neuro- science, 3(11):884–895, 2002. [41] Stefano Luccioli, Simona Olmi, Antonio Politi, and Alessandro Torcini. Collective dynamics in sparse net- works. Phys. Rev. Lett., 109:138103, Sep 2012. [42] Stefano Luccioli and Antonio Politi. Irregular Collective Behavior of Heterogeneous Neural Networks. Phys. Rev. Lett., 105(15):158104+, October 2010. [43] Nicolas Mallet, Benjamin R Micklem, Pablo Henny, Matthew T Brown, Claire Williams, J Paul Bolam, Kouichi C Nakamura, and Peter J Magill. Dichotomous organization of the external globus pallidus. Neuron, 74(6):1075–1086, 2012. [44] Nicolas Mallet, Robert Schmidt, Daniel Leventhal, Fujun Chen, Nada Amer, Thomas Boraud, and Joshua D Berke. Arkypallidal cells send a stop signal to striatum. Neuron, 89(2):308–316, 2016. [45] Eve Marder and Dirk Bucher. Central pattern generators and the control of rhythmic movements. Current biology, 11(23):R986–R996, 2001. [46] Hans Meinhardt. Models of biological pattern formation, volume 6. Academic Press London, 1982. [47] Benjamin R Miller, Adam G Walker, Anand S Shah, Scott J Barton, and George V Rebec. Dysregulated infor- mation processing by medium spiny neurons in striatum of freely behaving mouse models of huntington's disease. Journal of neurophysiology, 100(4):2205–2216, 2008. [48] Renato E Mirollo and Steven H Strogatz. Synchroniza- tion of pulse-coupled biological oscillators. SIAM Journal on Applied Mathematics, 50(6):1645–1662, 1990. [49] Michael Monteforte and Fred Wolf. Dynamic flux tubes form reservoirs of stability in neuronal circuits. Phys. Rev. X, 2:041007, Nov 2012. [50] Rub´en Moreno-Bote and N´estor Parga. Role of synaptic filtering on the firing response of simple model neurons. Physical review letters, 92(2):028102, 2004. [51] Rub´en Moreno-Bote and N´estor Parga. Response of integrate-and-fire neurons to noisy inputs filtered by synapses with arbitrary timescales: Firing rate and cor- relations. Neural Computation, 22(6):1528–1572, 2010. [52] Thomas Nowotny and Mikhail I Rabinovich. Dynam- ical origin of independent spiking and bursting activ- ity in neural microcircuits. Physical review letters, 98(12):128106, 2007. [53] S Olmi, A Politi, and A Torcini. Linear stability in net- works of pulse-coupled neurons. Front. Comput. Neu- rosci., 8(8), 2014. [54] Simona Olmi, David Angulo-Garcia, Alberto Imparato, and Alessandro Torcini. The influence of synaptic weight distribution on the activity of balanced networks. preprint, 2016. [55] Simona Olmi, Roberto Livi, Antonio Politi, and Alessan- dro Torcini. Collective oscillations in disordered neural networks. Phys. Rev. E, 81(4 Pt 2), April 2010. [56] Srdjan Ostojic. Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience, 17(4):594–600, 2014. [57] Andr´e Parent and Lili-Naz Hazrati. Functional anatomy of the basal ganglia. i. the cortico-basal ganglia-thalamo- cortical loop. Brain Research Reviews, 20(1):91–127, 1995. [58] Arkady Pikovsky and Michael Rosenblum. Dynamics of globally coupled oscillators: Progress and perspectives. Chaos: An Interdisciplinary Journal of Nonlinear Sci- ence, 25(9):097616, 2015. [59] Henrike Planert, Thomas K Berger, and Gilad Silber- berg. Membrane properties of striatal direct and indirect pathway neurons in mouse and rat slices and their mod- ulation by dopamine. PloS one, 8(3):e57054, 2013. [60] Dietmar Plenz. When inhibition goes incognito: feedback interaction between spiny projection neurons in striatal function. Trends in neurosciences, 26(8):436–443, 2003. [61] Dietmar Plenz and Stephen T Kitai. Up and down states in striatal medium spiny neurons simultaneously recorded with spontaneous activity in fast-spiking in- 20 terneurons studied in cortex–striatum–substantia nigra organotypic cultures. The Journal of Neuroscience, 18(1):266–283, 1998. [62] Antonio Politi and Michael Rosenblum. Equivalence of phase-oscillator and integrate-and-fire models. Physical Review E, 91(4):042916, 2015. [63] Antonio Politi and Alessandro Torcini. Stable chaos. In Nonlinear Dynamics and Chaos: Advances and Perspec- tives, pages 103–129. Springer, 2010. [64] Adam Ponzi and Jeff Wickens. Sequentially switching cell assemblies in random inhibitory networks of spiking neurons in the striatum. The Journal of Neuroscience, 30(17):5894–5911, 2010. [65] Adam Ponzi and Jeff Wickens. Input dependent cell as- sembly dynamics in a model of the striatal medium spiny neuron network. Frontiers in systems neuroscience, 6, 2012. [66] Adam Ponzi and Jeffery R Wickens. Optimal balance of the striatal medium spiny neuron network. PLoS compu- tational biology, 9(4):e1002954, 2013. [67] Mikhail I Rabinovich and Pablo Varona. Robust tran- sient dynamics and brain functions. Frontiers in compu- tational neuroscience, 5:24, 2011. [68] Alfonso Renart, Rub´en Moreno-Bote, Xiao-Jing Wang, and N´estor Parga. Mean-driven and fluctuation-driven persistent activity in recurrent networks. Neural. Com- put., 19(1):1–46, 2007. [69] Luigi M Ricciardi. Diffusion processes and related top- ics in biology, volume 14. Springer Science & Business Media, 2013. [70] Magnus JE Richardson and Rupert Swarbrick. Firing- rate response of a neuron receiving excitatory and in- hibitory synaptic shot noise. Physical review letters, 105(17):178102, 2010. [71] Francisco A Rodrigues, Thomas K DM Peron, Peng Ji, and Jurgen Kurths. The kuramoto model in complex networks. Physics Reports, 610:1–98, 2016. [72] Yousheng Shu, Andrea Hasenstaub, and David A Mc- Cormick. Turning on and off recurrent balanced cortical activity. Nature, 423(6937):288–293, 2003. [73] Peter Somogyi, Gabor Tamas, Rafael Lujan, and Eber- hard H Buhl. Salient features of synaptic organisation in the cerebral cortex. Brain research reviews, 26(2):113– 135, 1998. [74] Steven H Strogatz. Exploring complex networks. Nature, 410(6825):268–276, 2001. [75] Lorenzo Tattini, Simona Olmi, and Alessandro Torcini. Coherent periodic activity in excitatory erdos-renyi neu- ral networks: the role of network connectivity. Chaos, 22(2):023133, jun 2012. [76] Stefano Taverna, Ema Ilijic, and D James Surmeier. Re- current collateral connections of striatal medium spiny neurons are disrupted in models of parkinson's disease. The Journal of neuroscience, 28(21):5504–5512, 2008. [77] Stefano Taverna, Yvette C Van Dongen, Henk J Groe- newegen, and Cyriel MA Pennartz. Direct physiological evidence for synaptic connectivity between medium-sized spiny neurons in rat nucleus accumbens in situ. Journal of neurophysiology, 91(3):1111–1121, 2004. [78] James M Tepper, Tibor Ko´os, and Charles J Wilson. Gabaergic microcircuits in the neostriatum. Trends in neurosciences, 27(11):662–669, 2004. [79] LS Tsimring, NF Rulkov, ML Larsen, and Michael Gab- bay. Repulsive synchronization in an array of phase os- 21 cillators. Physical review letters, 95(1):014101, 2005. [86] Arthur T Winfree. The geometry of biological time, vol- [80] Henry C Tuckwell. Introduction to theoretical neurobi- ology: Volume 2, nonlinear and stochastic theories, vol- ume 8. Cambridge University Press, 2005. [81] Mark J Tunstall, Dorothy E Oorschot, Annabel Kean, and Jeffery R Wickens. Inhibitory interactions between spiny projection neurons in the rat striatum. Journal of Neurophysiology, 88(3):1263–1269, 2002. [82] Ekkehard Ullner and Antonio Politi. Self-sustained irreg- ular activity in an ensemble of neural oscillators. Physical Review X, 6(1):011015, 2016. [83] Carl van Vreeswijk. Partial synchronization in pop- Phys. Rev. E, ulations of pulse-coupled oscillators. 54(5):5522–5537, November 1996. [84] Vladimir K Vanag, Lingfa Yang, Milos Dolnik, Anatol M Zhabotinsky, and Irving R Epstein. Oscillatory cluster patterns in a homogeneous chemical system with global feedback. Nature, 406(6794):389–391, 2000. [85] Xiao-Jing Wang and Gyorgy Buzs´aki. Gamma oscilla- tion by synaptic inhibition in a hippocampal interneu- ronal network model. The journal of Neuroscience, 16(20):6402–6413, 1996. ume 12. Springer Science & Business Media, 2001. [87] Yuan Xiong, Chuan-Hsiang Huang, Pablo A Iglesias, and Peter N Devreotes. Cells navigate with a local-excitation, global-inhibition-biased excitable network. Proceedings of the National Academy of Sciences, 107(40):17079–17086, 2010. [88] Alan L Yuille and Norberto M Grzywacz. A winner-take- all mechanism based on presynaptic inhibition feedback. Neural Computation, 1(3):334–347, 1989. [89] Rudiger Zillmer, Nicolas Brunel, and David Hansel. Very long transients, irregular firing, and chaotic dynamics in networks of randomly connected inhibitory integrate- and-fire neurons. Phys. Rev. E, 79:031909, Mar 2009. [90] Rudiger Zillmer, Roberto Livi, Antonio Politi, and Alessandro Torcini. Desynchronization in diluted neural networks. Phys. Rev. E, 74(3):036203, 2006. [91] Rudiger Zillmer, Roberto Livi, Antonio Politi, and Alessandro Torcini. Stability of the splay state in pulse- coupled networks. Phys. Rev. E, 76:046102, Oct 2007.
1906.02863
2
1906
2019-11-22T10:34:16
Double Generalized Linear Model Reveals Those with High Intelligence are More Similar in Cortical Thickness
[ "q-bio.NC", "stat.AP" ]
Most studies indicate that intelligence (g) is positively correlated with cortical thickness. However, the interindividual variability of cortical thickness has not been taken into account. In this study, we aimed to identify the association between intelligence and cortical thickness in adolescents from both the group's mean and dispersion point of view, utilizing the structural brain imaging from the Adolescent Brain and Cognitive Development (ABCD) Consortium, the largest cohort in early adolescents around 10 years old. The mean and dispersion parameters of cortical thickness and their association with intelligence were estimated using double generalized linear models(DGLM). We found that for the mean model part, the thickness of the frontal lobe like superior frontal gyrus was negatively related to intelligence, while the surface area was most positively associated with intelligence in the frontal lobe. In the dispersion part, intelligence was negatively correlated with the dispersion of cortical thickness in widespread areas, but not with the surface area. These results suggested that people with higher IQ are more similar in cortical thickness, which may be related to less differentiation or heterogeneity in cortical columns.
q-bio.NC
q-bio
Double Generalized Linear Model Reveals Those with High Intelligence are More Similar in Cortical Thickness Qi Zhao1, Lingli Zhang2, Chun Shen3, Jie Zhang3, Jianfeng Feng3 1. School of Mathematical Sciences, Fudan University, Shanghai, China, 2. Shanghai Children's Medical Center, Shanghai Jiao Tong University School of Medicine, Shanghai, China, 3. Institute of Science and Technology for Brain Inspired Intelligence, Fudan University, Shanghai, China Abstract Most studies indicate that intelligence (g) is positively correlated with cortical thickness. However, the interindividual variability of cortical thickness has not been taken into account. In this study, we aimed to identify the association between intelligence and cortical thickness in adolescents from both the group's mean and dispersion point of view, utilizing the structural brain imaging from the Adolescent Brain and Cognitive Development (ABCD) Consortium, the largest cohort in early adolescents around 10 years old. The mean and dispersion parameters of cortical thickness and their association with intelligence were estimated using double generalized linear models (DGLM). We found that for the mean model part, the thickness of the frontal lobe like superior frontal gyrus was negatively related to intelligence, while the surface area was most positively associated with intelligence in the frontal lobe. In the dispersion part, intelligence was negatively correlated with the dispersion of cortical thickness in widespread areas, but not with the surface area. These results suggested that people with higher IQ are more similar in cortical thickness, which may be related to less differentiation or heterogeneity in cortical columns. Keywords: cortical thickness, intelligence, crystallized intelligence, fluid intelligence, MRI, structural dispersion Abbreviations: intelligence=g, crystallized intelligence=gC, fluid intelligence=gF, gyrus=G, sulcus=S, frontal=front, parietal=pariet, occipital=oc, temporal=temp, cingulate=cingul, transverse=transv, hippocampal=hip, inferior=inf, superior = sup, medial = med, anterior = ant, posterior = post, long=lg Introduction Intelligence is the most heritable of all cognitive measures, and develops rapidly during adolescence. General intelligence (g) can be defined as the weighted sum of fluid (gF) and crystallized (gC) intelligence. gF refers to reasoning ability, and is related to working memory(Deary and Caryl 1997); while gC is sometimes described as verbal ability, and is more dependent on accumulated experience. Gray matter is closely related to intellectual ability, and it has been shown that intelligence and cortical thickness are partly associated through shared genes (Brans, Kahn et al. 2010). Recent neuroimaging studies have investigated the correlation between human intelligence and cortical thickness. For example, in young adults (20.9 ±2.9 years), g was strongly related to cortical thickness in regions of the temporal lobe (Choi, Shamosh et al. 2008). However, a longitudinal study (Shaw, Greenstein et al. 2006) described that there is a developmental shift from a predominantly negative correlation between intelligence and cortical gray matter thickness in early childhood (3.8 to 8.4 years) to a pronounced positive correlation in late childhood (8.6 to 11.7 years) in superior frontal gyri, medial prefrontal cortex, middle and orbitofrontal cortices. As the neuroanatomical expression of intelligence in children and adolescents is dynamic, it might help better illustrated this relationship in the participants with narrow age distribution to avoid developmental periods in which the brain changes radically. Previous studies focusing on the correlation between brain morphology and IQ only explored the association between group mean of neuroimaging measures and IQ, and did not model the interindividual variability, or dispersion in the group, which carry relevant information regarding gene- environment interactions related to the individual sensitivity to environmental and genetic perturbation(Alnaes, Kaufmann et al. 2019).Therefore, we first using double generalized linear model (DGLM) to model the variability between brain cortical thickness and intelligence in preadolescents. We utilized the structural brain magnetic resonance imaging from a large sample of the ABCD Study (Casey, Cannonier et al. 2018), with an narrow age span (around 10 years old) and identify the association both from correlation and variability using DGLM, which allows the mean and dispersion to be modelled simultaneously in a generalized linear model context(Smyth and Verbyla 1999). Meanwhile, the association between intelligence and surface area were also analyzed to compare with thickness results. Results Study samples Our study included 10,666 subjects after quality control for FreeSurfer v5.3.0 (N=11,076) and remove subjects with lack cognitive score. The samples include 22 sites; sex, age and mean thickness information were listed in the Table1. We also analyzed the correlation between different scores and covariables (FigS1 in Supplementary). Mean model results In the mean model, we model the relationship between cortical thickness and g in the mean part. Cortical thickness were predominantly negatively correlated with g in some frontal lobe and limbic lobe (Fig1-2), including G_front_sup, G_front_middle, S_suborbital, right S_orbital_med-olfact, G_and_S_cingul-Ant,S_pericallosal and S_temporal_transverse. Meanwhile, g were predominantly positively correlated with some frontal lobe (S_precentral-inf_part, G_and_S_subcentral, G_orbital), temporal regions (S_oc-temp_med_and_Lingual, S_temporal_inf, S_temporal_sup, Pole_temporal), parietal lobe (G_and _S_postcentral), limbic lobe (G_oc- temp_med-parahip) and major divisions (G_Ins_lg_and_S_cent_ins, S_calcarine, S_parieto_occipital). However, cortical surface area were globally positively associated with g, gC and gF (FigS4 in Supplementray). The prominent brain regions correlates were almost symmetrically distributed in g, gF and gC. However, gC were predominantly more associated with cortical thickness than gF in the mean model (gC: 109 regions > gF: 65 regions). Moreover, the part that gC differ from gF was mostly in left hemisphere (left: 17, right: 7). For example, left S_circular_insula_sup, left Pole_occipital, left S_temporal_inf, left G_front_inf-Orbital, left S_front_inf and left Lat_Fis-ant-Horizont. Dispersion model results In the dispersion model, we model the relationship between cortical thickness and g in the dispersion/variance part. Interestingly, g, gC and gF were all negatively associated with cortical thickness dispersion (Fig3-4). However, g, gC and gF were not associated with surface area dispersion. gF correlated more regions than gC in the dispersion model (gF: 25 regions > gC: 11 regions). gF was negatively associated with dispersion in frontal lobe (G_and_S_frontomargin, S_front_middle, S_precentral- inf-part and S_suborbital), parietal lobe (S_parieto_occipital, G_and_S_postcentral, G_and_S_intrapariet_and_P_trans, G_and_S_postcentral, G_precuneus, S_subparietal), occipital lobe (S_oc_middle_and_Lunatus, S_oc_sup_and_transversal, G_oc-temp_lat- fusifor, S_oc-temp_lat, S_oc_sup_and_transversal, S_oc_middle_and_Lunatus, G_occipital_sup, S_oc- temp_med_and_Lingual), S_pericallosal, G_and_S_cingul-Ant and circular sulcus of the Insula. gC was negatively associated with dispersion in temporal lobe (Pole_temporal, S_collat_transv_ant, S_temporal_inf), limbic lobe (G_and_S_cingul-Ant, G_and_S_cingul- Mid-Ant), G_precuneus, G_orbital, S_oc_middle_and_Lunatus and S_parieto_occipital. Besides, gC was more associated with dispersion in left-hemi cortex regions (left: 7, right: 4), and gF are more associated with dispersion in right-hemi cortex regions (left: 12, right: 13). Cross-Validated Elastic Net Regression results We set the α value to 0.5 to take advantage of the relative strengths of the two above regression approaches, providing a no sparse solution with low variance among several correlated independent variables. The cortical thickness, surface area respectively accounted for about 10%, both accounted for about 14% of the total variance of cognition total composite score age-corrected standard score. This 𝑅2 was significantly higher than expected due to chance (P <0.001, compared with 𝑅2 from 500 randomly generated elastic net regressions). Correlations between actual standard score versus predicted cognitive standard scores, averaging across 10 folds of the cross-validation, were gF r = 0.25, gC r = 0.29 and g r = 0.32 using cortical thickness(Fig.5), gF r = 0.24, gC r = 0.31 and g r = 0.31 using cortical surface area and gF r = 0.30, gC r = 0.36 and g r = 0.37 using both cortical thickness and surface area (FigS5 in the supplementray). Discussion The most important finding of the present work was that, higher intelligence was associated with lower interindividual heterogeneity in cortical thickness, for example in parieto-occipital sulcus, anterior part of the cingulate gyrus and sulcus. This means higher IQ population has lower variation than lower IQ population in some regions (FigS3), possibly reflecting higher IQ population are more similar in brain structure than lower IQ population. However, intelligence was not associated with brain heterogeneity in cortical surface area. Cortical thickness and surface area were both highly heritable but were essentially unrelated genetically (Panizzon, Fennema-Notestine et al. 2009). From neuronal point of view, cortical thickness is associated with radial neuronal migration and number of neurons, dendritic arborizations, and glial support in cortical columns, while surface area is related to tangential neuronal migration and captures of mini-columnar units in the cortex (Chenn and Walsh 2003, Rakic 2009, Rakic, Ayoub et al. 2009, Tadayon, Pascual-Leone et al. 2019). Therefore, our results suggested that in the early stages of development (around 10 years old), those with higher IQ had less differentiation in cortical columns, and that their brain morphology developed following a similar trajectory that leads to higher IQ. Interestingly, these areas overlap substantially with the default network. A recent review paper proposed that distributed association networks in default network are supported by anatomical connectivity(Buckner and DiNicola 2019). Thus, this finding might reveal the underlying relation between morphology and default network in the development process. From the dispersion model, we further found significant asymmetry of left and right hemispheres in terms of the correlation of their cortex thickness with gF/gC. gC is associated with dispersion in the left hemisphere regions; the fluid intelligence is associated with dispersion in right hemisphere Regions. This phenomenon may reflect that gF/gC can be studied from another perspective using the DGLM model. For the mean model, we found that higher intelligence was associated with a decrease in thickness in frontal lobe, but an increase in other areas like calcarine sulcus, lingual sulcus, parahippocampal gyrus and central sulcus, which means higher performance was associated with cortical thickness related to working memory, attention, and visio-spatial processing. Interestingly higher intelligence was associated with increase in surface area in almost the whole brain, most prominently in frontal cortex. The negative correlation between thickness and IQ in frontal lobe was largely in line with recent studies showing cortical gray matter thinning in anterior and superior frontal areas was associated with superior arithmetic performance to 9- and 10-year-old children (Chaddock-Heyman, Erickson et al. 2015). Shaw's longitudinal study showed that superior intelligence clusters demonstrated a marked increase in cortical thickness peaking in superior frontal gyri at around 11 years old, later than average intelligence group (Shaw, Greenstein et al. 2006). This provided an explanation for the negative correlation between thickness and IQ in frontal lobe. Interestingly, the most significant positive correlation between surface area and IQ was also in frontal lobe. Considering the theories that the first step in the evolutionary ascent of the human cerebral cortex is its enlargement, which occurs mainly by expansion of the surface area without a comparable increase in its thickness(Rakic 2009). Taken together, these results suggested that the frontal lobe surface area enlarge at first, and then thickness increases later for preadolescents with higher IQ. The morphological correlates of subitems of IQ revealed significant difference between gC and gF. For gC, its two cognitive domain scores, Picture Vocabulary and Oral Reading Recognition task scores exhibited very similar patterns of association with cortical thickness (Fig S6). However, the subitems in gF had different associated with cortical thickness. Therein, working memory are mostly associated with cortical thickness, picture sequence memory, cognitive flexibility (Dimensional Change Card Sort Task) and flanker are less associated with cortical thickness. Pattern comparison processing speed test is negative associated with cortical thickness in the right hemisphere anterior cingulate gyrus and sulcus, parieto-occipital sulcus and temporal inferior sulcus, which all distributed in default network related regions (FigS7). Using cortical thickness and surface area accounted for about 14% of the total variance of cognition total composite score age-corrected standard score, more than using both respectively. It means cortical thickness and surface area contribute different aspects to g. The findings, based on harmonized analysis protocols for all included data sets, were robust to strict procedures for removing outliers and quality assessment and multisite case-control differences cannot be explained by scanning site. Excluding total brain thickness/area as a cofactor in the model did not influence the association between gF/gC and cortical thickness. Methods Samples The participants were recruited by the ABCD Study Release 2.0.1 after quality control for neuroimaging data and behavioral tests with an age span between 108-131 months (around 9-10 years old). The ABCD Consortium used NIH Toolbox Cognition battery (NIHTB-CB) composite scores(Luciana, Bjork et al. 2018), which include a Total Score Composite, a Crystalized Intelligence Composite (The Toolbox Picture Vocabulary Task and The Toolbox Oral Reading Recognition Task) and a Fluid Intelligence Composite (The Toolbox Pattern Comparison Processing Speed Test, The Toolbox List Sorting Working Memory Test, The Toolbox Picture Sequence Memory Test, The Toolbox Flanker Task, The Toolbox Dimensional Change Card Sort Task)(Akshoomoff, Beaumont et al. 2013). These composite scores also show good test -- retest reliabilities in both children and adults as well as validity in children(Akshoomoff, Beaumont et al. 2013, Heaton, Akshoomoff et al. 2014) and highly related (r=0.89) with scores measured with WAIS-IV(Heaton, Akshoomoff et al. 2014). Cortical thickness are measured using FreeSurfer 5.3.0 under Destrieux atlas, which include 148 regions. Multiple linear regression models were employed to model the relationship between brain cortical thickness and three cognition scores, separately. Although the age span is narrow, intelligence is significantly correlated with age. Therefore, age-corrected standard scores were used and meanwhile, age, gender and site were considered as nuisance variables in the models. Statistical Analysis Statistical analyses of demographic data and test scores were conducted using R software. The mean and variability parameters of cortical thickness and their association with intelligence were estimated using double generalized linear models(DGLM) (Efron 1986, Smyth 1989). Before this, age, sex and site were regressed as nuisance variables using generalized additive model(GAM) (Diederich 2007). Then, DGLM iteratively fit a generalized linear model of the mean parameter and a second generalized linear model of the variability parameter on the deviance of the first model and. Cortical thickness statistic map(t statistics) are submitted to correct for multiple comparisons using false discovery rate (FDR) correction(Benjamini and Hochberg 1995 ) and the brain regions with corrected p value less than 0.05 would survive. Finally, elastic net regression was employed to cortical thickness for predicting three kinds of intelligence. Generalized additive Model (GAM): In order to correct the data for site, age, sex effects, we ran generalized additive models on each ROI analyses using the following model: 𝑌 ~𝑠(𝐴𝑔𝑒) + 𝑆𝑒𝑥 + 𝑆𝑐𝑎𝑛𝑛𝑒𝑟. Where Y represents cortical thickness in each brain regions, s(.) is a smooth function, estimated from the data. Double Generalized Linear Model (DGLM): DGLM fitted using the following model for both the mean and dispersion part. Modeling the dispersion is important for obtaining correct mean parameter estimates if dispersion varies as a function of the predictor, and allows for systematic investigation into factors associated dispersion in observations. Mean model: 𝑚𝑖 = 𝜇 + 𝐴𝑔𝑒𝛽𝑎𝑔𝑒 + 𝑆𝑒𝑥𝛽𝑠𝑒𝑥 + 𝐼𝑄𝛽𝑐𝑠 𝑖 = 1,2, … , 𝑁 Dispersion model: 𝑣𝑖 = 𝑣 + 𝐴𝑔𝑒𝛾𝑎𝑔𝑒 + 𝑆𝑒𝑥𝛾𝑠𝑒𝑥 + 𝐼𝑄𝛾𝑐𝑠 𝑖 = 1,2, … , 𝑁 Here, we assume Y is cortical thickness regressed nuisance variables. It follows a normal distribution with expectation 𝑚𝑖 and variance 𝜎𝑖 2, and 𝜎𝑖 is also a function rather than a constant like 𝑚𝑖 . All 𝛽 are the parameters to be estimated. N is the number of brain regions. For a more intuitive explanation of the model, Figure S2 shows a general view of relationship between different kinds of data distribution and DGLM. Cross-Validated Elastic Net Regression(Zou and Hastie 2005): We used elastic net to test whether cortical thickness can predict different kind of intelligence across subjects. Elastic net enables data-driven regression analysis by enforcing sparsity of regression output values (i.e., reducing the number of final β regression values). In other words (Casey, Cannonier et al. 2018), it provides automatic variable selection by removing all independent variables not predicted dependent variable. We normalized all input data: 𝑋 = 𝑋 − 𝑚𝑒𝑎𝑛(𝑋) max(𝑋) − min (𝑋) This resulted in variables, x, with values between 0 and 1. The elastic net equation is then written as 𝑛 𝑝 2 0, 𝛽 = arg min 𝛽 𝛽0,𝛽 ∑ (𝑦𝑖 − 𝛽0 − ∑ 𝛽𝑗𝑋𝑖𝑗 ) 𝑖=1 𝑗=1 𝑝 + 𝜆 ∑ 𝑗=1 1 2 (1 − 𝛼)𝛽𝑗 2 + 𝛼𝛽𝑗 This is a doubly penalized regression model using both LASSO and Ridge regression. 𝛼 sets the degree of mixing between ridge regression and lasso. Meanwhile, 𝛽 is the shrinkage parameter. When 𝛽 = 0, no shrinkage is performed. Conclusion Ongoing efforts are attempting to account for brain cognitive function and brain morphology. Herein we report that intelligence appears to be associated with widespread increased mean differences and decreased heterogeneity in cortical thickness. The results seem to support the notion that cognitive function has high heterogeneity. Subjects with high IQ have lower heterogeneity in cortical thickness in widespread brain areas. Together these findings warrant future longitudinal studies that cortical thickness contributing to neurobiological heterogeneity. Table 1. Demographic and background characteristics of ABCD samples among 22 sites Site Count Sex(female/all) Age [SD] Mean Thickness [SD] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 345 529 602 643 360 559 325 265 392 621 437 564 574 526 366 990 530 306 486 662 551 33 0.48 0.46 0.47 0.49 0.50 0.50 0.46 0.46 0.49 0.48 0.49 0.49 0.50 0.46 0.47 0.45 0.50 0.46 0.52 0.50 0.44 0.58 118.58[7.62] 2.79[0.10] 120.91[7.51] 2.81[0.08] 118.34[7.42] 2.8[0.08] 117.66[7.82] 2.69[0.10] 118.72[7.41] 2.8[0.09] 119.23[7.16] 2.82[0.09] 118.38[7.52] 2.79[0.09] 119.72[7.42] 2.69[0.09] 119.38[7.36] 2.78[0.09] 118.19[7.57] 2.69[0.09] 117.68[7.63] 2.79[0.09] 118.28[7.4] 2.78[0.09] 117.45[7.29] 2.69[0.10] 122.01[6.79] 2.82[0.09] 118.55[7.34] 2.76[0.10] 118.54[7.88] 2.83[0.08] 117.64[7.57] 2.81[0.10] 119.44[7.57] 2.68[0.10] 120.66[6.62] 2.78[0.11] 120.69[5.86] 2.81[0.09] 118.78[7.53] 2.78[0.09] 122.55[6.49] 2.67[0.10] Fig 1. The mean model results of regions of interest, which the cortical thickness significantly correlation with g, gC and gF. Cortical regions of interest which p value through FDR (0.05) correction were shown based on Destrieux atlas. Fig 2. The general correlation between the three intelligence indicators and region thickness of cortical regions based on Destrieux atlas. The atlas are further broken down into limbic lobe and sulcus (LL), frontal lobe and sulcus(FL), temporal lobe and sulcus(TL), parietal lobe and sulcus (PL), occipital lobe and sulcus(OL), insular cortex(Ins) and sulci /spaces major divisions(SSmd). Fig 3. The dispersion model results of regions of interest, which the cortical thickness significantly correlation with total, crystallized and fluid intelligence. Cortical regions of interest which p value through FDR(0.05) correction were shown and p values through Bonferroni(0.05) correction were tagged based on Destrieux atlas. Fig 4. The variability correlation between the three intelligence scores and cortical thickness of ROI based on Destrieux atlas. Fig 5. Scatterplots between elastic net predicted and normalized intelligence scores (y axes) and original normalized values (x axes). Each dot is a single sample, and dashed lines denote the best linear fit between predicted and normalized intelligence scores. References Akshoomoff, N., J. L. Beaumont, P. J. Bauer, S. S. Dikmen, R. C. Gershon, D. Mungas, J. Slotkin, D. Tulsky, S. Weintraub and P. D. Zelazo (2013). "VIII. NIH Toolbox Cognition Battery (CB): composite scores of crystallized, fluid, and overall cognition." Monographs of the Society for Research in Child Development 78(4): 119-132. Alnaes, D., T. Kaufmann, D. van der Meer, A. Cordova-Palomera, J. Rokicki, T. Moberget, F. Bettella, I. Agartz, D. M. Barch, A. Bertolino, C. L. Brandt, S. Cervenka, S. Djurovic, N. T. Doan, S. Eisenacher, H. Fatouros-Bergman, L. Flyckt, A. Di Giorgio, B. Haatveit, E. G. Jonsson, P. Kirsch, M. J. Lund, A. Meyer- Lindenberg, G. Pergola, E. Schwarz, O. B. Smeland, T. Quarto, M. Zink, O. A. Andreassen, L. T. Westlye and C. Karolinska Schizophrenia Project (2019). "Brain Heterogeneity in Schizophrenia and Its Association With Polygenic Risk." JAMA Psychiatry. Brans, R. G., R. S. Kahn, H. G. Schnack, G. C. van Baal, D. Posthuma, N. E. van Haren, C. Lepage, J. P. Lerch, D. L. Collins, A. C. Evans, D. I. Boomsma and H. E. Hulshoff Pol (2010). "Brain plasticity and intellectual ability are influenced by shared genes." J Neurosci 30(16): 5519-5524. Buckner, R. L. and L. M. DiNicola (2019). "The brain's default network: updated anatomy, physiology and evolving insights." Nat Rev Neurosci 20(10): 593-608. Casey, B. J., T. Cannonier, M. I. Conley, A. O. Cohen, D. M. Barch, M. M. Heitzeg, M. E. Soules, T. Teslovich, D. V. Dellarco, H. Garavan, C. A. Orr, T. D. Wager, M. T. Banich, N. K. Speer, M. T. Sutherland, M. C. Riedel, A. S. Dick, J. M. Bjork, K. M. Thomas, B. Chaarani, M. H. Mejia, D. J. Hagler, M. D. Cornejo, C. S. Sicat, M. P. Harms, N. U. F. Dosenbach, M. Rosenberg, E. Earl, H. Bartsch, R. Watts, J. R. Polimeni, J. M. Kuperman, D. A. Fair, A. M. Dale and A. I. A. Workgrp (2018). "The Adolescent Brain Cognitive Development (ABCD) study: Imaging acquisition across 21 sites." Developmental Cognitive Neuroscience 32: 43-54. Chaddock-Heyman, L., K. I. Erickson, C. Kienzler, M. King, M. B. Pontifex, L. B. Raine, C. H. Hillman and A. F. Kramer (2015). "The role of aerobic fitness in cortical thickness and mathematics achievement in preadolescent children." PloS one 10(8): e0134115. Chenn, A. and C. A. Walsh (2003). "Increased neuronal production, enlarged forebrains and cytoarchitectural distortions in beta-catenin overexpressing transgenic mice." Cereb Cortex 13(6): 599- 606. Choi, Y. Y., N. A. Shamosh, S. H. Cho, C. G. DeYoung, M. J. Lee, J. M. Lee, S. I. Kim, Z. H. Cho, K. Kim, J. R. Gray and K. H. Lee (2008). "Multiple Bases of Human Intelligence Revealed by Cortical Thickness and Neural Activation." Journal of Neuroscience 28(41): 10323-10329. Deary, I. J. and P. G. Caryl (1997). "Neuroscience and human intelligence differences." Trends in Neurosciences 20(8): 365-371. Diederich, A. (2007). "Generalized additive models. An introduction with R." Journal of Mathematical Psychology 51(5): 339-339. Efron, B. (1986). "Double exponential families and their use in generalized linear regression." Journal of the American Statistical Association 81(395): 709-721. Heaton, R. K., N. Akshoomoff, D. Tulsky, D. Mungas, S. Weintraub, S. Dikmen, J. Beaumont, K. B. Casaletto, K. Conway, J. Slotkin and R. Gershon (2014). "Reliability and Validity of Composite Scores from the NIH Toolbox Cognition Battery in Adults." Journal of the International Neuropsychological Society 20(6): 588-598. Luciana, M., J. M. Bjork, B. J. Nagel, D. M. Barch, R. Gonzalez, S. J. Nixon and M. T. Banich (2018). "Adolescent neurocognitive development and impacts of substance use: Overview of the adolescent brain cognitive development (ABCD) baseline neurocognition battery." Developmental Cognitive Neuroscience 32: 67-79. Panizzon, M. S., C. Fennema-Notestine, L. T. Eyler, T. L. Jernigan, E. Prom-Wormley, M. Neale, K. Jacobson, M. J. Lyons, M. D. Grant, C. E. Franz, H. Xian, M. Tsuang, B. Fischl, L. Seidman, A. Dale and W. S. Kremen (2009). "Distinct genetic influences on cortical surface area and cortical thickness." Cereb Cortex 19(11): 2728-2735. Rakic, P. (2009). "Evolution of the neocortex: a perspective from developmental biology." Nat Rev Neurosci 10(10): 724-735. Rakic, P., A. E. Ayoub, J. J. Breunig and M. H. Dominguez (2009). "Decision by division: making cortical maps." Trends Neurosci 32(5): 291-301. Shaw, P., D. Greenstein, J. Lerch, L. Clasen, R. Lenroot, N. Gogtay, A. Evans, J. Rapoport and J. Giedd (2006). "Intellectual ability and cortical development in children and adolescents." Nature 440(7084): 676-679. Smyth, G. K. (1989). "Generalized linear models with varying dispersion." Journal of the Royal Statistical Society: Series B (Methodological) 51(1): 47-60. Smyth, G. K. and A. P. Verbyla (1999). "Adjusted likelihood methods for modelling dispersion in generalized linear models." Environmetrics 10(6): 695-709. Tadayon, E., A. Pascual-Leone and E. Santarnecchi (2019). "Differential Contribution of Cortical Thickness, Surface Area, and Gyrification to Fluid and Crystallized Intelligence." Cerebral Cortex. Zou, H. and T. Hastie (2005). "Regularization and variable selection via the elastic net." Journal of the Royal Statistical Society Series B-Statistical Methodology 67: 301-320. Supplementary ABCD Youth NIHTB-CB Summary Scores include two types of scores: age corrected scores and fully corrected T-Score. Age-corrected scores compare the score of the test-taker to others of the same age. For children, normative scores are provided separately for each year of age to consider expected developmental changes. These are presented as Standard Scores (mean=100, SD=15). Fully Corrected T-Scores (mean = 50, SD = 10) compare the score of the test-taker to those in the NIH Toolbox nationally representative normative sample, while adjusting for key demographic variables. These variables include age, gender, race/ethnicity and educational attainment (for ages 3-17, parent's education is used). All seven of the NIHTB-CB tests were included in this study. This resulted in two measures of crystallized abilities (the NIHTB Picture Vocabulary Test and Oral Reading Test), as well as five measures of fluid abilities: the NIHTB Dimensional Change Card Sort (DCCS) Test of Executive Function-Cognitive Flexibility, NIHTB Flanker Test of Executive Function- Inhibitory Control and Attention, NIHTB Picture Sequence Memory Test of Episodic Memory, NIHTB List Sorting Working Memory Test, and NIHTB Pattern Comparison Processing Speed Test. Figure S1. The correlation between different TB Summary Scores FigureS2. A general view of relationship between different kinds of data distribution and DGLM. The thick blue line represents the data mean, and the thin blue line represents the data variance. Figure S3. The standard deviation of four significant regions in variability model of two groups under different intelligence cut value. Fig S4. The mean model results of regions of interest, which the cortical surface area significantly correlation with total, crystallized and fluid intelligence. Fig S5. Scatterplots between elastic net predicted and normalized intelligence scores (y axes) and original normalized values (x axes). A) Surface area were used to predict the g, gC and gF. B) Cortical thickness and surface area were used to predict the g, gC and gF. Fig S6. The mean model results of regions of interest, which the cortical thickness significantly correlation with gC and its subitems. Fig S7. The mean model results of regions of interest, which the cortical thickness significantly correlation with gF and its subitems.
1712.02449
1
1712
2017-12-06T23:55:56
Quantifying how much sensory information in a neural code is relevant for behavior
[ "q-bio.NC", "cs.IT", "cs.IT", "physics.data-an" ]
Determining how much of the sensory information carried by a neural code contributes to behavioral performance is key to understand sensory function and neural information flow. However, there are as yet no analytical tools to compute this information that lies at the intersection between sensory coding and behavioral readout. Here we develop a novel measure, termed the information-theoretic intersection information $I_{II}(S;R;C)$, that quantifies how much of the sensory information carried by a neural response R is used for behavior during perceptual discrimination tasks. Building on the Partial Information Decomposition framework, we define $I_{II}(S;R;C)$ as the part of the mutual information between the stimulus S and the response R that also informs the consequent behavioral choice C. We compute $I_{II}(S;R;C)$ in the analysis of two experimental cortical datasets, to show how this measure can be used to compare quantitatively the contributions of spike timing and spike rates to task performance, and to identify brain areas or neural populations that specifically transform sensory information into choice.
q-bio.NC
q-bio
Quantifying how much sensory information in a neural code is relevant for behavior Giuseppe Pica1,2 [email protected] Eugenio Piasini1 [email protected] Houman Safaai1,3 [email protected] Caroline A. Runyan3,4 [email protected] Mathew E. Diamond5 [email protected] Tommaso Fellin2,6 [email protected] Christoph Kayser7,8 [email protected] Christopher D. Harvey3 [email protected] Stefano Panzeri1,2 [email protected] 1 Neural Computation Laboratory, Center for Neuroscience and Cognitive Systems@UniTn, Istituto Italiano di Tecnologia, Rovereto (TN) 38068, Italy 2 Neural Coding Laboratory, Center for Neuroscience and Cognitive Systems@UniTn, Istituto Italiano di Tecnologia, Rovereto (TN) 38068, Italy 3 Department of Neurobiology, Harvard Medical School, Boston, MA 02115, USA 4 Department of Neuroscience, University of Pittsburgh, Center for the Neural Basis of Cognition, Pittsburgh, USA 5 Tactile Perception and Learning Laboratory, International School for Advanced Studies (SISSA), Trieste, Italy 6 Optical Approaches to Brain Function Laboratory, Istituto Italiano di Tecnologia, Genova 16163, Italy 7 Institute of Neuroscience and Psychology, University of Glasgow, Glasgow, UK 8 Department of Cognitive Neuroscience, Faculty of Biology, Bielefeld University, Universitätsstr. 25, 33615 Bielefeld, Germany Abstract Determining how much of the sensory information carried by a neural code con- tributes to behavioral performance is key to understand sensory function and neural information flow. However, there are as yet no analytical tools to compute this infor- mation that lies at the intersection between sensory coding and behavioral readout. Here we develop a novel measure, termed the information-theoretic intersection information III(S; R; C), that quantifies how much of the sensory information carried by a neural response R is used for behavior during perceptual discrimi- nation tasks. Building on the Partial Information Decomposition framework, we define III(S; R; C) as the part of the mutual information between the stimulus S and the response R that also informs the consequent behavioral choice C. We compute III(S; R; C) in the analysis of two experimental cortical datasets, to show how this measure can be used to compare quantitatively the contributions of spike timing and spike rates to task performance, and to identify brain areas or neural populations that specifically transform sensory information into choice. 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. 1 Introduction Perceptual discrimination is a brain computation that is key to survival, and that requires both encoding accurately sensory stimuli and generating appropriate behavioral choices (Fig.1). Previous studies have mostly focused separately either on the former stage, called sensory coding, by analyzing how neural activity encodes information about the external stimuli [1–10], or on the latter stage, called behavioral readout, by analyzing the relationships between neural activity and choices in the absence of sensory signal or at fixed sensory stimulus (to eliminate spurious choice variations of neural response due to stimulus-related selectivity) [11–13]. The separation between studies of sensory coding and readout has led to a lack of consensus on what is the neural code, which here we take as the key set of neural activity features for perceptual discrimination. Most studies have in fact defined the neural code as the set of features carrying the most sensory information [1, 2, 8], but this focus has left unclear whether the brain uses the information in such features to perform perception [14–16]. Recently, Ref. [17] proposed to determine if neural sensory representations are behaviorally relevant by evaluating the association, in single trials, between the information about the sensory stimuli S carried by the neural activity R and the behavioral choices C performed by the animal, or, in other words, to evaluate the intersection between sensory coding and behavioral readout. More precisely, Ref. [17] suggested that the hallmark of a neural feature R being relevant for perceptual discrimination is that the subject will perform correctly more often when the neural feature R provides accurate sensory information. Ref.[17] proposed to quantify this intuition by first decoding sensory stimuli from single-trial neural responses and then computing the increase in behavioral performance when such decoding is correct. This intersection framework provides several advantages with respect to earlier approaches based on computing the correlations between trial-averaged psychometric performance and trial-averaged neurometric performance [13, 14, 18], because it quantifies associations between sensory information coding and choices within the same trial, instead of considering the similarity of trial-averaged neural stimulus coding and trial-averaged behavioral performance. However, the intersection information measure proposed in Ref.[17] relies strongly on the specific choice of a stimulus decoding algorithm, that might not match the unknown decoding algorithms of the brain. Further, decoding only the most likely stimulus from neural responses throws away part of the full structure in the measured statistical relationships between S, R and C [3]. To overcome these limitations, here we convert the conceptual notions described in [17] into a novel and rigorous definition of information-theoretic intersection information between sensory coding and behavioral readout III(S; R; C). We construct the information-theoretic intersection III(S; R; C) by building on recent extensions of classical information theory, called Partial Information Decomposi- tions (PID), that are suited to the analysis of trivariate systems [19–21]. We show that III(S; R; C) is endowed with a set of formal properties that a measure of intersection information should satisfy. Finally, we use III(S; R; C) to analyze both simulated and real cortical activity. These applications show how III(S; R; C) can be used to quantitatively redefine the neural code as the set of neural features that carry sensory information which is also used for task performance, and to identify brain areas where sensory information is read out for behavior. 2 An information-theoretic definition of intersection information Throughout this paper, we assume that we are analyzing neural activity recorded during a perceptual discrimination task (Fig.1). Over the course of an experimental trial, a stimulus s ∈ {s1, ..., sNs} is presented to the animal while simultaneously some neural features r (we assume that r either takes discrete values or is discretized into a certain number of bins) and the behavioral choice c ∈ {c1, ..., cNc} are recorded. We assume that the joint probability distribution p(s, r, c) has been empirically estimated by sampling these variables simultaneously over repeated trials. After the animal learns to perform the task, there will be a statistical association between the presented stimulus S and the behavioral choice C, and the Shannon information I(S : C) between stimulus and choice will therefore be positive. How do we quantify the intersection information between the sensory coding s → r and the consequent behavioral readout r → c that involves the recorded neural activity features r in the same trial? Clearly, the concept of intersection information must require the analysis of the full trivariate probability distribution p(s, r, c) during perceptual discriminations. The well-established, 2 Figure 1: Schematics of the information flow in a perceptual discrimination task: sensory infor- mation I(S : R) (light blue block) is encoded in the neural activity R. This activity informs the behavioral choice C and so carries information about it (I(R : C), green block). III(S; R; C) is both a part of I(S : R) and of I(R : C), and corresponds to the sensory information used for behavior. classical tools of information theory [22] provide a framework for assessing statistical associations between two variables only. Indeed, Shannon's mutual information allows us to quantify the sensory information I(S : R) that the recorded neural features carry about the presented stimuli [3] and, separately, the choice information I(R : C) that the recorded neural features carry about the behavior. To assess intersection information in single trials, we need to extend the classic information-theoretic tools to the trivariate analysis of S, R, C. More specifically, we argue that an information-theoretic measure of intersection information should quantify the part of the sensory information which also informs the choice. To quantify this concept, we start from the tools of the Partial Information Decomposition (PID) framework. This framework decomposes the mutual information that two stochastic variables (the sources) carry about a third variable (the target) into four nonnegative information components. These components characterize distinct information sharing modes among the sources and the target on a finer scale than Shannon information quantities [19, 20, 23, 24]. In our analysis of the statistical dependencies of S, R, C, we start from the mutual information I(C : (S, R)) that S and R carry about C. Direct application of the PID framework then leads to the following nonnegative decomposition: I(C : (S, R)) = SI(C : {S; R}) + CI(C : {S; R}) + U I(C : {S \ R}) + U I(C : {R \ S}), (1) where SI, CI and U I are respectively shared (or redundant), complementary (or synergistic) and unique information quantities as defined in [20]. More in detail, • SI(C : {S; R}) is the information about the choice that we can extract from any of S and R, i.e. the redundant information about C shared between S and R. • U I(C : {S \ R}) is the information about the choice that we can only extract from the stimulus but not from the recorded neural response. It thus includes stimulus information relevant to the behavioral choice that is not represented in R. • U I(C : {R \ S}) is the information about the choice that we can only extract from the neural response but not from the stimulus. It thus includes choice information in R that arises from stimulus-independent variables, such as level of attention or behavioral bias. • CI(C : {S; R}) is the information about choice that can be only gathered if both S and R are simultaneously observed with C, but that is not available when only one between S and R is simultaneously observed with C. More precisely, it is that part of I(C : (S, R)) which does not overlap with I(S : C) nor with I(R : C) [19]. Several mathematical definitions for the PID terms described above have been proposed in the literature [19, 20, 23, 24]. In this paper, we employ that of Bertschinger et al. [20], which is widely used for tripartite systems [25, 26]. Accordingly, we consider the space ∆p of all probability distributions q(s, r, c) with the same pairwise marginal distributions q(s, c) = p(s, c) and q(r, c) = 3 p(r, c) as the original distribution p(s, r, c). The redundant information SI(C : {S; R}) is then defined as the solution of the following convex optimization problem on the space ∆p [20]: SI(C : {S; R}) ≡ max q∈∆p CoIq(S; R; C), (2) where CoIq(S; R; C) ≡ Iq(S : R) − Iq(S : RC) is the co-information corresponding to the probability distribution q(s, r, c). All other PID terms are then directly determined by the value of SI(C : {S; R})[19]. However, none of the existing PID information components described above fits yet the notion of intersection information, as none of them quantifies the part of sensory information I(S : R) carried by neural activity R that also informs the choice C. The PID quantity that seems to be closest to this notion is the redundant information that S and R share about C, SI(C : {S; R}). However, previous works pointed out the subtle possibility that even two statistically independent variables (here, S and R) can share information about a third variable (here, C) [23, 27]. This possibility rules out using SI(C : {S; R}) as a measure of intersection information, since we expect that a neural response R which does not encode stimulus information (i.e., such that S ⊥⊥ R) cannot carry intersection information. We thus reason that the notion of intersection information should be quantified as the part of the redundant information that S and R share about C that is also a part of the sensory information I(S : R). This kind of information is even finer than the existing information components of the PID framework described above, and we recently found that comparing information components of the three different Partial Information Decompositions of the same probability distribution p(s, r, c) leads to the identification of finer information quantities [21]. We take advantage of this insight to quantify the intersection information by introducing the following new definition: III(S; R; C) = min{SI(C : {S; R}), SI(S : {R; C})}. (3) This definition allows us to further decompose the redundancy SI(C : {S; R}) into two nonnegative information components, as SI(C : {S; R}) = III(S; R; C) + X(R), (4) where X(R) ≡ SI(C : {S; R}) − III(S; R; C) ≥ 0. This finer decomposition is useful because, unlike SI(C : {S; R}), III(S; R; C) has the property that S ⊥⊥ R =⇒ III(S; R; C) = 0 (see Supp. Info Sec.1). This is a first basic property that we expect from a meaningful definition of intersection information. Moreover, III(S; R; C) satisfies a number of additional important properties (see proofs in Supp. Info Sec. 1) that a measure of intersection information should satisfy: 1. III(S; R; C) ≤ I(S : R): intersection information should be a part of the sensory informa- tion extractable from the recorded response R – namely, the part which is relevant for the choice; 2. III(S; R; C) ≤ I(R : C): intersection information should be a part of the choice information extractable from the recorded response R – namely, the part which is related to the stimulus; 3. III(S; R; C) ≤ I(S : C): intersection information should be a part of the information 4. III(S;{R1, R2}; C) ≥ III(S; R1; C), III(S; R2; C), as the task-relevant information that can be extracted from any recorded neural features should not be smaller than the task- relevant information that can be extracted from any subset of those features. between stimulus and choice – namely, the part which can be extracted from R; The measure III(S; R; C) thus translates all the conceptual features of intersection information into a well-defined analytical tool: Eq.3 defines how III(S; R; C) can be computed numerically from real data once the distribution p(s, r, c) is estimated empirically. In practice, the estimated p(s, r, c) defines the space ∆p where the problem defined in Eq.2 should be solved. We developed a gradient-descent optimization algorithm to solve these problems numerically with a Matlab package that is freely available for download and reuse through Zenodo and Github https://doi.org/10.5281/zenodo.850362 (see Supp. Info Sec. 2). Computing III(S; R; C) allows the experimenter to estimate that portion of the sensory information in a neural code R that is read out for behaviour during a perceptual discrimination task, and thus to quantitatively evaluate hypotheses about neural coding from empirical data. 4 S S R1 R2 R1 R2 C (a) C (b) S R1 C (c) S R1 R2 C (d) information about the stimulus; blue: Figure 2: Some example cases where III(S; R1; C) = 0 for a neural code R1. Each panel contains a probabilistic graphical model representation of p(s, r, c), augmented by a color code illustrating the nature of the information carried by statistical relationships between variables. Red: information about anything else (internal noise, distractors, and so on). III(Ri) > 0 only if the arrows linking Ri with S and C have the same color. a: I(S : R2) > I(S : R1) = 0. I(C : R2) = I(C : R1). III(R2) > III(S; R1; C) = 0. b: I(S : R2) = I(S : R1). I(C : R2) > I(C : R1) = 0. III(R2) > III(S; R1; C) = 0. c: I(S : R1) > 0, I(C : R1) > 0, I(S : C) = 0. d: I(S : R1) > 0, I(C : R1) > 0, I(S : C) > 0, III(S; R1; C) = 0. 2.1 Ruling out neural codes for task performance A first important use of III(S; R; C) is that it permits to rule out recorded neural features as candidate neural codes. In fact, the neural features R for which III(S; R; C) = 0 cannot contribute to task performance. It is interesting, both conceptually and to interpret empirical results, to characterize some scenarios where III(S; R1; C) = 0 for a recorded neural feature R1. III(S; R1; C) = 0 may correspond, among others, to one of the four scenarios illustrated in Fig.2: • R1 drives behavior but it is not informative about the stimulus, i.e. I(R1 : S) = 0 (Fig.2a); • R1 encodes information about S but it does not influence behavior, i.e. I(R1 : C) = 0 (Fig.2b); • R1 is informative about both S and C but I(S : C) = 0 (Fig.2c, see also Supp. Info Sec.2); • I(S : R1) > 0, I(R1 : C) > 0, I(S : C) > 0, but the sensory information I(S : R1) is not read out to drive the stimulus-relevant behavior and, at the same time, the way R1 affects the behaviour is not related to the stimulus (Fig.2d, see also Supp. Info Sec.2). 3 Testing our measure of intersection information with simulated data To better illustrate the properties of our measure of information-theoretic intersection information III(S; R; C), we simulated a very simple neural scheme that may underlie a perceptual discrimination task. As illustrated in Fig.3a, in every simulated trial we randomly drew a stimulus s ∈ {s1, s2} which was then linearly converted to a continuous variable that represents the neural activity in the simulated sensory cortex. This stimulus-response conversion was affected by an additive Gaussian noise term (which we term "sensory noise") whose amplitude was varied parametrically by changing the value of its standard deviation σS. The simulated sensory-cortex activity was then separately converted, with two distinct linear transformations, to two continuous variables that simulated two higher-level brain regions. These two variables are termed "parietal cortex" (R) and "bypass pathway" (R(cid:48)), respectively. We then combined R and R(cid:48) with parametrically tunable weights (we indicate the ratio between the R-weight and the R(cid:48)-weight with α, see Supp. Info Sec.4) and added Gaussian noise (termed "choice noise"), whose standard deviation σC was varied parametrically, to eventually produce another continuous variable that was fed to a linear discriminant. We took as the simulated behavioral choice the binary output of this final linear discriminant, which in our model was meant to represent the readout mechanism in high-level brain regions that inform the motor output. We ran simulations of this model by varying parametrically the sensory noise σS, the choice noise σC, and the parietal to bypass ratio α, to investigate how III(S; R; C) depended on these parameters. 5 Figure 3: a) Schematics of the simulated model used to test our framework. In each trial, a binary stimulus is linearly converted into a "sensory-cortex activity" after the addition of 'sensory noise'. This signal is then separately converted to two higher-level activities, namely a "parietal-cortex activity" R and a "bypass-pathway activity" R(cid:48). R and R(cid:48) are then combined with parametrically tunable weights and, after the addition of "choice noise", this signal is fed to a linear discriminant. The output of the discriminant, that is the decoded stimulus s, drives the binary choice c. We computed the intersection information of R to extract the part of the stimulus information encoded in the "parietal cortex" that contributes to the final choice. b-d) Intersection Information for the simulations represented in a). Mean ± sem of III(S; R; C) across 100 experimental sessions, each relying on 100 simulated trials, as a function of three independently varied simulation parameters. b) Intersection Information decreases when the stimulus representation in the parietal cortex R is more noisy (higher sensory noise σS ). c) Intersection Information decreases when the beneficial contribution of the stimulus information carried by parietal cortex R to the final choice is reduced by increasing choice noise σC. d) Intersection Information increases when the parietal cortex R contributes more strongly to the final choice by increasing the parietal to bypass ratio α. In each simulated session, we estimated the joint probability psession(s, r, c) of the stimulus S, the response in parietal cortex R, and the choice C, from 100 simulated trials. We computed, sepa- rately for each simulated session, an intersection information III(S; R; C) value from the estimated psession(s, r, c). Here, and in all the analyses presented throughout the paper, we used a quadratic extrapolation procedure to correct for the limited sampling bias of information [28]. In Fig.3b-d we show mean ± s.e.m. of III(S; R; C) values across 100 independent experimental sessions, as a function of each of the three simulation parameters. We found that III(S; R; C) decreases with increasing σS (Fig.3b). This result was explained by the fact that increasing σS reduces the amount of stimulus information that is passed to the simulated parietal activity R, and thus also reduces the portion of such information that can inform choice and can be used to perform the task appropriately. We found that III(S; R; C) decreases with increasing σC (Fig.3c), consistently with the intuition that for higher values of σC the choice depends more weakly on the activity of the simulated parietal activity R, which in turn also reduces how accurately the choice reflects the stimulus in each trial. We also found that III(S; R; C) increases with increasing α (Fig.3d), because when α is larger the portion of stimulus information carried by the simulated parietal activity R that benefits the behavioral performance is larger. 6 4 Using our measure to rank candidate neural codes for task performance: studying the role of spike timing for somatosensory texture discrimination The neural code was traditionally defined in previous studies as the set of features of neural activity that carry all or most sensory information. In this section, we show how III(S; R; C) can be used to quantitatively redefine the neural code as the set of features that contributes the most sensory information for task performance. The experimenter can thus use III(S; R; C) to rank a set of candidate neural features {R1, ..., RN} according to the numerical ordering III(S; Ri1 ; C) ≤ ... ≤ III(S; RiN ; C). An advantage of the information-theoretic nature of III(S; R; C) is that it quantifies intersection information on the meaningful scale of bits, and thus enables a quantitative comparison of different candidate neural codes. If for example III(S; R1; C) = 2III(S; R2; C) we can quantitatively interpret that the code R1 provides twice as much information for task performance as R2. This interpretation is not as meaningful, for example, when comparing different values of fraction-correct measures [17]. To illustrate the power of III(S; R; C) for evaluating and ranking candidate neural codes, we apply it to real data to investigate a fundamental question: is the sensory information encoded in millisecond- scale spike times used by the brain to perform perceptual discrimination? Although many studies have shown that millisecond-scale spike times of cortical neurons encode sensory information not carried by rates, whether or not this information is used has remained controversial [16, 29, 30]. It could be, for example, that spike times cannot be read out because the biophysics of the readout neuronal systems is not sufficiently sensitive to transmit this information, or because the readout neural systems do not have access to a stimulus time reference that could be used to measure these spike times [31]. To investigate this question, we used intersection information to compute whether millisecond- scale spike timing of neurons (n=299 cells) in rat primary (S1) somatosensory cortex provides information that is used for performing a whisker-based texture discrimination task (Figure 4a-b). Full experimental details are reported in [32]. In particular, we compared III(S; timing; C) with the intersection information carried by rate III(S; rate; C), i.e. information carried by spike counts over time scales of tens of milliseconds. We first computed a spike-timing feature by projecting the single-trial spike train onto a zero-mean timing template (constructed by linearly combining the first three spike trains PCs to maximize sensory information, following the procedure of [32]), whose shape indicated the weight assigned to each spike depending on its timing (Figure 4a). Then we computed a spike-rate feature by weighting the spikes with a flat template which assigns the same weight to spikes independently of their time. Note that this definition of timing, and in particular the fact that the timing template was zero mean, ensured that the timing variable did not contain any rate information. We verified that this calculation provided timing and rate features that had negligible (-0.0030 ± 0.0001 across the population) Pearson correlation. The difficulty of the texture discrimination task was set so that the rat learned the task well but still made a number of errors in each session (mean behavioral performance 76.9%, p<0.001 above chance, paired t-test). These error trials were used to decouple in part choice from stimulus coding and to assess the impact of the sensory neural codes on behavior by computing intersection information. We thus computed information across all trials, including both behaviorally correct and incorrect trials. We found that, across all trials and on average over the dataset, timing carried similar texture information to rate (Figure 4b) ((9 ± 2) × 10−3 bit in timing, (8.5 ± 1.1) × 10−3 bit in rate, p=0.78 two-sample t-test), while timing carried more choice information than rate ((16 ± 1) × 10−3 bit in timing, (3.0 ± 0.7) × 10−3 bit in rate, p<10−15 two-sample t-test). If we used only traditional measures of stimulus and choice information, it would be difficult to decide which code is most helpful for task performance. However, when we applied our new information-theoretic framework, we found that the intersection information III (Figure 4b) was higher for timing than for rate ((7 ± 1) × 10−3 bit in timing, (3.0 ± 0.6) × 10−3 bit in rate, p<0.002 two-sample t-test), thus suggesting that spike timing is a more crucial neural code for texture perception than spike rate. Interestingly, intersection information III was approximately 80% of the total sensory information for timing, while it was only 30% of the total sensory information for rate. This suggests that in somatosensory neurons timing information about the texture is read out, and influences choice, more efficiently than rate information, contrarily to what is widely assumed in the literature [34]. These results confirm early results that were obtained with a decoding-based intersection information measure [32]. However, the information theoretic results in Fig.4b have the advantage that they do 7 Figure 4: Intersection Information for two experimental datasets. a: Simplified schematics of the experimental setup in [32]. Rats are trained to distinguish between textures with different degrees of coarseness (left), and neural spiking data from somatosensory cortex (S1) is decomposed in independent rate and timing components (right). b: Stimulus, choice and intersection information for the data in panel a. Spike timing carries as much sensory information (p=0.78, 2-sample t-test), but more choice information (p<10−15), and more III (p<0.002) than firing rate. c: Simplified schematics of the experimental setup in [33]. Mice are trained to distinguish between auditory stimuli located to their left or to their right. Neural activity is recorded in auditory cortex (AC) and posterior parietal cortex (PPC) with 2-photon calcium imaging. d: Stimulus, choice and intersection information for the data in panel c. Stimulus information does not differ significantly between AC and PPC, but PPC has more choice information (p<0.05) and more III than AC (p<10−6, 2-sample t-test). not depend on the use of a specific decoder to calculate intersection information. Importantly, the new information theoretic approach also allowed us to quantify the proportion of sensory information in a neural code that is read out downstream for behavior, and thus to obtain the novel conclusion that only spike timing is read out with high efficiency. 5 Application of intersection information to discover brain areas transforming sensory information into choice Our intersection information measure III(S; R; C) can also be used as a metric to discover and index brain areas that perform the key computations needed for perceptual discrimination, and thus turn sensory information into choice. Suppose for example that we are investigating this issue by recording from populations of neurons in different areas. If we rank the neural activities in the recorded areas according to the sensory information they carry, we will find that primary sensory areas are ranked highly. Instead, if we rank the areas according to the choice information they carry, the areas encoding the motor output will be ranked highly. However, associative areas that transform sensory information into choice will not be found by any of these two traditional sensory-only and choice-only rankings, and there is no currently established metric to quantitatively identify such areas. Here we argue that III(S; R; C) can be used as such metric. To illustrate this possible use of III(S; R; C), we analyzed the activity of populations of single neurons recorded in mice with two-photon calcium imaging either in Auditory Cortex (AC, n=329 neurons) or in Posterior Parietal Cortex (PPC, n=384 neurons) while the mice were performing a sound location discrimination task and had to report the perceived sound location (left vs right) by the direction of their turn in a virtual-reality navigation setup (Fig.4c; full experimental details are available in Ref.[33]). AC is a primary sensory area, whereas PPC is an association area that has been described as a multisensory-motor interface [35–37], was shown to be essential for virtual-navigation tasks [36], and is implicated in the spatial processing of auditory stimuli [38, 39]. When applying our information theoretic formalism to these data, we found that similar stimu- lus (sound location) information was carried by the firing rate of neurons in AC and PPC (AC: (10 ± 3) × 10−3 bit, PPC: (5 ± 1) × 10−3 bit, p=0.17, two-sample t-test). Cells in PPC carried 8 00.0050.010.0150.02RoughSmoothTimeInst. rateInst. rateTimeRate templateTiming templateXLeftRightPPCAC(a)(b)(c)(d)***ACPPCRateTiming*****Information (bits)Information (bits)StimulusChoiceIntersection00.0040.0080.012StimulusChoiceIntersection* more choice information than cells in AC (AC: (2.8 ± 1.4) × 10−3 bit, PPC: (6.4 ± 1.2) × 10−3 bit, p<0.05, two-sample t-test). However, neurons in PPC had values of III ((3.6 ± 0.8) × 10−3 bit) higher (p<10−6, two-sample t-test) than those of AC ((2.3 ± 0.8) × 10−3 bit): this suggests that the sensory information in PPC, though similar to that of AC, is turned into behavior into a much larger proportion (Figure 4d). Indeed, the ratio between III(S; R; C) and sensory information was higher in PPC than in AC (AC: (24 ± 11) %, PPC: (73 ± 24) %, p<0.03, one-tailed z-test). This finding reflects the associative nature of PPC as a sensory-motor interface. This result highlights the potential usefulness of III(S; R; C) as an important metric for the analysis of neuro-imaging experiments and the quantitative individuation of areas transforming sensory information into choice. 6 Discussion Here, we derived a novel information theoretic measure III(S; R; C) of the behavioral impact of the sensory information carried by the neural activity features R during perceptual discrimination tasks. The problem of understanding whether the sensory information in the recorded neural features really contributes to behavior is hotly debated in neuroscience [16, 17, 30]. As a consequence, a lot of efforts are being devoted to formulate advanced analytical tools to investigate this question [17, 40, 41]. A traditional and fruitful approach has been to compute the correlation between trial- averaged behavioral performance and trial-averaged stimulus decoding when presenting stimuli of increasing complexity [13, 14, 18]. However, this measure does not capture the relationship between fluctuations of neural sensory information and behavioral choice in the same experimental trial. To capture this single-trial relationship, Ref.[17] proposed to use a specific stimulus decoding algorithm to classify trials that give accurate sensory information, and then quantify the increase in behavioral performance in the trials where the sensory decoding is correct. However, this approach makes strong assumptions about the decoding mechanism, which may or may not be neurally plausible, and does not make use of the full structure of the trivariate S, R, C dependencies. In this work, we solved all the problems described above by extending the recent Partial Information Decomposition framework [19, 20] for the analysis of trivariate dependencies to identify III(S; R; C) as a part of the redundant information about C shared between S and R that is also a part of the sensory information I(S : R). This quantity satisfies several essential properties of a measure of intersection information between the sensory coding s → r and the consequent behavioral readout r → c, that we derived from the conceptual notions elaborated in Ref.[17]. Our measure III(S; R; C) provides a single-trial quantification of how much sensory information is used for behavior. This quantification refers to the absolute physical scale of bit units, and thus enables a direct comparison of different candidate neural codes for the analyzed task. Furthermore, our measure has the advantages of information-theoretical approaches, that capture all statistical dependencies between the recorded quantities irrespective of their relevance to neural function, as well as of model-based approaches, that link directly empirical data with specific theoretical hypotheses about sensory coding and behavioral readout but depend strongly on their underlying assumptions (see e.g. [12]). An important direction for future expansions of this work will be to combine III(S; R; C) with interventional tools on neural activity, such as optogenetics. Indeed, the novel statistical tools in this work cannot distinguish whether the measured value of intersection information III(S; R; C) derives from the causal involvement of R in transmitting sensory information for behavior, or whether R only correlates with causal information-transmitting areas [17]. More generally, this work can help us mapping information flow and not only information represen- tation. We have shown above how computing III(S; R; C) separates the sensory information that is transmitted downstream to affect the behavioral output from the rest of the sensory information that is not transmitted. Further, another interesting application of III arises if we replace the final choice C with other nodes of the brain networks, and compute with III(S; R1; R2) the part of the sensory information in R1 that is transmitted to R2. Even more generally, besides the analysis of neural information processing, our measure III can be used in the framework of network information theory: suppose that an input X = (X1, X2) (with X1 ⊥⊥ X2) is encoded by 2 different parallel channels R1, R2, which are then decoded to produce collectively an output Y . Suppose further that experimental measurements in single trials can only determine the value of X, Y , and R1, while the values of X1, X2, Y1, Y2, R2 are experimentally unaccessible. As we show in Supp. Fig. 3, III(X; R1; Y ) allows us to quantify the information between X and Y that passes through the channel R1, and thus does not pass through the channel R2. 9 7 Acknowledgements and author contributions GP was supported by a Seal of Excellence Fellowship CONISC. SP was supported by Fondation Bertarelli. CDH was supported by grants from the NIH (MH107620 and NS089521). CDH is a New York Stem Cell Foundation Robertson Neuroscience Investigator. TF was supported by the grants ERC (NEURO-PATTERNS) and NIH (1U01NS090576-01). CK was supported by the European Research Council (ERC-2014-CoG; grant No 646657). Author contributions: SP, GP and EP conceived the project; GP and EP performed the project; CAR, MED and CDH provided experimental data; GP, EP, HS, CK, SP and TF provided materials and analysis methods; GP, EP and SP wrote the paper; all authors commented on the manuscript; SP supervised the project. References [1] W. Bialek, F. Rieke, R.R. de Ruyter van Steveninck, and D. Warland. Reading a neural code. Science, 252(5014):1854–1857, 1991. [2] A. Borst and F.E. Theunissen. Information theory and neural coding. Nat. Neurosci., 2(11):947–957, 1999. [3] R. Quian Quiroga and S. Panzeri. Extracting information from neuronal populations: information theory and decoding approaches. Nat. Rev. Neurosci., 10(3):173–185, 2009. [4] D. V. Buonomano and W. Maass. State-dependent computations: spatiotemporal processing in cortical networks. Nat. Rev. Neurosci., 10:113–125, 2009. [5] M. A. Harvey, H. P. Saal, J. F. III Dammann, and S. J. Bensmaia. Multiplexing stimulus information through rate and temporal codes in primate somatosensory cortex. PLOS Biology, 11(5):e1001558, 2013. [6] C. Kayser, M. A. Montemurro, N. K. Logothetis, and S. Panzeri. Spike-phase coding boosts and stabilizes information carried by spatial and temporal spike patterns. Neuron, 61(4):597–608, 2009. [7] A. Luczak, B. L. McNaughton, and K. D. Harris. Packet-based communication in the cortex. Nat. Rev. Neurosci., 16(12):745–755, 2015. [8] S. Panzeri, N. Brunel, N. K. Logothetis, and C. Kayser. Sensory neural codes using multiplexed temporal scales. Trends Neurosci., 33(3):111–120, 2010. [9] M. Shamir. Emerging principles of population coding: in search for the neural code. Curr. Opin. Neurobiol., 25:140–148, 2014. [10] S. Panzeri, J.H. Macke, J. Gross, and C. Kayser. Neural population coding: combining insights from microscopic and mass signals. Trends Cogn. Sci., 19(3):162–172, 2015. [11] K. H. Britten, W. T. Newsome, M. N. Shadlen, S. Celebrini, and J. A. Movshon. A relationship between behavioral choice and the visual responses of neurons in macaque MT. Vis. Neurosci., 13:87–100, 1996. [12] R. M. Haefner, S. Gerwinn, J. H. Macke, and M. Bethge. Inferring decoding strategies from choice probabilities in the presence of correlated variability. Nat. Neurosci., 16:235–242, 2013. [13] W. T. Newsome, K. H. Britten, and J. A. Movshon. Neuronal correlates of a perceptual decision. Nature, 341(6237):52–54, 1989. [14] C. T. Engineer, C. A. Perez, Y. H. Chen, R. S. Carraway, A. C. Reed, J. A. Shetake, V. Jakkamsetti, K. Q. Chang, and M. P. Kilgard. Cortical activity patterns predict speech discrimination ability. Nat. Neurosci., 11:603–608, 2008. [15] A. L. Jacobs, G. Fridman, R. M. Douglas, N. M. Alam, P. E. Latham, G. T. Prusky, and S. Nirenberg. Ruling out and ruling in neural codes. Proc. Natl. Acad. Sci. U.S.A., 106(14):5936–5941, 2009. [16] R. Luna, A. Hernandez, C. D. Brody, and R. Romo. Neural codes for perceptual discrimination in primary somatosensory cortex. Nat. Neurosci., 8(9):1210–1219, 2005. [17] S. Panzeri, C. D. Harvey, E. Piasini, P. E. Latham, and T. Fellin. Cracking the Neural Code for Sensory Perception by Combining Statistics, Intervention, and Behavior. Neuron, 93(3):491–507, 2017. [18] R. Romo and E. Salinas. Flutter discrimination: neural codes, perception, memory and decision making. Nat. Rev. Neurosci., 4(3):203–218, 2003. 10 [19] P. Williams and R. Beer. Nonnegative decomposition of multivariate information. arXiv:1004.2515, 2010. [20] N. Bertschinger, J. Rauh, E. Olbrich, J. Jost, and N. Ay. Quantifying unique information. Entropy, 16(4):2161–2183, 2014. [21] G. Pica, E. Piasini, D. Chicharro, and S. Panzeri. Invariant components of synergy, redundancy, and unique information among three variables. Entropy, 19(9):451, 2017. [22] C. E. Shannon. A mathematical theory of communication. Bell System Technical Journal, 27(3):379–423, 1948. [23] M. Harder, C. Salge, and D. Polani. Bivariate measure of redundant information. Phys. Rev. E, 87(1):012130, 2013. [24] V. Griffith and C. Koch. Quantifying synergistic mutual information. Inception, pages 159–190. Springer Berlin Heidelberg, 2014. In Guided Self-Organization: [25] A. Barrett. Exploration of synergistic and redundant information sharing in static and dynamical gaussian systems. Phys. Rev. E, 91(5):052802, 2015. [26] D. Chicharro. Quantifying multivariate redundancy with maximum entropy decompositions of mutual information. arXiv:1708.03845, 2017. [27] N. Bertschinger, J. Rauh, E. Olbrich, and J. Jost. Shared information – new insights and problems in decomposing information in complex systems. In Proceedings of the ECCS 2012, Brussels, Belgium, 2012. [28] S. P. Strong, R. Koberle, R. R. de Ruyter van Steveninck, and W. Bialek. Entropy and information in neural spike trains. Phys. Rev. Lett., 80:197–200, 1998. [29] J. D. Victor and S. Nirenberg. Indices for testing neural codes. Neural Comput., 20(12):2895–2936, 2008. [30] D. H. O' Connor, S. A. Hires, Z. V. Guo, N. Li, J. Yu, Q.-Q. Sun, D. Huber, and K. Svoboda. Neural coding during active somatosensation revealed using illusory touch. Nat. Neurosci., 16(7):958–965, 2013. [31] S. Panzeri, R. A. A. Ince, M. E. Diamond, and C. Kayser. Reading spike timing without a clock: intrinsic decoding of spike trains. Phil. Trans. R. Soc. Lond., B, Biol. Sci., 369(1637):20120467, 2014. [32] Y. Zuo, H. Safaai, G. Notaro, A. Mazzoni, S. Panzeri, and M. E. Diamond. Complementary contributions of spike timing and spike rate to perceptual decisions in rat S1 and S2 cortex. Curr. Biol., 25(3):357–363, 2015. [33] C. A. Runyan, E. Piasini, S. Panzeri, and C. D. Harvey. Distinct timescales of population coding across cortex. Nature, 548:92–96, 2017. [34] M. N. Shadlen and W. T. Newsome. The variable discharge of cortical neurons: Implications for connectiv- ity, computation, and information coding. J. Neurosci., 18(10):3870–3896, 1998. [35] J. I. Gold and M. N. Shadlen. The neural basis of decision making. Annu. Rev. Neurosci., 30(1):535–574, 2007. [36] C. D. Harvey, P. Coen, and D. W. Tank. Choice-specific sequences in parietal cortex during a virtual- navigation decision task. Nature, 484(7392):62–68, 2012. [37] D. Raposo, M. T. Kaufman, and A. K. Churchland. A category-free neural population supports evolving demands during decision-making. Nat. Neurosci., 17(12):1784–1792, 2014. [38] K. Nakamura. Auditory spatial discriminatory and mnemonic neurons in rat posterior parietal cortex. J. Neurophysiol., 82(5):2503, 1999. [39] J. P. Rauschecker and B. Tian. Mechanisms and streams for processing of "what" and "where" in auditory cortex. Proc. Natl. Acad. Sci. U.S.A., 97(22):11800–11806, 2000. [40] R. Rossi-Pool, E. Salinas, A. Zainos, M. Alvarez, J. Vergara, N. Parga, and R. Romo. Emergence of an abstract categorical code enabling the discrimination of temporally structured tactile stimuli. Proc. Natl. Acad. Sci. U.S.A., 113(49):E7966–E7975, 2016. [41] X. Pitkow, S. Liu, D. E. Angelaki, G. C. DeAngelis, and A. Pouget. How can single sensory neurons predict behavior? Neuron, 87(2):411–423, 2015. 11
1201.2845
2
1201
2012-04-03T10:22:29
Competition through selective inhibitory synchrony
[ "q-bio.NC", "cs.NE" ]
Models of cortical neuronal circuits commonly depend on inhibitory feedback to control gain, provide signal normalization, and to selectively amplify signals using winner-take-all (WTA) dynamics. Such models generally assume that excitatory and inhibitory neurons are able to interact easily, because their axons and dendrites are co-localized in the same small volume. However, quantitative neuroanatomical studies of the dimensions of axonal and dendritic trees of neurons in the neocortex show that this co-localization assumption is not valid. In this paper we describe a simple modification to the WTA circuit design that permits the effects of distributed inhibitory neurons to be coupled through synchronization, and so allows a single WTA to be distributed widely in cortical space, well beyond the arborization of any single inhibitory neuron, and even across different cortical areas. We prove by non-linear contraction analysis, and demonstrate by simulation that distributed WTA sub-systems combined by such inhibitory synchrony are inherently stable. We show analytically that synchronization is substantially faster than winner selection. This circuit mechanism allows networks of independent WTAs to fully or partially compete with each other.
q-bio.NC
q-bio
Competition through selective inhibitory synchrony∗ Department of Neural Systems, Max Planck Institute for Brain Research Ueli Rutishauser [email protected] Jean-Jacques Slotine Nonlinear Systems Laboratory, Massachusetts Institute of Technology [email protected] Rodney J. Douglas Institute of Neuroinformatics, ETH Zurich and University of Zurich [email protected] October 17, 2018 1 Abstract Models of cortical neuronal circuits commonly depend on inhibitory feedback to control gain, pro- vide signal normalization, and to selectively amplify signals using winner-take-all (WTA) dynamics. Such models generally assume that excitatory and inhibitory neurons are able to interact easily, because their axons and dendrites are co-localized in the same small volume. However, quanti- tative neuroanatomical studies of the dimensions of axonal and dendritic trees of neurons in the neocortex show that this co-localization assumption is not valid. In this paper we describe a simple modification to the WTA circuit design that permits the effects of distributed inhibitory neurons to be coupled through synchronization, and so allows a single WTA to be distributed widely in cor- tical space, well beyond the arborization of any single inhibitory neuron, and even across different cortical areas. We prove by non-linear contraction analysis, and demonstrate by simulation that distributed WTA sub-systems combined by such inhibitory synchrony are inherently stable. We show analytically that synchronization is substantially faster than winner selection. This circuit mechanism allows networks of independent WTAs to fully or partially compete with other. 2 Introduction Many models of neuronal computation involve the interaction of a population of excitatory neu- rons whose outputs drive inhibitory neuron(s), which in turn provide global negative feedback to ∗This article has been accepted by MIT Press for publication in a future issue of Neural Computation (2012). This is a pre-print version (as accepted). This version was updated on 04/03/12, correcting minor typos discovered during proof-reading. 1 the excitatory pool (Amari & Arbib, 1977; Douglas, Koch, Mahowald, Martin, & Suarez, 1995; Hahnloser, Sarpeshkar, Mahowald, Douglas, & Seung, 2000; Yuille & Geiger, 2003; Maass, 2000; Hertz, Krogh, & Palmer, 1991; Rabinovich et al., 2000; Rutishauser, Douglas, & Slotine, 2011; Coultrip, Granger, & Lynch, 1992). Practical implementation of such circuits in biological neural circuits depend on co-localization of the excitatory and inhibitory neurons, an assumption which studies of the extents of axonal and dendritic trees of neurons in the neocortex show is not valid (Katzel, Zemelman, Buetfering, Wolfel, & Miesenbock, 2011; Binzegger, Douglas, & Martin, 2004; Shepherd, Stepanyants, Bureau, Chklovskii, & Svoboda, 2005; Douglas & Martin, 2004). Firstly, a substantial fraction of the axonal arborization of a typical excitatory 'spiny' pyramidal neuron extends well beyond the range of the arborization of a typical 'smooth' inhibitory neuron, partic- ularly in the populous superficial layers of the neocortex (Yabuta, 1998; Binzegger et al., 2004). This spatial arrangement means that excitatory effects can propagate well outside the range of the negative feedback provided by a single inhibitory neuron. Secondly, the horizontally disposed 'basket' type of inhibitory neuron, which is a prime candidate for performing normalization, makes multiple synaptic contacts with its excitatory targets, so that even within the range of its axonal arborization, not all the members of an excitatory population can be covered by its effect. This connection pattern means that excitatory neurons within some local population must either be partitioned functionally, or multiple smooth cells must co-operate to cover the entire population of excitatory cells. In previous publications we have shown how winner-take-all (WTA) circuits composed of a small population of excitatory neurons and a single inhibitory neuron can be combined to construct super- circuits that exhibit finite state-machine (FSM) like behavior (Rutishauser & Douglas, 2009; Neftci, Chicca, Indiveri, Cook, & Douglas, 2010). The super-circuits made use of sparse excitatory cross- connections between WTA modules to express the required states of the FSM. These excitatory connections can extend well outside of the range of the local WTA connections, and so are consistent with the observed long-range lateral excitatory connections referred to above. On the other hand, we have not previously confronted the question of whether the WTA is necessarily localized to the extent of the smooth-cell arborization, or whether the WTA can itself be well distributed in space within or between cortical lamina, or even between cortical areas. In this paper we describe a simple modification to the WTA circuit design that couples the effects of distributed inhibitory neurons through synchronization, and so permits a WTA to be widely distributed in cortical space, well beyond the range of the axonal arborization of any single inhibitory neuron, and even across cortical areas. We also demonstrate that such a distributed WTA is inherently stable in its operation. 3 Results We have considered a number of circuits that could be used to distribute spatially the WTA behavior (Fig 1). However, we will describe and analyze only the circuit shown in Fig. 2, which we consider to be the most elegant of the distributive mechanisms (notice the similarity to Fig. 1B). The key insight is the following: Under normal operating conditions, all the participating distributed inhibitory neurons should receive the same summed excitatory input. We achieve this by interposing an excitatory neuron in the negative feedback loop from the excitatory population to its local inhibitory neuron. Instead of the local inhibitory neuron summing over its excitatory population, the interposed neuron performs the summing and passes its result to the inhibitory neuron. This result is also copied to the more distant inhibitory neurons in the spatially distributed 2 Figure 1: Circuits for distributing WTAs. (A) Illustration of the principal idea - mutual excitation of the inhibitory neurons. (B-E) Are are biologically plausible versions. (B) Implementation using intermediate excitatory neurons z1,2. This circuit will be considered in detail in Fig. 2B with more realistic connectivity. (C) Implementation using disinhibition of persistently active units z1,2 as illustrated by the step-functions. (D) Implementation with disinhibition and long-range excitatory units. (E) Implementation using multiplication of inhibitory neurons. Here, x3 = y3 at all times. The maximal excitatory projection length is double that of the inhibitory. 3 A)C)z1z2D)E)InhibitoryExcitatoryz1z2z2B)β2β1x2x1x3α1α1β1β2β1y2y1y3α1α1β1β4β4β2β1x2x1x3α1α1β1β2β1y2y1y3α1α1β1β4β4z1β2β1x2x1x3α1α1β1β2β1y2y1y3α1α1β1β4β4β3β3β2β1x2x1x3α1α1β1β2β1x2x1x3α1α1β1z3z4β4β4β2β1y2y1y3α1α1β1β2β1x2x1x3α1α1β1inhexc Figure 2: Schematic of connectivity. Gray units and dashed lines are inhibitory, white units and straight lines are excitatory. (A) Single WTA with two possible winners x1,2, inhibitory unit x3, and intermediate excitatory unit x4 that carries the average activity of x1,2. (B) Coupling of two WTAs to form a single WTA with 4 possible winners. β4 are excitatory long-range connections that serve to synchronize the two inhibitory units. (C-D) Are reduced versions for theoretical analysis. (C) Merged WTA with one winner each and thus two possible winners x1 and y1 in the merged network. (D) Reduced single WTA. WTA. In this way the inhibitory neuron of each sub-WTA sums over the projections from the interposed excitatory neurons of all other sub-WTAs, including its own one. Thus, each inhibitory neuron is able to provide feedback inhibition to its local sub-WTA that is proportional to the total excitation provided by all excitatory neurons participating in the entire distributed WTA. We will show that functionally this amounts to a form of synchrony between all the inhibitory units. 3.1 Connectivity and dynamics - single WTA All the circuits of Fig 1 can achieve a distributed WTA by merging several independent WTAs, but we consider the circuit shown in Fig 2B to be most feasible, and so our analysis will focus on this one. However, similar reasoning could be applied to all the variants shown. Note that our chosen circuit is similar to that of Fig 1B, but has a more realistic connectivity pattern in that the summed excitatory activity is only projected onto a single unit, which requires less wiring specificity than Fig 1B. The dynamics of a single WTA (Fig 2A) with in total N units, consisting of 1..N − 2 excitatory units, one inhibitory unit xN−1 and one intermediary interconnect excitatory unit xN , are τ xi + Gxi = f (Ii + αxi − β1xN−1 − Ti) τ xN−1 + GxN−1 = f (β3xN − TN−1) xj − TN ) τ xN + GxN = f (β2 N−2(cid:88) (1) Each excitatory unit receives recurrent input from itself (α1) and its neighbors (α2,3,..., see Fig 2A). For simplicity, only self-recurrence is considered here (α = α1 and α2,3,... = 0), but very similar j=1 4 A)B)β4β4C)β2β1x1x2x3y2y3β4β4β3β3β2β1y1x2β2β1x1D)β3β3β2β1β1x2x1x3α1α1x4β3β2β1β1x2x1x3α1α1x4β3β2β1β1y2y1y3α1α1y4α1α1α1α2α2 arguments obtain when recurrence from neighboring units is included. Using the weight matrix W the dynamics of this system is described as with W = τ x + Gx = f (Wx − T + I(t))  α 0 −β1 0 α −β1 0 0 0 0 β2 β2  0 0 β3 0 (2) (3) where the order of units is x1,2,3,4 (i.e. first column and row is x1, last column/row is x4). The firing rate activation function f (x) is a non-saturating rectification non-linearity max(0, x). We assume τ = 1 and G = 1 throughout unless mentioned otherwise. T = [T1, ..., TN−1, TN ] is a vector of the constant activation thresholds Ti ≥ 0. 3.2 Connectivity and dynamics - coupled WTA Two identical single WTAs, each described by weight matrices W1 = W2 = W can be combined into one distributed WTA that acts functionally as a single WTA by adding a recurrent excitatory feedback loop β4 between the two WTAs (Fig 2B). The weight matrix WC of the merged system is with interconnections C1 = WC = C1 W2 (cid:20) W1 C2  0 0 0 0 0 0 0 0 0 0 0 β4 0 0 0 0 (cid:21)  (4) (5) (6) The dynamics of this system are as shown in Eq 2 using W = WC. 3.3 Stability analysis The stability analysis, using non-linear contraction analysis (see Appendix) (Lohmiller & Slotine, 1998, 2000; Slotine, 2003; W. Wang & Slotine, 2005), consists of three steps: i) demonstrate con- traction of a single WTA, ii) merge two WTAs by demonstrating that inhibitory units synchronize, and iii) demonstrate contraction of the combined WTAs. We have previously shown how contrac- tion analysis can be applied to reasoning over the stability and functionality of WTA circuits (Rutishauser et al., 2011). Here, we apply and extend the same methods to this new circuit. Contraction analysis is based on the Jacobians of the system. For purposes of analysis, but without loss of generality, we will base this section on a reduced system with only 1 possible winner for each WTA as shown in Fig 2C. The Jacobian of a single system is τ J =  l1α − 1 l1 − β1 −1 0 0 l3β2  0 l2β3 −1 5 In a stable network, a constant external input to the first unit x1 will lead to a constant amplitude of x1 that is either an amplified or suppressed version of its input. The activation function f (x) = max(x, 0) is not continuously differentiable, but it is continuous in both space and time, so that contraction results can still be applied directly (Lohmiller & Slotine, 2000). Furthermore, the activation function is piecewise linear with a derivative of either 0 or 1. We exploit this property by inserting dummy terms lj, which can either be 0 or 1 according to the derivative of f (x): lj = d dx f (xj(t)). In this case, all units are active and thus l1 = l2 = l3 = 1. A system with Jacobian J is contracting if Θ J Θ−1 < 0 (7) where Θ is a constant transformation into an appropriate metric and F = Θ J Θ−1 is the gen- eralized Jacobian. If F is negative definite, the system is said to be contracting. We have shown previous (Rutishauser et al., 2011) (section 2.4) how to choose the constant transformation Θ and conditions that guarantee contraction for a WTA circuit where all excitatory units provide direct input to the inhibitory unit (Fig 1A). In summary, Θ = Q−1 where Q is defined based on the eigendecomposition J = QΛQ−1. In this case 0 < α < 2(cid:112)β1β2 0 < β1β2 < 1 (8) guarantee contraction for any such WTA of any size (Rutishauser et al., 2011). Structurally, the two versions of the WTA are equivalent in that an additional unit was added in the pathway of recurrent inhibition, but no inhibition is added or removed (Compare Fig 1A to Fig 2A). Thus, we can apply the same constraints by replacing the product β1β2 with β1β2β3 in above equations. This product is equivalent to the inhibitory loop gain. This reduction is verified as follows. Using the notation shown in Fig 2C, assume that T = 0 for x3 where T > 0 for the other units. Then, x3 + x3 = f (β2x1) x2 + x2 = f (β3x3 − T2) (9) At steady-state, x2 = f (β2β3x1), showing that x3 and x2 can be merged into a single unit x2 by providing input of weight β2β3 directly to unit x2 (Fig 2D). The first key result of this paper are the following limits for contraction of a single such WTA (Fig 2A): 0 < α < 2(cid:112)β1β2β3 0 < β1β2β3 < 1 (10) 3.3.1 Synchronizing two WTAs Next, we show that connecting two WTAs in the manner illustrated in Fig 2C results in synchro- nization of the two inhibitory units, which in turn leads to the two WTAs merging into a single WTA. Note that by synchronization, we mean that two variables have the same trajectory, or more generally that their difference is constant (in contrast to other meanings of synchronization i.e. in population coding). The approach is to show that adding excitatory connections β4 of sufficient strength will lead to the activity of the two inhibitory units x2 and y2 approaching a constant difference. 6 The Jacobian of the coupled system as shown in Fig 2C is with J1 = J2 = J (see Eq 6) and τ D1,2 = JC = (cid:21) D1 J2 (cid:20) J1 D2  0 0 0 0 0 l2,5β4 0 0 0  (11) (12) Following (Pham & Slotine, 2007; Rutishauser et al., 2011), synchronization occurs exponentially if the following holds: VJCVT < 0 (13) where V defines an invariant subset of the system such that Vx is constant, with x = (x1, x2, x3, y1, y2, y3). Here, we define synchrony as a regime where the differences between the inhibitory units x2 − y2 and between the interconnect units x3 − y3 are constant (although not necessarily zero). This results in (cid:20) 0 1 0 0 −1 0 0 1 0 0 0 −1 (cid:21) V = which embeds the two conditions. Condition (13) is satisfied if (14) (15) (16) (17) 1 < α 0 < β4 < β3 + 2 β3 < 2 The conditions on the interconnect-weight β4 guarantees that the dynamics are stable and that the inhibitory units synchronize. As long as β4 > 0 is sufficiently small but non-zero, the inhibitory parts of the system will synchronize. Realistically, β4 needs to be sufficiently large to drive the other inhibitory neuron above threshold and will thus be a function of the threshold T (see (Rutishauser et al., 2011), Eq 2.51). Here, synchrony is defined as their difference being constant. This in turn shows that the two WTAs have merged into a single WTA since the definition of a WTA is that each excitatory unit receives an equivalent amount of inhibition (during convergence but not necessarily afterwards, see simulations). This is our second key result. 3.3.2 Stability of pairwise combinations of WTAs The final step of the stability analysis are conditions for the coupling strength β4 > 0 such that the coupled system as shown in Fig 2C is contracting. The reasoning in this section assumes that the individual subsystems are contracting (as shown above). The Jacobian of the combined system remains Eq 11, where J1,2 are the Jacobians of the indi- vidual systems and C1,2 are the coupling terms. Rewriting the Jacobian of the second subsystem J2 by variable permutation y(cid:48) 3 = y2 allows expression of the system in the form of 2 = y3 and y(cid:48) (cid:21) (cid:20) J1 D2 D1 J2 (cid:21) (cid:20) J1 E ET J2 JC = = 7 (cid:21) (cid:20) J1 D DT J2 where E = D1 (Eq 12). This transformation1 of J2 is functionally equivalent to the original system (thus, its contraction limits remain) but it allows expression of the connection between the systems in the symmetric form of Eq 17. This functionally equivalent system can now be analyzed using the approach that follows. A matrix of the form is negative definite if the individual systems J1,2 are negative definite and if J2 < DT J−1 1 D (Horn, 1985) (Page 472). Following (Slotine, 2003) (Section 3.4) and (Rutishauser et al., 2011) (Section 2.8), this implies that a sufficient condition for contraction is σ2(D) < λ(J1)λ(J2) where σ(D) is the largest singular value of D and equivalent to β4 in our case (all other elements of D are zero) and λ is the contraction rate of the individual subsystems. Since the two subsystems are equivalent, the contraction rates are also the same λ1 = λ2 = λ(J1,2). It thus follows that the coupled systems are stable if β4 < λ1. The contraction rate (Slotine, 2003; Rutishauser et al., 2011) of an individual subsystem is It the absolute value of the largest eigenvalue of the Hermitian part of F (also see Eq (7)). is λ1,2 = 1 contracting reduces to 2(−2 + α) for our system. Thus the condition for the two coupled systems to be β4 < 1 − α 2 (18) 3.3.3 Summary of boundary conditions In summary, the following conditions guarantee stability of both the single and combined system as well as hard competition between the two coupled systems (that is, only one of the WTAs can have a winner). These conditions can be relaxed if α < 1, which will permit a soft winner-take-all. 1 < α < 2(cid:112)β1β2β3 0 < β1β2β3 < 1 0 < β4 < 1 − α 2 (19) The lower bound on β4 is from the synchronization analysis, whereas the upper bound is from the stability analysis. These results illustrate the critical tradeoff between having enough strength to ensure functionality, while being weak enough to exclude instability. 3.4 Speed of winner selection How quickly will a system select a winner? For a single WTA, this question is answered by how quickly a single system contracts toward a winner and for a coupled system how quickly the two systems synchronize. One of the key advantages of contraction analysis is that the rate of contraction, and in this case the rate of synchronization, can be calculated explicitly. We will express the contraction and synchronization rate in terms of the time constants τ and its inverse, the decay constant. τ refers to the mean lifetime of exponential decay x(t) = N0e−t 1 τ = N0e−λt. λ = 1 τ is the decay constant. Both the contraction and the synchronization rate are expressed 1The variable permutation is equivalent to a transformation of J1 by the permutation matrix Θ: J2 = ΘJ1Θ−1  1 0 0 0 0 1 . 0 1 0 with Θ = 8 in the form of a decay constant λ. For example, the contraction rate of a system of the form N = −λN is equivalent to λ. Physiologically, the time constants τ in our system are experimentally determined membrane time constants that are typically in the range of 10-50ms (Koch, 1998; McCormick, Connors, Lighthall, & Prince, 1985; Koch, Rapp, & Segev, 1996; Brown, Fricke, & Perkel, 1981). For simplicity, we assume that all excitatory and inhibitory units have the same, but different, time constants τE and τI, respectively. While the exact values depend on the cell type and state of the neurons, it is generally the case that τI < τE due to the smaller cell bodies of inhibitory neurons (McCormick et al., 1985; Koch, 1998). The bounds (19) were calculated assuming equal time constants for all units. However, the same calculations yield to very similar bounds when assuming different time constants for inhibitory and excitatory units (Appendix C, (Rutishauser et al., 2011)). In this case the ratio of the time constants τI τE becomes an additional parameter for the parameter bounds. 3.4.1 Speed of synchronization The synchronization rate is equivalent to the contraction rate of the system defined in Eq 13 (Pham & Slotine, 2007), which is the absolute value of the maximal eigenvalue of the Hermitian part of V(cid:48)(RJC)V(cid:48)T . Here, the original τ−1 is replaced by the diagonal matrix R, with the appropriate τE, τI terms on the diagonal 2. The matrix V(cid:48) is an orthonormal version of V as defined in Eq 14, which here is simply V(cid:48) = The resulting synchronization rate (sync rate) is a function of the weights β3 (local inhibitory loop) and β4 (remote inhibitory loop). We assume β4 ≤ β3, which means that remote connectivity is weaker than local connectivity. However, qualitatively similar results can be found using the opposite assumption. For τE = τI, the sync rate is √ 2−1V. λs = 1 2τ (2 − β3 + β4) (20) Note the tradeoff between local and remote connectivity: stronger remote connectivity increases and stronger local connectivity decreases the speed of synchronization (the larger λ, the quicker the system synchronizes). For approximately equal connectivity strength β3 (cid:39) β4 or in general for β3, β4 (cid:28) 1, the sync rate is approximately τ−1. In general for τE (cid:54)= τI, the sync rate is λs = √ τE +τI− (3 −(cid:112)1 + 4(β3 − β4)2). Again, for β3 (cid:39) β4, the sync rate is E−2τE τI +(1+(β3−β4)2)τ 2 τ 2 . For example, 2τE τI I 2 this reduces to λ = 1 2τI for τI = τE approximately τ−1 . In conclusion, the sync rate is thus approximately equal to the contraction rate of the inhibitory units. Thus, synchronization occurs very quickly (20-50ms for typical membrane time constants). I 3.4.2 Speed of contraction The speed of selecting a winner (the contraction rate) for a single WTA can similarly be calculated based on the absolute value of the maximal eigenvalue of the Hermitian part of Θ(RJ)Θ−1 (7). Assuming τ = τE = τI, the contraction rate is λc = (2 − α) 1 2τ (21) 2For the example of JC (Eq 11), the diagonal terms are τ−1 E , τ−1 E , τ−1 I , τ−1 E , τ−1 E , τ−1 I 9 Note that the larger α, the longer it takes till the system converges. Qualitatively similar findings result for other ratios of τE and τI. For a typical value of α = 1.2 (see simulations below) and τ = 20ms, the contraction rate would be 20s−1. This equals a half-way time (time constant) of λ−1 = 50ms. For α = 1.5, this would increase to 80ms. The time it takes to find a winner is thus a multiple of the membrane time constant (in this example 20ms) and substantially slower than the time it takes to synchronize the network. In conclusion, synchronization is achieved first which is then followed by winner selection. 3.5 Coupling more than two WTAs So far we have shown how two different WTAs compete with each other after their inhibitory neurons are coupled. Similarly, more than two WTAs can compete with each other by all-to-all coupling of the inhibitory units, i.e. every WTA is connected with two β4 connections from and to every other WTA. Thus, the wiring complexity of this system scales as O(M 2) where M is the number of WTAs in the system (note that M is not the number of units but the number of WTAs). Notice also that the all-to-all coupling concerns only the sparse long-range excitatory connections and not the internal connectivity of the WTAs them-self. The same principle can be used to embed hierarchies or sequences of competition. Thus, in a network of such WTAs, some WTAs could be in direct competition with each other while others are not. Thus, for example, in a network of three WTAs A, B, and C relationships such as A competes with B and B competes with C are possible. In this case A does not directly compete with C. So if A has a winner, C can also have a winner. If B has a winner, however, neither B nor C can have a winner (see Fig 4D-F for a demonstration). Regardless of how many WTAs are combined and whether all compete with all or more selec- tively, the stability of the aggregated system is guaranteed if the individual sub-systems are stable and the coupling strengths β4 observe the derived bounds. While in themselves combinations of stable modules have no reason to be stable, certain combinations (such as the one we utilize) of contracting systems are guaranteed to be stable (Slotine & Lohmiller, 2001). This is a key benefit of Contraction Analysis for the analysis of neural circuits. 3.6 Numerical simulations We simulated several cases of the network to illustrate its qualitative behavior. We used Euler integration with δ = 0.01s. The analytically derived bounds offer a wide range of parameters for which stability as well as function is guaranteed. For the simulations, we chose parameters that verify all bounds discussed. First, we explored a simple system consisting of two WTAs with two possible winners each (Fig 3). Parameters were α = 1.2, β1 = 2, β2 = 3, β3 = β4 = 0.1 and T = 0. We found that any of the four possible winners can compete with each other irrespective of whether they reside on the first or second WTA (Fig 3B-D shows an example). The inhibitory units quickly synchronized (Fig 3C) their activity and reached the same steady-state amplitude (because β3 = β4)3. Second, we simulated a system with 3 WTAs using the same parameters (Fig 4). For all-to-all coupling, all 3 WTAs directly compete with each other (Fig 4A,B), i.e. there can only be one winner across the entire system. Again, the inhibitory units all synchronize quickly during and after convergence (Fig 4C). We also simulated the same system with more selective connectivity, 3If β3 = β4, it can be verified that x3(t) = y4(t) for all t > 0 if the initial values at t = 0 are equal. Thus, the two inhibitory neurons become exactly equivalent in this special case. 10 Figure 3: Simulation of merged WTA consisting of two WTAs with two excitatory units each (four possible winners). (A) Illustration of connectivity and notation (color code). (B) Activity as a function of time, for two excitatory units on two different WTAs, together with the external input provided to the same units. Notice how the network selects the winner appropriately. (C) Activity of the inhibitory units in each WTA. Note: y3 is slightly delayed for plotting purposes. (D) Activity of the two interconnect units. Notice how the output of the losing WTA y4 descends to zero after the competition has been resolved and the network has contracted. Units of time are multiplies of the time constant τ . Notice that the same color indicates the same unit throughout this figure (notation is shown in A). 11 α1β105010002468time [t / τ]activity 05010001234activity 05010001020activity x1y1I to x1I to y2x3y3x4y4time [t / τ]time [t / τ]A)B)C)D)β3β2β2β3β4β4β1β1β1α1α1α1 eliminating competition between WTAs 1 and 3 (Fig 4D). This arrangement allows either one winner if it is on WTA 2, or two winners if they are on WTAs 1 and 3. If the maximal activity is not on WTA 2, then the network permits 2 winning states. Otherwise, if the maximal input is on WTA 2 only 1 winner is permitted (see Fig 4E for an illustration). This configuration allows for partial competition. 4 Discussion Neural circuits commonly depend on negative feedback loops. Such recurrent inhibition is a crucial element of microcircuits from a wide range of species and brain structures (Shepherd & Grillner, 2010) and enables populations of neurons to compute non-linear operations such as competition, decision making, gain control, filtering, and normalization. However, when considering biologically realistic versions of such circuits additional factors such as wiring length, specificity and complexity become pertinent. Here, we are principally concerned with the superficial layers of neocortex where the average distance of intracortical inhibitory connections is typically much shorter than the ex- citatory connections (Bock et al., 2011; Binzegger et al., 2004; Perin, Berger, & Markram, 2011; Katzel et al., 2011; Adesnik & Scanziani, 2010). In contrast, in invertebrates an inhibitory neuron has been identified that receives input from and projects back to all Kenyon cells (which are excita- tory) (Papadopoulou, Cassenaer, Nowotny, & Laurent, 2011). This neuron has been demonstrated to perform response normalization, making this system a direct experimental demonstration of competition through shared inhibition. No such system has yet been identified in the cortex. The number of excitatory neurons that can be contacted by an inhibitory neuron thus poses a limit on how many excitatory neurons can compete directly with one another (in terms of numbers and distance). Other models, such as those based on Mexican-hat type inhibitory surrounds (Hertz et al., 1991; Willshaw & Malsburg, 1976; Soltani & Koch, 2010), even require that inhibitory connectivity be longer range than the excitatory. These anatomical constraints have been used to argue that models such as the WTA are biologically unrealistic and as such of limited use. We have demonstrated here, by theoretical analysis and simulation, that it is possible to ex- tend such circuits by merging several independent circuits functionally, through synchronization of their inhibitory interneurons. This extension allows the construction of large, spatially distributed circuits that are composed of small pools of excitatory units that share an inhibitory neuron. We have applied and proved by non-linear contraction analysis that systems combined in this manner are inherently stable and that arbitrary aggregation by inhibitory synchrony of such sub-systems results in a stable system. This composition of subcircuits removes the limits on maximal circuit size imposed by anatomical wiring constraints on inhibitory connectivity, because the synchrony between local inhibitory neurons is achieved entirely by excitatory connectivity which can possi- bly be long-range so permitting competition between excitatory units that are separated by long distances; for example, in different cortical areas. We show that the time necessary to achieve sychronization is much shorter than the time required to select a winner. Thus, synchronization is faster than winner selection, which can thus proceed robustly across long-range connections that enforce synchronization. Further, selective synchronization between some WTAs but not oth- ers allows partial competition between some but not other WTAs (see Fig 4). The strength of these long-range connections could be modulated dynamically to enable/disable various competi- tions between two populations conditional on some other brain state. This modulation could be implemented by a state-dependent routing mechanism (Rutishauser & Douglas, 2009). There are several possibilities of mapping the abstract units in our model to real physiologi- 12 Figure 4: Simulation of merged WTA consisting of three WTAs with three possible winners each. (A-C) Case 1: Pairwise coupling allows all-to-all competition. (A) Illustration of connectivity. Filled circles are excitatory neurons, filled rectangles inhibitory. The activity of units with colored fills are shown as a function of time. (B) Activity of one excitatory unit for each WTA (bold lines) and the external input to each (dashed lines). Notice that there can be only one winner among the three WTAs. (C) Activity of the inhibitory units, shifted in time to each other slightly for plotting only. (D-F) Case 2: selective coupling, allowing partial competition only between 1&2 and 2&3 but not 1&3. (E) Activity of the excitatory units for different cases. Notice that WTA 1 and 3 can have winners simultaneously, but not 2. Numbers indicate the WTA the winner belongs to. (F) Activity of the inhibitory units, illustrating synchrony in the presence of different absolute amplitudes. Units of time are multiplies of the time constant τ . Notice that the same color indicates the same unit throughout this figure (notation is shown in A). 13 010020002468activity 010020001234activity WTA 1WTA 2WTA 3inhib 1inhib 2inhib 3time [t / τ]time [t / τ]010020002468activity 0100200012activity time [t / τ]time [t / τ]inhib 1inhib 2inhib 3WTA 1WTA 2WTA 3WTA 1WTA 2WTA 3WTA 3WTA 2WTA 1A)B)C)D)E)F) cal neurons. Our units are mean-rate approximations of a small group of neurons. In terms of intra-cortical inhibition, these would lie anatomically close to each other within superficial layers of neocortex. Since such inhibitory connectivity would have only limited reach, each inhibitory subunit can only enforce competition across a limited number of closeby excitatory units. Compe- tition between different areas is made possible by synchronizing remote populations by long-range excitatory mechanisms in the way we propose. Direct long-range inhibition, on the other hand, is unlikely both intracortically and subcortically, since all known connections from the thalamus and basal ganglia to cortex are excitatory. Networks such as the LEGION network (D. Wang & Terman, 1995) assume global inhibitory input to all excitatory units in the network, which for the reasons we discuss is unlikely in the case of cortex. It would, however, be possible to implement a feasible version of the global inhibitory input by synchronizing many local inhibitory neurons using the mechanism we describe, resulting in an anatomically realistic version of the LEGION network. Functionally, the model presented here makes several testable predictions. Consider a sensory area with clearly defined features as possible winners, such as orientations. The model predicts that the inhibitory units would not be tuned to these features, particularly if the number of possible winners is large. This is because the connectivity to the inhibitory units is not feature specific. Experimental studies indicate that this is indeed the case: units that functionally represent different tuning project to the same inhibitory unit, resulting in untuned inhibitory activity (Bock et al., 2011; Fino & Yuste, 2011; Kerlin, Andermann, Berezovskii, & Reid, 2010; Kuhlman, Tring, & Trachtenberg, 2011; Hofer et al., 2011). Secondly, this model predicts that inhibitory activity between two different areas or parts of the same area can either be highly synchronous or completely decoupled depending on whether at present the two are competing or functioning independently. This thus predicts that synchrony of inhibitory units should be affected by manipulations that manipulate competition, such as top-down attention. Our model suggests that synchronized populations of inhibitory neurons are crucial for enforcing competition across several subpopulations of excitatory neurons. It further suggests that the larger the number and spatial distribution of such synchronized inhibitory units, the larger the number of units that compete with each other. Experimentally, synchronized modulation of inhibitory neurons is a common phenomena that is believed to generate the prominent gamma rhythm trig- gered by sensory stimulation in many areas (Fries, Nikoli, & Singer, 2007; Whittington, Traub, & Jefferys, 1995; Traub, Whittington, Stanford, & Jefferys, 1996). Recent experiments have utilized stimulation of inhibitory neurons (Cardin et al., 2009; Sohal, Zhang, Yizhar, & Deisseroth, 2009; Szucs, Huerta, Rabinovich, & Selverston, 2009) to increase or decrease their synchronization with direct observable effects on nearby excitatory neurons such as, for example, increased or decreased amplitude and precision of evoked responses relative to how strongly the inhibitory neurons were synchronizing. Note that our proposal for this function of inhibitory synchrony is distinct and independent from the proposal that gamma-band synchrony serves to increase readout efficacy by making spikes arrive co-incidentally from a large number of distributed sources (Tiesinga, Fellous, & Sejnowski, 2008; Singer & Gray, 1995). Here, we propose that an additional function of such synchrony is to allow select populations of excitatory neurons to compete with each other because they each receive inhibition at the same time. 14 5 Appendix: Contraction Analysis This section provides a short summary of contraction analysis. We have previously published the detailed methods of applying contraction theory to WTA circuits (Rutishauser et al., 2011). Essentially, a nonlinear time-varying dynamic system will be called contracting if arbitrary initial conditions or temporary disturbances are forgotten exponentially fast, i.e., if trajectories of the perturbed system return to their unperturbed behavior with an exponential convergence rate. A relatively simple algebraic conditions can be given for this stability-like property to be verified, and this property is preserved through basic system combinations and aggregations. A nonlinear contracting system has the following properties (Lohmiller & Slotine, 1998, 2000; Slotine, 2003; W. Wang & Slotine, 2005) • global exponential convergence and stability are guaranteed • convergence rates can be explicitly computed as eigenvalues of well-defined Hermitian ma- trices • combinations and aggregations of contracting systems are also contracting • robustness to variations in dynamics can be easily quantified Before stating the main contraction theorem, recall first the following properties: The symmet- 2(A + A∗T ). A complex square matrix A is Hermitian if AT = A∗ ric part of a matrix A is AH = 1 , where T denotes matrix transposition and ∗ complex conjugation. The Hermitian part AH of any 2(A + A∗T ) . All eigenvalues of a Hermitian complex square matrix A is the Hermitian matrix 1 matrix are real numbers. A Hermitian matrix A is said to be positive definite if all its eigenvalues are strictly positive. This condition implies in turn that for any non-zero real or complex vector x, x∗T Ax > 0. A Hermitian matrix A is called negative definite if −A is positive definite. A Hermitian matrix A(x, t) dependent on state or time will be called uniformly positive definite if there exists a strictly positive constant such that for all states x and all t ≥ 0 the eigenvalues of A(x, t) remain larger than that constant. A similar definition holds for uniform negative definite- ness. Consider now a general dynamical system in Rn, x = f (x, t) (22) with f a smooth non-linear function. The central result of Contraction Analysis, derived in (Lohmiller & Slotine, 1998) in both real and complex forms, can be stated as: Theorem Denote by ∂f ∂x the Jacobian matrix of f with respect to x. Assume that there exists a complex square matrix Θ(x, t) such that the Hermitian matrix Θ(x, t)∗T Θ(x, t) is uniformly positive definite, and the Hermitian part FH of the matrix F = Θ + Θ ∂f ∂x Θ−1 (cid:18) (cid:19) is uniformly negative definite. Then, all system trajectories converge exponentially to a single trajectory, with convergence rate supx,t λmax(FH) > 0. The system is said to be contracting, F is called its generalized Jacobian, and Θ(x, t)∗T Θ(x, t) its contraction metric. The contraction rate is the absolute value of the largest eigenvalue (closest to zero, although still negative) λ = λmaxFH. 15 In the linear time-invariant case, a system is globally contracting if and only if it is strictly stable, and F can be chosen as a normal Jordan form of the system, with Θ a real matrix defining the coordinate transformation to that form (Lohmiller & Slotine, 1998). Alternatively, if the system is diagonalizable, F can be chosen as the diagonal form of the system, with Θ a complex matrix diagonalizing the system. In that case, FH is a diagonal matrix composed of the real parts of the eigenvalues of the original system matrix. References Adesnik, H., & Scanziani, M. (2010). Lateral competition for cortical space by layer-specific horizontal circuits. Nature, 464 , 1155 -- 1160. Amari, S., & Arbib, M. (1977). Competition and cooperation in neural nets. In J. Metzler (Ed.), Systems neuroscience (p. 119-165). San Diego, CA: Academic Press. Binzegger, T., Douglas, R. J., & Martin, K. A. (2004). A quantitative map of the circuit of cat primary visual cortex. J Neurosci, 24 (39), 8441-53. Bock, D. D., Lee, W. C., Kerlin, A. M., Andermann, M. L., Hood, G., Wetzel, A. W., et al. (2011). Network anatomy and in vivo physiology of visual cortical neurons. Nature, 471 , 177 -- 182. Brown, T. H., Fricke, R. A., & Perkel, D. H. (1981). Passive electrical constants in three classes of hippocampal neurons. Journal of Neurophysiology, 46 (6), 1360. Cardin, J. A., Carlen, M., Meletis, K., Knoblich, U., Zhang, F., Deisseroth, K., et al. (2009). Driving fast-spiking cells induces gamma rhythm and controls sensory responses. Nature, 459 , 663 -- 667. Coultrip, R., Granger, R., & Lynch, G. (1992). A cortical model of winner-take-all competition via lateral inhibition. Neural Networks, 5 , 47-54. Douglas, R., Koch, C., Mahowald, M., Martin, K., & Suarez, H. (1995). Recurrent excitation in neocortical circuits. Science, 269 (5226), 981-5. Douglas, R., & Martin, K. (2004). Neuronal circuits of the neocortex. Annu Rev Neurosci , 27 , 419 -- 451. Fino, F., & Yuste, R. (2011). Dense inhibitory connectivity in neocortex. Neuron, 69 , 1188-1203. Fries, P., Nikoli, D., & Singer, W. (2007). The gamma cycle. Trends Neurosci., 30 , 309 -- 316. Hahnloser, R., Sarpeshkar, R., Mahowald, M., Douglas, R., & Seung, H. S. (2000). Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit. Nature, 405 (6789), 947-51. Hertz, J., Krogh, A., & Palmer, R. (1991). Introduction to the theory of neural computation. Redwood City,CA: Addison-Wesley. Hofer, S. B., Ko, H., Pichler, B., Vogelstein, J., Ros, H., Zeng, H., et al. (2011). Differential connectivity and response dynamics of excitatory and inhibitory neurons in visual cortex. Nature Neuroscience, 14 , 1045-1052. Horn, R. (1985). Matrix analysis. Cambridge University Press. Katzel, D., Zemelman, B. V., Buetfering, C., Wolfel, M., & Miesenbock, G. (2011). The columnar and laminar organization of inhibitory connections to neocortical excitatory cells. Nature Neuroscience, 14 , 100 -- 107. Kerlin, A. M., Andermann, M. L., Berezovskii, V. K., & Reid, R. C. (2010). Broadly tuned response properties of diverse inhibitory neuron subtypes in mouse visual cortex. Neuron, 67 (5), 858-71. 16 Koch, C. (1998). Biophysics of Computation: Information Processing in Single Neurons (Compu- tational Neuroscience). Oxford University Press. Koch, C., Rapp, M., & Segev, I. (1996). A brief history of time (constants). Cerebral Cortex , 6 (2), 93 -- 101. Kuhlman, S. J., Tring, E., & Trachtenberg, J. T. (2011). Fast-spiking interneurons have an initial orientation bias that is lost with vision. Nat Neurosci, 14 , 1121-1123. Lohmiller, W., & Slotine, J. (1998). On contraction analysis for nonlinear systems. Automatica, 34 (6). Lohmiller, W., & Slotine, J. (2000). Nonlinear process control using contraction theory. AIChE Journal , 46 (3), 588-596. Maass, W. (2000). On the computational power of winner-take-all. Neural Computation, 12 , 2519-2536. McCormick, D. A., Connors, B. W., Lighthall, J. W., & Prince, D. A. (1985). Comparative electrophysiology of pyramidal and sparsely spiny stellate neurons of the neocortex. Journal of Neurophysiology, 54 (4), 782 -- 806. Neftci, E., Chicca, E., Indiveri, G., Cook, M., & Douglas, R. J. (2010). State-dependent sensory processing in networks of vlsi spiking neurons. In IEEE international symposium on circuits and systems, ISCAS 2010. Papadopoulou, M., Cassenaer, S., Nowotny, T., & Laurent, G. (2011). Normalization for sparse encoding of odors by a wide-field interneuron. Science, 332 , 721 -- 725. Perin, R., Berger, T. K., & Markram, H. (2011). A synaptic organizing principle for cortical neuronal groups. Proc. Natl. Acad. Sci. U.S.A., 108 , 5419 -- 5424. Pham, Q., & Slotine, J. (2007). Stable concurrent synchronization in dynamic system networks. Neural Networks, 20 (1), 62-77. Rabinovich, M. I., Huerta, R., Volkovskii, A., Abarbanel, H. D. I., Stopfer, M., & Laurent, G. (2000). Dynamical coding of sensory information with competitive networks. Journal of Physiology-Paris, 94 (5-6), 465 - 471. Rutishauser, U., & Douglas, R. (2009). State-dependent computation using coupled recurrent networks. Neural Computation, 21 (2). Rutishauser, U., Douglas, R., & Slotine, J. (2011). Collective stability of networks of winner-take- all circuits. Neural computation, 23 (3). Shepherd, G. M., & Grillner, S. (2010). Handbook of brain microcircuits. Oxford University Press. Shepherd, G. M., Stepanyants, A., Bureau, I., Chklovskii, D., & Svoboda, K. (2005). Geometric and functional organization of cortical circuits. Nat Neurosci, 8 (6), 782-90. Singer, W., & Gray, C. M. (1995). Visual feature integration and the temporal correlation hy- pothesis. Annu. Rev. Neurosci., 18 , 555 -- 586. Slotine, J. (2003). Modular stability tools for distributed computation and control. International Journal of Adaptive Control and Signal Processing, 17 , 397-416. Slotine, J., & Lohmiller, W. (2001). Modularity, evolution, and the binding problem: a view from stability theory. Neural Networks, 14 , 137-145. Sohal, V. S., Zhang, F., Yizhar, O., & Deisseroth, K. (2009). Parvalbumin neurons and gamma rhythms enhance cortical circuit performance. Nature, 459 , 698 -- 702. Soltani, A., & Koch, C. (2010). Visual saliency computations: mechanisms, constraints, and the effect of feedback. Journal of Neuroscience, 22 , 12831-43. Szucs, A., Huerta, R., Rabinovich, M. I., & Selverston, A. I. (2009). Robust microcircuit synchro- nization by inhibitory connections. Neuron, 61 , 439 -- 453. 17 Tiesinga, P., Fellous, J. M., & Sejnowski, T. J. (2008). Regulation of spike timing in visual cortical circuits. Nat. Rev. Neurosci., 9 , 97 -- 107. Traub, R. D., Whittington, M. A., Stanford, I. M., & Jefferys, J. G. (1996). A mechanism for generation of long-range synchronous fast oscillations in the cortex. Nature, 383 , 621 -- 624. Wang, D., & Terman, D. (1995). Locally excitatory globally inhibitory oscillator networks. IEEE Transactions on Neural Networks, 6 (1), 283-286. Wang, W., & Slotine, J. (2005). On partial contraction analysis for coupled nonlinear oscillators. Biological Cybernetics, 91 (1). Whittington, M. A., Traub, R. D., & Jefferys, J. G. (1995). Synchronized oscillations in interneuron networks driven by metabotropic glutamate receptor activation. Nature, 373 , 612 -- 615. Willshaw, D. J., & Malsburg, C. von der. (1976). How patterned neural connections can be set up by self-organization. Proc. R. Soc. Lond., B, Biol. Sci., 194 , 431 -- 445. Yabuta, N. a. C. E. (1998). Cytochrome-oxidase blobs and intrinsic horizontal connections of layer 2/3 pyramidal neurons in primate v1. Vis Neuroscience, 15 , 1007-1027. 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. 18
1506.01033
1
1506
2015-06-02T20:05:05
Multiplexed Neural Recording Down a Single Optical Fiber via Optical Reflectometry with Capacitive Signal Enhancement
[ "q-bio.NC", "physics.optics" ]
We introduce a fiber-optic architecture for neural recording without contrast agents, and study its properties theoretically. Our sensor design is inspired by electrooptic modulators, which modulate the refractive index of a waveguide by applying an electric field across an electrooptic core material, and allows recording of the activities of individual neurons located at points along a 10 cm length of optical fiber with 20 um axial resolution, sensitivity down to 100 uV and a dynamic range of up to 1V using commercially available optical reflectometers as readout devices. A key concept of the design is the ability to create an "intensified" electric field inside an optical waveguide by applying the extracellular voltage from a neural spike over a nanoscopic distance. Implementing this concept requires the use of ultrathin high-dielectric capacitor layers. If suitable materials can be found -- possessing favorable properties with respect to toxicity, ohmic junctions, and surface capacitance -- then such sensing fibers could, in principle, be scaled down to few-micron cross-sections for minimally invasive neural interfacing. Custom-designed multi-material optical fibers, probed using a reflectometric readout, may therefore provide a powerful platform for neural sensing.
q-bio.NC
q-bio
Multiplexed Neural Recording Down a Single Optical Fiber via Optical Reflectometry with Capacitive Signal Enhancement Samuel G. Rodriques,a,b,* Adam H. Marblestone,a,* Max Mankin,c Lowell Wood, Edward Boydend aMIT Media Lab bMIT Department of Physics cHarvard University Department of Chemistry dMIT Media Lab and McGovern Institute, Departments of Brain and Cognitive Science and Biological Engineering Abstract. We introduce a fiber-optic architecture for neural recording without contrast agents, and study its properties theo- retically. Our sensor design is inspired by electrooptic modulators, which modulate the refractive index of a waveguide by applying an electric field across an electrooptic core material, and allows recording of the activities of individual neurons located at points along a 10 cm length of optical fiber with 20 µm axial resolution, sensitivity down to 100 µV and a dynamic range of up to 1 V using commercially available optical reflectometers as readout devices. A key con- cept of the design is the ability to create an “intensified” electric field inside an optical waveguide by applying the extracellular voltage from a neural spike over a nanoscopic distance. Implementing this concept requires the use of ultrathin high-dielectric capacitor layers. If suitable materials can be found – possessing favorable properties with respect to toxicity, ohmic junctions, and surface capacitance – then such sensing fibers could, in principle, be scaled down to few-micron cross-sections for minimally invasive neural interfacing. Custom-designed multi-material optical fibers, probed using a reflectometric readout, may therefore provide a powerful platform for neural sensing. Keywords: reflectometry, neural recording, fiber-optic, electrooptic, nanophotonics. Correspondence to: [email protected] *These authors contributed equally to this work. 1 Introduction The extracellular electrode is a classic neural recording technology. The electrode is essentially a conductive wire, insulated except at its tip, placed in the extracellular medium within a few hundred microns of a neuron of interest, where it samples the local voltage relative to a ground lead.1 This voltage differential, typically on the order of 100 µV in response to an action potential from a nearby neuron,2 is then amplified and digitized. The virtues of the electrode are twofold. First, the technique can reach single neuron precision by virtue of the electrode being inserted close to the measured neuron. Second, no exogenous contrast 1 agents (i.e., genetically encoded fluorescent proteins, voltage sensitive nanoparticles, chemical dyes) are necessary: the endogenously generated electric currents in the brain are sensed directly in the form of a voltage. Ideally, for a neurotechnology to be medically valuable for a large number of human patients in a reasonably short timescale, it should not require modification of the neuron. Yet, while multi-electrode arrays allow the insertion of many electrodes into a brain, electrodes have limitations2 in scaling to the simultaneous observation of large numbers of neurons. The electrical Johnson noise scales as R 1 2 , leading to voltage noise levels on the order of 10 µV in standard practice. Moreover, the bandwidth of an electrical wire is limited by the cross-sectional area of the wire, due to capacitance. These factors pose limits on the scalability of electrode-based recording, as described further in Section 3.2. In order to maintain the advantages of electrodes – single neuron precision based on endogenous neural signals – while enabling improved scaling performance, we turn to photonics. Telecommu- nications has moved from electrical to optical data transmission because of the high bandwidths and low power losses enabled by optics in comparison to electrical conductors;3 the same may be helpful for neural readout technologies. Because optical radiation heats brain tissue and scatters off tissue inhomogeneities, a wired (i.e., fiber or waveguide based) optical solution may be de- sirable, i.e., using optical fibers to guide light so that it need not travel through the tissue itself. Second, to minimize volume displacement, signals from many neurons should be multiplexed into each optical fiber. Third, ideally the sensing mechanism would rely only on endogenous signals, e.g., electrical, magnetic or acoustic fields from the firing neurons, rather than imposing a need for exogenous protein or nanoparticle contrast agents. With on the order of 100,000 neurons per mm3 2 in the cortex, or a median spacing of roughly one neuron per cube of size 21.5 µm, we require an axial resolution of sensing in the range of 20 µm. Note that every neuron in a mammalian brain is within a few tens of microns of the nearest capillary,4 well within the distance necessary for direct electrical sensing of the action potential,2 and thus in principle the fine microvessels of the cerebral vasculature could serve as a delivery route for neural activity sensors, if the fibers could be made sufficiently thin,5 i.e, well below 10 µm for the smallest capillaries. Thus, the system should be compatible with a variety of form factors, e.g., thin flexible fibers suitable for minimally invasive endovascular delivery,5, 6 or rigid pillars suitable for direct penetration of the brain parenchyma.7 Our proposed architecture is based on two powerful technologies developed by the photonics industry: fiber optic reflectometry, which enables optical fibers to act as distributed sensors,8–11 and electrooptic modulators based on the plasma dispersion effect, which generate large changes in the index of refraction of a waveguide in response to relatively small applied voltages.12–16 By combining reflectometry with electro-optic modulation, we demonstrate that it would be possible to do spatially multiplexed neural recording in a single optical fiber. 2 Design Principles Reflectometers are capable of measuring changes in the index of refraction along the length of an optical fiber by sending optical pulses down the length of the fiber and recording the times and magnitudes of returning reflections.9 We propose to use reflectometry to sense neural activity at many points along the length of an optical fiber, as shown in figure 1A. The local voltage at a given position along the fiber would modulate its local index of refraction via the free carrier dispersion effect, giving rise to reflections. A reflectometer located outside the brain would then determine, 3 A) B) Brain Tissue (Extracellular Fluid) neuron brain tissue neuronal electric field Weak Outer Conductor activity-modulated local reflection coefficient in fiber Optical Reflectometer Nanosheet Capacitor +++++++ Si Core buildup of free charge carriers Weak Inner Conductor Metallic Ground Layer Insulator 500 nm 12000 nm 7000 nm 2.5 nm 500 nm 2000 nm 500 nm 1500 nm C) Cext Vext Rext Router RESR Cins Rins Rinner Rground Fig 1 A) High-level architecture. An optical fiber inserted into the brain acts as a distributed sensor for neuronal activity, which is read out reflectometrically. B) Structure of the reflectometric probe. When a voltage is applied across the nanosheet capacitor insulator layer, free holes in the inner conductor and in the silicon build up on the surface of the insulator layer and alter the refractive index. In order to magnify number of charge carriers accumulated on the surface of the insulator layer, an insulating film with an extremely large dielectric constant is used. C) Circuit diagram of the device. The circuit diagram of the device consists of a resistor representing each of the material layers between the neuron and ground, and two capacitors, one of which (Cext) represents the interfacial capacitance and the other of which (Cins) represents the capacitance of the insulating layer. The effective series resistance RESR of the insulating region capacitor can be ignored provided the insulating region has dielectric loss tan(δ) ≪ 1, and the parallel resistance of the insulating layer Rins can be ignored provided it is much larger than Rground. If in addition Rground is chosen to be larger than the other resistances in the circuit, the capacitances Cins and Cext may be treated as series capacitances (see text). 4 at each time, the spatial profile of extracellular voltage along the length of the fiber. 2.1 Fiber-Optic Reflectometry The sensitivity of a reflectometer is given in terms of fractional power reflected relative to the input power, measured in decibels. To determine the magnitude of the reflections generated by a change in local refractive index inside a fiber, note that when an electromagnetic plane wave propagates in a material with refractive index n1, and is normally incident on a material with refractive index n2, the intensity reflected is given by the Fresnel equation n2 + n1(cid:19)2 R = (cid:18)n1 − n2 . Assuming small changes in n, we can approximate R by 2n (cid:19)2 R = (cid:18)∆n . (1) (2) Expressed in decibels, the magnitude of the reflections generated by the change in refractive index is 10 log10(R). (3) Below, we will assume the use of reflectometers that can achieve 20 µm spatial resolution and sensitivities down to −130 dB (corresponding to R ∼ 10−13) with 12 Hz repetition rates over 8.5 m. This corresponds to a measurement time of 1 ms for any given 10 cm segment of fiber, so using a similar device we anticipate that it would be possible to sense reflections along the length of a 10 cm fiber with a repetition rate of 1 kHz, 20 µm spatial resolution and sensitivity of −130 dB. 5 For example, the commercially available Luna OBR 5T-50 and ODiSI B systems exhibit an optical noise floor of about −120 to −125 dB and a dynamic range of −60 dB, while the OBR 4600 gives a < −130 dB optical noise floor. Optical phase noise associated with the laser is limiting in such optical frequency domain reflectometry (OFDR) systems;17 Littman-Metcalf external cavity tunable lasers, with narrow linewidths and low phase noise, can be swept at 1 kHz repetition rates over an optical frequency range of several THz, leading to an OFDR spatial resolution of roughly 20 µm. If additional speed is required, even faster reflectometers are available that achieve similar sensitivities; one reflectometer has been demonstrated that achieves −130 dB sensitivity over 2 km with millimeter spatial resolution and a 10 Hz repetition rate,18 corresponding to a sensing rate of 2 × 107 measurements per second, a factor of 4 improvement over the commercial reflectometer mentioned above. If additional sensitivity is needed, reflectometers can achieve sensitivities as low as −148 dB,19 while systems using fiber-based optical amplifiers and super-fluorescent fiber sources can achieve sensitivities as low as −160 dB, which is nearly shot-noise limited.20 2.2 Electro-optic modulation Silicon electro-optic modulators are widely used in photonics to alter the propagation of light through a material in response to an applied voltage.15, 21 Typical applications of electrooptic mod- ulators take the form of electrically controlled optical switches: signals on the order of 5 V are used to drive optical phase shifts on the order of π. These devices are optimized for GHz bandwidths, with the goal of providing high speed, low power microchip interconnects,14 with bandwidths up to 30 GHz possible.22 Here, however, we are interested in the application of similar device physics to a very different problem: sensing extracellular neuronal voltages on the order of 100 µV at 1 kHz 6 rates. Thus, our required switching rate is 1 millionfold slower, yet our required sensitivity is on the order of 1 millionfold better. We are thus concerned with the design of electrooptic modulators optimized for sensitivity rather than bandwidth. 2.2.1 Insufficiency of endogenous electric fields The voltages present in the extracellular medium due to neural activity range from 10 µV to 1000 µV, with electric fields on the order of 1 V m−1.2 These electric field arise from the fact that the extracellular voltage from a spiking neuron decays over a distance on the order of 100 µm. Note that the transmembrane voltage during an action potential is much larger, on the order of 100 mV. Although it is possible to measure electric fields down to microvolts per meter using refractive index sensors, such sensors rely on very long sensing distances and can only measure the average field over the length of the fiber.8, 23–25 For measurements that preserve spatial information, sensi- tivities down to hundreds of volts per meter have been obtained using the Stark shift.26–29 Thus, the endogenous electric fields in the brain are too small to be directly read out using such physical coupling strategies in conjunction with optical reflectometers. For sensing, it will therefore be necessary to create an intensified electric field E = V /d by causing the extracellular voltage V to drop over a nanoscopically narrow region of insulating mate- rial of thickness d. This intensified electric field could then in principle be detected electrooptically. 2.2.2 The free carrier dispersion effect The design shown in figure 1B consists of an extended multi-layer semiconductor waveguide on a grounded metal substrate, surrounded on three sides by insulation and on the fourth side by brain 7 tissue or extracellular fluid. The “outer conductor,” “inner conductor,” and “silicon” core layers are weak conductors which function as resistive layers between the brain and ground. Between the core and outer conductor there is an ultra-thin insulator layer that functions as a capacitor with large capacitance, over which most of the voltage drops. Light propagates through the silicon waveguide core, which is doped to achieve sufficiently high conductivity to allow most of the voltage to drop over the insulator layer. Note that the inner conductor is chosen to be thick enough to prevent optical attenuation due to the metal substrate (although there are other possible methods to reduce attenuation due to the metal, e.g., by removing the metal from the region directly under the waveguide, as in13), and the metal substrate is chosen thick enough to provide a high-fidelity ground throughout the fiber. Clearly, the index of refraction of the inner and outer conductors must be smaller than that of the silicon core, which is assumed to be around 3 at telecommunications frequencies. The design relies on the free carrier dispersion effect (also known as the plasma dispersion effect) in silicon: the index of refraction within the silicon core changes due to the injection of charge carriers into the silicon when an electric field is applied across the thin insulator layer.13, 30, 31 In addition to the free carrier effect, there exist other modalities of electrooptic modulation, such as the linear electrooptic (Pockels) effect,32 the quadratic electrooptic effect (Kerr) effect33 and the Stark effect.26 All of these effects benefit from reducing the thickness d of the insulator layer to create a large electric field V /d.34 However, the magnitude of the free carrier dispersion effect may be increased further by increasing the dielectric constant of the insulator layer. For a material with a suitably large value of ǫ/d, the change in refractive index due to the free carrier effect will be much larger than the changes that can be obtained via the other electrooptic effects. Many current integrated semiconductor electrooptic modulators are based on the free carrier dispersion 8 effect.12, 13 Changes in the index of refraction in the free carrier modulated sub-layer of the silicon may be modeled as changes in the overall effective index of refraction of the fiber.35 The magnitude of this effective change is given by weighting the magnitude of the change in the free carrier modulated layer by the percentage of power contained in that layer, i.e., ∆neff = (1 − η)∆nactive, (4) where nactive is the index of refraction in the free carrier modulated layer and 1 − η is the fraction of the power in the beam contained in the active region. We will denote by d the thickness of the insulator layer, by b the thickness of the layer of injected charge carriers in the silicon, and by a the remaining thickness of the silicon layer. An order-of-magnitude approximation for η is then given by η ∼= a (a + b) and we have ∆neff = (cid:18)1 − a (a + b)(cid:19) ∆nactive. (5) (6) For this reason, the silicon waveguide is chosen to be thin to maximize the percentage of the optical wave contained in the layer containing the injected charges. Because of the deep subwavelength thickness of the active layer, a precise calculation of ∆neff could be done using a full vectorial Maxwell simulation of the waveguide modes,13 but for our purposes the approximation of equation 6 suffices to illustrate the basic scaling. 9 Upon applying a voltage across the insulating layer capacitor, the density of charge carriers injected into the active layer inside the silicon core, denoted by ∆c, is simply given by the equation for capacitance: ∆c = Cins Ab ∆Vins, (7) where Cins is the capacitance of the insulating region, A is the area over which the charges are distributed, b is the thickness of the layer of injected charge carriers in the silicon, and ∆Vins is the voltage dropped over the insulating region. Equation (7) may be recast in terms of the total voltage ∆V applied over the device by introducing an effective capacitance Ceff, such that ∆c = Ceff Ab ∆V. (8) In practice, Ceff will only deviate significantly from Cins when the capacitance of the brain-fiber interface is significant (discussed below). The change in refractive index in the region with the injected charge is related to the change in the carrier concentration by a power law.31 When the injected carriers are holes, the magnitude of the electrooptic effect is greatest. The relation is then, for the change in refractive index in the injected charge region, ∆nactive,h = Ch(cid:20) Ceff Ab ∆V(cid:21)0.8 (9) where Ch is an empirically defined constant. 10 For silicon, the value of Ch is given by,13 for 1.55 µm light, Ch = −8.5 × 10−18 cm2.4. (10) Similar values are obtained for other semiconductors and other wavelengths.31, 36 To find the effective refractive index within the silicon waveguide, we multiply equation (9) by the volume factor 1 − η from equation (5). Assuming b ≪ a (i.e., that the injected charge layer is deeply sub-wavelength while the waveguide core thickness is on the same order as the wavelength), we find ∆neff ∼= Ch b0.2 a (cid:20) Ceff A ∆V(cid:21)0.8 . (11) Since b will be on the order of 10 nm, we will henceforth assume b0.2 ≈ 0.01 m0.2. 2.2.3 Effect of the brain-electrode interface The brain-electrode interface also has a finite capacitance Cext of approximately 1 F m−2,37, 38 which arises due to the presence of a double-layer and which can be calculated via the Gouy-Chapman equation. Figure 1C shows a circuit diagram of the device which includes this interfacial capaci- tance. At < 1000 Hz, and subject to appropriate materials choices (see section 3.1), the impedance of the circuit is dominated by the two capacitors Cins and Cext rather than by purely resistive ele- ments of the circuit. For this reason, we may ignore the purely resistive elements and treat the two capacitances as though they were in series. To a good approximation, therefore, the charge that accumulates on the surface of the insulating region in response to a voltage ∆V across the entire device is given by Q = Ceff∆V, 11 (12) where the effective capacitance Ceff of the surface and insulating region capacitors in series is Ceff = 1 + 1 Cext . 1 Cins The capacitance of the insulating region is given by Cins = ǫ0A ǫ d , (13) (14) where d is the thickness of the insulating region, ǫ is the relative permittivity, A is the sensing area of the device, and ǫ0 is the permittivity of free space. The capacitance per unit area is specified by ǫ/d, which will be the primary figure of merit for determining the sensitivity, noise, and dynamic range of the device. The effective capacitance Ceff is shown in figure 2A as a function of ǫ/d, assuming a surface capacitance per unit area of approximately 1 F m−237, 38 and a sensing length of 20 µm. Note that because a pair of series capacitors is dominated by the smaller capacitor, the silicon waveguide core is kept only 500 nm wide in order to ensure that the capacitance per unit length of the insulating layer remains smaller than the capacitance per unit length of the 12 µm-wide brain-fiber interface. 2.2.4 Noise The RMS amplitude of the Johnson noise in the voltage across the capacitor is given by VRMS = p4kBT Rc∆F (15) 12 where ∆F = 1/(2πRcCeff) and Rc is the total resistance governing charging/discharging of the capacitor. Hence VRMS ∼ pkBT /Ceff (16) which does not depend on the resistance Rc. 2.2.5 Dynamic range In addition to the constraints on the minimum resolvable voltage described above, the maximum resolvable voltage is limited by the total number of charge carriers available in the outer conductor layer. In particular, if a voltage V is applied over the device capacitance Ceff, a number of charge carriers equal to V Ceff/e must be recruited to the surface of the insulator layer, where e is the electron charge. Hence, if the outer conductor has a volume of V associated with each recording site and a free charge carrier density equal to c, the saturation voltage is given by Vmax = cVe Ceff . (17) For this reason, the outer conductor layer is chosen to occupy most of the thickness of the fiber in order to maximize the dynamic range. 3 Predicted properties The key figure of merit determining the properties of the device is ǫ/d, where ǫ is the dielectric constant of the insulating layer and d is the thickness of the insulating layer. The figure of merit is proportional to the capacitance per unit area of the insulating region, and it determines the sensitivity, noise, and dynamic range via equations (11), (16), and (17). In figure 2B, the sensitivity 13 of the device is shown as a function of ǫ/d for devices of width 12 µm. The blue, orange, and green lines correspond to the voltages which would be sensed at −120 dB, −130 dB, and −140 dB respectively at the given value of ǫ/d. The red line corresponds to the level of the Johnson noise at the given value of ǫ/d. The black dashed lines indicate the 10 µV and 100 µV ranges. The power law region (a straight line on the log-log plot) corresponds to the region in which Ceff ≈ Cins, so that the reflection coefficient R ∝ (ǫ∆V /d)0.8. For values of ǫ/d much greater than 1012 m−1, we have Ceff ≈ Cext, so the sensitivity does not improve with increasing ǫ/d. Clearly, in order to be able to sense a 100 µV signal using a reflectometer with −130 dB sen- sitivity, a value of ǫ/d > 2 × 1011 m−1 is required. With a reflectometer of sensitivity −140 dB, a 100 µV signal can be sensed with a value of ǫ/d as low as 4 × 1010 m−1. However, even with a reflectometer of sensitivity −150 dB, a value of ǫ/d > 1010 m−1 would be required to overcome Johnson noise. Finally, if a value of ǫ/d on the order of 2 × 1012 m−1 could be obtained, it would be possible to detect signals as small as 10 µV with commercially available reflectometers with −130 dB sensitivity. 3.1 Required material properties The primary electrical constraint on the device is that the impedance at 1 kHz must be dominated by the insulator capacitance. If the capacitance per unit area of the capacitor is 0.1 F m−1, corre- sponding to a value of ǫ/d equal to 1010 m−1, then the capacitance of a region with width 12 µm and length 20 µm is 24 pF, corresponding to an impedance of 4 × 107 Ω at 1 kHz. Assuming that the metal layer has a resistivity no greater than 100 nΩ m (10x that of silver), if the metal layer is made at least 500 nm thick, it will have a resistance of 10 000 Ω along the entire length of the fiber, which is much less than that of a single sensing region of the capacitor, and which is also large enough for 14 A Capacitance (F) 10 10 10 10 B C 10 10 10 10 10 Voltage (V) 10 10 10 10 10 10 10 10 10 10 10 Voltage (V) 10 5 1 0.50 0.10 0.05 10 10 10 10 ε/d 10 Fig 2 Properties of the design parametrized by ǫ/d. A) The effective capacitance Ceff given in equation (13) is shown as a function of ǫ/d assuming the dimensions shown in figure 1, and a capacitance of 1 F m−2 at the surface of the device. B) The change in voltage required to give rise to a reflection at the −120 dB level (blue), the −130 dB level (orange), and the −140 dB level (green) is shown as function of ǫ/d, obtained by solving equation (11) for ∆V . Also shown is the magnitude of the Johnson noise given in equation (16) as a function of ǫ/d (red). The black dashed lines correspond to 10 µV and 100 µV. For example, detecting 100 µV signals at the −130 dB level requires a figure of merit ǫ/d > 1011. C) The saturation voltage Vmax given in equation (17) is shown as a function of ǫ/d assuming a carrier concentration of 5 × 1016 cm−3 in the outer conductor. 15 the input impedance of an implanted recording device.39 Moreover, in this case, the RC-time con- stant of the system consisting of the metal ground and the effective capacitor (for a single sensing region) will be 240 ns, which is much less than 1 ms. The resistance of the inner conductor, outer conductor, and core will be negligible compared to the huge capacitive impedance, provided they are chosen to be semiconductors. The inner conductor must be chosen to make ohmic junctions with the metal and silicon layers. In addition, the outer conductor is constrained by the requirement that its conductivity be less than or on the order of that of the extracellular medium (approximately 0.1 S m−1 to 0.3 S m−140, 41) to avoid lateral propagation of the electrical signals from the brain along the length of the fiber. Finally, to maximize the saturation voltage, it is necessary to choose the outer conductor to have a large free carrier density. These two properties can be satisfied simultaneously using a semicon- ductor with very low mobility. As an example, nickel oxide is a p-type semiconductor known to have a mobility between 0.1 cm2 V−1 s−1 and 1 cm2 V−1 s−1, but with a conductivity of 0.1 S m−1 to 1 S m−1 in the undoped state,42, 43 corresponding to a free carrier concentration on the order of 5 × 1016 cm−3, which is ideal for our application. With these material properties, the saturation voltage can be calculated from equation (17) as a function of the figure of merit ǫ/d, and is shown in figure 2C. Alternatively, films of semiconducting nanoparticles may provide similarly large car- rier concentrations at low mobilities, and may be significantly simpler to fabricate.44 16 3.2 Comparison of electrical and optical performances The bandwidth of an electrical wire is limited by its cross-sectional area by BW ≈ A log(1 + α) (1 + α)24π2ǫrǫ0ρL2 (18) where ρ is the resistivity of the wire, L its length, A its cross-sectional area, α is the thickness of the insulation as a fraction of the thickness of the wire, ǫr is the dielectric constant of the medium surrounding the conductor and ǫ0 is the permittivity of free space. For any given value of the bandwidth, equation (18) specifies a minimum total cross-sectional conductor area that must be available to convey the information. Note that, according to the Shannon-Hartley theorem, the bandwidth required to achieve an information rate I is given in terms of the signal power S and noise power N by BW = I log2(1 + S/N) . (19) Optical waves can be spatially confined to channels with thickness on the order of the wave- length of light (or below, e.g., using high-index-contrast polymer fibers or nanoslot waveguides45), yet can transmit data at rates of terabits per second over long distances. To be specific, an infor- mation rate of roughly I = 104 bits per second per neuron is required for analog neural recording. At a signal to noise ratio of 10, sensing 1000 neurons requires a bandwidth of roughly 3 × 107 Hz via equation (19). In theory, this is trivial even for an optical channel of 500 nm width and 10 cm length; yet for an electrical channel consisting of a 500 nm wide electrical conductor, 500 nm in- sulation and 10 cm length, this is already at the limit of achievable bandwidth as calculated from equations (18) and (19). In practice, the field of electrical recording has advanced greatly in re- 17 cent years, and efforts are underway to package ∼ 1000 electrical recording sites into shanks with 50 µm width.1, 46 However, our calculations suggest that with further miniaturization, optical meth- ods will surpass the fundamental limits of wired electrodes in terms of space efficiency. Thus, improvements in nanophotonic miniaturization could become a key enabler for the comprehensive optical readout of neural circuit dynamics. 4 Discussion Ultra-large-scale neural recording is highly constrained both by physics and by the biology of the brain.2 Here we have argued that an architecture for scalable neural recording could combine a) the use of optical rather than electronic signal transmission to maximize bandwidth, b) confined rather than free space optics to obviate the effects of light scattering and absorption in the tissue, c) spatial or wavelength multiplexing within each optical fiber in order to minimize total tissue volume displacement, d) a thin form factor to enable potential deployment of fibers via the cerebral vasculature, e) direct electrical sensing to remove the need for exogenous dyes or for genetically encoded contrast agents. In our proposed design, the 100 µV scale extracellular voltage resulting from a neuronal spike is applied across a nanoscopically thin, high-dielectric capacitor. Charging of the capacitor results in the buildup of free charge carriers in the neighboring silicon waveguide core, altering the lo- cal refractive index of the silicon and causing a detectable optical reflection. Reflectometry then enables multiplexed readout of these spike-induced reflections. A key challenge in implementing such a design is to achieve a figure of merit ǫ/d for the ca- pacitor in excess of 1011. While current nanosheet capacitors47–49 approach ǫ/d ≈ 1011, achieving sufficiently high figure of merit in a practical fabrication process may be difficult. Alternatively, 18 there exist ultra-high-dielectric (ǫ > 105) materials such as calcium copper titanate,50 but much of their dielectric behavior derives from boundaries between microscopic crystal grains, and whether these materials can maintain their dielectric properties in sufficiently thin nanoscopic layers is un- clear. Another challenge is to maintain high carrier concentrations in the outer conductor without creating high electron mobilities which could lead to attenuation of the optical wave propagating in the nearby core. The 12 µm thickness of the device presented here is likewise limited by material concerns. If the carrier concentration of the outer conductor could be increased while maintaining low mobility, further miniaturization of the device would be possible. If appropriate materials combinations can be fabricated, we have shown that the device could achieve the requisite sensitivity, noise level, dynamic range and response time for recording both neural spikes and local field potentials. More broadly, our results suggest that integrated photonics could enable highly multiplexed readout of neuronal electrical signals via purely optical channels. Acknowledgments We thank Jorg Scholvin, Michal Lipson, Mian Zhang, Chris Phare, Konrad Kording, Dario Amodei, Deblina Sarkar and George Church for helpful discussions. Funding was contributed by the Fan- nie and John Hertz Foundation (SR, AM, MM), the NIH Director’s Pioneer Award (ESB) and NIH grants 1U01MH106011 (ESB) and 1R24MH106075-01 (ESB, AM). All code will be posted at http://scalablephysiology.org/fiber for use by the community. References 1 G. Buzs´aki, E. Stark, A. Ber´enyi, D. Khodagholy, D. R. Kipke, E. Yoon, and K. D. Wise, “Tools for probing local circuits: High-density silicon probes combined with optogenetics,” Neuron 86(1), 92–105 (2015). 19 2 A. H. Marblestone, B. M. Zamft, Y. G. Maguire, M. G. Shapiro, T. R. Cybulski, J. I. Glaser, D. Amodei, P. B. Stranges, R. Kalhor, D. A. Dalrymple, et al., “Physical principles for scal- able neural recording,” Frontiers in computational neuroscience 7 (2013). 3 M. Hilbert and P. L´opez, “The worlds technological capacity to store, communicate, and compute information,” Science 332(6025), 60–65 (2011). 4 P. Blinder, P. S. Tsai, J. P. Kaufhold, P. M. Knutsen, H. Suhl, and D. Kleinfeld, “The corti- cal angiome: an interconnected vascular network with noncolumnar patterns of blood flow,” Nature neuroscience 16(7), 889–897 (2013). 5 R. R. Llinas, K. D. Walton, M. Nakao, I. Hunter, and P. A. Anquetil, “Neuro-vascular central nervous recording/stimulating system: Using nanotechnology probes,” Journal of Nanopar- ticle Research 7(2-3), 111–127 (2005). 6 M. R. Bower, M. Stead, J. J. Van Gompel, R. S. Bower, V. Sulc, S. J. Asirvatham, and G. A. Worrell, “Intravenous recording of intracranial, broadband eeg,” Journal of neuro- science methods 214(1), 21–26 (2013). 7 J. Du, T. J. Blanche, R. R. Harrison, H. A. Lester, and S. C. Masmanidis, “Multiplexed, high density electrophysiology with nanofabricated neural probes,” PLoS One 6(10), e26204 (2011). 8 J. Jarzynski and R. P. De Paula, “Fiber optic electric field sensor technology,” in Cambridge Symposium-Fiber/LASE’86, 48–55, International Society for Optics and Photonics (1987). 9 W. V. Sorin, “High-resolution optical fiber reflectometry techniques,” in Fibers’ 92, 109–118, International Society for Optics and Photonics (1993). 20 10 L. Palmieri, “Distributed polarimetric measurements for optical fiber sensing,” Optical Fiber Technology 19(6), 720–728 (2013). 11 X. Bao and L. Chen, “Recent progress in distributed fiber optic sensors,” Sensors 12(7), 8601–8639 (2012). 12 A. Liu, L. Liao, D. Rubin, H. Nguyen, B. Ciftcioglu, Y. Chetrit, N. Izhaky, and M. Paniccia, “High-speed optical modulation based on carrier depletion in a silicon waveguide,” Optics Express 15(2), 660–668 (2007). 13 A. Liu, R. Jones, L. Liao, D. Samara-Rubio, D. Rubin, O. Cohen, R. Nicolaescu, and M. Pan- iccia, “A high-speed silicon optical modulator based on a metal–oxide–semiconductor capac- itor,” Nature 427(6975), 615–618 (2004). 14 Q. Xu, B. Schmidt, S. Pradhan, and M. Lipson, “Micrometre-scale silicon electro-optic mod- ulator,” Nature 435(7040), 325–327 (2005). 15 R. Soref, “The past, present, and future of silicon photonics,” Selected Topics in Quantum Electronics, IEEE Journal of 12(6), 1678–1687 (2006). 16 B. Jalali and S. Fathpour, “Silicon photonics,” Lightwave Technology, Journal of 24(12), 4600–4615 (2006). 17 B. Soller, D. Gifford, M. Wolfe, and M. Froggatt, “High resolution optical frequency domain reflectometry for characterization of components and assemblies,” Optics Express 13(2), 666–674 (2005). 18 D. Gifford, M. Froggatt, M. Wolfe, S. Kreger, A. Sang, and B. Soller, “Millimeter resolution optical reflectometry over up to two kilometers of fiber length,” in Avionics, Fiber-Optics and Photonics Technology Conference, 2007 IEEE, 52–53, IEEE (2007). 21 19 W. V. Sorin and D. M. Baney, “Measurement of rayleigh backscattering at 1. 55 mu m with 32 mu m spatial resolution,” IEEE Photonics Technology Letters 4(4), 374–376 (1992). 20 K. Takada, T. Kitagawa, M. Shimizu, and M. Horiguchi, “High-sensitivity low coherence re- flectometer using erbium-doped superfluorescent fibre source and erbium-doped power am- plifier,” Electronics Letters 29(4), 365–367 (1993). 21 M. Lipson, “Guiding, modulating, and emitting light on silicon-challenges and opportuni- ties,” Journal of Lightwave Technology 23(12), 4222 (2005). 22 C. T. Phare, Y.-H. D. Lee, J. Cardenas, and M. Lipson, “30 ghz zeno-based graphene electro- optic modulator,” arXiv preprint arXiv:1411.2053 (2014). 23 S. T. Vohra and F. Bucholtz, “Fiber-optic ac electric-field sensor based on the electrostrictive effect,” Optics letters 17(5), 372–374 (1992). 24 C. Ku, R. DePaula, J. Jarzynski, and J. Bucaro, “High frequency response of a single mode fiber optical phase modulator utilizing a piezoelectric plastic jacket,” in 1983 Technical Sym- posium East, 178–185, International Society for Optics and Photonics (1983). 25 L. Donalds, W. French, W. Mitchell, R. Swinehart, and T. Wei, “Electric field sensitive optical fibre using piezoelectric polymer coating,” Electronics Letters 18(8), 327–328 (1982). 26 U. Czarnetzki and K. Sasaki, “Ultra-sensitive measurement of sheath electric fields by laser-induced fluorescence-dip spectroscopy,” Journal of Plasma and Fusion Research 83(3) (2007). 27 K. Takizawa, K. Sasaki, and A. Kono, “Sensitive measurements of electric field distributions in low-pressure ar plasmas by laser-induced fluorescence-dip spectroscopy,” Applied physics letters 84(2), 185–187 (2004). 22 28 K. Takizawa, K. Sasaki, and K. Kadota, “Observation of stark spectra of argon high ryd- berg states in well-defined electric fields by laser-induced fluorescence-dip spectroscopy,” Japanese journal of applied physics 41(11B), L1285 (2002). 29 U. Czarnetzki, D. Luggenholscher, and H. Dobele, “Sensitive electric field measurement by fluorescence-dip spectroscopy of rydberg states of atomic hydrogen,” Physical review letters 81(21), 4592 (1998). 30 R. A. Soref and B. R. Bennett, “Electrooptical effects in silicon,” Quantum Electronics, IEEE Journal of 23(1), 123–129 (1987). 31 R. A. Soref and B. R. Bennett, “Kramers-kronig analysis of electro-optical switching in sili- con,” in Cambridge Symposium-Fiber/LASE’86, 32–37, International Society for Optics and Photonics (1987). 32 L. R. Dalton, P. A. Sullivan, and D. H. Bale, “Electric field poled organic electro-optic mate- rials: state of the art and future prospects,” Chemical reviews 110(1), 25–55 (2009). 33 T.-Z. Shen, S.-H. Hong, and J.-K. Song, “Electro-optical switching of graphene oxide liquid crystals with an extremely large kerr coefficient,” Nature materials (2014). 34 M. Adams, S. Ritchie, and M. Robertson, “Optimum overlap of electric and optical fields in semiconductor waveguide devices,” Applied physics letters 48(13), 820–822 (1986). 35 P. H. Paul and G. Kychakoff, “Fiber-optic evanescent field absorption sensor,” Applied physics letters 51(1), 12–14 (1987). 36 B. R. Bennett, R. A. Soref, and J. A. del Alamo, “Carrier-induced change in refractive index of inp, gaas and ingaasp,” Quantum Electronics, IEEE Journal of 26(1), 113–122 (1990). 23 37 W. Franks, I. Schenker, P. Schmutz, and A. Hierlemann, “Impedance characterization and modeling of electrodes for biomedical applications,” Biomedical Engineering, IEEE Trans- actions on 52(7), 1295–1302 (2005). 38 S. K. Arfin, A miniature, implantable wireless neural stimulation system. PhD thesis, Mas- sachusetts Institute of Technology (2006). 39 S. F. Cogan, “Neural stimulation and recording electrodes,” Annu. Rev. Biomed. Eng. 10, 275–309 (2008). 40 J. B. Ranck Jr, “Specific impedance of rabbit cerebral cortex,” Experimental neurology 7(2), 144–152 (1963). 41 P. W. Nicholson, “Specific impedance of cerebral white matter,” Experimental neurology 13(4), 386–401 (1965). 42 H.-L. Chen, Y.-M. Lu, and W.-S. Hwang, “Characterization of sputtered nio thin films,” Sur- face and Coatings Technology 198(1), 138–142 (2005). 43 S. Liu, R. Liu, Y. Chen, S. Ho, J. H. Kim, and F. So, “Nickel oxide hole injection/transport layers for efficient solution-processed organic light-emitting diodes,” Chemistry of Materials 26(15), 4528–4534 (2014). 44 E. J. Klem, H. Shukla, S. Hinds, D. D. MacNeil, L. Levina, and E. H. Sargent, “Impact of dithiol treatment and air annealing on the conductivity, mobility, and hole density in pbs colloidal quantum dot solids,” Applied Physics Letters 92(21), 212105 (2008). 45 M. Hochberg, T. Baehr-Jones, G. Wang, J. Huang, P. Sullivan, L. Dalton, and A. Scherer, “Towards a millivolt optical modulator with nano-slot waveguides,” Optics Express 15(13), 8401–8410 (2007). 24 46 J. Scholvin, J. Kinney, J. Bernstein, C. Moore-Kochlacs, N. Kopell, C. Fonstad, and E. Boy- den, “Close-packed silicon microelectrodes for scalable spatially oversampled neural record- ing,” IEEE Transactions on Biomedical Engineering (2015). 47 Y.-H. Kim, H.-J. Kim, M. Osada, B.-W. Li, Y. Ebina, and T. Sasaki, “2d perovskite nanosheets with thermally-stable high-κ response: A new platform for high-temperature capacitors,” ACS applied materials & interfaces 6(22), 19510–19514 (2014). 48 C. Wang, M. Osada, Y. Ebina, B.-W. Li, K. Akatsuka, K. Fukuda, W. Sugimoto, R. Ma, and T. Sasaki, “All-nanosheet ultrathin capacitors assembled layer-by-layer via solution-based processes,” ACS nano 8(3), 2658–2666 (2014). 49 T. Q. Ngo, A. B. Posadas, M. D. McDaniel, C. Hu, J. Bruley, T. Y. Edward, A. A. Demkov, and J. G. Ekerdt, “Epitaxial c-axis oriented batio3 thin films on srtio3-buffered si (001) by atomic layer deposition,” Applied Physics Letters 104(8), 082910 (2014). 50 T. B. Adams, D. C. Sinclair, and A. R. West, “Giant barrier layer capacitance effects in cacu3ti4o12 ceramics,” Advanced Materials 14(18), 1321–1323 (2002). 25
1103.2373
1
1103
2011-03-11T21:05:18
Drive for Creativity
[ "q-bio.NC" ]
We advance a hypothesis that creativity has evolved with evolution of internal representations, possibly from amniotes to primates, and further in human cultural evolution. Representations separated sensing from acting and gave "internal room" for creativity. To see (or perform any sensing), creatures with internal representations had to modify these representations to fit sensor signals. Therefore the knowledge instinct, KI, the drive to fit representations to the world, had to evolve along with internal representations. Until primates, it remained simple, without language internal representations could not evolve from perceptions to abstract representations, and abstract thoughts were not possible. We consider creative vs. non-creative decision making, and compare KI with Kahneman-Tversky's heuristic thinking. We identify higher, conscious levels of KI with the drive for creativity (DC) and discuss the roles of language and music, brain mechanisms involved, and experimental directions for testing the advanced hypotheses.
q-bio.NC
q-bio
1 Drive for Creativity Leonid Perlovskya and Dan Levineb aHarvard University and AFRL [email protected], bDepartment of Psychology, University of Texas at Arlington, [email protected] Abstract—We advance a hypothesis that creativity has evolved with evolution of internal representations, possibly from amniotes to primates, and further in human cultural evolution. Representations separated sensing from acting and gave “internal room” for creativity. To see (or perform any sensing), creatures with internal representations had to modify these representations to fit sensor signals. Therefore the knowledge instinct, KI, the drive to fit representations to the world, had to evolve along with internal representations. Until primates, it remained simple, without language internal representations could not evolve from perceptions to abstract representations, and abstract thoughts were not possible. We consider creative vs. non-creative decision making, and compare KI with Kahneman-Tversky’s heuristic thinking. We identify higher, conscious levels of KI with the drive for creativity (DC) and discuss the roles of language and music, brain mechanisms involved, and experimental directions for testing the advanced hypotheses. I. INTRODUCTION PRIMITIVE creatures, such as lobsters, do not use extensive internal representations in the mind. Their neurons are “hardwired” for relatively few crudely defined situations and sensor signals almost directly wired to muscles. Although the intermediate neural circuits in invertebrates between sensory organs and muscles are plastic, the learning and adaptation that takes place amounts to forming simple stimulus-response connections that do not generalize to other situations. There is no creativity. With evolution of adaptive representations, the variety of objects and situations in the world that the mind can recognize grows tremendously. Adaptive representations imply that yesterday‟s representations may not fit today‟s reality, and learning mechanisms are necessary for adaptation. Moreover, adaptation becomes necessary, because without adaptation perception becomes impossible. Therefore the knowledge instinct (KI), that is, the drive to adapt mental representations to the surrounding reality, had to evolve along with adaptivity. We discuss this mechanism, the mathematical reasons and experimental evidence that learning proceeds from vague to crisp representations [1-5]. Without language, adaptivity is limited to what can be perceived by sensors and to a limited number of combinations of perceptions. Visual and other sensory cortex circuitries in higher animals are tremendously complex mechanisms. Still we suggest that animals cannot learn to generalize much beyond direct perceptions; consequently complex animal behavior has to be based on inborn mechanisms with minimal adaptations. We discuss the reason for this hypothesis and how it can be tested. With language, cultural evolution outpaced genetic evolution, knowledge accumulated fast, and a hierarchy of abstract representations evolved. Abstract thoughts in language are grounded in external society and kids by age 5 can talk about almost everything ([6], [7]). Learning language is guided by a variant of KI (the Language Instinct, LI; [8]), and proceeds from vague to crisp representations (including syntax and complex abstract thoughts). Whereas KI fits internal cognitive representations to sensor perceptions, LI fits internal representations of language (vague at birth) to external language. But LI does not connect language to the world. We discuss a mechanism connecting language to cognitive representations, so that matching cognitive representations to the world by KI is guided by language 2 ([6], [7], [9]). This hypothesis is fundamental to understanding creativity, by means of understanding the difference between creative and heuristic thinking. We discuss its theoretical bases, experimental evidence, and future tests [10]. We emphasize that creativity is a matter of degree. Autonomous and unconscious working of KI usually is not called creativity. Conscious mental effort resulting in new understanding-representations in the mind of a single person is a creative process, even if this understanding is not new for a society; this is the distinction Boden [11] and others draw between personal creativity and historical creativity. Every normal kid understands surrounding language by the age of 5; these mechanisms are autonomous and non-creative. Yet even at 5 sometimes kids use language creatively: could this be studied scientifically? Reading and writing are not autonomously acquired skills. Skills acquired at an average level, with much tutoring and rote learning, usually are not called creativity. Should a better memory, quicker reactions, or faster speaking be called more creative? We emphasize the creative aspect of connecting language with cognition. Since the time of Aristotle [12], the ability to speak well has been identified with intelligence. Yet some people can be cognitively creative while possessing only average ability for expressing their thoughts in language. Of course the two sides of SAT tests acknowledge this. We discuss scientific approaches to studying this phenomenon. Are the two parts of SAT adequate for testing creativity? What can be deduced from contemporary models of the mind? The mind architecture has hierarchical structure, from elementary percepts to abstract thoughts. We suggest that creativity is manifest at higher levels. As mentioned, language is acquired early in life crisply and consciously at many hierarchical levels including the abstract ones, but we discuss why cognition at higher levels may remain fuzzy-vague. We discuss that crispness of language masks vagueness of cognition from our consciousness. Subjectively, ideas may be perceived as consciously understood, if we can talk about them. The paper discusses a mechanism by which crispness of language hides from the consciousness the vagueness of corresponding cognitive representations. It is obvious when watching kids: they talk well, but they do not “really” know what it means. In regards to highly abstract ideas, most of the time most adults are like kids. Therefore it takes more creativity to make crisp and conscious cognitive model-representations at higher levels. This hypothesis can be tested experimentally. Creativity in a wide sense develops crisp cognitive representations corresponding to crispness of language, in other words, to understand the world and self at the level existing in language and culture. In a more narrow and refined sense creativity has several aspects. Scientific creativity develops cognitive representations, which surpass language in their adequacy to the world (and self). Writers formulate knowledge in language crisper than it has been formulated previously; new knowledge is created in this process if stronger connections to cognitive representations are achieved. As we discuss that requires connecting language with emotions. In this area, writers, poets, and composers use various means for similar goals: emotionally connecting language and cognition. Creativity, thus, involves not only conceptual knowledge but emotions as well. We discuss the role of emotions in cognition and identify specific emotions related to the KI and creat ivity ([2-4], [6], [13-17]). Higher up in the hierarchy of the mind, emotions and concepts are intermixed in a non- differentiated fashion. Development of differentiated knowledge, especially at higher levels of the hierarchy, creates psychological tensions, that is, cognitive dissonances. These dissonances in the psyche interfere with unity, which is necessary for achieving higher goals, psychic health and sometimes even survival. To continue developing knowledge and to be able to cope with this multitude of dissonances, humans need the corresponding multitude of conscious “higher” emotions that help one to bring the “feelings” of the dissonances into consciousness, and then to cope with them. Poetry and music evolved for this purpose [18-20]. Closely related is a need to continue connecting language to cognition, which requires this same multitude of emotions ([7], [15]). This is the cognitive role of 3 poetry, songs, and musical theater. Correspondingly, there are emotional creativities of poets, composers, musicians, singers and also writers – they create new emotions necessary for being able to cope with the diversity of cultures – to connect language with cognition, and to continue the process of cultural evolution [15]. Cognitive representations include understanding and behavior. We discuss that the highest conceptual understanding relates to the beautiful. The highest of behavioral representations relate to the spiritually sublime [10]. The next section discusses mechanisms of KI, a contradiction between knowledge-maximizing mechanisms of creativity driving cultural evolution of human highest spiritual abilities on one hand and, on the other, heuristic mechanisms minimizing cognitive-effort as revealed by Tversky and Kahneman [21, 22]. Then we discuss mechanisms of interaction between cognition and language and relate them to the contradiction between KI and heuristics. The discussion section addresses experimental approaches to demonstrating the difference between heuristic and KI mechanisms, the role of emotions in creativity, and experimental approaches to demonstrating drive for creativity [17]. How far toward understanding creativity can we move based on contemporary models of the mind? II. HEURISTICS VS. THE KNOWLEDGE INSTINCT Research initiated by Tversky and Kahneman [21, 22] demonstrated that decision making is often irrational and based on heuristics, which may or may not be suitable to a particular situation. Humans often prefer fast decisions, saving effort, including mental effort, to original thinking. According to Maimonides [23], refusing original thinking and choosing a shortcut to the ready-made knowledge was the reason for expelling Adam from paradise. Maimonides‟ explanation was connected to contemporary cognitive science in [10]. Fast decisions saving time and effort have clear evolutionary advantages. Efficient use of heuristics could be counted as “personal creativity.” But if irrationality and heuristics are fundamental to human decision making, what is the source of all human knowledge? Heuristics might be used well or poorly, but what is the source of heuristics? This source should be no less fundamental to human psyche than irrationality. We suggest that heuristics is a store of accumulated cultural knowledge, whereas the source of this knowledge is the KI. Heuristics originate in those relatively rare times when creative individuals use their KI and create new knowledge that turn useful for the rest of the society (creativity or “historical creativity”). Perlovsky and McManus [24], (see also [1] and [2]), developed a mathematical theory of the knowledge instinct (KI), the drive to make sense out of one‟s environment. Their mathematical description of the KI involves maximizing similarity between mental representations-models and patterns in sensory signals. More generally, KI maximizes similarity between bottom-up and top-down signals, which is interpreted as knowledge. Top-down signals produced by mental representations, indexed by m at each hierarchical level. In neural terminology they prime receiving neurons, n, to recognize pattern Mm(n) in the bottom-up signals X(n). Mathematically, they predict the expected patterns in the bottom-up signals; the similarity between bottom-up signals X(n) and top-down signals Mm(n) ([1], [2]) is expressed as L({X},{M}) = r(m) l(X(n) Mm(n)). (1) This similarity expression considers each top-level representation, m, as an alternative, a possibility of an object or situation m to be present among bottom-up signals n. Conditional similarities MmNn 4 l(X(n) Mm(n)) can be modeled [25, 26] as l(X(n) Mm(n)) = pmi xni(1 – pmi)(1-xni) .(2) Here, n indexes bottom-up signals, m indexes top-down signals, and i indexes their constituent components; e.g. features of a pattern, or objects making up a situation, or situations making up an abstract representations. A bottom-up signal xni indicates a presence of a component i in a bottom-up signal n, and a top-down signal pmi models a probability of a component i in a top-down signal m. Thus expressions (1) and (2) model similarity measures at every level of the mind hierarchy. The number of possible components, i, is denoted as D. All possible associations between bottom-up signals, N, and top-down signals, M, are accounted for in similarity (1). This is a huge combinatorial number on the order of NM. As discussed in [1] and [2], this is the reason for the computational complexity encountered by most artificial intelligence, pattern recognition, and neural network algorithms since the 1950s. Such complexity is related to classical logic [27, 28], which is a part not only of logical rule systems, but also of those approaches specifically designed to overcome logical limitations. Classical logic is used in training procedures of pattern recognition and neural network algorithms, and in fuzzy systems‟ selection of the degrees of fuzziness. This is the reason for the seemingly irresolvable computational complexity of most algorithms encountered for decades. The above references describe Neural Modeling Fields (NMF), a mathematical procedure that uses dynamic logic (DL) to overcome limitations of computational complexity. Resulting algorithms outperform previous approaches by orders of magnitude ([29-33]). Whereas classical logic involves crisp statements and thus is a static statement-logic, DL is a process-logic; it involves evolution of mental representations from vague-fuzzy to crisp. It describes emergence of approximately classical logic processes from illogical and fuzzy neural processes in the brain ([7], [35]). Neuroimaging experiments demonstrated that visual perceptions in the human brain follow predictions of NMF-DL ([1], [2], [5], [7], [16]). Thus at the basic level of visual perception, DL models a mechanism of KI. At lower levels of the brain hierarchy KI operates autonomously, serving as a foundation of perception. However, our hypothesis in this paper is that at higher levels KI operations are not autonomous, they require conscious effort, and they serve as foundations for creativity. KI at higher levels becomes a drive for creativity. This is because existing concepts or combinations of concepts are often inadequate for resolving cognitive and emotional dissonances between top-down and bottom-up signals between different high levels. The roles of several prefrontal and subcortical regions in forming, breaking, and changing behavioral and decision rules are discussed in [34]. At lower levels of perception there is no competition among mechanisms of decision making . Perception cannot operate heuristically; higher animals as well as human beings have to use DL mechanisms of KI to be able to orient in the surrounding world. In the next section we discuss decision making at higher cognitive levels and relate a competition between KI and Kahneman-Tversky heuristic mechanisms to a competition between cognitive and language mechanisms. III. KI AND HEURISTICS VS. COGNITION AND LANGUAGE Let us repeat that at lower levels of perception KI operates autonomously. This is similar in animals and in humans. Here we relate ability for higher level abstract thinking to language, and in turn, relate language to a competition between KI and heuristic thinking. As argued in [15] and [36], higher level abstract thinking is only possible due to language. To Di1 5 model this ability we have to consider mechanisms of interaction between cognition and language, and their parallel hierarchy, Fig. 1 ([2], [6], [14], [31]). NMF-DL describes language similar to cognition by the hierarchy of levels maximizing similarity (1). A newborn human mind does not have specific models for concrete words or syntactic rules. Instead, it contains vague “placeholders” that in the process of learning would acquire specific contents (words, rules) matching the language spoken around. Similarly, a newborn human mind does not have specific models for concrete objects or situations; it contains vague “placeholders” that in the process of learning would acquire specific contents (images, relations) matching surrounding world. Interaction between cognition and language is modeled mathematically by the mechanism of dual models. While language and cognitive models in the newborn brain are vague placeholders, the neural connections between these future models are inborn ([6], [15]). Therefore human brain-mind does not have to “figure out” which words go with which objects (this mathematically unsolvable problem can be solved using the dual model, [37-41]). As vague placeholders acquire concrete contents, connections between words and objects, phrases and situations are already in place. This is the mechanism of the dual model, which enables language-cognition interaction. Note that as Fig. 1 indicates, language is acquired from surrounding language ready-made at all hierarchical levels. This is the reason kids of 5 years old can talk about complex abstract ideas, which they do not yet have experience to “really understand.” But note also that cognitive experience is acquired from the surrounding world only at the very bottom of the hierarchy, at the level of sensory perceptions. Learning at this level (perception) is possible without language, and animals can perceive objects, and to some degree situations. But learning abstract ideas, which are combinations of elements of many situations, combinations that cannot be directly perceived, is impossible from direct experience. The reason is that the number of combinations is too large, much larger than any individual experience will enable to learn useful abstract ideas from random sets of objects and relations. Only due to guidance from language, abstract cognitive models could acquire concrete contents. KI drives every human brain - mind to develop abstract cognitive contents corresponding to language models. Higher up in the hierarchy cognitive models remain vaguer and less concrete. Learning- understanding and creating new models continue throughout life. Higher in the hierarchy the autonomous KI gradually turns into drive for creativity. It becomes less autonomous and more dependent on individual conscious effort. This is why there are significant differences between people in how well they “really understand” the contents of the spoken language. Our hypothesis is that one “really understands” when his or her cognitive models become crisp and specific corresponding to crispness of language. This hypothesis can be experimentally tested. We would emphasize that “real understanding” requires more than just crispness and concreteness. An abstract idea involves more than one crisp and concrete content; the drive for creativity strives for multiple differentiated contents, which have to be properly related and unified. It follows from the above discussion that the drive for creativity has an inborn autonomous component; however, this inborn component weakens at the higher levels. Language, while providing guidance to the development of abstract cognitive models, also creates impediments to this process. To understand the nature of this difficulty, let us consider cognition at the lower level of perception, where it does not involve language. Close your eyes for 2 seconds and try to imagine an object that you have clearly seen in front of your eyes. The imagined object is vague, not as crisp and clear and not as consciously perceived as the object with opened eyes. Bar et al. [5] performed a similar and much more detailed experiment using brain imaging. The imagined vague object is created on the visual cortex by projections of mental representations, top-down signals. When you open your eyes, the object again is perceived crisp and clear, and it is impossible to perceive the vague imagined object with opened eyes. With opened eyes, crisp and consciously perceived projections from the retina, which are created by interacting bottom-up and top-down signals, mask less conscious and vaguer top-down signals. Similar 6 processes occur at higher level. Language models act as “eyes” for cognition of abstract concepts. Because language models are crisp and conscious, they mask vague and unconscious abstract cogni tive models. Therefore it is difficult to perceive that cognition is vague and inadequate. Every step towards more crisp and conscious cognitive abstract models requires conscious creative effort. Still, evolution of human knowledge and culture confirms that drive for creativity makes this development possible. THOUGHTS THOUGHTS THOUGHTS LANGUAGE SURROUNDING LANGUAGE … abstract ideas abstract ideas abstract ideas … abstract words/phrases … situations situations situations phrases objects objects objects words sensor signals sensor signals sensor signals from the world from the world from the world language sensory-motor mechanisms language descriptions of abstract thoughts phrases for situations words for objects language sounds Fig. 1. Hierarchical integrated language-cognition MF system. At each level in a hierarchy there are similarities, models, and mechanisms of learning of language and cognitive models ([15]). The dual models (thick arrow) integrate language and cognition. Mathematically, similarities are integrated as products of language and cognition similarities. Initial models are fuzzy placeholders, so integration of language and cognition is sub-conscious. Association variables depend on both language and cognitive models and signals. Therefore language model learning helps cognitive model learning and v.v. High-layer abstract cognitive concepts are grounded in abstract language concepts similarities. Now we can better understand the interplay between drive for creativity and irrationality of 7 Kahneman-Tversky heuristics. Drive for creativity makes possible the improvement of abstract cognitive models. Yet, it takes much effort to overcome heuristic knowledge accumulated in language. Irrationality and heuristic thinking, according to the hypothesis of this paper, correspond to thinking with crisp language models, while the corresponding cognitive models remain vaguer and less conscious. IV. DISCUSSION Steps toward mathematical modeling of interacting heuristic thinking and cognitive drive have been taken in [10] and [42]. Based on various fMRI results, we suggested in [10] that operations of heuristic mechanisms tend to activate the amygdala, a part of the brain that has changed little from other mammals to humans. On the other hand, creative operations of the knowledge instinct tend to activate executive regions of prefrontal cortex, a part of the brain whose complexity has grown exponentially in humans. Yet the roles of those regions are complex because connections between the amygdala and two regions of prefrontal cortex (orbitofrontal and anterior cingulate) are also implicated in higher -level emotional responses based on actual or expected consequences of stimuli or actions (see [42] for review). Hence those same connections likely play a role in the aesthetic emotional responses that accompany the intuitive elements of creativity. Experimental tests of this hypothesis are ongoing. Here we continue developing this hypothesis by suggesting specific roles of cognition and language. We suggest that thinking by using ready-made heuristics involves language brain centers to larger degrees than original thinking using cognitive mechanisms. The connection between heuristics and language could be mediated by neural connections between cortical language centers and amygdala (either direct or through the thalamus). Whereas heuristics involve vague and less conscious cognitive representations, original thinking involves crisper and more conscious cognitive representations. This hypothesis can be experimentally tested using procedures similar to those used in referenced literature. According to the theory of drives and emotions [13], every drive involves specific emotional neural signals indicating satisfaction or dissatisfaction of this drive. Specific emotions associated with KI were identified as aesthetic emotions ([4], [6], [7], [16], [18-20], [43], [44]). Experimental confirmation of aesthetic emotions related to satisfaction of KI was documented in [17]. This hypothesis has been further developed with our theory of differentiated KI, which in this paper we identify with the drive for creativity. It has been suggested that all inconsistencies in the body of knowledge, as well as all contradictions between knowledge and other instincts, create cognitive dissonances which are perceived emotionally as a multiplicity of aesthetic emotions. Maintaining unity of the psyche along with differentiated knowledge requires highly developed differentiated emotions, which are created in individuals and cultures by music ([6], [18-20]). The related form of creativity is a province of poets, composers, and singers. Affinity between emotions of cognitive dissonance and musical emotions is being tested by several research groups of psychologists and cognitive musicologists. This is a wide field for future research. Also, an important part of the drive for creativity is the sense that we have achieved an understanding through our own efforts. Why should satisfaction of the KI through generating an idea on one‟s own provide us with a particular pleasure that we do not obtain as much from leaning about the same idea from someone else, or from a book or web site? The answer to this question probably involves a greater understanding of the nature of the self and how we differentiate the self from others. Aside from a general sense that the self is closely involved with prefrontal cortex, and that different and adjacent prefrontal regions are implicated in emotional and cognitive parts of the sense of self [43], the neural basis for the self-other distinction is largely unknown and is another field for future research. 8 The theory developed in this paper and others advances hypotheses about the nature of the highest aesthetic emotions of the beautiful and sublime ([2-4], [6], [7], [16-20], [44-47]. In the mind hierarchy, as shown in Fig. 1, concepts-representations at every higher level have evolved in cultural evolution for the purpose of creating a unified, general meaning of lower level concepts. Mental representations at the highest levels of the hierarchy strive for creating a unity of the entire human experience. While language representations at these highest levels have been developed by artists, philosophers, and theologians over millennia, cognitive representations remain vague and unconscious. Every step toward creating crisper contents of these representations improves understanding of the meaning of the entire life experience in its unity and is felt as emotions of the beautiful. Improving representations of behavior realizing the beautiful in one‟s life is felt as emotions of the sublime. Experimental tests of this hypothesis are a subject of future research. REFERENCES [1] L. I. Perlovsky, Neural Networks and Intellect: Using Model Based Concepts, Oxford University Press, New York, 2001. [2] L. I. Perlovsky, “Toward physics of the mind: Concepts, emotions, consciousness, and symbols,” Physics of Life Reviews, Vol. 3, pp. 22-55, 2006. [3] L. I. Perlovsky, “„Vague-to-crisp‟ neural mechanism of perception,” IEEE Transactions on Neural Networks, Vol. 20, pp. 1363-1367, 2009. [4] L. I. Perlovsky, “Neural mechanisms of the mind, Aristotle, Zadeh, & fMRI,” IEEE Transactions on Neural Networks, in press, 2010. [5] M. K. Bar, S. Kassam, A. S. Ghuman, J. Boshyan, A. M. Schmid, A. M. Dale, M. S. Hamalainen, K. Marinkovic, D. L. Schacter, B. R. Rosen, and E. Halgren, “Top-down facilitation of visual recognition,” Proceedings of the National Academy of Sciences USA, Vol 103, pp. 449-454, 2006. [6] L. I. Perlovsky. “Integrating language and cognition,” IEEE Connections, Vol. 2, pp. 8-12, 2004. [7] L. I. Perlovsky, “Language and cognition,” Neural Networks, Vol. 22, pp. 247-257, 2009. [8] S. Pinker, The Language Instinct: How the Mind Creates Language, Perennial Classics, New York, 2000. [9] L. I. Perlovsky, “Evolution of languages, consciousness, and cultures,” IEEE Computational Intelligence Magazine, Vol. 2, pp.25-39, 2007. [10] D. S. Levine and L. I. Perlovsky, “Neuroscientific insights on Biblical myths: Simplifying heuristics versus careful thinking: Scientific analysis of millennial spiritual issues,” Zygon, Journal of Science and Religion, Vol. 43, pp. 797-821, 2008. [11] M. Boden, “State of the art: Computer models of creativity,” The Psychologist, Vol. 13, 72-76, 2000. [12] Aristotle, Rhetoric to Alexander, In The Complete Works of Aristotle, Ed. J. Barnes, Princeton University Press, Princeton, NJ, IV BCE/1995. [13] S. Grossberg and D. S. Levine. “Neural dynamics of attentionally modulated Pavlovian conditioning: Blocking, inter - stimulus interval, and secondary reinforcement,” Applied Optics,Vol. 26, pp. 5015-5030, 1987. [14] L. I. Perlovsky, “Symbols: Integrated cognition and language,” in R. Gudwin and J. Queiroz (Eds.),Semiotics and Intelligent Systems Development,” Idea Group, Hershey, PA, pp.121-151, 2007. [15] L. I. Perlovsky, “Language and emotions: Emotional Sapir-Whorf hypothesis,” Neural Networks, Vol. 22, pp. 518-526, 2009. [16] L. I. Perlovsky, “Intersections of Mathematical, Cognitive, and Aesthetic Theories of Mind,” Psychology of Aesthetics, Creativity, and the Arts, in press, 2010. [17] L. I. Perlovsky, M.-C. Bonniot-Cabanac, and M. Cabanac. “Curiosity and pleasure,” IJCNN 2010, PNAS, submitted. 2010.R. Deming, S. Higbee, D. Dwyer, M. Welser, L. Perlovsky, and P. Pellegrini. “Robust detection and spectrum estimation of multiple sources from rotating-prism spectrometer images,” Proceedings of SPIE Remote Sensing Europe, Stockholm, Sweden, 2006. [18] L. I. Perlovsky, “Music–the first principles,” Musical Theater, 2006. http://www.ceo.spb.ru/libretto/kon_lan/ogl.shtml. [19] L. I. Perlovsky, “Music and consciousness,” Leonardo, Journal of Arts, Sciences and Technology, Vol. 41, pp. 420-421, 2008 [20] L. I. Perlovsky, “Musical emotions: Functions, origin, evolution,” Physics of Life Reviews, in press, 2010. [21] A. Tversky, and D. Kahneman, “Judgment under uncertainty: Heuristics and biases,” Science, Vol. 185, pp. 1124-1131, 1974. [22] A. Tversky, and D. Kahneman, “The framing of decisions and the rationality of choice,” Science, Vol. 211, pp. 453- 458, 1981. 9 [23] M. Maimonides, The Guide for the Perplexed, 1190, M. Friedlander, trans. 1956, Unabridged Dover Edition, New York. [24] L. I. Perlovsky and M. M. McManus. “Maximum likelihood neural networks for sensor fusion and adaptive classification,” Neural Networks, Vol. 4, pp. 89-102, 1991. [25] R. Ilin and L. I. Perlovsky, “Cognitively Inspired Neural Network for Recognition of Situations,” International Journal of Natural Computing Research, in press, 2010. [26] L. I. Perlovsky, and R. Ilin, “Neurally and Mathematically Motivated Architecture for Language and Thought.” Special Issue Brain and Language Architectures: Where We are Now? The Open Neuroimaging Journal, in press, 2010. [27] L. I. Perlovsky, “Gödel theorem and semiotics,” Proceedings of the Conference on Intelligent Systems and Semiotics '96. Gaithersburg, MD: v. 2, pp. 14-18, 1996. [28] L. I. Perlovsky, “Conundrum of combinatorial complexity,” IEEE Transactions on PAMI, Vol. 20, 666-670, 1998. [29] L. I. Perlovsky and R.W. Deming. “Neural networks for improved tracking,” IEEE Transactions on Neural Networks, Vol. 18, pp. 1854-1857, 2007. [30] R. Deming, S. Higbee, D. Dwyer, M. Welser, L. Perlovsky, and P. Pellegrini. “Robust detection and spectrum estimation of multiple sources from rotating-prism spectrometer images,” Proceedings of SPIE Remote Sensing Europe, Stockholm, Sweden, 2006. [31] R. W. Deming and L. I. Perlovsky, “Concurrent multi-target localization, data association, and navigation for a swarm of flying sensors, Information Fusion, Vol. 8, pp.316-330, 2007. [32] R. Linnehan, D. Brady, J. Schindler, L. Perlovsky, & M. Rangaswamy, “On the Design of SAR Apertures Using the Cram´er-Rao Bound,” IEEE Transactions on Aerospace and Electronic Systems, Vol. 43, pp. 344-355, 2007. [33] L. I. Perlovsky, web page, http://www.leonid-perlovsky.com, 2010. [34] D. S. Levine, “How does the brain create, change, and selectively override its rules of conduct?,” in L. I. Perlovsky and R. Kozma (Eds.), Neurodynamics of Higher-Level Cognition and Consciousness, Springer-Verlag, Heidelberg, pp. 163- 181, 2007. [35] L. I. Perlovsky and R. Kozma, Eds. Neurodynamics of Higher-Level Cognition and Consciousness, Springer-Verlag, Heidelberg, 2007. [36] L. I. Perlovsky, “Cognitive High Level Information Fusion,” Information Sciences, Vol. 177, pp. 2099-2118, 2007. [37] J. F. Fontanari and L.I. Perlovsky, “Evolving compositionality in evolutionary language games,” IEEE Transactions on Evolutionary Computations, Vol. 11, pp. 758-769, 2007 [38] J. F. Fontanari and L.I. Perlovsky, “How language can help discrimination in the Neural Modeling Fields framework,” Neural Networks, Vol. 21, 250–256, 2008. [39] J. F. Fontanari and L. I. Perlovsky, “A game theoretical approach to the evolution of structured communication codes,” Theory in Biosciences, Vol. 127, 205-214, 2008. [40] J. F. Fontanari, V. Tikhanoff, A. Cangelosi, R. Ilin, and L.I. Perlovsky, “Cross-situational learning of object–word mapping using Neural Modeling Fields,” Neural Networks,Vol. 22, pp. 579-585, 2009. [41] V. Tikhanoff, J. F. Fontanari, A. Cangelosi, & L. I. Perlovsky, “Language and cognition integration through modeling field theory: category formation for symbol grounding,” In Book Series in Computer Science, v. 4131, Springer, Heidelberg, 2006. [42] D. S. Levine, “Value maps, drives, and emotions,” to appear in V. Cutsuridis, J. G. Taylor, D. Polani, A. Hussain, and N. Tishby, eds., Perception-reason-action Cycle: Models, Algorithms, and Systems, Springer, Berlin, 2010. [43] J. M. Moran, C. N. Macrae, T. F. Heatherton, C. L. Wyland, and W. M. Kelley, “ Neuroanatomical evidence for distinct cognitive and affective components of self,” Journal of Cognitive Neuroscience, Vol. 18, pp. 1586-1594, 2006. [44] L. I. Perlovsky, “Aesthetics and mathematical theories of intellect,” Iskusstvoznanie, 2002, pp.558-594, Moscow, (Russian), 2002. [45] L. I. Perlovsky, “Modeling field theory of higher cognitive functions,” In A. Loula, R. Gudwin, J. Queiroz, Eds. Artificial cognition systems. Hershey, PA: Idea Group, pp. 64-105, 2007. [46] I. Kant, Critique of Judgment, J.H.Bernard, trans., Macmillan and Company, London, 1790/1914. [47] I. Kant, Anthropology from a Pragmatic Point of View. Tr. M. J. Gregor. Boston, MA: Kluwer Academic Publishers, 1798/1974.
1512.00810
3
1512
2016-08-10T17:00:26
Valid population inference for information-based imaging: From the second-level $t$-test to prevalence inference
[ "q-bio.NC", "stat.AP" ]
In multivariate pattern analysis of neuroimaging data, 'second-level' inference is often performed by entering classification accuracies into a $t$-test vs chance level across subjects. We argue that while the random-effects analysis implemented by the $t$-test does provide population inference if applied to activation differences, it fails to do so in the case of classification accuracy or other 'information-like' measures, because the true value of such measures can never be below chance level. This constraint changes the meaning of the population-level null hypothesis being tested, which becomes equivalent to the global null hypothesis that there is no effect in any subject in the population. Consequently, rejecting it only allows to infer that there are some subjects in which there is an information effect, but not that it generalizes, rendering it effectively equivalent to fixed-effects analysis. This statement is supported by theoretical arguments as well as simulations. We review possible alternative approaches to population inference for information-based imaging, converging on the idea that it should not target the mean, but the prevalence of the effect in the population. One method to do so, 'permutation-based information prevalence inference using the minimum statistic', is described in detail and applied to empirical data.
q-bio.NC
q-bio
Valid population inference for information-based imaging: From the second-level t-test to prevalence inference Carsten Allefelda* Kai Görgena John-Dylan Haynesa,b** a. Bernstein Center for Computational Neuroscience, Berlin Center of Advanced Neuroimaging, Department of Neurology, and Excellence Cluster NeuroCure, Charité – Universitätsmedizin Berlin, Germany b. Berlin School of Mind and Brain and Department of Psychology, Humboldt-Universität zu Berlin, Germany Address for all affiliations: Charité-Campus Mitte, Philippstr. 13, Haus 6, 10115 Berlin, Germany * E-mail: [email protected] ** E-mail: [email protected] Corresponding author: Carsten Allefeld, Tel. +49 30 2093 6766 preprint of C. Allefeld, K. Görgen, J.-D. Haynes. Valid population inference for information-based imaging: From the second-level t-test to prevalence inference. NeuroImage, 141: 378–392, 2016. doi:10.1016/j.neuroimage.2016.07.040 this version includes minor fixes and a note added after publication 2016-8-10 6 1 0 2 g u A 0 1 ] . C N o i b - q [ 3 v 0 1 8 0 0 . 2 1 5 1 : v i X r a 1 Abstract In multivariate pattern analysis of neuroimaging data, 'second-level' inference is often performed by entering classification accuracies into a t-test vs chance level across sub- jects. We argue that while the random-effects analysis implemented by the t-test does provide population inference if applied to activation differences, it fails to do so in the case of classification accuracy or other 'information-like' measures, because the true value of such measures can never be below chance level. This constraint changes the meaning of the population-level null hypothesis being tested, which becomes equiva- lent to the global null hypothesis that there is no effect in any subject in the population. Consequently, rejecting it only allows to infer that there are some subjects in which there is an information effect, but not that it generalizes, rendering it effectively equivalent to fixed-effects analysis. This statement is supported by theoretical arguments as well as simulations. We review possible alternative approaches to population inference for information-based imaging, converging on the idea that it should not target the mean, but the prevalence of the effect in the population. One method to do so, 'permutation- based information prevalence inference using the minimum statistic', is described in detail and applied to empirical data. Keywords information-based imaging, multivariate pattern analysis, t-test, population inference, effect prevalence 1 Introduction Since the seminal work of Haxby et al. (2001), an increasing number of neuroimaging studies have employed multivariate methods to complement the established mass- univariate approach (Friston et al., 1995) to the analysis of functional magnetic reso- nance imaging (fMRI) data, a field now known as multivariate pattern analysis (MVPA; Norman et al., 2006). Most MVPA studies use classification (Pereira et al., 2009) to exam- ine activation patterns; the accuracy of a classifier in distinguishing activation patterns associated with different experimental conditions serves as a measure of multivariate effect strength. Since the target of MVPA is not a generally increased or decreased level of activation but the information content of activation patterns (cf. Pereira and Botvinick, 2011), it has also been characterized as information-based imaging and distinguished from traditional activation-based imaging (Kriegeskorte et al., 2006). Many methodological aspects of MVPA have already been discussed in detail: what kind of classifier to use (Cox and Savoy, 2003; Norman et al., 2006), whether to adapt parametric multivariate statistics instead of classifiers (Allefeld and Haynes, 2014; Nili et al., 2014), how to understand searchlight-based accuracy maps (Etzel et al., 2013), or how classifier weights can be made interpretable (Haufe et al., 2014; Hoyos-Idrobo et al., 2015). By contrast, the topic of population inference based on per-subject measures of information content, i.e. the question whether an information effect observed in a sample of subjects generalizes to the population these subjects were recruited from, has not yet received sufficient attention (but see Brodersen et al., 2013). 2 In univariate analysis of multi-subject fMRI studies, the standard way to achieve population inference is to perform a 'second-level' null hypothesis test (Holmes and Friston, 1998). For each subject, a 'first-level' contrast (activation difference) is computed, and this contrast enters a second-level analysis, a t-test or an ANOVA. Specifically for a simple one-sided t-test vs 0, reaching statistical significance allows to infer that the experimental manipulation is associated with an increase of activation on average in the population of subjects. This is interpreted in such a way that the effect is 'common' or 'stereotypical' in that population (Penny and Holmes, 2007, p. 156). With the adoption of information-based imaging, it has become accepted practice to apply the same second-level inferential procedures to the results of first-level multivari- ate analyses, in particular classification accuracy (see e.g. Haxby et al., 2001; Spiridon and Kanwisher, 2002; Haynes et al., 2007): A classifier is trained on part of the data and is tested on another part, using each part for testing once (cross-validation), and the classification performance is quantified in the form of an accuracy, the fraction of correctly classified test data points. Applied for example to two different experimental conditions, if there was no multivariate difference in the data between conditions, the classifier would operate at 'chance level', i.e. it would on average achieve a classifica- tion accuracy of 50 %. At the second level, accuracies from different subjects are then entered into a one-sided one-sample t-test vs 50 %, in order to show that the ability to classify above chance and therefore the presence of an information effect is typical in the population the subjects were recruited from. In this paper we argue that despite of the seemingly analogous statistical procedure, a t-test vs chance level applied to accuracies cannot provide evidence that the corre- sponding effect is typical in the population. In contrast to other criticisms of this use of the t-test (see below), in our view the problem is not so much that the estimation distribution of cross-validated accuracies is not normal or even symmetric, or that a normal distribution model is generally inadequate for a quantity bounded to an interval [0 %, 100 %]. Rather, the problem is that other than estimated accuracies, the true single-subject accuracy can never be below chance level because it measures an amount of information.1 We will show that this restriction changes the meaning of the t-test: It now tests the global null hypothesis (Nichols et al., 2005) that there is no information in any subject in the population. As a consequence, achieving a significant test result allows us only to infer that there are people in which there is an effect, but not that the presence of information generalizes to the population. The argument does not only hold for classification accuracy, but also for other 'information-like' measures. The t-test on accuracies has been criticized before (Stelzer et al., 2013; Brodersen et al., 2013) on the grounds that its distributional assumptions are not fulfilled for cross- validated classification accuracies. Such a distributional error invalidates the calculation of critical values for the t-statistic and can therefore lead to an increased rate of false positives. This problem may be solved by better distribution models (Brodersen et al., 2013) or the use of non-parametric statistics (Stelzer et al., 2013). Our criticism goes 1Note that in this paper we only discuss the standard case of MVPA where the pair of experimental conditions is the same for training and test data. In 'cross-decoding' (cf. Haynes and Rees, 2005a), where it is tested whether a classifier trained on one pair of conditions is able to extract information corresponding to another pair of conditions, below-chance true accuracies may be possible. Cross-decoding targets not just the presence of information, but also the degree to which its neurophysiological representation is invariant with respect to another experimental manipulation. 3 significantly beyond that: Not only is the t-test quantitatively wrong, but it effectively tests a null hypothesis that is qualitatively different from its use with univariate statis- tics, with the consequence that rejection of this null hypothesis no longer supports population inference. Please note that our criticism pertains specifically to a second-level t-test applied to per-subject classification accuracies or similar measures. It does not apply to the classifi- cation of subjects, e.g. into different patient groups in medical applications (Sabuncu and Van Leemput, 2012; Sabuncu, 2014), or to the classification of condition-specific pat- terns across subjects (Mourao-Miranda et al., 2005). Moreover, it only concerns quantities that measure the information content of data, but not related quantities like classifier weights (Wang et al., 2007; Gaonkar and Davatzikos, 2013; Gaonkar et al., 2015, see below). The organization of the paper is as follows: In Part 2 we detail how a second-level t-test achieves population inference for univariate contrasts. We then explain that MVPA measures are 'information-like' and show, both theoretically and using simulations, that for such measures the t-test effectively tests the global null hypothesis that there is no effect in any subject. Part 3 reviews possible alternatives to the t-test on accuracies, converging on the idea that population inference for information-based imaging should target the proportion of subjects in the population with an effect. One way to implement such an 'information prevalence inference' is described in detail in Part 4, and results of its application to real data are compared with those of the t-test. We conclude with the discussion of a number of questions surrounding the problem of population inference for information-based imaging. 2 The problem with the t-test on accuracies 2.1 Population inference in univariate fMRI analysis To see why the t-test on accuracies cannot provide population inference, we briefly recapitulate how standard univariate analysis does achieve it. In a single subject, an activation difference or contrast ∆β is estimated based on the general linear model (GLM; Friston et al., 1995). Because it is obtained from noisy data, the estimate is itself noisy, (1) where σ2 1 denotes the estimation variance of the contrast (cf. Fig. 1a). If several subjects are included in a study, the true activation difference ∆β varies across subjects (Fig. 1a): (cid:99)∆β ∼ N (∆β, σ2 1 ), ∆βk ∼ N (∆µ, σ2 2 ) (2) where ∆µ is the average true activation difference in the population of subjects and σ2 2 the population variance of the effect (Fig. 1b). The added subscript k indicates that we now consider the subject as randomly sampled from the population. The estimated contrast in several subjects therefore shows variation for two reasons - they are noisy estimates (σ2 1), and different subjects respond differently (σ2 2): (cid:99)∆βk ∼ N (∆µ, σ2 4 1 + σ2 2 ). (3) The symbol (cid:99)∆βk indicates that this contrast is both estimated and sampled. A one-sided t-test applied to the (cid:99)∆βk from a sample of subjects k = 1 . . . N has the null hypothesis ∆µ = 0. If it can be rejected (∆µ > 0), this allows us to make a statement about the population of subjects because ∆µ is a parameter of a population model (Eq. 2). And this statement concerns a typical effect because ∆µ is the mean, median, and mode of the assumed normal distribution. This kind of test is also called random-effects analysis (RFX) because it treats subjects as as randomly sampled from a population (Searle et al., 1992). It was introduced into fMRI by Holmes and Friston (1998) to replace previous fixed-effects analyses (FFX), which did not account for population variation and therefore did not provide population inference.2 2.2 The t-test on accuracies Using a second-level t-test vs chance level with classification accuracies implies that an analogous random-effects model applies: In each subject (first level) we obtain an estimated accuracy which varies with an estimation variance ς2 1, a ∼ N (a, ς2 1). (4) The underlying true accuracy a (see App. A) varies across subjects (second level) with a population variance ς2 2, ak ∼ N (¯a, ς2 2). (5) 1 + ς2 2, Therefore estimated accuracies vary across subjects with the combined variance ς2 ak ∼ N (¯a, ς2 1 + ς2 2). (6) Here ¯a is the average true classification accuracy in the population, and the null hy- pothesis is that this population average is at chance level, ¯a = a0. Again, the symbol ak indicates that this accuracy is both estimated and sampled. Though a normal distribu- tion cannot hold exactly since both a and ak are limited to [0 %, 100 %], we can argue that the t-test is robust against violations of normality (Rasch and Guiard, 2004). It appears that we have a viable random-effects model to justify the application of a second-level t-test to accuracies. But there is a problem with this simple transfer: In contrast to estimated accuracies a, the true accuracy a can never be below the chance level a0. 2.3 There is no negative information To understand why a ≥ a0 must hold, it helps to recognize that MVPA aims at a generic kind of effect, namely, whether or not information about experimental conditions is present 2The specific RFX procedure to apply a second-level t-test or ANOVA to first-level contrast estimates is called the 'summary statistic' approach (Penny and Holmes, 2004) because it considers only the contrast estimates (cid:99)∆βk which summarize single-subject data. It is interesting to note that it is only this summary statistic approach that suggests other first-level summary statistics like classification accuracies could simply be plugged into a second-level t-test. For related attempts at a mixed-effects model for classification accuracies in a Bayesian setting see Olivetti et al. (2012) and Brodersen et al. (2012, 2013). 5 σ2 1 +σ2 Figure 1: Distributions of true (black) and estimated (red) values of contrasts and classification accuracies. - a) An activation difference estimated from a limited amount of noisy data shows variation (red curves; σ1 = 1) around the true value (black bars). Moreover, the true contrast is different in different subjects (Three panels S1, S2, S3; ∆β = −5, 1, 8. Values for these subjects are also indicated in the following panels by dots or bars.) - b) True activation differences (black bars) come from a distribution characterizing the population of possible subjects (black curve; ∆µ = 1, σ2 = 4). Estimated contrasts across subjects show the combined effect of both sources of variation (population + estimation, red curve; 2 = 4.12). - c) The amount of information single-trial data provide about the trial class (condition) and vice versa, as a function of the true activation difference. A negative contrast provides positive information. - d) True accuracy a of classification of univariate single-trial data as a function of the true activation difference. It is a symmetric function of ∆β, which makes accuracy an information-like measure, with a minimal value a0 = 50 %. - e) Estimation variation of accuracy (6-fold cross-validation) in the three subjects (red curves). As apparent for subject 2, the distributions can deviate strongly from normality. While estimated accuracies can be below chance (gray line), true accuracies (black bars; a = 67.7, 51.5, 61.3 %) cannot. - f) The population variation (black curve) and combined variation (population + estimation, red curve) of accuracy that result from the population distribution of contrasts ∆βk (b, black line), the functional relationship between ∆β and a (d), and the estimation distributions (e). The population distribution is restricted to a ≥ a0 and in this example shows a spike at 50 % and a weaker maximum at 56 %. - For further details, see App. B. √ 6 c∆β∆β−20−1001020activationdifferenceS3S2S1a)c∆βk∆βkS3S2S1−20−1001020activationdifferenceb)−60−40−20020406000.20.40.60.81trueactivationdifference∆βmut.information/bitc)chancelevela0−60−40−2002040605060708090100trueactivationdifference∆βtrueaccuracya/%d)aa0255075100accuracy/%S3S2S1e)akakS3S2S10255075100accuracy/%f) in the experimental data (Pereira and Botvinick, 2011). Instead of making the common distinction between univariate and multivariate fMRI analysis, we follow Kriegeskorte et al. (2006; 2007) in distinguishing between activation-based and information-based imaging. Activation-based analysis is interested in whether there is a specific change in the activation of a voxel (average BOLD signal) corresponding to an experimental manipulation, normally an increase. Information-based analysis determines whether there is any change at all, be it an increase or decrease. It looks for brain areas where the difference of conditions makes any kind of difference with respect to the fMRI signal, i.e. for information (Bateson, 1972). Because information-based analysis disregards the sign of activation differences, it has itself an unsigned outcome: There either is a difference, then there is above-zero informa- tion, or there is no difference, then there is zero information ('chance level'). This holds for information-theoretic measures in the strict sense (cf. Cover and Thomas, 2012), in particular mutual information (Fig. 1c), but also for the true value of information-like mea- sures including averaged absolute t and Mahalanobis distance ∆ (used by Kriegeskorte et al., 2006), Wilks' Λ (used by Haynes and Rees, 2005b), linear discriminant t (LD-t, Nili et al., 2014), pattern distinctness D (Allefeld and Haynes, 2014), or classification accuracy a.3 The true single-subject accuracy is either above chance level, a > a0, if it is possible to extract information, or it is at chance level, a = a0, if not - but never below (Fig. 1d). Estimated accuracies can be below chance, but only due to imprecise estimation of a by a, i.e. below-chance accuracies are accounted for by the first-level model of Eq. 4, not the second level of Eq. 5. 2.4 What does a t-test on accuracies mean? This restriction creates a problem for the second-level null hypothesis, which qualita- tively changes its meaning. If (7) is true, then ak ∼ N (a0, ς2 2) (Eq. 5), which means that while half the people in the population exhibit true above-chance classification, the other half is assumed to system- atically exhibit true below-chance classification, contradicting our insight that a ≥ a0. The null hypothesis of the t-test can only be made compatible with this constraint by ad- ditionally assuming that there is no population variation at all, H0 : ¯a = a0 ∧ ς2 2 = 0.4 And this means that the the true accuracy is at chance level for everybody, H0 : ¯a = a0 (8) - there is no information in any subject in the population. Such a null hypothesis, which is a logical conjunction of many simpler null hypotheses, has been called 'global null hypothesis' by Nichols et al. (2005). H0 : ∀k ak = a0, 3We define as 'information-like' those measures that covary with mutual information. Note that this excludes other quantities that may be of interest in MVPA, in particular classifier weights (see Gaonkar and Davatzikos, 2013). Correspondingly, a classifier weight can be negative (cf. Todd et al., 2013), in the simplest case if information is coded in the corresponding voxel by a decrease in activation (but cf. Haufe et al., 2014). 4We would like to thank an anonymous reviewer for pointing out that this poses an additional distributional problem for the use of the t-test: The standard derivation of the null distribution of test statistics becomes invalid when the null hypothesis lies on the border of the parameter space (see Fahrmeir et al., 2013, Sec. 7.3.4). 7 Note that this conclusion does not depend on the assumption of normality in Eq. 5 (which entered through the analogy to Eq. 2): For any distribution of true accuracies, if its mean is at chance but none of its realizations can be below chance, it follows that there can be no above-chance realizations either. Therefore the only possible form in which the null hypothesis formulated in Eq. 7 can hold is given by Eq. 8. The seemingly small constraint a ≥ a0 has strong consequences for inference. If the t-test allows us to reject the null hypothesis, this provides evidence for the alternative, its logical negation. Since the global null hypothesis (Eq. 8) is a universal statement, its negation is a statement of existence: ¬H0 : ∃k ak > a0. (9) This means we have reason to believe that there are some people in the population whose fMRI data carry information about the experimental condition - but we have no grounds to believe that we have found an effect that is typical in the population. The constraint on a neutralizes the RFX modeling of between-subject variation, making the t-test applied to accuracies effectively an FFX analysis.5 Since the constraint that the true value cannot be below chance level applies to all information-like measures, the problems for population inference demonstrated here hold for information-based imaging in general. However, in the interest of conciseness our discussion focuses on cross-validated classification accuracy as the most commonly used measure in MVPA. At this point, the main argument of this paper is concluded: The t-test on accuracies does not provide population inference, but effectively implements fixed-effects analysis. The following section further illustrates and practically demonstrates this fact using simulated data. Readers which are already convinced by the theoretical argument may skip forward to Part 3 which discusses information prevalence inference as an alternative to the t-test. 2.5 What does a t-test on accuracies do? In the previous sections we gave a theoretical argument that the null hypothesis of a second-level t-test changes its meaning under the constraint that holds for information- like measures including classification accuracy. But does this argument have practical relevance - after all, we never see 'true accuracies' but only estimates, which can be below chance? Does our observation actually affect how a t-test on accuracies behaves? To investigate this question, we simulate fMRI data according to the first-level GLM (see Eq. 28) of a simplified experiment containing two experimental conditions, for a sample of subjects. A simulation is parametrized by the population mean activation difference between conditions, ∆µ, and the population variation σ2 (Eq. 2). The estima- tion variation (Eq. 1) is kept at σ1 = 1. Accuracies for the classification of single-trial data by a linear support vector machine are estimated using run-wise cross-validation, 5Todd et al. (2013) point out a related but different problem in applying a second-level test to a summary statistic that estimates an unsigned quantity: The traditional strategy of balancing or random- izing confounds across subjects (cf. Fisher, 1935) does no longer work, because after removing signs confounding effects of different direction cannot cancel each other out. 8 Figure 2: Rejection probability as a function of simulation parameters ∆µ (the population mean true contrast) and σ2 (the population variance of true contrasts), for different null hypothesis tests at significance level α = 0.05. - a) Second-level one-sided one-sample t-test applied to estimated classification accuracies vs chance level a0 = 50 %. The smallest rejection probability is reached if both ∆µ = 0 and σ2 = 0. - b) Second-level two-sided one-sample t-test applied to estimated activation differences vs 0 (univariate RFX analysis). The smallest rejection probability is reached for ∆µ = 0. - c) Fixed-effects analysis of activation differences. The smallest rejection probability is reached if both ∆µ = 0 and σ2 = 0. - d) Test based on classification across subjects. The smallest rejection probability is reached for ∆µ = 0. 9 −1−0.500.5100.511.52∆µσ2a)t-testonaccuracies−1−0.500.5100.511.52∆µσ2b)t-testonactivationdifferences(RFX)−1−0.500.5100.511.52∆µσ2c)FFXanalysis−1−0.500.5100.511.52∆µσ2d)classificationacrosssubjects0.050.10.20.51rejectionprobability and these accuracies are entered into a one-sided one-sample t-test vs chance level a0 = 50 % across subjects at α = 0.05. This is repeated many times to determine the probability to reject the null hypothesis. For full details of the simulations, see App. B and C. 2.5.1 Classification of univariate data with a normal population model For simplicity, we first examine the rejection probability of the t-test on accuracies for the classification of univariate data; Fig. 1 illustrates the distributions arising in this case. The resulting rejection probability as a function of the simulation parameters ∆µ and σ2 is shown in Fig. 2a. The rejection probability function is a standard tool to check whether a test is valid for a given null hypothesis and how powerful it is (Lehmann and Romano, 2005). Here, we use the function in the opposite direction: We define the test's effective null hypothesis as that set of parameter values where the rejection probability remains at or below the specified significance level α. For the t-test on accuracies (Fig. 2a) the result is that strictly there are no such parameter values: the smallest rejection probability is 0.055. This is most likely because the nor- mality assumption of Eq. 6 holds only approximately; the slight increase of the α-error is however within the bounds used by Rasch and Guiard (2004) when stating that the t-test is robust against violations of normality. If we disregard this deviation, the effec- tive null hypothesis turns out to be ∆µ = 0 ∧ σ2 2 = 0: no population variation, and no activation effect in any subject. This means that the true activation difference is zero in all subjects, ∀k ∆βk = 0, and since the true accuracy corresponding to a zero activation difference is at chance level, a(∆β = 0) = a0, it is equivalent to H0 : ∀k ak = a0 - no information in any subject in the population. The simulation confirms the practical relevance of our previous theoretical conclusion that the t-test on accuracies tests the global null. For comparison, Fig. 2b shows the rejection probability function of a second-level two- sided t-test on first-level activation differences (univariate RFX analysis). In accordance with the design of the test, its rejection probability remains at 0.05 if and only if ∆µ = 0, and increases monotonically with ∆µ. Though σ2 2 is not part of the specification of the null hypothesis, we observe that for stronger population variation it becomes harder to reach a significant effect. This is in stark contrast to the t-test on accuracies (Fig. 2a), where increasing σ2 2 makes it easier to reject the null hypothesis. The more inconsistent activation differences (or by extension: patterns; see below) are in the population, the more likely it is that the t-test on accuracies will indicate the presence of an information effect that is supposedly 'typical' in the population!6 As we noted above, univariate fixed-effects analysis does not provide population- level inference simply because its null hypothesis (that the sample mean is 0) does not reference a population distribution. We can however apply an FFX test to our simulated univariate RFX data and thereby determine the null hypothesis it effectively implements in this context. The result shown in Fig. 2c demonstrates that the effective null hypothesis of FFX analysis is ∆µ = 0 ∧ σ2 2 = 0. Though their rejection probability 6Davis et al. (2014) make a similar observation, but without noting its consequences for population inference. 10 functions are different in detail, a t-test on accuracies and a fixed-effects analysis of activation differences operate qualitatively identical: they both detect deviations from a zero population mean contrast ∆µ as well as from a zero population variance of constrasts σ2 2. This agreement supports our earlier statement that even though a t- test on accuracies formally acknowledges population variation, it does not provide population inference any more than FFX analysis does. To complete the picture, Fig. 2d shows the rejection probability function of a test based on classification across subjects (cf. Mourao-Miranda et al., 2005). The single-subject univariate contrast estimates in each of the two conditions form the data set that is used to train and test a classifier in leave-one-subject-out cross-validation, and the distribution of accuracies is determined for different simulation parameters. For a given simulated data set the null hypothesis (true accuracy = chance level) is rejected if the accuracy reaches or exceeds the critical value of 67.6 %, which is the approximate 95th percentile of the distribution of accuracies under the null hypothesis. The result demonstrates that this test behaves in a way that is very similar to the second-level t-test applied to the same first-level activation differences (Fig. 2b). Both provide population inference because they both implement the null hypothesis ∆µ = 0 for arbitrary population variance σ2 2. The difference is that while the second-level t-test is limited to univariate contrasts, the test based on classification across subjects can just as well be applied to multivariate data, where the null hypothesis becomes ∆(cid:126)µ = (cid:126)0. If a significant effect is found, this provides evidence that there is a pattern difference ∆(cid:126)µ that is typical in the population. Since this implies that the presence of information is typical in the population, it appears that the problem of population inference for information-based imaging could simply be solved by applying classifiers or other MVPA methods always across subjects. Unfortunately, spatial normalization algorithms do not achieve precise voxel-level anatomical alignment (Thirion et al., 2006), and moreover we cannot assume that a one- to-one correspondence of informative patterns between subjects always exists (Haynes and Rees, 2006; Kriegeskorte and Bandettini, 2007; Haxby, 2012), so that classification across subjects is often bound to fail for a trivial reason.7 2.5.2 Classification of multivariate data with a normal population model An important observation in the last section was that the larger the population variance σ2 2, i.e. the more inconsistent activation differences are across subjects, the easier it becomes for the t-test on accuracies to achieve significance. To make sure that this effect is not limited to univariate data, Fig. 3a shows the rejection probability function for different numbers of dimensions, p = 1, 2, and 10. Data are generated as before, but for p different voxels in parallel, with both first- and second-level variation uncorrelated between voxels. Accordingly, the classifier is trained and tested in a p-dimensional space. To facilitate visualization, we kept the population mean activation difference at zero in all voxels, ∆µ = 0. The result demonstrates that the effect of population 7A solution might be provided by 'hyperalignment' (Haxby et al., 2011) which attempts to establish a fine-grained functional correspondence between different subjects' brains. However, as Todd et al. (2013) point out, aligning patterns effectively discards sign (direction) information, too, unless the hyperalignment parameters are determined from separate data. 11 Figure 3: Rejection probability of a second-level one-sided one-sample t-test applied to esti- mated classification accuracies vs chance level, as a function of simulation parameters. The significance level α = 0.05 is shown as a gray horizontal line. - a) Multivariate normally distributed contrasts in p voxels, with variation uncorrelated between voxels and a population mean activation difference of ∆µ = 0 everywhere. The rejection probability increases with the population variance σ2 2 , and the increase is stronger for higher dimensionality p. The line for univariate data (p = 1) corresponds to a central vertical section through the rejection probability function of Fig. 2a. - b) Population proportion model: Fixed true contrast ∆β∗ in a proportion γ of the population, and 0 in the rest. The rejection probability always reaches α for γ = 0. variance on the rejection probability stays qualitatively the same, but is even stronger in higher dimensions. We can therefore assume that the effective null hypothesis of the t-test on accuracies generalizes from the univariate ∆µ = 0 ∧ σ2 2 = 0 to the multivariate ∆(cid:126)µ = (cid:126)0 ∧ Σ2 = 0, where Σ2 is the multivariate population variance (variance–covariance matrix). That implies ∀k ∆(cid:126)βk = (cid:126)0 - there is no informative pattern in any subject in the population - which is again equivalent to H0 : ∀k ak = a0, the global null. 2.5.3 Classification of univariate data with a population proportion model Our simulation-based finding that the effective null hypothesis of a t-test applied to first-level accuracies is H0 : ∀k ak = a0 (the global null) was so far obtained using the standard normal distribution population model for activation differences (Eq. 2), which may not be correct (cf. Rosenblatt et al., 2014). With respect to true accuracies, this model for true contrasts has a peculiar consequence: As soon as there is any population variation, i.e. σ2 2 > 0, the probability that the true contrast in a given subject k is exactly ∆βk = 0 becomes zero, which implies that almost always ∀k ∆βk (cid:54)= 0. Since for non- zero true contrast the true accuracy is above chance, a(∆β (cid:54)= 0) > a0 (cf. Fig. 1d), this further implies that almost always ∀k ak > a0. Therefore the assumption of a normal population distribution of activation differences allows only two possibilities: Either there is no information in the data of any subject (the effective null hypothesis), or there is information in the data of every subject. There is nothing in between. To see how the t-test on accuracies reacts to a situation between these extremes, we use an alternative population model for activation differences (replacing Eq. 2; modified 12 p=1p=2p=1000.511.5200.20.40.60.81σ2rejectionprobabilitya)∆β∗=1∆β∗=2∆β∗=500.20.40.60.8100.20.40.60.81γrejectionprobabilityb) from Rosenblatt et al., 2014): Assume that the true contrast has a fixed value ∆β∗ in a certain proportion γ ∈ [0, 1] of the population and the fixed value 0 in the rest. If now a subject k is randomly selected from the population, this means that ∆βk =(cid:26) 0 with probability 1 − γ ∆β∗ with probability γ . (10) Fig. 3b shows the behavior of the t-test in a simulation using this model. The result is that for different values of ∆β∗, the rejection probability reaches the significance level α always at γ = 0. This effective null hypothesis is again equivalent to the global null, H0 : ∀k ak = a0. Moreover, the simulation demonstrates that the t-test on accuracies may with high probability declare an information effect to be 'typical' even though it is only present in a small minority of subjects in the population! 3 An alternative: Information prevalence inference In the previous part we established that the second-level t-test applied to accuracies is not able to provide population inference. We now discuss alternative approaches, leading us to the idea that population inference for information-based imaging should target the proportion of people in the population in which there is an information effect. Within the MVPA literature, there are three alternative proposals. First, Kriegeskorte and Bandettini (2007) recommend to apply the methods for 'combining brains' collected by Lazar et al. (2002). However, except for the summary-statistic approach of Holmes and Friston (1998), all of these methods8 are meta-analytic procedures which explicitly test the global null. Second, Stelzer et al. (2013) propose to generate single-subject permutation statistics, and then to construct a second-level permutation distribution by randomly selecting first-level permutations in each subject. Because each permutation realizing the second-level null hypothesis is a combination of permutations realizing the first-level null hypotheses in every subject, this again tests the global null hypothesis of no information in any subject. And third, Brodersen et al. (2013) follow Olivetti et al. (2012) and Brodersen et al. (2012) in describing a mixed-effects analysis for MVPA, introducing explicit estimation and population models for classification accuracies. This approach offers several improvements over the t-test on accuracies: Their model can account for different estimation variances (ς2 1) in different subjects, and it uses more realistic distributional assumptions. However, the authors do not consider the fact that true accuracies are limited to the range [a0, 1] (for binary classification: [50 %, 100 %]). Unfortunately, this renders their approach only an improved version of the t-test on accuracies. As detailed in Sec. 2.4, the flaw of the t-test on accuracies lies in the fact that the distributional assumption of the underlying second-level model (Eq. 5) is incompatible with the restriction a ≥ a0, unless ς2 = 0. This could conceivably be fixed by using another population model which adheres to that restriction by design. However, the simulated population distribution of true accuracies for univariate classification in 8Fisher's (1925) combined probability test, Tippett's (1931) minimum p-value, the conjunction test of Worsley and Friston (2000), Stouffer et al.'s (1949) combined z-value, Mudholkar and George's (1979) logit method, and fixed-effects analysis. 13 Fig. 1f does not look like it could be appropriately captured by a parametric model. And even if that were the case, the distribution would likely have a different shape for classification in higher dimensions or for multi-class classification. It might additionally depend on the specific classification algorithm, and it would certainly be different again for other information-like measures. Moreover, inference with respect to the population mean is generally inadequate for information-like measures no matter how well the actual population distribution can be modeled. The reason is that the true mean is above chance as soon as there is a true above-chance effect in a small fraction of the population (cf. Fig. 3b; because there can be no true below-chance effect). This problem can be resolved by inference that does not target the mean effect, but the proportion of subjects in the population in which there is an information effect, i.e. the prevalence of information. Such an approach is followed by Rouder et al. (2007) in the context of signal detection theory with their 'mass-at-chance model'. They assume a probit-normal population distribution of true detection accuracies, which however is truncated at 50 %, such that all subjects that are nominally below chance are instead located at chance. Inference with respect to the prevalence of an effect has also been advocated by Rosenblatt et al. (2014) for mass-univariate analysis. The authors argue that imperfect alignment of single-subject activation maps leads to areas where only a subset of subjects have a non-zero activation, and propose to replace the standard normal population model (Eq. 2) by a mixture model. We used a simplified version of this in the population proportion simulation (Eq. 10 and Fig. 3b). Stephan et al. (2009) propose a Bayesian form of prevalence inference as RFX analysis for dynamic causal models (DCM), ex- tending first-level model selection to a posterior distribution over the space of different model frequencies. In a specific application this discrete distribution may describe the distinction between a zero effect and a generic non-zero effect.9 Friston et al. (1999a) build upon their previous idea of a conjunction test (Price and Friston, 1997; Worsley and Friston, 2000) and introduce the minimum-statistic approach. Their analysis pro- ceeds in two steps: first, the minimum statistic is used to derive a p-value for the null hypothesis that there is no effect in any subject in the population, the global null. In a second step, a correction is applied that allows to test the null hypothesis that the proportion of subjects in which there is an effect, γ, is at or below a threshold γ0. The rejection of this null hypothesis therefore allows to infer that γ > γ0. The approaches of Rouder et al. (2007), Rosenblatt et al. (2014), Stephan et al. (2009), and Friston et al. (1999a) are all candidates to be adapted for information-based imaging, to provide population inference with respect to the prevalence of an information effect. In the following we demonstrate this in detail for the method of Friston and colleagues. 9Stephan et al.'s Bayesian RFX analysis for DCM has been adapted for GLM model selection by Soch et al. (2016), and may be adapted for MGLM (cf. Allefeld and Haynes, 2014) model selection to support MVPA. 14 4 Permutation-based information prevalence inference using the minimum statistic In this part we recapitulate the minimum-statistic approach to prevalence inference developed by Friston et al. (1999a), adapt it to be based on permutation statistics, and detail the resulting algorithm. Applied to information-like measures this method allows us to achieve information prevalence inference, i.e. inference with respect to the proportion of subjects in the population that exhibit an information effect. We demonstrate the method using an example data set. The advantage of Friston et al.'s approach is that it can be implemented based on known permutation methods at the single-subject level (Golland and Fischl, 2003; Etzel and Braver, 2013; Schreiber and Krekelberg, 2013; Stelzer et al., 2013; Allefeld and Haynes, 2014; see also Ernst, 2004). The method is discussed with respect to classification accuracy, but the test logic applies equally to other information-like measures. A note on notation: We previously used variables without index (e.g. a) when talking about the 'first level' of a given single subject, and symbols with index (e.g. ak) when considering a subject as randomly selected from the population, or referring to all members of the population (∀k). This notation is still followed, but we now extend it such that an index associated with an explicit range, k = 1 . . . N, refers to the specific subjects included in a given sample. 4.1 The minimum statistic and the global null In a single subject, an estimated classification accuracy a has an associated p-value p(a), which is the probability to observe an accuracy that large or larger given that the true accuracy is at chance level (a = a0). For a sample of N subjects with estimated classification accuracies ak, k = 1 . . . N, we choose the minimum statistic (smallest observed accuracy) as the second-level test statistic, m = N min k=1 ak. (11) (12) As a second-level null hypothesis we first consider the global null hypothesis, H0 : ∀k ak = a0, that there is no effect in any subject in the population. In order to test this null hypothe- sis, we need the p-value of m, pN(m), with respect to the global null. To say that the minimum of estimated accuracies is at or larger than a given value m is the same as saying that all of the estimated accuracies ak, k = 1 . . . N, are at or larger than m. Since subjects are independently drawn from the population, the probability to observe a minimum of m or larger under the global null is the product of probabilities to observe an estimated accuracy of m or larger in each subject in the sample: pN(m) = N∏ k=1 p(m) = p(m)N, 15 (13) where p(m) is the single-subject p-value for a = m. If pN(m) ≤ α, then we can reject H0 and infer that there are some subjects in the population in which there is an above-chance effect. Since this is a statement of existence, this does not provide evidence that the effect is typical in the population. 4.2 The prevalence null We now consider a population model which contains the global null as a special case: The information effect targeted by the classification procedure has a prevalence γ, i.e. a proportion γ ∈ [0, 1] of subjects in the population have an above-chance effect, the others no effect. If a subject k is selected at random from this population, then ak = a0 with probability 1 − γ, ak > a0 with probability γ. (14) This is similar to the population proportion model for activation differences used above (Eq. 10), but in contrast no assumption is made about the size and distribution of above-chance effects. An estimated accuracy a in a single subject can be larger than or equal to m either purely by chance (p(m), with probability 1 − γ) or because there is actually an effect in that subject (p(ma > a0), with probability γ). The probability to observe a sample minimum of m or larger if the prevalence is γ is therefore pN(mγ) = N∏ [(1 − γ) p(m) + γ p(ma > a0)] = [(1 − γ) p(m) + γ p(ma > a0)]N. k=1 (15) Here p(ma > a0) is the probability to observe an estimated accuracy of m or larger in a single subject given a true accuracy of a, where we only know that a > a0. Because the size of the above-chance effect a > a0 is not specified, we cannot know this probability precisely; but because it is a probability, we know that it is smaller than or equal to one, p(ma > a0) ≤ 1, and therefore pN(mγ) ≤ [(1 − γ) p(m) + γ]N. (16) The reason we chose the minimum statistic as the second-level test statistic is that it enables us to formulate this inequality for the prevalence model, i.e. to determine a p-value without specifying the size and distribution of above-chance effects. The prevalence model allows us to formulate the prevalence null hypothesis, H0 : γ ≤ γ0, (17) that the prevalence is smaller than or equal to a threshold prevalence γ0. The global null (Eq. 12) is a special case of the prevalence null where γ0 = 0. The prevalence null hypothesis is a complex null hypothesis, i.e. it can be realized by different values of the parameter γ. In such a case, the p-value associated with a 16 test statistic is the probability to observe the given value or larger, maximized over all situations consistent with H0. Therefore pN(mγ ≤ γ0) = [(1 − γ0) p(m) + γ0]N. (18) In a permutation approach, pN(m) can be more precisely determined than p(m) (see step 3 under Algorithm). We therefore express the prevalence null p-value pN(mγ ≤ γ0) in terms of the global null p-value pN(m), using the relation pN(m) = p(m)N (Eq. 13): pN(mγ ≤ γ0) = [(1 − γ0) N(cid:113)pN(m) + γ0]N. (19) If pN(mγ ≤ γ0) ≤ α, then we can reject H0 and infer that the prevalence γ is sig- nificantly larger than γ0, i.e. more than a proportion γ0 of the population have an effect. As an alternative to fixing a threshold prevalence γ0 in advance and then testing the cor- responding prevalence null, we can compute the largest γ0 such that the corresponding null hypothesis can still be rejected at the given significance level α: γ0 = . (20) N√ α − N(cid:112)pN(m) 1 − N(cid:112)pN(m) Note that this is not an estimator for the true prevalence γ, but ]γ0, 1] is a one-sided (1 − α)-confidence interval for it. (Eq. 16). Moreover, even for the strongest possible effect (pN(m) = 0) it holds γ0 = N√ This confidence interval will often be too wide because of the inequality used above α, meaning that the strength of the population inference is limited by the number of subjects N and the chosen significance level α. 4.3 Information maps So far we have considered a single second-level test based on classification accuracies ak from N different subjects. But in a common variant of MVPA, searchlight analysis (Kriegeskorte et al., 2006), we have maps of classification accuracies and perform a test at each voxel in these maps, i.e. we have to adjust for multiple comparisons. To do so, we need to specify a spatially extended version of the prevalence null. Again following Friston et al. (1999a), our spatially extended null hypothesis is: - there is an effect with prevalence γ ≤ γ0 in a small area, - and no effect everywhere else. The justification for this is that in experiments investigating the localization of in- formation we normally expect this information to be restricted to specialized brain areas. Under this null hypothesis, a sample minimum of m or larger can occur either purely by chance (no true effect), with a probability that is increased because we examine many voxels at once (p∗ N(m)) it can occur because there actually is an effect in a sub-threshold proportion of the population, with a probability that is not increased because the effect is only present in a small area (pN(mγ ≤ γ0)). N(m)); or if that is not the case (1 − p∗ 17 Here p∗ N(m) is the p-value for the spatially extended global null, corrected for multiple comparisons using a standard method (see step 4 under Algorithm). Taken together, the probability to observe a sample minimum of m or larger at a given voxel, corrected for multiple comparisons according to the spatially extended prevalence null hypothesis is p∗ N(mγ ≤ γ0) = p∗ N(m) + [1 − p∗ N(m)] pN(mγ ≤ γ0). (21) For given threshold γ0, the spatially extended prevalence null can be rejected at a N(mγ ≤ γ0) ≤ α. Equivalently, we can define a significance level particular voxel if p∗ that is corrected for multiple comparisons, ∗ = α α − p∗ 1 − p∗ N(m) N(m) , (22) and reject the spatially extended prevalence null if the uncorrected p-value is at or below the corrected level, pN(mγ ≤ γ0) ≤ α∗. Note that because m is voxel-specific, α∗ is, too. Again, we can alternatively compute the largest γ0 such that the corresponding spatially extended prevalence null can still be rejected: ∗ 0 = γ N√ α∗ − N(cid:112)pN(m) 1 − N(cid:112)pN(m) , (23) which results in a map of lower confidence bounds of the prevalence of the effect, an information prevalence map. 4.4 Algorithm We now explain in detail how the computations derived above can be implemented based on first-level permutation statistics. Step 1: For each subject, classification accuracies av are computed for each voxel v. Additionally, classification accuracies are computed for data where the class labels have been permuted, avi with i = 1 . . . P1, where P1 is the number of available first-level permutations. i = 1 denotes the neutral permutation, i.e. av1 = av. Step 2: The minimum classification accuracy mv across subjects (Eq. 11) is computed at each voxel v. Additionally, the minimum accuracy is computed for each second-level permutation, mvj with j = 1 . . . P2. A second-level permutation is a combination of first- level permutations, and therefore there are PN 1 possible second-level permutations. If there are too many combined permutations, a subset can be selected randomly (Monte Carlo estimation), but it has to be made sure that j = 1 denotes the combination of first-level neutral permutations (i = 1 in all subjects). This procedure of combined permutations is identical to the one employed by Stelzer et al. (2013). Step 3: The uncorrected p-value for the global null hypothesis is determined at each voxel (Eq. 13) as [mv ≤ mvj] (24) pN(mv) = 1 P2 P2∑ j=1 18 where the so-called Iverson bracket [·] has the value 1 for a true condition and the value 0 for a false condition. That is, pN(mv) is the fraction of combined-permutation values of the minimum statistic larger than or equal to the actual value. Because mv1 = mv, the smallest possible p-value is 1 P2 Step 4: To correct pN(mv) for multiple comparisons, the maximum statistic across voxels is computed for each combined permutation (see Nichols and Holmes, 2001), . Mj = max v mvj, and then p∗ N(mv) = 1 P2 [mv ≤ Mj] P2∑ j=1 (25) (26) N(mvγ ≤ γ0) ≤ α. N(mvγ ≤ γ0) from p∗ is determined. p∗ N(mv) is the p-value for the spatially extended global null hypothesis. Step 5a: To determine where the spatially extended prevalence null hypothesis for a given threshold γ0 can be rejected, at each voxel Eq. 19 is used to compute pN(mvγ ≤ N(mv) and pN(mvγ ≤ γ0), γ0) from pN(mv), Eq. 21 to compute p∗ and it is checked whether p∗ Step 5b: Alternatively, to determine for each voxel the largest threshold γ∗ 0 at which the spatially extended prevalence null hypothesis can be rejected, Eq. 22 is used to compute α∗ v from p∗ v is checked to see whether the spatially extended global null hypothesis can be rejected. If that is not the case, the largest threshold γ∗ 0 is not defined for that voxel (the prevalence null cannot be rejected, even at γ∗ 0 = 0). For all voxels where the spatially extended global null hypothesis can be rejected, Eq. 23 is used to compute γ∗ v and pN(mv). Note that the maximally possible γ∗ 0v determined this way is limited not just by the chosen significance level α and the number of subjects N, but also by the number of second-level permutations P2; it is N(mv). Then pN(mv) ≤ α∗ 0v from α∗ max − N√ α∗ 1 − N√ 1/P2 N√ ∗ 0max = γ 1/P2 with α ∗ max = . (27) α − 1/P2 1 − 1/P2 max is approximately equal to α. For large P2, α∗ A problem for this method may arise from the fact that both the minimum statistic underlying prevalence inference and the maximum statistic used to correct for multiple comparisons do not produce new values (unlike e.g. the mean, which in general differs from all the values it is calculated from). Since the number of possible classification accuracies is limited because of a limited number of data points, this may lead to a large number of permutations where the statistic attains the same value (tied permutation values), which inflates the p-values computed in steps 2 and 3 above. This problem can be solved by using spatially smoothed accuracy maps as inputs (which is also advisable to reduce residual anatomical misalignment between subjects), or by using a continuously-valued information-like measure like pattern distinctness (Allefeld and Haynes, 2014) instead. 4.5 Application In order to illustrate our permutation-based implementation of information prevalence inference, we re-use the data of Cichy et al. (2011). Twelve different visual stimuli 19 belonging to four different categories were presented either to the left or the right of fixation (24 experimental conditions) to N = 12 subjects. There were four different trials per condition in each of five different runs. fMRI data were recorded from a field of view covering the ventral visual cortex at an isotropic resolution of 2 mm. Data were preprocessed and normalized to the MNI template. A linear SVM with parameter C = 1 was trained on GLM parameter estimates from four of the runs, and tested on the fifth run, in a leave-one-run-out cross-validation scheme. Classification was pairwise (24 · 23 / 2 = 276 pairs) and accuracies were averaged across pairs combining different factor levels, so that the chance-level accuracy was a0 = 50 %. For permutation statistics, class labels were exchanged in each of the five runs separately, which lead to P1 = 25−1 = 16 unique first-level permutations. The analysis was performed using a searchlight of radius 4 voxels (comprising 257 voxels). The resulting accuracy maps were smoothed with a Gaussian kernel of 6 mm FWHM. For more details, see Cichy et al. (2011) and Allefeld and Haynes (2014). For information prevalence inference, we 1 = 2.8 · 1014 possible combined permutations at randomly selected P2 = 107 out of PN the second level. All the following results are corrected for multiple comparisons, but for simplicity we omit the superscript '∗'. Figure 4: Second-level results for the classification of object category of a visual stimulus (see Cichy et al., 2011) shown in a sagittal slice through right lateral occipital cortex and fusiform gyrus. Statistics are corrected for multiple comparisons. - a) Information prevalence inference. Highlighted areas are those where the global null hypothesis (prevalence γ = 0) can be rejected at a level of α = 0.05. Colors visualize a lower bound γ0 on the prevalence of category information (confidence level 0.95). - b) For those areas where the prevalence null hypothesis γ ≤ γ0 can be rejected at γ0 = 0.5, i.e. where it can be inferred that the majority of subjects in the population have an effect, colors visualize the median classification accuracy across subjects. - c) t-test on accuracies vs chance level a0 = 50 %. For those areas where the null hypothesis ¯a = a0 can be rejected at a level of α = 0.05, colors visualize the underlying t-value. The results for the classification of stimulus category are shown in Fig. 4. The spatially extended global null hypothesis of no information in any subject in the population can be rejected at a level of α = 0.05 in about 27 % of in-mask voxels. For those voxels, the largest lower bound γ0 at which the spatially extended prevalence null hypothesis can be rejected is shown in Fig. 4a. In about half of those voxels, γ0 reaches the maximally possible value (for the given sample size N, significance level α, and number of second- level permutations P2) of γ0max = 0.701 (Eq. 27). For those voxels where the largest lower bound γ0 is larger than or equal to 0.5, 20 x=32prevalenceγ0a)00.20.40.60.81medianaccuracy/%b)5060708090100t-valuec)0510152025 i.e. where we can infer that in the majority of subjects in the population the data contain information about the stimulus category, the median estimated accuracy is shown in Fig. 4b. We chose the median across all subjects in the sample as a descriptive statistic to accompany our prevalence results, because it can be considered as a cautious and robust estimator of the typical above-chance classification accuracy.10 For comparison, the result of a second-level t-test on accuracies vs chance level is shown in Fig. 4c. Although the effective null hypothesis of this test is identical to the global null hypothesis γ = 0 explicitly tested in Fig. 4a, the number of voxels at which it can be rejected is much smaller (about 14 % of in-mask voxels at α = 0.05, FWE-corrected), indicating that in this case the t-test is less sensitive. The picture is the same for the comparison with the test of the prevalence null hypothesis γ ≤ γ0 at γ0 = 0.5 shown in Fig. 4b. Information prevalence inference therefore allows us to draw conclusions that are stronger than those provided by the t-test on accuracies, concerning both interpretation - population inference - and, for this data set, statistical power. However, the result of Fig. 4b also calls into question whether the assumption that an effect is constrained to a 'small area' (adopted from Friston et al., 1999a) is generally adequate. 5 Discussion In this paper we have shown that the t-test on accuracies commonly used in MVPA studies is not able to provide population inference because the true single-subject accuracy a can never be below chance level. This constraint makes the effective null hypothesis of the test the global null hypothesis that there is no effect in any subject in the population, which means that in rejecting that null hypothesis we can only infer that there are some subjects in which there is an effect. This is in stark contrast to the standard interpretation of a significant result of a second-level t-test. We supported our statement both by theoretical arguments, in particular detailing that classification accuracy in MVPA is an information-like measure, as well as by simulations of the rele- vant distributions to investigate the practical behavior of the t-test applied to accuracy data. Finally, we reviewed possible alternative inference methods and described one approach that can be implemented based on known first-level permutation statistics in combination with the minimum statistic as the second-level test statistic. In the following we discuss a number of possible counter-arguments and questions. Does this mean that the t-test fails, is not robust enough? - Not really. In all the instances we examined, the t-test does what it is supposed to do: Check whether the population mean is increased. The point is that under the constraint a ≥ a0, rejecting H0 : ¯a = a0 no longer tells us that the effect is typical in the population; an increased accuracy in a small fraction of the population is sufficient to increase the population mean (cf. Fig. 3b). Average information content is not typical information content. Concerning robustness, 10For γ > 0.5 we can consider subjects where the true accuracy is at chance (no information) as 'outliers'. The median is a robust estimator with a breakdown point of 0.5, i.e. it can handle samples where up to half of the values are outliers (Huber and Ronchetti, 2009). It is cautious in the sense that it will under- rather than overestimate the the true median above-chance accuracy, because all 'outliers' are on the side of small values. 21 for the case we simulated (Fig. 2a) the α-error was only slightly increased, consistent with the notion that the t-test is indeed robust. Is it wrong to test the global null hypothesis? - No. The decision which null hypothesis to test rests with the researchers performing a study. The problem lies with the inter- pretation of a significant result, which for second-level analysis - at least tacitly - is to infer that the effect is typical in the population. Moreover, the step from sample to population inference (FFX to RFX) as the gold standard has been taken already a long time ago for univariate analyses (Holmes and Friston, 1998) and more recently for DCM (Stephan et al., 2009); it would therefore be natural to expect that it also becomes the standard for information-based imaging. In particular, any claim that MVPA is more sensitive than univariate analysis (Norman et al., 2006) is meaningless if MVPA is not held up to the same inferential standards. We observe below-chance accuracies all the time. - There are two aspects to this. For one, our statement that accuracies cannot be below chance refers to true accuracies, not to estimated accuracies. Second, it is sometimes the case that estimated accuracies strongly suggest that the true accuracy is below-chance, too.11 This is most likely due to the circumstance that a crucial assumption of cross-validation is not met, namely that the different parts of the data (here: fMRI recording sessions) come from the same distribution (Efron and Tibshirani, 1994). If there are systematic changes across data parts, for example because of a confound either in the data or in the experimental design (Görgen et al., 2014), this can induce a negative bias, including the possibility that classifier performance lies systematically below chance. (For another tentative explanation, see Kowalczyk, 2007.) This does of course not mean that now there is 'negative information', but only that cross-validated accuracy is not a meaningful information measure under such circumstances. This possibility does not invalidate the argument made in this paper, but points to another problem that needs a separate remedy. A classifier could be designed to systematically give the wrong answer, leading to a true accuracy below chance. - Yes, but in that case the accuracy of this classifier would no longer be a measure of the information content of the data. For example in the simple case where the output of a working classifier is falsified by always returning the opposite classification result (A instead of B and B instead of A), information content would no longer be quantified by a − a0, but by −(a − a0). The argument in this paper does not refer to classifiers in general, but to the use of classification accuracy and other measures to quantify the information content of data. What does it mean for an effect to be 'typical' in the population? - This question essentially asks about the scientific content of statistical inference at the population level. Surpris- ingly, the topic is almost never discussed in statistical scientific papers and textbooks, including those aimed at psychologists and other cognitive scientists. Our use of the term 'typical' was inspired by Penny and Holmes (2007), who state that in population inference 'one is interested in what is common to the subjects' or in a 'stereotypical effect in the population' (p. 156). In this paper, we use the term mainly in a negative way: An effect that is only present in a small fraction of the population can hardly be 11With respect to fMRI–MVPA, this phenomenon is informally discussed (see e.g. J. Etzel's blog, http://mvpa.blogspot.de/2013/04/below-chance-classification-accuracy.html), but appears to not yet have given rise to a peer-reviewed publication. 22 considered typical (or 'common', or 'stereotypical'). In standard univariate analysis concerning the mean of a normal distribution population model, a positive use of the term can be motivated by the fact that the mean is also the mode of the distribution (the most frequent realization) as well as its median (the value that sits 'in the middle' of the distribution). But there is another aspect: If in this context we can reject the null hypothesis H0 : ∆µ = 0 in a one-sided t-test, the inference ∆µ > 0 also means that more than half of the population - the majority - have a positive effect (and less than half a negative effect). This statement can be extended to the non-normal case if there is a test that can show that the population median is above zero. If we take this observation as a guideline, the natural choice for the threshold in prevalence inference is γ0 = 0.5 (see also Friston et al., 1999b). If we can reject this null hypothesis, we can again infer that there is an effect in the majority of subjects in the population, which also implies that the median true effect strength is above-chance (motivating Fig. 4b).12 We therefore propose to call an effect 'typical' if it is present in the majority of subjects in the population. But don't we want to show that the effect generalizes in the sense that it is present in every subject in the population? - Maybe ideally, but in practice this is impossible. First, it is not known whether any of the effects that are of interest in functional neuroimaging do actually exhibit such an extreme degree of generalization, in particular with respect to a specific anatomical location. Second, statistical inference is unable to provide support for such a statement on principle. As pointed out above, a univariate one-sided t-test for a normal distribution only provides evidence that a majority of subjects in the population have an effect, not that it generalizes to everyone. And in population prevalence inference, at best we could test the prevalence null hypothesis for γ0 = 0.99 or similar, at the price of very low sensitivity. Concluding we would like to point out that we do not consider the method of 'permutation-based information prevalence inference using the minimum statistic' put forward in the last part to be the definitive solution to the problem of population inference in information-based imaging. The main aim of this paper was to demonstrate conclusively that the t-test on accuracies does not provide population inference, and that prevalence inference is an alternative. This particular method was presented in order to show that population inference is possible, even based on established methodology only (minimum statistic, first-level permutations). However, it does have several shortcomings: The use of the minimum statistic limits the highest γ0 for which the prevalence null hypothesis can be rejected depending on the number of subjects, and the permutation-based implementation imposes an even lower limit depending on the number of second-level permutations that can be performed. Moreover, this method does not provide a way to estimate (instead of just bound) the true population prevalence γ. We do however believe that methods focusing on population prevalence are the most promising approach, not just for information- but also for activation-based imaging, because they explicitly provide information about the degree to which an effect generalizes. And while we hope that this paper will motivate further methodological work on population inference for information-based imaging, our method does provide 12Additionally, this choice is consistent with the use of the 'exceedance probability' in Bayesian RFX (Stephan et al., 2009), which is the posterior probability that a given model is more frequent in the population than all other models, if only two models are considered. 23 a way to improve upon the commonly used t-test on accuracies that is available now.13 Acknowledgments Kai Görgen was supported by the German Research Foundation (DFG grants GRK1589/1 and FK:JA945/3-1). The authors would like to thank Tom Nichols, Jakob Heinzle, Jörn Diedrichsen, Will Penny, María Herrojo Ruiz, Joram Soch, Martin Hebart, Jo Etzel, Yaroslav Halchenko, and Thomas Christophel for discussions, comments, and hints. Appendix A What is a true accuracy? The first-level model (Eq. 4) for the RFX analysis of accuracies assumes that there is a 'true accuracy' a that underlies the random estimated accuracies a that we actually observe. But what does this true value stand for? In the case of a true activation difference ∆β, the answer is simple: The first-level GLM yt = βA xAt + βB xBt + et (28) where xAt and xBt are regressor functions and et is noise, provides a statistical descrip- tion of the (neural and hemodynamic) process by which we assume the observed fMRI data are generated (a generative model). This model includes parameters βA and βB which govern the generative process, and which characterize the respective subject. From a given data set yt, we can estimate these parameters, βA and βB, but because the data are noisy, these estimates will not coincide with the underlying true values. How- ever, if we could repeat the experiment an infinite number of times with the same subject, the mean of estimates across these repetitions would recover the true values, because the estimators are unbiased (provided the parameters are estimable). This mean is also called expectation value; (cid:104) βA(cid:105) = βA, (cid:104) βB(cid:105) = βB, and therefore (cid:104)(cid:99)∆β(cid:105) = ∆β = βB − βA (cf. Eq. 1). For a true accuracy a the situation is not that simple, because we do not have a genera- tive model of the data yt that is parametrized by a. But, consistent with the RFX model for accuracies and in analogy to GLM parameters, we can define the true accuracy as the expectation value of estimated accuracies, a = (cid:104)a(cid:105), i.e. the mean across an infinite number of repetions of the experiment with the same subject. This true accuracy a is a complex function of the true GLM parameters for the included voxels and conditions, as well as the error variance and covariance, and may also depend on the classification algorithm. 13An implementation for Matlab can be obtained from the corresponding author or at http://github. com/allefeld/prevalence-permutation/releases, and is also included in The Decoding Toolbox (TDT; Hebart et al., 2015) from version 3.8. 24 This definition is in line with the use of the term 'true accuracy' by Brodersen et al. (2013). It goes beyond the definition of the true accuracy of a particular trained classifier given by Pereira et al. (2009) as the expectation of a across test data sets (or equivalently, the accuracy determined on an infinite amount of test data). Like Brodersen and colleagues, we are not interested in characterizing the performance of a classifier trained on specific random training data, but in characterizing the subject as the source of both training and test data.14 B Simulated distributions of contrasts and accuracies We here explain the basic outline of the simulations underlying Sec. 2.5; for full details of the implementation, see App. C. The distributions arising in this simulation are illustrated in Fig. 1, so that this appendix also serves as an extended explanation of that figure. Single-subject data are generated according to the first-level model of Eq. 28. There are n = 25 trials for each condition in each of m = 6 runs. At the second level, true activation differences ∆βk for a sample of subjects are generated as independent normally-distributed numbers with mean ∆µ and standard deviation σ2 (Eq. 2). A sample consists of N = 17 subjects. We used a relatively large number of trials and runs in combination with trial-wise classification to obtain fine-grained accuracy estimation distributions well-suited for graphical display (red lines in Fig. 1e and f), and we chose a sample of 17 subjects to be able to closely approximate the standard significance level of α = 0.05 for subject-wise classification (Fig. 2d). The results remain qualitatively the same with other parameter choices and for classification of run-wise parameter estimates. Since multiplying all parameters by a constant factor changes only the scale but not the structure of the data, we choose σ1 = 1. The distributions of activation differences are illustrated in Figure 1a and b, for the case ∆µ = 1 and σ2 = 4 and for three subjects with true contrasts ∆β = −5, 1, and 8, respectively. Based on the generative model for single-subject data (Eq. 28), the amount of informa- tion the data contain about the experimental condition can be precisely calculated as a function of the true activation difference ∆β in that subject (Fig. 1c); it is a symmetric function which reaches its minimum value of 0 bit for ∆β = 0 and saturates towards 1 bit for large ∆β. This demonstrates that in the calculation of information the sign of the true activation difference is discarded, and that there never is negative information. The data are entered into a classification procedure for each subject separately. Single- trial data from 5 of the 6 runs are used to train a linear support vector machine (C-SVM with parameter C = 1; implementation by Chang and Lin, 2011, see http://www.csie. ntu.edu.tw/~cjlin/libsvm/) which is applied to trials from the left-out run, and this is repeated such that each run is once used for testing (6-fold cross-validation). The proportion of correctly classified trials gives the estimated classification accuracy a. Fig. 1d shows the true accuracy a (calculated as the mean of a across many simulations) in a single subject as a function of the true activation difference: a(∆β). The shape of 14Note that a classifier trained on a limited amount of training data will perform worse than optimally possible. This effect brings the accuracy closer to (but not below) chance level, i.e. what we consider as the true accuracy is generally smaller than the 'optimal true accuracy': aopt ≥ a ≥ a0 (cf. Wyman et al., 1990). 25 this function shows that classification accuracy is an information-like measure: It is a symmetric function of ∆β with a minimum at ∆β = 0, where it reaches the 'chance level' a0 = 50 %. Actual estimation distributions of accuracies a for the three simulated subjects are shown in Fig. 1e (red lines). Note that in contrast to the assumption of Eq. 4, these dis- tributions are not necessarily close to normal (or binomial; cf. Schreiber and Krekelberg, 2013; Noirhomme et al., 2014; Jamalabadi et al., 2016) and may not even be symmetric around the true value a (black bars); this becomes particularly apparent for ∆β = 1 (Fig. 1e, S2). However, the mean of the estimation distribution is by definition identical to the true accuracy, which makes a an unbiased estimator of a. The function a(∆β) (Fig. 1d) associates each true activation difference ∆β with a corre- sponding true accuracy a. The population distribution of true activation differences ∆βk (Fig. 1b, black line) therefore determines the population distribution of true accuracies ak (Fig. 1f, black line). For the case illustrated here, this distribution is far from normal (cf. Eq. 5); it is limited to a ≥ a0 but also has a spike at a = a0 due to the flat minimum of a(∆β), while a weaker secondary maximum at 56 % is due to the nonlinear increase of that function. Because a is just an estimate of a, the pronounced profile of this pop- ulation distribution ak gives rise to a smoother distribution of estimated accuracies across subjects ak (Fig. 1f, red line), for which the normality assumption of Eq. 6 might be accepted as a rough approximation. C Simulation implementation We here fill in technical details of the numerical calculations and simulations described in App. B and presented in Figs. 1–3. Single-subject data were simulated using the first-level GLM (Eq. 28). For simplicity, each trial lasted for one time unit (repetition time, 'TR'), each time unit belonged to a trial of one of the two conditions, and the hemodynamic response was assumed to be instantaneous, i.e. xAt and xBt were sequences of 0s and 1s. There were 2 · 25 · 6 = 300 time units in total (conditions · trials per run · runs). A given activation difference ∆β was implemented by setting βA = 0 and βB = ∆β. The error et consisted of independent (no temporal autocorrelation) normally-distributed pseudo-random numbers with standard deviation The mutual information (Cover and Thomas, 2012) between the data y and the trial type T ∈ {A, B} encoded in the regressors shown in Fig. 1c was calculated as 2 σ1, to ensure that Eq. 1 holds. √ mn I(y, T) = H(y) − H(yT), where the conditional entropy of y can be determined analytically as but the marginal entropy H(yT) = 1 2 log2(πe mn), H(y) =(cid:90) − fy(y) log2 fy(y) dy 26 (29) (30) (31) was computed by numerical integration based on the marginal density 1 2 1 2 mn mn (32) fy(y) = 2 (cid:17) + N(cid:16)y; 0, N(cid:16)y; ∆β, 2 (cid:17) . Here N (x; µ, σ2) denotes the density of the normal distribution. For the univariate classification results, data at the first level were simulated for 100 values of the true contrast ∆β from 0 to 86.6, at steps that were linearly increasing from 0.00883 to 1.74 to achieve better coverage close to ∆β = 0. For each value, 400,000 time series yt were randomly generated and the classification accuracy a was determined by cross-validation. Histograms of the resulting estimation distribution of accuracy for three different values of ∆β were used for Fig. 1e. For each ∆β, the true accuracy a was estimated by the mean, and the width of the estimation distribution ς1 by the standard deviation of generated values of a. To obtain smooth functional relationships used for Fig. 1d and as a basis for later calculations, both the mean and standard deviation were modeled as functions of the contrast, a(∆β) and ς1(∆β), using a cubic smoothing spline. The population density of the true accuracy fa(a) in Fig. 1f (black line) was computed by transforming the normal population density of the true contrast f∆β(∆β) (Eq. 2 and Fig. 1b, black line) through the modeled function a(∆β) using the standard rules of change of variables for densities, a−1− (a)(cid:12)(cid:12)(cid:12)(cid:12) f∆β(a−1− (a)), (33) d da fa(a) =(cid:12)(cid:12)(cid:12)(cid:12) a−1 + (a)(cid:12)(cid:12)(cid:12)(cid:12) f∆β(a−1 + (a)) +(cid:12)(cid:12)(cid:12)(cid:12) d da + (a) and a−1− (a) denote the positive and negative solution of a(∆β) = a. The where a−1 combined population-and-estimation distribution of accuracies in Fig. 1f (red line) was computed as an average of histograms of a for different values of ∆β, weighted by the population density of ∆β. The rejection probability of a second-level two-sided t-test on contrast estimates as a function of ∆µ and σ2 in Fig. 2b was computed as follows. Since the summary statistic (cid:99)∆βk is normally distributed with expectation ∆µ and combined variance σ2 1 + σ2 (Eq. 3), the t-statistic from N samples is noncentrally t-distributed with N − 1 degrees 2 of freedom and noncentrality parameter ∆µ(cid:112)N/σ2 c . The rejection probability is the probability mass of this distribution exceeding the critical value of the two-sided t-test with N − 1 degrees of freedom in either direction. In a first approximation, the rejection probability of a second-level one-sided t-test on accuracies vs a0 in Fig. 2a was computed in the same way, by assuming that the distribution of the summary statistic ak is exactly normal (Eq. 6 holds), with parameters ¯a and ς2 2. Note that we do not assume this to result from a normal distribution of true accuracies in the population (Eq. 5), but we treat Eq. 6 solely as an approximate description of the actual combined estimation and population variation (Fig. 1f, red line). Under this approximation, the t-statistic from N samples is noncentrally t-distributed c. The distribution moments of the summary statistic ak were numerically calculated from the population distribution of ∆βk (Eq. 2) combined with the modeled parameters of the with N − 1 degrees of freedom and noncentrality parameter (¯a − a0)(cid:112)N/ς2 c = σ2 c = ς2 1 + ς2 27 estimation distribution of a (Eq. 4), a(∆β) and ς1(∆β), by evaluating ¯a =(cid:90) a(∆β) N (∆β; ∆µ, σ2 c =(cid:90) (cid:16)ς2 1(∆β) + (a(∆β) − ¯a)2(cid:17) N (∆β; ∆µ, σ2 2 ) d∆β and ς2 2 ) d∆β. (34) The rejection probability is the probability mass of the resulting distribution exceeding the critical value of the one-sided t-test with N − 1 degrees of freedom. This approximation was finessed by comparing its results with those of a simulation of t-tests at 26 values of ∆µ (0 to 0.5, equidistant) combined with 101 values of σ2 (0 to 2, equidistant). At each parameter setting, N = 17 values of ∆β were sampled from the population distribution (Eq. 2), and for each a value of a was drawn from the corresponding estimation distribution. For this purpose, we did not simulate time series data and classification again, but re-used the result of the univariate classification simulation described above. For each ∆β, the closest of the 100 values realized in that simulation was determined, and a was drawn randomly from the pool of correspond- ing 400,000 simulation results. A one-sided t-test vs a0 was applied to the N drawn accuracies and it was recorded whether the null hypothesis could be rejected or not. This was repeated 5,000,000 times for each parameter setting. The resulting simulated rejection probabilities were compared to the approximated rejection probabilities and it was found that the approximation overestimates the rejection probability for larger values (largest discrepancy 0.7270 instead of 0.7169) and underestimates it for smaller values (largest discrepancy 0.1530 instead of 0.1604). The relation between simulated and approximated rejection probabilities was modeled using a 4th-order polynomial, which was then used to correct the approximation results for Fig. 2a. This resulted in the smallest occurring approximated rejection probability of 0.05 to be corrected to 0.055. The advantage of this combination of approximation and simulation is that the semi-analytic part provides a smooth function of the simulation parameters well- suited for graphical display, which is guaranteed to be precise by calibrating it using simulation. The rejection probability of FFX analysis applied to contrast estimates in Fig. 2c was computed as follows. FFX analysis also uses a t-statistic, but it compares the mean estimated activation difference 1 σ2 c = σ2 1 is estimated from the first- level GLM residuals with 2(mn − 1) degrees of freedom and the estimate is pooled across N subjects, such that the resulting statistic is t-distributed with 2(mn − 1)N degrees of freedom under the FFX null hypothesis. However, under the RFX model including random population variation, the variance of the statistic is actually larger 1. For arbitrary ∆µ, the distribution of the FFX t-statistic is a scaled by a factor σ2 noncentral t-distribution, with 2(mn− 1)N degrees of freedom, noncentrality parameter ∆µ(cid:112)N/σ2 c , and scaling factor σc/σ1. The rejection probability is the probability mass of this distribution exceeding the critical value of a two-sided t-test with 2(mn − 1)N degrees of freedom in either direction. The distribution of accuracies for classification across subjects depends on the ratio δ = 2 ∆µ/σc (the subject-level Mahalanobis distance). For 100 values of δ from 0 to 10, at steps that were linearly increasing from 0.00102 to 0.2 to achieve better coverage k=1(cid:99)∆βk not to an estimate of the combined variance 2, but of the estimation variance σ2 1 only. σ2 1 + σ2 N ∑N c /σ2 28 close to 0, second-level univariate data βAk and βBk were generated for N = 17 sub- jects. For 400,000 realizations, the resulting classification accuracy was determined by leave-one-subject-out cross-validation. The null hypothesis of no population difference between conditions A and B is realized at δ = 0; from the corresponding simulated null distribution the critical value was determined to be an accuracy of 67.6 %. This gives the best possible approximation of the 95th percentile of the discrete distribution of accuracies with 2N + 1 = 35 possible outcomes and leads to a test at a significance level of α = 0.051. The rejection probability for all simulated ratios was calculated as the fraction of simulated accuracies that reach or exceed this value, and the rejection probability was modeled as a function of δ using a cubic smoothing spline. The rejection probability of classification across subjects in Fig. 2d was then computed by applying this function to δ(∆µ, σ2) = 2 (35) ∆µ 1 + σ2 2 (cid:113)σ2 1 = 1. with σ2 For the multivariate classification results, simulations were performed in the same way as above, but with 10,000 simulated multi-voxel time series for each value of the true contrast, with dimensions p = 2 and 10. In contrast to the univariate case, no semi-analytic approximation was used, but the rejection probabilities in Fig. 3a were directly estimated from 1,000,000 simulated t-tests applied to accuracies resampled from this multivariate simulation, for each of 100 equidistant values of σ2 from 0 to 2. The rejection probabilities in Fig. 3b were directly estimated from 5,000,000 simulated t-tests applied to univariate classification accuracies. Again, the simulation of time series and classification was not repeated, but the result of the univariate classification simulation described above was re-used. N = 17 values of ∆β were sampled from the alternative population model of Eq. 10, and for each a value of a was drawn from the corresponding estimation distribution. This was performed for ∆β∗ = 1, 2, and 5, and for 101 equidistant values of γ from 0 to 1. Note Though the spatially extended prevalence null hypothesis (PN) at γ0 = 0 is equivalent to the spatially extended global null hypothesis (GN), the stated significance criterion for PN (cf. Eq. 21) p∗ N(mγ ≤ γ0) = p∗ N(m)]pN(mγ ≤ γ0) ≤ α N(m) + [1 − p∗ (36) at γ0 = 0, i.e. p∗ N(m) + [1 − p∗ N(m)]pN(m) ≤ α is not equivalent to the significance criterion for GN p∗ N(m) ≤ α. Rather, it is conservative due to the additional term [1 − p∗ N(m)]pN(m). (37) (38) 29 This discrepancy carries over into the algorithm Step 5b, where the criterion pN(m) ≤ α∗ (corresponding to Eq. 37) for the prevalence lower bound (Eq. 23) to be defined is not equivalent to the significance criterion for GN. N(m) ≤ α), the effective number of tests is moderately However, if GN can be rejected (p∗ large (~ 10), and a standard significance level α is used (e.g. 0.05), it holds pN(m) (cid:28) 1, so that the difference between the criteria becomes negligible. Is summary, Eq. 21 is not exact but rather formulates (yet another) upper bound, but the error introduced this way is small under normal circumstances. In any case, the criterion is conservative and therefore the test remains valid. References Allefeld, C., Haynes, J.-D., 2014. Searchlight-based multi-voxel pattern analysis of fMRI by cross-validated MANOVA. Neuroimage 89, 345–357. Bateson, G., 1972. Steps to an ecology of mind. University of Chicago Press, Chicago. Brodersen, K., Daunizeau, J., Mathys, C., Chumbley, J., Buhmann, J., Stephan, K., 2013. Variational Bayesian mixed-effects inference for classification studies. Neuroimage 76, 345–361. Brodersen, K., Mathys, C., Chumbley, J., Daunizeau, J., Ong, C., Buhmann, J., Stephan, K., 2012. Bayesian mixed-effects inference on classification performance in hierarchical data sets. The Journal of Machine Learning Research 13, 3133–3176. Chang, C.-C., Lin, C.-J., 2011. LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology 2, 27:1–27:27. Cichy, R.M., Chen, Y., Haynes, J.-D., 2011. Encoding the identity and location of objects in human LOC. Neuroimage 54, 2297–2307. Cover, T., Thomas, J., 2012. Elements of information theory. John Wiley & Sons, Hobo- ken. Cox, D., Savoy, R., 2003. Functional magnetic resonance imaging (fMRI) 'brain reading': Detecting and classifying distributed patterns of fMRI activity in human visual cortex. Neuroimage 19, 261–270. Davis, T., LaRocque, K., Mumford, J., Norman, K., Wagner, A., Poldrack, R., 2014. What do differences between multi-voxel and univariate analysis mean? How subject-, voxel-, and trial-level variance impact fMRI analysis. Neuroimage 97, 271–283. Efron, B., Tibshirani, R., 1994. An introduction to the bootstrap. Chapman & Hall, London. Ernst, M., 2004. Permutation methods: A basis for exact inference. Statistical Science 19, 676–685. Etzel, J., Braver, T., 2013. MVPA permutation schemes: Permutation testing in the land of cross-validation, in: International Workshop on Pattern Recognition in Neuroimaging. pp. 140–143. 30 Etzel, J.A., Zacks, J.M., Braver, T.S., 2013. Searchlight analysis: Promise, pitfalls, and potential. Neuroimage 78, 261–269. Fahrmeir, L., Kneib, T., Lang, S., Marx, B., 2013. Regression: Models, methods and applications. Springer, Berlin. Fisher, R., 1935. The design of experiments. Oliver & Boyd, Edinburgh. Fisher, R., 1925. Statistical methods for research workers. Oliver & Boyd, Edinburgh. Friston, K., Holmes, A., Price, C., Büchel, C., Worsley, K., 1999a. Multisubject fMRI studies and conjunction analyses. Neuroimage 10, 385–396. Friston, K., Holmes, A., Worsley, K., 1999b. How many subjects constitute a study? Neuroimage 10, 1–5. Friston, K., Holmes, A., Worsley, K., Poline, J.-P., Frith, C., Frackowiak, R., 1995. Statisti- cal parametric maps in functional imaging: A general linear approach. Human Brain Mapping 2, 189–210. Gaonkar, B., Davatzikos, C., 2013. Analytic estimation of statistical significance maps for support vector machine based multi-variate image analysis and classification. Neu- roimage 78, 270–283. Gaonkar, B., Shinohara, R.T., Davatzikos, C., ADNI, 2015. Interpreting support vector machine models for multivariate group wise analysis in neuroimaging. Medical Image Analysis 24, 190–204. Golland, P., Fischl, B., 2003. Permutation tests for classification: Towards statistical significance in image-based studies, in: Information Processing in Medical Imaging. Springer, Berlin, pp. 330–341. Görgen, K., Hebart, M., Allefeld, C., Haynes, J.-D., 2014. Detecting, avoiding & elim- inating confounds in MVPA / decoding studies. Poster presented at the 20th an- nual meeting of the Organization for Human Brain Mapping (OHBM). available at http://dx.doi.org/10.7490/f1000research.1111808.1. Haufe, S., Meinecke, F., Görgen, K., Dähne, S., Haynes, J.-D., Blankertz, B., Biessmann, F., 2014. On the interpretation of weight vectors of linear models in multivariate neu- roimaging. Neuroimage 87, 96–110. Haxby, J., 2012. Multivariate pattern analysis of fMRI: The early beginnings. Neuroim- age 62, 852–855. Haxby, J., Gobbini, M., Furey, M., Ishai, A., Schouten, J., Pietrini, P., 2001. Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science 293, 2425–2430. Haxby, J., Guntupalli, J., Connolly, A., Halchenko, Y., Conroy, B., Gobbini, M., Hanke, M., Ramadge, P., 2011. A common, high-dimensional model of the representational space in human ventral temporal cortex. Neuron 72, 404–416. Haynes, J.-D., Rees, G., 2006. Decoding mental states from brain activity in humans. Nature Reviews Neuroscience 7, 523–534. Haynes, J.-D., Rees, G., 2005a. Predicting the stream of consciousness from activity in human visual cortex. Current Biology 15, 1301–1307. 31 Haynes, J.-D., Rees, G., 2005b. Predicting the orientation of invisible stimuli from activity in human primary visual cortex. Nature Neuroscience 8, 686–691. Haynes, J.-D., Sakai, K., Rees, G., Gilbert, S., Frith, C., Passingham, R., 2007. Reading hidden intentions in the human brain. Current Biology 17, 323–328. Hebart, M., Görgen, K., Haynes, J.-D., 2015. The Decoding Toolbox (TDT): A versatile software package for multivariate analyses of functional imaging data. Frontiers in Neuroinformatics 8, 88. Holmes, A., Friston, K., 1998. Generalisability, random effects & population inference. Neuroimage 7, S754. Hoyos-Idrobo, A., Schwartz, Y., Varoquaux, G., Thirion, B., 2015. Improving sparse recovery on structured images with bagged clustering, in: International Workshop on Pattern Recognition in Neuroimaging. pp. 73–76. Huber, P., Ronchetti, E., 2009. Robust statistics, 2nd ed. John Wiley & Sons, Hoboken. Jamalabadi, H., Alizadeh, S., Schönauer, M., Leibold, C., Gais, S., 2016. Classification based hypothesis testing in neuroscience: Below-chance level classification rates and overlooked statistical properties of linear parametric classifiers. Human Brain Mapping 37, 1842–1855. Kowalczyk, A., 2007. Classification of anti-learnable biological and synthetic data, in: Knowledge Discovery in Databases: PKDD 2007. Springer, Berlin, pp. 176–187. Kriegeskorte, N., Bandettini, P., 2007. Analyzing for information, not activation, to exploit high-resolution fMRI. Neuroimage 38, 649–662. Kriegeskorte, N., Goebel, R., Bandettini, P., 2006. Information-based functional brain mapping. Proceedings of the National Academy of Sciences of the United States of America 103, 3863–3868. Lazar, N., Luna, B., Sweeney, J., Eddy, W., 2002. Combining brains: A survey of methods for statistical pooling of information. Neuroimage 16, 538–550. Lehmann, E., Romano, J., 2005. Testing statistical hypotheses, 3rd ed. Springer, Berlin. Mourao-Miranda, J., Bokde, A.L.W., Born, C., Hampel, H., Stetter, M., 2005. Classifying brain states and determining the discriminating activation patterns: Support Vector Machine on functional MRI data. Neuroimage 28, 980–995. Mudholkar, G., George, E., 1979. The logit method for combining probabilities, in: Symposium on Optimizing Methods in Statistics. Academic Press, New York, pp. 345–366. Nichols, T., Brett, M., Andersson, J., Wager, T., Poline, J.-B., 2005. Valid conjunction inference with the minimum statistic. Neuroimage 25, 653–660. Nichols, T., Holmes, A., 2001. Nonparametric permutation tests for functional neu- roimaging: A primer with examples. Human Brain Mapping 15, 1–25. Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., Kriegeskorte, N., 2014. A toolbox for representational similarity analysis. PLoS Computational Biology 10, e1003553. 32 Noirhomme, Q., Lesenfants, D., Gomez, F., Soddu, A., Schrouff, J., Garraux, G., Luxen, A., Phillips, C., Laureys, S., 2014. Biased binomial assessment of cross-validated esti- mation of classification accuracies illustrated in diagnosis predictions. Neuroimage: Clinical 4, 687–694. Norman, K., Polyn, S., Detre, G., Haxby, J., 2006. Beyond mind-reading: Multi-voxel pattern analysis of fMRI data. Trends in Cognitive Sciences 10, 424–430. Olivetti, E., Veeramachaneni, S., Nowakowska, E., 2012. Bayesian hypothesis testing for pattern discrimination in brain decoding. Pattern Recognition 45, 2075–2084. Penny, W., Holmes, A., 2007. Random effects analysis, in: Friston, K., others (Eds.), Statistical Parametric Mapping. Academic Press, London, pp. 156–165. Penny, W., Holmes, A., 2004. Random-effects analysis, in: Frackowiak, R., others (Eds.), Human Brain Function. Academic Press, London, pp. 843–850. Pereira, F., Botvinick, M., 2011. Information mapping with pattern classifiers: A com- parative study. Neuroimage 56, 476–496. Pereira, F., Mitchell, T., Botvinick, M., 2009. Machine learning classifiers and fMRI: A tutorial overview. Neuroimage 45, S199–S209. Price, C., Friston, K., 1997. Cognitive conjunction: A new approach to brain activation experiments. Neuroimage 5, 261–270. Rasch, D., Guiard, V., 2004. The robustness of parametric statistical methods. Psychology Science 46, 175–208. Rosenblatt, J., Vink, M., Benjamini, Y., 2014. Revisiting multi-subject random effects in fMRI: Advocating prevalence estimation. Neuroimage 84, 113–121. Rouder, J., Morey, R., Speckman, P., Pratte, M., 2007. Detecting chance: A solution to the null sensitivity problem in subliminal priming. Psychonomic Bulletin & Review 14, 597–605. Sabuncu, M.R., 2014. A universal and efficient method to compute maps from image- based prediction models, in: Medical Image Computing and Computer-Assisted Inter- vention. Springer, Berlin, pp. 353–360. Sabuncu, M.R., Van Leemput, K., 2012. The relevance voxel machine (RVoxM): A self- tuning Bayesian model for informative image-based prediction. IEEE Transactions on Medical Imaging 31, 2290–2306. Schreiber, K., Krekelberg, B., 2013. The statistical analysis of multi-voxel patterns in functional imaging. PLOS one 8, e69328. Searle, S., Casella, G., McCulloch, C., 1992. Variance components. John Wiley & Sons, Hoboken. Soch, J., Haynes, J.-D., Allefeld, C., 2016. How to avoid mismodelling in GLM- based fMRI data analysis: Cross-validated Bayesian model selection. Neuroimage. doi:10.1016/j.neuroimage.2016.07.047 Spiridon, M., Kanwisher, N., 2002. How distributed is visual category information in human occipito-temporal cortex? An fMRI study. Neuron 35, 1157–1165. 33 Stelzer, J., Chen, Y., Turner, R., 2013. Statistical inference and multiple testing correction in classification-based multi-voxel pattern analysis (MVPA): Random permutations and cluster size control. Neuroimage 65, 69–82. Stephan, K., Penny, W., Daunizeau, J., Moran, R., Friston, K., 2009. Bayesian model selection for group studies. Neuroimage 46, 1004–1017. Stouffer, S., Lumsdaine, A., Lumsdaine, M., Williams Jr., R., Smith, M., Janis, I., Star, S., Cottrell Jr., L., 1949. The American soldier: Combat and its aftermath. Princeton University Press, Princeton. Thirion, B., Flandin, G., Pinel, P., Roche, A., Ciuciu, P., Poline, J.-B., 2006. Dealing with the shortcomings of spatial normalization: Multi-subject parcellation of fMRI datasets. Human Brain Mapping 27, 678–693. Tippett, L., 1931. The methods of statistics. Williams & Norgate, London. Todd, M., Nystrom, L., Cohen, J., 2013. Confounds in multivariate pattern analysis: Theory and rule representation case study. Neuroimage 77, 157–165. Wang, Z., Childress, A.R., Wang, J., Detre, J.A., 2007. Support vector machine learning- based fMRI data group analysis. Neuroimage 36, 1139–1151. Worsley, K., Friston, K., 2000. A test for a conjunction. Statistics & Probability Letters 47, 135–140. Wyman, F., Young, D., Turner, D., 1990. A comparison of asymptotic error rate expan- sions for the sample linear discriminant function. Pattern Recognition 23, 775–783. 34
1508.04624
2
1508
2017-12-19T11:19:46
Fast {\large\it Cl-}type inhibitory neuron with delayed feedback has non-markov output statistics
[ "q-bio.NC" ]
For a class of fast {\it Cl-}type inhibitory spiking neuron models with delayed feedback fed with a Poisson stochastic process of excitatory impulses, it is proven that the stream of output interspike intervals cannot be presented as a Markov process of any order.
q-bio.NC
q-bio
Fast Cl-type inhibitory neuron with delayed feedback has non-Markov output statistics Alexander K.Vidybida Bogolyubov Institute for Theoretical Physics Metrologichna str., 14-B, Kyiv 03680, Ukraine June 9, 2021 Abstract For a class of fast Cl-type inhibitory spiking neuron models with de- layed feedback stimulated with a Poisson stochastic process of excitatory impulses, it is proven that the stream of output interspike intervals cannot be presented as a Markov process of any order. Keywords: Poisson stochastic process; spiking neuron; probability den- sity function; delayed feedback; fast Cl-type inhibition; non-Markov stochas- tic process 1 Introduction Spiking statistics of various neuronal models under a random stimulation has been studied in the framework of two main approaches. The first one is named in [1] as "Gaussian", because it describes random stimulation by means of Gaussian noise, see e.g. [2]. This approach has developed into the well-known diffusion approximation methodology, see [3]. The second approach is named in [1] as "quantal", because it takes into account the discrete nature of the influence any input impulse may have on its target neuron. The wide area of research and applications known as spiking neural networks, see [4] for a review, could be considered as utilizing the quantal approach. For a recent review of mathematically rigorous results regarding neuronal We study here spiking statistics in the both approaches see [5]. mathematically the framework of quantal approach, spiking statistics of inhibitory neuron model belonging to a class of models (see Sec. 2, below) with fast Cl-type inhibitory delayed feedback (see Fig. 1, below). rigorously, in 1.1 Biological inspiration Neurons, which send inhibitory impulses onto their own body or dendrites are known in real nervous systems, see [20, 21, 22, 23]. 1 The chief in inhibitory neurotransmitter the nervous system is Gamma-aminobutyric acid (GABA). The GABA can acti- vate several types of receptors, the main of which are GABAa and GABAb. If GABAa receptors are activated, the excitable membrane becomes permeable for Cl− ions. If a neuron is partially excited, that is its membrane is depolar- ized, the Cl− current cancels this depolarization since the Cl− reversal potential is close/equal to the resting potential. For the same reason, the Cl− current through open GABAa channels does not appear, if the membrane is at its resting potential. Another case is with GABAb receptors activation. This causes K + ions permeability. The outward K + current is able to hyperpolarize membrane even below its resting potential. The remarkable difference between GABAa and GABAb mediated inhibition is rather different kinetics of the corresponding Cl− and K + currents. Namely, according to [24], the Cl− current rise time is 1 - 5 ms, and the decay time constant is about 10 - 25 ms. The K + current rise time is 10 - 120 ms, and the decay time constant is about 200 - 1600 ms. The K + current can be even slower, see [22, 25, 26]. Inspired by this contrast in the speed of Cl− and K + transients, we idealize the Cl− current kinetics as having infinitesimally short rise time and infinitely fast decay, both can be achieved with infinitely large Cl− conductance at the moment of receiving inhibitory impulse. This kind of the Cl− current kinetics does ensure the perfect reset of the membrane voltage to the resting state at the moment when inhibitory impulse arrives. Within the limited experimental data set available for inhibitory autapses, see [24], a single impulse delivered through a single synapse ensures only partial reset. In this point, our state- ment of the problem diverges from the current data. At the same time, in the artificial neuromorphic systems, see [6, 7], the complete reset may well be re- alized. Considering a partial reset would inappropriately increase the paper's dimensions. Below, it is expected that a neuron sends back to itself its output impulses through a delayed feedback line, which ends with a GABAa autapse, see Prop1- Prop3, in Sec. 2.2. This construction is stimulated with a Poisson stream of excitatory input impulses. For this configuration it has been proven in the pre- vious paper [8] for the case of Poisson input stream and for a concrete neuronal model -- the inhibitory binding neuron1 with threshold 2 -- , that statistics of its interspike intervals (ISIs) is essentially non-Markov2. In paper [13], it 1Detailed description of the binding neuron model can be found in [9]. See also https://en.wikipedia.org/wiki/Binding neuron. 2Sometimes, a concept of a point stochastic Markov process is confused with a process whose consecutive realizations are uncorrelated. Actually, the latter one is a renewal process, which is a specific case of Markov process. As regards a general Markov process, its realizations can well be correlated, see e.g. [10]. In this paper, we do not study correlations (which itself is an interesting topic, which could be addressed separately), but prove that the output statistics does not have the Markov property defined, e.g., in [11, Ch.2 §6]. Some interesting remarks about non-markoviannes can be found in [12]. 2 has been proven for the Poisson input and for a class of excitatory neuronal models that the presence of delayed feedback makes their activity non-Markov. In this paper, we use the approach developed in [13] in order to refine and extend methods of [8] making them applicable to any inhibitory neuron with fast Cl-type inhibition satisfying a number of simple and natural conditions, see Cond0-Cond4, below. The stimulation is assumed to be a Poisson stochastic process. Under those conditions, we prove rigorously that ISI statistics of a neuron with delayed fast Cl-type inhibitory feedback stimulated with a Poisson point stochastic process of input impulses cannot be represented as a Markov chain of any finite order. Finally, it should be mentioned that our consideration is valid also for artificial hardware neurons, see [27, 28], and abstract neurons used in mathematical studies, provided that Cond0-Cond4 and Prop1-Prop3, below, are satisfied. 2 Definitions and assumptions 2.1 Neuron without feedback We assume that a neuron satisfies the following conditions: • Cond0: Neuron is deterministic: trains from the same neuron. Identical stimuli elicit identical spike • Cond1: Neuron is stimulated with input Poisson stationary stream of excitatory impulses. • Cond2: Neuron may fire a spike only at a moment when it receives an input impulse. • Cond3: Just after firing, neuron appears in its resting state. • Cond4: The output interspike interval (ISI) distribution can be character- ized with a probability density function (pdf) p0(t), which is continuous with positive: and bounded: p0(0) = 0, t > 0 ⇒ p0(t) > 0, p0(t) < ∞. sup t>0 (1) (2) (3) By t we denote the ISI's length. Also, we impose on the function p0(t) the following condition: t < 0 ⇒ p0(t) = 0 in order to have it defined for all real numbers. These conditions are, with some modifications, similar to those assumed in [13] for a class of excitatory neurons. The modifications are as follows: 3 Cond3 -- we assume that after firing a neuron appears in its resting state with all excitation canceled, while in [13] it is a standard state, which not necessarily is the resting one. Cond4 -- the requirement of continuity of p0(t) is added as compared to [13]. This addition has a pure mathematical nature and seems to be valid for any "good" neuronal model (without feedback). The subsequent proof of non- markoviannes relies on it. The Cond3, above, limits the set of models as compared to [13]. Namely, it claims that the standard state of [13, Cond3] must be exactly the resting state of neuron. This requirement is imposed due to the specifics of Cl-type fast inhibition. For our approach, it is important that after receiving inhibitory impulse, the neuron appears in exactly the same state as immediately after firing. And the state after receiving Cl-type inhibitory impulse can be only the resting state, see Sec. 1.1, above. It seems that these conditions are satisfied for many threshold-type neuronal models known in the literature, see [14, 15, 16, 17] and citations therein. But this still has to be proven by calculating corresponding p0(t). At least, all the five conditions are satisfied for the binding neuron model and for the basic leaky integrate-and-fire (LIF) model, both for Poisson stimulation. See [18, 19], where p0(t) is calculated exactly for each model, respectively. 2.2 Feedback line action We expect that the feedback line satisfies Prop1, Prop2 of [13], which are re- produced below for completeness. The Prop3 of [13] should be modified for the Cl-type fast inhibition as shown below: • Prop1: The time delay in the line is ∆ > 0. • Prop2: The line is able to convey no more than one impulse. • Prop3: The impulse conveyed to the neuronal input is the fast Cl-type inhibitory impulse. This means that after receiving such an impulse, the neuron appears in its resting state. This exhausts the action of the in- hibitory impulse in a sense that it has no influence on further neuronal states created by next excitatory impulses. It as well does not affect neu- ron being in its resting state. The Prop1 expects that the delay is always the same and each impulse, entered the line is delivered to its output and effects the neuron. Thus, we do not consider here cases when the transmission is unreliable, or the delay time is not constant. The validity of the Prop2 depends on relation between the conduction ve- locity, recovery time and the line's length. Also Prop2 seems plausible if the firing frequency is low. The Prop3 just characterizes a neuronal model as inhibitory with GABAa- type autapse. Its validity depends on the fact that the Cl− reversal potential 4 is identical to the resting potential. In same cases this is fulfilled, see [24]. It is also expected that a single AP delivered by a feedback line is potent enough for canceling any excitation present. Taking into account that observed single GABAa IPSP peak value rare exceeds 6 mV, this may require the delay line sprouting into several autaptic endings. The important for us consequence of the Prop2, above, is that at any moment of time the feedback line is either empty, or conveys a single impulse. If it does convey an impulse, then its state can be described with a stochastic variable s, s ∈]0; ∆], which we call further "time to live", see Fig. 1. The variable s denotes the exact time required by the impulse to reach the output end of the line, which is the neuron's input for inhibitory impulses, and to leave the line with the consequences described in the Prop3, above. Here, ∆ denotes the delay duration in the feedback line. Notice, that at the beginning of any ISI, the line is never empty. Figure 1: Neuron with delayed feedback. As neuron in the figure, we consider any neuronal model, which satisfies the set of conditions Cond0 - Cond4, above. 3 Results Our purpose here is to prove the following Theorem3: Theorem 1 Let a neuronal model satisfies conditions Cond0-Cond4, above. Suppose that the model is extended by introducing a delayed fast Cl-type in- hibitory feedback line, which satisfies the Prop1-Prop3, above. Then, in the stationary regime, the output stream of ISIs of the neuron cannot be presented as a Markov chain of any finite order. 3.1 Proof outline Let pinh(tn+1 tn, . . . , t0)dtn+1 denote the conditional probability to get the duration of (n + 2)-nd ISI in the interval [tn+1; tn+1 + dtn+1[ provided that previous n + 1 ISIs have duration tn, . . . , t0, respectively. From the definition in [11, Ch.2 §6], one can obtain the necessary condition pinh(tn+1 tn, . . . , t1, t0) = pinh(tn+1 tn, . . . , t1), (4) 3A similar theorem for the excitatory feedback line has been proven in [13]. 5 ✲inputstream(Poisson)neuron✲✛✲0r∆rsr✲delayedfeedback{z}toutputstream(non-Markov) -- ISIduration required for the stochastic process {tj} to be nth order Markov chain. Notice, that (4) must be satisfied for any values of the variables ti, i = 0, . . . , n + 1. purpose we calculate exact expression for pinh(tn+1 tn, . . . , t0) as We intend to prove that the relation (4) does not hold for any n. For this pinh(tn+1 tn, . . . , t0) = pinh(tn+1, tn, . . . , t0) pinh(tn, . . . , t0) (5) from which it will be clearly seen that the t0-dependence in pinh(tn+1 tn, . . . , t0) cannot be eliminated whatever large the n is. (5), expression pinh(tn, . . . , t1) denotes the joint probability density function of ISIs duration of neuron with the fast Cl-type inhibitory delayed feedback. Let us introduce the conditional joint probability density pinh(tn+1, . . . , t0 s), which denotes the conditional probability density to get n + 2 consecutive ISIs {tn+1, . . . , t0} provided that at the beginning of the first ISI (t0) the time to live of impulse in the feedback line is equal to s. This conditional probability can be used to calculate required joint pdfs as follows In the Eq. pinh(tn+1, . . . , t0) = ∆(cid:90)0 pinh(tn+1, . . . , t0 s)f inh(s) ds, (6) where f inh(s) is the stationary pdf which describes distribution of times to live at the beginning of any ISI in the stationary regime. In what follows we analyze the structure of functions f inh(s) and pinh(tn+1, . . . , t0 s). It appears that f inh(s) has a singular component aδ(s − ∆) with a > 0, and pinh(tn+1, . . . , t0 s) has jump discontinuities at definite hyper- planes in the (n + 3)-dimensional space of its variables (tn+1, . . . , t0, s). After integration in (6), some of those discontinuities will survive in the (n + 2)- dimensional space of variables (tn+1, . . . , t0), and exactly one of those survived has its position depending on t0. The t0-dependent jump discontinuity will as well survive in the pinh(tn+1 tn, . . . , t0) for any n, provided that tn, . . . , t0 satisfy the following condition: ti < ∆, n(cid:88)i=0 (7) where ∆ > 0 is the full delay time in the feedback line. Taking into account that the equation in the necessary condition (4) must hold for any set of tn+1, . . . , t0, we conclude that (4) cannot be satisfied for any n. 3.2 The proof 3.2.1 Structure of functions pinh(tn+1, . . . , t0 s) Specifics of the feedback line action together with condition (7) results in a very simple of structure 6 pinh(tn+1, . . . , t0 s) at different parts of the integration domain in (6). Those parts are defined as follows: Dk = {s k−1(cid:88)i=0 ti < s ≤ k(cid:88)i=0 ti}, k = 0, . . . , n, Dn+1 = {s n(cid:88)i=0 ti < s ≤ ∆} . As regards the structure itself, the following representation can be derived sim- ilarly as it was done in [13]: pinh(tn+1, . . . , t0 s) = pinh(tn+1, . . . , tk+1 ∆)× × pinh(cid:32)tk s − k−1(cid:88)i=0 ti(cid:33) k−1(cid:89)i=0 p0(ti), s ∈ Dk, k = 0, . . . , n, (8) pinh(tn+1, . . . , t0 s) = = pinh(cid:32)tn+1 s − n(cid:88)i=0 ti(cid:33) n(cid:89)i=0 p0(ti), s ∈ Dn+1. (9) pinh(tn+1, . . . , tk+1 ∆) = = pinh(cid:32)tn+1 ∆ − n(cid:88)i=k+1 ti(cid:33) n(cid:89)i=k+1 p0(ti). (10) Here pinh(t s) denotes the conditional pdf to get ISI of duration t if at its beginning, time to live of impulse in the feedback line is s. Representation of pinh(tn+1, . . . , t0 s) by means of p0(t) and pinh(t s) found here is similar to that found in [13] for the excitatory case. But the structure of function pinh(t s), used in that representation, is different. 3.2.2 Structure of function pinh(t s) Expect that at the beginning of an ISI, there is an impulse in the feedback line with time to live s. Then the probability that this ISI will have its duration t < s does not depend on the feedback line presence. Therefore, t < s ⇒ pinh(t s) = p0(t). (11) In the opposite situation, receiving of an ISI duration greater than s hap- pens if (i) the neuron is not firing during interval ]0; s[ and (ii) the neuron starts at its resting state (Prop3, above) at the moment s and fires at t > s. Real- izations of events (i) and (ii) depend on disjoint segments of the input Poisson 7 stream (Cond1, above). Therefore, (i) and (ii) are statistically independent. The probability of (i) is as follows: P0(s) = 1 − s(cid:90)0 p0(t)dt. The probability of (ii) is p0(t − s). This gives t > s ⇒ pinh(t s) = P0(s)p0(t − s). It can be concluded from (11) and (13) that pinh(t s) = p0(s) lim t↑s and lim t↓s pinh(t s) = 0. (12) (13) Now, taking into account (1) and (2) from Cond4, above, we conclude that the function pinh(t s) considered as a function of two variables (t, s), t ≥ 0, s ∈ ]0; ∆] has a jump discontinuity along the straight line t = s. The magnitude of this jump is p0(s), and it is strictly positive for positive t. Concrete values of pinh(t s) along the line t = s does not matter and can be chosen arbitrarily. Finally, for pinh(t s) we have4 pinh(t s) = χ(s − t)p0(t) + P0(s)p0(t − s), (14) where χ(s) is the Heaviside step function. 3.2.3 Structure of probability density function f inh(s) Everywhere in this paper we expect that all pdfs pinh(tn+1, . . . , t0) have achieved their stationary form, and we analyze the stationary regime. But any station- ary regime arises from some initial distribution. In principle, different initial distributions may result in different final stationary distributions. As it can be concluded from (8)-(10), the only quantity, which might de- pend on initial conditions in the right-hand side of representation (6) is the pdf f inh(s). Before the stationary regime is achieved, f inh(s) is changed after each firing: fn+1(s) = ∆(cid:90)0 P(s s(cid:48))fn(s(cid:48))ds(cid:48), (15) where the transition function P(s s(cid:48)) gives the probability density to find at the beginning of an ISI an impulse in the line with time to live s provided that 4Compare this with [29, Eq. (11)], where pinh(t s) is calculated exactly for the binding neuron model stimulated with Poisson stream. 8 at the beginning of the previous ISI, there was an impulse with time to live s(cid:48). In the stationary regime, the pdf f (s) must satisfy the following equation f inh(s) = ∆(cid:90)0 P(s s(cid:48))f inh(s(cid:48))ds(cid:48), (16) Now, the question of existence and uniqueness of the stationary regime might be resolved by analyzing Eqs. (15) and (16) for convergence and uniqueness. This is expected to do in another paper. In this paper we assume that sequence n (s)} of pdfs generated by Eq. (15) converges to some pdf for any initial {f inh f0(s), and admit that there might be different limiting distributions for different f0(s). The exact expression for P(s s(cid:48)) is found in [13, Eqs.(11)-(13)]. It appears that the structure of f inh(s), which follows from (16) is exactly the same as it has been found in [13] for the excitatory case. This structure is as follows5 f inh(s) = g(s) + aδ(s − ∆), (17) where a > 0 and g(s) is bounded continuous function vanishing out of interval ]0; ∆[. 3.2.4 Form of pinh(tn+1, . . . , t0) and pinh(tn, . . . , t0) after integration in (6) Let D = Dk. At D, representations (8) and (10) are valid. Also at D, n(cid:83)k=0 f inh(s) reduces to g(s). Therefore, pinh(tn+1, . . . , t0 s)f inh(s) ds = (cid:90)D = pinh(cid:32)tn+1 ∆ − n(cid:88)i=k+1 n(cid:88)k=0 × n(cid:89)i = 0 p0(ti)(cid:90)Dk i (cid:54)= k ti(cid:33)× pinhtk s − k−1(cid:88)j=0 ti(cid:33) . pinh(cid:32)tn+1 ∆ − n(cid:88)i=k+1 tj g(s)ds. (18) The first factor (with fixed k, 0 ≤ k ≤ n) in the r.h.s. of Eq. (18) is as follows: 5Compare this with [30, Eqs. (14)-(16)], where f (s) is calculated exactly for the binding neuron model. 9 Due to Eq. (14), this factor does have a jump discontinuity along the hyperplane ti = ∆ in the space of variables (t0, . . . , tn+1). Notice, that the position of this hyperplane does not depend on t0 for any k ∈ {0, . . . , n}. The second factor in the r.h.s. of Eq. (18) is as follows: is continuous. p0(ti), and it n(cid:81)i = 0 i (cid:54)= k The third factor in the r.h.s. of Eq. (18) can be transformed as follows: n+1(cid:80)i=k+1 (cid:90)Dk = n(cid:89)i=0 n(cid:81)i=0 10 tj j=0 = pinhtk s − k−1(cid:88)j=0 tj g(s)ds = k(cid:80) tj g(s)ds = pinhtk s − k−1(cid:88)j=0 (cid:90)k−1(cid:80) tj ds = pinh(tk s)gs + k−1(cid:88)j=0 tk(cid:90)0 P0(s)p0(tk − s)gs + tk(cid:90)0 = tj j=0 = k−1(cid:88)j=0 tj ds. (19) The last expression is continuous with respect to variables (t0, . . . , tn+1). There- fore, one can conclude that expression (18) does not have a jump discontinuity, which position depends on t0. Consider now the remaining part of integral in (6). With (9) taken into account one has: pinh(tn+1, . . . , t0 s)f inh(s) ds = (cid:90)Dn+1 p0(ti) (cid:90)Dn+1 pinh(cid:32)tn+1 s − n(cid:88)i=0 ti(cid:33) f inh(s)ds. (20) Here, the first factor, p0(ti) is continuous and strictly positive for positive ti. The second factor can be transformed as follows: (cid:90)Dn+1 pinh(cid:32)tn+1 s − n(cid:88)i=0 ∆(cid:90)n(cid:80) ti(cid:33) f inh(s)ds = pinh(cid:32)tn+1 s − n(cid:88)i=0 = ti ti(cid:33) f inh(s)ds = i=0 = ∆− n(cid:80) (cid:90)0 i=0 ti pinh(tn+1 s)f inh(cid:32)s + ti(cid:33) ds. n(cid:88)i=0 (21) Now, let us use representations (14) and (17) in order to figure out which kind of discontinuities does the expression (21) have. Due to (14) and (17), expression (21) will have four terms. The first one we get by choosing the first term both in (14) and (17): ti ∆− n(cid:80) (cid:90)0 i=0 A11 = χ(s − tn+1)p0(tn+1)g(cid:32)s + This term is either equal to zero, if tn+1 > ∆ − n(cid:80)i=0 into a continuous function of variables (t0, . . . , tn+1). Moreover, ti(cid:33) ds. n(cid:88)i=0 ti, or otherwise transforms tn+1↑∆− n(cid:80) lim ti i=0 A11(tn+1) = 0. The second one we get by choosing the second term in (14) and the first term in (17): ∆− n(cid:80) (cid:90)0 i=0 A21 = ti P0(s)p0(tn+1 − s)g(cid:32)s + ti(cid:33) ds. n(cid:88)i=0 This (t0, . . . , tn+1). is as well a continuous function of variables The third one we get by choosing the first term in (14) and the second term 11 in (17): A12 = ∆− n(cid:80) (cid:90)0 i=0 = a ti χ(s − tn+1)p0(tn+1)δ(cid:32) n(cid:88)i=0 ti + s − ∆(cid:33) ds = = aχ(cid:32)∆ − n+1(cid:88)i=0 ti(cid:33) p0(tn+1). (22) (23) ti = ∆ . n+1(cid:88)i=0 The forth one we get by choosing the second term in (14) and the second This term has a jump discontinuity along the hyperplane term in (17): A22 = = a ∆− n(cid:80) (cid:90)0 i=0 ti P0(s)p0(tn+1 − s)δ(cid:32) n(cid:88)i=0 ti + s − ∆(cid:33) ds = ti(cid:33) p0(cid:32)n+1(cid:88)i=0 = P0(cid:32)∆ − n(cid:88)i=0 continuous function of ti − ∆(cid:33) . variables This (t0, . . . , tn+1). is as well a After taking into account the above reasoning, we conclude that the required joint probability density has the following form pinh(tn+1, . . . , t0) = pw(tn+1, . . . , t0) + aχ(cid:32)∆ − n+1(cid:88)i=0 ti(cid:33) n+1(cid:89)j=0 p0(tj). (24) where function pw(tn+1, . . . , t0) does not have a jump discontinuity depending on t0, and the second term in (24) does have such a discontinuity along the hyperplane (23). Form of pinh(tn, . . . , t0) after integration If (7) is satisfied, then we have similarly to (8), (9) pinh(tn, . . . , t0 s) = pinh(tn, . . . , tk+1 ∆)× × pinh(cid:32)tk s − k−1(cid:88)i=0 ti(cid:33) k−1(cid:89)i=0 p0(ti), s ∈ Dk, k = 0, . . . , n − 1, 12 pinh(tn, . . . , t0 s) = pinh(cid:32)tn s − n−1(cid:88)i=0 ti(cid:33) n−1(cid:89)i=0 p0(ti), s ∈ Dn. Again due to (7), and in analogy with (10) we have instead of the last two equations the following one: pinh(tn, . . . , t0 s) = pinh(cid:32)tk s − k−1(cid:88)i=0 ti(cid:33) n(cid:89)i = 0 i (cid:54)= k p0(ti), s ∈ Dk, k = 0, . . . , n. (25) It is clear that expression similar to (9) turns here into the following pinh(tn, . . . , t0 s) = n(cid:89)i=0 p0(ti), s ∈ Dn+1. (26) Now, due to (25), (26) we have pinh(tn, . . . , t0) = pinh(tn, . . . , t0 s)f inh(s)ds = ∆(cid:90)0 n(cid:89)i = 0 n(cid:88)k=0 i (cid:54)= k = p0(ti)(cid:90)Dk pinh(cid:32)tk s − k−1(cid:88)i=0 n(cid:89)i=0 + ti(cid:33) g(s)ds+ p0(ti) (cid:90)Dn+1 f inh(s)ds. (27) From calculations similar to those made in Eq. (19) it can be concluded that pinh(tn, . . . , t0) is continuous at the domain defined by (7). 3.2.5 t0-dependence cannot be eliminated in pinh(tn+1 tn, . . . , t0) Now, with representations (24) for pinh(tn+1, . . . , t0) and (27) for pinh(tn, . . . , t0) we can pose a question about the form of pinh(tn+1 tn, . . . , t0). The latter can be found as defined in (5). First of all notice that due to (27) and Cond4, pinh(tn, . . . , t0) is strictly positive for positive ISIs. This allows us to use it as denominator in the definition (5). Second, it can be further concluded from (27) and Cond4, that pinh(tn, . . . , t0) is bounded. The latter together with continuity of pinh(tn, . . . , t0) means that any discontinuity of jump type present in the pinh(tn+1, . . . , t0) appears as well in the pinh(tn+1 tn, . . . , t0). It follows from the above and from Eq. (24) that the conditional pdf pinh(tn+1 tn, . . . , t0) 13 can be represented in the following form: pinh(tn+1 tn, . . . , t0) = pw(tn+1 tn, . . . , t0)+ + Z(tn+1, . . . , t0)χ(cid:32)∆ − n+1(cid:88)i=0 ti(cid:33) , (28) where pw(tn+1 tn, . . . , t0) does not have any jump type discontinuity which position depends on t0, and Z(tn+1, . . . , t0) is strictly positive function: Z(tn+1, . . . , t0) = p0(ti) a n+1(cid:81)i=0 p(tn, . . . , t0) . Thus the representation (28) proves that for any n, conditional pdf pinh(tn+1 tn, . . . , t0) does depend on t0 (the second term in (28)) and this dependence cannot be eliminated. (cid:3) See also Appendix, below, where the above general reasoning is illustrated for the LIF neuronal model with threshold 2 (that is two input impulses applied in a short succession are able to trigger, see (31), below). 4 Discussion and Conclusions The question as to what extent the stream of neuronal output impulses can be modeled as Poisson stream has been discussed in neuroscience, see [31]. The experimentally observed presence of memory in the ISIs output of real neurons has been reported many times, see [32, 33, 34, 35, 36]. Also several theoretical models of how the memory could appear are offered, see [15, 37, 38, 39, 40, 41, 43]. In this paper we use the quantal approach, as it is defined in [1] in order to prove that the Markov property is broken in the ISI output stream of a neuronal model belonging to a defined class of models, equipped with delayed fast Cl- type inhibitory feedback, which is stimulated with a Poisson stochastic process of input excitatory impulses. Several previous results obtained in the quantal approach are used in this paper, see [8, 13, 18, 29, 30]. In all these papers (as in many other computational neuroscience works) it is assumed that input impulse has zero duration and is modeled as the Dirac δ- function in time. On the other hand, the data of papers [24, 21], which inspired this study, suggest that observed there width of inhibitory impulse is comparable with or even longer than the delay time for autaptic selfinhibition. In the case of that extended impulses it is not clear which figure should be considered as the delay. In this connection, it should be taken into account that any extended in time impulse can be represented as a sum of short impulses precisely positioned at different moments of time. For each short impulse component the delay value, either axonal, or synaptic, or both is precisely defined. Of course, we cannot expect here that each short impulse component performs a complete reset of 14 membrane potential. Therefore, additional analysis is required, which is out of scope of this paper. On the other hand, for neurocybernetical artificial devices, see, e.g. [6, 7], situation with short impulse and long delay seems to be more natural due to specifics of digital devices. The first results this paper is based on are obtained for the binding neuron (BN) model. Namely, in [18] under Poisson stimulation the output ISI pdf and mean ISI are obtained for the BN with threshold (Th) 2, and the mean ISI for Th = 3. In [8, 29, 30] a BN model with Th = 2 and with delayed feedback, either excitatory or inhibitory, stimulated with Poisson stream is considered. For this case, the ISI pdf is found and also it is proven that the output ISI stream is non-Markov. In [13], any neuronal model from a defined class is considered. A delayed feedback is assumed excitatory, and stimulation is Poissonian. For this case, it is proven that the output stream is non-Markov. In this paper, a class of neuronal models with delayed Cl-type inhibitory feedback is considered. The memory property in output ISI streams is often discussed in terms of correlation coefficient (CC), e.g. [42]. Unfortunately, the expressions obtained in this paper cannot be used for conclusions made in terms of CC. This is because all expressions, including p(t1, t0) are obtained under restriction (7), whereas in order to calculate CC one needs to know p(t1, t0) for all t0 > 0, t1 > 0. Nevertheless, expressions derived in this paper under restriction (7) allows one to show that the Markov property is broken in the output ISI due to delayed feedback. Another reason for neuronal activity to be non-Markov in a network is offered in [43]. In further work, it is expected to extend obtained here exact expressions to the full range of ISI values and to compare our findings with those obtained in terms of CC. This includes also a quantitative estimation of how much the statistics is non-Markov and to what extent it might be approximated by a Markow/renewal process. Also, a general renewal stochastic process can be considered as a stimulus instead of Poisson one. The latter can be achieved if to find adequate expression for Eq. (13), which in its current form is valid for Poisson stimulation only. Acknowledgements This paper was partially supported by the Program "Struc- ture and Dynamics of Statistical and Quantum-Field Systems" of the National Academy of Science of Ukraine, Project PK No 0117U000240. Some calculations in the Appendix are made with the help of free Computer Al- gebra System Maxima, see http://maxima.sourceforge.net/. A Appendix Here we give a simple example of the proven property. Namely, we consider pinh(t2 t1, t0) and show that t0-dependence cannot be eliminated for a LIF model, stimulated with Poisson stream. 15 For the two values of n = 1, 2, Eq. (6) due to (7), (14), (18)-(20), (27), turns into the following two equations: pinh(t1, t0) = p0(t1) P0(s)p0(t0 − s)g(s)ds+ t0(cid:90)0 + p0(t0) t1(cid:90)0 P0(s)p0(t1 − s)g(s + t0)ds+ + p0(t1)p0(t0) f (s + t0 + t1)ds, (29) ∆−t0−t1 (cid:90)0 pinh(t2, t1, t0) = = pinh(t2 ∆ − t1)p0(t1) P0(s)p0(t0 − s)g(s)ds+ t0(cid:90)0 + pinh(t2 ∆)p0(t0) t1(cid:90)0 + p0(t1)p0(t0) P0(s)p0(t1 − s)g(s + t0)ds+ ∆−t0−t1 (cid:90)0 pinh(t2 s)f (s + t0 + t1)ds, (30) Now, let the neuronal model be the basic LIF model characterized with the firing threshold V0, input impulse height h and relaxation time τ . Assume that 0 < h < V0 < 2h. (31) Assume also that the neuron is stimulated with a Poisson stream of intensity λ. For this case, it is proven in ([19, Eqs. (14),(21)] that p0(t) = λ2te−λt, provided t ∈ [0; T2], (32) where T2 = τ log h V0 − h . Assume, for simplicity, that ∆ < T2. This allows to obtain exact expressions for P0(s) and pinh(t s): P0(s) = e−λs(λs + 1), (33) pinh(t s) = λ2e−λt(tχ(s − t) + χ(t − s)(t − s)(λs + 1)). 16 If we put χ(0) = 0.5, then χ(−x) = 1 − χ(x) and the last expression can be transformed as follows pinh(t s) = λ2e−λt(t + χ(t − s)s(λ(t − s) − 1)). (34) Under the assumptions of this Appendix, it appears that the kernel of inte- gral equation (16), above, is exactly the same as for the binding neuron model with excitatory feedback. The latter case has been studied in [30, Eqs. (14)- (16)], where the unique solution to Eq. (16) is found. The unknown in general case quantities from Eq. (17), above, under assumptions of this Appendix can be taken from [30]: g(s) = a = aλ 2 (cid:16)1 − e−2λ(∆−s)(cid:17) , 4e2λ∆ (2λ∆ + 3)e2λ∆ + 1 . (35) (36) (37) After substituting (32)-(36) into (29) one obtains pinh(t1, t0) = λ4e−λ(t0+t1)t0t1 6((2λ∆ + 3)e2λ∆ + 1) 0 + t2 ×(cid:16)2λe2λ∆(λ(t2 × 1) + 3(2∆ − t0 − t1))+ + 3(e2λ(t1+t0) + 6e2λ∆ + 1)(cid:17) . Notice, that pinh(t1, t0) is strictly positive for strictly positive t1, t0. This allows one to use it safely as denominator in the definition of conditional probability (5), above. Also, as it may be observed from (37), the pinh(t1, t0) is continuous and bounded. This means, that pinh(t2 t1, t0) as it is defined in (5), will preserve any discontinuity which may appear in the pinh(t2, t1, t0), which is the numerator in (5) for n = 1. Consider now Eq.(30) for pinh(t2, t1, t0). After partial simplifications, it turns into the following: pinh(t2, t1, t0) = = pinh(t2 ∆ − t1) + pinh(t2 ∆) λ4t0t1e−λ(t0+t1) 6((2λ∆ + 3)e2λ∆ + 1) + λ2t2e−λt2p0(t1)p0(t0)× λ4t0t1e−λ(t0+t1) 6((2λ∆ + 3)e2λ∆ + 1) × × ×(cid:0)3(cid:0)1 − e2λt0(cid:1) + 2λt0e2λ∆(λt0 + 3)(cid:1) + (a) ×(cid:0)3e2λt0(cid:0)1 − e2λt1(cid:1) + 2λt1e2λ∆(λt1 + 3)(cid:1) + (b) g(s + t0 + t1)ds+ (c) ∆−t0−t1 (cid:90)t2 × χ(∆ − t0 − t1 − t2) 17 + λ2e−λt2p0(t1)p0(t0)× × min(t2,∆−t0−t1) (cid:90)0 (t2 − s)(λs + 1)g(s + t0 + t1)ds+ (d) + a p0(t1)p0(t0)pinh(t2 ∆ − t0 − t1). (38) The summands (a) and (b) in Eq. (38) correspond to the first and second term of the right-hand side in Eq.(30), respectivly. The remaining three ones correspond to the the third term of the right-hand side in Eq.(30). It is easily seen that both (a) and (b) are continuous with respect to t0. The term (d) is as well continuous in t0, because function min(x, y) is continuous on x and y. The term (c) is as well continuous because lim ∆−t0−t1−t2→0 χ(∆ − t0 − t1 − t2) ∆−t0−t1 (cid:90)t2 g(s + t0 + t1)ds = 0. Consider the final term in Eq. (38). For the sake of clarity, we omit the factor a p0(t1)p0(t0) having in mind that it is continuous and strictly positive for t0 > 0, t1 > 0. The remaining expression is as follows λ2e−λt2(t2 − χ(t2 + t1 + t0 − ∆)(∆ − t0 − t1)(λ(∆ − t0 − t1 − t2) + 1)). (39) This expression, if considered as a function of t0, t1, t2 has a step-like disconti- nuity along the hyperplane t2 + t1 + t0 = ∆. (40) Indeed, if t2 + t1 + t0 < ∆, then (39) turns into λ2e−λt2 t2. Otherwise, if t2 + t1 + t0 > ∆, then (39) turns into λ2e−λt2 (t2 − (∆ − t0 − t1)(λ(∆ − t0 − t1 − t2) + 1)). The difference between the two expressions is as follows λ2e−λt2 (∆ − t0 − t1)(λ(∆ − t0 − t1 − t2) + 1). (41) This difference vanishes along the hyperplane t2 + t1 + t0 = ∆ + 1 λ only (due to (7), we do not consider the case t0 + t1 = ∆). Comparing the last equation with (40), we see that the jump (41) is strictly positive along the hyperplane (40). The same is valid for the last term in (38). Taking into account that the other four terms in (38) are continuous in t0, and what is said after Eq. 18 Figure 2: Different values of pinh(t2 t1, t0) for different t0. Here ∆ = 7 ms, λ = 0.3 ms−1, t1 = 3 ms both for (a) and (b). The t0 = 3.5 ms for (a) and t0 = 1.5 ms for (b). The curves are calculated based on Eqs. (34)-(38). (37), we conclude that pinh(t2 t1, t0) has a nonzero jump along the hyperplane (40). For a fixed t1, t2 and infinitesimally small  > 0 consider two different 0 = ∆ − t1 − t2 ± . It is clear from the above that when t0 value values of t0: t± 0 , the pinh(t2 t1, t0) gets finite obtains infinitesimally small change from t+ change due to the jump (41), which means that t0-dependence in pinh(t2 t1, t0) is indeed present (at least due to the discovered jump discontinuity) and cannot be eliminated. This is illustrated in Fig. 2. 0 to t− References [1] R. B. Stein, Some models of neuronal variability, Biophysical Journal 7(1) (1967) 37 -- 68. [2] H. L. Bryant and J. P. Segundo, Spike initiation by transmembrane current: a white-noise analysis, The Journal of Physiology 260(2) (1976) 279 -- 314. [3] R. M. Capocelli and L. M. Ricciardi, Diffusion approximation and first passage time problem for a model neuron, Kybernetik 8(6) (1971) 214 -- 223. [4] S. Ghosh-Dastidar and H. Adeli, Spiking neural networks, International Journal of Neural Systems, 19(4) (2009) 295 -- 308. [5] L. Sacerdote and M. T. Giraudo, Stochastic integrate and fire models: A re- view on mathematical methods and their applications, Stochastic Biomath- ematical Models, Lecture Notes in Mathematics 2058., eds. M. Bachar, J. J. Batzel and S. Ditlevsen (Springer-Verlag, 2013), pp. 99 -- 148. [6] R. Sarpeshkar, Neuromorphic and biomorphic engineering systems (New York: McGraw-Hill Yearbook of Science & Technology 2009 , McGraw-Hill, 2009), pp. 250 -- 252. 19 0 0.02 0.04 0.06 0.08 0.1 0.12 0 1 2 3 4 5 6 7 8 9 10p(t2 t1, t0) / ms-1t2 / msa 0 0.02 0.04 0.06 0.08 0.1 0.12 0 1 2 3 4 5 6 7 8 9 10p(t2 t1, t0) / ms-1t2 / msb [7] D. Bruderle, M.A. Petrovici, B. Vogginger, B. et al., A comprehensive work- flow for general-purpose neural modeling with highly configurable neuro- morphic hardware systems Biol Cybern, 104(4) (2011) 263 -- 296. [8] K. G. Kravchuk and A. K. Vidybida, Firing statistics of inhibitory neuron with delayed feedback. II: Non-markovian behavior, BioSystems 112(3) (2013) 233 -- 248. [9] A. K. Vidybida, Binding neuron, Encyclopedia of information science and technology, ed. M. Khosrow-Pour (IGI Global, 2014), pp. 1123 -- 1134. [10] B. Lindner, A brief introduction to some simple stochastic processes Stochastic Methods in Neuroscience, ed. C. Laing and G.J. Lord (Oxford University Press, 2010), pp. 1 -- 28. [11] J. L. Doob, Stochastic processes (Wiley, 1953). [12] N. G. van Kampen, Remarks on Non-Markov Processes, Brazilian Journal of Physics 28(2) (1998) 90 -- 96. [13] A. K. Vidybida, Activity of excitatory neuron with delayed feedback stim- ulated with Poisson stream is non-Markov, Journal of Statistical Physics 160 (2015) 1507 -- 1518. [14] A. N. Burkitt, A review of the integrate-and-fire neuron model: I. homo- geneous synaptic input, Biological Cybernetics 95(1) (2006) 1 -- 19. [15] M. J. Chacron, K. Pakdaman and A. Longtin, Interspike interval correla- tions, memory, adaptation, and refractoriness in a leaky integrate-and-fire model with threshold fatigue, Neural Computation 15(2) (2003) 253 -- 278. [16] R. Jolivet, T. J. Lewis and W. Gerstner, 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) (2004) 959 -- 976. [17] R. Jolivet, A. Rauch, H. Luscher and W. Gerstner, Predicting spike timing of neocortical pyramidal neurons by simple threshold models, J Comput Neurosci 21(1) (2006) 35 -- 49. [18] O. Vidybida, Output stream of a binding neuron, Ukrainian Mathematical Journal 59(12) (2007) 1819 -- 1839. [19] O. K. Vidybida, Output stream of leaky integrate and fire neuron, Reports of the National Academy of Science of Ukraine 2014(12) (2014) 18 -- 23. [20] J. M. Bekkers, Neurophysiology: Are autapses prodigal synapses?, Current Biology 8(2) (1998) R52 -- R55. 20 [21] A. Bacci, J. R. Huguenard and D. A. Prince, Functional autaptic neuro- transmission in fast-spiking interneurons: A novel form of feedback inhibi- tion in the neocortex, The Journal of Neuroscience 23(3) (2003) 859 -- 866. [22] A. Bacci, J. R. Huguenard and D. A. Prince, Long-lasting self-inhibition of neocortical interneurons mediated by endocannabinoids, Nature 431 (2004) 312 -- 316. [23] T. C. Smith and C. E. Jahr, Self-inhibition of olfactory bulb neurons, Na- ture Neuroscience 5(8) (2002) 760 -- 766. [24] L. S. Benardo, Separate activation of fast and slow inhibitory postsynaptic potentials in rat neocortex in vitro, Journal of Physiology 476.2 (1994) 203 -- 215. [25] J. F. Storm, Four voltage-dependent potassium currents in adult hippocam- pal pyramidal cells, Biophys.J. 53 (1988) p. 148a. [26] J. Storm, Potassium currents in hippocampal pyramidal cells, Progress in Brain Research 83 (1990) 161 -- 187. [27] J. L. Rossell´o, V. Canals, A. Morro and A. Oliver, Hardware implementa- tion of stochastic spiking neural networks, International Journal of Neural Systems 22(4) (2012) p. 1250014. [28] R. Wang, G. Cohen, K. M. Stiefel, T. J. Hamilton, J. Tapson and A. van Schaik, An FPGA implementation of a polychronous spiking neural net- work with delay adaptation, Frontiers in Neuroscience 7(14) (2013). [29] A. K. Vidybida and K. G. Kravchuk, Firing statistics of inhibitory neuron with delayed feedback. I. output ISI probability density, BioSystems 112(3) (2013) 224 -- 232. [30] A. K. Vidybida, Output stream of binding neuron with delayed feedback, 14th International Congress of Cybernetics and Systems of WOSC, Wro- claw, Poland, September 9-12, 2008 , eds. J. J´ozefczyk, W. Thomas and M. Turowska (Oficyna Wydawnicza Politechniki Wroclawskiej, 2008), pp. 292 -- 302. [31] B. B. Averbeck, Poisson or not Poisson: Differences in spike train statistics between parietal cortical areas, Neuron 62(3) (2009) 310 -- 311. [32] S. B. Lowen and M. C. Teich, Auditory-nerve action potentials form a nonrenewal point process over short as well as long time scales, The Journal of the Acoustical Society of America 92(2) (1992) 803 -- 806. [33] S. Shinomoto, Y. Sakai and S. Funahashi, The Ornstein-Uhlenbeck process does not reproduce spiking statistics of cortical neurons, Neural Computa- tion 11 (1999) 935 -- 951. 21 [34] R. Ratnam and M. E. Nelson, Nonrenewal statistics of electrosensory af- ferent spike trains: implications for the detection of weak sensory signals, The Journal of Neuroscience 20(17) (2000) 6672 -- 6683. [35] M. P. Nawrot, C. Boucsein, V. Rodriguez-Molina, A. Aertsen, S. Grun and S. Rotter, Serial interval statistics of spontaneous activity in cortical neurons in vivo and in vitro, Neurocomputing 70 (2007) 1717 -- 1722. [36] G. Maimon and J. A. Assad, Beyond Poisson: Increased spike-time regu- larity across primate parietal cortex, Neuron 62(3) (2009) 426 -- 440. [37] J. P. Rospars and P. L´ansk´y, Stochastic model neuron without resetting of dendritic potential: application to the olfactory system, Biol. Cybern. 69(4) (1993) 283 -- 294. [38] P. L´ansk´y and R. Rodriguez, Two-compartment stochastic model of a neu- ron, Physica D: Nonlinear Phenomena 132(1 -- 2) (1999) 267 -- 286. [39] E. Benedetto and L. Sacerdote, On dependency properties of the ISIs gen- erated by a two-compartmental neuronal model, Biol Cybern 107(1) (2013) 95 -- 106. [40] R. E. Kass, V. Ventura and E. N. Brown, Statistical issues in the analysis of neuronal data, Journal of Neurophysiology 94(1) (2005) 8 -- 25. [41] O. Avila-Akerberg and M. J. Chacron, Nonrenewal spike train statistics: causes and functional consequences on neural coding, Experimental Brain Research 210(3-4) (2011) 353 -- 371. [42] F. Farkhooi, E. Muller and M.P. Nawrot, Adaptation reduces variability of the neuronal population code, Phys Rev E 83(5) (2011) : 050905. [43] B. Cessac, A discrete time neural network model with spiking neurons: II: Dynamics with noise, Journal of Mathematical Biology 62(6) (2011) 863 -- 900. 22
1807.02741
1
1807
2018-07-08T02:43:04
Algebraic signatures of convex and non-convex codes
[ "q-bio.NC", "cs.DM", "math.CO" ]
A convex code is a binary code generated by the pattern of intersections of a collection of open convex sets in some Euclidean space. Convex codes are relevant to neuroscience as they arise from the activity of neurons that have convex receptive fields. In this paper, we use algebraic methods to determine if a code is convex. Specifically, we use the neural ideal of a code, which is a generalization of the Stanley-Reisner ideal. Using the neural ideal together with its standard generating set, the canonical form, we provide algebraic signatures of certain families of codes that are non-convex. We connect these signatures to the precise conditions on the arrangement of sets that prevent the codes from being convex. Finally, we also provide algebraic signatures for some families of codes that are convex, including the class of intersection-complete codes. These results allow us to detect convexity and non-convexity in a variety of situations, and point to some interesting open questions.
q-bio.NC
q-bio
Algebraic signatures of convex and non-convex codes Carina Curtoa, Elizabeth Grossb, Jack Jeffriesc,d, Katherine Morrison1,e,a, Zvi Rosenf,g,a, Anne Shiuh, and Nora Youngsi,j a Department of Mathematics, The Pennsylvania State University, University Park, PA 16802 b Department of Mathematics, San Jos´e State University, San Jos´e, CA 95192 c Department of Mathematics, University of Utah, Salt Lake City, UT 84112 d Department of Mathematics, University of Michigan, Ann Arbor, MI 48109 e School of Mathematical Sciences, University of Northern Colorado, Greeley, CO 80639 f Department of Mathematics, University of California, Berkeley, Berkeley, CA 94720 g Departments of Mathematics & Biology, University of Pennsylvania, Philadelphia, PA 19104 h Department of Mathematics, Texas A&M University, College Station, TX 77843 i Department of Mathematics, Harvey Mudd College, Claremont, CA 91711 j Department of Mathematics and Statistics, Colby College, Waterville, Maine 04901 Abstract A convex code is a binary code generated by the pattern of intersections of a collection of open convex sets in some Euclidean space. Convex codes are relevant to neuroscience as they arise from the activity of neurons that have convex receptive fields. In this paper, we use algebraic methods to determine if a code is convex. Specifically, we use the neural ideal of a code, which is a generalization of the Stanley-Reisner ideal. Using the neural ideal together with its standard generating set, the canonical form, we provide algebraic signatures of certain families of codes that are non-convex. We connect these signatures to the precise conditions on the arrangement of sets that prevent the codes from being convex. Finally, we also provide algebraic signatures for some families of codes that are convex, including the class of intersection-complete codes. These results allow us to detect convexity and non-convexity in a variety of situations, and point to some interesting open questions. Keywords: neural coding, convex codes, neural ideal, local obstructions, simplicial complexes, links Contents 1 Introduction 1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Summary of main results 1.3 Examples illustrating main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Detecting local obstructions 2.1 Local obstructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Algebraic detection of local obstructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Detecting convex codes 4 Examples 5 Supplemental Text 1 Introduction 1 2 6 8 9 9 10 12 16 20 A convex code is a binary code generated by the pattern of intersections of a collection of open convex sets in some Euclidean space (see Section 1.1 for a precise definition and example). Convex codes 1corresponding author: Katherine Morrison, University of Northern Colorado, [email protected] 1 have been experimentally observed in sensory cortices [8] and hippocampus [11], where they arise from convex receptive fields; this connection has previously been described in detail in [2, 3, 5]. Given their relevance to neuroscience, it is valuable to further understand the intrinsic structure of convex codes. In particular, how can we detect if a neural code is convex? We have previously found combinatorial constraints that must be satisfied by any code that is convex [3]. In this work, we further address the question of convexity via an algebraic object known as the neural ideal JC, first introduced in [5], which is a generalization of the well-studied Stanley- Reisner ideal. We first present conditions, which we refer to as algebraic signatures, on JC and its standard generating set the canonical form CF(JC) that detect that a code is not convex. We also connect these signatures to precise conditions on the arrangement of sets that prevent a code from being convex. Finally, we also provide algebraic signatures of certain combinatorial families of convex codes, including intersection-complete codes, first introduced in [1]. In Section 1.1, we provide some background on the algebra of neural codes, convexity of codes, receptive field relationships, and local obstructions to convexity. Next, Section 1.2 highlights the main results of the paper. Specifically, Theorem 1.7 provides algebraic signatures of two classes of local obstructions; codes satisfying these signatures are thus guaranteed to be non-convex. Theorem 1.9 gives an algebraic signature for the class of intersection-complete codes, which have been proven to be convex. Section 1.3 illustrates these main results through a series of example codes satisfying these algebraic signatures. The remainder of the paper is organized as follows: Section 2.1 formalizes the notion of local obstruction, and Section 2.2 provides further results on detecting local obstructions algebraically, in- cluding the proof of Theorem 1.7. Section 3 focuses on algebraic signatures guaranteeing convexity, and includes the proof of Theorem 1.9. Finally, Section 4 collects all the algebraic signatures presented in this paper and provides additional examples of codes satisfying these signatures. 1.1 Background In this paper, we develop algebraic tools for analyzing neural codes, which are collections of binary patterns. A binary pattern on n neurons is a string of 0s and 1s of length n, with a 1 for each active neuron and a 0 denoting silence. We can also view a binary pattern as the subset of active neurons σ ⊆ [n] def= {1, . . . , n}, so that i ∈ σ precisely when there is a 1 in the ith entry of the binary pattern; thus, we will consider 0/1 strings of length n and subsets of [n] interchangeably. For example, 1011 and 0100 are also denoted {1, 3, 4} and {2}, respectively. A neural code on n neurons, C ⊆ 2[n], is a collection of binary patterns. Such a code is also referred to as a combinatorial code in the neuroscience literature [4]. The elements of a code are called codewords. For convenience, we will always assume a neural code C includes the all-zeros codeword, 00··· 0 ∈ C; the presence or absence of the all-zeros codeword has no effect on the code's convexity (see Definition 1.5, below), which is the main focus of this paper. Algebra of neural codes In order to represent a neural code algebraically, it is useful to consider binary patterns of length n as 2 , where F2 is the finite field of two elements: 0 and 1. Polynomials f ∈ F2[x1, . . . , xn] can elements of Fn be evaluated on a binary pattern of length n by evaluating each indeterminate xi at the 0/1 value of the ith neuron. For example, if f = x1x3(1 − x2) ∈ F2[x1, . . . , x4], then f (1011) = 1 and f (1100) = 0. It is natural to then consider the ideal IC def= {f ∈ F2[x1, . . . , xn] f (c) = 0, ∀ c ∈ C} of polynomials that vanish on a neural code C. However, this ideal contains extraneous Boolean relations B = (cid:104)xi(1− xi)(cid:105) that do not capture any information specific to the code. Thus we turn instead to the neural ideal JC, first introduced in [5], which captures all the information in IC that is specific to the 2 code, thus omitting the Boolean relations. More precisely, the neural ideal can be defined in terms of characteristic functions of non-codewords: JC def= (cid:104)χv v ∈ Fn 2 \ C(cid:105) where χv is the characteristic function χv (cid:89) def= xi {ivi=1} (cid:89) (1 − xj). {jvj =0} (1) Note that the variety of both IC and JC is precisely the code C [5]. polynomials f ∈ F2[x1, . . . , xn] that can be written in the form The characteristic functions used to define the neural ideal are examples of pseudo-monomials, (cid:89) j∈τ f = xσ (1 − xj), def= (cid:81) where xσ i∈σ xi and σ, τ ⊂ [n] with σ ∩ τ = ∅. Pseudo-monomials in JC come in two types2: (cid:81) • Type 1: xσ, for σ (cid:54)= ∅, and • Type 2: xσ For any ideal J ⊆ F2[x1, . . . , xn], a pseudo-monomial f ∈ J is called minimal if there does not exist another pseudo-monomial g ∈ J with deg(g) < deg(f ) such that f = hg for some h ∈ F2[x1, . . . , xn]. If J is an ideal generated by a set of pseudo-monomials, the canonical form of J is the set of all minimal pseudo-monomials of J: i∈τ (1 − xi), for σ, τ (cid:54)= ∅, with σ ∩ τ = ∅. CF(J) def= {f ∈ J f is a minimal pseudo-monomial}. For any neural code C, the neural ideal JC is generated by pseudo-monomials, and thus has a canonical form CF(JC).3 We denote the Type 1 and Type 2 pseudo-monomials of CF(JC) by CF1(JC) and CF2(JC), respectively, so that: CF(JC) = CF1(JC) ∪ CF2(JC). Example 1.1. Consider the code C = {0000, 0100, 0010, 0001, 1100, 1010, 0110, 1011}. The neural ideal JC is given by JC = (cid:104)x1(1 − x2)(1 − x3)(1 − x4), x1x4(1 − x2)(1 − x3), x2x4(1 − x1)(1 − x3), x3x4(1 − x1)(1 − x2), x2x3x4(1 − x1), x1x2x4(1 − x3), x1x2x3(1 − x4), x1x2x3x4(cid:105), which has canonical form CF(JC) = CF1(JC) ∪ CF2(JC), where CF1(JC) = {x1x2x3, x2x4} and CF2(JC) = {x1(1 − x2)(1 − x3), x1x4(1 − x3), x3x4(1 − x1)}. Note that CF(JC) is a generating set for JC, as every pseudo-monomial of JC is a multiple of an element in CF(JC). Furthermore, CF1(JC) generates the ideal of monomials in JC, which is precisely the Stanley-Reisner ideal of the associated simplicial complex ∆(C), where ∆(C) def= {σ ⊆ [n] σ ⊆ c for some c ∈ C} is the smallest abstract simplicial complex on [n] that contains all elements of C [5]. In particular, if C is a simplicial complex, then JC is precisely the Stanley-Reisner ideal of C. Note that the facets 2There is a third type (see [5]), but this is eliminated by our convention that 00··· 0 ∈ C. 3Furthermore, every ideal generated by pseudo-monomials is actually the neural ideal of some neural code [9]. 3 of ∆(C), which are maximal elements of the simplicial complex under inclusion, correspond to the maximal codewords of C. The canonical form of a code C can be computed algorithmically; for example, [5, Section 4.5] provides an algorithm using primary decompositions of pseudo-monomial ideals. A more efficient algorithm has since been proposed in [12], with software publicly available [14]. Supplemental Text S1 gives full details for computing the canonical form of an example code by hand; for information on using software to compute CF(JC), see [12]. The code of a cover Let X be a topological space. A collection of non-empty open sets U = {U1, . . . , Un}, where each Ui ⊂ X, is called an open cover. Given an open cover U, the code of the cover is the neural code C(U) def= {σ ⊆ [n] Uσ \ (cid:91) Uj (cid:54)= ∅}, j∈[n]\σ def= (cid:84) i∈σ Ui. We say that a code C is realized by U if C = C(U). Observe that X is subdivided where Uσ into regions defined by intersections of the open sets in U. Each codeword in C(U) then corresponds to a non-empty intersection that is not covered by other sets in U (see Example 1.2). By convention, i∈[n] Ui (cid:40) X. We will It is important to note that C(U) is not the same as the nerve N (U ) of the cover, which consists of i∈∅ Ui equals X, so that ∅ ∈ C(U) if and only if(cid:83) the empty intersection U∅ =(cid:84) assume(cid:83) i∈[n] Ui (cid:40) X, so that 00··· 0 ∈ C (i.e., ∅ ∈ C), in agreement with our convention. all non-empty intersections, regardless of whether the intersection region is covered by other sets: N (U) def= {σ ⊆ [n] Uσ (cid:54)= ∅}. In fact, N (U) = ∆(C(U)), the simplicial complex of the code [5]. The nerve of any cover U such that C = C(U) can thus be recovered directly from the code as ∆(C), without reference to a specific cover. The code C(U), however, contains additional information about U that is not captured by the nerve alone (see [5, Section 2.3.2]). Example 1.2. Consider the configuration of sets U = {U1, . . . , U4} shown in Figure 1. The code of the cover is C = C(U) = {0000, 1000, 0100, 0010, 1100, 1001, 0110, 0101, 1101}. Note that from C alone, we can detect that any realization must have U4 ⊆ U1 ∪ U2, since every codeword with a 1 in the 4th position has a 1 in the 1st or 2nd position as well. However, this containment information is not available from the nerve N (U). Figure 1: Code of the cover U = {U1, . . . , U4}. RF relationships and the neural ideal. Any realization of a code C by an open cover will satisfy relationships among the Ui that are intrinsic to the code itself. Because of the neuroscience motivation, where the Ui model receptive fields, we call these receptive field relationships [5]. 4 000001010010110111001001100001100100 Definition 1.3. For σ, τ ⊆ [n] with σ (cid:54)= ∅ and σ ∩ τ = ∅, we say that (σ, τ ) is a receptive field (RF) relationship of a code C if Uσ ⊆(cid:91) i∈τ and Uσ ∩ Ui (cid:54)= ∅ for all i ∈ τ, Ui for any U = {U1, . . . , Un} where C = C(U). RF(C) denotes the collection of RF relationships of C. It is important to note that the receptive field relationships RF(C) are strictly a function of the code itself and do not depend on any particular realization of C as C(U). Specifically, RF relationships correspond to pseudo-monomials in JC as shown in Table 1, and thus are detectable algebraically without reference to a specific cover U [5]. Relation type Type 1 Type 2 Pseudo-monomial xσ ∈ JC xσ i∈τ (1 − xi) ∈ JC (cid:81) ⇔ ⇔ RF condition Uσ = ∅ Uσ ⊆(cid:83) i∈τ Ui (cid:81) (cid:81) Uσ ⊆(cid:83) Table 1: Types of pseudo-monomials in JC and the corresponding conditions on receptive fields. Note i∈τ (1 − xi) ∈ JC is not sufficient to guarantee that the presence of a Type 2 pseudo-monomial xσ that (σ, τ ) is actually an RF relationship. Such a pseudo-monomial ensures the covering relationship i∈τ Ui, but to guarantee that (σ, τ ) ∈ RF(C) for τ (cid:54)= ∅ we must also have xσxi /∈ JC for all i ∈ τ . The RF relationships of the form (σ,∅) capture when Uσ = ∅, and thus σ /∈ N (U), yielding a complete description of N (U) = ∆(C). In contrast, the RF relationships (σ, τ ) for τ (cid:54)= ∅ capture when an intersection is covered so that σ /∈ C despite σ ∈ ∆(C), thus measuring how C deviates from its simplicial complex. destroying the containment Uσ ⊆ (cid:83) A RF relationship (σ, τ ) is called minimal if no neuron can be removed from σ or τ without i∈τ Ui. The following useful fact is a direct consequence of [5, Theorem 4.3], which allows us to interpret the elements of CF(JC) as minimal RF relationships. Lemma 1.4. The pseudo-monomial xσ relationship of C. i∈τ (1 − xi) ∈ CF(JC) if and only if (σ, τ ) is a minimal RF Thus, the canonical form gives a compact description of JC that captures all the minimal intersection and containment relations that must exist among sets that give rise to the code. Convex codes When the open cover U is contained in Rd for some d, the sets Ui may (for some codes) be chosen to all be convex. If this is possible, we say that the code is convex : Definition 1.5. Let C be a neural code on n neurons. If there exists an open cover U = {U1, . . . , Un} such that C = C(U) and every Ui is a convex subset of Rd for a fixed d, then we say that C is convex. Note that the code in Example 1.2 is convex since it can be realized via the convex sets shown in Figure 1. In contrast, the code from Example 1.1 is not convex, as the following example shows. Example 1.6. Recall the code C = {0000, 0100, 0010, 0001, 1100, 1010, 0110, 1011} from Exam- ple 1.1. Neuron 1 always co-fires with neuron 2 or neuron 3 since a 1 only occurs in the first entry when it is accompanied by a 1 in the second or third entry. This forces the RF relationship U1 ⊆ U2 ∪ U3 to hold in any realization of the code. But neurons 1, 2, and 3 never co-fire, so U1 ∩ U2 ∩ U3 = ∅. Thus U1 is the disjoint union of non-empty open sets U1 ∩ U2 and U1 ∩ U3, and so U1 is disconnected. Since any convex set is connected, we conclude that U1 cannot be convex, and thus C is not convex. 5 This topological mismatch between the underlying set U1 and its cover by U1∩ U2 and U1∩ U3 is an example of a local obstruction [3, 6]; we define local obstructions precisely in Section 2.1. Notice that this local obstruction is immediately identifiable from the canonical form CF(JC) seen in Example 1.1: the RF relationship U1 ⊆ U2 ∪ U3 is detectable from x1(1 − x2)(1 − x3) ∈ CF2(JC) and the RF relationship U1 ∩ U2 ∩ U3 = ∅ is captured by x1x2x3 ∈ CF1(JC). 1.2 Summary of main results Detecting non-convex codes. Example 1.6 shows that some local obstructions to convexity can be detected algebraically from the neural ideal of a code. In particular, any code satisfying the algebraic signature xσ(1 − xi)(1 − xj) ∈ CF2(JC) and xσxixj ∈ CF1(JC) is guaranteed to be non-convex. This is because Uσ is forced to be disconnected since it is the disjoint union of the nonempty sets Uσ ∩ Ui and Uσ ∩ Uj. Theorem 1.7 gives two additional algebraic signatures of local obstructions that force a code to be non-convex. The first signature captures more generally when the nerve of a cover of Uσ is disconnected, thus forcing Uσ to be disconnected and non-convex. The second signature captures cases when the nerve is a hollow simplex, thus forcing Uσ to contain a hole. In other words, these signatures capture when the nerve of the cover of Uσ has a nontrivial 0th homology group and nontrivial top homology group, respectively. It remains an open question to identify algebraic signatures that can detect when a relevant nerve has an intermediate homology group that is nontrivial. Theorem 1.7. Let C be a code with neural ideal JC and canonical form CF(JC) = CF1(JC) ∪ CF2(JC), and let GC(σ, τ ) be the simple graph on vertex set τ with edge set {(ij) ∈ τ × τ xσxixj /∈ JC}. The following algebraic signatures imply that C is not convex. Property of C (cid:81) (cid:81) Algebraic signature of JC ∃ xσ ∃ xσ (i) (ii) i∈τ (1 − xi) ∈ CF2(JC) s.t. GC(σ, τ ) is disconnected ⇒ non-convex ⇒ non-convex i∈τ (1 − xi) ∈ CF2(JC) s.t. xσxτ ∈ CF1(JC) Table 2: Algebraic signatures of non-convex codes. It is important to note that although signature (i) in Table 2 requires the construction of a graph based on the absence of pseudo-monomials from all of JC, this condition can actually be checked in a straightforward manner from CF1(JC) alone (see Lemma 2.5 in Section 1.7). The signatures of local obstructions in Theorem 1.7 can thus be directly detected from the canonical form of the code. The proof of Theorem 1.7 is given in Section 2.2. Our previous work has given an alternative method of identifying the full set of local obstructions; however, the recasting of those local obstructions in terms of RF relationships is less well understood. A characterization of the full set of local obstructions of a code is given in Theorem 1.3 of [3]. In general, however, the absence of local obstructions does not guarantee that C is convex [10]. Thus, it is essential to have other methods of identifying convexity. Detecting convex codes. Currently the only known method for proving a code is convex is to produce a convex realization or establish that it belongs to a combinatorial family of codes for which a construction of a convex realization is known. In the following, we give algebraic signatures for identi- fying when a code belongs to any of four combinatorial families of codes for which convex constructions are known. The simplest algebraic signatures of families of convex codes are CF1(JC) = ∅ or CF2(JC) = ∅. Since CF1(JC) captures minimal subsets missing from ∆(C), the signature CF1(JC) = ∅ implies ∆(C) 6 (cid:81) is the full simplex, and so C must contain the all-ones word. Convex realizations of such codes were given in [3]. When C contains the all-ones word (CF1(JC) = ∅), ∆(C) has a single facet, and this fact is exploited in the construction of convex realizations of these codes. More generally, if ∆(C) has disjoint facets this same construction can be employed in parallel for each facet, ensuring these codes are also convex [3]. These codes can also be detected algebraically, but the signature is more complicated, so we save the statement and proof of the signature for Section 3. On the other hand, CF2(JC) = ∅ implies that C is a simplicial complex, which is guaranteed to have a convex realization [3, 13]. These codes can be generalized to a broader family of codes known as intersection-complete codes, which are also known to be convex [1]. Definition 1.8. A code C is intersection-complete (∩-complete) if every intersection of codewords is also a codeword in C; i.e. σ, ω ∈ C implies that σ ∩ ω ∈ C. The algebraic signature for ∩-complete codes is given in the following theorem, whose proof appears in Section 3. Theorem 1.9. A code C is ∩-complete if and only if every pseudo-monomial xσ has τ ≤ 1. If C is ∩-complete, then C is convex. i∈τ (1− xi) ∈ CF(JC) Note that if τ = 0 for all elements of CF(JC), then CF2(JC) = ∅, which is the signature for simplicial complex codes. Using Table 1, the algebraic signature in Theorem 1.9 can be reinterpreted for any realization of an ∩-complete code C = C(U), every in terms of receptive fields as follows: intersection Uσ for σ /∈ C is minimally covered by a single set Ui for some i /∈ σ. The families of codes presented above, for which we have algebraic signatures, are special cases of max ∩-complete codes: codes for which every intersection of a collection of facets of ∆(C) is also a codeword in C. In [1], convex realizations of max ∩-codes were constructed, guaranteeing their convexity. Theorem 1.10. [1, Theorem 4.4] If a code C is max ∩-complete, then C is convex. Finding an algebraic signature of max ∩-complete codes remains an open question. Given that these codes generalize ∩-complete codes, one might hope to generalize the algebraic signature of ∩-complete codes to obtain a signature for this broader class. One natural generalization is the class of codes for i∈τ (1 − xi) ∈ CF(JC) has τ ≤ 2. Unfortunately, Example 1.11 which every pseudo-monomial xσ (below) shows that codes with this property need not be max ∩-complete and vice versa. In particular, the code in Example 1.11(b) has the τ ≤ 2 property, but is not even convex. Example 1.11. (a) Consider the code (cid:81) C1 = {0000, 0100, 0010, 0001, 1100, 1010, 1001, 0110, 0011, 1110, 1011} with maximal codewords 1110 and 1011. This code is max ∩-complete because it would in fact be a simplicial complex except that it is missing 1000, which is not an intersection of maximal codewords. However, C1 does not satisfy τ ≤ 2, since CF2(JC1) = {x1(1 − x2)(1 − x3)(1 − x4)}. (b) Consider the code C2 = {00000, 00100, 00010, 10100, 10010, 01100, 00110, 00011, 11100, 10110, 10011, 01111} with maximal codewords 11100, 10110, 10011, 01111 (i.e. facets {123, 134, 145, 2345}). C2 is not max ∩-complete since it does not contain the triple intersection of facets 1 = 123 ∩ 134 ∩ 145. However, C2 satisfies τ ≤ 2 for all xσ i∈τ (1 − xi) ∈ CF2(JC2) since (cid:81) CF2(C2) = {x2(1 − x3), x5(1 − x4), x2x4(1 − x5), x3x5(1 − x2), x1(1 − x3)(1 − x4)}. Interestingly, this code is not convex, although it has no local obstructions [10]. Note that the code from Example 1.6 also satisfies τ ≤ 2 and is not convex, but it has a local obstruction. Thus, the signature τ ≤ 2 does not ensure convexity or provide guarantees about the presence/absence of local obstructions. 7 1.3 Examples illustrating main results This section gives examples of codes satisfying each of the algebraic signatures presented in Theo- rems 1.7 and 1.9 together with an analysis of the implications of these signatures for RF relationships. We begin with an example of a code on 5 neurons that satisfies the first signature in Theorem 1.7. Example 1.12 (Theorem 1.7, signature (i)). Consider the code C = {00000, 11100, 10011, 01111} ∪ {all binary patterns with exactly two 1s}. This code has CF1(JC) = {x1x2x4, x1x2x5, x1x3x4, x1x3x5} and CF2(JC) = {xi1(1 − xi2)(1 − xi3)(1 − xi4)(1 − xi5) i1, . . . , i5 ∈ [5]} ∪ {xi1xi2xi3(1 − xi4) i1, . . . , i4 ∈ [5] \ {1}}, where all the indices in the pseudo-monomials of CF2(JC) are distinct. Consider x1(1 − x2)(1 − x3)(1 − x4)(1 − x5) ∈ CF2(JC), where σ = {1} and τ = {2, 3, 4, 5}. We will construct the graph G = GC(σ, τ ) whose vertices are precisely the elements of τ . By definition, whenever xσxixj /∈ JC for i, j ∈ τ , then (ij) is an edge in G. Using CF1(JC), we immediately see that (24), (25), (34), and (35) are not edges in G, and that (23) and (45) are edges in G (see Lemma 2.5). Thus G consists only of two disjoint edges, and is disconnected. (Note that this implies that U1 ∩ (U2 ∪ U3) and U1 ∩ (U4 ∪ U5) are disjoint, and so U1 is disconnected, as it is covered by the disjoint union of nonempty open sets.) Therefore, signature (i) of Theorem 1.7 is satisfied and C is not convex. The next example gives a code on 4 neurons satisfying the second signature of Theorem 1.7. Example 1.13 (Theorem 1.7, signature (ii)). Consider C = {0000, 1110, 1101, 1011, 0111, 1100, 1010, 1001}. Then CF1(JC) = {x1x2x3x4} and CF2(JC) = {xi(1−x1)(1−xj) i, j = 2, 3, 4; i (cid:54)= j} ∪ {x1(1−x2)(1−x3)(1−x4)}. Since x1(1 − x2)(1 − x3)(1 − x4) ∈ CF2(JC) and x1x2x3x4 ∈ CF1(JC), we see that signature (ii) of Theorem 1.7 applies. Thus C is not convex. To see the obstruction to convexity here, note that since x1(1 − x2)(1 − x3)(1 − x4) ∈ CF2(JC) we have from Table 1 that U1 is minimally covered by U2 ∪ U3 ∪ U4. Also, since x1x2x3x4 ∈ CF1(JC), the full intersection U1 ∩ U2 ∩ U3 ∩ U4 is empty, but the minimality of elements in CF(JC) guarantees that every other intersection is non-empty. This forces U1 to contain a hole (see Figure 2), and so U1 cannot be convex, and hence C cannot be convex. Figure 2: For the code in Example 1.13, the set U1 is the union of the shaded regions shown since it is covered by (U1 ∩ U2)∪ (U1 ∩ U3)∪ (U1 ∩ U4). U1 must contain a hole since the covering sets all pairwise intersect, but the full intersection is missing. 8 Finally, the following example shows how to use the neural ideal to detect that a code is ∩-complete, and thus convex. Example 1.14 (Theorem 1.9). Consider C = {00000, 11110, 10111, 01111, 10110, 01110, 00111, 00110}. This code has CF1(JC) = {x1x2x5} and CF2(JC) = {xi(1 − xj) i ∈ [5]; j = 3, 4; i (cid:54)= j}. We immediately see that all elements of CF(JC) satisfy τ ≤ 1, and so the signature from Theorem 1.9 applies. Thus, C is ∩-complete. 2 Detecting local obstructions The primary method for showing that a code is not convex is to show that it has a local obstruction. Section 2.1 defines local obstructions and connects them to links of certain restricted simplicial com- plexes. Section 2.2 shows how to detect certain classes of local obstructions via JC and CF(JC) and provides the proof of Theorem 1.7. 2.1 Local obstructions Recall that the code in Example 1.6 failed to have a convex realization because the receptive field U1 was covered by a pair of disjoint nonempty open sets U1 ∩ U2 and U1 ∩ U3, and thus no realization of C could have U1 as a convex set. In this case, the restricted cover of U1 by U1 ∩ U2 and U1 ∩ U3 had a nerve that was disconnected and thus, if U1 were convex, there would be a topological mismatch between U1 and the nerve of its restricted cover. This topological mismatch is an example of a local obstruction. Specifically, the Nerve Lemma [7, Corollary 4G.3] guarantees that if U is a convex open cover (and thus a "good cover"), then Uσ must have the same homotopy type as N ({Uσ ∩ Ui}i∈τ ) i∈τ Ui, i.e. whenever (σ, τ ) ∈ RF(C). In particular, since Uσ is the intersection of convex sets, it must be convex and hence contractible 4, and thus N ({Uσ ∩ Ui}i∈τ ) must also be contractible. Thus, if the nerve of such a restricted cover is not contractible, then a local obstruction is present. This restricted nerve has an alternative combinatorial formulation; specifically, whenever Uσ is non-empty and covered by a union of sets (cid:83) N ({Uσ ∩ Ui}i∈τ ) = Lkσ(∆σ∪τ ), where ∆σ∪τ is the restricted simplicial complex ∆σ∪τ def= {ω ∈ ∆ ω ⊆ σ ∪ τ} and the link Lkσ(∆σ∪τ ) is given by Lkσ(∆σ∪τ ) = {ω ∈ ∆σ∪τ σ ∩ ω = ∅ and σ ∪ ω ∈ ∆σ∪τ}. This alternative characterization of the nerve yields the following formal definition of local obstruction. For more details about local obstructions, see [3, Section 3]. Definition 2.1. Let C be a code on n neurons with simplicial complex ∆. For σ, τ ⊆ [n] with τ (cid:54)= ∅, we say that (σ, τ ) is a local obstruction of C if (σ, τ ) ∈ RF(C) and the link Lkσ(∆σ∪τ ) is not contractible. As an immediate consequence of the Nerve Lemma, as described above, we obtain Lemma 2.2. Lemma 2.2. [3, Lemma 1.3] If C has a local obstruction, then C is not a convex code. 4A set is contractible if it is homotopy-equivalent to a point, and every convex set is contractible [7]. 9 2.2 Algebraic detection of local obstructions (cid:81) (cid:81) i∈τ (1 − xi) ∈ JC is not sufficient to guarantee that In general, the presence of a pseudo-monomial xσ (σ, τ ) is a RF relationship (see Table 1), and thus a possible candidate for a local obstruction. This is i∈τ (1− xi) is minimal, because we cannot guarantee that Uσ ∩ Ui (cid:54)= ∅ for all i ∈ τ . However, when xσ i∈τ (1 − xi) ∈ CF2(JC), these conditions are guaranteed and (σ, τ ) ∈ RF(C). Thus, we i.e. when xσ focus on the canonical form to algebraically detect local obstructions. Lemma 2.3. For a code C, if there exists (σ, τ ) such that xσ is not contractible, then C is not convex. i∈τ (1 − xi) ∈ CF2(JC) and Lkσ(∆σ∪τ ) (cid:81) (cid:81) With this result, we can now prove Theorem 1.7. Specifically, we prove Theorem 2.4, a broader result that also characterizes relevant RF conditions corresponding to these signatures. Theorem 2.4. If C has any of the algebraic signatures in rows A-1, A-2, A-3, or A-4 of Table 3, then C is not convex. More precisely, each algebraic signature corresponds to a RF condition (as illustrated in Figure 3), which implies that C is not convex. A-1 A-2 A-3 A-4 Algebraic signature ∃ xσ(1 − xi)(1 − xj) ∈ CF2(JC) s.t. xσxixj ∈ JC ∃ xσ i∈τ (1 − xi) ∈ CF2(JC) s.t. GC(σ, τ ) is disconnected ∃ xσ i∈τ (1 − xi) ∈ CF2(JC) s.t. xσxτ ∈ CF1(JC) (cid:81) (cid:81) (cid:81) i∈τ (1 − xi) ∈ CF2(JC), ∃ xσ ∃ ∅ ⊆ σ ⊆ σ s.t. xσxτ ∈ CF1(JC) but xσ(cid:48)xτ(cid:48) /∈ CF1(JC) ∀ σ(cid:48) ⊆ σ, τ(cid:48) (cid:40) τ Receptive field condition Property of C GU (σ, τ ) is disconnected ⇒ (σ,{i, j}) ∈ RF(C) and Uσ ∩ Ui ∩ Uj = ∅ ⇒ (σ, τ ) ∈ RF(C) and ⇒ (σ, τ ) ∈ RF(C) and Uσ ∩ Uτ = ∅ but Uσ ∩ Uτ(cid:48) (cid:54)= ∅ ∀ τ(cid:48) (cid:40) τ ⇒ (σ, τ ) ∈ RF(C) and Uσ ∩ Uτ ⊆ Uσ ∩ Uτ = ∅ but Uσ ∩ Uτ(cid:48) (cid:54)= ∅ ∀ τ(cid:48) (cid:40) τ ⇒ non-convex ⇒ non-convex ⇒ non-convex ⇒ non-convex Table 3: Algebraic signatures and receptive field conditions for non-convex codes. GC(σ, τ ) is the simple graph on vertex set τ with edge set {(ij) ∈ τ × τ xσxixj /∈ JC}. The graph GU (σ, τ ) has vertex set τ and edge set {(ij) ∈ τ × τ Uσ ∩ (Ui ∩ Uj) (cid:54)= ∅}. Figure 3: Illustrations of the RF conditions implied by signatures A-1 through A-4 in Theorem 2.4 (see Table 3). In each picture, Uσ is the union of the shaded regions; thus Uσ is not contractible and hence not convex. Proof of Theorem 2.4. (A-1) By Lemma 1.4, xσ(1 − xi)(1 − xj) ∈ CF2(JC) implies (σ,{i, j}) ∈ RF(C), and thus Uσ ⊆ Ui ∪ Uj and both Uσ ∩ Ui and Uσ ∩ Uj are non-empty. Recall from Table 1 that xσxixj ∈ JC implies Uσ ∩ Ui ∩ Uj = ∅. Thus, Uσ is the disjoint union of non-empty open sets Uσ ∩ Ui k∈σ Uk is not convex, and so some Uk is not convex. Hence C is not convex. and Uσ ∩ Uj, and so Uσ is disconnected. Thus, Uσ =(cid:84) 10 A-1A-2disconnectedA-3 & A-4 (A-2) By Table 1 if C = C(U), then GC(σ, τ ) = GU (σ, τ ) since xσxixj /∈ JC precisely when Uσ ∩ Ui ∩ Uj (cid:54)= ∅. Furthermore, this graph is precisely the 1-skeleton5 of Lkσ(∆σ∪τ ). Since we assume this is disconnected, it follows that Lkσ(∆σ∪τ ) is not contractible, and hence C is non-convex by Lemma 2.3. Alternatively, GU (σ, τ ) disconnected implies that Uσ is disconnected, and hence C cannot be convex. (A-3) The signature for A-3 is a special case of that for A-4 since xσxτ ∈ CF1(JC) guarantees xσ(cid:48)xτ(cid:48) /∈ CF1(JC) for all σ(cid:48) ⊆ σ, τ(cid:48) (cid:40) τ by minimality of the elements in the canonical form. Thus, we prove non-convexity of these codes via the following proof of A-4. (A-4) Note that xσxτ ∈ CF1(JC) implies that Uσ ∩ Uτ = ∅ by Table 1 and thus Uσ ∩ Uτ = ∅ as well. Thus, σ ∪ τ /∈ ∆(C) and so τ /∈ Lkσ(∆σ∪τ ). For every τ (cid:40) τ , we have xσxτ /∈ JC, since if it were in JC then some factor of it must be in CF1(JC), but xσ(cid:48)xτ(cid:48) /∈ CF1(JC) for every σ(cid:48) ⊆ σ and τ(cid:48) ⊆ τ . Thus, for all τ (cid:40) τ , we have σ∪ τ ∈ ∆(C) and so τ ∈ Lkσ(∆σ∪τ ); equivalently Uσ ∩ Uτ (cid:54)= ∅ for all τ (cid:40) τ . This means Lkσ(∆σ∪τ ) is a simplex missing only the top dimensional face τ (i.e. a hollow simplex), and so is homotopy-equivalent to a sphere, and thus is not contractible. At the level of RF relationships, this implies that Uσ is not contractible since it must contain a hole. Thus, C is non-convex. As the proof of Theorem 2.4 illustrates, signature A-1 captures cases where Uσ is disconnected by a pair of sets. Signature A-2 generalizes A-1 and detects all cases where Uσ is minimally covered by a collection of sets Uσ ∩ Ui for i ∈ τ in a way that forces Uσ to be disconnected. Note that A-2 is signature (i) from Theorem 1.7 in the main results (Section 1.2). Signature A-3 captures a particular case when Uσ is minimally covered by a collection of sets Uσ∩Ui for i ∈ τ and Lkσ(∆σ∪τ ) = N ({Uσ∩Ui}i∈τ ) is a hollow simplex. Specifically, in the case of A-3, Uσ∩Uτ is the minimal missing intersection in that for all σ (cid:40) σ, we have Uσ ∩ Uτ (cid:54)= ∅; thus everywhere outside of Uσ, Uτ has a non-empty intersection with each subcollection of sets from σ. More generally, signature A-4 captures all cases when Lkσ(∆σ∪τ ) is a hollow simplex. Specifically, the signature for A-4 does not require the minimality of the empty intersection Uσ ∩ Uτ , and so there may be a σ (cid:40) σ such that Uσ∩Uτ = ∅, in particular we may have Uτ = ∅. All that is required is that every intersection of Uσ with each proper subcollection of sets in τ is non-empty, which is guaranteed by ensuring that Uσ(cid:48) ∩ Uτ(cid:48) (cid:54)= ∅, for all σ(cid:48) ⊆ σ and τ(cid:48) (cid:40) τ . Signature (ii) from Theorem 1.7 is A-3, a special case of A-4, and so the proof of Theorem 2.4 completes the proof of Theorem 1.7. Note that although checking signatures A-1 and A-2 requires determining the absence of pseudo- monomials from JC, these conditions can actually be checked from CF1(JC) alone as Lemma 2.5 shows. Thus, all the algebraic signatures in Theorem 1.7 can be checked directly via CF(JC). (cid:81) Lemma 2.5. Suppose xσ k∈τ (1 − xk) ∈ CF2(JC). Then for any i, j ∈ τ with i (cid:54)= j, xσxixj ∈ JC ⇔ xσ(cid:48)xixj ∈ CF1(JC) for some ∅ ⊆ σ(cid:48) ⊆ σ. Proof. The backward direction (⇐) is immediate since CF(JC) ⊆ JC. To see the forward direction (⇒), suppose xσxixj ∈ JC. Then it is a multiple of some monomial xω ∈ CF1(JC) with ω ⊆ σ ∪ {i, j}. There are four possibilities: (1) ω ⊆ σ. Then xω divides xσ (2) ω = σ(cid:48) ∪ {i} for some σ(cid:48) ⊆ σ. Then xσxi hence (1 − xk) + xσ k∈τ (1−xk) ∈ CF2(JC). k∈τ\{i}(1 − xk) ∈ JC, since it is a multiple of xω, and (1 − xk) = xσ k∈τ (1−xk), contradicting the minimality of xσ (1 − xk) ∈ JC, (cid:89) (cid:81) (cid:81) (cid:89) k∈τ\{i} (cid:81) (cid:89) k∈τ xσxi k∈τ\{i} 5The 1-skeleton of a simplicial complex is the subcomplex consisting of all faces of dimension at most 1, i.e. the vertices and edges of the simplicial complex; thus the 1-skeleton is the underlying graph of the simplicial complex (see e.g. [7]). 11 (cid:81) contradicting the minimality of xσ (3) ω = σ(cid:48)∪{j} for some σ(cid:48) ⊆ σ. This argument is identical to the previous, leading to a contradiction. k∈τ (1 − xk) ∈ CF2(JC). Thus, the only viable possibility is: (4) ω = σ(cid:48) ∪ {i, j} for some σ(cid:48) ⊆ σ, and thus xω = xσ(cid:48)xixj ∈ CF1(JC). (cid:81) As mentioned at the beginning of this section, the algebraic signatures in Theorem 1.7 only consider minimal RF relationships as detectable by the canonical form. The motivation for this is that other i∈τ (1 − xi) in the full ideal JC are not guaranteed to correspond to RF rela- pseudo-monomials xσ tionships as Uσ ∩ Ui is not necessarily non-empty for all i ∈ τ . This begs the question of whether it is sufficient to only consider these minimal pseudo-monomials and minimal RF relationships. Specifically, if for every pseudomonomial in CF2(JC), we find that the links Lkσ(∆σ∪τ ) are all contractible, does that guarantee that C has no local obstructions? Unfortunately, this is not the case. Example 2.6 shows that there exist codes that have no local i∈τ (1− xi) ∈ CF2(JC), and yet the codes still have local obstructions arising from pairs (σ, τ ) with xσ obstructions. Example 2.6. Consider the code C = {0000, 1110, 1101, 1011, 0111}, where ∆ = ∆(C) is the hollow simplex on four vertices, missing only the top-dimensional face (see Figure 4A). The canonical form is (cid:81) CF(JC) = {x1x2x3x4} ∪ {xi(1 − xj)(1 − xk) i, j, k ∈ [4] with i (cid:54)= j (cid:54)= k}. The minimal RF relationships (σ, τ ) = ({i},{j, k}) detected by the canonical form all have corre- sponding links Lki(∆{i,j,k}) that are equivalent to the simplex shown in Figure 4B, and hence are contractible. However, C has multiple local obstructions, and thus is not convex. For example, observe that ({1, 2},{3, 4}) ∈ RF(C) since (U1 ∩ U2) ⊆ U3 ∪ U4 and (U1 ∩ U2) ∩ Ui (cid:54)= ∅ for i = 3, 4, and Lk12(∆[4]) is the non-contractible disconnected graph in Figure 4C. Thus, ({1, 2},{3, 4}) is a local ob- struction, and so C cannot be convex. Note that since ({1, 2},{3, 4}) is a non-minimal RF relationship, its corresponding pseudo-monomial x1x2(1 − x3)(1 − x4) is only in JC and not in CF(JC). Similarly, ({1},{2, 3, 4}) gives another local obstruction that is not detectable from the canonical form. Specifically, ({1},{2, 3, 4}) ∈ RF(C) since U1 ⊆ U2 ∪ U3 ∪ U4 with U1 ∩ Ui (cid:54)= ∅ for each i = 2, 3, 4, and Lk1(∆[4]) is the non-contractible hollow simplex shown in Figure 4D. In fact, it turns out that every non-maximal σ ∈ ∆ has a related RF relationship that is a local obstruction (see [3, Table 2 in Supplementary Text S1]). Figure 4: Simplicial complexes in Example 2.6. Note that the simplicial complex in (A) is missing the top-dimensional face {1, 2, 3, 4} and thus is a hollow simplex. 3 Detecting convex codes Recall that to prove a code is convex, it is not sufficient to show that it has no local obstructions [10]. Currently the only known method for proving convexity is to construct a convex realization or 12 124334234ADCB show that the code belongs to a combinatorial family of codes that have been proven to be convex. The broadest such family of codes is max ∩-complete codes, for which every intersection of facets of ∆(C) is a codeword in C. Currently, however, there is no efficient way to determine if a code is max ∩-complete. Thus, we instead provide algebraic signatures of four combinatorial families of codes that all happen to be max ∩-complete, and thus are guaranteed to be convex by Theorem 1.10. Moreover, these signatures can be checked efficiently. Table 4 summarizes these signatures together with the combinatorial property implied by each signature. Section 1.2 provided sketches of proofs for signatures B-1 and B-2. In this section, we prove B-3 and B-4. Algebraic signature of JC B-1 CF1(JC) = ∅ B-2 CF2(JC) = ∅ B-3 B-4 (cid:81) ∀ xσ ∈ CF1(JC), σ = 2, and if xixj ∈ CF1(JC), then xixk or xjxk ∈ CF1(JC) for all k ∈ [n] ∀ xσ i∈τ (1 − xi) ∈ CF2(JC), τ = 1 Property of C ⇒ convex (11··· 1 ∈ C) ⇒ convex (C = ∆(C)) ⇒ convex (∆(C) has disjoint facets) ⇒ convex (C is ∩–complete) Table 4: Algebraic signatures of convex codes. Proof of B-3. Recall that signature B-1 captures when a code contain the all-ones word, and thus the corresponding simplicial complex has a single facet. As a generalization, we consider codes whose simplicial complexes have disjoint facets, which are also provably convex [3]. In the following, we show these codes can be algebraically detected via the signature B-3, but first we need the following lemma. Lemma 3.1. The graph, or 1-skeleton, of ∆(C) is a disjoint union of maximal cliques if and only if the following property holds: if xixj ∈ CF1(JC), then xixk or xjxk ∈ CF1(JC) for all k ∈ [n]. (∗) Proof. Let G be the underlying graph of ∆ = ∆(C), i.e. its 1-skeleton. Observe that xixj ∈ CF1(JC) precisely when {i, j} is a minimal set missing from ∆, and so i and j are vertices of G, but the edge (ij) is missing from G. (⇒) If G is a disjoint union of maximal cliques, then whenever two vertices i and j are in distinct maximal cliques, no other vertex k can be adjacent to both i and j. This means that whenever xixj ∈ CF1(JC), for every k ∈ [n], at least one of xixk or xjxk ∈ JC. Since these elements are minimal, we must have xixk or xjxk ∈ CF1(JC) (because xk /∈ JC for any k since we assume Uk (cid:54)= ∅). (⇐) We prove this by contrapositive. Suppose that G is not the disjoint union of maximal cliques. Then there exist distinct vertices i, j, k ∈ G that are connected, but do not form a clique; specifically, (ik) and (jk) are edges in G, but (ij) is not. Since (ij) /∈ G, xixj ∈ CF1(JC), but neither xixk nor xjxk is in CF1(JC); thus violating the condition on CF1(JC) from the statement. Satisfying property (*) from Lemma 3.1 alone is not sufficient to guarantee convexity, as the fol- lowing example shows. Example 3.2. Consider C = {000000, 111000, 110100, 101100, 000011, 110000, 101000, 100100} with CF1(JC) = {x1x5, x1x6, x2x5, x2x6, x3x5, x3x6, x4x5, x4x6, x2x3x4} and CF2(JC) = {x1(1 − x2)(1 − x3)(1 − x4), x2(1 − x1), x3(1 − x1), x4(1 − x1), x5(1 − x6), x6(1 − x5)}. Observe that property (∗) from Lemma 3.1 holds for CF1(JC), and the graph of ∆(C) is the disjoint union of the 4-clique on {1, 2, 3, 4} and the edge (56). C is not convex, however, because it satisfies signature A-4 via x1(1 − x2)(1 − x3)(1 − x4) ∈ CF2(JC) and x2x3x4 ∈ CF1(JC) which forces a local obstruction since Lk1(∆[4]) is the non-contractible hollow triangle. 13 Despite not necessarily being convex, codes satisfying property (∗) from Lemma 3.1 display an interesting relation. Consider the co-firing relation defined by i ∼ j if and only if neurons i and j "co-fire" together in C, i.e. {i, j} ⊆ σ for some σ ∈ C. It is straightforward to check that this is an equivalence relation precisely when the code satisfies property (∗), as that condition on the 1-skeleton of ∆(C) ensures transitivity. We now turn to the subclass of codes described in Lemma 3.1 that have simplicial complexes with disjoint facets, and thus are guaranteed to be convex [3]. This will complete the proof of B-3. Proposition 3.3. Given a code C, its simplicial complex ∆(C) has disjoint facets if and only if the following two properties hold: 1. For all xσ ∈ CF1(JC), we have σ = 2; and 2. If xixj ∈ CF1(JC), then xixk or xjxk ∈ CF1(JC) for all k ∈ [n]. Proof. Observe that ∆ = ∆(C) has disjoint facets if and only if ∆ is a disjoint union of simplices. This occurs precisely when (a) the 1-skeleton of ∆ is a disjoint union of maximal cliques and (b) each maximal clique is in ∆ (i.e. each maximal clique yields a simplex in ∆). By Lemma 3.1, (a) holds if and only if Property 2 is satisfied. Note that (b) holds if and only if every clique of the 1-skeleton is in ∆, not just the maximal cliques, since ∆ is closed under taking subsets. Property 1 guarantees that xω ∈ JC (i.e. ω /∈ ∆) if and only if xω is a multiple of xixj ∈ CF1(JC) for some i, j ∈ ω, and so {i, j} /∈ ∆. Thus Property 1 ensures ω /∈ ∆ if and only if ω is missing an edge (ij), and thus is not a clique. Hence (b) holds if and only if Property 1 is satisfied. Note that the signature in B-3 relies solely on CF1(JC), which is a generating set for the Stanley- Reisner ideal of ∆(C); thus, this property can be read off from the Stanley-Reisner ideal alone. Proof of B-4. Recall from B-2 that a code C is a simplicial complex precisely when all pseudo- monomials in CF(JC) have τ = 0, and in this case C is provably convex, and consequently has no local obstructions. It turns out that when all pseudo-monomials in CF2(JC) have τ = 1, Corollary 3.6 shows that no local obstructions can arise involving the neurons in σ and {i}. To prove this, we first need the following two lemmas. Lemma 3.4. Let ρ ∈ F2[x1, . . . , xn] and i ∈ [n]. If ρ /∈ JC but ρ(1 − xi) ∈ JC, then ρxi /∈ JC. Proof. If ρxi ∈ JC, then the sum ρ(1 − xi) + ρxi ∈ JC, but ρ(1 − xi) + ρxi = ρ /∈ JC by hypothesis. Thus, ρxi /∈ JC. Lemma 3.5. For a code C, if xσ /∈ JC but xσ(1 − xi) ∈ JC, then for any σ ⊇ σ and any τ ⊇ {i}, the link Lkσ(∆σ∪τ ) is contractible. Proof. To show the link Lkσ(∆σ∪τ ) is contractible, we show it is a cone. Let τ(cid:48) ∈ Lkσ(∆σ∪τ ); note τ(cid:48) ⊆ τ . This implies σ ∪ τ(cid:48) ∈ (∆σ∪τ ) ⊆ ∆(C), and so ρ = xσxτ(cid:48) /∈ JC. But ρ(1 − xi) ∈ JC since it is a multiple of xσ(1 − xi) ∈ JC. Thus by Lemma 3.4, ρxi /∈ JC which implies σ ∪ τ(cid:48) ∪ {i} ∈ ∆(C), and so τ(cid:48) ∪ {i} ∈ Lkσ(∆σ∪τ ). Thus, i is a cone point of Lkσ(∆σ∪τ ), and so the link is contractible. Corollary 3.6. For a code C, if xσ(1− xi) ∈ CF(JC), then (σ, τ ) is not a local obstruction of C for any σ ⊇ σ and any τ ⊇ {i}. Proof. Observe that if xσ(1 − xi) ∈ CF(JC), then by minimality xσ /∈ JC. Thus by Lemma 3.5, for any (σ, τ ), the link Lkσ(∆σ∪τ ) is contractible, and thus (σ, τ ) cannot be a local obstruction. Corollary 3.6 shows that no RF relationship containing a minimal RF relationship (σ,{i}) (i.e. when xσ(1 − xi) ∈ CF(JC)) can produce a local obstruction. Thus, if a code only has minimal RF relationships of the form (σ,∅) or (σ,{i}), then it has no local obstructions, since every RF relationship 14 (σ, τ ) must contain one of these minimal RF relationships. Such a code can be immediately identified from its canonical form, as every pseudo-monomial will satisfy τ ≤ 1. Furthermore, Proposition 3.7 shows that these codes are precisely ∩-complete codes. Since ∩-complete codes are max ∩-complete, Theorem 1.10 guarantees these codes are in fact convex beyond simply having no local obstructions. This completes the proof of B-4 and Theorem 1.9 from Section 1.2. Proposition 3.7. For a code C, the following are equivalent: (1) Every pseudo-monomial xσ (2) For each RF relationship (σ, τ ) with τ (cid:54)= ∅, there exists an i ∈ τ such that (σ,{i}) is also an RF i∈τ (1 − xi) in CF(JC) has τ ≤ 1, (cid:81) relationship, and (3) C is ∩-complete. (cid:81) In the proof, we will use the notation Uσ∪τ to denote the subcover Uσ∪τ def= {Ui i ∈ σ ∪ τ}. def= {ω ∈ C ω ⊆ σ ∪ τ}. Note that if C = C(U) Also we use Cσ∪τ to denote the restricted code Cσ∪τ then Cσ∪τ = C(Uσ∪τ ). We also use the straightforward fact that if C is ∩-complete, then Cσ∪τ is ∩-complete for any σ ∪ τ ⊂ [n]. Proof. We prove (1) ⇔ (2), and then (1) ⇔ (3). (1)⇔(2): By Lemma 1.4, there is a one-to-one correspondence between pseudo-monomials in CF(JC) i∈τ (1 − xi) in CF(JC) has τ ≤ 1 if and only if the and minimal RF relationships. Thus, every xσ only minimal RF relationships are those of the form (σ,∅) and (σ,{i}) for some i ∈ [n]. Recall that if (σ, τ ) is a RF relationship with τ (cid:54)= ∅, then Uσ (cid:54)= ∅ since Uσ ∩ Ui (cid:54)= ∅ for all i ∈ τ ; hence (σ, τ ) contains a minimal RF relationship (σ(cid:48), τ(cid:48)) for some non-empty σ(cid:48) ⊆ σ and non-empty τ(cid:48) ⊆ τ . We see that (1) holds if and only if each such RF relationship (σ, τ ) contains a minimal relationship of the form (σ(cid:48),{i}) for some i ∈ τ , in which case, (σ,{i}) is also an RF relationship since Uσ ⊆ Uσ(cid:48). Thus (1) and (2) are equivalent. (1)⇒(3): Suppose CF(JC) only contains pseudo-monomials with τ ≤ 1. Consider a pair of overlap- ping codewords ω1, ω2 ∈ C and let σ = ω1 ∩ ω2. To obtain a contradiction, suppose σ /∈ C. Then σ ∈ ∆(C) \ C, since it is a subset of ω1, ω2 ∈ ∆(C). This implies that xσ i∈τ (1 − xi) ∈ JC for some τ (cid:54)= ∅ with σ ∩ τ = ∅. It follows that xσ i∈τ (1− xi) must be a multiple of a Type 2 pseudo-monomial in the canonical form, say xσ(cid:48)(1 − xi) ∈ CF(JC), where i ∈ τ and σ(cid:48) ⊆ σ is nonempty. In particular, xσ(1 − xi) ∈ JC, which implies Uσ ⊆ Ui by Table 1. Since σ ⊆ ω1, Uω1 ⊆ Uσ ⊆ Ui which implies i ∈ ω1 since otherwise the region corresponding to ω1 would be covered and thus could not produce a codeword. Similarly, i ∈ ω2, and so i ∈ ω1 ∩ ω2 = σ. But then i ∈ σ ∩ τ contradicting σ ∩ τ = ∅. Hence σ ∈ C and C is ∩-complete. (3)⇒(1): Suppose C is ∩-complete, and consider an element xσ contradiction, assume τ > 1 and let j, k ∈ τ with j (cid:54)= k. i∈τ (1 − xi) ∈ CF(JC). Note that i∈τ Ui in Uσ∪τ , and so σ /∈ Cσ∪τ . To obtain a (cid:81) (cid:81) (cid:81) i∈τ\{j} Ui. This would imply that Uσ ⊆(cid:83) i∈τ\{j} Ui. Then it would follow that the portion of Uσ that is covered i∈τ\{j} Ui, i∈τ (1 − xi) ∈ CF(JC). i∈τ Ui guaranteed by xσ (cid:81) for any cover U with C = C(U), this implies Uσ ⊆ (cid:83) Suppose that Uσ∪{j} ⊆ (cid:83) by Uj (i.e. Uσ ∩ Uj = Uσ∪{j}) is also covered by(cid:83) but that would contradict the minimality of Uσ ⊆(cid:83) Hence, Uσ∪{j} (cid:42)(cid:83) Uσ∪{j} \ (cid:91) i∈τ\{j} Ui, and so Ui (cid:54)= ∅ i∈τ\{j} 15 ensuring that σ ∪ {j} ∈ Cσ∪τ .6 By the same argument, we have σ ∪ {k} ∈ Cσ∪τ . σ = ωj ∩ ωk must be in Cσ∪τ . But this contradicts the hypothesis that xσ j, k ∈ τ , which guaranteed σ /∈ Cσ∪τ . Thus, we conclude that τ ≤ 1 and (1) holds. Let ωj = σ ∪ {j} and ωk = σ ∪ {k}. Since the restricted code Cσ∪τ is ∩-complete, this implies that i∈τ (1 − xi) ∈ CF(JC) with (cid:81) 4 Examples In this section, we provide examples to illustrate how one can use the algebraic signatures summarized in Table 5 to detect convexity or non-convexity. In Section 1.3, examples of codes satisfying A-2, A-3, and B-4 were given, and so we do not give examples for those signatures here. (cid:81) (cid:81) (cid:81) Algebraic signature of JC ∃ xσ(1 − xi)(1 − xj) ∈ CF2(JC) s.t. xσxixj ∈ JC ∃ xσ ∃ xσ ∃ xσ but xσ(cid:48)xτ(cid:48) /∈ CF1(JC) for all σ(cid:48) ⊆ σ, τ(cid:48) (cid:40) τ ⇒ non-convex ⇒ non-convex i∈τ (1 − xi) ∈ CF2(JC) s.t. GC(σ, τ ) is disconnected i∈τ (1 − xi) ∈ CF2(JC) s.t. xσxτ ∈ CF1(JC) ⇒ non-convex i∈τ (1 − xi) ∈ CF2(JC), and ∃ σ ⊆ σ s.t. xσxτ ∈ CF1(JC) ⇒ non-convex Property of C A-1 A-2 A-3 A-4 B-1 CF1(JC) = ∅ B-2 CF2(JC) = ∅ ∀ xσ ∈ CF1(JC), σ = 2, and if xixj ∈ CF1(JC), then xixk or xjxk ∈ CF1(JC) for all k ∈ [n] ∀ xσ i∈τ (1 − xi) ∈ CF2(JC), τ = 1 (cid:81) B-3 B-4 ⇒ convex (11··· 1 ∈ C) ⇒ convex (C = ∆(C)) ⇒ convex (∆(C) has disjoint facets) ⇒ convex (C is ∩–complete) Table 5: Algebraic signatures of convex and non-convex codes. GC(σ, τ ) is the simple graph on vertex set τ with edge set {(ij) ∈ τ × τ xσxixj /∈ JC}. Example 4.1 (signature A-1). Consider C1 = {000, 110, 101, 011}. This code has CF1(JC1) = {x1x2x3} and CF2(JC1) = {xi(1 − xj)(1 − xk) i, j, k = 1, 2, 3; and all indices distinct}. Observe that x1(1− x2)(1− x3) ∈ CF2(JC1) and x1x2x3 ∈ CF1(JC1) ⊆ JC1. Thus signature A-1 applies, and so C1 is not convex. U1 ∩ U2 and U1 ∩ U3; hence U1 is disconnected and not convex. At the level of receptive fields, the obstruction is that U1 is the disjoint union of nonempty sets Example 4.2 (signature A-4). Consider C2 = {0, 1}5 \ {11000, 10111, 11111}. Then CF1(JC2) = {x1x3x4x5} and CF2(JC2) = {x1x2(1 − x3)(1 − x4)(1 − x5)}. i∈τ (1 − xi) ∈ CF2(JC2). For σ = {1}, we see Consider σ = {1, 2} and τ = {3, 4, 5}, so that xσ xσxτ ∈ CF1(JC2) and for all σ(cid:48) ⊆ σ, τ(cid:48) (cid:40) τ , xσ(cid:48)xτ(cid:48) /∈ CF1(JC2). Thus A-4 applies, and so C2 is not convex. (cid:81) (cid:81) i∈τ\j(1 − xi). We have ρ(1 − xj) ∈ CF(JC) while ρ /∈ JC, by minimality of elements in CF(JC), and so Lemma 3.4 guarantees ρxj = xσxj i∈τ\j(1− xi) /∈ JC. Thus σ ∪ {j} ∈ Cσ∪τ , since the absence of that pseudo-monomial from JC implies Uσ∪j is nonempty and is not covered sets in τ \ {j}. 6Alternatively, we can see σ ∪ {j} ∈ Cσ∪τ algebraically, by considering ρ = xσ (cid:81) 16 In terms of receptive fields, we have that U1 ∩ Uτ = ∅, and so Uσ ∩ Uτ = ∅. But for all σ(cid:48) ⊆ σ and (cid:54)= ∅. Hence, the collection of sets Uσ ∩ Ui for i ∈ τ form a hollow simplex τ(cid:48) (cid:40) τ , we have Uσ(cid:48) ∩ Uτ(cid:48) covering Uσ, forcing Uσ to contain a hole. Example 4.3 (signature B-1). Consider C4 = {000, 110, 101, 111}. Then CF1(JC4) = ∅ and CF2(JC4) = {x2(1 − x1), x3(1 − x1), x1(1 − x2)(1 − x3)}. Since CF1(JC4) is empty, signature B-1 applies, and so C4 is convex. We could have also seen this directly from the fact that 111 ∈ C4. Example 4.4 (signature B-2). Consider C5 = {0000, 1000, 0100, 0010, 0001, 1100, 1010, 0110, 0011, 1110}. This code has CF1(JC5) = {x1x4, x2x4} and CF2(JC5) = ∅. Since CF2(JC5) is empty, signature B-2 applies, and so C5 is a simplicial complex, and hence is convex. Example 4.5 (signature B-3). Consider C6 = {0000, 0100, 0001, 1100, 1010, 1110}. This code has CF1(JC6) = {x1x4, x2x4, x3x4} and CF2(JC6) = {x3(1 − x1), x1(1 − x2)(1 − x3)}. Observe that all the elements of CF1(JC6) have σ = 2, satisfying the first part of signature B-3. For x1x4 ∈ CF1(JC6), we have x2x4 ∈ CF1(JC6) and x3x4 ∈ CF1(JC6) so the second condition holds for i = 1, j = 4, and k = 2, 3. It is easy to see the condition also holds for i = 2, j = 4 and for i = 3, j = 4. Thus, signature B-3 applies, so ∆(C6) has disjoint facets (specifically, {1, 2, 3} and {4}) and C6 is convex. The signatures A-1 through A-4 guarantee non-convexity by way of a local obstruction that can be detected from the canonical form, while signatures B-1 through B-4 guarantee convexity, and thus no local obstructions. However, Example 2.6 showed that even in the absence of local obstructions corresponding to elements of CF(JC), a code C may still have local obstructions and thus be provably non-convex. Additionally, even when a code has no local obstructions of any type, it may still be non- convex (see C2 in Example 1.11(b), first observed to be non-convex in [10]). Despite these complicating factors, it may still be useful to identify when a code cannot have any local obstructions "arising" from canonical form elements; more precisely, it has no CF-detectable local obstructions, as defined below. Definition 4.6. A local obstruction (σ, τ ) is CF-detectable if there exists a local obstruction (σ(cid:48), τ(cid:48)) with σ(cid:48) ⊆ σ and τ(cid:48) ⊆ τ, such that (σ(cid:48), τ(cid:48)) is a minimal RF relationship. C-1 and C-2 give two algebraic signatures of codes with no CF-detectable local obstructions. Sup- plemental Text S2 provides more background on CF-detectable local obstructions and Theorem 5.4 proving these signatures. Example 4.7 (signature C-1). Consider the code C7 = {0000000, 0100000, 0010000, 0001000, 0000100, 0000010, 1100000, 1010000, 1001000, 0110000, 0101000, 0011000, 0001100, 0000110, 0000101, 0000011, 1110000, 1101000, 1011000, 0111000, 0000111, 1111000}. More compactly, we can describe the codewords as subsets of active neurons, and we obtain C7 = {∅, 2, 3, 4, 5, 6, 12, 13, 14, 23, 24, 34, 45, 56, 57, 67, 123, 124, 134, 234, 567, 1234}, 17 (cid:81) (cid:81) Algebraic signature of JC ∀ xσ ∀ xσ ∃ i ∈ τ s.t. xixω /∈ CF1(JC) for all ω ⊆ σ ∪ τ i∈τ (1 − xi) ∈ CF2(JC), xσxτ /∈ JC i∈τ (1 − xi) ∈ CF2(JC), Property of C ⇒ no CF-detectable local obstructions ⇒ no CF-detectable local obstructions C-1 C-2 Table 6: Algebraic signatures of codes with no CF-detectable local obstructions. These codes are not guaranteed to be convex or non-convex, but do not have any local obstructions that can be detected from the canonical form. with maximal codewords 1234, 45, and 567. This code has CF1(JC7) = {x1x5, x1x6, x1x7, x2x5, x2x6, x2x7, x3x5, x3x6, x3x7, x4x6, x4x7}, and CF2(JC7) = {x1(1 − x2)(1 − x3)(1 − x4), x7(1 − x5)(1 − x6)}. Since all the elements of CF1(JC7) have σ ≤ 2, we might attempt to apply signature B-3 to guarantee convexity; however, that signature fails here since x1x5 ∈ CF1(JC7) but neither x1x4 ∈ CF1(JC7) nor x4x5 ∈ CF1(JC7). Thus, we must turn to CF2(JC7). For x1(1 − x2)(1 − x3)(1 − x4), we see that x1x2x3x4 /∈ JC7 since no factor of it is in CF1(JC7). Similarly, for x7(1 − x5)(1 − x6), we see that x5x6x7 /∈ JC7 since it has no factors in CF1(JC7). Thus, signature C-1 is satisfied and C7 has no CF-detectable local obstructions. This signature does not enable us to conclude anything about the convexity of C7; however, C7 is in fact convex, as it is max ∩-complete. Figure 5: Links from Example 4.7. (A) Lk1(∆1234) is a simplex. (B) Lk7(∆567) is a simplex. (cid:81) Recall from Section 2.1 that the source of CF-detectable local obstructions is non-contractible links of the form Lkσ(∆σ∪τ ), where xσ i∈τ (1−xi) ∈ CF2(JC). For C7, the relevant links occur when σ = {1} and τ = {2, 3, 4} and when σ = {7} and τ = {5, 6}. Since x1x2x3x4 /∈ JC7, the link Lk1(∆1234) contains the top-dimensional face 234, and thus is a simplex, which is contractible (see Figure 5A). Similarly, Lk7(∆567) (shown in Figure 5B) is a simplex since x5x6x7 /∈ JC7, and so is contractible. Signature C-1 precisely characterizes when all the relevant links Lkσ(∆σ∪τ ) are simplices, and hence contractible, for minimal receptive field relationships (σ, τ ). This ensures the absence of any CF-detectable local obstructions. The previous example showed that some codes satisfying signature C-1 are convex; however, this signature does not guarantee convexity. Specifically, code C2 from Example 1.11 also satisfies this signature, and in fact has no local obstructions, yet that code is not convex [10]. Example 4.8 (signature C-2). Consider the code C8 = {0000000, 0100000, 0010000, 0001000, 0000100, 0000010, 1100000, 1010000, 1001000, 0101000, 0011000, 0010100, 0001100, 0000110, 0010001, 0001001, 0000101, 0000011, 1101000, 1011000, 0011100, 0011001, 0010101, 0001101, 0000111, 0011101}. More compactly, C8 = {∅, 2, 3, 4, 5, 6, 12, 13, 14, 24, 34, 35, 45, 56, 37, 47, 57, 67, 124, 134, 345, 347, 357, 457, 567, 3457}. 18 A423B56 This code has CF1(JC8) = {x1x5, x1x6, x1x7, x2x3, x2x5, x2x6, x2x7, x3x6, x4x6} and CF2(JC8) = {x1(1 − x2)(1 − x3)(1 − x4), x7(1 − x3)(1 − x4)(1 − x5)(1 − x6)}. For x1(1 − x2)(1 − x3)(1 − x4) ∈ CF2(JC8), we have σ = {1} and τ = {2, 3, 4}. Observe that x4 does not appear together with x1, x2, or x3 in CF1(JC8), so for i = 4 ∈ τ , we have xixω /∈ CF1(JC8) for every ω ⊆ σ ∪ τ . For x7(1−x3)(1−x4)(1−x5)(1−x6) ∈ CF2(JC8), we have σ = {7} and τ = {3, 4, 5, 6}. Observe that x5 does not appear with any of x3, x4, x6, or x7 in CF1(JC8), so for i = 5 ∈ τ , we have xixω /∈ CF1(JC8) for every ω ⊆ σ∪ τ . Thus, signature C-2 is satisfied, and so C8 has no CF-detectable local obstructions. In fact, C8 is convex, as it is max ∩-complete. Figure 6: Links from Example 4.8. (A) Lk1(∆1234) is a cone with respect to 4. (B) Lk7(∆34567) is a cone with respect to vertex 5. (cid:81) As noted in Example 4.7, to understand the absence of CF-detectable local obstructions we need i∈τ (1 − xi) ∈ CF2(JC). For C8, the relevant links occur to consider links for pairs (σ, τ ) where xσ when σ = {1} and τ = {2, 3, 4} and when σ = {7} and τ = {3, 4, 5, 6}. The link Lk1(∆1234) is shown in Figure 6A and is a cone with respect to vertex 4, so is contractible. Similarly, Lk7(∆34567) is shown in Figure 6B and is a cone with respect to vertex 5, so is contractible. In fact, signature C-2 characterizes when all the relevant links Lkσ(∆σ∪τ ) are cones, and hence contractible, for minimal receptive field relationships (σ, τ ). Thus, signature C-2 generalizes C-1. This again ensures the absence of any CF-detectable local obstructions, but does not necessarily ensure convexity (e.g. code C2 from Example 1.11 satisfies this signature, but is not convex). It is worth noting though that ∩-complete codes (characterized by signature B-4) are a special class of codes satisfying signature C-2 that are guaranteed to be convex. Specifically, if a code satisfies B-4, then every element of CF2(JC) has the form xσ(1 − xi), and so Lemma 3.4 guarantees that xσxi /∈ JC. Thus, no factor of xσxi can be in CF1(JC), and so signature C-2 holds. Acknowledgements This work began at a 2014 AMS Mathematics Research Community, "Algebraic and Geometric Meth- ods in Applied Discrete Mathematics," which was supported by NSF DMS-1321794. CC was supported by NIH R01 EB022862 and NSF DMS-1516881; EG was supported by NSF DMS-1620109; JJ was sup- ported by NSF DMS-1606353; KM was supported by NIH R01 EB022862; and AS was supported by NSF DMS-1312473/1513364 and Simons Foundation grant 521874. We thank Mohamed Omar and Caitlin Lienkaemper for numerous discussions. References [1] J. Cruz, C. Giusti, V. Itskov, and W. Kronholm. On open and closed convex codes. Available online at https://arxiv.org/abs/1609.03502. 19 A423B3564 [2] C. Curto. What can topology tells us about the neural code? Bull. Amer. Math. Soc., 54(1):63–78, 2017. [3] C. Curto, E. Gross, J. Jeffries, K. Morrison, M. Omar, Z. Rosen, A. Shiu, and N. Youngs. What makes a neural code convex? SIAM J. of Appl. Algebra and Geometry, 1:222–238, 2017. [4] C. Curto, V. Itskov, K. Morrison, Z. Roth, and J. L. Walker. Combinatorial neural codes from a mathematical coding theory perspective. Neural Comput., 25(7):1891–1925, 2013. [5] C. Curto, V. Itskov, A. Veliz-Cuba, and N. Youngs. The neural ring: an algebraic tool for analyzing the intrinsic structure of neural codes. Bull. Math. Biol., 75(9):1571–1611, 2013. [6] C. Giusti and V. Itskov. A no-go theorem for one-layer feedforward networks. Neural Comput., 26(11):2527–2540, 2014. [7] A. Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002. [8] D. H. Hubel and T. N. Wiesel. Receptive fields of single neurons in the cat's striate cortex. J. Physiol., 148(3):574–591, Oct 1959. [9] R. A. Jeffs. Convexity of neural codes. Undergraduate thesis. Harvey Mudd College, 2016. [10] C. Lienkaemper, A. Shiu, and Z. Woodstock. Obstructions to convexity in neural codes. Advances in Appl. Math., 85:31–59, 2017. [11] J. O'Keefe and J. Dostrovsky. The hippocampus as a spatial map. Preliminary evidence from unit activity in the freely-moving rat. Brain Res., 34(1):171–175, 1971. [12] E. Petersen, N. Youngs, R. Kruse, D. Miyata, R. Garcia, and L.D. Garcia Puente. Neural ideals in SageMath. Available online at https://arxiv.org/abs/1609.09602. [13] M. Tancer. Intersection patterns of convex sets via simplicial complexes: a survey. In Thirty essays on geometric graph theory, pages 521–540. Springer, New York, 2013. [14] N. Youngs. Neural ideal: a Matlab package for computing canonical forms. https://github. com/nebneuron/neural-ideal, 2015. 5 Supplemental Text S1: Computing the canonical form CF(JC) In the following two examples, we illustrate how to compute the canonical form by hand. For details on how to algorithmically calculate CF(JC) and software to support this, see [12]. Example 5.1 (CF(JC5) from Example 4.4). Consider the code C5 = {0000, 1000, 0100, 0010, 0001, 1100, 1010, 0110, 0011, 1110} from Example 4.4. Here we show how to compute CF(JC5) by hand. Recall the neural ideal JC def= (cid:104)χv v ∈ Fn 2 \ C(cid:105), where χv is the characteristic pseudo-monomial of v (as defined in Equation (1)). The non-codewords are 1001, 0101, 1101, 1011, 0111, 1111, and so JC5 = (cid:104){x1x4(1 − x2)(1 − x3), x2x4(1 − x1)(1 − x3), x1x2x4(1 − x3), x1x3x4(1 − x2), x2x3x4(1 − x1), x1x2x3x4}(cid:105). Since x1x4 is a minimal divisor of generators in JC5 that vanishes on all codewords (so it is in JC5), we have x1x4 ∈ CF1(JC5). Similarly, x2x4 ∈ CF1(JC5). Since all the generators of JC5 are multiples of x1x4 20 and x2x4, both of which are monomials, it follows that CF2(JC5) is empty and CF1(JC5) = {x1x4, x2x4}. Thus, CF(JC5) = CF1(JC5) ∪ CF2(JC5) where CF1(JC5) = {x1x4, x2x4} and CF2(JC5) = ∅. Example 5.2 (CF(JC6) from Example 4.5). Consider the code C6 = {0000, 0100, 0001, 1100, 1010, 1110} from Example 4.5. The non-codewords are 1000, 0010, 1001, 0110, 0101, 0011, 1101, 1011, 0111, 1111. Thus, JC6 = (cid:104){x1(1 − x2)(1 − x3)(1 − x4), x3(1 − x1)(1 − x2)(1 − x4), x1x4(1 − x2)(1 − x3), x2x3(1 − x1)(1 − x4), x2x4(1 − x1)(1 − x3), x3x4(1 − x1)(1 − x2), x1x2x4(1 − x3), x1x3x4(1 − x2), x2x3x4(1 − x1), x1x2x3x4}(cid:105). Since x1x4 is a minimal divisor of generators in JC6 that vanishes on all codewords, x1x4 ∈ CF1(JC6). Similarly, x2x4, x3x4 ∈ CF1(JC6). Since 1110 ∈ C6, none of x1x2, x1x3, nor x1x2x3 is in JC6, and so CF1(JC6) = {x1x4, x2x4, x3x4}. Every pseudo-monomial in JC6 is a multiple of one of the monomials in CF1(JC6) except for x1(1− x2)(1− x3)(1− x4), x3(1− x1)(1− x2)(1− x4), and x2x3(1− x1)(1− x4). The minimal pseudo-monomials in JC6 that generate these are x3(1 − x1) and x1(1 − x2)(1 − x3), and so CF2(JC6) ⊇ {x3(1 − x1), x1(1 − x2)(1 − x3)}. In fact, one can check that this is the complete set of generators of CF2(JC) [5]. Thus, CF(JC6) = CF1(JC6) ∪ CF2(JC6) where CF1(JC6) = {x1x4, x2x4, x3x4} and CF2(JC6) = {x3(1 − x1), x1(1 − x2)(1 − x3)}. S2: CF-dectectable local obstructions Some local obstructions (σ, τ ) correspond to minimal RF relationships. Among those that do not, we distinguish local obstructions that can be "stripped down" (by removing neurons from σ and/or τ ) to local obstructions corresponding to minimal RF relationships. We refer to local obstructions that correspond to minimal RF relationships or that can be stripped down to such as CF-detectable local obstructions (precise definition was given in Definition 4.6). As we will see, both these types of local obstructions can be detected directly from CF(JC). Since every minimal RF relationship corresponds to a pseudo-monomial in CF(JC) (Lemma 1.4), Lemma 5.3 shows that all CF-detectable local obstructions can be determined solely from the canonical form. Lemma 5.3. Given a code C, the following are equivalent: (1) The link Lkσ(∆σ∪τ ) is contractible for every (σ, τ ) such that xσ (2) The link Lkσ(∆σ∪τ ) is contractible for every minimal RF relationship (σ, τ ) with τ (cid:54)= ∅, and (3) C has no CF-detectable local obstructions. i∈τ (1 − xi) ∈ CF2(JC), (cid:81) i∈τ (1 − xi) ∈ CF2(JC) if and only if (σ, τ ) Proof. It is clear that (1) and (2) are equivalent since xσ is a minimal RF relationship with τ (cid:54)= ∅ by Lemma 1.4 and the definition of CF2(JC). We now prove (2) ⇔ (3) by contrapositive. If C has a CF-detectable local obstruction (σ, τ ), then by definition there exist σ(cid:48) ⊆ σ and τ(cid:48) ⊆ τ such that (σ(cid:48), τ(cid:48)) is a minimal RF relationship that gives a local obstruction. Thus for that (σ(cid:48), τ(cid:48)), Lkσ(cid:48)(∆σ(cid:48)∪τ(cid:48)) is not contractible, and so (2) does not hold. Conversely, if there exists a minimal RF relationship (σ, τ ) such that Lkσ(∆σ∪τ ) is not contractible, then (σ, τ ) is itself a CF-detectable local obstruction, and so (3) does not hold. Theorem 5.4. If C has either of the algebraic signatures in rows C-1 or C-2 of Table 7, then C has no CF-detectable local obstructions. (cid:81) 21 (cid:81) (cid:81) Algebraic signature of JC ∀ xσ ∀ xσ ∃ i ∈ τ s.t. xixω /∈ CF1(JC) for all ω ⊆ σ ∪ τ i∈τ (1 − xi) ∈ CF2(JC), xσxτ /∈ JC i∈τ (1 − xi) ∈ CF2(JC), Property of C ⇒ no CF-detectable local obstructions ⇒ no CF-detectable local obstructions C-1 C-2 Table 7: Algebraic signatures of codes with no CF-detectable local obstructions. These codes are not guaranteed to be convex or non-convex, but do not have any local obstructions that can be detected from the canonical form. (C-2) We will show that signature C-2 guarantees that for every (σ, τ ) with xσ Proof. (C-1) Observe that xσxτ /∈ JC implies that xσxτ(cid:48) /∈ JC for all τ(cid:48) ⊆ τ since JC is an ideal. Thus, τ(cid:48) ∈ Lkσ(∆σ∪τ ) for all τ(cid:48) ⊆ τ , and so Lkσ(∆σ∪τ ) is the full simplex on the vertex set τ . i∈τ (1 − xi) ∈ CF2(JC), Thus Lkσ(∆σ∪τ ) is contractible. Since this holds for all (σ, τ ) such that xσ Lemma 5.3 guarantees that C has no CF-detectable local obstructions. i∈τ (1 − xi) ∈ CF2(JC), the link Lkσ(∆σ∪τ ) is a cone, and hence is contractible. Consider τ ∈ Lkσ(∆σ∪τ ), so that τ ⊆ τ and σ ∪ τ ∈ ∆(C). Since σ ∪ τ ∈ ∆(C), we have xσxτ /∈ JC. By hypothesis, there exists an i such that for every σ(cid:48) ⊆ σ and τ(cid:48) ⊆ τ , xixσ(cid:48)xτ(cid:48) /∈ CF1(JC). Since CF1(JC) generates the monomials of JC, this condition together with xσxτ /∈ JC guarantees that xixσxτ /∈ JC, and so {i}∪σ∪ τ ∈ ∆(C) implying that {i} ∪ τ ∈ Lkσ(∆σ∪τ ). Hence Lkσ(∆σ∪τ ) is a cone with respect to i, and so is contractible. Since i∈τ (1 − xi) ∈ CF2(JC), Lemma 5.3 guarantees that C has no this holds for all (σ, τ ) such that xσ CF-detectable local obstructions. (cid:81) (cid:81) (cid:81) As mentioned at the end of Section 4, C-1 is just a special case of C-2. Specifically, if a code satisfies C-1, then every i ∈ τ will satisfy the conditions of C-2, since C-1 guarantees that each link is a simplex, and thus also is a cone with any vertex acting as a cone point. We nevertheless include the proof of C-1 to clarify the structure of these links. 22
1912.02040
1
1912
2019-12-04T15:06:17
Universal principles justify the existence of concept cells
[ "q-bio.NC", "math.DS", "nlin.AO" ]
It is largely believed that complex cognitive phenomena require the perfect orchestrated collaboration of many neurons. However, this is not what converging experimental evidence suggests. Single neurons, the so-called concept cells, may be responsible for complex tasks performed by an individual. Here, starting from a few first principles, we layout physical foundations showing that concept cells are not only possible but highly likely, given that neurons work in a high dimensional space.
q-bio.NC
q-bio
Universal principles justify the existence of concept cells Carlos Calvo Tapia1, Ivan Tyukin2,3, Valeri A. Makarov1,3∗ 1Instituto de Matem´atica Interdisciplinar, Faculty of Mathematics, Universidad Complutense de Madrid, Plaza Ciencias 3, 28040 Madrid, Spain 2University of Leicester, Department of Mathematics, University Road, LE1 7RH, United Kingdom 3Lobachevsky State University of Nizhny Novgorod, Gagarin Ave. 23, 603950 Nizhny Novgorod, Russia (Dated: December 5, 2019) It is largely believed that complex cognitive phenomena require the perfect orchestrated collab- oration of many neurons. However, this is not what converging experimental evidence suggests. Single neurons, the so-called concept cells, may be responsible for complex tasks performed by an individual. Here, starting from a few first principles, we layout physical foundations showing that concept cells are not only possible but highly likely, given that neurons work in a high dimensional space. Brain is undoubtedly high-dimensional [1, 2]. Even in the simplest animal, the rotifer, it has 200 neurons acting as coupled dynamical systems, while in the human brain this figure rises to billions. The huge range of the number of neurons in different species has been related to the high variety of their cognitive abilities [3, 4]. Here, however, we assess the implication of another brain dimension, the number of synaptic inputs, n, a single neuron receives. Recent empirical evidence shows that a variation in the dendrite length and hence in the number of synapses n can explain up to 25% of the vari- ance in IQ scores between individuals [5]. However, no rigorous physical theory explaining how n affects high- level cognitive abilities has been put forward yet. The importance of such a theory and of the underlying universal principles is difficult to overestimate. For ex- ample, the design of modern artificial neural networks (ANNs) copy the converging architecture of biological sensory systems [6]. As a result, they already outper- form humans in pattern recognition benchmarks, yet re- maining far behind in cognition [7, 8]. Thus, the next qualitative leap requires novel biophysical insights on the functional architecture of higher brain stations. A step towards may reside in recent mathematical studies of the so-called "grandmother" cells [1, 9]. Con- verging experimental evidence suggests that some pyra- midal neurons in the medial temporal lobe can exhibit remarkable selectivity and invariance to complex stim- uli. In particular, it has been shown that the so-called concept cells (CCs) can fire when a subject sees one of 7 different pictures of Jennifer Aniston but not other 80 pictures of other persons and places [10]. CCs can also fire to the spoken or written name of the same person [11]. Thus, a single CC responds to an abstract concept but not to sensory features of the stimuli. This casts doubts on the widespread belief that complex cognitive phenomena require the perfectly orchestrated collabora- tion of many neurons. Moreover, CCs are relatively easily ∗ Corresponding author: [email protected] recorded in the hippocampus [12]. Thus, they must be abundant in the brain, contrary to the common opinion that their existence is highly unlikely [13]. Presumably, CCs play a role in episodic memory [11]. Memory formation and retrieval has been in the centre of attention for several decades, starting from the semi- nal Hopfield's work [14]. Recently, the linear scaling of the memory capacity with a low factor of 0.14 has been overcome [15]. Yet, as has been found, memory retrieval is inherently unstable due to complex network dynam- ics [16]. Thus, a "single" neuron approach to memory functions can also be useful. FIG. 1. Hypothesis of concept cells. (a) Model mimicking the information flow in the hippocampus. A stimulus (sound wave) activates the concept of a musical note. (b) Rearrange- ment of the neuronal "receptive fields" leads to formation of note-specific concept cells (different colors correspond to dif- ferent neurons). arXiv:1912.02040v1 [q-bio.NC] 4 Dec 2019 1 A concept stratum M z <latexit sha1_base64="HE4lduW21epI6Y7tpUnb8KglfIc=">AAAB9XicbVDLSgMxFL1TX7W+qi7dBIvgqsxUUZcFNy4r2Ae0Y8mkmTY0kwxJRqlD/8ONC0Xc+i/u/Bsz7Sy09UDI4Zx7yckJYs60cd1vp7Cyura+UdwsbW3v7O6V9w9aWiaK0CaRXKpOgDXlTNCmYYbTTqwojgJO28H4OvPbD1RpJsWdmcTUj/BQsJARbKx03wskH+hJZK/0adovV9yqOwNaJl5OKpCj0S9/9QaSJBEVhnCsdddzY+OnWBlGOJ2WeommMSZjPKRdSwWOqPbTWeopOrHKAIVS2SMMmqm/N1Ic6SyanYywGelFLxP/87qJCa/8lIk4MVSQ+UNhwpGRKKsADZiixPCJJZgoZrMiMsIKE2OLKtkSvMUvL5NWreqdVWu355X6RV5HEY7gGE7Bg0uoww00oAkEFDzDK7w5j86L8+58zEcLTr5zCH/gfP4AVAiTAw==</latexit> note !A = 440Hz <latexit sha1_base64="yway9rBsTNW9vBHC3mEq2jsxnO0=">AAAB/3icbVDLSsNAFJ34rPUVFdy4GSyCq5LUgm6EipsuK9gHNCFMppN26EwmzEzEGrvwV9y4UMStv+HOv3HaZqGtBy4czrmXe+8JE0aVdpxva2l5ZXVtvbBR3Nza3tm19/ZbSqQSkyYWTMhOiBRhNCZNTTUjnUQSxENG2uHweuK374hUVMS3epQQn6N+TCOKkTZSYB96gpM+Cq4uq1UHejwU91n9YRzYJafsTAEXiZuTEsjRCOwvrydwykmsMUNKdV0n0X6GpKaYkXHRSxVJEB6iPukaGiNOlJ9N7x/DE6P0YCSkqVjDqfp7IkNcqREPTSdHeqDmvYn4n9dNdXThZzROUk1iPFsUpQxqASdhwB6VBGs2MgRhSc2tEA+QRFibyIomBHf+5UXSqpTds3LlplqqVfI4CuAIHINT4IJzUAN10ABNgMEjeAav4M16sl6sd+tj1rpk5TMH4A+szx/Fl5VB</latexit> x <latexit sha1_base64="IJI05O5Ar/dlPNkVt1A//j0Tj3I=">AAAB9XicbVDLSgMxFL1TX7W+qi7dBIvgqsxUUZcFNy4r2Ae0Y8mkmTY0kwxJRi1D/8ONC0Xc+i/u/Bsz7Sy09UDI4Zx7yckJYs60cd1vp7Cyura+UdwsbW3v7O6V9w9aWiaK0CaRXKpOgDXlTNCmYYbTTqwojgJO28H4OvPbD1RpJsWdmcTUj/BQsJARbKx03wskH+hJZK/0adovV9yqOwNaJl5OKpCj0S9/9QaSJBEVhnCsdddzY+OnWBlGOJ2WeommMSZjPKRdSwWOqPbTWeopOrHKAIVS2SMMmqm/N1Ic6SyanYywGelFLxP/87qJCa/8lIk4MVSQ+UNhwpGRKKsADZiixPCJJZgoZrMiMsIKE2OLKtkSvMUvL5NWreqdVWu355X6RV5HEY7gGE7Bg0uoww00oAkEFDzDK7w5j86L8+58zEcLTr5zCH/gfP4AUP6TAQ==</latexit> stimulus 1 2 selective stratum m INITIAL concept stratum 831<latexit sha1_base64="z2WIefaj3ao11VfUDgzbbiiQiRc=">AAAB6nicbVDLTgJBEOzFF+IL9ehlIjHxRHbBRI5ELx4xyiOBDZkdemHC7OxmZtaEED7BiweN8eoXefNvHGAPClbSSaWqO91dQSK4Nq777eQ2Nre2d/K7hb39g8Oj4vFJS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj27nffkKleSwfzSRBP6JDyUPOqLHSQ63q9Yslt+wuQNaJl5ESZGj0i1+9QczSCKVhgmrd9dzE+FOqDGcCZ4VeqjGhbEyH2LVU0gi1P12cOiMXVhmQMFa2pCEL9ffElEZaT6LAdkbUjPSqNxf/87qpCWv+lMskNSjZclGYCmJiMv+bDLhCZsTEEsoUt7cSNqKKMmPTKdgQvNWX10mrUvaq5cr9Val+k8WRhzM4h0vw4BrqcAcNaAKDITzDK7w5wnlx3p2PZWvOyWZO4Q+czx9ok404</latexit> 784<latexit sha1_base64="XYR7uu7cfMjpVtZmypgBHfaOHWM=">AAAB6nicbVDLTgJBEOz1ifhCPXqZSEw8kV0kgSPRi0eM8khgQ2aHXpgwO7uZmTUhhE/w4kFjvPpF3vwbB9iDgpV0UqnqTndXkAiujet+OxubW9s7u7m9/P7B4dFx4eS0peNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPx7dxvP6HSPJaPZpKgH9Gh5CFn1FjpoVqr9AtFt+QuQNaJl5EiZGj0C1+9QczSCKVhgmrd9dzE+FOqDGcCZ/leqjGhbEyH2LVU0gi1P12cOiOXVhmQMFa2pCEL9ffElEZaT6LAdkbUjPSqNxf/87qpCWv+lMskNSjZclGYCmJiMv+bDLhCZsTEEsoUt7cSNqKKMmPTydsQvNWX10mrXPKuS+X7SrF+k8WRg3O4gCvwoAp1uIMGNIHBEJ7hFd4c4bw4787HsnXDyWbO4A+czx9zMo0/</latexit> t1<latexit sha1_base64="nDqZhfR1lMYiu0EYlDMoUhaBC74=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lqQY8FLx4r2g9oQ9lsN+3SzSbsToQS+hO8eFDEq7/Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzjwBuWKW3UXIOvEy0kFcjQH5a/+MGZpxBUySY3peW6CfkY1Cib5rNRPDU8om9AR71mqaMSNny1OnZELqwxJGGtbCslC/T2R0ciYaRTYzoji2Kx6c/E/r5dieONnQiUpcsWWi8JUEozJ/G8yFJozlFNLKNPC3krYmGrK0KZTsiF4qy+vk3at6l1Va/f1SqOWx1GEMziHS/DgGhpwB01oAYMRPMMrvDnSeXHenY9la8HJZ07hD5zPHwIWjZA=</latexit> tn<latexit sha1_base64="GOa1II0f0YM7QfMSUaX8pIyfW0I=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lqQY8FLx4r2g9oQ9lsN+3SzSbsToQS+hO8eFDEq7/Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzhQg3LFrboLkHXi5aQCOZqD8ld/GLM04gqZpMb0PDdBP6MaBZN8VuqnhieUTeiI9yxVNOLGzxanzsiFVYYkjLUthWSh/p7IaGTMNApsZ0RxbFa9ufif10sxvPEzoZIUuWLLRWEqCcZk/jcZCs0ZyqkllGlhbyVsTDVlaNMp2RC81ZfXSbtW9a6qtft6pVHL4yjCGZzDJXhwDQ24gya0gMEInuEV3hzpvDjvzseyteDkM6fwB87nD16Kjc0=</latexit> f1 s<latexit sha1_base64="z+t2abNGi30+DoxSXzRIJDgLunc=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoPgxbAbBT0GvXiMYB6QxDA76U2GzM6uM7NCWPITXjwo4tXf8ebfOEn2oIkFDUVVN91dfiy4Nq777eRWVtfWN/Kbha3tnd294v5BQ0eJYlhnkYhUy6caBZdYN9wIbMUKaegLbPqjm6nffEKleSTvzTjGbkgHkgecUWOlVtDTD+mZN+kVS27ZnYEsEy8jJchQ6xW/Ov2IJSFKwwTVuu25semmVBnOBE4KnURjTNmIDrBtqaQh6m46u3dCTqzSJ0GkbElDZurviZSGWo9D33aG1Az1ojcV//PaiQmuuimXcWJQsvmiIBHERGT6POlzhcyIsSWUKW5vJWxIFWXGRlSwIXiLLy+TRqXsnZcrdxel6nUWRx6O4BhOwYNLqMIt1KAODAQ8wyu8OY/Oi/PufMxbc042cwh/4Hz+AKy/j7o=</latexit> FINAL +⇡<latexit sha1_base64="jsWtubxZ0wRynetquFp0D4wfrq8=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBZBUEpSBT0WvHisYNpCG8pmu2mX7m7C7kYooX/BiwdFvPqHvPlv3LQ5aOuDgcd7M8zMCxPOtHHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFaE+iXmsuiHWlDNJfcMMp91EUSxCTjvh5C73O09UaRbLRzNNaCDwSLKIEWxy6aKfsEG15tbdOdAq8QpSgwKtQfWrP4xJKqg0hGOte56bmCDDyjDC6azSTzVNMJngEe1ZKrGgOsjmt87QmVWGKIqVLWnQXP09kWGh9VSEtlNgM9bLXi7+5/VSE90GGZNJaqgki0VRypGJUf44GjJFieFTSzBRzN6KyBgrTIyNp2JD8JZfXiXtRt27qjcermvNyyKOMpzAKZyDBzfQhHtogQ8ExvAMr/DmCOfFeXc+Fq0lp5g5hj9wPn8AsQKN8A==</latexit> +⇡<latexit sha1_base64="jsWtubxZ0wRynetquFp0D4wfrq8=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBZBUEpSBT0WvHisYNpCG8pmu2mX7m7C7kYooX/BiwdFvPqHvPlv3LQ5aOuDgcd7M8zMCxPOtHHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFaE+iXmsuiHWlDNJfcMMp91EUSxCTjvh5C73O09UaRbLRzNNaCDwSLKIEWxy6aKfsEG15tbdOdAq8QpSgwKtQfWrP4xJKqg0hGOte56bmCDDyjDC6azSTzVNMJngEe1ZKrGgOsjmt87QmVWGKIqVLWnQXP09kWGh9VSEtlNgM9bLXi7+5/VSE90GGZNJaqgki0VRypGJUf44GjJFieFTSzBRzN6KyBgrTIyNp2JD8JZfXiXtRt27qjcermvNyyKOMpzAKZyDBzfQhHtogQ8ExvAMr/DmCOfFeXc+Fq0lp5g5hj9wPn8AsQKN8A==</latexit> (a) y<latexit sha1_base64="51ub15LVnoLcX4rNyKM7u8P4jEQ=">AAAB9XicbVDLSgMxFL3js9ZX1aWbYBFclZkq6rLgxmUF+4B2LJlMpg3NJEOSUcrQ/3DjQhG3/os7/8ZMOwttPRByOOdecnKChDNtXPfbWVldW9/YLG2Vt3d29/YrB4dtLVNFaItILlU3wJpyJmjLMMNpN1EUxwGnnWB8k/udR6o0k+LeTBLqx3goWMQINlZ66AeSh3oS2yubTAeVqltzZ0DLxCtIFQo0B5WvfihJGlNhCMda9zw3MX6GlWGE02m5n2qaYDLGQ9qzVOCYaj+bpZ6iU6uEKJLKHmHQTP29keFY59HsZIzNSC96ufif10tNdO1nTCSpoYLMH4pSjoxEeQUoZIoSwyeWYKKYzYrICCtMjC2qbEvwFr+8TNr1mndeq99dVBuXRR0lOIYTOAMPrqABt9CEFhBQ8Ayv8OY8OS/Ou/MxH11xip0j+APn8wdSg5MC</latexit> (b) 831<latexit sha1_base64="z2WIefaj3ao11VfUDgzbbiiQiRc=">AAAB6nicbVDLTgJBEOzFF+IL9ehlIjHxRHbBRI5ELx4xyiOBDZkdemHC7OxmZtaEED7BiweN8eoXefNvHGAPClbSSaWqO91dQSK4Nq777eQ2Nre2d/K7hb39g8Oj4vFJS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj27nffkKleSwfzSRBP6JDyUPOqLHSQ63q9Yslt+wuQNaJl5ESZGj0i1+9QczSCKVhgmrd9dzE+FOqDGcCZ4VeqjGhbEyH2LVU0gi1P12cOiMXVhmQMFa2pCEL9ffElEZaT6LAdkbUjPSqNxf/87qpCWv+lMskNSjZclGYCmJiMv+bDLhCZsTEEsoUt7cSNqKKMmPTKdgQvNWX10mrUvaq5cr9Val+k8WRhzM4h0vw4BrqcAcNaAKDITzDK7w5wnlx3p2PZWvOyWZO4Q+czx9ok404</latexit> 784<latexit sha1_base64="XYR7uu7cfMjpVtZmypgBHfaOHWM=">AAAB6nicbVDLTgJBEOz1ifhCPXqZSEw8kV0kgSPRi0eM8khgQ2aHXpgwO7uZmTUhhE/w4kFjvPpF3vwbB9iDgpV0UqnqTndXkAiujet+OxubW9s7u7m9/P7B4dFx4eS0peNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPx7dxvP6HSPJaPZpKgH9Gh5CFn1FjpoVqr9AtFt+QuQNaJl5EiZGj0C1+9QczSCKVhgmrd9dzE+FOqDGcCZ/leqjGhbEyH2LVU0gi1P12cOiOXVhmQMFa2pCEL9ffElEZaT6LAdkbUjPSqNxf/87qpCWv+lMskNSjZclGYCmJiMv+bDLhCZsTEEsoUt7cSNqKKMmPTydsQvNWX10mrXPKuS+X7SrF+k8WRg3O4gCvwoAp1uIMGNIHBEJ7hFd4c4bw4787HsnXDyWbO4A+czx9zMo0/</latexit> 523<latexit sha1_base64="0vP3tuWhYBPmE1ZZlhqJVtxvoxM=">AAAB6nicbVDLTgJBEOzFF+IL9ehlIjHxRHZBo0eiF48Y5ZHAhswOszBhdnYz02tCCJ/gxYPGePWLvPk3DrAHBSvppFLVne6uIJHCoOt+O7m19Y3Nrfx2YWd3b/+geHjUNHGqGW+wWMa6HVDDpVC8gQIlbyea0yiQvBWMbmd+64lrI2L1iOOE+xEdKBEKRtFKD5eVaq9YcsvuHGSVeBkpQYZ6r/jV7ccsjbhCJqkxHc9N0J9QjYJJPi10U8MTykZ0wDuWKhpx40/mp07JmVX6JIy1LYVkrv6emNDImHEU2M6I4tAsezPxP6+TYnjtT4RKUuSKLRaFqSQYk9nfpC80ZyjHllCmhb2VsCHVlKFNp2BD8JZfXiXNStmrliv3F6XaTRZHHk7gFM7BgyuowR3UoQEMBvAMr/DmSOfFeXc+Fq05J5s5hj9wPn8AZYSNNg==</latexit> 494<latexit sha1_base64="rEHQWdBAdGindePqK3V/m0nzLqw=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoPgKezGgHoLevEY0TwgWcLspDcZMju7zMwKIeQTvHhQxKtf5M2/cZLsQRMLGoqqbrq7gkRwbVz328mtrW9sbuW3Czu7e/sHxcOjpo5TxbDBYhGrdkA1Ci6xYbgR2E4U0igQ2ApGtzO/9YRK81g+mnGCfkQHkoecUWOlh+p1tVcsuWV3DrJKvIyUIEO9V/zq9mOWRigNE1Trjucmxp9QZTgTOC10U40JZSM6wI6lkkao/cn81Ck5s0qfhLGyJQ2Zq78nJjTSehwFtjOiZqiXvZn4n9dJTXjlT7hMUoOSLRaFqSAmJrO/SZ8rZEaMLaFMcXsrYUOqKDM2nYINwVt+eZU0K2Xvoly5r5ZqN1kceTiBUzgHDy6hBndQhwYwGMAzvMKbI5wX5935WLTmnGzmGP7A+fwBcCWNPQ==</latexit> 466<latexit sha1_base64="U3d9HU3+W+utCo1pc7TOo3eM46E=">AAAB6nicbVDLTgJBEOz1ifhCPXqZSEw8kV0k6JHoxSNGeSSwIbPDLEyYnd3M9JoQwid48aAxXv0ib/6NA+xBwUo6qVR1p7srSKQw6Lrfztr6xubWdm4nv7u3f3BYODpumjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGtzO/9cS1EbF6xHHC/YgOlAgFo2ilh0q12isU3ZI7B1klXkaKkKHeK3x1+zFLI66QSWpMx3MT9CdUo2CST/Pd1PCEshEd8I6likbc+JP5qVNybpU+CWNtSyGZq78nJjQyZhwFtjOiODTL3kz8z+ukGF77E6GSFLlii0VhKgnGZPY36QvNGcqxJZRpYW8lbEg1ZWjTydsQvOWXV0mzXPIuS+X7SrF2k8WRg1M4gwvw4ApqcAd1aACDATzDK7w50nlx3p2PReuak82cwB84nz9uno08</latexit> 440<latexit sha1_base64="I8p99MXj3nTAzfIRQez5/FO/EFU=">AAAB6nicbVBNS8NAEJ34WetX1aOXxSJ4Kkkt6LHoxWNF+wFtKJvtpF262YTdjVBCf4IXD4p49Rd589+4bXPQ1gcDj/dmmJkXJIJr47rfztr6xubWdmGnuLu3f3BYOjpu6ThVDJssFrHqBFSj4BKbhhuBnUQhjQKB7WB8O/PbT6g0j+WjmSToR3QoecgZNVZ6qNXcfqnsVtw5yCrxclKGHI1+6as3iFkaoTRMUK27npsYP6PKcCZwWuylGhPKxnSIXUsljVD72fzUKTm3yoCEsbIlDZmrvycyGmk9iQLbGVEz0sveTPzP66YmvPYzLpPUoGSLRWEqiInJ7G8y4AqZERNLKFPc3krYiCrKjE2naEPwll9eJa1qxbusVO9r5fpNHkcBTuEMLsCDK6jDHTSgCQyG8Ayv8OYI58V5dz4WrWtOPnMCf+B8/gBifI00</latexit> frequency (Hz) ⇡<latexit sha1_base64="QIXphsQ/mi/7rGW0zYsZvRqA8Ow=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBbBg5akCnosePFYwbSFNpTNdtMu3d2E3Y1QQv+CFw+KePUPefPfuGlz0NYHA4/3ZpiZFyacaeO6305pbX1jc6u8XdnZ3ds/qB4etXWcKkJ9EvNYdUOsKWeS+oYZTruJoliEnHbCyV3ud56o0iyWj2aa0EDgkWQRI9jk0mU/YYNqza27c6BV4hWkBgVag+pXfxiTVFBpCMda9zw3MUGGlWGE01mln2qaYDLBI9qzVGJBdZDNb52hM6sMURQrW9Kgufp7IsNC66kIbafAZqyXvVz8z+ulJroNMiaT1FBJFouilCMTo/xxNGSKEsOnlmCimL0VkTFWmBgbT8WG4C2/vErajbp3VW88XNeaF0UcZTiBUzgHD26gCffQAh8IjOEZXuHNEc6L8+58LFpLTjFzDH/gfP4AtBCN8g==</latexit> 0<latexit sha1_base64="KJa6pDgxxeogIDVidSgL+lobpZs=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBg5SkCnosePHYgq2FNpTNdtKu3WzC7kYoob/AiwdFvPqTvPlv3LY5aOuDgcd7M8zMCxLBtXHdb6ewtr6xuVXcLu3s7u0flA+P2jpOFcMWi0WsOgHVKLjEluFGYCdRSKNA4EMwvp35D0+oNI/lvZkk6Ed0KHnIGTVWarr9csWtunOQVeLlpAI5Gv3yV28QszRCaZigWnc9NzF+RpXhTOC01Es1JpSN6RC7lkoaofaz+aFTcmaVAQljZUsaMld/T2Q00noSBbYzomakl72Z+J/XTU1442dcJqlByRaLwlQQE5PZ12TAFTIjJpZQpri9lbARVZQZm03JhuAtv7xK2rWqd1mtNa8q9Ys8jiKcwCmcgwfXUIc7aEALGCA8wyu8OY/Oi/PufCxaC04+cwx/4Hz+AHPJjKI=</latexit> phase +⇡<latexit sha1_base64="jsWtubxZ0wRynetquFp0D4wfrq8=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBZBUEpSBT0WvHisYNpCG8pmu2mX7m7C7kYooX/BiwdFvPqHvPlv3LQ5aOuDgcd7M8zMCxPOtHHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFaE+iXmsuiHWlDNJfcMMp91EUSxCTjvh5C73O09UaRbLRzNNaCDwSLKIEWxy6aKfsEG15tbdOdAq8QpSgwKtQfWrP4xJKqg0hGOte56bmCDDyjDC6azSTzVNMJngEe1ZKrGgOsjmt87QmVWGKIqVLWnQXP09kWGh9VSEtlNgM9bLXi7+5/VSE90GGZNJaqgki0VRypGJUf44GjJFieFTSzBRzN6KyBgrTIyNp2JD8JZfXiXtRt27qjcermvNyyKOMpzAKZyDBzfQhHtogQ8ExvAMr/DmCOfFeXc+Fq0lp5g5hj9wPn8AsQKN8A==</latexit> selective stratum ⇡<latexit sha1_base64="QIXphsQ/mi/7rGW0zYsZvRqA8Ow=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBbBg5akCnosePFYwbSFNpTNdtMu3d2E3Y1QQv+CFw+KePUPefPfuGlz0NYHA4/3ZpiZFyacaeO6305pbX1jc6u8XdnZ3ds/qB4etXWcKkJ9EvNYdUOsKWeS+oYZTruJoliEnHbCyV3ud56o0iyWj2aa0EDgkWQRI9jk0mU/YYNqza27c6BV4hWkBgVag+pXfxiTVFBpCMda9zw3MUGGlWGE01mln2qaYDLBI9qzVGJBdZDNb52hM6sMURQrW9Kgufp7IsNC66kIbafAZqyXvVz8z+ulJroNMiaT1FBJFouilCMTo/xxNGSKEsOnlmCimL0VkTFWmBgbT8WG4C2/vErajbp3VW88XNeaF0UcZTiBUzgHD26gCffQAh8IjOEZXuHNEc6L8+58LFpLTjFzDH/gfP4AtBCN8g==</latexit> 0<latexit sha1_base64="KJa6pDgxxeogIDVidSgL+lobpZs=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBg5SkCnosePHYgq2FNpTNdtKu3WzC7kYoob/AiwdFvPqTvPlv3LY5aOuDgcd7M8zMCxLBtXHdb6ewtr6xuVXcLu3s7u0flA+P2jpOFcMWi0WsOgHVKLjEluFGYCdRSKNA4EMwvp35D0+oNI/lvZkk6Ed0KHnIGTVWarr9csWtunOQVeLlpAI5Gv3yV28QszRCaZigWnc9NzF+RpXhTOC01Es1JpSN6RC7lkoaofaz+aFTcmaVAQljZUsaMld/T2Q00noSBbYzomakl72Z+J/XTU1442dcJqlByRaLwlQQE5PZ12TAFTIjJpZQpri9lbARVZQZm03JhuAtv7xK2rWqd1mtNa8q9Ys8jiKcwCmcgwfXUIc7aEALGCA8wyu8OY/Oi/PufCxaC04+cwx/4Hz+AHPJjKI=</latexit> phase +⇡<latexit sha1_base64="jsWtubxZ0wRynetquFp0D4wfrq8=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBZBUEpSBT0WvHisYNpCG8pmu2mX7m7C7kYooX/BiwdFvPqHvPlv3LQ5aOuDgcd7M8zMCxPOtHHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFaE+iXmsuiHWlDNJfcMMp91EUSxCTjvh5C73O09UaRbLRzNNaCDwSLKIEWxy6aKfsEG15tbdOdAq8QpSgwKtQfWrP4xJKqg0hGOte56bmCDDyjDC6azSTzVNMJngEe1ZKrGgOsjmt87QmVWGKIqVLWnQXP09kWGh9VSEtlNgM9bLXi7+5/VSE90GGZNJaqgki0VRypGJUf44GjJFieFTSzBRzN6KyBgrTIyNp2JD8JZfXiXtRt27qjcermvNyyKOMpzAKZyDBzfQhHtogQ8ExvAMr/DmCOfFeXc+Fq0lp5g5hj9wPn8AsQKN8A==</latexit> In this letter, we report the first theoretical justifica- tion of the existence of CCs. It is based on a few first principles, and the neuronal dimension n is the major factor. We suggest that the evolution of neurons led to an increase of their input dimension n, which trig- gered a qualitative leap in their function and emergence of episodic memory. Figure 1(a) illustrates the model mimicking basic sig- naling pathways in the hippocampus. It takes into ac- count the stratified structure of the hippocampus that facilitates ramification of axons conveying the same high- dimensional input to multiple pyramidal cells, which is supported by electrophysiological observations of spatial modules of coherent activity [17]. For simplicity, we con- sider one "selective" and one "concept" neural strata only, without connections between neurons within the strata. Moreover, as we will see below, learning is local- ized within individual neurons. It is unsupervised and hence no global fitness function usually used in ANNs is required. Thus, the model discards all a priori assump- tions on the local network structure and dynamics. The first stratum contains m neurons receiving in a sequence L n-dimensional (nD) stimuli xi ∈ [−1, 1]n (i = In general, 1, 2, . . . , L), e.g. sound waves (Fig. 1(a)). m (cid:29) L (e.g. in the CA1 region of the hippocampus there are 1.4 × 107 pyramidal cells). The mD output from the first stratum yi, as a response to stimulus xi, goes to the second stratum. There, K consecutive signals {yi} can overlap in time due to short-term memory, implemented through e.g. synaptic integration, and we get output z. We note that neurons in the concept stratum asso- ciate several items and then respond to groups of stim- uli, which form concepts (in our case, concepts of musical notes). Stimuli within a group can be uncorrelated and even represent different sensory modalities, which gives rise to complex concepts as experimentally observed [11]. Although the nature of stimuli xi can be arbitrary, we illustrate this model on a simple example of acquiring "musical memory". To follow music, the system must be able to recognize unambiguously the tones of sound or notes. To preserve generality, we do not apply any algo- rithmic pre-processing of sound signals, largely extended in ANNs. A piece of a sound wave sampled at fs = 213 Hz is represented as a "raw" nD stimulus (Fig. 1(a)): (1) xi =(cid:0)Ai cos(2πfit/fs + φi)(cid:1)n t=1, where Ai, fi, and φi are the amplitude, frequency, and phase, respectively. Let's assume that Ai is fixed, i.e., the amplitude is normalized by a sensory organ. Then, Ω = {(f, φ)} defines the set of primary stimuli. In this set, musical note A corresponds to frequency 440 Hz, i.e., to the subset ωA = {(440, φ) : ∀φ} ⊂ Ω. At the beginning, all neurons in both strata are initial- ized randomly and hence their "receptive fields" (areas in the sensory domain Ω invoking a response of a neuron) form a disordered mixture of random regions (see the car- toon in Fig. 1(b), left). Thus, the output of the concept stratum is random and the system cannot follow music. 2 The purpose of learning is to organize receptive fields in such a way that the concept cells become note-specific, i.e., fire on a presentation of a given tone regardless of its phase (Fig. 1(b), right). In this case, each concept cell will not be a stimulus-specific but represent a set of associated stimuli or a concept, e.g. note A. To enable such a learning, we need at least a two- stratum system. Neurons in the sensory stratum learn to respond selectively to all sound waves, while neurons in the concept stratum associate stimuli with different phases but with the same frequency. Such an association cannot be done within the first stratum, since raw signals can be anti-correlated, e.g. φ1 = 0 and φ2 = π and then cancel each other on a neuron x1 + x2 = 0. We now assess the implication of the input neuronal di- mensions n and m on the emergence of concept cells. All neurons in both strata are described by the same model, which captures the threshold nature of the neuronal ac- tivation but disregards the dynamics of spike generation. The response of the j-th neuron yj(t) in the selective stratum to the external input sext(t) is given by: sext = xiσik(t), L i=1Xk r 3 X n (2a) (2b) (2c) yj = H(vj − θj), vj = hwj, sexti, wj = αyj(cid:0)β2sext − vjwj(cid:1) , where σik(t) are disjoint rectangular time windows defin- ing the k-th appearance of the i-th stimulus, H(u) = max{0, u} is the transfer function, vj(t) is the membrane potential, θj ≥ 0 is the "firing" threshold, wj(t) ∈ Rn is the vector of the synaptic weights, h·,·i is the standard inner product, α > 0 defines the relaxation time, and β > 0 is an order parameter that will be defined later. Equation (2c) simulates the Hebbian type of synaptic plasticity. The term yjsext forces plastic changes when a stimulus evokes a non-zero neuronal response only, simi- lar to the classical Oja rule [18]. The second term ensures boundness of wj to conform with physical plausibility. At t = 0, the synaptic weights of all neurons, wj(0), are randomly initialized in the hypercube U n([−1, 1]). The threshold values θj can also be chosen arbitrary. Then, neuron j "fires" at the presentation of stimulus xi if its membrane potential is higher than the threshold, vj > θj. In this case, we say that the neuron detects the stimulus. Let dj ∈ {0, 1, . . . , L} be the number of stimuli the j- th neuron can detect. Then, if dj = 0, the neuron is inactive for sext, it is selective if dj = 1, and non-selective otherwise. To quantify the performance of the selective stratum, we introduce the ratios of selective neurons Rslctv (i.e., the number of selective neurons over m), inactive neu- rons Rinact, and "lost" stimuli Rlost (stimuli that ex- cite no neurons). To estimate the expected values of these indexes, we note that a random stimulus xi, taken from U n([−1, 1]), elicits random membrane potential in each neuron, which will be normally distributed as 3 (4) 2Φ−1(psl) √5n . , δ =s1 − θ δ βsl = by neurons. At t = 0, all neurons in the aggregate are not selective and respond in average to d = 25 stimuli, while at t = 80 they are absolutely selective, d = 1. To extend this numerical observation, we first find the condition that a neuron, started firing to a stimulus xi, keeps firing in forward time with a probability no smaller than some constant 0 < psl < 1. This condition is fulfilled by choosing the order parameter [20]: Note that the higher the neuronal dimension n, the higher psl can be chosen. Under condition (4) the synaptic weights converge [20]: ), up to an error term of order O(1/√n). vj ∼ N (0, 1√3 For n large enough (n & 10 − 20), the error decays ex- ponentially [19, 20]. Then, we can estimate the firing probability P(vj > θ) = 1 − Φ(√3θ) [20], where Φ(·) is the normal cumulative distribution function. By using a binomial distribution, we get: Rslctv =L(1 − Φ(√3θ))Φ(√3θ)L−1, ¯ Rinact =Φ(√3θ)L, ¯ Rlost = Φ(√3θ)m. ¯ (3) w(t) = βsl lim t→∞ xi kxik . (5) Thus, given that α is large enough, learning forces the neuron to "align" along its "preferable" stimulus: w∞ (cid:20) xi. At the same time, for high n, we have the property: hxi, xki ≈ 0, i 6= k (for details see e.g. [9]). Thus, after a transient, the neuronal membrane potential will be close to zero for all stimuli except xi and hence the neuron will become selective. To estimate the probability that a neuron will be se- lective after learning, i.e. S := P(dj = 1), we evaluate the probability that the neuron will be silent to another arbitrary stimulus xk (k 6= i) [20]: Φ(δ√ns)κ(s; µ, σ) ds, (6) P =Z ∞ 0 FIG. 2. Poor initial brain performance and the emergence of selectivity by learning. (a) Performance indexes [Eq. (3), thick curves] of the selective stratum at t = 0. Insets show raster plots of stimuli detected by neurons (m = 300, L = 100). (b) Median number of stimuli detected by a neuron, d, vs time (n = 30, L = 400, θ = 0.5, α = 20, psl = 0.95). Figure 2(a) illustrates the performance measures and raster plots of stimuli detected by neurons (a black dot at position (i, j) means that neuron j detects stimulus i). The ratio of selective neurons has a modest peak of height e−1 at θ∗ = 1√3 t = 0 a randomly initialized selective stratum can have at most 37% of selective neurons, independently on n. Thus, the first universal property of different "brains" is their poor initial performance, regardless of the neuronal dimension n. L (cid:1) ≈ 1.35. Therefore, at Φ−1(cid:0) L−1 As we show now, learning can dramatically improve the performance. We choose the firing threshold small enough, e.g. θ = 1. Then, with high probability, all neurons are active, i.e., dj ≥ 1, and there are no lost stimuli (Fig. 2(a)), i.e., Hebbian learning (2c) is activated for all neurons and all stimuli. Figure 2(b) illustrates the dynamics of the median number of the stimuli detected where κ(·; µ, σ) is the normal pdf with the mean µ = 1 and the standard deviation σ = 2√5n . Note that with an increase of n, κ concentrates around 1 and we can roughly evaluate P ≈ Φ(δ√n), which tends to 1 for high n. Finally, the neuronal selectivity, S(n, L) = P L−1, (7) depends on the number of stimuli L and the neuronal dimension n only. Figure 3(a) shows the selectivity S as a function of the neuronal dimension n. Learning yields a step-like depen- dence of S on n. For small n, there is no any improvement of the selectivity by learning (S ≈ 0), while for higher n, it rapidly reaches 100%. Insets in Fig. 3(a) illustrate an example of raster plots of stimuli detected by the neurons at the beginning and the end of learning. We observe that almost all neurons become selective to single information items (area shadowed by blue). Thus, the second univer- sal property is that an increase in the input dimension n provokes an explosive emergence of selective behavior in "brains" composed of high-dimensional neurons at criti- cal dimensions 10 -- 20. From Eq. (7) we can also estimate the maximal num- ber of stimuli that a big enough stratum can work with: Lmax = 1 + ln(pL) ln(P) , (8) 100% 0% 100% neuron 100% 0% neuron 100% 0% stimulus 0.6 0.4 0.2 0% stimulus selective neur. inactive neur. lost stimuli (a) 0.8 performance indexes 20 40 time (a.u.) 60 0.4 0.6 0.8 1 firing threshold, 1.2 1.4 1.6 median, d<latexit sha1_base64="uys1wdrAZQgcCw/FzUwyBxxqmtE=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBg5SkCnosePHYgq2FNpTNZtKu3WzC7kYopb/AiwdFvPqTvPlv3LY5aOuDgcd7M8zMC1LBtXHdb6ewtr6xuVXcLu3s7u0flA+P2jrJFMMWS0SiOgHVKLjEluFGYCdVSONA4EMwup35D0+oNE/kvRmn6Md0IHnEGTVWaob9csWtunOQVeLlpAI5Gv3yVy9MWBajNExQrbuemxp/QpXhTOC01Ms0ppSN6AC7lkoao/Yn80On5MwqIYkSZUsaMld/T0xorPU4DmxnTM1QL3sz8T+vm5noxp9wmWYGJVssijJBTEJmX5OQK2RGjC2hTHF7K2FDqigzNpuSDcFbfnmVtGtV77Jaa15V6hd5HEU4gVM4Bw+uoQ530IAWMEB4hld4cx6dF+fd+Vi0Fpx85hj+wPn8AcKZjNY=</latexit> (25,75)% interval selective, d = 1 <latexit sha1_base64="8R8itZeK5GP34VFiLbwtjdUbEkU=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBg5SkCnoRCl48VjBtoQ1ls9m0S3c3YXcjlNDf4MWDIl79Qd78N27bHLT1wcDjvRlm5oUpZ9q47rdTWlvf2Nwqb1d2dvf2D6qHR22dZIpQnyQ8Ud0Qa8qZpL5hhtNuqigWIaedcHw38ztPVGmWyEczSWkg8FCymBFsrORH6BZ5g2rNrbtzoFXiFaQGBVqD6lc/SkgmqDSEY617npuaIMfKMMLptNLPNE0xGeMh7VkqsaA6yOfHTtGZVSIUJ8qWNGiu/p7IsdB6IkLbKbAZ6WVvJv7n9TIT3wQ5k2lmqCSLRXHGkUnQ7HMUMUWJ4RNLMFHM3orICCtMjM2nYkPwll9eJe1G3busNx6uas2LIo4ynMApnIMH19CEe2iBDwQYPMMrvDnSeXHenY9Fa8kpZo7hD5zPH2CVjaw=</latexit> 30 20 10 (b) detected stimuli,d <latexit sha1_base64="B1AteRWHl5459edupMcsDkxCD58=">AAAB6HicbVBNS8NAEJ34WetX1aOXxSJ4Kkkt6LHgxWML9gPaUDababt2swm7G6GE/gIvHhTx6k/y5r9x2+agrQ8GHu/NMDMvSATXxnW/nY3Nre2d3cJecf/g8Oi4dHLa1nGqGLZYLGLVDahGwSW2DDcCu4lCGgUCO8Hkbu53nlBpHssHM03Qj+hI8iFn1FipGQ5KZbfiLkDWiZeTMuRoDEpf/TBmaYTSMEG17nluYvyMKsOZwFmxn2pMKJvQEfYslTRC7WeLQ2fk0iohGcbKljRkof6eyGik9TQKbGdEzVivenPxP6+XmuGtn3GZpAYlWy4apoKYmMy/JiFXyIyYWkKZ4vZWwsZUUWZsNkUbgrf68jppVyvedaXarJXrtTyOApzDBVyBBzdQh3toQAsYIDzDK7w5j86L8+58LFs3nHzmDP7A+fwBxQGM3g==</latexit> 4 representing coherent stimuli (sound waves indistinguish- able for neurons due to some tolerance). Thus, neurons in the selective stratum learn individual sound waves and we observed spontaneous formation of neuronal "clusters" (Fig. 4(b)). Neurons within a cluster detect sound waves with different phases corresponding to a single note, while rejecting the other stimuli. Let us now consider the second stratum composed of concept cells (Fig. 1(a)). The dynamics of concept cells is also described by Eq. (2) but now as an input we use the output from the first stratum y ∈ Rm + within one time window: sint(t) = K X i=1 yiχi(t), t ∈ [0, K∆], (9) where χi(t) are overlapping rectangular time windows [(i−1)∆, K∆). At the stratum output z, we then expect to obtain codification of concepts, which are associations of K individual stimuli. The coding is now sparse. Only a little portion of the neurons in the selective stratum responds to a single stimulus xi, i.e., supp(yi) (cid:28) m. Besides, yi ≥ 0 and hence hyj, yii ≥ 0. We also assume that after learning neurons in the first stratum are selective, i.e., dj = 1, j = 1, . . . , m. Then, supp(yj) ∩ supp(yi) = ∅ and hence hyj, yii = 0 (j 6= i), which facilitates learning. Repeating similar arguments for selecting the order pa- rameter β as provided above, after tedious calculations [20], we find: βcn = θcn√LδKΓ(K + 1 2 ) θsl(1 − pcn)(1 − δ)(K − 1)!√m , (10) where Γ is the gamma function and pcn is the probability that after learning, a neuron will fire to all K stimuli, i.e., will become a concept cell. Inequality (10) yields the following approximate con- dition on the neuronal dimension of CCs: m ∝ K 3/β2 cn. (11) Figure 5(a) shows how the number of stimuli K, which can be associated in a concept, scales with the neuronal dimension m. An increase in the association depth K re- quires cubic increase of the input dimension m, which can be balanced by an increase of the order parameter βcn. In consequence, overloaded associations with high K can result in detection of "wrong" stimuli as being within a concept. Such an observation has been reported experi- mentally, when a Jennifer Aniston neuron the next day also detected Lisa Kudrow from the TV series "Friends" [12]. Figure 5(b) illustrates the process of formation of selec- tive and concept cells responding to musical notes. At the beginning, both selective and concept cells have messy re- ceptive fields (see also Fig. 4(a)). Learning organizes the strata as it was advanced in Fig. 1(b). Finally, concept cells are activated to particular notes regardless of the FIG. 3. Emergence of extreme selectivity in high-dimensional brains. (a) Step-like increase of the ratio of selective neurons (experiment: black circles, estimate (7): red curve). Insets show paster plots of detected stimuli for n = 30 (compare to Fig. 2(a)). (b) Exponential growth of the memory capacity (experiment: blue circles, theoretical estimate: red curve). Green area marks the working zone. where pL is the lower bound of the probability that the stratum detects all L stimuli. Figure 3(b) shows the theo- retical and experimental estimates of the stratum capac- ity. Even for a rather moderate dimension n = 60, the capacity goes beyond 1010 (numerical estimate beyond n = 30 was not calculated due to exponential growth of the computational load). In practical terms, it means that a big enough "brain" consisting of high-dimensional neurons can selectively detect all stimuli existing in the world. FIG. 4. Learning musical stimuli. (a) Receptive fields of two neurons in the "sound" space before and after learning. (b) Raster plot at t = 0 (top) shows random response of the stratum to 48 stimuli representing 12 musical notes from A to G#. After learning (bottom), neurons have grouped into clusters selective to individual stimuli. To illustrate how the selective stratum can deal with "real-world" stimuli, we simulated learning of 48 sound waves corresponding to 12 musical notes from A to G# [see Eq. (1) and Fig. 1]. Figure 4(a) shows the recep- tive fields of two arbitrary chosen neurons before and after learning. At the beginning, the neurons have wide random receptive fields, as it was hypothesized (see Fig. 1(b)). Learning reduces the receptive field to tiny elipses Computed Eq. (8) (n, L ) = (30,400) <latexit sha1_base64="VYfkOlQaFQQSOdLS7P4xH7eBXd0=">AAAB+XicbVBNS8NAEJ34WetX1KOXxSKkUEpiC3oRCl48eKhgP6ANZbPdtks3m7C7KZTQf+LFgyJe/Sfe/Ddu2xy09cHA470ZZuYFMWdKu+63tbG5tb2zm9vL7x8cHh3bJ6dNFSWS0AaJeCTbAVaUM0EbmmlO27GkOAw4bQXju7nfmlCpWCSe9DSmfoiHgg0YwdpIPdt2ROmhiG6RU3FLVdct9uyCW3YXQOvEy0gBMtR79le3H5EkpEITjpXqeG6s/RRLzQins3w3UTTGZIyHtGOowCFVfrq4fIYujdJHg0iaEhot1N8TKQ6VmoaB6QyxHqlVby7+53USPbjxUybiRFNBlosGCUc6QvMYUJ9JSjSfGoKJZOZWREZYYqJNWHkTgrf68jppXpW9Stl7rBZqThZHDs7hAhzw4BpqcA91aACBCTzDK7xZqfVivVsfy9YNK5s5gz+wPn8ArfSQXA==</latexit> 20 60 neuronal dimension, n<latexit sha1_base64="jxEcJ/8byvx9+0M9USZVv2gLxRQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI8FLx5bsB/QhrLZTtq1m03Y3Qgl9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RS2tnd294r7pYPDo+OT8ulZR8epYthmsYhVL6AaBZfYNtwI7CUKaRQI7AbTu4XffUKleSwfzCxBP6JjyUPOqLFSSw7LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JlVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VaN5VGLY+jCBdwCdfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx/SQYzi</latexit> 40 <latexit sha1_base64="jxEcJ/8byvx9+0M9USZVv2gLxRQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI8FLx5bsB/QhrLZTtq1m03Y3Qgl9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RS2tnd294r7pYPDo+OT8ulZR8epYthmsYhVL6AaBZfYNtwI7CUKaRQI7AbTu4XffUKleSwfzCxBP6JjyUPOqLFSSw7LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JlVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VaN5VGLY+jCBdwCdfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx/SQYzi</latexit> <latexit sha1_base64="jxEcJ/8byvx9+0M9USZVv2gLxRQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI8FLx5bsB/QhrLZTtq1m03Y3Qgl9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RS2tnd294r7pYPDo+OT8ulZR8epYthmsYhVL6AaBZfYNtwI7CUKaRQI7AbTu4XffUKleSwfzCxBP6JjyUPOqLFSSw7LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JlVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VaN5VGLY+jCBdwCdfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx/SQYzi</latexit> 10 8 10 6 10 4 10 2 stratum capacity (b) 40 20 neuronal dimension, n<latexit sha1_base64="jxEcJ/8byvx9+0M9USZVv2gLxRQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI8FLx5bsB/QhrLZTtq1m03Y3Qgl9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RS2tnd294r7pYPDo+OT8ulZR8epYthmsYhVL6AaBZfYNtwI7CUKaRQI7AbTu4XffUKleSwfzCxBP6JjyUPOqLFSSw7LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JlVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VaN5VGLY+jCBdwCdfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx/SQYzi</latexit> <latexit sha1_base64="jxEcJ/8byvx9+0M9USZVv2gLxRQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI8FLx5bsB/QhrLZTtq1m03Y3Qgl9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RS2tnd294r7pYPDo+OT8ulZR8epYthmsYhVL6AaBZfYNtwI7CUKaRQI7AbTu4XffUKleSwfzCxBP6JjyUPOqLFSSw7LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JlVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VaN5VGLY+jCBdwCdfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx/SQYzi</latexit> <latexit sha1_base64="jxEcJ/8byvx9+0M9USZVv2gLxRQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI8FLx5bsB/QhrLZTtq1m03Y3Qgl9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RS2tnd294r7pYPDo+OT8ulZR8epYthmsYhVL6AaBZfYNtwI7CUKaRQI7AbTu4XffUKleSwfzCxBP6JjyUPOqLFSSw7LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JlVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VaN5VGLY+jCBdwCdfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx/SQYzi</latexit> n = 30 <latexit sha1_base64="ouqFOzVZMexJ+cFOYX7u31JFsU8=">AAAB7XicbVBNSwMxEJ34WetX1aOXYBE8SNm1gl6EghePFewHtEvJptk2NpssSVYoS/+DFw+KePX/ePPfmLZ70NYHA4/3ZpiZFyaCG+t532hldW19Y7OwVdze2d3bLx0cNo1KNWUNqoTS7ZAYJrhkDcutYO1EMxKHgrXC0e3Ubz0xbbiSD3acsCAmA8kjTol1UlPiG1z1eqWyV/FmwMvEz0kZctR7pa9uX9E0ZtJSQYzp+F5ig4xoy6lgk2I3NSwhdEQGrOOoJDEzQTa7doJPndLHkdKupMUz9fdERmJjxnHoOmNih2bRm4r/eZ3URtdBxmWSWibpfFGUCmwVnr6O+1wzasXYEUI1d7diOiSaUOsCKroQ/MWXl0nzouJXK/79Zbl2nsdRgGM4gTPw4QpqcAd1aACFR3iGV3hDCr2gd/Qxb11B+cwR/AH6/AHhZY3x</latexit> INITIAL FINAL (a) 1 0.8 0.6 0.4 0.2 selective rate 1 41 neuron 81 121 INITIAL # FINAL # GG EFF DD ABCC EFF DD ABCC # # A# # GG # # # A# (b) stimulus (a) 1.2 INITIAL FINAL 1.1 1.0 0.9 1.2 1.1 1.0 frequency (kHz) 0.9 ⇡<latexit sha1_base64="mSebUVwgEPOjUmah8FjR81TxX+8=">AAAB63icbVBNSwMxEJ34WetX1aOXYBG8WHaroMeCF48V7Ae0S8mm2TY0yS5JVihL/4IXD4p49Q9589+YbfegrQ8GHu/NMDMvTAQ31vO+0dr6xubWdmmnvLu3f3BYOTpumzjVlLVoLGLdDYlhgivWstwK1k00IzIUrBNO7nK/88S04bF6tNOEBZKMFI84JTaXLvsJH1SqXs2bA68SvyBVKNAcVL76w5imkilLBTGm53uJDTKiLaeCzcr91LCE0AkZsZ6jikhmgmx+6wyfO2WIo1i7UhbP1d8TGZHGTGXoOiWxY7Ps5eJ/Xi+10W2QcZWklim6WBSlAtsY54/jIdeMWjF1hFDN3a2Yjokm1Lp4yi4Ef/nlVdKu1/yrWv3hutqoF3GU4BTO4AJ8uIEG3EMTWkBhDM/wCm9Iohf0jj4WrWuomDmBP0CfP7Xejfg=</latexit> 0 phase ⇡<latexit sha1_base64="06QdeKKY9Q7Fl56iS0XAFf7L6Rk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lqQY8FLx4r2g9oQ9lsJ+3SzSbsboQS+hO8eFDEq7/Im//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ7dzvPKHSPJaPZpqgH9GR5CFn1FjpoZ/wQbniVt0FyDrxclKBHM1B+as/jFkaoTRMUK17npsYP6PKcCZwVuqnGhPKJnSEPUsljVD72eLUGbmwypCEsbIlDVmovycyGmk9jQLbGVEz1qveXPzP66UmvPEzLpPUoGTLRWEqiInJ/G8y5AqZEVNLKFPc3krYmCrKjE2nZEPwVl9eJ+1a1buq1u7rlUYtj6MIZ3AOl+DBNTTgDprQAgYjeIZXeHOE8+K8Ox/L1oKTz5zCHzifP0w7jcE=</latexit> 5 FIG. 5. Emergence of concept cells. (a) Working zone (shadowed by green) for association of stimuli in concepts. Insets show raster plots of the concept cells' response: a random mixture in the red zone and correct association in the green zone. (b) Formation of "musical memory". Receptive fields in the selective and concept strata are organized into wave and note specific structures, respectively [see also the hypothesis in Fig. 1(b)]. (c) Perception of a fragment of the 9-th Symphony by Beethoven: "Ode to joy". phase of sound waves. Thus, the "brain" can now follow music. Figure 5(c) illustrates the system response to a song. As expected, the selective stratum detects individ- ual sound waves, while the concept stratum puts them together and forms the note-specific output. In conclusion, we have shown that the emergence of concept cells is conditioned by the synaptic (input) di- mension of principal neurons in feedforward connected strata. The evolution of living organisms (and ANNs) towards more complex cognitive functions requires an in- crease of the neuronal input dimension. A concept capa- ble brain (or ANN) should meet the following require- ments: a) At least one selective and one concept strata; b) The adequate neuronal dimensions, e.g. n ≈ 102 and m ≈ 104 for the selective and concept strata, respec- tively; c) The order parameter β properly chosen for dif- ferent strata. These conditions are fundamental and have no specific relations to the fine features of the model. Encephal- ized animals and humans satisfy constraints (a) and (b). Thus, our results support the hypothesis of a strong cor- relation between the level of the neural connectivity in living organisms and different cognitive behaviors such organisms can exhibit (cf. [3, 21]). Condition (c) is re- lated to the learning rate and hence to the magnitude of the synaptic plasticity, which differs significantly among neurons [22]. It defines whether a neuron can be selec- tive or associative. We thus suggest a hierarchy of cog- nitive functionality. First relay stations in the informa- tion processing, i.e., selective strata, gain extreme selec- tivity at intermediate dimensions (n ≈ 30 − 100). The second critical transition occurs at much higher dimen- sions m ≈ 500 − 1000 (cf. [1]). Then, neurons located in concept strata become capable of associating multiple uncorrelated inputs of different sensory modalities into concepts. A straightforward extension of our model is an inclusion of more strata, which could encode associ- ation of primarily concepts into compound ones, as was observed experimentally [11]. Acknowledgements. This work was supported by the Russian Science Foundation (19-12-00394) and by the Spanish Ministry of Science, Innovation and Universities (FIS2017-82900P). [1] I.Y. Tyukin et al., Bull. Math. Biol. doi: 10.1007/s11538- [12] R. Quian Quiroga, Phys. Life Rev. 018-0415-5 (2018). [2] A. Tozzi, Phys. Life Rev. doi: 10.1016/j.plrev.2018.12.004 (2019). [3] S. Herculano-Houzel, PNAS 109, 1066110668 (2012). [4] E.L. MacLean et al., PNAS 111, E2140E2148 (2014). [5] N.A. Goriounova et al., eLife 7, e41714 (2018). [6] B.A. Olshausen, D.J. Field, Curr. Opin. Neurobiol. 14, 481-487 (2004). [7] I. Goodfellow, Y. Bengio, A. Courville, Deep Leaning (MIT Press, Cambridge, 2016). [8] J. Schmidhuber, Neur. Netw. 61, 85 (2015). [9] A.N. Gorban, V.A. Makarov, I.Y. Tyukin, Phys. Life doi: 10.1016/j.plrev.2019.02.014 (2019). [13] J. Bowers, Psychol. Rev. 116, 220 (2009). [14] J.J. Hopfield, PNAS 79(8), 2554 (1982). [15] V. Folli, M. Leonetti, G. Ruocco, Front. Comp. Neurosci. 10, 144 (2017). [16] J. Rocchi, D. Saad, D. Tantari, J. Phys. A: Math. Theor. 50(46), 465001 (2017). [17] N. Benito et al., Cereb. Cort. 24(7), 1738 (2014). [18] E. Oja, J. Math. Biol. 15, 267 (1982). [19] W. Hoeffding, J. Amer. Statist. Assoc. 301, 13 (1963). [20] See Supplemental material for more details. [21] S.A. Lobov et al., Math. Mod. Nat. Phenom. 12(4), 109 Rev. doi: 10.1016/j.plrev.2018.09.005 (2018). (2017). [10] R. Quian Quiroga et al., Nature 435(7045), 1102 (2005). [11] R. Quian Quiroga, Nat. Rev. Neurosci. 13: 587 (2012). [22] C. Schmidt-Hieber, P. Jonas, J. Bischofberger, Nature 429, 184 (2004). (cid:1)(cid:1) (cid:9)(cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8)(cid:10) (cid:8)(cid:8) (cid:8) (cid:8) (cid:8) (cid:8)(cid:8)(cid:8)(cid:8) (cid:2) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8)(cid:8)(cid:8) (cid:8) (cid:8) (cid:8) (cid:8)(cid:8) (cid:6)(cid:6) (cid:6)(cid:6) (cid:6)(cid:6) (cid:7)(cid:7)(cid:7) (cid:1)(cid:1) 9 5 Vln. (cid:9)(cid:8)(cid:8)(cid:10) Vln. (cid:8)(cid:10) (cid:9)(cid:8) 3 2 (cid:3) (cid:4) (cid:8) (cid:8) 0 (cid:3) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) 2 (cid:4) (cid:8) f (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8) (cid:8)(cid:8) # # A FGG DEF D# (cid:6)(cid:6) (cid:6)(cid:6) (2%) stratum selective 5 (cid:7)(cid:7)(cid:7) INITIAL FINAL (c) Violin (cid:6)(cid:6) 4 4 (cid:7) stratum concept Vln. phase phase (b) frequency (Hz) Eq. (14) Computed 1000 3000 neuronal dimension, m 2000 2468 (a) stimuli, K num. of associated Universal principles justify the existence of concept cells Supplemental material for: Carlos Calvo Tapia, Ivan Tyukin, Valeri A. Makarov (Dated: December 5, 2019) Appendix A: Firing probability in selective stratum tribution of v satisfies at t = 0 Before discussing the firing probability, let us mention useful results. 1. Hoeffding's inequality Let Xi, i = 1, . . . , n be random variables with compact support P(Xi ∈ [a, b]) = 1, and ¯X = 1 nP Xi. Then (A1) P( ¯X − E[ ¯X] ≥ t) ≤ e− 2nt2 P(− ¯X + E[ ¯X] ≥ t) ≤ e− 2nt2 (b−a)2 (b−a)2 for some t > 0 [1]. 2. Central Limit Theorem Let {Xi}n i=1 be n independent random variables with i=1. We intro- zero means and standard deviations {σi}n duce new random variable: X = Pn pPn i=1 Xi i=1 σ2 i (A2) (A3) with the cdf Fn(·). Then, we have [2]: Fn( X) − Φ( X) ≤ C Pn (Pn i=1 ρi i=1 σ2 i )3/2 , where Φ is the cdf of the standard normal distribution and ρi = E[Xi3]. Moreover, the constant is bounded by [3, 4]: √10 + 3 6√2π ≤ C ≤ 0.56. 0.4097 ' (A4) Property (A3), being a version of the Central Limit Theorem, implies that empirical averages of independent random variables with zero means and finite second and third moments are asymptotically normally distributed as n → ∞. If no further assumptions are imposed then the convergence rate is O(1/√n). 3. Decay of tails of the membrane potential Employing (A3) and noting that σ2 = E[x2 i ] = 1/9 and ρ = E[xi3]E[wi3] = 1/16, the cumulative dis- i ]E[w2 (cid:12)Fn(cid:16)√3v(cid:17) − Φ(cid:16)√3v(cid:17)(cid:12)(cid:12)(cid:12) ≤ (cid:12) (cid:12) 27C 16√n ≤ 0.945 √n (A5) where, as above, Fn(·) is the corrected distribution. probability to a random stimulus x: Using (A1) we get the following estimate on the firing P(v > θ) < e−θ2/6. (A6) Now by employing this concentration inequality together with (A5), we find the bounds: where pdw ≤ P(v > θ) ≤ pup, pup = min(cid:26)e−θ2/6, 1 − Φ(√3θ) + pdw = max(cid:26)0, 1 − Φ(√3θ) − 0.945 √n (cid:27) . 0.945 √n (cid:27) , (A7) (A8) For high n the bounds converge to the probability value 1− Φ(√3θ), provided in the main text. We also note the exponential convergence of pup(θ, n) to zero as a function of θ. This is a direct consequence of measure concentra- tion effects. Appendix B: Conditions of neuronal firing in forward time: Selection of βsl We set the firing threshold small enough, e.g. θ = 1. Then, with high probability, all neurons are active, i.e., dj ≥ 1, and there are no lost stimuli (Fig. 2(a), main text). For convenience, we denote by h = q 3 n xi the first stimulus activating the j-th neuron at t∗, i.e., yj(t < t∗) = 0, yj(t∗) > 0. Let us now find the condition that the neuron keeps "firing" for t > t∗. We decompose wj into vectors parallel and orthogonal to h (by omitting the index j): w = wk + w⊥, where wk := q(t) h and hwk, w⊥i = 0. Then, Eq. (2c) yields: khk w⊥ = − αkhkyqw⊥, q =αkhky(β2 − q2). (B1) By construction, at t = t∗ the neuron fires, i.e., v(t∗) = q(t∗)khk > θ. Note that q(t ≥ t∗) > 0, otherwise y = 0 and there is no dynamics. Selecting β > θ/khk we ensure arXiv:1912.02040v1 [q-bio.NC] 4 Dec 2019 2 2. Probability by normal distribution By employing normal distribution, from (C1) we get: P(y = 0h) = Φ(δkhk√n). (C6) Then, we extend it to arbitrary h as above: PN =Z ∞ 0 Φ(δ√ns)κ(s; µ, σ) ds, (C7) where κ(·; µ, σ) is the normal pdf with the mean µ = 1 and the standard deviation σ = 2√5n . Equation (C7) corresponds to Eq. (6) in the main text. 3. Comparison of two approaches The neuronal selectivity is given by [Eq. (7) in the main text]: S(n, L) = P L−1, (C8) where P can be taken either from (C5) or from (C7). 2Φ−1(psl) √5n . (B3) , δ =s1 − θ δ βsl = the firing condition y(t ≥ t∗) > 0. Then, w⊥(t) → 0 and q → β, which implies: w(t) = β lim t→∞ . h khk (B2) Note that the value of β should not be too high, since it can diminish the neuronal selectivity (see below). Choos- ing β = θ/khk + , where 0 <  (cid:28) 1, ensures activity of the neuron but it requires knowledge of khk, inaccessible a priori. Then, by using khk2 ∼ N (1, ) and requiring P(khk2 > δ2) = psl, where δ ∈ (0, 1) is a lower bound of khk, we can set: 2√5n This guarantees faring of the neuron to the stimulus h in forward time with a probability no smaller than psl. Note that the higher the neuronal dimension n, the higher psl can be chosen. Appendix C: Selectivity after learning We assume that a neuron has learnt an arbitrary stim- n xio. Then, after learning w = β h . We now estimate the probability khk that the neuron is silent to another arbitrary stimulus ulus, which we denote by h ∈ nq 3 g ∈ {q 3 n xi} (g 6= h) given h: P(y = 0h). This can be done in several ways. 1. Probability by Hoeffding's inequality From (C1) we have: P(y = 0h) = P(cid:0)hh, gi ≤ δkhk(cid:12)(cid:12)h(cid:1) . By employing the Hoeffding's inequality (A1) we get (C1) FIG. 1. Two estimates of S (see also Fig. 3(a) in the main text, all parameter values are the same). (C2) Thus P(hh, gi ≥ τ ) ≤ e−nτ 2/18. P(y = 0h) > 1 − e−nkhk2δ2/18. (C3) (C4) Figure 1 shows the neuronal selectivity estimated by two methods. The lower bound estimated from inequali- ties (C5) is too conservative, while Eq. (C7) matches well the numerical results (see Fig. 3(a) in the main text). Now, by recalling khk2 ∼ N (1, we obtain: 2√5n ) for high enough n, Appendix D: Order parameter βcn for the concept stratum where γ(n) = δ2n PH = P(y = 0) > 1 − e−γ(n). 18 (1 − δ2 45 ) is an increasing function of n. (C5) A neuron in the concept stratum receives as an input i=1 yi, where k defines the time the stimulus hk = Pk window [see Eq. (9) in the main text]. Hoeffding Normal 10 1 neuronal dimension, n 10 2 1 0.8 0.6 0.4 0.2 0 selectivity, S 1. Learning condition For further calculations, we assume that z := kyk2 is exponentially distributed: 3 At t = 0 we assume that the neuron detects the first stimulus h1 = y1, i.e. hw(0), h1i > θcn, which is equiva- lent to q(0) > θ/kh1k in Eq. (B1). Thus, to keep firing we require βcn > θcn kh1k . (D1) By using Eq. (B1) we get that, at the end of the first interval ∆, w → βcnh1/kh1k. In general, the initial con- dition for the k-th interval is wk0 = lim t→(k−1)∆ w(t) ≈ βcn hk−1 khk−1k . This is equivalent to qk0 = βcn hhk−1, hki khk−1kkhkk . (D2) (D3) To meet the firing condition at t = (k−1)∆, we require qk0 > θcn/kh1k which yields βcn > θcn kh1k khkk khk−1k > θcn kh1k , (D4) where we used hyi, yji = 0 for j 6= i [see the main text]. Thus, given that α is big enough, the neuron will fire during the whole process of learning. Once the learning is finished, w = βcn neuron is conceptual if hK khKk . Then, the βcn > θcnkhKk kyik2 which is equivalent to , i = 1, 2, . . . , K, (D5) fz(z) =(cid:26) λe−λz, 0 z > 0 otherwise (D9) Then, S follows the Erlang distribution: fS(s) = λKsK−1e−λs (K − 1)! , s > 0. (D10) To find the distribution of M we write: FM (m) = P(min{zi} ≤ m) = 1 − (1 − Fz(m))K. (D11) where Fz is the cdf of z. Thus, FM (m) = 1 − e−Kλm, m > 0. (D12) We now can assume that S and M are independent and hence f (m, s) = fM (m)fS(s). Then, Eq. (D8) yields fS(s)(1 − FM (√s/Ψ)) ds. (D13) By using (D10), (D12), (D13), and operating, we get uK−1e−u−a√u (K − 1)! du, a = K√λ Ψ . (D14) We now note that a is a small parameter. Thus, we can approximate e−u−a√u ≈ e−u(1 − √ua) and evaluate the integral (D14): Γ(K + 1 2 ) (K − 1)! This equation provides the estimate: pcn = 1 − a . (D15) βcn = θcn√λ KΓ(K + 1 2 ) (1 − pcn)(K − 1)! . (D16) 0 pcn =Z ∞ pcn =Z ∞ 0 β2 > θ2 cn i=1 kyik2 PK mini∈{1,...,K}{kyik4} 2. Estimate of βcn For convenience, let's denote: . (D6) We now note that λ = 1/E[z] and assume that all neurons in the selective stratum have learnt stimuli, i.e., z = k(βslkhk − θsl)bk2, (D17) where b is a binary vector representing neurons activated by the stimulus h. Thus, S = K X i=1 kyik2, M = min i {kyik2}. (D7) We then set βcn = θcnΨ, where Ψ satisfies [Eq. (D6)]: P(M 2Ψ2 > S) = pcn, (D8) where pcn is the lower probability bound. This equation ensures that the concept stratum learns at least K inputs with the probability not smaller than pcn. E[z] = β2 slE[(khk − δ)2]E[kbk2]. (D18) Then, we note that kbk2 ∼ B(m, p) and hence E[kbk2] = mp. In the case that all L stimuli have been learnt p = L−1. Now we have E[(khk − δ)2] = 1 − 2δE[khk] + δ2. In the first order approximation, E[khk] ≈ 1. Thus, we have λ ≈ L β2 sl(1 − δ)2m . (D19) Substituting approximation (D19) into Eq. (D16) we ob- tain Eq. (10) in the main text. [1] W. Hoeffding, J. Amer. Statist. Assoc. 301, 13 (1963). [2] C.G. Esseen. Arkiv for matematik, astronomi och fysik. [3] C.G. Esseen. Skand. Aktuarietidskr. 39, 160 (1956). [4] I.G. Shevtsova. Doklady Mathematics. 82(3), 862 (2010). A28, 1 (1942). 4
1811.05403
2
1811
2019-08-19T15:42:23
First passage time distribution for spiking neuron with delayed excitatory feedback
[ "q-bio.NC" ]
A class of spiking neuronal models with threshold 2 is considered. It is defined by a set of conditions typical for basic threshold-type models, such as the leaky integrate-and-fire (LIF) or the binding neuron model and also for some artificial neurons. A neuron is stimulated with a Poisson stream of excitatory impulses. Each output impulse is conveyed through the feedback line to the neuron input after finite delay $\Delta$. This impulse is identical to those delivered from the input stream. We have obtained a general relation allowing calculating exactly the probability density function (PDF) $p(t)$ for distribution of the first passage time of crossing the threshold, which is the distribution of output interspike intervals (ISI) values for this neuron. The calculation is based on known PDF $p^0(t)$ for that same neuron without feedback, intensity of the input stream $\lambda$ and properties of the feedback line. Also, we derive exact relation for calculating the moments of $p(t)$ based on known moments of $p^0(t)$. The obtained general expression for $p(t)$ is checked numerically using Monte Carlo simulation for the case of LIF model. The course of $p(t)$ has a $\delta$-function type peculiarity. This fact contributes to the discussion about the possibility to model neuronal activity with Poisson process, supporting the "no" answer.
q-bio.NC
q-bio
August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback 9 1 0 2 g u A 9 1 ] . C N o i b - q [ 2 v 3 0 4 5 0 . 1 1 8 1 : v i X r a Fluctuation and Noise Letters c(cid:13) World Scientific Publishing Company First passage time distribution for spiking neuron with delayed excitatory feedback OLHA SHCHUR and ALEXANDER VIDYBIDA Bogolyubov Institute for Theoretical Physics, Metrologichna str., 14-B, Kyiv, 03680, Ukraine [email protected] Received (received date) Revised (revised date) A class of spiking neuronal models with threshold 2 is considered. It is defined by a set of conditions typical for basic threshold-type models, such as the leaky integrate-and- fire (LIF) or the binding neuron model and also for some artificial neurons. A neuron is stimulated with a Poisson stream of excitatory impulses. Each output impulse is conveyed through the feedback line to the neuron input after finite delay ∆. This impulse is identical to those delivered from the input stream. We have obtained a general relation allowing calculating exactly the probability density function (PDF) p(t) for distribution of the first passage time of crossing the threshold, which is the distribution of output interspike intervals (ISI) values for this neuron. The calculation is based on known PDF p0(t) for that same neuron without feedback, intensity of the input stream λ and properties of the feedback line. Also, we derive exact relation for calculating the moments of p(t) based on known moments of p0(t). The obtained general expression for p(t) is checked numerically using Monte Carlo simulation for the case of LIF model. The course of p(t) has a δ-function type peculiarity. This fact contributes to the discussion about the possibility to model neuronal activity with Poisson process, supporting the "no" answer. Keywords. spiking neuron; Poisson stochastic process; probability density function; delayed feedback; interspike interval statistics 1. Introduction Activity of neurons either in the brain or at the periphery is highly irregular and is normally treated as random, [1 -- 11]. For description/modeling of this activity simple stochastic processes are used, like Poisson [1, Fig.12], [6, 8, 9], gated Poisson [4], Erlang [7]. It seems that a Poisson-type statistics is adequate at the periphery of the nervous system, e.g. at the neuromuscular junctions, [1], the olfactory receptor neurons, [8], the chemoreceptor neurons, [9], the mechanoreceptor neurons, [7]. For the primary visual cortex, already in 1966, it was observed that activity of only 1/7th portion of recorded neurons could be modeled by Poisson process, [6]. In the more central areas, like the parietal lobe, neuronal activity exhibits a remarkable regularity incompatible with Poisson-type statistics, [10, 11]. Deviation from Poisson statistics manifests itself in the deviation of the inter- 1 August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback 2 Olha Shchur and Alexander Vidybida spike intervals probability distribution from the exponential form. Many different reasons may bring on such a deviation. The most evident reason is the requirement to have more than a single input impulse for generating a single output spike. This makes the shortest ISIs less probable, see Fig. 2,a, Fig. 3, a,b. Other straightforward reasons might be a variable input, [12, 13], or adaptation, [13]. One more reason is feedback, either positive or negative. Such a feedback is often present if a neuron is embedded into a neuronal network with non-zero transmission times, capable of reverberating dynamics. In this case, the feedback is normally mediated by other neurons and this may have a pronounced effect on the firing statistics of any of them. Our purpose is to check what might be that effect. Rigorous mathematical treatment of stochastic behavior of neurons in a network with delayed connections is a difficult task. In order to obtain an exact result, we take the simplest possible case of feedback, namely, a neuron sends its output impulses onto its input, see Fig. 1, below. This can be considered as the simplest, "caricature network" with delayed feedback, composed of a single neuron. It was our surprise that this kind of neuronal organization has been observed experimentally, see Sec. 2.3. In this paper, we consider a neuron with delayed excitatory feedback. The pur- pose of this work is to determine which effect the delayed feedback of this type might have on the probability density function (PDF) of output interspike intervals (ISI). Mathematically, the problem consists of calculation of PDF p(t) for the first passage times of crossing the threshold for a neuron with the excitatory feedback, provided that the PDF of the same type for the neuron without feedback, p0(t), and the properties of the feedback line are given. It is worth noticing that in the paper [14] the similar aim has been achieved but in the case of instantaneous excitatory feedback. As a result, the relation between three PDFs, namely for stimulus, for output stream without feedback and for output stream with instantaneous feedback, has been obtained. This relation is used farther in this paper. In our previous paper [15], we have considered the case of delayed fast inhibitory feedback. We have obtained the relation between the PDF for the neuron with delayed fast inhibitory feedback stimulated with Poisson stream and the PDF for the same neuron without feedback. In that work, we have come to the conclusion that the presence of delayed fast inhibitory feedback results in jump discontinuity of PDF at the point where ISI value equals the delay in the feedback line. In [15], the problem has been stated for a whole class of threshold-type neuronal models that are determined through specifying a threshold level, a height of the input impulses, and a time course of excitation decay. It has been shown there that for any neuronal model (with and without feedback) that satisfies the imposed conditions there is invariant with respect to the way of excitation decay the initial segment of PDF. It is completely determined by the input stream and a ratio between the threshold level and the height of input impulses. In addition to this, it has appeared that the August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback Firing statistics of spiking neuron with delayed feedback 3 initial segment found for PDF for a neuron with feedback p(t) is enough to express statistical moments of p(t) through moments of p0(t), provided the time delay in the feedback line is shorter than the length of the above-mentioned initial segment. In this work, we study the case of excitatory feedback with a non-zero delay ideologically similarly to our preceding work [15] dealing with delayed fast inhibitory feedback. Biological justification of this case of feedback is given in Sec. 2.3. We consider a subclass of neuronal models of the class considered in [15], which is composed of models having the firing threshold 2 (see Sec. 2.1, below). That is, the imposed in [15] conditions on neuronal models, are valid here, see Cond0-Cond4, below. We assume that the neuron is stimulated with the input Poisson stream of excitatory impulses. For this type of stimulation, we derive a general relation between the ISI values PDF p(t) for a neuron with feedback based on known PDF p0(t) for the same neuron without feedback, see Eqs. (13), (16),(17) and (18). Further, as it was mentioned previously, that for any neuronal model satisfying imposed conditions, there exists initial interval ]0; T2[ of ISI values at which p0(t) and, consequently, p(t) does not depend on the neuronal model chosen. We calculate exactly that model-independent initial segment of p(t), see Sec. 3.2. As it can be clearly seen in Fig. 2 b,c, the p(t) has peculiarity at the point equal to the delay time. Moreover, we obtain in our approach the model-independent relations between the moments of PDF of a neuron with and without feedback, see (20), (21), (23). The first moment, the mean ISI value, was known before for the binding neuron model only, see Eq. (21). The general expressions obtained here confirm what has been found previously for the binding neuron. Finally, the exact expressions found for p(t) are verified by means of Monte Carlo simulation for the LIF neuron model, see Fig. 2 b,c. 2. Methods 2.1. Class of neuronal models We consider a class of neuronal models (without feedback) with threshold 2 stim- ulated with Poisson stream of input impulses of intensity λ. This class is a subset of a class of neuronal models considered in our preceding work [15], where any value for the threshold was considered. The class considered here includes the leaky integrate-and-fire (LIF) and the binding neuron models,a both with threshold 2. The neuronal state at the moment t is described by the depolarization voltage V (t), which considered to be biased by the resting potential value. Thus V = 0 at the resting state and depolarization voltage is positive. The input impulse increases the depolarization voltage by h: V (t) → V (t) + h, (1) aDefinition of the binding neuron model can be found in [16]. See also https://en.wikipedia.org/wiki/Binding neuron. August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback 4 Olha Shchur and Alexander Vidybida where h > 0. The input impulse decay is determined by a decreasing function y(u), which is different for different neuronal models. It means that if an impulse is received at the moment t, then for any u > 0 V (t + u) = v(t + u) + hy(u), where v(t + u) denotes depolarization voltage without that impulse involved. For example, for the LIF model one has y(u) = e− u τ , (2) (3) where τ is the relaxation time. The neuron is characterized by a firing threshold value V0: as soon as V (t) > V0, the neuron generates a spike and V (t) becomes zero. Instead of specifying any concrete neuronal model (through specifying V0, h and y(u)), we consider a class of neuronal models, which (without feedback) satisfy the following conditions: • Cond0: Neuron is deterministic: Identical stimuli elicit identical spike trains • Cond1: Neuron is stimulated with input Poisson stream of excitatory im- from the same neuron. pulses. PDF of intervals between those impulses is pin(t) = λe−λt, (4) where t means an input ISI duration. • Cond2: Just after firing, neuron appears in its resting state. • Cond3: The function y(u), which governs decay of excitation, see Eq. (2), is continuous and satisfies the following conditions: y(0) = 1, 0 < u1 < u2 ⇒ y(u1) ≥ y(u2). ration, exists together with all its moments. (5) • Cond4: The PDF for output ISIs, p0(t), where t means an output ISI du- • Cond5: Neuron has a threshold 2. It means that in order to be triggered neuron should obtain at least 2 impulses, or, in other words, the following relation between the threshold V0 and the input impulse height h is fulfilled: 1 < V0 h < 2. (6) Notice, that Cond0-4, above, are the same as in [15]. As it has been shown in [15], for any neuronal model that satisfies Cond0-5 there is the initial segment of PDF p0(t) of length T2 where the PDF p0(t) on that segment looks as followsb p0(t)dt = λte−λtλdt, t < T2. (7) bIn the (7), we have taken into account that the threshold equals two. August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback Firing statistics of spiking neuron with delayed feedback 5 The concrete value of T2 is determined by a way of excitation decay (the explicit form of y(u), see [15, Sec. 3.3]). 2.2. Type of feedback Fig. 1. Neuron with delayed feedback. As neuron in the figure, we consider any neuronal model, which satisfies the set of conditions Cond0 - Cond5, above. To the neuronal model which satisfies the Cond0-Cond5 mentioned above we add the delayed excitatory feedback line and expect that it has the following properties: • Prop1: The time delay in the line is ∆ > 0. • Prop2: The line is able to convey no more than one impulse. • Prop3: The impulse conveyed to the neuronal input is excitatory impulse • Prop4: The delay in the feedback line ∆ satisfies the following condition: (8) identical to those from the input stream. ∆ < T2. Notice, that Prop1-2, above, are the same as in [15]. 2.3. Biological justification In the natural conditions, neurons are embedded into reverberating networks. Since output neuronal activity can be fed back transformed and transmitted by other neurons, most neurons can feel their own activity produced some time earlier. This justifies consideration of a feedback in neural systems. At the same time, it is known that excitatory neurons can form synapses (au- tapses) on its own body or dendritic tree [17, 18]. This substantiates consideration of a single neuron with excitatory feedback not only as the simplest reverberating network possible but also as an independent biologically relevant case. The delay ∆ comprises the time required by the output spike to pass the distance from axonal hillock, where it is generated, to the autapse and the synaptic delay. 3. Results 3.1. Pdf: general relation The strategy for finding the PDF for a neuron with delayed excitatory feedback is the same as in our previous work [15] on delayed fast inhibitory feedback. ✲✐♣✉tstr❡❛♠✭P♦✐ss♦✮♥✁✂✄☎♥✲✛✲✵✆✝✆✞✆✲❞❡❧❛②❡❞❢❡❡❞❜❛❝❦⑤④③⑥✟♦✉t♣✉tstr❡❛♠✠■❙■❞✉r❛t✐♦ August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback 6 Olha Shchur and Alexander Vidybida Let us introduce a time-to-live s -- the time at the beginning of ISI needed for an impulse in the feedback line to reach a neuron. f (s) is a distribution of such times-to-live in a stationary regime. If p(ts) is the conditional PDF that allows calculating the probability to obtain an ISI duration t given that the time-to-live at the beginning of the ISI is s, then in case of stationary regime the wanted PDF for a neuron with feedback can be found in the following way: p(t) = ds p(ts)f (s). (9) ∆(cid:90) The conditional PDF p(ts) depends on the type of feedback. Now we need to figure out what p(ts) looks like in case of excitatory feedback. 0 3.1.1. Conditional probability p(ts) Below we will scrutinize different cases for the sake of obtaining the exact expression for p(ts). Firstly, the probability to get an ISI t shorter than time-to-live s does not depend on the feedback line presence and is the same as in the case of the feedback absence. Thus t < s ⇒ p(ts) dt = p0(t) dt. (10) Secondly, there is nonzero probability to obtain ISI with exact duration t = s. In order to get such an ISI duration, a neuron must receive one input impulse before receiving an impulse from the feedback line. The condition (8) ensures that in this case at the moment t ∈]s− ; s + [ there will be enough excitation in the neuron to surpass the threshold. The probability of this event, i.e. to obtain one impulse (and no more) from input Poisson stream on time interval ]0; t[, is λte−λt. Therefore for the conditional probability p(ts) in this case: t ∈]s − ; s + [⇒ p(ts)ds = λte−λtδ(t − s)ds. (11) Thirdly, if we obtain an ISI duration t > s, it means that before receiving an impulse from the feedback line, due to (8), there has not been any excitation on the neuron, i.e. the neuron has not received an impulse from the input line. The probability of this event is simply e−λs. Right away after the receiving that impulse, the neuron is in the same state as at the beginning of ISI if there was an instantaneous feedback instead of delayed. Let us denote PDF for the same neuron but with instantaneous feedback (i.e. when ∆ = 0) as po if (t). Thus, in this case, we have the following: t > s ⇒ p(ts) dt = e−λspo if (t − s) dt. (12) In [14] the relation between PDF for a neuron without feedback p0(t) and PDF for the same neuron with instantaneous feedback po if (t), stimulated with Poisson stream, has been derived. It looks as follows: po if (t) = p0(t) + 1 λ d dt p0(t). (13) August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback Firing statistics of spiking neuron with delayed feedback 7 To summarize, the conditional probability p(ts) can be written in the following form: p(ts) = χ(s − t)p0(t) + λte−λtδ(t − s) + e−λspo if (t − s), (14) where χ(t) is the Heaviside step function. 3.1.2. Distribution of times-to-live f (s) As regards the distribution of times-to-live f (s), in [15], it was proved that f (s) has the following form: f (s) = g (s) + aδ (s − ∆) , where and a = g(s) = aλ 2 3.1.3. General relation for PDF p(t) 4e2λ∆ , 1 + e2λ∆(2λ∆ + 3) (cid:16) 1 − e−2λ(∆−s)(cid:17) (15) (16) (17) . After substituting (14) and (15) into (9), one can obtain the following general formulae for different domains of ISI duration: ds e−λspo if (t − s)g(s)+ t(cid:82) ∆(cid:82) 0  ∆(cid:82) 0 ds g(s) + a) + g(t)λte−λt, p(t) = t +p0(t)( aλ∆e−λ∆δ(t − ∆), ae−λ∆po if (t − ∆)+ t < ∆; t ∈]∆ − ; ∆ + [; (18) ds e−λspo if (t − s)g(s), t > ∆. + As it can be seen from right above, the PDF p(t) has a Dirac δ-function type peculiarity at the point corresponding the ISI duration that equals the delay in the feedback line ∆. To sum up, we have the following algorithm for finding the PDF p(t) for a neuron with feedback, if the PDF p0(t) for the same neuron without feedback is known: (1) find po if (t) by using (13); calculate p(ts); (2) substitute po if (t) and components of the distribution function f (s) (16) and (17) into (18) and take integrals. August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback 8 Olha Shchur and Alexander Vidybida Fig. 2. Example of ISI PDF for the LIF neuron with threshold 2 without feedback (a, used Eq. (7)) and with delayed excitatory feedback (b, used Eq. 19)). For both panels: τ = 20 ms, V0 = 20 mV, h = 11.2 mV, λ = 62.5 s−1. ∆ = 4 ms in (b). c -- Monte Carlo simulation for p(t) (used 1 000 000 output ISIs). The total probability under the curves is 0.03726 (a) and 0.42205 (b). 3.2. Model-invariant initial segment of p(t) We can find the initial segment ]0; T2[ of p(t) by substituting (7), (13), (16), (17) into (18):  λe−λt 3 + 2∆λ + e−2λ∆ (e−2λ∆(cid:0)1 − e2λt(1 + λt)(cid:1) + λt(7 + 2∆λ) − 2λ2t2), 3 + 2∆λ + e−2λ∆ δ(t − ∆), λe−λt, 4λ∆e−λ∆ p(t) = t < ∆; t ∈]∆ − ; ∆ + [; t ∈]∆; T2[. (19) This is in agreement with [19, Eq. (18-20)] and [20, Eq. (12-14)] where PDF has been obtained for the binding neuron model only. It is necessary to emphasize that in the present work the exact expression for p(t) (19) on the initial segment ]0; T2[ of ISI values is the same for the whole class of neuronal models with threshold 2 stimulated with Poisson input stream (invariant with respect to a way of excitation decay). The validity of the obtained formulas (19) has been verified using numerical simulation for the LIF model by means of the Monte Carlo method, see Fig. 2 c. 3.3. Moments of PDF Let us denote the nth moment of PDF for a neuron with delayed feedback p(t) and without feedback p0(t) as Wn and W 0 n respectively. Using the definitions of the moments W 0 n and (18) within different time domains, one can obtain the following expression for the nth moment of PDF of ISI duration for a neuron stimulated with Poisson stream: ∆(cid:90) g(t)λ∆e−λ∆ + ap0(t) + p0(t) ae−λ∆∆n−k + ∆(cid:90) k−1) ∆(cid:90) t ds g(s) ds g(s)e−λssn−k 0   (20) Wn = a∆nλ∆e−λ∆ + dt tn (cid:18)n (cid:19) n(cid:88) + k k=0 0 (W 0 k − 1 λ kW 0 024t, ms0.0000.0060.012p0(t), ms1024t, ms0.000.060.120.18p(t), ms1024t, ms0.000.060.120.18p(t), ms1abc August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback Firing statistics of spiking neuron with delayed feedback 9 Fig. 3. Examples of spike trains for the input Poisson stream with intensity λ = 130 s−1, a, the LIF neuron with threshold 2 without feedback (λ = 300 s−1), b, and with delayed excitatory feedback (λ = 220 s−1), c. The λ values are chosen in such a way that the total number of spikes at the 0 -- 500 ms interval is the same. It may be observed that in b the spikes are distributed more evenly than in a. In c, notice the series of ISIs equal ∆ in periods between † and ‡. W 0 0 is a normalization coefficient and equals 1. Notice that the explicit expres- sion of p0(t) in formula right above is used only for values on interval ]0; ∆[ where it is given by (7). The exact expression for the first moment of p(t) looks as follows: 2(cid:0)−1 + W 0 W1 = 1 λ + e2∆λ(−1 + W 0 λ(1 + e2λ∆(2λ∆ + 3)) 1 λ + 2∆λ)(cid:1) . (21) The expression right above is in agreement with that found for binding neuron model in [19] and [20]. In the case of instantaneous feedback, i.e. when ∆ = 0, for the mean of p(t) we have: W 0 if 1 = W 0 1 − 1 λ . As regards the second moment, W2, the exact expression is: W2 = +e2∆λλ(−1 + W 0 2 λ2(1 + e2λ∆(2λ∆ + 3)) 2 λ2 2 λ + 2∆λ) + 2e∆λ(−2 − ∆2λ3 + ∆3λ3)). 1 λ + W 0 (4 + λ − 4W 0 In case of the instantaneous feedback the second moment is: W 0 if 2 = W 0 2 − 2 λ W 0 1 . (22) (23) (24) (22) and (24) are in concordance with [14, Eq. (7),(8)] where relations between the first two moments for a neuron with instantaneous feedback and for the same neuron without feedback have been obtained. 4. Conclusions and Discussion In the current paper, we study the effect of the presence of the delayed excitatory feedback line on the neuronal activity. For a class of neural models with threshold 2 stimulated with Poisson process and satisfying the imposed conditions (see Sec. ‡†‡†‡†‡†abc✲0500ms August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback 10 Olha Shchur and Alexander Vidybida 2.1), we have derived the general relation between the PDF p(t) of ISI values for a neuron with excitatory feedback and the corresponding PDF p0(t) for the same neuron, but without feedback, see Eqs. (13), (16), (17) and (18). Furthermore, the initial model-invariant segment of PDF p(t) has been found in the explicit form, see (19). Also, we express the moments of PDF p(t) for a neuron with feedback through the corresponding moments of PDF p0(t) for the neuron without feedback. The obtained relations between moments must be met for any neuronal model with threshold 2 stimulated with Poisson stream (i. e. that satisfies the imposed conditions Cond0-5, above). Notice that we consider here a deterministic threshold. In reality, the exact value of the firing threshold may fluctuate for different reasons. Results for fluctuating threshold obtained in diffusion approximation can be found in [21]. All obtained results are mathematically rigorous. Other rigorous results as regards statistics of neuronal activity can be found in the review [22]. As it can be clearly seen in Fig. 2 b,c, the course of the PDF found has a peculiarity -- a Dirac δ-function for ISI t = ∆, which makes it impossible to describe the output neuronal stream by a Poisson-like or other simple distribution. If the output of a neuron is fed to another neuron, the spiking pattern of the output determines the reaction of the post-synaptic neuron, see e.g. [2, 5]. The δ-function presence in the p(t) may rearrange the spiking pattern by adding some regularity, as illustrated in Fig. 3 c. Acknowledgments This research was supported by theme grant of the department of physics and astronomy of NAS of Ukraine: "Dynamics of formation of spatially non-uniform structures in many-body systems", PK 0118U003535. References [1] A. W. Liley, "An investigation of spontaneous activity at the neuromuscular junction of the rat", Journal of Physiology 132 (1956) 650 -- 666. [2] J. P. Segundo, G. P. Moore, L. J. Stensaas and T. H. Bullock, "Sensitivity of neurons in aplysia to temporal pattern of arriving impulses", Journal of Experimental Biology 40 (1963) 643 -- 667. [3] G. L. Gerstein and B. Mandelbrot, "Random walk models for the spike activity of a single neuron", Biophysical Journal 4 (1964) 41 -- 68. [4] D. R. Smith and G. K. Smith, "A statistical analysis of the continual activity of single cortical neurones in the cat unanaesthetized isolated forebrain", Biophysical Journal 5 (1965) 47 -- 74. [5] J. P. Segundo, D. H. Perkel and G. P. Moore, "Spike probability in neurones: influence of temporal structure in the train of synaptic events", Kybernetic 3 (1966) 67 -- 82. [6] J. S. Griffith and G. Horn, "An analysis of spontaneous impulse activity of units in the striate cortex of unrestrained cats", Journal of Physiology 186 (1966) 516 -- 534. [7] V. S. Korolyuk, P. G. Kostyuk, B. Y. Pjatigorskii and E. P. Tkachenko, "Mathemat- ical model of spontaneous activity of some neurons of the central nervous system", Biofizika 12 (1967) 895 -- 899. August 20, 2019 2:15 WSPC/INSTRUCTION FILE delayexcfeedback Firing statistics of spiking neuron with delayed feedback 11 [8] W. Drongelen, "Unitary recordings of near threshold responses of receptor cells in the olfactory mucosa of the frog", Journal of Physiology 277 (1978) 423 -- 435. [9] D. F. Donnelly, J. M. Panisello and D. Boggs, "Effect of sodium perturbations on rat chemoreceptor spike generation: implications for a poisson model", Journal of Physiology 511 (2004) 301 -- 311. [10] G. Maimon and J. A. Assad, "Beyond poisson: Increased spike-time regularity across primate parietal cortex", Neuron 62 (2009) 426 -- 440. [11] B. B. Averbeck, "Poisson or not poisson: Differences in spike train statistics between parietal cortical areas", Neuron 62 (2009) 310 -- 311. [12] G. D'Onofrio and E. Pirozzi, "Successive spike times predicted by a stochastic neu- ronal model with a variable input signal", Mathematical Biosciences and Engineering 13 (2016) 495 -- 507. [13] E. Pirozzi, "Colored noise and a stochastic fractional model for correlated inputs and adaptation in neuronal firing", Biological Cybernetics 112 (2018) 25 -- 39. [14] A. K. Vidybida, "Relation between firing statistics of spiking neuron with instan- taneous feedback and without feedback", Fluctuation and Noise Letters 14 (2015) 1550034. [15] A. Vidybida and O. Shchur, "Relation between firing statistics of spiking neuron with delayed fast inhibitory feedback and without feedback", Fluctuation and Noise Letters 17 (2018) 1850005. [16] A. K. Vidybida, "Binding neuron", in Encyclopedia of Information Science and Tech- nology, Third Edition, ed. M. Khosrow-Pour (IGI Global, 2015), pp. 1123 -- 1134. [17] J. M. Bekkers, "Neurophysiology: Are autapses prodigal synapses?", Current Biology 8 (1998) R52 -- R55. [18] V. Aroniadou-Anderjaska, M. Ennis and M. T. Shipley, "Dendrodendritic recurrent excitation in mitral cells of the rat olfactory bulb", Journal of Neurophysiology 82 (1999) 489 -- 494. [19] A. K. Vidybida, "Output stream of binding neuron with delayed feedback", in Pro- ceedings of the 14th International Congress of Cybernetics and Systems of WOSC, Wroclaw, Poland, September 9-12, 2008, eds. J. Jzefczyk, W. Thomas and M. Tur- owska (Oficyna Wydawnicza Politechniki Wroclawskiej, 2008), pp. 292 -- 302. [20] A. K. Vidybida and K. G. Kravchuk, "Output stream of binding neuron with de- layed feedback", The European Physical Journal B - Condensed Matter and Complex Systems 72 (2009) 279 -- 287. [21] M. Abundo and E. Pirozzi, "Integrated stationary Ornstein-Uhlenbeck process, and double integral processes", Physica A: Statistical Mechanics and its Applications 494 (2018) 265 -- 275. [22] L. Sacerdote and M. T. Giraudo, Stochastic Integrate and Fire Models: A Review on Mathematical Methods and Their Applications (Springer Berlin Heidelberg, Berlin, Heidelberg, 2013), pp. 99 -- 148.
1701.02272
2
1701
2017-03-04T18:20:10
Morphognosis: the shape of knowledge in space and time
[ "q-bio.NC", "cs.AI" ]
Artificial intelligence research to a great degree focuses on the brain and behaviors that the brain generates. But the brain, an extremely complex structure resulting from millions of years of evolution, can be viewed as a solution to problems posed by an environment existing in space and time. The environment generates signals that produce sensory events within an organism. Building an internal spatial and temporal model of the environment allows an organism to navigate and manipulate the environment. Higher intelligence might be the ability to process information coming from a larger extent of space-time. In keeping with nature's penchant for extending rather than replacing, the purpose of the mammalian neocortex might then be to record events from distant reaches of space and time and render them, as though yet near and present, to the older, deeper brain whose instinctual roles have changed little over eons. Here this notion is embodied in a model called morphognosis (morpho = shape and gnosis = knowledge). Its basic structure is a pyramid of event recordings called a morphognostic. At the apex of the pyramid are the most recent and nearby events. Receding from the apex are less recent and possibly more distant events. A morphognostic can thus be viewed as a structure of progressively larger chunks of space-time knowledge. A set of morphognostics forms long-term memories that are learned by exposure to the environment. A cellular automaton is used as the platform to investigate the morphognosis model, using a simulated organism that learns to forage in its world for food, build a nest, and play the game of Pong.
q-bio.NC
q-bio
Morphognosis: the shape of knowledge in space and time Thomas E. Portegys [email protected] Ernst & Young LLP, New York, NY, USA Abstract Artificial intelligence research to a great degree focuses on the brain and behaviors that the brain generates. But the brain, an extremely complex structure resulting from millions of years of evolution, can be viewed as a solution to problems posed by an environment existing in space and time. The environment generates signals that produce sensory events within an organism. Building an internal spatial and temporal model of the environment allows an organism to navigate and manipulate the environment. Higher intelligence might be the ability to process information coming from a larger extent of space-time. In keeping with nature's penchant for extending rather than replacing, the purpose of the mammalian neocortex might then be to record events from distant reaches of space and time and render them, as though yet near and present, to the older, deeper brain whose instinctual roles have changed little over eons. Here this notion is embodied in a model called morphognosis (morpho = shape and gnosis = knowledge). Its basic structure is a pyramid of event recordings called a morphognostic. At the apex of the pyramid are the most recent and nearby events. Receding from the apex are less recent and possibly more distant events. A morphognostic can thus be viewed as a structure of progressively larger chunks of space-time knowledge. A set of morphognostics forms long-term memories that are learned by exposure to the environment. A cellular automaton is used as the platform to investigate the morphognosis model, using a simulated organism that learns to forage in its world for food, build a nest, and play the game of Pong. Keywords: Knowledge representation, Machine learning, Cellular automata Introduction The human brain is the seat of intelligence. Thus when we attempt to craft intelligence, naturally we turn to it as a guide. Fortunately, neuroscience is proceeding at an astounding pace (Kaiser, 2014; Stetka, 2016), methodically unpacking its mysteries. Yet the complexity of the brain, with billions of neurons and trillions of synapses, remains daunting. Teasing apart which aspects and features of the brain are essential to the function of intelligence and which are incidental is a crucial and difficult task. Unfortunately, the prospects of understanding complex systems through examination and dissection are questionable (Jonas and Kording, 2016). And as for constructing a complete precise brain model, it is possible, as John von Neumann believed (Mühlenbein, 2009), that at a certain level of complexity the simplest precise description of a thing is the thing itself. In reaction to this, some efforts, such as The Human Brain Project (2015) and Numenta (Hawkins, 2004; White paper, 2011), have taken the position that analysis must be complemented with synthesis and simulation to achieve a satisfactory level of understanding. From an artificial intelligence (AI) viewpoint, we must keep in mind that the purpose of a brain is to allow an organism to navigate and manipulate its environment. Thus it is a solution to problems posed by the environment. The approach of this project is to model the environment as something that could plausibly be in turn modeled by an artificial brain. Some researchers maintain that the environment largely consists of a body for the brain to interact with. The embodied brain will thus leverage the sensory and motor capabilities of a body that are adapted to an environment. Robotics researchers such as Brooks (1999), Hoffmann and Pfeifer (2011) have argued that true artificial intelligence can only be achieved by machines that have sensory and motor skills and are connected to the world through a body. However, this approach belies the problem since the body, like the brain, is also a solution to its environment. Determining a model of an organism's environment is more tractable than creating a brain model of an environmental model. But it requires settling on what is in the world that produces sensory events and reacts to motor responses. Confounding this is that we of course must use our brains to do this. There is a common and somewhat ironic tendency to describe AI inputs and outputs in human cognitive terms, i.e. post-processed brain output, such as symbolic variables. Hoffman (2009) argues that evolution has shaped our senses and perceptual machinery to only provide information on events that are ancestrally significant, such as finding food and safety. Other events in the environment that we cannot directly sense must be mapped through technology onto our sensory capabilities. For example, in the age of science the existence and use of X-rays is important, but we sense them only indirectly, as shadows on photographic film. Indeed, Hoffman argues that reality may be more radically alien than we can imagine. Epistemological offerings would seem at best too abstract to be useful for framing a sensory-response environment, and at worst useless, as in the cases of nihilism and solipsism. And physics has in recent times become increasingly muddier on the "true" nature of reality:  The arrow of time may be related to the perception of entropy (Halliwell, 2011).  String theory demands a number of extra infinitesimal dimensions (Rickles, 2014).  The perception of space may be a holographic projection (Bousso, 2002).  Reality could be a cellular automaton (Wolfram, 2002), a graph (Wolfram, 2015), or a simulation (Moskowitz, 2016). Despite these hazards, people universally experience the environment as a space-time structure. And even if there is a different underlying substructure, the model is empirically effective. The presence of mammalian brain structures for mapping spatial events (Vorhees and Williams, 2014) provides evidence for the processing of this type of information. Similarly, brain structures for sensing the passage of time have also found support (Sanders, 2015). Using space-time as a model, it can be speculated that higher intelligence is the ability to process information arising from a larger extent of space-time. And in keeping with nature's penchant for extending rather than replacing, the purpose of the mammalian neocortex might then be to record events from distant reaches of space and time and render them, as though yet near and present, to the older, deeper brain whose instinctual roles have changed little over eons. If this is so, these structures would be repurposed to embody language and abstract concepts. Building an internal spatial and temporal model of the environment allows an organism to navigate and manipulate the environment. This paper introduces a model called morphognosis (morpho = shape and gnosis = knowledge). Its basic structure is a pyramid of event recordings called a morphognostic, as shown in Figure 1. At the apex of the pyramid are the most recent and nearby events. Receding from the apex are less recent and possibly more distant events. Figure 1 – Morphognostic event pyramid Morphognosis is partially inspired by an abstract morphogenesis model called morphozoic (Portegys et al., 2017). Morphogenesis is the process of generating complex structures from simpler ones within an environment. Morphozoic is based on hierarchically nested neighborhoods within a cellular automaton. Morphozoic was found to be robust and noise tolerant in reproducing a number of morphogenesis-like phenomena, including Turing diffusion-reaction systems (Turing, 1952), gastrulation, and neuron pathfinding. It is also capable of image reconstruction tasks. Description The morphognosis model is demonstrated in three 2D cellular environments: (1) a food foraging task, (2) a nest building task, and (3) the game of Pong. The food foraging task is used as a venue to further define the model. Food foraging In this task a virtual creature called a mox finds itself in a 2D cellular world as shown in Figure 2. To find food the mox must navigate around various obstacles of various types (colors). Figure 2 - Mox food foraging in a 2D cellular world. Figure 3 shows a snapshot of a morphognostic describing the space-time events, in this case obstacle encounters, while the mox forages. In this case a neighborhood is configured as a 3x3 set of sectors. Figure 3 - Pyramid of obstacle type densities arranged as hierarchy of 3x3 cell neighborhoods. Cell type densities are stored instead of raw cell values to allow linear scaling of information as the hierarchy increases since storing individual cell values would result in a geometric growth. The cell type density is only one of a number of possible statistical or aggregation functions that could be used. An alternative might be to look at the distribution of cell types as an image processing operation, such as taking a Laplacian, Sobel or other image operator. Morphognostic spatial neighborhoods A cell defines an elementary neighborhood: neighborhood0 = cell A non-elementary neighborhood consists of an NxN set of sectors surrounding a lower level neighborhood: neighborhoodi = NxN(neighborhoodi-1) where N is an odd positive number. The value of a sector is a vector representing a histogram of the cell type densities contained within it: value(sector) = (density(cell-type0), density(cell-type1), … density(cell-typen)) The number of cells contributing to the density histogram of a sector of neighborhoodi = Ni-1xNi-1 Morphognostic temporal neighborhoods A neighborhood contains events that occur between time epoch and epoch + duration: t10 = 0 t20 = 1 t1i = t2i-1 t2i = (t2i-1 * 3) + 1 epochi = t1i durationi = t2i - t1i Metamorphs In order to navigate and manipulate the environment, it is necessary for an agent to be able to respond to the environment. A metamorph embodies a morphognostic→response rule. A set of metamorphs can be learned from a manual or programmed sequence of responses within a world. Metamorphs establish an important feedback: Learned morphognostics shape responses. • • Responses shape the learning of morphognostics. Metamorph "execution" consists generating a morphognostic for the current mox position and orientation then finding the closest morphognostic contained in the learned metamorph set, where: 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒(𝑚𝑒𝑡𝑎𝑚𝑜𝑟𝑝ℎ𝑖, 𝑚𝑒𝑡𝑎𝑚𝑜𝑟𝑝ℎ𝑗) = 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟ℎ𝑜𝑜𝑑𝑠 𝑠𝑒𝑐𝑡𝑜𝑟𝑠 𝑐𝑒𝑙𝑙 𝑡𝑦𝑝𝑒𝑠 ∑ 𝑥 ∑ 𝑦 ∑ 𝑎𝑏𝑠(𝑐𝑒𝑙𝑙 𝑡𝑦𝑝𝑒 𝑑𝑒𝑛𝑠𝑖𝑡𝑦𝑖,𝑥,𝑦,𝑧 − 𝑐𝑒𝑙𝑙 𝑡𝑦𝑝𝑒 𝑑𝑒𝑛𝑠𝑖𝑡𝑦𝑗,𝑥,𝑦,𝑧) 𝑧 Artificial neural network implementation In a complex environment, generating a large number of metamorphs may be prohibitive in terms of storage and search processing. Alternatively, metamorphs can be used to train an artificial neural network (ANN), as shown in Figure 4, to learn responses associated with morphognostic inputs. During operation, a current morphognostic can be input to the ANN to produce a learned response. The ANN also has these advantages: Faster. • • More compact. • More noise tolerant. Figure 4 – Metamorph artificial neural network. Results The mox were trained in worlds featuring a number of randomly placed obstacles of various types. Training was done by "autopiloting" the mox along an optimal path to the food. This generated a set of metamorphs suitable for testing. Table 1 shows the results of varying the neighborhood hierarchy depth in a 10x10 world. Success indicates the mean amount of food eaten, so 1 is a perfect score. It can be observed that more obstacles tend to improve performance. This is because they tend to form unique landmark configurations to guide the mox. Larger neighborhoods also tend to improve performance. Neighborhoods Obstacle types Obstacles 1 1 1 1 1 2 10 20 10 Food 0.1 0.2 0 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 2 4 4 1 1 2 2 4 4 1 1 2 2 4 4 20 10 20 10 20 10 20 10 20 10 20 10 20 10 20 0 0 0 0.3 0.4 0.2 0.6 0.2 0.6 1 0.9 1 1 1 1 Table 1 – Foraging in a 10x10 world. The next test examines how well the model performs when the test world is not a duplicate of a training world, but is similar to a set of training worlds. Thus for this, multiple training runs are used. Before each training run, the cell types of all the cells are probabilistically modified to a random value. A successful test run must then rely on a composite of multiple training runs. The results are shown in Table 2. Of note is how performance only begins to falter under heavy noise and few training runs. Noise #Train Food 1 0.1 1 0.1 1 0.1 0.9 0.25 1 0.25 1 0.25 0.6 0.5 0.8 0.5 0.5 0.9 Table 2 – Foraging with noise. 1 5 10 1 5 10 1 5 10 Nest building This task illustrates how the morphognosis model can be used to not only navigate but also manipulate the environment. Figure 5 left shows an environment in which a nest will be constructed out of 4 stones (reddish circles) on top of an elevation depicted by the shaded cells. The mox must seek out the stones, pick them up, and assemble them into the completed nest shown in Figure 5 right. Figure 5 – Nest building with gathered stones. Left: scattered stones. Right: completed nest. For this task, the mox is capable of sensing the presence of a stone immediately in front of it, and sensing the elevation gradient both laterally and in the forward-backward direction. In addition to the forward and turning movements used by the foraging task, the mox is capable of picking up a stone in front of it and dropping the stone onto an unoccupied cell in front of it. It also senses whether it is carrying a stone. Training was done by running the mox through 10 repetitions on "autopilot" to build a set of metamorphs. The environment was then reset and the mox tested to discover whether it is capable of building the nest. Over 50 trials were performed with 100% success. Internally, the sensory information from the stone, gradient and stone carry states were sufficient to achieve success with a neighborhood hierarchy of only one level. Pong game Much of the real world is nondeterministic, taking the form of unpredictable or probabilistic events that must be acted upon. If AIs are to engage such phenomena, then they must be able to learn how to deal with nondeterminism. In this task the game of Pong poses a nondeterministic environment. The learner is given an incomplete view of the game state and underlying deterministic physics, resulting in a nondeterministic game. Game details • The goal of the game is to vertically move a paddle to prevent a bouncing ball from striking the right wall, as shown in Figure 6. • Ball and paddle move in a cellular grid. • Unseen deterministic physics moves ball in grid. • Cell state: (ball state, paddle state) • Ball state: (empty, present, moving left/right/up/down) • Paddle state: (true false) Learner orientation: (north, south, east, west) • • Responses: (wait, forward, turn right/left) • If paddle present and orientation north or south, then forward response moves paddle also. Figure 6 – The game of Pong. Learner was trained with multiple randomly generated initial ball velocities. Procedure and results • • When the ball moved left and right, the learner moved with the ball. • When the ball moved up or down, the learner moved to the paddle and moved it up or down. • This was the challenge: remembering ball state while traversing empty cells to the paddle so as to move it correctly, then to turn and return to ball for next input. Testing on random games: 100% successful. Conclusion This is an early exploration of the morphognosis model. The positive results on the three tasks prompt future investigation. Possible next tasks include: • Web building. Can a space-time memories of building one or more training webs allow one to be built in a quasi-novel environment? Food foraging social signaling. Bees retain memories of foraging food sources that they communicate to other bees through instinctive dancing. Can this task be cast into the model? • The Java code is available at https://github.com/portegys/MoxWorx References Bousso, R. (2002). "The holographic principle". Reviews of Modern Physics. 74 (3): 825–874. arXiv:hep- th/0203101 Brooks, R. (1999). Cambrian Intelligence: The Early History of the New AI. Cambridge MA: The MIT Press. ISBN 0-262-52263-2 Halliwell, J.J.; et al. (1994). Physical Origins of Time Asymmetry. Cambridge. ISBN 0-521-56837-4. Hawkins, J. (2004). On Intelligence (1 ed.). Times Books. p. 272. ISBN 0805074562 Hoffman, D.D. (2009). The interface theory of perception: Natural selection drives true perception to swift extinction. In: Object Categorization: Computer and Human Vision Perspectives. Ed.: S.J. Dickinson, A. Leonardis, B. Schiele & M.J. Tarr. Cambridge, Cambridge University Press: 148-165. Hoffmann, M. & Pfeifer, R. (2011). The implications of embodiment for behavior and cognition: animal and robotic case studies, in W. Tschacher & C. Bergomi, ed., 'The Implications of Embodiment: Cognition and Communication', Exeter: Imprint Academic, pp. 31-58. Human Brain Project, Framework Partnership Agreement (2015). https://www.humanbrainproject.eu/documents/10180/538356/FPA++Annex+1+Part+B/41c4da 2e-0e69-4295-8e98-3484677d661f Jonas, E., Kording, K. (2016). Could a neuroscientist understand a microprocessor? bioRxiv 055624; doi: https://doi.org/10.1101/055624 Kaiser, U. B. (2014). Editorial: Advances in Neuroscience: The BRAIN Initiative and Implications for Neuroendocrinology. Molecular Endocrinology, 28(10), 1589–1591. http://doi.org/10.1210/me.2014- 1288 Moskowitz, C. (2016). Are We Living in a Computer Simulation? Scientific American. Mühlenbein, H. (2009). Computational Intelligence: The Legacy of Alan Turing and John von Neumann, in Computational Intelligence Collaboration, Fusion and Emergence. Editors: Mumford, C. L. (Ed.) Volume 1 of the series Intelligent Systems Reference Library pp 23-43. Numenta White paper (2011). http://numenta.org Portegys, T., Pascualy, G., Gordon, R., McGrew, S., Alicea, B. (2017). Morphozoic: cellular automata with nested neighborhoods as a metamorphic representation of morphogenesis. In Multi-Agent Based Simulations Applied to Biological and Environmental Systems. ISBN: 978-1-5225-1756-6. Rickles, D. (2014). A Brief History of String Theory: From Dual Models to M-Theory. Springer Science & Business Media. ISBN 978-3-642-45128-7. Sanders, L. (2015). How the brain perceives time. ScienceNews. https://www.sciencenews.org/article/how-brain-perceives-time Stetka, B. (2016). From Psychedelics To Alzheimer's, 2016 Was A Good Year For Brain Science. http://www.npr.org/sections/health-shots/2016/12/31/507133144/from-psychedelics-to-alzheimers- 2016-was-a-good-year-for-brain-science Turing, A.M. (1952). The chemical basis of morphogenesis. Phil. Trans. Roy. Soc. London B237, 37-72. Vorhees, C. V., & Williams, M. T. (2014). Assessing Spatial Learning and Memory in Rodents. ILAR Journal, 55(2), 310–332. http://doi.org/10.1093/ilar/ilu013. Wolfram, S. (2002). A New Kind of Science. Wolfram Media. ISBN-10: 1579550088. Wolfram, S. (2015). What Is Spacetime, Really? Stephen Wolfram Blog. http://blog.stephenwolfram.com/2015/12/what-is-spacetime-really/
1112.1330
1
1112
2011-12-06T15:59:30
Emotional control - conditio sine qua non for advanced artificial intelligences?
[ "q-bio.NC", "cs.AI" ]
Humans dispose of two intertwined information processing pathways, cognitive information processing via neural firing patterns and diffusive volume control via neuromodulation. The cognitive information processing in the brain is traditionally considered to be the prime neural correlate of human intelligence, clinical studies indicate that human emotions intrinsically correlate with the activation of the neuromodulatory system. We examine here the question: Why do humans dispose of the diffusive emotional control system? Is this a coincidence, a caprice of nature, perhaps a leftover of our genetic heritage, or a necessary aspect of any advanced intelligence, being it biological or synthetic? We argue here that emotional control is necessary to solve the motivational problem, viz the selection of short-term utility functions, in the context of an environment where information, computing power and time constitute scarce resources.
q-bio.NC
q-bio
Emotional control - conditio sine qua non for advanced artificial intelligences? Claudius Gros Abstract Humans dispose of two intertwined information processing pathways, cognitive information processing via neural firing patterns and diffusive volume control via neuromodulation. The cognitive information processing in the brain is traditionally considered to be the prime neural correlate of human intelligence, clin- ical studies indicate that human emotions intrinsically correlate with the activation of the neuromodulatory system. We examine here the question: Why do humans dispose of the diffusive emo- tional control system? Is this a coincidence, a caprice of nature, perhaps a leftover of our genetic heritage, or a necessary aspect of any advanced intelligence, being it biological or synthetic? We argue here that emotional control is necessary to solve the motivational prob- lem, viz the selection of short-term utility functions, in the context of an environ- ment where information, computing power and time constitute scarce resources. 1 Introduction The vast majority of research in artificial intelligences is devoted to the study of algorithms, paradigms and philosophical implications of cognitive information pro- cessing, like conscious reasoning and problem solving [1]. Rarely considered is the motivational problem - a highly developed AI needs to set and select its own goals and tasks autonomously. We believe that it is necessary to consider the motivational problem in the context of the observation that humans are infused with emotions, possibly to a greater extend than any other species [2]. Emotions play a very central role in our lives, in literature and human culture in general. Is this predominance of emotional states a 1 1 0 2 c e D 6 ] . C N o i b - q [ 1 v 0 3 3 1 . 2 1 1 1 : v i X r a Claudius Gros Institute for [email protected] Theoretical Physics, Goethe University Frankfurt, e-mail: 1 2 Claudius Gros coincidence, a caprice of nature, perhaps a leftover from times when we were still 'primitives and brutes', or perhaps a necessary aspect of any advanced intelligence? The motivational problem is about the fundamental conundrum that all living in- telligences face. From the myriads of options and behavioral strategies it needs to select a single route of action at any given time. These decisions are to be taken considering three limited resources, the information disposed of about the present and the future state of the world, the time available to take the decision and the com- putational power of its supporting hard- or wetware. Here we argue that emotional control is deeply entwined with both short- and long-term decision making and al- lows to compute in real time approximate solutions to the motivational problem. When considering the relation between emotional control and the motivational problem one needs to discuss the nature of non-biological intelligences for which this issue is of relevance. We believe that, in the long term, there will be two major developmental tracks in AI research - focused artificial intelligences and organismic universal synthetic intelligences. We believe that the emotional control constitutes an inner core functionality for any universal intelligence and not a secondary adden- dum. 2 Intelligent Intelligences We start with some terminology and a loose categorization of possible forms of intelligence. Focused Artificial Intelligences We will use the term focused AI for what consti- tutes today's mainstream research focus in artificial intelligence and robotics. These are highly successful and highly specialized algorithmic problem solvers like the chess playing program Deep Blue [8], the DARPA-like autonomous car driving sys- tems [9] and Jeopardy software champion Watson [10]. Focused artificial intelligences are presently the only type of artificial intelli- gences suitable for commercial and real-world applications. In the vast majority of today's application scenarios a focused intelligence is exactly what is needed, a reliable and highly efficient solution solver or robotic controller. Focused AIs may be able to adapt to changing demands and have some forms of built-in, application specific learning capabilities. They are however characterized by two features. • Domain specificity A chess playing software is not able to steer a car. It is much more efficient to develop two domain specific softwares, one for chess and one for driving, than to develop a common platform. • Maximal a priori information The performance real-world applications are gen- eraly greatly boosted when incorporating a maximal amount of a priori infor- mation into the architecture. Deep Blue contains the compressed knowledge of hundreds of years of human chess playing, the DARPA racing car software the Diffusive emotional control 3 Newton laws of motion and friction, the algorithms do not need to discover and acquire this knowledge from proper experiences. Focused AI sees a very rapid development, increasingly driven by commercial ap- plications. They will become extremely powerful within the next decades and it is questionable whether alternative forms of intelligences, whenever the may be avail- able in the future, will ever be able to compete with focused AI on economical grounds. It may very well be, though difficult to foretell, that focused AI will al- ways yield a greater return on investment than more general types of intelligences with their motivational issues. Synthetic Intelligences The term 'artificial intelligence' has been used and abused in myriads of ways over the past decades. It is standardly in use for mainstream AI research, or focused AI as described above. We will use here the term synthetic intelligence for alternative forms of intelligences, distinct from todays mainstream route of AI and robotics research. Universal Intelligences It is quite generally accepted that the human brain is an exemplification of 'universal' or 'generic' intelligence. The same wetware and neu- ral circuitry can be used in many settings - there are no new brain protuberances being formed when a child learns walking, speaking, operating his fairy-tale player or the alphabet at elementary school. There are parts of the brain more devoted to vi- sual, auditory or linguistic processing, but rewiring of the distinct incoming sensory data streams will lead to reorganization processes of the respective cortical neural circuitry allowing it to adapt to new tasks and domains. The human brain is extremely adaptive, a skilled car driver will experience, to a certain extend, its car as an extension of his own body. A new brain-computer inter- ference, when available in the future, will be integrated and treated as a new sensory organ, on equal footing with the biological pre-existing senses. Human intelligence is to a large extend not domain specific, its defining trait is universality. Organismic Intelligences An 'organismic intelligence' is a real-world or simu- lated robotic system which has the task to survive. It is denoted organismic since the survival task is generically formulated as the task to keep the support unit, the body, functional [3, 4]. Humans are examples of organismic intelligences. An organismic synthetic in- telligence may be universal, but not necessarily. The term 'organismic' is not to be confused with 'embodiment'. Embodied AI deals with the question whether consid- ering the physical functionalities of robots and bodies is helpful, of even essential, for the understanding of cognitive information processing and intelligence in gen- eral [5, 6, 7]. Cognitive System The term 'cognitive system' is used in various ways in the lit- erature, mostly as a synonym for a cognitive architecture, viz for an information processing domain-specific software. I like to reserve the term cognitive system for an intelligence which is both universal and organismic, may it be biological or syn- thetic. 4 Fig. 1 Illustration of the (hypothetical) complexity co- nundrum, which regards the speculation that the mental capabilities of biological or synthetic intelligences (right) might be systematically too low to fully understand the complexity of their own supporting cognitive architec- tures (left). In this case the singularity scenario would be void. complexity complexity Claudius Gros s e i t i l i b a p a c l a u t c e l l t e n i f o e r u t c e t i h c r a e v i t i n g o c f o Humans are biological cognitive systems in this sense and most people would expect, one can however not foretell with certainty, that 'true' or 'human level AI' would eventually be realized as synthetic cognitive systems. It is an open and unre- solved questions, as a matter of principle, whether forms of human level AI which are not cognitive systems in above sense, are possible at all. Human Level Artificial Intelligences An ultimate goal of research in artificial and synthetic intelligences is to come up with organizational principles for intelligences of human or higher level. How and when this goal will be achieved is presently in the air, a few aspects will be discussed in the next section. This has not precluded an abundance of proposals on how to test for human-level intelligences, like the Turing test [11] or the capability to perform scientific research. Some people believe that human intelligence will have been achieved when we do not notice it. The Complexity Conundrum Regarding the issue when and how humanity will develop human level intelligences we discuss here shortly the possible occurance of a 'complexity paradox', for which we will use the term complexity conundrum. Every intelligence arises form a highly organized soft- or wetware. One may as- sume, though this is presently nothing more than a working hypothesis, that more and more complex brains and software architectures are needed for higher and higher intelligences. The question is than, whether a brain with a certain degree of complexity will give raise to a level on intelligence capable to understand its own wetware, compare Fig. 1. It may be, as a matter of principle, that the level of com- plexity a certain level of intelligence is a able to handle is always below the level of complexity of its own supporting architecture. This is really a handwaving and rather philosophical question with many open ends. Nevertheless one may speculate whether the apparent difficulties of present- day neuroscience research to carve out the overall working principles of the brain may be in part due to a complexity conundrum. Equivalently, considering the suc- cesses and the failures of over half a century of AI research, our present near-to complete ignorance of the overall architectural principles necessary for the develop- ment of eventual human level AI may be routed similarly in either a soft or a strong version of the complexity conundrum. Diffusive emotional control 5 Fig. 2 Mainstream architecture for a hypothetical human-level artificial intelligence. The motiva- tional problem would be delegated to a secondary level responsible of selecting appropriate mod- ules for problems and tasks which are not autonomously generated but presumably presented to the AI by human supervisors. Higher cognitive states like consciousness are sometimes postulated to emerge spontaneously with raising complexity from self-organizational principles, emotional control is generically regarded as a later-stage add-on, if at all. The complexity conundrum would however not, even if true, preclude humanity to develop human level artificial or synthetic intelligences in the end. As a last re- sort one may proceed by trial and error, viz using evolutionary algorithms, or via brute force reverse engineering, if feasible. The notion of a complexity conundrum is relevant also to the popular concept of a singularity, a postulated runaway self improving circle of advanced intelligences [12, 13]. The complexity conundrum, if existing in any form, would render the notion of a singularity void, as it would presumably apply to intelligences at all levels. 3 Routes to Intelligence There are presently no roadmaps, either individually proposed or generally ac- cepted, for research and development plans leading to the ultimate goal of highly advanced intelligences. Nevertheless there are two main, conceptually distinct, ap- proaches. 3.1 From focused to general intelligence? The vast majority of present-day research efforts is devoted to the development of high-performing focused intelligences. It is to be expected that we will see advances, within the next decades, along this roadmap for hundreds and many more applica- tion domains. There is no generally accepted blueprint on how to go beyond focused intelli- gences, a possible scenario is presented in Fig. 2. A logical next step would be to 6 Claudius Gros hook up a vast bank of specialized algorithms, the focused intelligences, adding a second layer responsible for switching between them. This second layer would then select the algorithm most appropriate for the problem at hand and could contain suitable learning capabilities. This kind of selection layer constitutes a placebo for the motivational problem, the architecture presented in Fig. 2 would not be able to autonomously generate its own goals. This is however not a drawback for industrial and for the vast majority of real-world applications, for which the artificial intelligence is expected just to efficiently solve problems and tasks presented to it by human users and supervisors. In a third step it is sometimes expected that cognitive architectures may develop spontaneously consciousness with raising levels of complexity. This speculation, particularly popular with science-fiction media, is presently void of any supporting or contrarian scientific basis [14, 15]. Interesting is the tendency of mainstream AI to discuss emotions as secondary features, mostly useful to facilitate human-robot interactions [16]. Emotions are generically not attributed a central role in cognitive architectures withing mainstream AI. One could imagine that the kind of cognitive architecture presented in Fig. 2 ap- proaches, with the expansion of its basis of focused intelligences, step by step the goal of a universal intelligence able to handle nearly any conceivable situation. It is unclear however which will be the pace of progress towards this goal. It may be that progress will be initially very fast, slowing then however down substantially when artificial intelligence with elevated levels of intellectual capabilities have been successfully developed. This kind of incremental slowing-down is not uncommon for the pace of scientific progress in general. Life expectancy has been growing lin- early, to give an example, over the last two centuries. The growth in life expectancy is extremely steady and still linear nowadays, despite very rapidly growing med- ical research efforts. Not only in economics, but also in science there are generic decreasing returns on growing investments. Similarly, vast increases in the number and in the power of the underlying array of focused intelligences may, in the end, lead to only marginal advances towards universality. 3.2 Universal learning systems The only real-world existing example of an advanced cognitive system is the mam- malian brain. It is hence reasonable to consider biologically inspired cognitive ar- chitectures. Instead of reverse engineering the human brain, one tries then to deeply understand the general working principles of the human brain. There are good arguments that self-organization and general working principles are indeed dominant driving forces both for the development of the brain and for its ongoing functionality [17, 18]. Due to the small number of genes in the human genome, with every gene encoding only a single protein, direct genetic encoding of specific neural algorithms has either to be absent all together in the brain or be limited to only a very small number of vitally important features. Diffusive emotional control 7 Fig. 3 Architecture for biologically inspired universal synthetic intelligences, viz of cognitive sys- tems. The basis would be given by a relatively small number of genetically encode universal op- erating principles, with emotional control being central for the further development through self- organized learning processes. How consciousness would arise in this setting is not known presently, it is however regarded as a prerequisite for higher intellectual capabilities such as abstract reason- ing and knowledge specialization. It is hence plausible that a finite number of working principles, possibly as small as a few hundred, may be enough for a basic understanding of the human brain, with higher levels of complexity arising through self-organization. Two examples for general principles are 'slowness' [19] for view-invariant object recognition and 'universal prediction tasks' [3] for the autonomous generation of abstract concepts. Universality, in the form of operating principles, lies therefore at the basis of highly developed cognitive systems, compare Fig. 3. This is in stark contrast to mainstream AI, where universality is regarded as the long-term goal, to be reached when starting from advanced focused intelligences. One of the genetically encoded control mechanisms at the basis of a cognitive system is emotional control, which we will discuss in more detail in the next section. Emotional control is vitally important for the functioning of a universal learning system, and not a secondary feature which may be added at a later stage. • Learning In the brain two dominant learning mechanisms are known. Hebbian- type synaptic plasticity which is both sub-conscious and automatic, and reward- induced learning, with the rewards being generated endogenously through the neuromodulatory control system, the later being closely associated with the ex- perience of motions. • Goal selection Advanced cognitive systems are organismic and hence need to constantly select their short- and long term goals autonomously, with emotional weighing of action alternatives playing a central role. It is not a coincidence, that the emotional control system is relevant for above two functionalities, which are deeply inter-dependent. There can be no efficient goal selection without learning from successes and failure, viz without reward induced learning processes. 8 Claudius Gros motor action output autonomous sensory data input stream Fig. 4 Fast and slow variables have distinct functionalities in the brain, with the operating modus (mood) being set by the slow variables and the actual cognitive processes, which are either input induced or autonomous [20, 21], being performed by the fast variables. The adaption of the slow variables (metalearning) is the task of the diffusive neuromodulatory system (emotional control). 4 Emotional Control Emotions are neurobiologically not yet precisely defined. There are however sub- stantial indications from clinical studies that emotions are intrinsically related to either the tonic or the phasic activation of the neuromodulatory system [22]. For this reason we will denote the internal control circuit involving neuromodulation, compare Fig. 4, emotional control. We will also use the expression diffusive emo- tional control since neuromodulation acts as a diffusive volume effect. One needs to differentiate between the functionality of emotions in the con- text of cognitive system theory, discussed here, and the experience (the qualia) of emotions. It is presently an open debate whether the body is necessary for the experience of emotions and moods, which may be induced by the propri- oceptual sensing of secondary bodily reactions [23]. The origin of emotional experience is not subject of our deliberations. 4.1 Neuromodulation and metalearning Animals dispose of a range of operating modi, which one may identify with moods or emotional states. A typical example of a set of two complementary states is ex- ploitation vs. exploration: When exploitive the animal is focused, concentrated on a given task and decisive. In the explorative state the animal is curious, easily dis- tracted and prone to learn about new aspects of his environment. These moods are Diffusive emotional control 9 induced by the tonic, respectively the phasic activation of the neuromodulatory sys- tem [24], the main agents being Dopamine, Serotonin, Norepinephrine and Acetyl- choline. When using the language of dynamical system theory we can identify the task of the neuromodulatory system with metalearning [25]. Any complex system disposes of processes progressing on distinct time scales. There may be in principle a wide range of time scales, the simplest classification is to consider slow and fast processes driven respectively by slow and fast variables. Cognitive information processing is performed in the brain through neural fir- ing and synaptic plasticity, corresponding to the fast variables in terms of dynam- ical system theory [3]. The general operating modus of the neural circuitry, like the susceptibility to stimuli, the value of neural thresholds or the pace of synaptic plasticities are slow degrees of freedom. The adaption of slow degrees of freedom to changing tasks is the realm of metalearning, which in the brain is preformed through the neuromodulatory system, compare Fig. 4. Metalearning is a necessary component of any complex dynamical system and hence also for any evolved synthetic or biological intelligence. It is therefore not sur- prising that the human brain disposes of a suitable mechanism. Metalearning is also intrinsically diffusive, as it involves the modulation not of individual slow variables, metalearning is about the modulation of the operating modus of entire dynamical subsystems. It is hence logical that the metalearning circuitry of the brain involves neuromodulatory neurons which disperse their respective neuromodulators, when activated, over large cortical or subcortical areas, modulating the behavior of down- stream neural populations in large volumes. An interesting and important question regards the guiding principles for meta- learning. An animal has at its disposal a range of distinct behaviors and moods, foraging, social interaction, repose, exploration, and so on. Any cognitive system is hence faced with a fundamental time allocation problem, what to do over the course of the day. The strategy will in general not be to maximize time allocation of one type of behavior, say foraging, at the expense of all others, but to seek an equilibrated distribution of behaviors. This guiding principle of metalearning has been denoted 'polyhomeostatic optimization' [26]. 4.2 Emotions and the motivational problem It is presently unclear what distinguishes metalearning processes which are expe- rienced as emotional from those which are unconscious and may hence be termed 'neutral'. It has been proposed that the difference may be that emotional control has a preferred level of activation, neutral control not [27, 28]. When angry one generally tries behavioral strategies aimed at reducing the level of angriness and internal rewards are generated when successful. In this view emotional control is in- trinsically related to behavior and learning, in agreement with neuro-psychological observations [24, 2, 29]. 10 Claudius Gros Emotional states induce, quite generically, problem solving strategies. The cog- nitive system either tries to stay in its present mood, in case it is associated with positive internal rewards, or looks for ways to remove the causes for its current emotional state, in case it is associated with negative internal rewards. Emotional control hence represents a way, realized in real-world intelligences, to solve the mo- tivational problem, determining the utility function the intelligence tries to optimize at any given point of time. A much discussed alternative to emotional control is straightforward maximiza- tion of an overall utility function [30]. This paradigm is highly successful when applied to limited and specialized tasks, like playing chess, and is as such important for any advanced intelligence. Indeed we argue that emotional control determines the steady-state utility function. As an example consider playing chess. Your util- ity function may either consist in trying to beat the opponent chess player or to be defeated by your opponent (in a non-so-evident way) when playing together with your son or daughter. These kinds of utility functions are shaped in real life by our emotional control mechanisms. It remains however doubtful whether it would be possible to formulate an overall, viz a long-term utility function for a universal intelligence and to compute in real time its gradients. Even advanced hyper-intelligences will dispose of only an expo- nentially small knowledge about the present and the future state of the world, pre- diction tasks and information acquisition is generically NP-hard (non-polynomial) [31, 32, 33]. Time and computing power (however large it may be) will forever remain, relatively seen, scarce resources. It is hence likely that advanced artificial intelligences will be endowed with 'true' synthetic emotions, the perspective of a hyper-intelligent robot waiting emotionless in its corner, until its human boss calls him to duty, seems implausible [34, 35, 36]. Any advanced intelligence needs to be a twofold universal learning system. The intelligent system needs to be on one side able to acquire any kind of information in a wide range of possible environments and on the other side to determine au- tonomously what to learn, viz solve the time allocation problem. The fact that both facets of learning are regulated through diffusive emotional control in existing ad- vanced intelligences suggests that emotional control may be a conditio sine qua non for any, real-world or synthetic, universal intelligence. 5 Hyper-emotional trans-human intelligences? Looking around at the species on our planet one may surmise that increasing cog- nitive capabilities go hand in hand with rising complexity and predominance of emotional states [2]. The rational is very straightforward. An animal with say only two behavioral patterns at its disposition, e.g. sleeping and foraging, does not need dozens of moods and emotions, in contrast to animals with a vast repertoire of po- tentially complex behaviors. Diffusive emotional control 11 This observation is consistent with the theory developed here, that metalearning as a diffusive emotional control system is a necessary component for any synthetic and biological intelligence. It is also plausible that the complexity the metalearning control needs to increase adequately with increasing cognitive capacities. It is hence amusing to speculate, whether synthetic intelligences with higher and higher cognitive capabilities may also become progressively emotional. Super- human intelligences would then also be hyper-emotional. An outlook in stark con- trast to the mainstream view of hyper-rational robots, which presumes that emo- tional states will be later-stage addendums to high performing artificial intelli- gences. Acknowledgements I acknowledge lively discussions and feedback at the conference on the Phi- losophy and Theory of Artificial Intelligence, PT-AI, Thessaloniki, October 3-4 (2011). References 1. S.J. Russell, P. Norvig, Artificial intelligence: a modern approach, Prentice Hall (2010). 2. R.J. Dolan, Emotion, cognition, and behavior, Science 298, 1191 (2002). 3. C. Gros, Complex and Adaptive Dynamical Systems, A Primer, Springer (2008); second edi- tion 2010. 4. E. Di Paolo, Organismically-inspired robotics: Homeostatic adaptation and natural teleology beyond the closed sensorimotor loop, in Dynamical Systems Approach to Embodiment and Sociality, ed. by K. Murase, T Asakura, T., pages 19-42. Advanced Knowledge International (2003). 5. M.L. Anderson, Embodied cognition: A field guide, Artificial intelligence 149, 91–130 (2003). 6. R. Pfeifer, J. Bongard, S. Grand, How the body shapes the way we think: a new view of intel- ligence, MIT Press (2007). 7. T. Froesea, T. Ziemkeb, Enactive artificial intelligence: Investigating the systemic organiza- tion of life and mind, Artificial Intelligence 173, Pages 466-500 (2009). 8. M. Campbell, Deep Blue, Communications of the ACM 42, 65 (1999). 9. S. Thrun, Winning the darpa grand challenge, Machine Learning: ECML 2006, 4–4 (2006). 10. D. Ferrucci et al., Building Watson: An overview of the DeepQA project, AI Magazine 31, 59–79 (2010). 11. A. Turing, Can machines think?, Mind 59, 433–460 (1950). 12. V. Vinge, The coming technological singularity, Feedbooks (1993). 13. D. Chalmers, The Singularity: A philosophical analysis, Journal of Consciousness Studies 17, 7–65 (2010). 14. G. Tononi, G.M. Edelman, Consciousness and complexity, Science 282, 1846 (1998). 15. C. Koch, G. Laurent, Complexity and the nervous system, Science 284, 96 (1999). 16. J. Vallverdu, D. Casacuberta (Eds), Handbook of Research on Synthetic Emotions and Socia- ble Robotics: New Applications in Affective Computing and Artificial Intelligence, IGI-Global (2009). 17. T. Kohonen, Self-organized formation of topologically correct feature maps, Biological Cy- bernetics 43, 59–69 (1982). 18. H. Haken, Self-organization of brain function, Scholarpedia 3, 2555 (2008). 19. P. Foldi´ak, Learning invariance from transformation sequences, Neural Computation, 3, 194– 200 (1991). 12 Claudius Gros 20. C. Gros, Cognitive computation with autonomously active neural networks: An emerging field, Cognitive Computation 1, 77–99 (2009). 21. C. Gros, G. Kaczor, Semantic learning in autonomously active recurrent neural networks, Logic Journal of IGP 18, 686 (2010). 22. J.M. Fellous, Neuromodulatory basis of emotion, The neuroscientist 5, 283 (1999). 23. L.F. Barrett, B. Mesquita, K.N. Ochsner, J.J. Gross, The experience of emotion, Annual re- view of psychology 58, 373 (2007). 24. J.L. Krichmar, The neuromodulatory system: A framework for survival and adaptive behavior in a challenging world, Adaptive Behavior 16, 385 (2008). 25. K. Doya, Metalearning and neuromodulation, Neural Networks, 15, 495–506 (2002). 26. D. Markovic, C. Gros, Self-organized chaos through polyhomeostatic optimization, Physical Review Letters 105, 068702 (2010). 27. C. Gros, Emotions, diffusive emotional control and the motivational problem for autonomous cognitive systems, in Handbook of Research on Synthetic Emotions and Sociable Robotics: New Applications in Affective Computing and Artificial Intelligence, ed. by J. Vallverdu, D. Casacuberta, IGI-Global (2009). 28. C. Gros, Cognition and Emotion: Perspectives of a Closing Gap, Cognitive Computation 2, 78 (2010). 29. R.F. Baumeister, K.D. Vohs, C. Nathan DeWall, How emotion shapes behavior: Feedback, anticipation, and reflection, rather than direct causation, Personality and Social Psychology Review 11, 167 (2007). 30. M. Hutter, Universal artificial intelligence: Sequential decisions based on algorithmic prob- ability, Springer (2005). 31. D.M. Chickering, D. Heckerman, C. Meek, D. Madigan, Learning Bayesian networks is NP- hard, Microsoft Research, TechReport MSR-TR-94-17 (1994). 32. Z. Nikoloski, S. Grimbs, P. May, J. Selbig, Metabolic networks are NP-hard to reconstruct, Journal of theoretical biology 254, 807–816 (2008). 33. D. Sieling, Minimization of decision trees is hard to approximate, Journal of Computer and System Sciences 74, 394–403 (2008). 34. M.A. Arbib, J.M. Fellous, Emotions: from brain to robot, Trends in cognitive sciences 8, 554–561 (2004). 35. T. Ziemke, R. Lowe, On the role of emotion in embodied cognitive architectures: From or- ganisms to robots, Cognitive computation, 1, 104–117 (2009). 36. D. Parisi, G. Petrosino, Robots that have emotions, Adaptive Behavior 18, 453 (2010).
1203.3966
1
1203
2012-03-18T15:53:01
Grid Alignment in Entorhinal Cortex
[ "q-bio.NC", "cond-mat.dis-nn", "nlin.AO", "nlin.PS", "physics.bio-ph" ]
The spatial responses of many of the cells recorded in all layers of rodent medial entorhinal cortex (mEC) show a triangular grid pattern, and once established might be based in part on path-integration mechanisms. Grid axes are tightly aligned across simultaneously recorded units. Recent experimental findings have shown that grids can often be better described as elliptical rather than purely circular and that, beyond the mutual alignment of their grid axes, ellipses tend to also orient their long axis along preferred directions. Are grid alignment and ellipse orientation the same phenomenon? Does the grid alignment result from single-unit mechanisms or does it require network interactions? We address these issues by refining our model, to describe specifically the spontaneous emergence of conjunctive grid-by-head-direction cells in layers III, V and VI of mEC. We find that tight alignment can be produced by recurrent collateral interactions, but this requires head-direction modulation. Through a competitive learning process driven by spatial inputs, grid fields then form already aligned, and with randomly distributed spatial phases. In addition, we find that the self-organization process is influenced by the behavior of the simulated rat. The common grid alignment often orients along preferred running directions. The shape of individual grids is distorted towards an ellipsoid arrangement when some speed anisotropy is present in exploration behavior. Speed anisotropy on its own also tends to align grids, even without collaterals, but the alignment is seen to be loose. Finally, the alignment of spatial grid fields in multiple environments shows that the network expresses the same set of grid fields across environments, modulo a coherent rotation and translation. Thus, an efficient metric encoding of space may emerge through spontaneous pattern formation at the single-unit level.
q-bio.NC
q-bio
Grid Alignment in Entorhinal Cortex Bailu Si,1 Emilio Kropff,2 and Alessandro Treves1, 2 1Sector of Cognitive Neuroscience, International School for Advanced Studies, via Bonomea 265, 34136 Trieste, Italy∗ 2Kavli Institute for Systems Neuroscience and Center for the Biology of Memory, Norwegian University of Science and Technology, 7489 Trondheim, Norway† The spatial responses of many of the cells recorded in all layers of rodent medial entorhinal cortex (mEC) show a triangular grid pattern, which appears to provide an accurate population code for position, and once established might be based in part on path-integration mechanisms. Competing models, each partially contradicted by experimental observations, try to explain how the grid-like pattern emerges in terms of network interactions, or of interactions with theta oscillations or, the one we have proposed, of mere single-unit mechanisms. Grid axes are tightly aligned across simultaneously recorded units. Recent experimental findings have shown that grids can often be better described as elliptical rather than purely circular and that, beyond the mutual alignment of their grid axes, ellipses tend to also orient their long axis along preferred directions. Are grid alignment and ellipse orientation the same phenomenon? Does the grid alignment result from single-unit mechanisms or does it require network interactions? We address these issues by refining our model, to describe specifically the spontaneous emergence of conjunctive grid-by-head-direction cells in layers III, V and VI of mEC. We find that tight alignment can be produced by recurrent collateral interactions, but this requires head-direction modulation. Through a competitive learning process driven by spatial inputs, grid fields then form already aligned, and with randomly distributed spatial phases. In addition, we find that the self- organization process is influenced by the behavior of the simulated rat. The common grid alignment often orients along preferred running directions. The shape of individual grids is distorted towards an ellipsoid arrangement when some speed anisotropy is present in exploration behavior. Speed anisotropy on its own also tends to align grids, even without collaterals, but the alignment is seen to be loose. Finally, the alignment of spatial grid fields in multiple environments shows that the network expresses the same set of grid fields across environments, modulo a coherent rotation and translation. Thus, an efficient metric encoding of space may emerge through spontaneous pattern formation at the single-unit level, but it is coherent, hence context-invariant, if aided by collateral interactions. Keywords: Hippocampus; Entorhinal cortex; Grid cells; Conjunctive grid-by-head-direction cells; Firing rate adaptation; Competitive network; Remapping I. INTRODUCTION Internal representations of space appear necessary for any agent, such as a rat or a robot, to move around in a spatial context and to distinguish between different con- texts to which food or danger, for example, may be as- sociated. Spatial cognition and memory have been long investigated in rodents, and an impressive body of results point at the major role being played by the hippocampus and related cortices, a region that in humans has been as- sociated with the neural basis of episodic memory forma- tion, since1. Forty years ago, place cells were discovered in the rat hippocampus, showing specific firing activity whenever the rat enters a specific portion of the environ- ment, the place field2. Head direction (HD) cells were first found in the rat postsubiculum, firing steadily when the animal points its head towards a specific direction in the environment3. These two distinct systems provide simple, distributed population coding of the location and heading direction of the animal. In recent years, cells with more complex spatial codes have been discovered in the rat medial entorhinal cortex (mEC), which sends strong projections to the hippocam- pus. Grid cells, found to be particularly abundant in layer II of mEC (perhaps about half of stellate cells there) have multiple firing fields positioned on the vertices of remarkably regular triangular grids, spanning the envi- ronment which the animal explores4,5. Conjunctive grid- by-head-direction cells, found along a smaller proportion of pure grid cells in the deeper layers of mEC, show fir- ing selectivity to HD in addition to (perhaps slightly less precise) spatial tuning as grid cells6. The precise geomet- ric tessellation provided by the activity of grid cells has stimulated a series of experimental and theoretical stud- ies on the mechanisms underlying the emergence and the function of grid cells7. Theoretical models of grid cell formation may be grouped in three main categories. The first type of mod- els shows how grid fields may emerge from the attractor states induced, in continuous attractor networks, from structured recurrent connectivity8–12. The spatial lay- out of recurrent collateral connections of the network, assumed to be permanent or at least present during the developmental stage of grid cell formation, ensures that triangular spatial firing patterns are stable states of re- current dynamics. Grid units in a continuous attractor network model are able to perform path integration by propagating the activity in the network in correspon- dence with the movement of the animal, in a way similar to previous models for place cells or HD cells13–15. The second class of models relates the periodic firing of grid cells to sub-threshold membrane potential oscil- lations, and proposes that grid fields may result from interference between a theta-related baseline oscillator and other velocity-controlled oscillations which originate either in different dendrites of a neuron or in different neurons16–20. The frequency difference between these velocity-controlled oscillators is small, so that the low frequency "envelope" of the interference pattern corre- sponds to the spatial periodicity of grid cells. This hy- pothesis is in part supported by recent findings that the spatial periodicity of grid cells is susceptible to the sup- pression of theta oscillations by pharmacological silencing of the medial septum21,22. The third class of models on grid cell formation argues that grid fields may not require detailed ad hoc mecha- nisms like structured connectivity or theta oscillations, rather they may emerge spontaneously from a general feature of cortical cell activity, like firing rate adapta- tion23 or, equivalently, other types of temporal modula- tion24. Such temporal modulation is shown in computer simulations to sculpt the spatial modulation of grid cells through a self-organization process that, averaged over a long developmental time of one or two weeks25,26, leaves as a footprint on each unit the regular periodicity found in real grid cells. A simple analytical model "explains" this spontaneous pattern formation as an unsupervised optimization process at the single-unit level23. An interesting aspect of this model, which motivates the present study, is that the emergence of perfect grid symmetry requires a perfectly isotropic distribution of rat trajectories and speed, once averaged over the long (but not infinite) learning period. Any deviation from perfect isotropy, for example because the animal spends the rele- vant developmental period (for a rat, somewhere between P15-P35, say) mainly in a rectangular cage and tends to move along the walls, or because it runs a bit faster along some particular directions, would be expected to induce distortions in all grids units. Excitingly, such de- viations in the geometry of grid maps have been observed recently and described as an ellipticity effect27. This phe- nomenon cannot be explained as small random deviations from the perfect symmetry, because of its extent and re- markable consistency across the population. Not only, as observed earlier, do grid axes show nearly the same alignment across all simultaneously recorded grid units, but also the long axes of the corresponding ellipses ap- pear to loosely orient with either one of two major ellipse clusters, depending on their spacings27. Can the mutual alignment result from the same single- unit mechanisms that produce the individual grid pat- terns? With perfect isotropic grids, the answer is ob- viously negative, and in fact Kropff and Treves (2008) pointed at a mechanism that can align grid cells at a population level, but based on collateral interactions be- tween them. With the observed anisotropies, however, the situation might be different, as both alignment and common orientation might develop along the (common) 2 anisotropy axes. In the computer simulation study reported here we find, again, that developing a tight alignment requires network interactions, through recurrent collateral con- nections with a specific structure, modulated by HD. As a first focus of this paper, we then describe how collat- eral connections may align grid cells, and then link this to the emergence of a coherent population ellipticity when the behavior of the animal is biased by running direction or speed. We then argue that anisotropy at the single- unit level is in principle sufficient to also produce some alignment, but fails in practice to make different units as tightly aligned as experimentally observed. Finally, we also investigate how well the same model describes global remapping28. When a rat is taken into a new environment, or when the environment is manip- ulated in a way that it looks new, place maps in the hippocampus undergo an apparently random shuffling, or switch from active to inactive and vice versa. Grid maps, however, behave coherently at the population level, un- dergoing a common rotation and spatial shift, in such a way that spatial overlaps between maps are preserved across rooms29. We thus address the question of how, within our model, coherent grid fields may emerge in the same network for multiple environments. The rest of the article is organized as follows. The net- work model is first introduced in Section II. In Section III grid alignment in a cylinder environment is studied, and it is compared to what occurs in a square environment in Section IV. We investigate in Section V the effects of exploration with speed anisotropy. Grid realignment in multiple environments is discussed in Section VI. Finally, the results of the study are discussed in Section VII. II. NETWORK MODEL In Fig. 1 we present a diagram of the network that we use for simulations. It is intended to model conjunc- tive grid-by-head-direction cells in layers III, V and VI of mEC. In layer V, which receives strong projections from the subiculum and the CA1 region of the hippocampus, about 20% of the putative pyramidal cells are estimated to be conjunctive grid-by-head-direction cells, along with a very small proportion of pure grid cells and more than 60% of HD cells30. In layer III, which projects to CA1, the proportion of conjunctive cells is similar to that of layer V, but the proportion of HD cells is much lower (20%) and there is an extra 20% of pure grid cells. These layers present a prominent recurrent connectivity, denser in layer V (around 12%) but also important in layer III (around 9%)31. Layer II of mEC, where the highest pro- portion and the best quality of pure grid cells are found (around 50%), lacks two critical elements that in our model go hand by hand: recurrent collateral connections and head direction information. In this paper, we assume that layer II maps can self-organize using inputs which in- clude conjunctive maps, an assumption that is discussed Wik Head direction units i k Conjunctive units Wij j "Place" units Figure 1: A sketch of the network model for conjunctive cells in deep layers of mEC. Connections are shown only for unit i. Place units are fully connected to conjunctive units. Con- junctive units are connected to all other conjunctive units without self-connections. Each conjunctive unit is assumed to be modulated by one HD unit, representing the overall effect of angular modulation from the local network. in a separate study currently in progress. Here, we focus therefore on the learning process in layers III to VI, and neglect the possible influence there of the feedback from layer II. The critical difference with previous versions of the model23 is that now we assign a central role to head direction information, which, as we show, is a key ele- ment for grid alignment. A preferred head direction, θi, is introduced arbitrarily for each conjunctive unit i, to modulate its inputs. θi is uniformly sampled from all angles. This assumption is based on our own analysis of the HD selectivity of the conjunctive cells recorded by6, in which we did not find significant clustering of HDs with respect to either the reference frame or grid axes. Each conjunctive unit receives afferent spatial inputs, which as discussed in23 we take for simplicity to arise from regu- larly arranged "place" units, and collateral inputs from other conjunctive units (Fig. 1). The overall input to conjunctive unit i at time t is then given by ht i ht i = fθi(ωt)(Xj W t−1 ij rt j + ρXk WikΨt−τ k ), (1) with ρ = 0.2 a factor weighing, relative to feed-forward inputs, collateral inputs relayed with a delay τ = 25 steps. In the model, each time step corresponds to 10 msec in real time, so the collateral interaction describe temporally diffuse and delayed processes, rather than straightforward AMPA-mediated excitation. fθi(ωt) is a tuning function that has maximal value when the cur- rent HD ωt of the simulated rat is along the preferred direction θi of the unit, as in13 fθ(ω) = c + (1 − c) exp[γ(cos(θ − ω) − 1)], (2) where c = 0.2 and γ = 0.8 are parameters determining the minimal value and the width of HD tuning. The weight W t ij connects place unit j to conjunctive unit i, while Wik connects conjunctive unit k to unit i. For simplicity, throughout the paper we only consider 3 the self-organization of the feed-forward weights, keeping the collateral weights fixed at convenient values, which we will discuss in Sect. II B. A model in which collat- eral weights are also the result of a perhaps slower self- organizing learning process will be discussed elsewhere. It is possible that conjunctive cells receive place-cell- like inputs from the hippocampus already when they de- velop their firing maps. Studies on the development of the spatial representation system in the rat may be taken to show that place cells and head-direction cells develop adult-like spatial and directional codes somewhat earlier than grid cells do25,26. The firing rate of a place unit j is modeled by an exponential function centered in its preferred firing location ~xj0 rt j = exp(−~xt − ~xj02 2σ2 p ), (3) where ~xt is the current location of the simulated rat. σp = 5cm is the width of the firing field. The place field of a place unit is arranged such that the distances to neighboring fields of other place units are about 5cm. Note that the precise and regularly arranged location- specific inputs of place units are used only to speed up the learning process. The network model works in qual- itatively the same way with broad spatial inputs, but at the cost of longer learning time23. The firing rate Ψt i of conjunctive unit i is determined by Ψt i = Ψsat arctan[gt(αt i − µt)]Θ(αt i − µt), (4) where Ψsat = π/2, so that the maximal firing rate is 1 (in arbitrary units). Θ(·) is the Heaviside function. The variable αt i represents a time-integration of the input hi, adapted by the dynamical threshold βi i − βt−1 i − βt−1 i + b1(ht−1 i + b2(ht−1 i = αt−1 αt i = βt−1 βt i − αt−1 (5) ), i ), i where βi has slower dynamics than αi (b2 is set to b2 = b1/3, and in our time steps corresponds to a time scale of roughly 300 (ms)−1, with b1 = 0.1 ≃ 100(ms)−1). Due to the adaptation dynamics, a unit that fires strongly in the recent past tends not to fire in the near future, because of a high threshold. In Eq. 4, the gain gt and threshold µt are two param- eters chosen at each time step to keep the mean activity i/NmEC of the conjunctive units and the spar- 2) within a 10% relative error bound from pre-specified values, set at a0 = 0.1 and s0 = 0.3 respectively. The values of gt and µt are determined through iterations a =Pi Ψt sity s = (Pi Ψt i)2/(NmECPi Ψt i µt,l+1 = µt,l + b3(al − a0), gt,l+1 = gt,l + b4gt,l(sl − s0). (6) Here l is the index of the iteration within simulation step t. b3 = 0.01 and b4 = 0.1 are positive step sizes in it- eration. al and sl are the mean and the sparsity of the conjunctive units when the gain gt,l and threshold µt,l are applied. The gain and threshold at the end of the iteration are used in Eq. 4 to determine the activity of conjunctive units. With constant running speed, neural fatigue is invari- ant with respect to all directions. Conjunctive units self- organize firing fields into hexagonal/triangular grids in two-dimensional space, as the virtual rat explores the environment. This configuration of fields corresponds to the minimum of an energy function23, consistent with an hexagonal tiling being the most compact one to arrange circles in two-dimensional space32. A. Learning feed-forward weights The feed-forward weights are adaptively modified ac- cording to a Hebbian rule W t ij = W t−1 ij + ǫ(Ψt j − ¯Ψt−1 irt i ¯rt−1 j ). (7) Here ¯Ψt junctive unit i and place unit j i and ¯rt j are estimated mean firing rates of con- i = ¯Ψt−1 ¯Ψt j = ¯rt−1 ¯rt i i + η(cid:0)Ψt j + η(cid:0)rt i − ¯Ψt−1 (cid:1) , j − ¯rt−1 (cid:1) , j (8) while ǫ = 0.005 is a moderate learning rate, intended to produce gradual weight change, and η = 0.05 is a time averaging factor. After each learning step, the new weights are normal- ized into unitary norm 2 W t ij = 1. Xj (9) Through competitive learning, conjunctive units that win the competition are associated to the input units that provide strong inputs. As a result, firing fields of conjunc- tive units are anchored to places where inputs are strong simultaneously with recovery from adaptation, and are stabilized as the learning proceeds. There is experimen- tal evidence of the dependence of grid fields on sensory information33,34. For example, grids can expand and con- tract in response to expansions and contractions of a fa- miliar environment. Some still unpublished observations indicate that the learning of stable grid maps in a novel environment for non-over-trained adult rats could take several days of training35. B. Collateral weights 4 ωki θk θi (xi, yi) (xk, yk) Figure 2: Assignment of the fixed collateral weight from unit k to unit i. In our network model, we include these two constraints on grid cell population activity with the help of collat- eral weights. Head direction information is used to gate the interaction between two units that fire in sequence, so that the second unit will develop fields along its preferred direction close to those of the first unit, which fires ear- lier. If the interaction were not gated by HD, the second unit would tend to form a field that is a ring surrounding the field of the first unit. The collateral weights are assigned before any learning takes place in the feed-forward weights. The weight Wik for the connection from unit k to i is calculated in the following way. Each conjunctive unit i is temporarily as- signed an auxiliary field, the location (xi, yi) of which is randomly chosen among the place fields in the environ- ment. Once the weights are fixed, these auxiliary fields have nothing to do with the conjunctive units any longer. The collateral weight between unit k and i is calculated as Wik = [fθk (ωki)fθi(ωki) exp(− d2 ki 2σ2 f ) − κ]+, (10) where [·]+ is a threshold function, with [x]+ = 0 for x < 0, and [x]+ = x otherwise. κ = 0.05 is an inhibition parameter to favor sparse weights. fθ(ω) is the HD tuning function defined in Eq. 2. ωki is the direction of the line from field k to i. dki = σf = 10cm is the width of spatial tuning. p[xi − (xk + ℓ cos ωki)]2 + [yi − (yk + ℓ sin ωki)]2 is the distance between field k and i with offset ℓ = 10cm. Normalization is performed on Wik similarly as in Eq.9. The resulting weight structure allows strong col- lateral interactions between units that have similar HDs, and meanwhile avoids co-activation, to produce grids with certain spatial shifts. In the following sections, we are going to show how, with the same set of fixed collateral weights and adapt- able feed-forward weights, units in the network develop grid fields in single and multiple environments with vari- ous boundary shapes, and with different exploration be- haviors of the virtual rat. Two characteristic features of grid cells are that they align their axes and maintain fixed spatial phases relative to each other. Such alignment and relative phases are in- variant to environmental changes, including those under which place cells in the hippocampus remap completely. III. GRID ALIGNMENT IN CYLINDER ENVIRONMENTS To test whether triangular grid fields can emerge and mutually align in our network model, we first simulate the network for a virtual rat randomly exploring a cylinder environment, where trajectories are completely isotropic. For simplicity, we assume that when the rat runs in a certain direction, it points its head toward the running direction (RD) most of the time. Therefore, in our simu- lations the head direction and the running direction are not distinguished. Each step in the simulation is taken to correspond to 10msec of real time. Each run of a simulation lasts for 8 × 106 steps, or over 22h real time. The standard speed of the rat is vs = 0.4m/s, towards the peak speed of real rats in active exploration, so that, considering the time a real rat spends not in exploration, a run of the simula- tion is intended to correspond to the developmental time scale for the emergence of grid cells25,26. At each step, the running direction is chosen by using a pseudorandom Gaussian distribution with mean equal to the direction in the previous time step and an angular standard de- viation σRD = 0.2 radians. Very importantly, the new direction cannot lead the rat outside the limits of the en- vironment. If so, the selection process is repeated until a valid direction is chosen. This has the effect of favor- ing trajectories along the walls of the environment. The overall appearance of the trajectory is comparable with those of actual rats, reflecting also the natural tendency of naıve animals to run along the walls of the enclosure. The number of place units is 500 and the number of con- junctive units is 250. Place units are fully connected to conjunctive units. Each conjunctive unit receives con- nections from all other conjunctive units except itself. In the following simulations, we use the above mentioned default parameters, including the number of units in the network, the number of simulation steps, the speed and the noise in RD, unless stated differently. In the first simulation covered in this section, the rat runs in a cylinder environment with a diameter of 125cm. Fig. 3a plots part of a typical trajectory. The RD of the rat in this environment is uniformly distributed (Fig. 3b), making the cylinder environment a suitable control con- dition to compare to anisotropy later. Fig. 4a shows four examples of the spatial maps of the conjunctive units in the network. The multiple fir- ing fields of the units locate on the vertices of triangular grids. To better reveal grid structure, the autocorrelo- grams of the spatial firing maps are calculated (Fig. 4b). The {x,y} pixel in an autocorrelogram is the correlation between the original firing map and its shifted version by the vector {x,y}. The peaks away from the center of the autocorrelogram indicate the directions and dis- tances at which the shifted map has a high overlap with the unshifted one. The six maxima that are closer to the center of each autocorrelogram are then detected, and the three maxima with positive y coordinates (white mark- ers in Fig. 4b) define the orientations of the three grid axes (the autocorrelogram is redundant by definition, the lower part resulting from the rotation of the upper part by 180 degrees). The orientation of the first grid axis, i.e. the one with the lowest angle with respect to the x axis, 5 is used as the orientation of a grid. The mean distance of the three maxima away from the center is defined as the spacing of a grid. The average spacing of the grids that self-organize in our network is about 58cm, compa- rable to the grid spacings observed in experiments5. As noted previously, the spacing results from the parameters describing neuronal fatigue in the model23. Averaged into angular bins, the angular map of the firing of each unit shows HD selectivity, matching the HD tuning curve of the unit (Fig. 4c). This obviously results from the HD modulation on the inputs to conjunctive units (Eq. 1-2). The spatial periodicity of conjunctive units is measured by the gridness score, as in6. We thus cut a ring area in each autocorrelogram with inner and outer radii chosen so as to contain the six maxima closest to the center and exclude the rest. We then correlate the original ring with its rotated versions, expecting to obtain for a perfect grid a very positive (negative) correlation coefficient for rota- tions at even (odd) multiples of 30 degrees. The gridness index is defined as the difference between the mean cor- relation for even rotations (60 and 120 degrees), and odd rotations (30, 90 and 150 degrees). Note that the grid- ness score is in the range [-2, 2]. The histogram of the gridness scores of all units is pre- sented in Fig. 3c. Most units have high gridness score. Fig. 3d shows for all the 250 conjunctive units in the simulation the locations of the three maxima defining the grid axes. They clearly cluster into three clouds, indicating similar orientations and spacings for all units. The orientation histograms of the three grid axes are dis- played in Fig. 3e, one color for each. To quantify the coherence in aligning grids, we define a cross-population alignment coherence score by the standard deviation of the orientations averaged over the three grid axes, with low values indicating tight grid alignment. In the simu- lation shown in Fig.3, the alignment coherence score is 2.967 degrees, indicating (very) tight grid alignment. While grids are aligned, the spatial phases of all grids (shown relative to the best grid, as a conventional refer- ence) are distributed in an area with diameter similar to the grid spacing (Fig. 3f). This means that grids cover the whole environment more or less randomly. The collateral weights are sparse and are stronger for units with similar HDs (broken lines in Fig. 4d), a re- lationship given by Eq. 10. When the connectivity of collateral connections is reduced to 50%, the network is still able to align grids (data not shown). Therefore, the exact strength of the collateral connectivity is not critical for the model. Collateral weights are strongest between units that are active in locations a small distance apart and that are moderately correlated in firing, due to the delay parameter τ in their inputs (Fig. 4e). The net- work is not too sensitive to τ as long as τ ≥ 13 steps, i.e. 130msec in real time. This is the time it takes the virtual rat to travel the minimal distance between two input place fields. Smaller delays do not change the grid- ness scores or the alignment of grid maps, but cause the Trajectory a 100 y 50 0 0 50 x 100 b ) % ( e g a t n e c r e P 4 3 2 1 0 0 Alignment of grid axes, std:2.967deg; m spacing:57.999cm d 125 75 25 y d −25 −125 −75 −25 25 75 125 dx 6 Gridness score histogram. m:1.592 Running direction histogram c s t i n u f o r e b m u N 50 40 30 20 10 90 Running direction (degrees) 180 270 360 0 −1 −0.5 0 0.5 Gridness 1 1.5 2 Alignment of grid axes. std:2.967 deg e 150 100 50 s t i n u f o r e b m u N f 40 20 y d 0 −20 −40 0 0 30 60 120 Orientation (degrees) 90 150 180 −40 −20 0 dx 20 40 Phase histogram 4 2 0 Figure 3: Grid alignment in a cylinder environment. (a) A trajectory of the virtual rat with 3 ×104 steps; (b) RD distribution of the trajectory in the simulation; (c) Histogram of the gridness scores of all conjunctive units in the simulation; (d) Scatter plot of the locations of the three peaks found in autocorrelograms (as shown by the white markers in Fig. 4b); (e) Histograms of the angles of the three grid axes, plotted with different colors for each axis. The alignment coherence score, i.e. standard deviation (in degrees) of the orientation averaged over the three grid axes, is indicated at the top of the panel; (f) Two dimensional histogram of spatial phases relative to the best grid. The grayscale encodes the number of units, the phases of which fall into a 2.5cm × 2.5cm spatial bin. White color represents zero, and darker colors represent larger numbers. spatial phases of the grid fields to collapse, showing a non-random phase distribution (data not shown). The phase collapse due to small τ can be understood by the associative learning between conjunctive units and place units. Small τ would cause two strongly connected con- junctive units to fire consecutively in short intervals, dur- ing which the place units show similar population activ- ity. Associative learning therefore would produce similar grid fields for both units. To avoid that, the delay be- tween two strongly connected conjunctive units needs to be long enough so that grid fields are able to anchor to different locations, resulting in random spatial phases. The exact mechanism that generates this delay is some- thing that our simple rate-based model does not attempt to describe, and it could go from several types of slow synaptic potentials to interactions with rhythms, such as phase precession. In summary, with fixed collateral weights, the grid fields of model units align to a common orientation while the relative spatial phases between them are randomly distributed in the environment. The firing of each unit is conjunctively modulated by a spatial triangular grid and by its HD input. A. Grid development with speed variation Constant running speed is not a requirement of the model. To check this, we have simulated a rat with vari- able speeds, and left unchanged the method of choosing running directions. The trajectory of the simulated rat is then composed of epochs with positive or negative ac- celerations. The lengths of epochs are Poisson random numbers with mean length 3, roughly matching avail- able behavioral data6. The speed at the end of each epoch is drawn from a truncated Gaussian distribution. A two-sided truncation is applied to keep the speed both positive and symmetrically distributed around the mean. The mean of the truncated Gaussian distribution is the same as the constant speed used in most simulations in this paper, i.e. 0.4m/s. The speed within each epoch is a linear interpolation between the speeds at the start and end of the epoch. Regardless of different levels of speed standard devia- tion, grid fields form with high gridness scores (Fig. 5a), tight grid alignment and similar spacings (data not shown). Even for large speed standard deviation, 16.1 cm/s, i.e. 40% relative to the mean speed, the changes in mean gridness, mean standard deviation of grid orien- tations and mean gridness are all within ± one standard 7 100 y 50 a b 100 50 y d 0 −50 −100 unit 10, 0.6. unit 74, 0.6. unit 139, 0.6. unit 203, 0.6. 100 50 100 50 100 50 50 x 100 50 x 100 50 x 100 50 x 100 g:1.8, o:7deg. g:1.7, o:9deg. g:1.8, o:7deg. g:1.8, o:8deg. 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx c unit 10 unit 74 unit 139 unit 203 0.2 0.1 0 −0.1 −0.2 0.2 0.1 0 −0.1 −0.2 0.2 0.1 0 −0.1 −0.2 0.2 0.1 0 −0.1 −0.2 −0.2 −0.1 0 0.1 0.2 −0.2 −0.1 0 0.1 0.2 −0.2 −0.1 0 0.1 0.2 −0.2 −0.1 0 0.1 0.2 unit 10 unit 74 unit 139 unit 203 0.6 0.4 0.2 0 d t i h g e w l a r e t a l l o C e t i h g e w l a r e t a l l o C 0.6 0.4 0.2 0 0.6 0.4 0.2 0 0.6 0.4 0.2 0 0 90 180 270 360 0 90 180 270 360 0 90 180 270 360 0 90 180 270 360 Preferred HD Preferred HD Preferred HD Preferred HD unit 10 unit 74 unit 139 unit 203 0.6 0.4 0.2 0 −1 −0.5 0 0.5 1 Spatial correlation 0.6 0.4 0.2 0 −1 −0.5 0 0.5 1 Spatial correlation 0.6 0.4 0.2 0 −1 −0.5 0 0.5 1 Spatial correlation 0.6 0.4 0.2 0 −1 −0.5 0 0.5 1 Spatial correlation Figure 4: Collateral connections align grid fields. (a) Spatial firing rate maps of example units in a cylinder environment. Unit number and maximal firing rate (in arbitrary units) are indicated above each rate map; (b) Autocorrelograms of the maps shown in a. Gridness score and grid orientation (in degrees) are indicated above each autocorrelogram; (c) Angular firing maps (blue solid lines) of the same units as in a with, in green, the tuning curve of the HD input to each unit shown. The red dash-dot lines indicate preferred HD of the units; (d) Collateral weight Wik as a function of θk. The broken line is θi; (e) Scatter plots between collateral weights and the spatial correlations of the fields between pre- and post-synaptic units. deviation error bar when averaged across simulations. B. Gradual grid development Although grid formation is a gradual process, grids are expressed at relatively early stages of development25,26. In adult rats, grids appear after a first exposure to a novel environment, but need several days of experience to become stable35. Consistent with these findings, con- junctive units in our network also show early expression of grids, within a gradual formation process. As shown in Fig. 5b, grids can be observed already 20 minutes af- ter the simulated rat has started to explore the envi- ronment. With longer exploration, the grid fields be- come both more triangular and more coherently aligned to a common orientation, as quantified by the increasing mean gridness and the decreasing mean standard devia- tions of grid orientations (Fig. 5c,d). Grids stabilize af- ter about 14 hours of continuous exploration. The mean spacing of the grids does not show big changes during development (Fig. 5e). Both the time scales of early grid appearance and of grid stabilization at the population level are comparable with experimental results, consider- ing the time spent by a real rat in rest and sleep. b Time 1000 s 0.3 Time 2000 s 0.4 Time 5000 s 0.5 Time 10000 s 0.5 Time 20000 s 0.5 Time 50000 s 0.5 8 100 50 1 1 t i n U y 0.4 y d 100 50 0 −50 −100 100 50 x g:0.6,o:50deg. 100 50 x g:0.8,o:38deg. 100 50 100 50 0 −50 −100 100 50 100 50 0 −50 −100 100 50 x g:0.7,o:43deg. 100 50 x g:0.6,o:44deg. 100 50 100 50 0 −50 −100 100 50 100 50 0 −50 −100 100 50 x g:0.1,o:48deg. 100 50 x g:1.6,o:59deg. 100 50 100 50 0 −50 −100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 dx 0.3 dx 0.4 dx 0.5 dx 0.5 dx 0.6 dx 0.5 0.1 s.d./mean speed 0.2 0.3 a s s e n d i r G 2 1 0 −1 −2 c s s e n d i r G 2 1 0 −1 0 −2 1250 20000 40000 60000 80000 20 15 10 5 0 1250 20000 40000 60000 80000 d ) s e e r g e d ( s n o i t a t n e i r o d i r g . d . s e 100 ) m c ( g n c a p S i 80 60 40 20 0 1250 20000 40000 Time (s) 60000 80000 100 50 9 1 t i n U y y d 100 50 0 −50 −100 100 50 x g:0.6,o:15deg. 100 50 x g:0.8,o:14deg. 100 50 100 50 0 −50 −100 100 50 100 50 0 −50 −100 100 50 x g:1.2,o:10deg. 100 50 100 50 0 −50 −100 100 50 x g:1.2,o:8deg. 100 50 100 50 0 −50 −100 100 50 x g:0.1,o:68deg. 100 50 x g:1.3,o:57deg. 100 50 100 50 0 −50 −100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 dx 0.2 dx 0.4 dx 0.5 dx 0.5 dx 0.5 dx 0.6 100 50 1 4 t i n U y y d 100 50 0 −50 −100 100 50 x g:0.1,o:36deg. 100 50 100 50 0 −50 −100 100 50 x g:0.4,o:1deg. 100 50 100 50 0 −50 −100 100 50 x g:0.7,o:72deg. 100 50 x g:1.0,o:65deg. 100 50 100 50 0 −50 −100 100 50 100 50 0 −50 −100 100 50 x g:1.3,o:48deg. 100 50 x g:1.6,o:55deg. 100 50 100 50 0 −50 −100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 dx dx dx dx dx dx Figure 5: The variation in speed does not influence grid formation. (a) Grids show similar mean gridness scores at the end of simulations performed with different speed standard deviation relative to mean speed. Error bars indicate ± standard deviation across 5 simulations; (b) Grid fields appear in the early phases of the simulation and stabilize with more experience in the environment. Grid fields and the corresponding autocorrelograms are ordered in rows for three example units from the same network. Maps formed after the same amount of exploration are arranged in the same column; (c) Mean gridness increases with respect to the amount of exploration (speed standard deviation is 40% relative to the mean speed). Gray area is the ± standard deviation; (d) Standard deviation of grid orientations; (e) Mean spacing of the grids with gridness larger than 0.25. IV. THE EFFECT OF THE SHAPE OF THE ENVIRONMENT As shown in the previous section, grid fields align mu- tually to each other in cylinder environments. As cylin- der has no preferred direction, and the common align- ment emerging in one simulation bears no relation with the one emerging in another, leading to the expectation that different animals would show differently aligned grid units, if these were to form prevalently in cylinder envi- ronments. Rodent cages are usually rectangular, how- ever, and a rectangle does have preferred directions. Is the orientation of grid units in different rats influenced by the shape of the training environment? The current sec- tion is devoted to this issue, namely, to the coherence of grid orientation across rats, i.e., in our case, across sim- ulations. In contrast to the standard simulations in the previous section, we simulate virtual rats in a 125 cm × 125 cm square environment, which leads to a (simulated) anisotropic exploration. We also vary the standard devi- ation σRD in running directions, which affects the degree of behavioral anisotropy. Rats, and especially naıve ones, have a natural ten- dency to run along the walls of the environment. There- fore their trajectories would reflect the anisotropy of the enclosing perimeter. Our virtual rat in a square box has a probability higher than chance of following the walls, simply because of the random walk mechanism described earlier. While in the cylinder environment the trajectory is necessarily isotropic, following the walls in a square box will result in the emergence of four preferred running di- rections (Fig. 6 left column). The RD distributions for square environments exhibit four peaks centered around the directions parallel to the walls (Fig. 6 right column). In contrast, in a cylinder environment, the RD distribu- tion is uniform (Fig. 3b). The non-uniformity of the RD Trajectory Running direction histogram 9 a 100 y 50 0 0 b 100 y 50 4 3 2 1 ) % ( e g a t n e c r e P 0 0 4 3 2 1 ) % ( e g a t n e c r e P 0 0 90 180 270 Running direction (degrees) Running direction histogram 90 180 270 Running direction (degrees) 360 360 50 x 100 Trajectory 0 0 50 x 100 Figure 6: Square environments induce more running along the sides. The left panels show parts of typical trajectories. The corresponding distributions of RDs in simulations are depicted in the right panels. (a) The default standard deviation in RD, σRD = 0.2 radians; (b) A simulation with σRD = 0.15 radians, demonstrating stronger anisotropy in the trajectories. distribution in square environments is stronger when the standard deviation in RDs is smaller (Fig. 6b). In an environment with anisotropic boundary, if one running direction is preferred systematically, the network would associate the activity of conjunctive units more strongly to places following the preferred RD, and would effectively orient one of the grid axes along this direction, forcing the other two grid axes to follow. We would ex- pect to see coherent grid orientation across rats in such conditions. As in cylinder environments, gird fields show coher- ent alignment in individual simulations in square en- vironments, irrespective of the RD standard deviation (Fig. 7a-c). To see any common orientation across sim- ulations, we performed multiple simulations in square and cylinder environments with RD standard deviation σRD = 0.2 radians and in a square environment also with σRD = 0.15 radians. In each of the three conditions, 70 independent simulations were conducted with different seeds for the random number generator. Averaged and normalized over the 70 trials, the sum of the mean ori- entation distributions of the three grid axes in square environments shows a significant concentration at mul- tiples of 30 degrees (Fig. 7d-e). That is, the common orientation of the grid fields is more likely to align along the walls of box environments. To quantify the coherence in orienting grids, a cross-trial grid orientation coherence score is defined for the one-dimensional mean orientation distribution of grid axes. The coherence score is com- puted in a similar way as the gridness score in Section III but assuming 30-degree periodicity, thus evaluating even and odd multiples of 15 degrees. Note that the range of the cross-trial grid orientation coherence score is [-2,2]. The cross-trial coherence of grid orientation in square environments is much larger than that in cylinder envi- ronments (Fig. 7d-e vs. f). Results in Fig. 7 indicate that although the running- direction anisotropy associated with training in square environments does not influence the coherence in grid alignment, it may strongly modulate the coherence in grid orientation across rats. That is, running-direction anisotropy produces a single orientation for the aligned grids in each simulation, but two clusters of prevailing common orientations across simulations, at 30 (or rather, 90) degrees of each other, i.e. with one grid axis aligned to a wall of the square environment. In the real system, two main orientations are observed in the same experiment (across grid units with differ- ent characteristic grid spacings27), not quite orthogonal to each other, but also not aligned to the walls of the testing environment. It could well be that these prevail- ing orientations reflect other environments, where rats were caged during development. A rat who develops in a square or rectangular cage, however, is expected to expe- rience not only RD anisotropy, as tested in the model, but also speed anisotropy. To disentangle a potential contri- bution of the latter, in the following section we introduce speed anisotropy, and compare its effects with those of running-direction anisotropy. Alignment of grid axes. std:2.157 deg a 200 150 100 50 s t i n u f o r e b m u N b 200 150 100 50 s t i n u f o r e b m u N Alignment of grid axes. std:2.049 deg c 150 100 50 s t i n u f o r e b m u N 10 Alignment of grid axes. std:2.967 deg 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 Grid orientation. coherence: 0.993 d 15 10 5 ) % ( e g a t n e c r e P e 15 ) % ( e g a t n e c r e P 10 5 Grid orientation. coherence: 0.750 Grid orientation. coherence: 0.237 f 15 10 5 ) % ( e g a t n e c r e P 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 Figure 7: Square environments tend to orient the aligned grid fields along the directions of the sides, compared with cylinder environments. (a-c) Grid fields align to each other in square environments (left column: σRD = 0.2 radians, middle column: σRD = 0.15 radians) and in a cylinder environment (right column: σRD = 0.2 radians) in three individual simulations. The alignment coherence score, i.e. the mean standard deviation (in degrees) of the orientations averaged over the three grid axes, is indicated at the top of each panel. The same data in Fig. 3e is shown in c again; (d-f) The sum of the orientation distributions of the three grid axes averaged over 70 simulations. In square environments, the orientation distribution of the three grid axes shows periodic clusters, which are distinct from the random fluctuations of the average orientation distribution. The number at the top of each panel indicates the coherence of grid orientation. V. SPEED ANISOTROPY Having observed that grid orientation is influenced by the non-uniformity of running directions, in this section we focus on another factor in the trajectories of the simu- lated rat, namely running speed, and investigate whether the shape of the grids can be influenced by an anisotropic speed distribution in exploration behavior, even with no anisotropy in the boundary conditions (i.e. in cylinder environments). In this section, the simulations are iden- tical to those in Section III, except that the virtual rat explores the cylinder environment running faster in four preferred directions, as visualized in Fig. 8 vt = vs[q + (1− q) sin(ωt)3 + cos(ωt)3 − 1/√2 1 − 1/√2 ]. (11) Here ωt is the current running direction of the rat. The speed now can go from a minimum of vs = 24cm/s to a maximum of vs = 40cm/s depending on the running direction. q = 0.6 is the ratio between the two speed extremes. As in the standard simulations in Section III, the RD distribution of the trajectory is quite uniform (Fig.9a,b). In the RD distribution there is a tiny over representation of running directions parallel to the diagonals. This tiny distortion is originated by the fact that when the rat is 0.4 0.2 0 −0.2 −0.4 −0.4 −0.2 0 0.2 0.4 Figure 8: Polar plot of the speed of the simulated rat with re- spect to running directions. The maximal speed is the default speed 0.4m/s, and the minimal speed is 60% of vs. running faster it takes a shorter time for it to reach the boundary, where it is forced to turn. However, the effect appears much smaller than that due to the square shape of the environment in Section IV, and its direct influence on grid alignment is negligible. Fig. 9j shows the spatial maps as well as the auto- correlograms of four example conjunctive units from the network. Similar to the case with constant speed, grids align with each other, and the phases are kept broadly distributed in the environment (Fig. 9d-f). However, the average gridness score is somewhat lower as compared to the case with constant speed (Fig. 9c). The reason is that, with speed anisotropy, the grids are distorted, showing grid axes with non-equal lengths, as can been seen carefully in Fig. 9j. Among the three axes of a grid, the axis that passes through the maximum farthest from the origin is identified as the long grid axis. In our simu- lations with speed anisotropy, conjunctive units tend to develop maps that share the same grid axis as the long one (in the simulation of Fig. 9e this corresponds to the axis with orientation around 100 degrees), indicating that grid distortion happens predominantly along one direc- tion. Another way to quantify this distortion, introduced in27, is to fit an ellipse that passes the six maxima close to the center of the grid autocorrelogram (white curves in Fig. 9j). These ellipses deviate from perfect circles, indicating modified grid structures. We find that such ellipses are roughly aligned in each map with the corre- sponding long grid axis, laying within 30 degrees of it (Fig. 9g-h). The flattening of an ellipse is measured by ellipticity, i.e. the ratio between the major and minor axes. Note that the ellipticity of a perfect circle is 1. In the simulation, the median of the ellipticity distribution of the grids is about 1.15 (Fig. 9i). Averaged over 70 independent simulations in cylin- der environments, the orientation of the long grid axis concentrates along the preferred directions of the speed profile, i.e. 0 and 90 degrees (gray bars in Fig. 10a). The other two grid axes only broadly orient to 60/120 or 30/150 degrees respectively (light blue bars in Fig. 10a), with low coherence of grid orientation. In contrast, with constant speed in cylinder environments, both the aver- age orientation distribution of grid axes and of the long grid axis are fairly uniform, reflected in even lower coher- ence score of grid orientation (Fig. 10d). With speed anisotropy in cylinder environments, we observe that grid maps are more elliptical than those de- veloped from exploration with constant speed (Fig. 10c vs. Fig. 10f). The difference is rather subtle, however, as the fluctuations in the length of the grid axes, and our choosing always the major axis of the best fitting ellipse, whichever it is, obviously produce an ellipticity measure distributed above 1, even with constant speed. The effect of anisotropy is more salient in orienting the ellipses, Fig. 10b, around the preferred directions of the speed profile, a clear difference from the uniformly dis- tributed ellipse orientation of simulations with isotropic speed, Fig. 10e. The difference in orienting ellipses can be quantitatively measured by defining cross-trial ellipse orientation coherence score for the one-dimensional av- erage distribution of ellipse orientations. The cross-trial ellipse orientation coherence score is calculated the same way as the cross-trial grid orientation coherence score, except that 90-degree periodicity is assumed, instead of 30-degree periodicity. With speed anisotropy in cylinder environments, the cross-trial coherence in ellipse orienta- tion is as high as 1.533. However, with constant speed, still in cylinder environments, the corresponding coher- 11 ence score is close to zero, indicating a uniform ellipse orientation distribution. Speed anisotropy orients ellipses rather loosely, with a width at half-peak around 30 de- grees. It is reflected also in which of the three grid axes tends to be the longest, but without apparently forcing a rigid orientation of the grids themselves, Fig. 10a. Fig. 11 shows a summary of the differences between speed anisotropy and RD anisotropy, once coupled with RC interactions, with respect to three measures, namely, cross-trial coherence in grid orientation, cross-trial coher- ence in ellipse orientation, and mean ellipticity. One can see that both anisotropies slightly distort the shape of grid maps into a somewhat more pronounced elliptical ar- rangement of the firing fields. Speed anisotropy induces on average more distortion (with our parameters), but the average effect is overshadowed by the large variabil- ity from unit to unit, in what is an asymmetric ellipticity distribution with a long tail. The qualitative difference is in what common orientation emerges: speed anisotropy loosely orients ellipses toward the fast directions (larger cross-trial coherence in ellipse orientation), but it does not enforce a tight grid orientation, across simulations, on the aligned grids (dark circle in Fig. 11a-b). In con- trast, RD anisotropy orients grids toward preferred RDs (larger cross-trial coherence in grid orientation, white and gray squares in Fig. 11a-b), but not ellipses. When com- bined with RD anisotropy, speed anisotropy prevails (the dark square in Fig. 11a-b). This is likely because speed aniso-tropy induces the long grid axes to lie along either 0 or 90 degrees, and it constraints the lengths (shorter) and orientation (not at 60 degrees) of the other two grid axes; with RD anisotropy, the arrangement of the fields maybe more dependent on the details of the trajectory taken in each simulation, which does not appear to re- strict much the orientation of the ellipses, seemingly leav- ing the grids free to orient themselves. We may conclude that the shape of the environment and speed anisotropy cause apparently similar, but subtly distinct effects. A. Grid alignment without collaterals In the simulations above, collateral interactions among would-be conjunctive units tend to align their develop- ing grid fields along common axes. We have considered two additional factors that influence the alignment, the shape of the environment, if different from cylindrical, and anisotropy in running speed. The latter factor, in particular, enhances the ellipticity of the resulting grids, and affects their ellipse orientation. Its effects are how- ever secondary, in the simulations, to the primary effect produced by the collaterals. It is important to note that ellipticity and some degree of common orientation, how- ever, can be produced also by speed anisotropy on its own, in the absence of collaterals. This can be understood by considering a simple abstract model, which extends the one earlier considered in23 to account for the devel- opment of grid fields at the single unit level. The model Running direction histogram 90 180 270 Running direction (degrees) c s t i n u f o r e b m u N 30 25 20 15 10 5 0 −1 360 12 Gridness score histogram. m:1.327 −0.5 0 0.5 Gridness 1 1.5 2 Phase histogram 4 Trajectory a 100 y 50 0 0 50 x 100 b ) % ( e g a t n e c r e P 4 3 2 1 0 0 Alignment of grid axes, std:2.909deg; m spacing:51.264cm d 125 75 25 y d −25 −125 −75 −25 25 75 125 dx Histogram of ellipse orientations 50 40 30 20 10 g s t i n u f o r e b m u N j Alignment of grid axes. std:2.909 deg e 150 100 50 s t i n u f o r e b m u N f 40 20 y d 0 −20 −40 0 0 30 120 60 Orientation (degrees) 90 150 180 −40 −20 0 dx 20 40 Histogram of the angles between grid and ellipse axes h 50 40 30 20 10 s t i n u f o r e b m u N i s t i n u f o r e b m u N 100 80 60 40 20 0 1 120 180 Histogram of ellipticity 1.2 1.8 Ellipticity (long/short ellipse axis) 1.4 1.6 2 0 2 0 0 30 60 120 Orientation (degrees) 90 150 180 0 −180 −120 −60 0 60 Angle (degrees) unit 11, 0.6. unit 75, 0.6. unit 139, 0.6. unit 203, 0.6. g:0.9, o:47,e:1.21 g:1.9, o:43,e:1.02 g:1.4, o:47,e:1.12 g:1.5, o:47,e:1.11 100 y 50 100 50 100 50 100 50 100 50 y d 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 50 x 100 50 x 100 50 x 100 50 x 100 −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx Figure 9: Grid alignment during exploration in a cylinder environment with anisotropic speed. (a) A trajectory of the rat with 2 × 104 steps in a cylinder environment; (b) RD distribution of an entire trajectory in a 8 × 106-step simulation; (c) Histogram of the gridness scores of all conjunctive units in the simulation; (d) The scatter plot of the locations of the three peaks found in autocorrelograms; (e) Histogram of the orientations of the three grid axes. Shown in front in gray is the orientation histogram of the long grid axes. Indicated at the top of the panel is the coherence score in grid alignment (mean standard deviation, in degrees, of the orientation averaged over the three grid axes), similar as the coherence in grid alignment in the cylinder environment with constant speed; (f) The histogram of spatial phases (again, relative to that of the best grid); (g) Histogram of the orientations of the major axes of the ellipses determined from each autocorrelogram; (h) Histogram of the angles between the long grid axis and the ellipse major axis; (i) Histogram of the ratios between the major and minor axes of ellipses. The white broken line indicates the median; (j) Examples of the fields of conjunctive units (left) and the corresponding autocorrelograms (right). The white curves show the ellipses determined from the three maxima found in autocorrelograms. Unit number, and maximal firing rate (in arbitrary units) are indicated above each rate map. Gridness score,orientation (in degrees) and ellipticity are indicated above each autocorrelogram. describes a single unit in a very large environment, which then for all practical purposes has no shape. Hence the abstract model focuses on the effect of speed anisotropy, which is modeled similarly as in the simulations, extri- cating it from the other two factors, the presence of col- laterals, and the shape of the environment. The model, described in the Appendix, leads to a relation between the degree of anisotropy and the degree of ellipticity, and Grid orientation. coherence: 0.267 a 15 ) % ( e g a t n e c r e P 10 5 5 4 3 2 1 ) % ( t e g a n e c r e P 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 e Ellipse orientation. coherence: −0.032 Grid orientation. coherence: 0.237 d 15 ) % ( e g a t n e c r e P 10 5 5 4 3 2 1 ) % ( t e g a n e c r e P 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 b Ellipse orientation. coherence: 1.533 c Ellipticity histogram 13 ) % ( t e g a n e c r e P ) % ( t e g a n e c r e P 14 12 10 8 6 4 2 0 14 12 10 8 6 4 2 0 1 1.2 1.4 Ellipticity 1.6 1.8 2 f Ellipticity histogram 1 1.2 1.4 Ellipticity 1.6 1.8 2 Figure 10: Speed anisotropy in cylinder environments distorts the shape of the grids (top row) as compared to exploration with isotropic speed (bottom). (a,d) Average orientation distribution of long grid axes (gray) plotted in front of the average orientation distribution of the three grid axes (blue). The coherence of grid orientation is indicated at the top of each panel. The blue bars in d are the same measure as the distribution shown in Fig. 7f; (b,e) Average orientation distribution of the major axes of the ellipses. At the top of each panel, the coherence in ellipse orientation is noted. It is calculated similarly as the coherence in grid orientation, but assuming 90-degree periodicity instead of 30-degree; (c,f) Average distribution of ellipticity. The distribution in c is plotted again as the red empty bars in f. a 1.5 n o i t a t n e i r o d i r g n i e c n e r e h o C 1 0.5 0 −0.5 1.05 1.1 1.15 1.2 Ellipticity b 2 1.5 1 0.5 0 −0.5 −1 1.05 1.1 n o i t a t n e i r o e s p i l l e n i e c n e r e h o C 1.25 1.3 RD=0.2 Cylinder, σ Cylinder, SA, σ Square, SA, σ Square, σ Square, σ RD=0.2 RD=0.15 RD=0.2 RD=0.2 1.25 1.3 1.15 1.2 Ellipticity d Ellipse orientation. coherence: −0.339 Grid orientation. coherence: 0.993 c 15 ) % ( e g a t n e c r e P 10 5 5 4 3 2 1 ) % ( e g a t n e c r e P 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 Figure 11: Speed anisotropy and RD anisotropy have different effects on grid shape and orientation. (a,b) Five situations are compared with respect to coherence in grid/ellipse orientation and to ellipticity, based on the average over 70 independent simulations. The horizontal lines indicate ± one standard deviation in ellipticity. The two panels share the same legend to the right; (c) The average orientation distribution of the long grid axes (gray, in front) and the average orientation distribution of all three grid axes (blue, in the back) for the simulations in square environments without speed anisotropy, the same data as shown also in Fig. 7d. The coherence in grid orientation is noted at the top of the panel. (d) The average distribution of ellipse orientation, also for the simulations in square environments without speed anisotropy, with the corresponding coherence at the top. it indicates two orientations, at roughly 90 degree of each other, of the grids and of the ellipses that best match the assumed quadrupole anisotropy. Hence it predicts that grids and ellipses would have one of two common orien- tations even in the absence of collateral interactions. The effect predicted by the analytical model is repro- duced in ad hoc simulations. Fig.12 shows that in the presence of speed anisotropy only, without collateral in- teractions, grid units develop with an enhanced elliptic- ity, and with one of the grid axes preferentially aligned along one of the two orthogonal directions with higher mean speed (in the simulation of Fig. 12, this happens to be the one at 90 degrees). The variance in ellipse orien- tation is considerable, larger than what is observed with speed anisotropy and collateral interactions (Fig. 12b vs. Fig. 9g). However, the degree of ellipticity is sim- ilar, irrespective of the existence of collateral interac- tions (Fig. 12c and Fig. 9i). The first element that is missing, without collateral interactions, is crucially the tight alignment of the grids with each other. This is because, in our model, collateral interactions align the grids with each other through mutual iterative conver- gence, whereas without collaterals the alignment only re- flects single unit adaptation to what is effectively a broad shallow valley in a free energy landscape. We present the abstract model in the Appendix for clarity, but we believe it to be unlikely that, in the absence of collat- eral interactions, speed anisotropy alone can establish a common grid orientation with the tight alignment seen in experimental data. The second element that is miss- ing without collateral interactions is the invariance in the relative phase of grid maps across multiple environments, which we will examine in the next section. VI. GRID REALIGNMENT IN MULTIPLE ENVIRONMENTS Until now, we have only considered grid alignment in a single environment. Under dramatic environmen- tal changes, both the hippocampal and entorhinal neu- rons develop new maps, in a process called global remap- ping28. While hippocampal place fields seem to shuffle randomly during global remapping, grid cells behave in a population-coherent manner. The new maps preserve the same relative phases as the old maps, and maintain a common grid orientation, not necessarily the same one as before29. In this section, we address the question of how grid fields may possibly align in multiple environments. We simulate the network in two different cylinder en- vironments, with identical conditions as in Section III, except for the number of units. The total number of conjunctive units is 500. The total number of place units is 900. In each environment, 500 of the place units are active. The number of place units that are active in both environments is 100, i.e. 20% of the active place units in one environment. The place fields of the place units in the first environment are completely different 14 from those in the second environment. However the pre- ferred HDs of the conjunctive units are the same across environments. The training is interleaved in the two en- vironments, with 2000 epochs in each environment, 3000 steps in each epoch. This leads to 1.2×107 training steps in total. Fig. 13a,f compares the spatial fields and correspond- ing autocorrelograms of four example conjunctive units in two environments. In both environments, the grids are aligned, but to different orientations (Fig. 13b,g). The angular offset of the grids in the second environment rel- ative to the grids in the first environment is about 9 de- grees counterclockwise. The spatial phases of the units relative to the best grid in each environment do not show any clear pattern or clustering (Fig. 13c,h), as in the sim- ulations with a single learned environment. The distribution of ellipse orientations is quite differ- ent across environments (Fig. 13d,i), but the ellipticity distribution is comparable, with a similar median at 1.16 (Fig. 13e,j). This value is similar to that of the simula- tions with speed anisotropy. This is because 20% of the place units are active in both environments but in ran- domly different positions. Since conjunctive cells need to maintain population coherence, this randomness acts as a source of noise, imposing additional constrains on con- junctive units that move their maps away from perfect gridness. This can be seen as a third source of ellipticity. The other two sources discussed in previous sections, i.e. speed anisotropy and RD anisotropy, introduce elliptic- ity during grid development by breaking the symmetry of the trajectories of the simulated rat. The ellipticity caused by grid realignment in multiple environments is due to the overlap of the population codes of contextual information, and is imposed by the structure of the net- work. We then determine whether the relative spatial phases of the units are kept invariant across environments. For this we select a sub-population, by taking into account only units that have good grid maps in both environ- ments (gridness score > 0.75). The firing map in environ- ment B of each unit in the population is rotated counter- clockwise at multiples of 4.5 degrees and cross-correlated with the corresponding map in environment A, as in29. The cross-environment crosscorrelograms thus obtained are averaged over the population, to get the mean cross- environment crosscorrelograms for this sub-population. The mean crosscorrelogram of two groups of grid maps shows whether the grid maps are related by the same shift transformation. When two grids have the same orientation and spacing, the crosscorrelogram between them shows grid structure. In addition, the relative spatial shift between them determines the direction and distance that the corresponding crosscorrelogram is dis- placed away from the center. Cross-correlating one group of grids, which have the same orientation and spacing but distributed spatial phases, with a second group of grids, which are a shifted version of the first group for the same amount, will result in a set of crosscorrelograms with grid a 40 s t i n u f o r e b m u N 30 20 10 Alignment of grid axes. std:15.869 deg Histogram of ellipse orientations b 30 20 10 s t i n u f o r e b m u N 0 0 30 60 120 Orientation (degrees) 90 150 180 0 0 30 60 120 Orientation (degrees) 90 150 180 c Histogram of ellipticity s t i n u f o r e b m u N 70 60 50 40 30 20 10 0 1 1.2 1.8 Ellipticity (long/short ellipse axis) 1.4 1.6 15 2 Figure 12: Grids show a weak tendency to align due to speed anisotropy alone, without collateral connections in the network (ρ = 0). (a) Histogram of the orientations of the grid axes (blue, in the back) together with the histogram of the long grid axes (light gray, in the front). The coherence in grid alignment is low (the mean standard deviation, in degrees, of the orientation averaged over the three grid axes is shown at the top of the panel, and it is much higher than with collaterals, shown in Fig. 9e); (b) Orientation histogram of the major axes of the ellipses; (c) Histogram of ellipticity. pattern and identical offset away from the center. The offset is just the common shift between the two groups. Adding the crosscorrelograms between the grids in the two groups preserves a clear grid structure. If a com- mon rotation of multiples of 60 degrees is applied, instead of a common shift, the second group of grids still have the same orientation as the original grids, and each indi- vidual crosscorrelogram still keeps a grid structure, but the phases are changed differently for each grid (Fig.15), producing crosscorrelograms with peaks appearing in dis- tributed locations. Thus, the mean crosscorrelogram of the two groups does not maintain a grid pattern. For a common rotation other than multiples of 60 degrees, the rotated grids do not align to the same orientation as the original grids any more, making it impossible for the mean crosscorrelogram to be a grid. Therefore, for two groups of grids, the mean crosscorrelogram appears to be a grid only when there is a coherent shift between the two groups. Fig.14a shows that only with a counterclockwise rota- tion close to 9 degrees does the mean cross-environment crosscorrelogram show a grid structure. This is the an- gle that aligns the grid axes of both environments in Fig.13b,g. Fig. 14a thus indicates that after the angu- lar offset is counter-balanced, the spatial shifts between the grid fields in the two environments are the same (Fig. 14b). For other rotations, in contrast, the spatial shifts are different across units, giving rise to an increas- ingly flat mean cross-environment crosscorrelogram. The absolute angular offset between the grid orientations in two environments is not important. Note that our sim- ulations have fixed HD selectivity across environments, whereas in rats the preferred HDs of both conjunctive cells and HD cells appear to rotate coherently by the same amount in different environments6. Therefore the angular offset between the two sets of grid maps in our simulations can be considered as the angular shift af- ter correcting for the HD selectivity shift across environ- ments. Fig.14c shows the "gridness" scores of the mean cross-environment crosscorrelogram with respect to the amount of rotation when a different number of units from the network is included in the analysis. The "gridness" score defined here is similar to the one in the method described in Section III. The difference is that the ring circling the six maxima is centered on the closest maxima from the center of the crosscorrelogram, and that the dif- ference in covariance instead of the correlation coefficient is used to define gridness. This is because most of the mean cross-environment crosscorrelograms are quite flat, giving unstable correlation coefficients when normalizing with small variance. As can be seen in Fig.14c, with 20 units as a population, it is already possible to infer the angular mismatch between the fields in two environ- ments. When only one unit is considered, the gridness of the cross-environment crosscorrelogram is roughly pe- riodic with period 60 degrees. This indicates that with the population code expressed by grid cells it is possible to keep a unique orientation correspondence across en- vironments in spite of the orientation symmetry of the grids, and to differentiate locations in an environment by the relative spatial phases of the grid fields. To repre- sent places on a larger scale, multiple grid spacings are needed37–40. VII. DISCUSSION Several conclusions can be extracted from these simu- lations. In the first place, we have described how align- ment can be achieved in the model through recurrent collaterals, as long as they are associated with HD infor- mation. Secondly, we have shown that running-direction anisotropy in behavior, induced by the shape of the envi- ronment, is enough to distort the grid maps into ellipses, but it has the side-effect, of orientating one of the grid axes with one of the walls. At the same time, whether because too weak or localized, this kind of anisotropy fails to align the distortion, measured by the long grid axis or ellipse direction, into a common orientation. We c 40 20 y d 0 −20 −40 Phase histogram −40 −20 0 dx 20 40 Histogram of ellipse orientations d 30 s t i n u f o r e b m u N 20 10 0 0 30 60 120 Orientation (degrees) 90 150 180 6 4 2 0 Histogram of ellipticity 1.2 1.8 Ellipticity (long/short ellipse axis) 1.4 1.6 2 e 1 s t i n u f o r e b m u N 80 60 40 20 0 g Alignment of grid axes, std:4.762deg; m spacing:64.688cm 16 b Alignment of grid axes, std:5.136deg; m spacing:62.075cm a 100 y 50 100 50 y d 0 −50 −100 unit 15, 0.5. unit 140, 0.6. unit 290, 0.6. unit 413, 0.6. 100 50 100 50 100 50 50 x 100 50 x 100 50 x 100 50 x 100 g:1.6, o:21,e:1.05 g:1.2, o:24,e:1.22 g:1.0, o:21,e:1.14 g:1.3, o:34,e:1.19 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 125 75 25 y d −25 −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −125 −75 −25 25 75 125 dx f 100 y 50 100 50 y d 0 −50 −100 unit 15, 0.7. unit 140, 0.5. unit 290, 0.6. unit 413, 0.6. 100 50 100 50 100 50 50 x 100 50 x 100 50 x 100 50 x 100 g:1.0, o:8,e:1.22 g:1.1, o:15,e:1.32 g:1.2, o:61,e:1.11 g:1.4, o:14,e:1.16 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 125 75 25 y d −25 −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −125 −75 −25 25 75 125 dx h 40 20 y d 0 −20 −40 Phase histogram i Histogram of ellipse orientations 6 4 2 0 −40 −20 0 dx 20 40 30 20 10 s t i n u f o r e b m u N 0 0 30 60 120 Orientation (degrees) 90 150 180 j 1 s t i n u f o r e b m u N 80 60 40 20 0 Histogram of ellipticity 1.2 1.8 Ellipticity (long/short ellipse axis) 1.4 1.6 2 Figure 13: Grid realignment in environment A (top) and environment B (bottom). (a,f) Examples of the spatial fields and autocorrelograms of four different units in the two environments; The unit number and maximal firing rate (in arbitrary units) are indicated above each rate map. The gridness score, grid orientation (in degrees, between the x-axis and the nearest grid axis) and ellipticity are indicated above each autocorrelogram; (b,g) Scatter plots of the three peaks found in autocorrelograms. The angular offset of the grids in environment B relative to the grids in environment A is 9 degrees clockwise. Noted at the top of the panel are the mean spacing and coherence score in grid alignment (mean standard deviation, in degrees, of the orientations averaged over the three grid axes); (c,h) Two dimensional histograms of spatial phases relative to the best grid in each environment. Each spatial bin represents a 2.5cm × 2.5cm area; white indicates zero units with a phase in that spatial bin, with more units indicated by progressively darker gray; (d,i) Histograms of the ellipse orientations; (e,j) Histograms of the ellipticity, i.e. the ratio between the length of the major and minor axes of an ellipse, across all units. 17 a 100 50 y d 0 −50 −100 rotation: 0 rotation: 4.5 rotation: 9 rotation: 18 rotation: 27 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 rotation: 36 rotation: 45 rotation: 54 rotation: 63 rotation: 72 100 50 y d 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 −100 −50 0 50 100 rotation: 81 rotation: 90 rotation: 99 rotation: 108 rotation: 117 100 50 y d 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100 −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx −100 −50 0 50 100 dx b 100 50 y d 0 −50 −100 −100 −50 0 dx 50 100 #units= 1 #units= 10 #units= 20 #units= 50 #units= 174 c " s s e n d i r G " 0.3 0.25 0.2 0.15 0.1 0.05 0 −0.05 0 60 120 180 Rotation (degrees) 240 300 355.5 Figure 14: The phases of the grids in the two environments rotate and shift together. (a) The mean cross-environment crosscorrelogram of the population after the optimal counterclockwise rotation of 9 degrees is shown, together with some other rotations. The rotation angle is noted on the top of each panel. The mean cross-environment crosscorrelogram shows the best grid structure in the case of the specific rotations that (roughly) match the angular offset between the grid axes in the two environments. The Mann-Whitney test36 shows that the mean cross-environment crosscorrelogram after 9-degree rotation has larger mean than the crosscorrelogram after zero-degree rotation (p=0.0015); (b) After the optimal rotation, the phase shifts of the same units concentrate in a small region with size similar to the field of a grid. Note that the phase shifts of a small number of units fall in a different area, due to the periodicity of the grid and to small fluctuations in the peaks of the individual crosscorrelograms. The gray crosscorrelogram shown in the background is the mean cross-environment crosscorrelogram after the optimal rotation, as shown in a; (c) The "gridness" of the mean cross-environment crosscorrelogram when the population set is increased from a single grid (# units 1) to the whole selected population. The gray broken line indicates the optimal rotation leading to the highest gridness of the cross-environment crosscorrelogram. e2 e1 (0, 0) e3 Figure 15: Rotating a grid map by 60 degrees results in a non-linear spatial phase transform. The spatial periodicity of a grid can be decomposed into the periodicity along the projection axes ei, defined perpendicularly to the three grid axes. The orientation of and the period/wavelength along the third projection axis is constrained by the other two projec- tion axes. The spatial phases of grids with the same spacing and orientation can be uniquely represented by points in a rhombus (area in gray). The sides of the rhombus are paral- lel to two of the grid axes, and the length of the side is equal to grid spacing. The spatial phases are only distinguishable in a rhombus area, because the rhombus is exactly one period length if projected onto two of the projection axes, e1 and e2. Points outside the rhombus are equivalent to those in the rhombus, modulo the wavelength along each projection axis. A grid with non-zero phase (shown in brown) is rotated by 60 degrees (indicated by the dotted arc) around the origin. The phase of the rotated grid (in gray) is equivalent to the white dot in the gray rhombus, since the coordinates (dashed arrows) of the two points along the projection axes are the same, modulo the wavelength. The phase change caused by the rotation depends on the initial phase of the grid. have then explored the effect of running speed anisotropy, which has a milder influence on grid orientation but a stronger one on the orientation of the ellipses. In this scenario, recurrent collaterals have the effect of amplify- ing through population coherence what could be seen as a distortion of the idealized grid map already at a sin- gle cell level. This, however, may be regarded as a side effect of the function of recurrent collaterals in aligning the grid maps. Such a function may be useful, we hy- pothesize, in order to achieve global remapping with an invariant structure of relative phases, which, as we show in the last part of the paper, our model reproduces when the virtual rat explores two different rooms. Whether grid alignment results from network interac- tions or as a consequence of a single unit process is a very interesting and complex question. In this respect, the literature presents opposing views. In models of interference between multiple oscillators, the grid maps are the result of information processing within a single neuron16–19. Each grid cell is then inde- pendent, and the common alignment comes from the fact that they all share the same path integration inputs. The different spatial phases are achieved by anchoring each grid cell to a different set of spatial inputs. This selective anchoring is in a way analogous to what happens with 18 our place inputs, but it plays in the oscillator-interference models the minor role of correcting the errors of path in- tegration, which is the main source of information used to construct the grid maps. A completely opposite view to the single unit perspec- tive comes from the attractor models8–12. In these, grid maps do not emerge individually, but as an emergent property of the population, so that grid maps and grid alignment are intrinsically inseparable. In our model, as shown by Kropff and Treves (2008), grid cells with no recurrent collaterals can develop grid maps with in- dependent random orientations if they are left to inter- act weakly through inhibition, a situation that reminds us of the anatomy of layer II of medial entorhinal cor- tex. The introduction of collateral connections with the necessary head direction information to disambiguate fir- ing sequences, as suggested earlier and proved here, can achieve grid alignment, as observed in experiments, thus offering a possible explanation of why these elements are found together in layers III to VI and are both absent in layer II. In our model, the common grid orientation lies on the attractor manifold created by the collateral interactions. This attracting property is a feature shared by the attractor models. In the latter, however, the col- lateral interactions have the full job of creating both grid maps and grid alignment. The alignment of grid cells in layer II is perhaps inher- ited from the deeper layers, which send strong projections to superficial layers41. This hypothesis, not developed in the present work, represents an interesting direction of modeling research, which could focus on how conjunc- tive cells and grid cells may develop together in different layers, and perhaps even self-organize into laminar struc- tures23,42. From the point of view of our model, the best way to address the question of a single unit versus a population origin of grid alignment is to interfere with the head di- rection source of information, assuming with some plau- sibility that this signal is not directly generated in the entorhinal cortex. Without head direction information, sequences in the grid cell network would be ambiguous. It would be interesting to study experimentally the effects that this kind of manipulation would produce in layers III to VI, where we assume that collateral connections work hand by hand with head direction information. The functional difference between feed-forward connec- tions and collateral connections can be appreciated in global remapping. In each environment, only a subset of place units is active. Feed-forward connections there- fore relay environment-specific inputs to the conjunctive units. The collateral connections, however, signal the in- teractions between conjunctive units by the environment- independent geometrical structures they encode. With the same feed-forward and collateral connections, the network is able to provide an efficient metric encoding of space for all environments, since the emerging grid maps are essentially the same set of grids, almost identical af- ter a unique rotation and translation. A natural ques- b b b b b b b b b b b b b b c tion to ask is how the collateral weights can be learned. Addressing this important issue will be very instructive for the attractor models as well, and therefore requires a separate article to be elaborated in depth. As we show in the present work, a weaker form of alignment of grid cells can be produced with indepen- dent single units, if the behavior of the rat includes some form of anisotropy. This alignment, however, would im- ply no population coherence in relative phases with global remapping. Still, anisotropy might be the reason why grid maps present distortion or ellipticity. This is not the first time that behavior is proposed to influence the firing of grid cells. In the hairpin maze experiment43, rats covered a two dimensional space but walking through a series of corridors in such a way that the behavioral con- straints were those typical of exploration in one dimen- sion. As a result, the grid cells showed a fragmented map with several one-dimensional patches instead of the canonical triangular structure. In our model, behavior alone creates distortion in the symmetry of grid maps, which is amplified and made coherent by recurrent col- laterals. In a way, the distortion observed by Stensland and colleagues (2010), coherent at the population level, might be the price that the system has to pay in order to get a metric system of space that is consistent across environments. It is poorly understood how much the geometry of a map in a new environment depends on the previous ex- perience of the rat. It has been shown that in some over- trained animals grid maps emerge very fast in a novel environment5, but we are far from being able to gener- alize this observation to all rats, experienced and naıve 35. It is possible that the perfect grid pattern observed in some rats is a result of the successful generalization of an isotropic distribution of trajectories in 2 dimensions, related to a vast experience in open field environments at the right age of development, possibly starting around P20. Such extensive exposure to the open field is far from being an exclusive natural condition for all rats. It would be interesting to study the adult grid maps of rats raised in a different environment, for example in a system of tunnels. If there is a critical time window when the geometry of the grid system is developed, these animals might have, as adults, grid cells with peculiar properties even after prolonged training in the usual 2 dimensional square boxes. Another interesting aspect of the problem is what happens during this training period. Perhaps ex- tensive learning can compensate for part of the distortion observed in grid maps. The main problem of such a study is that untrained animals exhibit a very poor coverage of 2 dimensional environments, making it hard to compare the geometry of their spatial maps with trained controls. The network model produces qualitatively similar re- sults over a large portion of parameter space. The delay parameter τ is critical to avoid the collapse of the sub- population with similar HD into a regime of synchronous dynamics. One wonders whether this delay in synaptic transmission might be realized by NMDA synapses, or 19 λ1 e1 e2 λ2 λ3 e3 Figure 16: Non-evenly separated grid axes, i.e. not at π/3 of each other, imply non-even lengths of the axes themselves, to preserve periodicity, resulting in an elliptical arrangement of the vertices. In this example, the grid axes (indicated by the broken lines) are oriented towards at 90, 45 and 157.5 degrees. Orthogonal to them are the three projection axes entering the 2D Fourier transform, at angles ei of 0, 135, and 247.5 degrees, respectively. The vertices of the grid are arranged on lines at 3 different distances λi from each other (and from the origin). other complex form of coupling among conjunctive cells. It is possible to extend the network model to incorpo- rate path integration cues into the inputs. Grids in such a network would still be stable even with week spatial inputs, and can serve as spatial codes useful in robot ex- ploration tasks, where the robot has to rely on its own location estimation44–47. Acknowledgements This work is supported by the European 7th Frame- work Program SpaceBrain and the Norwegian NOTUR project. We are grateful for enlightening discussions to all colleagues in the Kavli Institute and in the Spacebrain EU collaboration. Appendix A: Abstract ellipse model What does a non-circular, but elliptical arrangement of the grid map imply, geometrically? If any two of the three grid axes are not separated by 60 degrees, the lengths of the grid axes cannot be equal any more, for the grid to retain its canonical periodicity. As Fig. 16 illustrates, the grid vertices occur periodically along lines (broken in the Figure; those passing through the origin are called grid axes) orthogonal to the three projection axes (the black axes ei in Fig. 16), provided such lines share 3-way intersections. If the grid is mod- eled as a simple sum of 3 cosines, the distance between b c b c b c the broken lines of each orientation is the wavelength λi ≡ 2π/ki of the corresponding cosine, which has a ~k- vector of length ki oriented at ei from the x-axis. To preserve grid structure, the inverse wavelength k3 along the third projection axis e3 is constrained so that the peaks (red broken lines in Fig. 16) coincide with those along each of the other two projection axes (blue and green broken lines). This leads to two equations relating λ3 to λ1 and λ2 and to the angles between them. In the Figure, it is assumed that e1 = 0, but it is easy to write the following equations considering the general case of nonzero e1 λ3 = λ3 = λ1 cos(e2 − π/2 − e1) λ2 cos(e2 − π/2 − e1) cos(e3 − e2 − π/2), cos(e3 − e1 + π/2). They lead to a solution for e3 and λ3 which may be ex- pressed for example as e3 = arccos" λ3 = 1 + λ2 pλ2 2 + 2λ1λ2 cos(e2 − e1)# + π + e1, λ1 cos(e2 − e1) + λ2 1 + λ2 λ1λ2 pλ2 2 + 2λ1λ2 cos(e2 − e1) . It is straightforward to see that these relations translate into an equivalent, but simpler relation between the ~k- vectors ~k1 + ~k2 + ~k3 = 0. (A1) What determines the exact position of the ~k-vectors? Considering the abstract model discussed in detail in23, of the development of an individual grid, the model does not distinguish between a would-be pure grid and a con- junctive unit. It assumes that at the end of the develop- mental process the firing map Ψi(~x) of the unit minimizes a functional L, which in its basic version takes the form d~x′Ψ(~x′)K(~x′−~x), L = 1 AZA d~x[∇Ψ(~x)]2+ γ AZA d~xΨ(~x)ZA (A2) where K(∆~x) is a kernel expressing neuronal fatigue. The first term of the cost function expresses a penalty for maps that vary too much across space, and a prefer- ence for smooth maps. This can be understood from the point of view of the smoothness of the spatial inputs and the smoothness of the neuronal transfer function. The second term of the cost function expresses a penalty for maps in which a neuron has to fire for very long periods of time, subject to neural fatigue. In23 the grid pattern emerges as a compromise solution between these oppo- site requirements. There we give two examples of the kernel, which can be treated analytically. The analysis involves going into Fourier space, where the firing map is decomposed into 2D Fourier modes Ψi(~x) = a0 +Xi ai cos(~ki · ~x + φi) (A3) and the functional becomes 20 L = 1 2Xi a2 i k2 i + γ 2 0 K(0) + γ 2 Xi a2 i K(ki) (A4) where K simply denotes the 2D Fourier transform of the kernel, and k, again, is the length of the 2D vector ~k. As discussed in23, among 2D periodic solutions the fa- vored ones are those with a superposition of 3 cosines Ψ(~x) = 1 + (2/3) 3 Xi=1 cos(~ki · ~x) (A5) where for simplicity we have omitted the phases by suit- ably defining the origin. To model a degree of speed anisotropy, with faster runs e.g. along the axes or along the diagonals of a square en- vironment, one can simply add a quadrupole term to the functional L, the precise form of which is not essential, like the precise form of the kernel K is not essential ei- ther. One simple choice is Lanis = a2 i k2 i + γ 2 0 K(0) + 1 2Xi + ζXi h(k3 γ a2 i 2 Xi i − 1/√2i , K(ki) (A6) xi + k3 yi)/k3 where the quadrupole term, with a strength ζ, adds a cost ζ(√2 − 1)/√2 for each ~k aligned with the x- or y-axis, and no cost for those oriented at ±π/4, hence favoring ~k vectors along the diagonals (rotating the quadrupole by π/4 one can favor instead those aligned with the axes). The addition of the extra term makes it hard to get an- alytical solutions, but numerical minima of the cost func- tion can be found with a gradient descent algorithm that takes into account the additional constraint on the triplet of ~ki vectors and their amplitudes ai, i = 1, 2, 3. We per- formed several numerical minimizations of Lanis using as initial conditions perfect grids oriented in all possible di- rections , with the 3 ~ki at 2π/3 of each other, their length such as to minimize L, and ai ≡ 2/3. The numerical algo- rithm rapidly converges to a solution with one somewhat shorter ~k vector aligned with one of the two orthogonal directions of faster speed (in Fig.17, 3π/4) and the other two vectors somewhat longer and more than 2π/3 of each other. Whatever the initial orientation of the original symmetric grid, the algorithm finds a solution very close to either of the two "exact" solutions, the one shown in Fig.17 and the equivalent reflection along the vertical or horizontal axis. In this way, when only the single cell level effect is considered, the shorter vector, indicating roughly the direction of the ellipse, stands at ±π/4 (or at 0, π/2 if the faster speed are along the axes). a b c d 21 Figure 17: The input (a) and output (c) produced by the gradient descent algorithm that adjusts the ~ki vectors to minimize a cost function Lanis that includes a quadrupole term (panel b, in Fourier space). Depending on the exact initial condition, the algorithm converges towards the theoretically optimal solution (which has angles ei = −110, 20, −45 deg and the two longer ~ki vectors of identical length) but it stops before reaching it, as solutions with a jitter of ±2 deg have the same cost, to the fourth significant digit. Panel c shows in different colors the orientation distribution of the three eis, clustering into two modes. However, the shallow landscape of the cost function (for any reasonable form of the quadrupole anisotropy) indicates that the tight alignment seen experimentally cannot result from single-unit mechanisms alone. ∗ Electronic address: bailusi,[email protected] † Electronic address: [email protected] 1 W. B. Scoville and B. Milner, J. Neurol. Neurosurg. Psy- chiatry 20, 11 (1957). 19 M. E. Hasselmo, Hippocampus 18, 1213 (2008). 20 E. A. Zilli and M. E. Hasselmo, The Journal of neuro- science 30, 13850 (2010). 21 J. Koenig, A. N. Linder, J. K. Leutgeb, and S. Leutgeb, 2 J. O'Keefe and J. Dostrovsky, Brain research 34, 171 Science 332, 592 (2011). (1971). 3 J. S. Taube, R. U. Muller, and J. B. Ranck, The Journal of neuroscience 10, 420 (1990). 4 M. Fyhn, S. Molden, M. Witter, E. Moser, and M.-B. Moser, Science 305, 1258 (2004). 22 M. P. Brandon, A. R. Bogaard, C. P. Libby, M. A. Con- nerney, K. Gupta, and M. E. Hasselmo, Science 332, 595 (2011). 23 E. Kropff and A. Treves, Hippocampus 18, 1256 (2008). 24 D. L. Garden, P. D. Dodson, C. O'Donnell, M. D. White, 5 T. Hafting, M. Fyhn, S. Molden, M.-B. B. Moser, and E. I. and M. F. Nolan, Neuron 60, 875 (2008). Moser, Nature 436, 801 (2005). 6 F. Sargolini, M. Fyhn, T. Hafting, B. L. Mc- naughton, M. P. Witter, M.-B. Moser, and E. I. Moser, Science 312, 758 (2006), data available at: http://www.ntnu.no/cbm/gridcell. 7 L. M. Giocomo, M.-B. Moser, and E. I. Moser, Neuron 71, 589 (2011). 25 R. F. Langston, J. A. Ainge, J. J. Couey, C. B. Canto, T. L. Bjerknes, M. P. Witter, E. I. Moser, and M.-B. B. Moser, Science 328, 1576 (2010). 26 T. J. Wills, F. Cacucci, N. Burgess, and J. O'Keefe, Science 328, 1573 (2010). 27 H. Stensland, T. Kirkesola, E. Moser, and M.-B. Moser, Society for Neuroscience abstract 101.14 (2010). 8 B. McNaughton, F. Battaglia, O. Jensen, E. Moser, and 28 L. L. Colgin, E. I. Moser, and M.-B. Moser, Trends in M.-B. Moser, Nature Reviews Neurosci. 7, 663 (2006). 9 M. Fuhs and D. Touretzky, J. Neurosci. 26, 4266 (2006). 10 A. Guanella, D. Kiper, and P. Verschure, International journal of neural systems 17, 231 (2007). 11 Y. Burak and I. R. Fiete, PLoS Comput Biol 5, e1000291 (2009). 12 Z. Navratilova, L. M. Giocomo, J.-M. M. Fellous, M. E. Hasselmo, and B. L. McNaughton, Hippocampus (2011). 13 K. Zhang, The Journal of neuroscience 16, 2112 (1996). 14 A. Samsonovich and B. L. McNaughton, The Journal of neuroscience 17, 5900 (1997). 15 D. M. Walters and S. M. Stringer, Biological Cybernetics 103, 21 (2010). Neurosciences 31, 469 (2008). 29 M. Fyhn, T. Hafting, A. Treves, M.-B. Moser, and E. Moser, Nature 446, 190 (2007). 30 C. N. Boccara, F. Sargolini, V. H. H. Thoresen, T. Solstad, M. P. Witter, E. I. Moser, and M.-B. B. Moser, Nature neuroscience 13, 987 (2010). 31 A. Dhillon and R. S. G. Jones, Neuroscience 99, 413 (2000). 32 M. S. Petkovic, Famous Puzzles of Great Mathematicians (American Mathematical Society, College Station, Texas, 2009). 33 C. Barry, R. Hayman, N. Burgess, and K. Jeffery, Nat. Neurosci. 10, 682 (2007). 16 N. Burgess, C. Barry, and J. O'Keefe, Hippocampus 17(9), 34 F. Savelli, D. Yoganarasimha, and J. Knierim, Hippocam- 901 (2007). pus 18(12), 1270 (2008). 17 L. M. Giocomo, E. A. Zilli, E. Frans´en, and M. E. Has- 35 C. Barry, J. O'Keefe, and N. Burgess, Society for Neuro- selmo, Science 315, 1719 (2007). 18 N. Burgess, Hippocampus 18, 1157 (2008). science abstract 101.24 (2009). 36 J. H. Zar, Biostatistical Analysis (4th Edition) (Prentice 22 Hall, 1998). 37 T. Solstad, E. Moser, and G. Einevoll, Hippocampus 16, 1026 (2006). 38 E. Rolls, S. Stringer, and T. Elliot, Network: Comput Neu- ral Sys 15, 447 (2006). 39 I. R. Fiete, Y. Burak, and T. Brookings, The Journal of neuroscience 28, 6858 (2008). 43 D. Derdikman, J. R. Whitlock, A. Tsao, M. Fyhn, T. Haft- ing, M.-B. B. Moser, and E. I. Moser, Nature neuroscience 12, 1325 (2009). 44 B. Si, J. M. Herrmann, and K. Pawelzik, in Proceedings of the International Conference on Natural Computation (2007), pp. 177–182. 45 M. Franzius, R. Vollgraf, and L. Wiskott, J. Comput. Neu- 40 B. Si and A. Treves, Cognitive Neurodynamics 3, 177 rosci. 22, 297 (2007b). (2009). 41 T. van Haeften, L. Baks-Te-Bulte, P. H. Goede, F. G. Wouterlood, and M. P. Witter, Hippocampus 13, 943 (2003). 42 A. Treves, Journal of Computational Neuroscience 14, 271 (2003). 46 D. Samu, P. Eros, B. Ujfalussy, and T. Kiss, Biological Cybernetics 101, 19 (2009). 47 M. J. Milford, J. Wiles, and G. F. Wyeth, PLoS Comput Biol 6, e1000995 (2010).
1809.09757
1
1809
2018-09-25T23:28:32
Towards Subject and Diagnostic Identifiability in the Alzheimer's Disease Spectrum based on Functional Connectomes
[ "q-bio.NC" ]
Alzheimer's disease (AD) is the only major cause of mortality in the world without an effective disease modifying treatment. Evidence supporting the so called disconnection hypothesis suggests that functional connectivity biomarkers may have clinical potential for early detection of AD. However, known issues with low test-retest reliability and signal to noise in functional connectivity may prevent accuracy and subsequent predictive capacity. We validate the utility of a novel principal component based diagnostic identifiability framework to increase separation in functional connectivity across the Alzheimer's spectrum by identifying and reconstructing FC using only AD sensitive components or connectivity modes. We show that this framework (1) increases test-retest correspondence and (2) allows for better separation, in functional connectivity, of diagnostic groups both at the whole brain and individual resting state network level. Finally, we evaluate a posteriori the association between connectivity mode weights with longitudinal neurocognitive outcomes.
q-bio.NC
q-bio
Towards Subject and Diagnostic Identifiability in the Alzheimer's Disease Spectrum based on Functional Connectomes Diana O. Svaldi*1[0000-0001-6265-1931], Joaquín Goñi*1,2[0000-0003-3705-1955], Apoorva Bhar- thur Sanjay1, Enrico Amico2, Shannon L. Risacher1, John D. West1, Mario Dzemi- dzic1, Andrew Saykin1, and Liana Apostolova1 1 Indiana University School of Medicine, Indianapolis IN 46202, USA 2 Purdue University, Lafayette IN 47907, USA [email protected], [email protected] Abstract. Alzheimer's disease (AD) is the only major cause of mortality in the world without an effective disease modifying treatment. Evidence supporting the so called "disconnection hypothesis" suggests that functional connectivity bi- omarkers may have clinical potential for early detection of AD. However, known issues with low test-retest reliability and signal to noise in functional connectivity may prevent accuracy and subsequent predictive capacity. We validate the utility of a novel principal component based diagnostic identifiability framework to in- crease separation in functional connectivity across the Alzheimer's spectrum by identifying and reconstructing FC using only AD sensitive components or con- nectivity modes. We show that this framework (1) increases test-retest corre- spondence and (2) allows for better separation, in functional connectivity, of di- agnostic groups both at the whole brain and individual resting state network level. Finally, we evaluate a posteriori the association between connectivity mode weights with longitudinal neurocognitive outcomes. Keywords: Alzheimer's Disease, Functional Connectivity, Principal Compo- nent Analysis, resting state fMRI 1 Introduction Developing biomarkers for early detection of Alzheimer's disease (AD) is of critical importance as researchers believe clinical trial failures are in part due to testing of ther- apeutic agents too late in the disease [1]. The AD disconnection syndrome hypothesis [2] posits that AD spreads via propagation of dysfunctional signaling, indicating that functional connectivity (FC) biomarkers have potential for early detection. Despite this potential, known issues with high amounts of variability in acquisition and prepro- cessing of resting state fMRI, and ultimately low disease-related signal to noise ratio in FC [3], remain a critical barrier to incorporating FC as a clinical biomarker of AD. Recent work validated the utility of group level principal component analysis (PCA) to denoise FC by reconstructing subject level FC using PCs which optimized test-retest reliability through a measurement denominated differential identifiability [4]. Building 2 on this work, we expand the utility of the framework to increase separation across di- agnostic groups in the AD spectrum by reconstructing individual FC using AD sensitive PCs. We identify AD sensitive PCs using a novel diagnostic identifiability metric (D). We evaluate the proposed method with data from the Alzheimer's Disease Neuroimag- ing Initiative (ADNI2/GO) using group balanced, bootstrapped random sampling. 2 Methods Subject Demographics 2.1 Of the original 200 ADNI2/GO individuals with resting state fMRI scans, subjects were excluded if they (1) had only extended resting state scans, (2) had no Amyloid status provided, (3) were cognitively impaired, but Amyloid-beta protein negative (Ab-) neg- ative, and/or had (4) over 30% of fMRI time points censored (see 2.2). The final sample included 82 individuals. Only Ab positive (Ab+) individuals were included in cogni- tively impaired groups to avoid confounding by non-AD neurodegenerative patholo- gies. Subjects were sorted into 5 diagnostic groups using criterion from ADNI2/GO and Ab positivity: (1) normal controls (CNAss-, n = 15), (2) pre-clinical AD (CNAss+, n = 12), (3) early mild cognitive impairment (EMCIAss+, n = 22), (4) late mild cognitive impairment (LMCIAss+, n = 12), and (5) dementia (ADAss+, n = 21). Ab status was deter- mined using either mean PET standard uptake value ratio cutoff (Florbetapir > 1.1, University of Berkley) or CSF Ass levels [5]. Composite scores were calculated for visuospatial, memory, executive function, and language domains [6] from the ANDI2/GO battery. No demographic group effects were observed. All neurocognitive domain scores exhibited a significant group effect (Table1). Table 1. Demographics and Neurocognitive Comparisons of Diagnostic Groups. Variable Age (Years) (SD) Sex (% F) Years of Education (SD) Visuospatial Domain Score (SD)** CNAss- (n = 14) 74.2 (8.8) 64.2 16.7 (2.3) 9.7 (0.61) CNAss+ (n = 12) 75.9 (7.0) 41.7 15.8 (2.6) 9.3 (0.9) EMCIAss+ (n = 22) 72.6 (5.2) 50 15.2 (2.6) 9.4 (0.9) LMCIAss+ (n = 13) 73.3 (6.1) 61.6 16 (1.8) 83 (2.3) ADAss+ (n = 21) 73.5 (7.6) 42.9 15.4 (2.6) 7.4 (2.1) Language Domain Score (SD)** 49.2 (4.2) 48.8 (4.4) 46.2 (5.8) 43.1 (8.0) 34.8 (9.6) Memory Domain Score (SD)** 125.4 (41.1) 142 (34.5) 104.9 (46.6) 81.0 (36.7) 34.2 (21.8) Executive Function Domain Score (SD)** 99.0 (26.8) 117.6 (27.4) 135.0 (48.6) 166.3 (102.0) 284.6 (101.0) ** Significant group effect (Chi-square or ANOVA as appropriate, a = 0.05) fMRI Data Processing 2.2 3 MRI scans used for construction of FC matrices included T1-weighted MPRAGE scans and EPI fMRI scans from the initial visit in ADNI2/GO (www.adni-info.org for protocols). fMRI scans were processed in MATLAB using an FSL based pipeline fol- lowing processing guidelines by Power et al [7] and described in detail in Amico et al. [8]. Subjects with over 30% of volumes censored due to motion were discarded to en- sure data quality. For purposes of denoising FC matrices [4], processed fMRI time se- ries were split into halves, representing "test" and "retest" sessions. 2.3 Test-Retest Identifiability and Construction and of Individual FC Matrices For each subject, two FC matrices were created from the "test" and "retest" halves of the fMRI time-series. FC nodes were defined using a 286 region parcellation [9], as detailed in Amico et al. [8]. Functional connectivity matrices were derived by calculat- ing the pairwise Pearson correlation coefficient (𝑟"#) between the mean fMRI time-se- ries of all nodes. "Test" and "retest" FCs were de-noised by using group level PCA to maximize test-retest differential identifiability (Idiff) [4]. The "identifiability matrix" I was defined as the matrix of pairwise correlations (square, non-symmetric) between the subjects' FCtest and FCretest. The dimension of I is N2 where N is the number of subjects in the cohort. Self-identifiability, (Iself, Eqn. 1), was defined to be the average of the main diagonal elements of I, consisting of correlations between FCtest and FCretest from the same subjects. Iothers (Eqn. 2), was defined as average of the off-diagonal elements of matrix I, consisting of correlations between FCtest and FCretest of different subjects. Differential identifiability (Idiff, Eqn. 3) was defined as the difference between Iself and Iothers. 𝐼%&'(=1𝑁 𝑰",# (1) ".# 𝐼123&4%=1𝑁 (2) 𝑰",# "5# 𝐼7"((=100∗ (𝐼%&'(−𝐼123&4%) (3) Group level PCA [10] was applied in the FC domain, on a data matrix (Y1) contain- ing vectorized FCtest and FCretest (upper triangular) from all subjects. PCs throughout this paper will be numbered in order of variance explained. The number of PCs esti- mated was constrained to 2*N, the rank of the data matrix Y1. Following decomposi- tion, PCs were iteratively added in order of variance explained. Denoised FCtest and FCretest matrices were reconstructed using the number of PCs (n) that maximized Idiff (Eqn3), while maintaining a minimum Iothers value of 0.4, such that between-subject FC was neither overly correlated (loss of valid inter-subject variability) nor overly orthog- onal (inter-subject variability dominated by noise). This was done because the ADNI2/GO fMRI data was noisier than data on which this method was previously im- plemented, as evidenced by a much lower original between-subject FC correlation 4 (Iothers 0.22 ADNI vs. 0.4 Human Connectome Project rs-fMRI). Therefore, not setting a minimum threshold for Iothers led to the algorithm picking PCs that were "specialized" to specific subjects. The threshold 0.4 was specifically chosen because it reflected av- erage Iothers values seen in FCs from previous data, on which this method was imple- mented [4]. Final, de-noised FC matrices were computed as the average of FCtest and FCretest. Nodes were assigned to 9 resting state subnetworks (RSN/RSNs), visual (VIS), somato- motor (SM), dorsal attention (DA), ventral attention (VA), limbic (L), fronto-parietal (FP), and default mode network (DMN) [11] with the additional subcortical (SUB) and cerebellar (CER). 2.4 Diagnostic Identifiability With the goal of early detection in mind, we hypothesized that FC in non-dementia groups would become significantly less identifiable from FC in 𝐴𝐷?@A with increased diagnostic proximity to 𝐴𝐷?@A. Figure 1 delineates the work flow for finding AD sen- entiability in connectivity between each non-dementia group (g) and 𝐴𝐷?@A and is cal- corr(g,g), minus the average correlation between that non-dementia group and 𝐴𝐷?@A, corr(g, 𝐴𝐷?@A). D, rather than variance explained, was used to filter components, as it culated from the correlation matrix (I) of Y2, containing final, de-noised FC from all subjects. Dg was defined as the average correlation within a non-dementia group, sitive PCs using a novel diagnostic identifiability metric (D), which quantifies differ- was hypothesized that early disease changes likely do not account for a large portion of between subject variance. 𝐷B=𝑐𝑜𝑟𝑟𝑔,𝑔 −𝑐𝑜𝑟𝑟𝑔,𝐴𝐷?@A (4) Group level PCA was again performed on the matrix Y2. Here, the number of PCs was constrained to n = 35 PCs, the rank of the Y2 matrix. Y2 was iteratively reconstructed using a subset of the n PCs, selected based on maximizing Dg. Starting with PC1, PC2…n were iteratively added based on their influence on average(Dg). At each iteration, the PCj* which most improved average(Dg) upon its addition to previously selected PCs, was selected. To avoid results driven by a subset of the population or by differences in sample sizes between groups, the cohort was randomly sampled 30 times, following total cohort PCA, in a group balanced fashion (nsample = 50; ng = 10). The number of bootstraps was chosen to allow adequate estimation of the Dg distribution while keeping run-time of the algorithm, reasonable. Bootstrapped distributions of Dg were generated for each number of PCs. The number of PCs (n*) which maximized average(Dg) was found (Eqn. 5). AD sensitive PCs were defined as those which appeared within the n* most influential PCs with the greatest frequency across samples. Final FC matrices were re-constructed using only AD sensitive PCs. 𝑛∗= 𝑛,𝑎𝑡 𝑎𝑟𝑔𝑚𝑎𝑥L(1𝑔 𝐷B(n)) (5) B Additionally, Dg curves were estimated and disease sensitive PCs were identified for the 9 RSNs individually, by calculating IRSN using the subset of connections where at least one of the nodes in the connection was part of the RSN. 5 Fig. 1. Diagnostic Identifiability workflow. 2.5 Statistical Validation and Association with Neurocognitive Outcomes Due to the small number of bootstraps, differences between Dg distributions were as- sessed at n* PCs by checking if the median of one distribution was an outlier relative to a reference distribution using non-parametric confidence intervals defined with the median and interquartile range (IQR). First, Dg distributions from each RSN were com- pared to those from WB. Next, WB and RSN Dg distributions were compared to a cor- responding null model. Null models for the WB and each RSN were constructed by randomly permuting diagnostic group membership among individuals selected at each bootstrap, such that Dg for the null model represented identifiability of a random heter- ogeneous group from a random heterogeneous reference group. Finally, individual D values (Di) were calculated for each subject using FC reconstructed with the n* PCs. ANOVA (a < 0.05) with follow up pairwise tests, was performed on WB Di distribu- tions to test for a group effect. Stepwise regressions (F-test, a = 0.05), starting with gender, age and education, were be used to test for associations between the n* PC weights and longitudinal changes in neurocognitive outcomes (0, 1, 2 years post imag- ing). 3 Results 3.1 Test-Retest Identifiability Figure 2 details the results of denoising FC using differential test-retest differential identifiability. An optimal reconstruction based on the first n = 35 PCs (in decreasing order of explained variance) was chosen (Figure 2A). Iself increased from 0.52 to 0.92 (Figure2A-B) while Iothers increased from 0.20 to 0.40 (Figure 2A-B). Idiff increased from 38% to 57% (Figure2A-B). 6 Fig. 2. (A) Iself, Iothers, and Idiff across the range of # PCs. (B) I matrices for original and denoised FC matrices. (C) Example original FC matrix versus denoised FC matrix. 3.2 Diagnostic Identifiability WB average(Dg) peaked at n* = 11 components which explained 58.82% of the vari- ance in the denoised FC data (Figure 3A, Table2). At n* PCs, LMCIAss+ was the only group who that did not exhibit significantly increased Dg from the null model. At n* components, Di distributions exhibited a significant group effect. Di decreased with diagnostic proximity to ADAss+ (Figure 3B). Between-subject correlation in FC in- creased from 0.41 to 0.71 after reconstruction with n* PCs (Figure 3B). Of the 9 RSNs, the L network exhibited significantly greater DRSN as compared to WB (Table2). Like WB, LMCIAss+ was the only group that did not exhibit significantly greater RSN Dg than the null model, with the exceptions of SM where EMCIAss+ was additionally not signif- icantly different from the null model and L where all non-dementia groups exhibited greater Dg than the null model (Table2). Eight of eleven PCs were identified as disease sensitive in all 9 RSNs and WB (Table 2). 7 Fig. 3. (A-left) Whole brain Dg across all possible number of PCs. (A-right) Individual Di val- ues at n* = 11 PCs. Distributions showing significant differences (t-test, p < 0.05) are deline- ated using lines. (B) Denoised I matrix (n = 35 PCs) versus I matrix reconstructed using disease sensitive PCs (n* = 11 PCs). (C) Example denoised FC matrix versus FC matrix reconstructed using disease sensitive PCs. Table 2. Diagnostic Identifiability Summary. RSN WB VIS SM DA VA L FP DMN SUB CER CNAss- 11.35** 13.21** 9.82** 12.16** 10.74** 17.18** 12.07** 12.09** 14.17** 13.29** CNAss+ 9.75** 10.33** 12.96** 11.24** 11.65** 13.76** 10.25** 10.20** 11.66** 12.85** EMCIAss+ 7.85** 8.11** 7.30 8.09** 7.96** 11.97** 9.34** 8.39** 9.93** 9.72** LMCIAss+ 2.60 2.75 4.43 3.37 2.33 6.28** 2.70 2.84 4.33 5.67 Mean 7.89 8.60 8.62 8.71 8.17 12.30 8.59 8.38 10.02 10.38 n 11 10 10 13 10 8 11 11 9 10 Var (%) 58.82 57.23 57.26 62.19 57.26 54.50 58.82 58.82 55.78 57.26 **Median outside CI null model, Median outside CI WB mean(Dg) Four PCs exhibited significant associations with various neurocognitive do- main scores (Table 3). Visuospatial domain scores were associated with PC 17 at 1 year post imaging and PC 9 at 2 years post imaging. Memory domain scores were associated with PC 32 at 1 year post imaging and PC 7 at 2 years post imaging. Language domain scores were associated with PC 23 at 0 year post imaging and PC 7 at 1 years post imaging. Finally, PC 17 was associated with executive domain scores at 1 and 2 years post imaging. Table 3. Associations of n* PC weights with neurocognitive composite domain scores. Step- wise regressions (F-test, a < 0.05) were used to assess the relationship of neurocognitive com- posite domain scores with PC weights, with age, gender, and education starting in the base model; p values are reported for the whole model, adjusted-R2 is reported for the model. Time Points 0 1 2 4 Visuospatial Memory Language Executive PC - 17 9 p - 0.001 0.020 R2 - 0.53 0.46 PC - 32 7 p - 0.032 0.044 R2 - 0.31 0.20 PC 23 7 - p 0.040 0.025 - R2 0.19 0.31 - PC - 17 17 p - 0.004 0.013 R2 - 0.48 0.36 Limitations, Future Work, and Conclusions We present here a two stage PCA based framework to improve the detection of AD signatures in whole-brain functional connectivity. We first use recently proposed test- retest differential identifiability to denoise subject-level functional connectomes and 8 consequently reduce dimensionality of functional connectomes. We subsequently in- troduce and validate the concept of PCA based differential diagnostic identifiability to increase AD signal to background in functional connectivity. The result of a significant diagnostic group effect in diagnostic differential identifiability shows that FC contains AD signature, even at early stages of disease. The finding of increased diagnostic iden- tifiability in Limbic regions, known to be associated with memory processes and known to be affected in AD, further validates this finding. Finally, we show that PC weights from AD sensitive principal components are correlated to longitudinal neurocognitive outcomes. In addition to the work presented here, we plan to delve further into the meaning of the PCs themselves. AD sensitive PCs did not appear to be specific to indi- vidual RSNs, as the same PCs were consistently AD sensitive across RSNs. Further- more, several PCs were associated with multiple neurocognitive domains. Therefore, AD sensitive PCs may characterize global brain changes related to AD. However, spa- tial representation of PCs and relationship of PCs with network properties need to be explored to further assess this. Finally, to further validate these promising results, this methodology needs to be applied to a larger cohort. With ADNI3 data becoming avail- able (~300 subjects already scanned), on which all subjects underwent resting state fMRI, we will be able to further validate findings and further improve identification and characterization of AD sensitive PCs based on whole brain functional connectomes. This dual decomposition/reconstruction framework makes forward progress in exploit- ing the clinical potential of functional connectivity based biomarkers. 5 References 1. Sperling, R.A., J. Karlawish, and K.A. Johnson, Preclinical Alzheimer disease-the chal- lenges ahead. Nat Rev Neurol, 2013. 9(1): p. 54-8. 2. Brier, M.R., J.B. Thomas, and B.M. Ances, Network dysfunction in Alzheimer's disease: refining the disconnection hypothesis. Brain Connect, 2014. 4(5): p. 299-311. 3. Braun, U., et al., Test-retest reliability of resting-state connectivity network characteristics using fMRI and graph theoretical measures. Neuroimage, 2012. 59(2): p. 1404-12. 4. Amico, E. and J. Goni, The quest for identifiability in human functional connectomes. Sci Rep, 2018. 8(1): p. 8254. 5. Shaw, L.M., et al., Cerebrospinal fluid biomarker signature in Alzheimer's disease neuroim- aging initiative subjects. Ann Neurol, 2009. 65(4): p. 403-13. 6. Wilhalme, H., et al., A comparison of theoretical and statistically derived indices for pre- dicting cognitive decline. Alzheimers Dement (Amst), 2017. 6: p. 171-181. 7. Power, J.D., et al., Spurious but systematic correlations in functional connectivity MRI net- works arise from subject motion. Neuroimage, 2012. 59(3): p. 2142-54. 8. Amico, E., et al., Mapping the functional connectome traits of levels of consciousness. Neu- roimage, 2017. 148: p. 201-211. 9. Shen, X., et al., Groupwise whole-brain parcellation from resting-state fMRI data for net- work node identification. Neuroimage, 2013. 82: p. 403-15. 10. Hotelling, H., Analysis of complex variables into principal components. Journal of Educa- tional Psychology, 1933. 24: p. 417-441. 11. Yeo, B.T.T., et al., Estimates of segregation and overlap of functional connectivity networks in the human cerebral cortex. NeuroImage, 2014. 88: p. 212-227.
1803.04364
1
1803
2018-02-13T01:04:40
Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band
[ "q-bio.NC", "cs.DM", "stat.ML" ]
The functional significance of resting state networks and their abnormal manifestations in psychiatric disorders are firmly established, as is the importance of the cortical rhythms in mediating these networks. Resting state networks are known to undergo substantial reorganization from childhood to adulthood, but whether distinct cortical rhythms, which are generated by separable neural mechanisms and are often manifested abnormally in psychiatric conditions, mediate maturation differentially, remains unknown. Using magnetoencephalography (MEG) to map frequency band specific maturation of resting state networks from age 7 to 29 in 162 participants (31 independent), we found significant changes with age in networks mediated by the beta (13-30Hz) and gamma (31-80Hz) bands. More specifically, gamma band mediated networks followed an expected asymptotic trajectory, but beta band mediated networks followed a linear trajectory. Network integration increased with age in gamma band mediated networks, while local segregation increased with age in beta band mediated networks. Spatially, the hubs that changed in importance with age in the beta band mediated networks had relatively little overlap with those that showed the greatest changes in the gamma band mediated networks. These findings are relevant for our understanding of the neural mechanisms of cortical maturation, in both typical and atypical development.
q-bio.NC
q-bio
Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 1 of 41 Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band S. Khan1,4,5‡*, J. A. Hashmi1,4‡, F. Mamashli1,4, K. Michmizos1,4, M. G. Kitzbichler1,4, H. Bharadwaj1,4, Y. Bekhti1,4, S. Ganesan1,4, K. A Garel1, 4, S. Whitfield-Gabrieli5, R. L. Gollub2,4, J. Kong2,4, L. M. Vaina4,6, K. D. Rana6, S. S. Stufflebeam3,4, M. S. Hämäläinen3,4, T. Kenet1,4 1Department of Neurology, MGH, Harvard Medical School, Boston, USA. 2Department of Psychiatry MGH, Harvard Medical School, Boston, USA. 3Department of Radiology, MGH, Harvard Medical School, Boston, USA. 4Athinoula A. Martinos Center for Biomedical Imaging, MGH/HST, Charlestown, USA 5McGovern Institute for Brain Research, Massachusetts Institute of Technology, Cambridge, USA 6Department of Biomedical Engineering, Boston University, Boston, USA ‡equal contribution Running Title: Multifaceted Development of Brain Connectivity with Age *Corresponding author: Sheraz Khan, Ph.D. Athinoula A. Martinos Center for Biomedical Imaging Massachusetts General Hospital Harvard Medical School Massachusetts Institute of Technology 149 13th Street, CNY-2275 Boston, MA-02129, USA Phone: +1 617-643-5634 Fax: +1 617-948-5966 E-mail: [email protected] Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 2 of 41 ABSTRACT The functional significance of resting state networks and their abnormal manifestations in psychiatric disorders are firmly established, as is the importance of the cortical rhythms in mediating these networks. Resting state networks are known to undergo substantial reorganization from childhood to adulthood, but whether distinct cortical rhythms, which are generated by separable neural mechanisms and are often manifested abnormally in psychiatric conditions, mediate maturation differentially, remains unknown. Using magnetoencephalography (MEG) to map frequency band specific maturation of resting state networks from age 7 to 29 in 162 participants (31 independent), we found significant changes with age in networks mediated by the beta (13-30Hz) and gamma (31-80Hz) bands. More specifically, gamma band mediated networks followed an expected asymptotic trajectory, but beta band mediated networks followed a linear trajectory. Network integration increased with age in gamma band mediated networks, while local segregation increased with age in beta band mediated networks. Spatially, the hubs that changed in importance with age in the beta band mediated networks had relatively little overlap with those that showed the greatest changes in the gamma band mediated networks. These findings are relevant for our understanding of the neural mechanisms of cortical maturation, in both typical and atypical development. Keywords: development, brain connectivity, rhythms, graph theory, magnetoencephalography Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 3 of 41 1. INTRODUCTION Synchronous neuronal activity in the brain gives rise to rhythms, that are known to be functionally significant. These rhythms are commonly divided into five fundamental frequency bands, most commonly classified as delta (1-2 Hz), theta (3-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (31-80 Hz) (Buzsáki G 2006). One of the hypothesized roles of these rhythms is in forming neuronal ensembles, or networks, via local and longer-range synchronization, across spatially distributed regions (Fries P 2005; Siegel M et al. 2012; Bastos AM et al. 2015; Fries P 2015). Brain networks that emerge in the absence of any directive task or stimulus, referred to as resting state networks (Raichle ME et al. 2001; Raichle ME 2015), have attracted particular interest due to their consistency across and within individuals. Abnormalities in these networks are also emerging as a hallmark of psychiatric and developmental disorders (Broyd SJ et al. 2009; Toussaint PJ et al. 2014; Kitzbichler MG et al. 2015), further underscoring their functional significance. While resting state networks have been studied extensively using fcMRI (functional connectivity MRI), a technique that relies on the slow hemodynamic signal and thus has a maximal temporal resolution of about 1Hz, studies using high temporal resolution magnetoencephalography (MEG), have confirmed that the five fundamental faster rhythms mediate these networks in non-overlapping patterns (de Pasquale F et al. 2010; Hipp JF et al. 2012) As part of understanding the function of resting state networks in general, and their role in cognitive development and neurodevelopmental disorders in particular, it is important to map their maturational trajectories, from childhood to adulthood. To date, our knowledge of maturational changes in macro-scale functional networks in the developing brain is largely based on task-free fcMRI studies. Several such studies have shown developmental changes in resting state brain networks, where regions associated with separate networks become connected while closely linked local subnetworks lose some of their connections with maturation (Dosenbach NU et al. 2010; Sato JR et al. 2014; Sato JR et al. 2015). Most of these studies have concluded that network integration, how well different Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 4 of 41 components of the network are connected, increases with maturation, while network segregation, the differentiation of the network into modules, or clusters, decreases with maturation. The spatial distribution of hubs, the most highly connected brain regions, also changes with maturation. Another feature examined in prior studies is the small-world property of brain networks. Small world networks optimize the balance between local and global efficiency. fcMRI studies have not documented a change in the small world property of brain networks with maturation from childhood, around age 7, to adulthood, around age 31 (Fair DA et al. 2009). Network resilience, a measure of the robustness of the network as hubs are removed, which has been used to assess robustness in psychiatric disorders (Lo CY et al. 2015), has been shown to be age dependent in infants (Gao W et al. 2011), but age dependency through maturation has not been studied. It has also been shown that the association between global graph metrics characterizing network properties and the ages of the participants follows an asymptotic growth curve (Dosenbach NU et al. 2010). While fMRI studies have greatly increased our understanding of the development of resting state networks from childhood to adulthood, the relative temporal coarseness of fcMRI makes it impossible to differentiate maturational trajectories by frequency bands (Hipp JF and M Siegel 2015). Mapping the contributions of distinct frequency bands to maturational trajectories is critical because these rhythms are associated with distinct neurophysiological generators (Uhlhaas PJ et al. 2008; Ronnqvist KC et al. 2013), have been mapped to a multitude of cognitive functions (Harris AZ and JA Gordon 2015), are known to themselves change in power and phase synchrony with maturation (Uhlhaas PJ et al. 2009; Uhlhaas PJ et al. 2010). To better understand the contribution of individual rhythms to network maturation, we used MEG, which measures magnetic fields associated with neural currents with millisecond time resolution, and has a spatial resolution on the order of a centimeter (Lin FH, JW Belliveau, et al. 2006). We chose to use graph theory with connectivity measured using envelope correlations (Hipp JF et al. 2012) as the core metric, to analyze cortical resting state (relaxed fixation) MEG signals from 131 individuals (64 Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 5 of 41 females), ages 7 to 29, in each of the five fundamental frequency bands. We focused on five well- studied graph theory metrics because the approach is well-suited for studying global network properties also in the functional domain (Bullmore E and O Sporns 2009; Rubinov M and O Sporns 2010; Bullmore E and O Sporns 2012; Misic B et al. 2016; Bassett DS and O Sporns 2017). The results were then validated using similar data from 31 individuals (16 females, ages 21-28) from an independent early adulthood resting state data set (Niso G et al. 2015). The full distribution of participants is shown in Figure S1 in SM. Lastly, to determine the relevance of these graph metrics to the maturation of resting state networks within each frequency band, we used machine learning to quantify the extent to which the MEG derived graph metrics can be used to predict age, similarly to a prior resting state networks study that used fMRI data (Dosenbach NU et al. 2010). We then assessed whether the data from the independent dataset fit on the same curves. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 6 of 41 2. MATERIALS AND METHODS The analysis stream we followed is illustrated in Figure 1. Figure 1: Schematic illustration of pipeline. From top left in a clockwise direction: Resting state data are acquired using MEG, and then mapped to the cortical surface. The surface is then divided into regions (parcellated), and envelopes are calculated for each frequency band, in each region. The connectivity between the regions is then computed from the envelopes, and, finally, connectivity metrics are derived. 2.1 Experimental Paradigm The resting state paradigm consisted of a red fixation cross at the center of the screen, presented for 5 minutes continuously, while participants were seated and instructed to fixate on the cross. The fixation stimulus was generated and presented using the psychophysics toolbox (Brainard, 1997; Pelli, 1997), and projected through an opening in the wall onto a back-projection screen placed 100 cm in front of the participant, inside a magnetically shielded room. 2.2 Participants 2.2.1 Massachusetts General Hospital (MGH) Based Participants Our primary data were collected from 145 healthy subjects, ages 7-29, at MGH. Due to excessive motion, data from 14 subjects were discarded. Because datasets from different MGH based studies Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 7 of 41 were combined here, no uniform behavioral measures were available across all participants. IQ measured with the Kaufman Brief Intelligence Test – II (Kaufman and Kaufman, 2004) was available for 68 of the participants. Within this subgroup, no significant change in IQ with age was observed (Figure S2), as expected, given that IQ is normalized by age. All the studies that were pooled for this analysis screened for typical development and health, but the approach varied. The full age and gender distribution of the participants is shown in Figure S1-A. 2.2.2 OMEGA Project Participants (McGill University) To test our results on an independent dataset, resting-state MEG scans from 31 additional young adult participants (ages 21-28) were obtained from the OMEGA project (Niso G et al. 2015), and chosen by order with gender matching to the MGH cohort in that same age range, subject to age restrictions. Note that the OMEGA project spans ages 21-75. While we would have liked to test our results on data from younger subjects, no pediatric MEG resting state data are currently openly available, so this was not possible. The age and gender distributions of the participants are shown in Figure S1-B. 2.3 MRI/MEG Data Acquisition 2.3.1 MRI Data Acquisition and Processing T1-weighted, high resolution MPRAGE (Magnetization Prepared Rapid Gradient Echo) structural images were acquired on either a 1.5T or a 3.0-T Siemens Trio whole-body MRI (magnetic resonance) scanner (Siemens Medical Systems) using either 12 channels or 32 channel head coil at MGH and MIT. The structural data was preprocessed using FreeSurfer (Dale AM et al. 1999; Fischl B et al. 1999). After correcting for topological defects, cortical surfaces were triangulated with dense meshes with ~130,000 vertices in each hemisphere. To expose the sulci in the visualization of cortical data, we used the inflated surfaces computed by FreeSurfer. For the 31 OMEGA dataset participants, templates were constructed Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 8 of 41 from age and gender matched subjects from our data. The same processing steps are followed on this template MRIs. 2.3.2 MEG Data Acquisition and Processing MEG data were acquired inside a magnetically shielded room (Khan S and D Cohen 2013) using a whole-head Elekta Neuromag VectorView system composed of 306 sensors arranged in 102 triplets of two orthogonal planar gradiometers and one magnetometer. The signals were filtered between 0.1 Hz and 200 Hz and sampled at 600 Hz. To allow co-registration of the MEG and MRI data, the locations of three fiduciary points (nasion and auricular points) that define a head-based coordinate system, a set of points from the head surface, and the locations of the four HPI coils were determined using a Fastrak digitizer (Polhemus Inc., Colchester, VT) integrated with the VectorView system. ECG as well as Horizontal (HEOG) and Vertical electro-oculogram (VEOG) signals were recorded. The position and orientation of the head with respect to the MEG sensor array were recorded continuously throughout the session with the help of four head position indicator (HPI) coils (Cheour M et al. 2004). At this stage, we monitored the continuous head position, blinks, and eye movements in real time, and the session was restarted if excessive noise due to the subject's eyes or head movement is recorded. In particular, the subjects and head coils position were carefully visually monitored continusouly, and the session was also restarted if any slouching in the seated position was observed. Pillows, cushions, and blankets were fitted to each individual to address slouching, and readjusted as needed if any slouching was observed. In addition to the human resting state data, 5 min of data from the room void of a subject were recorded before each session for noise estimation purposes. The acquisition parameters are published elsewhere for the 31-subjects from the OMEGA dataset (Niso G et al. 2015). The OMEGA CTF files were converted into a .fif file using mne_ctf2fif function from MNE (Gramfort A et al. 2014). After this conversion, the subsequent processing was the same as for the Vector View data. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 9 of 41 2.4 Noise suppression and motion correction The data were spatially filtered using the signal space separation (SSS) method (Taulu S et al. 2004; Taulu S and J Simola 2006) with Elekta Neuromag Maxfilter software to suppress noise generated by sources outside the brain. Since shielded room at MGH is three layers and we have exclusion criteria for subject having dental artifact, only SSS is applied and it temporal extension tSSS was not used. This procedure also corrects for head motion using the continuous head position data described in the previous section. Since SSS is only available for Electa MEG systems, it was not applied for OMEGA subjects, where data were collected with a CTF MEG system. The heartbeats were identified using in-house MATLAB code modified from QRS detector in BioSig (Vidaurre C et al. 2011). Subsequently, a signal- space projection (SSP) operator was created separately for magnetometers and gradiometers using the Singular Value Decomposition (SVD) of the concatenated data segments containing the QRS complexes as well as separately identified eye blinks (Nolte G and MS Hämäläinen 2001). Data were also low-pass filtered at 144 Hz to eliminate the HPI coil excitation signals. 2.5 Cortical Space Analysis 2.5.1 Mapping MEG Data onto Cortical Space The dense triangulation of the folded cortical surface provided by FreeSurfer was decimated to a grid of 10,242 dipoles per hemisphere, corresponding to a spacing of approximately 3 mm between adjacent source locations. To compute the forward solution, a boundary-element model with a single compartment bounded by the inner surface of the skull was assumed (Hämäläinen MS and J Sarvas 1989). The watershed algorithm in FreeSurfer was used to generate the inner skull surface triangulations from the MRI scans of each participant. The current distribution was estimated using the regularized minimum-norm estimate (MNE) by fixing the source orientation to be perpendicular to the cortex. The regularized (regularization = 0.1) noise covariance matrix that was used to calculate the inverse operator was estimated from data acquired in the absence of a subject before each session. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 10 of 41 This approach has been validated using intracranial measurements (Dale AM et al. 2000). To reduce the bias of the MNEs toward superficial currents, we incorporated depth weighting by adjusting the source covariance matrix, which has been shown to increase spatial specificity (Lin FH, T Witzel, et al. 2006). All forward and inverse calculations were done using MNE-C (Gramfort A et al. 2014). 2.5.2 Correlation between age, and the norms of the columns of the gain matrix We also examined the correlation between age, and the norms of the columns of the gain matrix, separately for magnetometers and gradiometers. The magnetometers showed no significant correlation with age. The gradiometers showed a small correlation in the frontal pole, in a region that was not identified as a significant hub in our results (see Figure S3). This is most likely due to the different in head size between the youngest participants and the oldest ones. To minimize the effect of these differences, we use both magnetometer and gradiometer data in our source estimation procedure. 2.5.3 Cortical Parcellation (Labels) FreeSurfer was used to automatically divide the cortex into 72 regions (Fischl B et al. 2004). After discarding "medial wall" and "corpus callosum", these regions were further divided in to a total of N=448 cortical "labels (Figure S4)", so that each label covers a similar area again using FreeSurfer. We employed this high-resolution parcellation scheme because cortical surface is very convoluted and averaging across a large label, which crosses multiple sulci and gyri, can result in signal cancellation across the parcel. Lastly, a high-resolution parcellation also reduces the dependence of the results on the specific selection of the parcels. We also checked that field spread (spatial signal leakage) between labels is not impacted by age by examining the correlation between age and the mean cross-talk across labels (Hauk O et al. 2011). The results showed no correlation (Figure S5). 2.5.4 Averaging the time series across a label Owing to the ambiguity of individual vertex (dipole) orientations, these time series were not averaged directly but first aligned with the dominant component of the multivariate set of time series Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 11 of 41 before calculating the label mean. In order to align sign of the time series across vertices, we used SVD of the data 𝑿𝑿𝑇𝑇=𝑼𝑼Σ𝑾𝑾𝑇𝑇. The sign of the dot product between the first left singular vector 𝑼𝑼 and all other time-series in a label was computed. If this sign was negative, we inverted the time-series before averaging. To verify that the label time series are meaningful, we computed the Power Spectral Density (PSD, see Spectral Density, below) for occipital, frontal, parietal, and temporal cortical regions within each age group, see Figure S6. 2.6 Time Series Analysis 2.6.1 Filtering and Hilbert transform The time series were band-pass filtered and down sampled for faster processing, while making sure that the sampling frequency was maintained at 𝑓𝑓𝑠𝑠>3𝑓𝑓hi (obeying the Nyquist theorem and avoiding aliasing artifacts). The chosen frequency bands were delta (1-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (31-80 Hz). The line frequency at 60 Hz was removed with a notch filter of bandwidth 1 Hz. Hilbert transform was then performed on this band pass data. 2.6.2 Hilbert Transform time series with its Hilbert transform into a complex time series: For each individual frequency band the analytic signal 𝑋𝑋�(𝑡𝑡) was calculated by combining the original The resulting time series 𝑋𝑋�(𝑡𝑡) can be seen as a rotating vector in the complex plane whose length corresponds to the envelope of the original time series 𝑥𝑥(𝑡𝑡) and whose phase grows according to the 𝑋𝑋�(𝑡𝑡)=𝑥𝑥(𝑡𝑡)+𝚥𝚥 ℋ[𝑥𝑥(𝑡𝑡)}]−−−(1) dominant frequency. Figure 1, step 4 shows an example of a modulated envelope on the top of the band pass data (carrier). An example of envelope PSD for the gamma frequency band is shown in Figure S7. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 12 of 41 At this stage further artifact cleaning was performed as follows: signal spikes where the amplitude is higher than 5σ over the course were identified and dropped over a width of 5 periods. To remove the effect of microsaccades, HEOG and VEOG channels were filtered at a pass-band of 31–80 Hz. The envelope was then calculated for the filtered signals and averaged to get REOG. Peaks exceeding three standard deviations above the mean calculated over the whole-time course, were identified and the corresponding periods were discarded from subsequent analysis. Head movement recordings from the HPI coils were used to drop these one second blocks where the average head movement exceeded 1.7 mm/s (empirical threshold). The amount of data lost through cleaning was well below 10% and did not differed significantly with age. 2.6.3 Orthogonal correlations We used a method based upon envelope correlations to reliably estimate synchronicity between different cortical labels (Colclough G et al. 2016). In contrast to phase-based connectivity metrics envelope correlations measure how the amplitude of an envelope within a frequency band is synchronously modulated over time across distinct cortical regions, as illustrated in the fourth panel of Figure 1. Previous studies (humans and primates) have demonstrated the validity and functional significance of these synchronous envelope amplitude modulations (Brookes MJ et al. 2011; Vidal JR et al. 2012; Wang L et al. 2012; Brookes MJ et al. 2016; Colclough G et al. 2016) for both oscillatory and broadband signals. To address the field-spread problem associated with MEG data (Sekihara K et al. 2011), we used the previously described orthogonal (Hipp JF et al. 2012) variation of envelope correlation metric. This method requires any two putatively dependent signals to have non-zero lag and is thus insensitive to the zero-lag correlations stemming from field-spread. Mathematically, the connectivity between two complex signals 𝑋𝑋� and 𝑌𝑌� is calculated by "orthogonalizing" one signal with respect to the other 𝑌𝑌�(𝑡𝑡,𝑓𝑓)→𝑌𝑌�⊥𝑋𝑋(𝑡𝑡,𝑓𝑓), and subsequently taking the Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 13 of 41 Pearson correlation between their envelopes. This is done in both directions and the two results are averaged to give the final connectivity measureC⊥(𝑋𝑋�,𝑌𝑌�;𝑡𝑡,𝑓𝑓). 𝑌𝑌�⊥𝑋𝑋(𝑡𝑡,𝑓𝑓)=ℑ�𝑌𝑌�(𝑡𝑡,𝑓𝑓)𝑋𝑋�†(𝑡𝑡,𝑓𝑓) 𝑋𝑋�(𝑡𝑡,𝑓𝑓)�𝑒𝑒⊥𝑋𝑋(𝑡𝑡,𝑓𝑓)−−−(2) C⊥�𝑋𝑋�,𝑌𝑌�;𝑡𝑡,𝑓𝑓�=𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶��𝑋𝑋��,�𝑌𝑌�⊥𝑋𝑋��+𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶��𝑌𝑌��,�𝑋𝑋�⊥𝑌𝑌�� −−(3) 2 Due to the slow time course of these envelopes and to ensure enough independent samples are available in the correlation window (Hipp JF et al. 2012), we calculated the orthogonal connectivity using an overlapping sliding window of 30 seconds with a stride of 1/8 of the window size. Figure S7 demonstrate this method on gamma envelopes. Lastly, note that this method does not address the problem of spurious correlations appearing due to source spread (Sekihara K et al. 2011), which is difficult to address. Therefore, spurious correlations coupled with the use of high resolution cortical parcellation can skew the topological metrics used in this study. Please see (Palva JM et al. 2017) for a comprehensive and detailed discussion of these confounds. 2.7 The Connectivity and Adjacency Matrices As a starting point for calculating the graph theoretic metrics we used the connectivity matrix, which separate matrix was computed for each frequency band. The final result of the processing pipeline is a contained the orthogonal correlations between all 𝑁𝑁×𝑁𝑁 node pairs and at each time window. A connectivity array of dimension 𝑁𝑁×𝑁𝑁×𝑁𝑁Time×𝑁𝑁Bands for each subject. In order to increase signal to Thresholding and binarizing the connectivity matrix results in the adjacency matrix 𝔸𝔸. noise, we collapsed the connectivity array along the temporal dimension by taking the median of each pairwise orthogonal correlation across time windows. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 14 of 41 We used a threshold proportional scheme to retain a given proportion of the strongest connectivity matrix entries in 𝔸𝔸. Specifically, adjacency matrix 𝔸𝔸, were constructed using a fixed cost threshold, ensuring that the density or number of connections of the network is equated across all individuals and age groups. Cost is a measure of the percentage of connections for each label in relation to all connections of the network. Since the total number of connections is the same for all participants, and is determined by the number of nodes being considered, the use of a fixed cost, i.e. fixed percentage threshold, allows for exactly equal numbers of connections across participants. This is important to ensure graph metrics can be compared across all individuals and age groups. As there is no rationale for using a cost threshold, therefore we compared graph network properties for a wide range of cost, we used thresholding range from 5% to 30% at increments of 5%. For the graph metrics to be reliable, it should be consistent over range of thresholds. Please see Table S1 for example of consistency of our results. We also checked that important hubs for all frequency bands are in line with previous published studies (Brookes MJ et al. 2011; Hipp JF et al. 2012; Colclough G et al. 2016) as shown in Figure S8. The adjacency matrix 𝔸𝔸 defines a graph 𝒢𝒢 in the form of pairs of nodes that are connected by an edge. Thus, 𝔸𝔸 is defined such that its binary elements 𝔸𝔸𝑖𝑖𝑖𝑖 are either 1 or 0 depending whether the edge 𝑒𝑒𝑖𝑖𝑖𝑖 between nodes 𝑣𝑣𝑖𝑖 and 𝑣𝑣𝑖𝑖 exists or not: 𝔸𝔸𝑖𝑖𝑖𝑖=�1 if ∃ 𝑒𝑒𝑖𝑖𝑖𝑖 0 if ∄ 𝑒𝑒𝑖𝑖𝑖𝑖 In addition, we also computed a weighted adjacency matrix which preserves the correlation values above the same thresholds. A more comprehensive description can be found elsewhere (Watts DJ and SH Strogatz 1998). For Figures 3B,D and 4B,D, the original adjacency matrices were averaged within age-groups and then thresholded for visualization purposes. 2.7.1 Path length The average shortest path length between all pairs of nodes was calculated as: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 15 of 41 𝐿𝐿= −−−(4) 1𝑛𝑛(𝑛𝑛−1) � 𝑑𝑑𝑖𝑖𝑖𝑖 𝑖𝑖≠𝑖𝑖;𝑣𝑣𝑖𝑖,𝑣𝑣𝑗𝑗∈𝒢𝒢 𝒅𝒅𝒊𝒊𝒊𝒊=𝐦𝐦𝐦𝐦𝐦𝐦 {𝒏𝒏𝔸𝔸𝒏𝒏[𝒊𝒊,𝒊𝒊]≠𝟎𝟎} edges one has to traverse in order to get from one node to the other where the topological distance 𝒅𝒅𝒊𝒊𝒊𝒊 between nodes 𝒗𝒗𝒊𝒊 and 𝒗𝒗𝒊𝒊 is defined as the minimum number of where 𝔸𝔸𝑛𝑛 denotes the 𝑛𝑛th power of the adjacency matrix 𝔸𝔸 and 𝑖𝑖 and 𝑗𝑗 are row and column indices of The degree of a node 𝒗𝒗𝒊𝒊 in a Graph 𝓖𝓖 is defined as 𝑛𝑛 𝐷𝐷𝑖𝑖= � 𝑒𝑒𝑖𝑖𝑖𝑖 𝑖𝑖=1,𝑖𝑖≠𝑖𝑖 the resulting matrix. 2.7.2 Degree −−−(5) th column edge of adjacency matrix 𝔸𝔸. where ije is the i th row and j The degree maps, averaged across all participants in the adult group (ages 22-29), for all bands, are shown in Figure S8. We also computed weighted degree (unthresholded mean connectivity with respect to age, also known as mean node strength) for alpha and beta (Figure S9), and, as expected, the results are in line with prior findings showing increase in overall connectivity with age (Schäfer CB et al. 2014). 2.7.3 Clustering coefficient The local clustering coefficient in the neighborhood of vertex 𝒗𝒗𝒊𝒊 is defined as the ratio of actual and maximally possible edges in the graph 𝓖𝓖𝒊𝒊, which is equivalent to the graph density of 𝓖𝓖𝒊𝒊: 𝐶𝐶𝑖𝑖= 2��𝑒𝑒𝑖𝑖𝑗𝑗�� 𝑘𝑘𝑖𝑖(𝑘𝑘𝑖𝑖−1):𝑣𝑣𝑖𝑖,𝑣𝑣𝑗𝑗∈𝒢𝒢𝑖𝑖 −−−(6) Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 16 of 41 2.7.4 Global and local efficiencies Global efficiency measures the efficiency of information transfer through the entire network, and is assessed by mean path length. While the concept of path length is intuitive in anatomical networks, it is also relevant for functional networks, since a particular functional connection may travel different anatomical paths, and while the correspondence between the two is generally high, it is not necessarily identical (Bullmore E and O Sporns 2009; Misic B et al. 2016; Bassett DS and O Sporns 2017). Local efficiency is related to the clustering of a network, i.e. the extent to which nearest neighbors are 𝐸𝐸glob=𝐸𝐸(𝒢𝒢)= interconnected. Thus, it assesses the efficiency of connectivity over adjacent brain regions. The average global efficiency of information transfer in graph 𝒢𝒢 having 𝑛𝑛 nodes can be calculated from the inverse of the edge distances 𝑑𝑑𝑖𝑖,𝑖𝑖 The quantity above is a measure of the global efficiency of information transfer for the whole graph 𝒢𝒢. There is also a local efficiency for each vertex 𝑣𝑣𝑖𝑖 measuring how efficiently its neighbours can communicate when vertex 𝑣𝑣𝑖𝑖 is removed. If the subgraph of all neighbors of 𝑣𝑣𝑖𝑖 is denoted by 𝒢𝒢𝑖𝑖, then its local efficiency 𝐸𝐸(𝒢𝒢𝑖𝑖) is approximately equivalent to the clustering coefficient 𝐶𝐶𝑖𝑖 (Achard S and E 1𝑛𝑛(𝑛𝑛−1) � 1𝑑𝑑𝑖𝑖𝑖𝑖 −−−(7) 𝑖𝑖≠𝑖𝑖;𝑣𝑣𝑖𝑖,𝑣𝑣𝑗𝑗∈𝒢𝒢 Bullmore 2007). 𝐸𝐸loc=1𝑛𝑛� 𝑣𝑣𝑖𝑖∈𝒢𝒢 𝐸𝐸(𝒢𝒢𝑖𝑖) −−−(8) We also computed weighted analogues of local and global efficiencies which used C⊥-weighted edge distances 𝑑𝑑𝑖𝑖,𝑖𝑖. This weighted analogue shows the same trend as a function of age as the unweighted one, please see Figure 2 and Figure S10. 2.7.5 Small World Small world property is a measure of optimization of the balance between short and long-range connections (Bassett DS and E Bullmore 2006). When graph 𝒢𝒢 that provides optimal balance between Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 17 of 41 local and global information transfer, it is called a small-world graph. The small-worldness of a network world network if it has a similar path length but greater clustering of nodes than an equivalent Erdös- is often characterized by two key metrics: the clustering coefficient 𝐶𝐶 and the characteristic path length 𝐿𝐿. To evaluate the small-world topology of brain networks, these topological parameters must be benchmarked against corresponding mean values of a null random graph. A network 𝒢𝒢 is is a small- Rényi (E–R) random graph 𝒢𝒢𝑟𝑟𝑟𝑟𝑛𝑛𝑟𝑟: where 𝐶𝐶𝑟𝑟𝑟𝑟𝑛𝑛𝑟𝑟 and 𝐿𝐿𝑟𝑟𝑟𝑟𝑛𝑛𝑟𝑟 are the mean clustering coefficient and the characteristic path length of the equivalent 𝒢𝒢𝑟𝑟𝑟𝑟𝑛𝑛𝑟𝑟 graph. 𝑆𝑆𝑆𝑆=𝐶𝐶/𝐶𝐶𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 𝐿𝐿/𝐿𝐿𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 ≥ 1 −−−(9) 2.7.6 Betweenness Centrality Betweenness centrality pertains to individual nodes in the network, rather than the network as a whole, and assesses how many of the shortest paths between all other node pairs in the network pass through that node. Nodes with high betweenness centrality (hubs) are therefore more important for overall network efficiency. The betweenness centrality of node i is defined as: 𝑏𝑏𝑖𝑖=∑ 𝑚𝑚≠𝑖𝑖≠𝑛𝑛∈𝐺𝐺 −−−(10) 𝜎𝜎𝑚𝑚𝑟𝑟(𝑖𝑖) 𝜎𝜎𝑚𝑚𝑟𝑟 where σmn is the total number of shortest paths (paths with the shortest path length) from node m to node n, and σmn(i) is the number of shortest paths from node m to node n that pass through node i. Betweenness centrality of a node thus reflects the control and influence of that node on other nodes. Nodes with high betweenness centrality have a high impact on information transferal and collaboration between disparate sub-networks. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 18 of 41 2.7.7 Resilience Resilience measures the robustness of the network if the most heavily connected nodes (hubs) are removed. This measure is inversely related to small world property (Peng G-S et al. 2016). We chose this measure because it has been studied, mostly using fMRI, in the context of psychiatric disorders, where multiple hubs might be functioning abnormally (Achard S et al. 2006; Lo CY et al. 2015). It has also been shown that greater resilience in a functionally derived task-driven network is associated with greater inhibitory control cognitively (Spielberg JM et al. 2015), a function that is often impaired in neurodevelopmental and psychiatric disorders. Importantly, the measure incorporates network topology in conjunction with the spatial distribution of hubs, because it takes the degree, i.e. the number of connections, of individual nodes into account. Resilience quantifies the Graph 𝒢𝒢's robustness to targeted or random attacks. Targeted attacks remove nodes in the descending order of degree. At each attack, global efficiency is computed. Robustness is defined as the ratio of the original efficiency with efficiency calculated after attack. This process is repeated until all nodes are removed. Graph where the connectivity probability follows a power law distribution (i.e. scale free networks) are very robust in the face of random failures. This property of networks mathematically described by a power law function 𝑝𝑝(𝑘𝑘)∝ 𝑘𝑘−𝑦𝑦 where 𝑝𝑝(𝑘𝑘) is the probability of a node having k links is and y is the exponent. When plotted in a log-log plot, this relationship follows a straight line with slope -y, i.e. resilience is the slope of the degree distribution. 2.8. Spectral Analysis The Power Spectral Density (PSD) (0.1 to 80Hz, logarithmically distributed) in Figure S6 was computed on single trial data from each label using the multitaper method (MTM) based on discrete prolate spheroidal sequences (Slepian sequences) taper(Thomson DJ 1982) with 1.5 Hz smoothing as implemented in MNE-Python. PSD (0.033 to 1Hz, logarithmically distributed) in Figure S7 of gamma envelope was also computed using multitaper method with spectral smoothing of 0.02 Hz. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 19 of 41 2.9. Correlation To evaluate the relationship between a network quantity and age, we used Spearman correlation (degree of freedom = 129). The p-values were computed after correcting for multiple comparisons across the correction space of frequency bands, thresholds, and graph metrics by controlling for family- wise error rate using maximum statistics through permutation testing (Groppe DM et al. 2011). The corrected p-values are shown in Table S1. Specifically, the correction for multiple comparisons and frequency bands, and finally taking maximum correlation value across this permuted correction were computed by first randomizing age and then correlating it with all graph metrics at all thresholds was done by constructing an empirical null distribution. For this purpose, 𝑛𝑛𝑝𝑝 = 1000,000 realizations space. This null distribution is shown in Figure S11. The corrected p-values (𝑝𝑝𝑐𝑐) were calculated as: where 𝑛𝑛 is the number values in the empirical null distribution greater or lower than the observed 𝑝𝑝𝑐𝑐= 2(𝑛𝑛+1) 𝑛𝑛𝑝𝑝+1 −−−(11) positive or negative correlation value, respectively. The factor of two stems from the fact that the test is two-tailed. Correlations resulting in significant p-values were then again tested using Robust Correlation (Pernet CR et al. 2013), which strictly checks for false positive correlations using bootstrap resampling. Effect size for correlation was computed using cohen's 𝑑𝑑: 2𝐶𝐶√1−𝐶𝐶2�−−−(12) where 𝐶𝐶 is the correlation coefficient. 𝑑𝑑= � 2.10. Bootstraping For the purposes of visualizing the significance of age effects and assessing uncertainties in the graph metrics with respect to age, we used nested bootstrapping with 1024 realizations. The nested bootstrap procedure approximates the joint distribution of age x with the age-dependent Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 20 of 41 network metric f (yx), where f (yx) is the average network metric over many subjects of age x (see notes below). We observed 𝑛𝑛 pairs (𝑥𝑥𝑖𝑖,𝑦𝑦𝑖𝑖), where 𝑥𝑥𝑖𝑖 is the age and 𝑦𝑦𝑖𝑖 the corresponding imaging data for the ith subject. Ideally, we would like to observe (xi, y¯x), where y¯x denotes the (conceptual) average of subjects chosen at random from a population, where each subject is of age x. Let f (y) denote the function which maps imaging data to a scalar metric describing some aspect of a network. Since yi contains noise, to visualize and estimate uncertainties in graph metrics we can approximate (xi, y¯x) by (x¯∗, y¯∗), where the ∗ denotes a bootstrap sample. We can then evaluate f (yx¯∗) instead of f (yi). Each realization of bootstraping yielded one average network metric and one value for the mean age of the group. Each data point on the normalized density colormap corresponds to one realization of the bootstrap (Figure 2, Figure 4A, B and Figure 5 inserts). Lastly, the correlation values were computed using the original data only, not on the bootsrapped data described here. 2.11 LOWESS Regression We used the non-parametric LOESS regression to fit a curve to the data (Cleveland WS and C Loader). To protect against overfitting in estimating bandwidth, we used 10-fold cross validation. We generated our predictive model using the data in the training set, and then measured the accuracy of the model using the data in the test set. We tested a range of bandwidths from .01 to .99 with .01 step. The bandwidth resulting in least sum of squares error was then selected (Webel K 2006). Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 21 of 41 2.12. Machine Learning (Multivariate regression using Random Forest) All the graph metrics calculated for each band were combined under a single non-parametric multivariate regression model using Random-Forest (Breiman L 2001; Liaw A and M Wiener 2002). More specifically: a) Maturity Index (prediction of age using random forest regression model) for beta band (MI-beta) was computed using the beta band network graph metrics: local efficiency, small world property, and resilience. b) Maturity Index for gamma band (MI-gamma) was computed using the gamma band network graph metrics: global efficiency, small world property and resilience. c) Maturity Index for both beta and gamma band (MI-combined) was computed using all of the above features. Lastly, we fit a parametric curve to each of the maturity index, parametric model for curve fitting was selected using Akaike information criterion (AIC). Below we provide details for each step of this procedure. 2.12.1 Random Forest Regression Analysis Random forest (RF), a method for non-parametric regression which is robust in avoiding overfitting (Breiman L 2001), was implemented in R using the randomForest package (Liaw A and M Wiener 2002). As a further step to avoid overfitting, each RF regressor model was trained on the stratified 70% of the data (training set) and then tested in remaining stratified 30% test-set. The possibility of bias introduced by the random choice of the test set was avoided by repeating the sampling 1000 times. The final model represents the aggregate of the 1000 sampling events. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 22 of 41 2.12.2 Random Forest Parameter optimization The optimal number of variables randomly sampled as candidates at each split was selected as p/3 where p is the number of features in the model (Breiman L 2001). A total of 1,000 decision trees were grown to ensure out-of-bag (OOB) error converges (Breiman L 2001). 2.12.3 Features (MI-beta, MI-gamma, MI-combined) MI-beta (Figure 6A) was derived using machine learning with three features computed in the beta band networks: local efficiency, small world property, and resilience as manifested by the slope of the degree distribution. MI-gamma (Figure 6B) was derived using machine learning with three features computed in the gamma band networks: global efficiency, small world property and resilience as manifested by the slope of the degree distribution. MI-combined (Figure 6C), as its name implies, was computed using machine learning with all six of the above features. The predicted ages for all subjects from the random forest regression model were converted to the maturation indices defined above, using a published scaling scheme (Dosenbach NU et al. 2010), by setting the mean predicted brain age to 1, for typically developed young adults. Feature (variable) importance was also computed using random forest regression model by estimating Percent Increase Mean Square Error (%IncMSE). %IncMSE was obtained by permuting the values of each features of the test set and comparing the prediction with the unpermuted test set prediction of the feature (normalized by the standard error)., where %IncMSE is the average increase in squared residuals of the test set when the feature is permuted. A higher %IncMSE value represents a higher variable importance (Figure 6D). 2.12.4 Curve Fitting The models (Linear, exponential, quadratic, Von Bertalanffy) for best fitted curve (Matlab's Curve Fitting Toolbox) were compared using Akaike information criterion (AIC). Given a set of models for the data, AIC is a measure that assesses the quality of each model, relative to the remaining models in the set. The chosen model minimizes the Kullback-Leibler distance between the model and the ground truth. The model with the lowest AIC is considered the best model among all models specified for the data at Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 23 of 41 hand. The absolute AIC values are not particularly meaningful since they are specific to the data set being modeled. The relative AIC value (ΔAICi = AICi - min[AICp]) is used to rank models (Akaike H 1974). The model with the minimum AIC was selected as the best model. To quantify goodness of fit, we also computed R-squared (R2; coefficient of determination) for best fitted regression models. 2.12.5 Independent Data Set Verification To verify whether the regression models obtained here were consistent with data collected independently, in a different site, using a different MEG machine (CTF 275), MEG resting state scans from 31 participants from OMEGA project (Niso G et al. 2015) were also examined. Random forest regression models learned using our primary, MGH based, dataset, were then applied to the previously unseen OMEGA project data, with no additional training. The predicted maturity indices of the OMEGA sourced subject are shown as orange dots in Figure 6. To quantify goodness of fit on this independent dataset, we also computed R2 between the values from the OMEGA dataset and predictions from models learned from the MGH dataset. 2.12.6 Prediction Interval Calculation The prediction interval for the best fitted curve in Figure 6 was obtained using Scheffe's method (Maxwell SE and HD Delaney 1990). In this method, the prediction interval 𝑠𝑠(𝛼𝛼,𝑔𝑔) is defined as: where g is the model order (2), n is the number of subjects (131), 𝛼𝛼 is the significance level chosen as 0.05/𝑛𝑛𝑐𝑐 (𝑛𝑛𝑐𝑐=3; total number of curves), and 𝐹𝐹(1−𝛼𝛼,𝑔𝑔,𝑛𝑛−2) is the F-distribution. 𝑠𝑠(𝛼𝛼,𝑔𝑔)= �𝑔𝑔 𝐹𝐹(1−𝛼𝛼,𝑔𝑔,𝑛𝑛−2) −−−(11), Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 24 of 41 2.13 Graphical Representation Custom code was written in MATLAB to build circle plots for graphically representing connectivity in the three age groups. To represent nodes on the brain, custom code in PySurfer was written using Python. The increases and decreases in connections and topography were represented by using connectivity patterns shown as graphs, with the tool Gephi. This method uses a spring embedding data-driven technique to align regions in two dimensional space based on strength of connections (Fair DA et al. 2008). For Figure 3B, 3D, 4B and 4D, original adjacency matrices were averaged within age-groups and then thresholded for visualizations Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 25 of 41 3. RESULTS 3.1 Age-dependent trajectories of network integration by frequency band The local and global efficiency graph theory metrics were used to evaluate the age- dependent trajectory of local and global network integration, respectively. These metrics were evaluated in each of the five frequency bands. Significant age- dependent changes in local, but not global, efficiency emerged only in the beta band (Figure 2A). In Figure 2: Network efficiency increases locally in beta band networks and globally in gamma band networks. A. LOESS plot (solid white line) for the relationship between age and local network efficiency of beta band mediated networks. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure. B. Same, for gamma band mediated networks, and global efficiency. parallel, significant age-dependent changes in global, but not local, efficiency emerged only in the gamma band (Figure 2B). These changes were significant across multiple thresholds (Table S1-A,B). No other significant age dependent changes emerged in any of the other frequency bands (Figure S12). We also tested the weighted network analogues for the same metrics, and the results followed a similar trend (Figure S10). Therefore, the remaining computations were carried out in unweighted networks. 3.2 Age-dependent trajectories of cortical hubs by frequency band To assess age dependent changes in spatial distribution of hubs, we measured correlation between age and the betweenness centrality of nodes. In networks mediated by the beta band, loss of betweenness centrality score with age was seen mostly in frontal and temporal hubs, while gain of betweenness centrality score with age was seen mostly in parietal hubs (Figure 3A,B). In networks mediated by the gamma bands, loss of betweenness centrality score with age was seen mostly in Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 26 of 41 occipital hubs, while gain of betweenness centrality score with age was seen mostly in frontal and parietal hubs (Figure 3C,D). Figure 3: Spatial distribution and connectivity patterns of growing and shrinking betwnness centrality of hubs. A. Hub regions with growing (red) and shrinking (blue) betweenness centrality scores, in beta band mediated networks. B. Visual representation of the connections from 4 hubs with shrinking betweenness centrality scores (blue) and 1 hub with growing betweenness centrality scores (red), averaged for children (7-13), adolescents (14-21) and young adults (22-29), displayed at 0.25 thesholding in beta band mediated networks. C. Same as A, for the gamma band mediated networks. D. Same as B, for 3 growing and 1 shrinking hubs, for the gamma band mediated networks. Notation: IPS: right intraparietal sulcus, PFC: right dorsolateral prefrontal cortex, IFG: right inferior frontal gyrus, OFC: right orbitofrontal cortex, STG: superior temporal gyrus, FEF: right frontal eye field, IPS: right intraparietal sulcus, SMA: supplementary motor areas, Occ: Occipital cortex. 3.3 Age-dependent trajectories of small world property by frequency band Given the differentiation in age-dependent trajectories between the beta and gamma bands, we next examined the small world property of the networks mediated by each of these frequency bands. While network coefficients in both the beta and gamma bands met small world criteria, we found a significant increase with age in small world property with maturation in the beta band (Figure 4A,B), alongside a significant decrease with age in small world property in the gamma band (Figure 4C,D). Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 27 of 41 These changes in beta and gamma bands were also consistent across multiple thresholds (Table S1-C for threshold specific p-values). Figure 4: Small world property is age and frequency band dependent. A. LOESS plot (solid white line) for the relationship between age and small world property for beta band mediated networks. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure (same colorbar as in Figure 2A). B. A visual representation of connections and hubs in the three age groups, averaged for children (7-13), adolescents (14-21) and young adults (22-29), in the beta band mediated networks, displayed at 0.25 thesholding. Degree represents the size of the hub. C. Same as A, but for the gamma band mediated networks. D. Same as B, but for the gamma band mediated networks. 3.4 Age-dependent trajectories of network resilience by frequency band We next evaluated how network resilience, a graph theoretical metric which measures the vulnerability of the network to attacks (by removal) on the most connected hubs, changed with age in beta and gamma band mediated networks. To assess resilience, we quantified the reduction in global efficiency as hubs were removed in order of connectedness, from largest to smallest. We found significant age dependent differences in network resilience in both the beta band (Figure 5A) and the gamma band (Figure 5B) mediated networks. While the significance of the age effect differed by percentage of nodes removed, whenever there was a significant age effect, the trend was similar; resilience weakened with age in the beta band mediated networks, but strengthened with age in the Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 28 of 41 gamma band mediated networks, as shown in the insets of Figure 5. We repeated the same analysis with local efficiency as the parameter, as well as with attacking the connections rather than the hubs. The results followed the same trajectory in all cases, with resilience weakening with age in the beta band, and increasing with age in the gamma band. Figure 5: Resilience in Gamma and beta mediated networks follows opposite developmental trajectories. A: Main plot - Solid line mean resilience, combined for all ages, plotted as the decrease in global efficiency as a function of percent nodes removed in beta band mediated networks, from largest to smallest. Dashed lines mark confidence interval at two standard deviations. For each point on the average curve, we computed whether there was a significant age effect. When age was a factor in resilience, that line between the upper and lower confidence interval for that point on the curve was assigned a color. The color marks the correlation coefficient of the effect of age, thresholded at p<0.05 corrected (i.e. at significance). The colorbar at the bottom left shows the colormap of strength of the correlation coefficient. A-Inset: LOESS plot of one instance of the effect of age, at 54% of nodes removed, where significance of age effect was maximal. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure (same colorbar as in Figure 2A). While the exact numbers differed for different percent nodes removed, the pattern was always identical, showing a negative correlation with age for the gamma band mediated networks. B: Main plot - Same as A, but for gamma band mediated networks. The white patch between 60% and 80% nodes removed indicates there was no impact of age in that range. B-Inset: Same as A-Inset, for the beta band, at 19% of nodes removed, where significance of age effect was maximal. Again, whenever there was a significant age effect, marked by colors on the main curve, the pattern was identical, showing a positive correlation with age for the beta band mediated networks. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 29 of 41 3.5 Age prediction by frequency band specific properties We next tested whether the graph metrics assessed here could be used to predict individual brain maturity. Given the different trajectory observed in the beta and gamma band mediated networks, we began by assessing age-based prediction using random forest regression within each set of networks separately. For the beta band, we defined a beta Maturity Index, MI-beta, using local efficiency, small world property, and resilience parameters from the beta mediated networks. When MI- beta was plotted relative to age, we found that age prediction follows a linear trajectory (Figure 6A). For the gamma band, likewise, we defined a gamma maturity index, MI-gamma, using global efficiency, small world property, and resilience parameters from the gamma mediated networks. When MI-gamma was plotted relative to age, we found that prediction followed a non-linear quadratic asymptotic growth curve trajectory (Figure 6B). Combining the two sets of measures resulted in a non-linear Von Bertalanffy growth curve that yielded significantly increased prediction accuracy (Figure 6C). The relative information (explained variance) of each of the parameters is shown in Figure 6D. MI-beta and MI-gamma together accounted for 52% of the variance observed in the data. To further test the reliability of the model, which was trained solely on MGH data set, it was applied blindly to 31 participants from an independent dataset (OMEGA), with no additional training. The data were plotted alongside the MGH data, and indeed follow the same trajectories (Figure 6A-C). Figure 6: Classification by maturation curve. Green circles represent results for individual study participants ("MGH") Orange circles represent values for participants from the independent OMEGA database (OMEGA), that were not used during the learning phase in the machine learning analysis. A. MI-beta (R2=0.39 MGH, R2 = 0.34 OMEGA) plotted relative to age. B. MI-gamma (R2=0.48 MGH, R2=0.33 OMEGA) plotted relative to age. C. The combined MI for beta and gamma (R2=0.52 MGH, R2=0.41 OMEGA) plotted relative to age. D. The relative contribution (variable importance computed using random forest regression) of each of the parameters to the model. Notation: GE: global efficiency; LE: local efficiency; RI: Resilience index (see SM). SW: Small world property. Orange circles in panels A-C mark the participants from the OMEGA data set. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 30 of 41 4. DISCUSSION We found that from age 7 to 29, resting state networks mediated by the beta and gamma frequency bands underwent marked topological reorganization, while resting state networks mediated by the slower alpha, theta and delta frequency bands showed no significant age dependent changes in network topology, for the examined graph-theoretical metrics. Importantly, the patterns of age- dependent changes for the beta and gamma mediated networks differed substantially. Beta band mediated networks became more locally efficient, i.e. tending towards clustering and more connections with adjacent regions with age, while gamma band mediated networks became more globally efficient, i.e. tending towards shorter overall path lengths and thus faster communication across larger cortical distances, with age. Additionally, the contribution and importance of many hubs to the overall network efficiency, measured using betweenness centrality, grew or shrunk with age, but a different set of hubs showed this pattern for beta and gamma mediated networks, with relatively little overlap. Since small world property and resilience are inversely proportional to one another and both depend on the relative magnitude of local and global efficiencies, these measures presented opposite age-dependent trajectories for the beta and gamma mediated networks. Specifically, small world property, i.e. overall network optimization in balancing short and long connections, increased with age, while resilience, i.e. robustness of the network, decreased with age in the beta band. The pattern was exactly opposite in the gamma band mediated networks. Remarkably, the two sets of networks followed different growth trajectories, with the beta band mediated networks best described with a linear, rather than an asymptotic, growth curve, and the gamma band mediated networks best described by a more expected asymptotic growth curve (Dosenbach NU et al. 2010). These results extend prior fMRI based findings in several ways, most notably by determining that only two out of the five fundamental frequency bands, beta and gamma, mediated the resting state networks that showed age-related changes, and that each followed a distinct trajectory, and in the case of the beta band, that trajectory was unexpectedly linear. Furthermore, contrary to prior suggestions from fMRI based studies (Fair DA et al. 2009; Hwang K et al. 2013), we found that the small world Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 31 of 41 property, which assesses the overall balance of the network in optimizing local versus distant connections, did not remain constant through this age range, and similarly, network resilience at a given age depended on the underlying frequency. Lastly, as reported with fMRI (Fransson P et al. 2011; Menon V 2013), gamma band mediated networks showed development of hubs in heteromodal regions such as posterior parietal, posterior cingulate and the anterior insula. But unlike observations in fMRI studies, beta band mediated networks showed a loss of hubs in heteromodal-frontal regions, alongside growth in hubs in posterior parietal regions. The observation that only the resting state networks that were mediated by the gamma and beta frequency bands showed significant topological reorganization with age may be driven by the fact that these two bands are strongly associated with cognitive control (Buschman TJ and EK Miller 2014; Roux F and PJ Uhlhaas 2014), which matures over adolescence (Luna B et al. 2015). It is likely also related to the fact that both of these high frequency rhythms are heavily dependent on GABAergic systems (Uhlhaas PJ et al. 2008; Sohal VS et al. 2009), which themselves undergo extensive changes during development, well into adolescence. The pattern of reduced frontal hubs observed in the beta band is in line with observations showing reduced frontal task related activation with maturation, for instance for inhibitory control, potentially due to increased efficiency of top-down communication, putatively mediated by the beta band (Ordaz SJ et al. 2013). In particular, the linear growth trajectory of the beta band mediated networks could be the result of the continuing maturation and development of top-down projections, which may be more likely to be mediated via the beta band (Buschman TJ and EK Miller 2007; Wang XJ 2010). Indeed, many processes that are heavily mediated via top-down connections, such as attention and verbal functioning, peak past the age range examined here (Peters BD et al. 2014). In contrast, the gamma band mediated networks followed the more expected asymptotic trajectory, which may reflect the completion of maturation of bottom-up projections, which have been associated with greater probability with the gamma band (Buschman TJ and EK Miller 2007; Wang XJ 2010). Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 32 of 41 The results using the resilience metric, i.e the measure of the robustness of networks, are particularly intriguing in the context of psychiatric disorders (Lo CY et al. 2015). In our prior studies, we have found that resting state networks in ASD, ages 8-18, showed increased efficiency in the gamma band, but decreased efficiency in the beta band (Kitzbichler MG et al. 2015). In parallel, studies have shown reduced efficiency in the alpha band in bipolar disorder and schizophrenia (Hinkley LB et al. 2011; Kim DJ et al. 2013; Kim JS et al. 2014), and abnormal resting state network connectivity in the gamma band (Andreou C et al. 2015). The observation of resilience, a measure that depends on network efficiency, increased with age in the beta band but decreased with age in the gamma band, diverges from our original hypothesis of minimal resilience during adolescence. However, it is possible that greater vulnerability during adolescence arises from the fact that resilience is not optimized in this age range in either network, and thus both are relatively more vulnerable. The study does have several limitations the merit noting. One limitation is that we chose to focus on eyes open with relaxed fixation as our resting state paradigm, rather than eyes closed, thus minimizing alpha power. This was done to best align with parallel prior fMRI studies (Fair DA et al. 2007; Fair DA et al. 2009; Dosenbach NU et al. 2010; Grayson DS et al. 2014), and to follow the guidelines of the Human Connectome Project. Eyes-open resting state networks derived using MEG also have greater test-retest reliability than eyes-closed derived networks (Jin SH et al. 2011). While some MEG/EEG studies do find differences between the two conditions (Jin SH et al. 2011; Tan B et al. 2013; Tagliazucchi E and H Laufs 2014; Miraglia F et al. 2016; Yu Q et al. 2016), overall, the differences between the eyes closed and eyes open conditions in all of these studies were small. Another limitation of the proposed study is that we only had IQ measures available for a subset of the sample (N=68), and no other behavioral measures uniformly across the sample. While we were able to show that there is no relationship to IQ in the subset of the sample for which IQ was available (Figure S2), the absence of behavioral assessments means we were not able to link the measures to any specific cognitive measures. A minor limitation is a different in head size across development. Given Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 33 of 41 that brain size reaches 95% of its maximum size by age 6 (Giedd JN et al. 1999; Lenroot RK and JN Giedd 2006), and our minimum age is 7, the impact of changing brain size is likely slight (see also Figure S3 and methods 2.5.2), but cannot be completely dismissed. Another important limitation is that this study focuses solely on topological network properties. Developmental studies of coherence for instance, clearly show increased coherence in the beta band as well in the alpha band (Schäfer CB et al. 2014). Indeed, when we look at changes with age of "degree", which measures the mean functional connectivity of each node, we find age dependent changes in the alpha band as well (Figure S9), reproducing these prior results. Further studies will need to be carried out to elucidate the contributions and relevance of topological versus more direct non-topological properties such as coherence, to cognitive development. Lastly, the independent data set only spans ages 21-28, and not the full age range studied here. As of now, unfortunately, there are no pediatric shared data sets of resting state MEG data. Therefore, our independent data set was by necessity limited in age range. In summary, we show that developmental refinement of resting state networks as assessed by graph metrics is dependent on the mediating frequency band, and age dependent changes in global network properties occur only in the beta and gamma bands between the ages of 7 and 29. Specifically, we show that gamma band mediated networks become more globally efficient with maturation, while beta band mediated networks become more locally efficient with maturation. Accordingly, the small world property, which measures how optimally balanced local and global efficiencies are, increased with age in beta band mediated networks, and decreased with age in gamma mediated networks. To reconcile our results with prior fMRI findings on the development of topological network properties, we need to consider the fact that since fMRI signal cannot be used to distinguish signals from different frequency bands (Hipp JF and M Siegel 2015), prior fMRI-based studies observed results from all frequency bands combined. In such a scenario, it might indeed appear that the small world property remains unchanged with age, as would resilience, if the signals from the beta and gamma band were weighed roughly equally in the fMRI signal. Similarly, because the Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 34 of 41 combination of a linear and a non-linear function results in a non-linear function, as shown in Figure 6C, the linear maturation trajectory observed here in beta mediated networks would likely be missed by fMRI. This observed differentiation between beta and gamma band mediated networks could hint at underlying neural mechanisms in case of abnormal maturation. For instance, disorders that are more impacted in the gamma band might be more related to dysfunction in PV+ interneurons (Takada N et al. 2014). In contrast, disturbances in maturation in the beta band might be more attributable to inhibitory-inhibitory connections (Jensen O et al. 2005). In combination, our findings significantly advance our understanding of the complex dynamics behind oscillatory interactions that subserve the maturation of resting state cortical networks in health, and their disruptions in developmental and psychiatric or neurological disorders. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 35 of 41 5. ACKNOWLEDGMENTS: This work was supported by grants from the Nancy Lurie Marks Family Foundation (TK, SK, MGK), Autism Speaks (TK), The Simons Foundation (SFARI 239395, TK), The National Institute of Child Health and Development (R01HD073254, TK), The National Center for Research Resources (P41EB015896, MSH), National Institute for Biomedical Imaging and Bioengineering (5R01EB009048, MSH), and the Cognitive Rhythms Collaborative: A Discovery Network (NFS 1042134, MSH). 6. REFERENCES: Achard S, Bullmore E. 2007. Efficiency and cost of economical brain functional networks. PLoS Comput Biol 3:e17-e17. 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. Akaike H. 1974. A new look at the statistical model identification. Automatic Control, IEEE Transactions on 19:716-723. Andreou C, Nolte G, Leicht G, Polomac N, Hanganu-Opatz IL, Lambert M, Engel AK, Mulert C. 2015. Increased Resting-State Gamma-Band Connectivity in First-Episode Schizophrenia. Schizophr Bull 41:930-939. Bassett DS, Bullmore E. 2006. Small-World Brain Networks. The Neuroscientist 12:512-523. Bassett DS, Sporns O. 2017. Network neuroscience. Nature neuroscience 20:353-364. Bastos AM, Vezoli J, Fries P. 2015. Communication through coherence with inter-areal delays. Current opinion in neurobiology 31:173-180. Breiman L. 2001. Random forests. Machine learning 45:5-32. Brookes MJ, Tewarie PK, Hunt BA, Robson SE, Gascoyne LE, Liddle EB, Liddle PF, Morris PG. 2016. A multi-layer network approach to MEG connectivity analysis. NeuroImage 132:425-438. Brookes MJ, Woolrich M, Luckhoo H, Price D, Hale JR, Stephenson MC, Barnes GR, Smith SM, Morris PG. 2011. Investigating the electrophysiological basis of resting state networks using magnetoencephalography. Proceedings of the National Academy of Sciences 108:16783-16788. Broyd SJ, Demanuele C, Debener S, Helps SK, James CJ, Sonuga-Barke EJ. 2009. Default-mode brain dysfunction in mental disorders: a systematic review. Neurosci Biobehav Rev 33:279-296. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 36 of 41 Bullmore E, Sporns O. 2009. Complex brain networks: graph theoretical analysis of structural and functional systems. Nature reviews Neuroscience 10:186-198. Bullmore E, Sporns O. 2012. The economy of brain network organization. Nature reviews Neuroscience 13:336-349. Buschman TJ, Miller EK. 2007. Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. Science 315:1860-1862. Buschman TJ, Miller EK. 2014. Goal-direction and top-down control. Philosophical transactions of the Royal Society of London Series B, Biological sciences 369. Buzsáki G. 2006. Rhythms of the Brain: Oxford University Press. Cleveland WS, Loader C. 1996. Smoothing by local regression: Principles and methods. In. Statistical theory and computational aspects of smoothing Springer p 10-49. Colclough G, Woolrich M, Tewarie P, Brookes M, Quinn A, Smith S. 2016. How reliable are MEG resting-state connectivity metrics? NeuroImage 138:284-293. Dale AM, Fischl B, Sereno MI. 1999. Cortical surface-based analysis. I. Segmentation and surface reconstruction. NeuroImage 9:179-194. Dale AM, Liu AK, Fischl BR, Buckner RL, Belliveau JW, Lewine JD, Halgren E. 2000. Dynamic statistical parametric mapping: combining fMRI and MEG for high-resolution imaging of cortical activity. Neuron 26:55-67. de Pasquale F, Della Penna S, Snyder AZ, Lewis C, Mantini D, Marzetti L, Belardinelli P, Ciancetta L, Pizzella V, Romani GL, Corbetta M. 2010. Temporal dynamics of spontaneous MEG activity in brain networks. Proceedings of the National Academy of Sciences 107:6040-6045. Dosenbach NU, Nardos B, Cohen AL, Fair DA, Power JD, Church JA, Nelson SM, Wig GS, Vogel AC, Lessov-Schlaggar CN, Barnes KA, Dubis JW, Feczko E, Coalson RS, Pruett JR, Jr., Barch DM, Petersen SE, Schlaggar BL. 2010. Prediction of individual brain maturity using fMRI. Science 329:1358- 1361. Fair DA, Cohen AL, Dosenbach NU, Church JA, Miezin FM, Barch DM, Raichle ME, Petersen SE, Schlaggar BL. 2008. The maturing architecture of the brain's default network. Proceedings of the National Academy of Sciences 105:4028-4032. Fair DA, Cohen AL, Power JD, Dosenbach NU, Church JA, Miezin FM, Schlaggar BL, Petersen SE. 2009. Functional brain networks develop from a "local to distributed" organization. PLoS Comput Biol 5:e1000381. Fair DA, Dosenbach NU, Church JA, Cohen AL, Brahmbhatt S, Miezin FM, Barch DM, Raichle ME, Petersen SE, Schlaggar BL. 2007. Development of distinct control networks through segregation and integration. Proc Natl Acad Sci U S A 104:13507-13512. Fischl B, Sereno MI, Dale AM. 1999. Cortical surface-based analysis. II: Inflation, flattening, and a surface-based coordinate system. NeuroImage 9:195-207. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 37 of 41 Fischl B, van der Kouwe A, Destrieux C, Halgren E, Segonne F, Salat DH, Busa E, Seidman LJ, Goldstein J, Kennedy D, Caviness V, Makris N, Rosen B, Dale AM. 2004. Automatically parcellating the human cerebral cortex. Cereb Cortex 14:11-22. Fransson P, Aden U, Blennow M, Lagercrantz H. 2011. The functional architecture of the infant brain as revealed by resting-state fMRI. Cereb Cortex 21:145-154. Fries P. 2005. A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. Trends Cogn Sci 9:474-480. Fries P. 2015. Rhythms for Cognition: Communication through Coherence. Neuron 88:220-235. Gao W, Gilmore JH, Giovanello KS, Smith JK, Shen D, Zhu H, Lin W. 2011. Temporal and spatial evolution of brain network topology during the first two years of life. PloS one 6:e25278. Giedd JN, Blumenthal J, Jeffries NO, Castellanos FX, Liu H, Zijdenbos A, Paus T, Evans AC, Rapoport JL. 1999. Brain development during childhood and adolescence: a longitudinal MRI study. Nature neuroscience 2:861-863. Gramfort A, Luessi M, Larson E, Engemann DA, Strohmeier D, Brodbeck C, Parkkonen L, Hamalainen MS. 2014. MNE software for processing MEG and EEG data. NeuroImage 86:446-460. Grayson DS, Ray S, Carpenter S, Iyer S, Dias TG, Stevens C, Nigg JT, Fair DA. 2014. Structural and functional rich club organization of the brain in children and adults. PloS one 9:e88297. Groppe DM, Urbach TP, Kutas M. 2011. Mass univariate analysis of event-related brain potentials/fields I: a critical tutorial review. Psychophysiology 48:1711-1725. Hämäläinen MS, Sarvas J. 1989. Realistic conductivity geometry model of the human head for interpretation of neuromagnetic data. IEEE Trans Biomed Eng BME-36:165-171. Harris AZ, Gordon JA. 2015. Long-Range Neural Synchrony in Behavior. Annual review of neuroscience 38:171-194. Hauk O, Wakeman DG, Henson R. 2011. Comparison of noise-normalized minimum norm estimates for MEG analysis using multiple resolution metrics. NeuroImage 54:1966-1974. Hinkley LB, Vinogradov S, Guggisberg AG, Fisher M, Findlay AM, Nagarajan SS. 2011. Clinical symptoms and alpha band resting-state functional connectivity imaging in patients with schizophrenia: implications for novel approaches to treatment. Biol Psychiatry 70:1134-1142. Hipp JF, Hawellek DJ, Corbetta M, Siegel M, Engel AK. 2012. Large-scale cortical correlation structure of spontaneous oscillatory activity. Nature neuroscience 15:884-890. Hipp JF, Siegel M. 2015. BOLD fMRI Correlation Reflects Frequency-Specific Neuronal Correlation. Current biology : CB 25:1368-1374. Hwang K, Hallquist MN, Luna B. 2013. The development of hub architecture in the human functional brain network. Cereb Cortex 23:2380-2393. Jensen O, Goel P, Kopell N, Pohja M, Hari R, Ermentrout B. 2005. On the human sensorimotor-cortex beta rhythm: sources and modeling. NeuroImage 26:347-355. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 38 of 41 Jin SH, Seol J, Kim JS, Chung CK. 2011. How reliable are the functional connectivity networks of MEG in resting states? J Neurophysiol 106:2888-2895. Khan S, Cohen D. 2013. Note: Magnetic noise from the inner wall of a magnetically shielded room. The Review of scientific instruments 84:056101. Kim DJ, Bolbecker AR, Howell J, Rass O, Sporns O, Hetrick WP, Breier A, O'Donnell BF. 2013. Disturbed resting state EEG synchronization in bipolar disorder: A graph-theoretic analysis. NeuroImage Clinical 2:414-423. Kim JS, Shin KS, Jung WH, Kim SN, Kwon JS, Chung CK. 2014. Power spectral aspects of the default mode network in schizophrenia: an MEG study. BMC neuroscience 15:104. Kitzbichler MG, Khan S, Ganesan S, Vangel MG, Herbert MR, Hämäläinen MS, Kenet T. 2015. Altered development and multifaceted band-specific abnormalities of resting state networks in autism. Biological psychiatry 77:794-804. Lenroot RK, Giedd JN. 2006. Brain development in children and adolescents: insights from anatomical magnetic resonance imaging. Neuroscience & Biobehavioral Reviews 30:718-729. Liaw A, Wiener M. 2002. Classification and regression by randomForest. R news 2:18-22. Lin FH, Belliveau JW, Dale AM, Hämäläinen MS. 2006. Distributed current estimates using cortical orientation constraints. Hum Brain Mapp 27:1-13. Lin FH, Witzel T, Ahlfors SP, Stufflebeam SM, Belliveau JW, Hämäläinen MS. 2006. Assessing and improving the spatial accuracy in MEG source localization by depth-weighted minimum-norm estimates. NeuroImage 31:160-171. Lo CY, Su TW, Huang CC, Hung CC, Chen WL, Lan TH, Lin CP, Bullmore ET. 2015. Randomization and resilience of brain functional networks as systems-level endophenotypes of schizophrenia. Proc Natl Acad Sci U S A 112:9123-9128. Luna B, Marek S, Larsen B, Tervo-Clemmens B, Chahal R. 2015. An Integrative Model of the Maturation of Cognitive Control. Annual review of neuroscience 38:151-170. Maxwell SE, Delaney HD. 1990. Designing experiments and analyzing data: A model comparison approach. Belmont, CA: Wadsworth. Menon V. 2013. Developmental pathways to functional brain networks: emerging principles. Trends Cogn Sci 17:627-640. Miraglia F, Vecchio F, Bramanti P, Rossini PM. 2016. EEG characteristics in "eyes-open" versus "eyes- closed" conditions: Small-world network architecture in healthy aging and age-related brain degeneration. Clin Neurophysiol 127:1261-1268. Misic B, Betzel RF, de Reus MA, van den Heuvel MP, Berman MG, McIntosh AR, Sporns O. 2016. Network-Level Structure-Function Relationships in Human Neocortex. Cereb Cortex 26:3285-3296. Niso G, Rogers C, Moreau JT, Chen L-Y, Madjar C, Das S, Bock E, Tadel F, Evans A, Jolicoeur P. 2015. OMEGA: The Open MEG Archive. NeuroImage. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 39 of 41 Nolte G, Hämäläinen MS. 2001. Partial signal space projection for artefact removal in MEG measurements: a theoretical analysis. Phys Med Biol 46:2873-2887. Ordaz SJ, Foran W, Velanova K, Luna B. 2013. Longitudinal growth curves of brain function underlying inhibitory control through adolescence. J Neurosci 33:18109-18124. Palva JM, Wang SH, Palva S, Zhigalov A, Monto S, Brookes MJ, Schoffelen J-M, Jerbi K. 2017. Ghost interactions in MEG/EEG source space: A note of caution on inter-areal coupling measures. bioRxiv. Peng G-S, Tan S-Y, Wu J, Holme P. 2016. Trade-offs between robustness and small-world effect in complex networks. Scientific Reports 6. Pernet CR, Wilcox R, Rousselet GA. 2013. Robust Correlation Analyses: False Positive and Power Validation Using a New Open Source Matlab Toolbox. Front Psychol 3. Peters BD, Ikuta T, DeRosse P, John M, Burdick KE, Gruner P, Prendergast DM, Szeszko PR, Malhotra AK. 2014. Age-related differences in white matter tract microstructure are associated with cognitive performance from childhood to adulthood. Biol Psychiatry 75:248-256. Raichle ME. 2015. The Brain's Default Mode Network. Annual review of neuroscience 38:433-447. Raichle ME, MacLeod AM, Snyder AZ, Powers WJ, Gusnard DA, Shulman GL. 2001. A default mode of brain function. Proc Natl Acad Sci U S A 98:676-682. Ronnqvist KC, McAllister CJ, Woodhall GL, Stanford IM, Hall SD. 2013. A multimodal perspective on the composition of cortical oscillations. Frontiers in human neuroscience 7:132. Roux F, Uhlhaas PJ. 2014. Working memory and neural oscillations: alpha-gamma versus theta- gamma codes for distinct WM information? Trends Cogn Sci 18:16-25. Rubinov M, Sporns O. 2010. Complex network measures of brain connectivity: Uses and interpretations. NeuroImage 52:1059-1069. Sato JR, Salum GA, Gadelha A, Picon FA, Pan PM, Vieira G, Zugman A, Hoexter MQ, Anes M, Moura LM, Gomes Del'Aquilla MA, Amaro E, Jr., McGuire P, Crossley N, Lacerda A, Rohde LA, Miguel EC, Bressan RA, Jackowski AP. 2014. Age effects on the default mode and control networks in typically developing children. Journal of psychiatric research 58:89-95. Sato JR, Salum GA, Gadelha A, Vieira G, Zugman A, Picon FA, Pan PM, Hoexter MQ, Anes M, Moura LM, Del'Aquilla MA, Crossley N, Amaro Junior E, McGuire P, Lacerda AL, Rohde LA, Miguel EC, Jackowski AP, Bressan RA. 2015. Decreased centrality of subcortical regions during the transition to adolescence: a functional connectivity study. NeuroImage 104:44-51. Schäfer CB, Morgan BR, Ye AX, Taylor MJ, Doesburg SM. 2014. Oscillations, networks, and their development: MEG connectivity changes with age. Human brain mapping 35:5249-5261. Sekihara K, Owen JP, Trisno S, Nagarajan SS. 2011. Removal of spurious coherence in MEG source- space coherence analysis. IEEE Trans Biomed Eng 58:3121-3129. Siegel M, Donner TH, Engel AK. 2012. Spectral fingerprints of large-scale neuronal interactions. Nature reviews Neuroscience 13:121-134. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 40 of 41 Sohal VS, Zhang F, Yizhar O, Deisseroth K. 2009. Parvalbumin neurons and gamma rhythms enhance cortical circuit performance. Nature 459:698-702. Spielberg JM, Miller GA, Heller W, Banich MT. 2015. Flexible brain network reconfiguration supporting inhibitory control. Proc Natl Acad Sci U S A 112:10020-10025. Tagliazucchi E, Laufs H. 2014. Decoding wakefulness levels from typical fMRI resting-state data reveals reliable drifts between wakefulness and sleep. Neuron 82:695-708. Takada N, Pi HJ, Sousa VH, Waters J, Fishell G, Kepecs A, Osten P. 2014. A developmental cell-type switch in cortical interneurons leads to a selective defect in cortical oscillations. Nature communications 5:5333. Tan B, Kong X, Yang P, Jin Z, Li L. 2013. The difference of brain functional connectivity between eyes- closed and eyes-open using graph theoretical analysis. Comput Math Methods Med 2013:976365. Taulu S, Kajola M, Simola J. 2004. Suppression of interference and artifacts by the Signal Space Separation Method. Brain Topogr 16:269-275. Taulu S, Simola J. 2006. Spatiotemporal signal space separation method for rejecting nearby interference in MEG measurements. Phys Med Biol 51:1759-1768. Thomson DJ. 1982. Spectrum estimation and harmonic analysis. Proceedings of the IEEE 70:1055- 1096. Toussaint PJ, Maiz S, Coynel D, Doyon J, Messe A, de Souza LC, Sarazin M, Perlbarg V, Habert MO, Benali H. 2014. Characteristics of the default mode functional connectivity in normal ageing and Alzheimer's disease using resting state fMRI with a combined approach of entropy-based and graph theoretical measurements. NeuroImage 101:778-786. Uhlhaas PJ, Haenschel C, Nikolic D, Singer W. 2008. The role of oscillations and synchrony in cortical networks and their putative relevance for the pathophysiology of schizophrenia. Schizophr Bull 34:927- 943. Uhlhaas PJ, Roux F, Rodriguez E, Rotarska-Jagiela A, Singer W. 2010. Neural synchrony and the development of cortical networks. Trends Cogn Sci. Uhlhaas PJ, Roux F, Singer W, Haenschel C, Sireteanu R, Rodriguez E. 2009. The development of neural synchrony reflects late maturation and restructuring of functional networks in humans. Proc Natl Acad Sci U S A 106:9866-9871. Vidal JR, Freyermuth S, Jerbi K, Hamamé CM, Ossandon T, Bertrand O, Minotti L, Kahane P, Berthoz A, Lachaux J-P. 2012. Long-distance amplitude correlations in the high gamma band reveal segregation and integration within the reading network. Journal of Neuroscience 32:6421-6434. Vidaurre C, Sander TH, Schlogl A. 2011. BioSig: the free and open source software library for biomedical signal processing. Comput Intell Neurosci 2011:935364. Wang L, Saalmann YB, Pinsk MA, Arcaro MJ, Kastner S. 2012. Electrophysiological low-frequency coherence and cross-frequency coupling contribute to BOLD connectivity. Neuron 76:1010-1020. Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 41 of 41 Wang XJ. 2010. Neurophysiological and computational principles of cortical rhythms in cognition. Physiological reviews 90:1195-1268. Watts DJ, Strogatz SH. 1998. Collective dynamics of 'small-world'networks. nature 393:440-442. Webel K. 2006. K. Takezawa: Introduction to Nonparametric Regression. Allgemeines Statistisches Archiv 90:625-626. Yu Q, Wu L, Bridwell DA, Erhardt EB, Du Y, He H, Chen J, Liu P, Sui J, Pearlson G, Calhoun VD. 2016. Building an EEG-fMRI Multi-Modal Brain Graph: A Concurrent EEG-fMRI Study. Frontiers in human neuroscience 10:476. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Supplementary Materials: Page 1 of 1 Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band S. Khan1,4,5‡*, J. A. Hashmi1,4‡, F. Mamashli1,4, K. Michmizos1,4, M. G. Kitzbichler1,4, H. Bharadwaj1,4, Y. Bekhti1,4, S. Ganesan1,4, K. A Garel1, 4, S. Whitfield-Gabrieli5, R. L. Gollub2,4, J. Kong2,4, L. M. Vaina4,6, K. D. Rana6, S. S. Stufflebeam3,4, M. S. Hämäläinen3,4, T. Kenet1,4 1Department of Neurology, MGH, Harvard Medical School, Boston, USA. 2Department of Psychiatry MGH, Harvard Medical School, Boston, USA. 3Department of Radiology, MGH, Harvard Medical School, Boston, USA. 4Athinoula A. Martinos Center for Biomedical Imaging, MGH/HST, Charlestown, USA 5McGovern Institute for Brain Research, Massachusetts Institute of Technology, Cambridge, USA 6Department of Biomedical Engineering, Boston University, Boston, USA ‡equal contribution Running Title: Multifaceted Development of Brain Connectivity with Age *Corresponding author: Sheraz Khan, Ph.D. Athinoula A. Martinos Center for Biomedical Imaging Massachusetts General Hospital Harvard Medical School Massachusetts Institute of Technology 149 13th Street, CNY-2275 Boston, MA-02129, USA Phone: +1 617-643-5634 Fax: +1 617-948-5966 E-mail: [email protected] SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 2 of 2 Table S1: Corrected (see methods) correlation p-values over the range of thresholds used, for each frequency band. Colormap codes for significant p-values. Cohen's d for correlations effect sizes are shown in blue. (A) P-values for local efficiency, per band, per threshold value. The p-value in the blue rectangle corresponds to Figure 2A. (B) P-values for global efficiency, per band, per threshold value. The p-value in the blue rectangle corresponds to Figure 2B. (C) P-values for small world property, per band, per threshold value. The p-values in the blue rectangles correspond to Figure 4A (beta band) and 4C (gamma band). All graph derived figures throughout the manuscript and SI are shown at 15% thresholding (cost). P-Values were corrected for multiple comparisons (Frequency bands, thresholds, graph metrics) by controlling for family-wise error rate using maximum statistics through permutation testing (see also section "Correction of correlation p-values for multiple comparisons", and Figure S8 for the null distribution). SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 3 of 3 Figure S1: Subject distribution for (A) our data (MGH), and (B) the OMEGA data SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 4 of 4 Figure S2: Correlation between age and Composite IQ (r = 0.18, p=0.16). Linear regression line (solid black line) for the relationship between age and IQ. The individual data points are represented using a scatter plot. There was no significant relationship between age and IQ, as expected. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 5 of 5 Figure S3: Correlation between age and norm of the lead field for gradiometers. Thresholded p-value for this correlation are shown as textured colormap on the cortical manifold. Magenta line outlines the only areas with significant correlation, at p < 0.05. These areas did not overlap with any of the identified hubs in the study, meaning age effects due to head size were likely negligible in the presented analyses. Note that no significant correlation with age was observed in the magnetometer data. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 6 of 6 Figure S4: Parcellation Scheme on the cortical surface showing 448 labels. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 7 of 7 Figure S5: Correlation between age and mean cross-talk for 448 Labels (p = 0.36). A. LOESS plot (solid white line) for the relationship between age and mean cross-talk between labels. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 8 of 8 Figure S6: Power Spectral Density in different cortical areas (source space data). The PSD plots are averaged across all participants in each of the three age groups noted above. Legend shows cortical area from which data was derived (source space). The lines show the mean PSD within that age group and cortical region, averaged over the entire region. Standard error is shown as shaded area. Empty room data mapped onto the cortical space is shown as comparison for estimated SNR. (A) Ages 7 – 13. (B) Ages 14 – 21. (C) Ages 22 – 29. Red: Occipital Lobe; Green: Parietal Lobe; Blue: Temporal Lobe; Blue: Temporal Lobe; Magenta: Frontal Lobe, Empty Room: Mapped onto the cortical Manifold. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 9 of 9 Figure S7: Example of the envelope approach applied to gamma frequency band (31Hz- 80Hz): Envelopes of gamma power in two cortical regions (left and right lateral orbitofrontal cortex) in one participant, marked in red (right) and blue (left). The peak frequency of the two independently derived envelopes is around 0.125Hz. The correlation between two regions is then measured via the correlation in these envelopes, not the carrier frequency bands. This is also illustrated in the fourth panel (fourth step) in Figure 1. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 10 of 10 Figure S8: Degree of each node for age group 22-29, for each frequency band. The textured colormap on cortical manifold shows mean degree across subjects in age group 22-29 for each label. There maps look similar to prior studies (Hipp et al. 2012). Note that the original Hipp et al study was done in cortical volume space and then projected onto the surface, whereas our study directly computes all the results on the surface. Both approaches show sensible networks. Please see methods for additional details. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 11 of 11 Figure S9: Average connectivity increases with age in the alpha and beta band mediated networks. A. LOESS plot (solid white line) for the relationship between mean connectivity (unthresholded weighted degree, also known as node strength) and age for alpha band mediated networks. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure. B. Same, for the beta band mediated networks. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 12 of 12 Figure S10: Weighted Networks - local efficiency increases in beta band networks and global efficiency increases in gamma band networks. A. LOESS plot (solid white line) for the relationship between age and weighted local network efficiency of beta band mediated networks. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure. B. Same, for gamma band mediated networks, and weighted global efficiency. The results show the same trend as unweighted networks, shown in Figure 2. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 13 of 13 Figure S11: Calculating Corrected Correlation p-value. The null distribution (1000,000 Permutations) was computed as described in the method section, and is shown. For each actual correlation value in the data shown as magenta line, we compute the corrected p-value by comparing it against the empirical null distribution. SUPPLEMENT: Maturation Trajectories of Cortical Resting-State Networks Depend on the Mediating Frequency Band Page 14 of 14 Figure S12: Local and global network efficiencies did not show age dependent changes in the remaining cases. A. LOESS plot (solid white line) for the relationship between age and local network efficiency of Delta band mediated networks. The individual data points are represented using a normalized density colormap, where each data point corresponds to one realization of the bootstrap procedure. B. Same, for delta band mediated networks, and global efficiency. C. Same, for Theta band mediated networks, and local efficiency. D. Same, for theta band mediated networks, and global efficiency. E. Same, for alpha band mediated networks, and local efficiency. F. Same, for alpha band mediated networks, and global efficiency. G. Same, for gamma band mediated networks, and local efficiency. H. Same, for beta band mediated networks, and global efficiency. Threshold=0.15%, please see Table S1 for p-values.
1904.10498
1
1904
2019-04-17T15:05:40
Information Based Centralization of Locomotion in Animals and Robots
[ "q-bio.NC", "physics.bio-ph" ]
Movement in biology is often achieved with distributed control of coupled subcomponents, e.g. muscles and limbs. Coupling could range from weak and local, i.e. decentralized, to strong and global, i.e. centralized. We developed a model-free measure of centralization that compares information shared between control signals and both global and local states. A second measure, called co-information, quantifies the net redundant information the control signal shares with both states. We first validate our measures through simulations of coupled oscillators and show that it successfully reconstructs the shift from low to high coupling strengths. We then measure centralization in freely running cockroaches. Surprisingly, extensor muscle activity in the middle leg is more informative of movements of all legs combined than the movements of that particular leg. Cockroach centralization successfully recapitulates a specific model of a strongly coupled oscillator network previously used to model cockroach leg kinematics. When segregated by stride frequency, slower cockroach strides exhibit more shared information per stride between control and output states than faster strides, indicative of an information bandwidth limitation. However, centralization remains consistent between the two groups. We then used a robotic model to show that centralization can be affected by mechanical coupling independent of neural coupling. The mechanically coupled bounding gait is decentralized and becomes more decentralized as mechanical coupling decreases while internal parameters of control remain constant. The results of these systems span a design space of centralization and co-information that can be used to test biological hypotheses and advise the design of robotic control.
q-bio.NC
q-bio
Information Based Centralization of Locomotion in Animals and Robots Izaak D. Nevelna,1, Amoolya Tirumalaia, and Simon Sponberga,b aSchool of Physics, Georgia Institute of Technology; bSchool of Biology, Georgia Institute of Technology This manuscript was compiled on April 17, 2019 Movement in biology is often achieved with distributed control of coupled subcomponents, e.g. muscles and limbs. Coupling could range from weak and local, i.e. decentralized, to strong and global, i.e. centralized. We developed a model-free measure of central- ization that compares information shared between control signals and both global and local states. A second measure, called co- information, quantifies the net redundant information the control sig- nal shares with both states. We first validate our measures through simulations of coupled oscillators and show that it successfully re- constructs the shift from low to high coupling strengths. We then measure centralization in freely running cockroaches. Surprisingly, extensor muscle activity in the middle leg is more informative of movements of all legs combined than the movements of that partic- ular leg. Cockroach centralization successfully recapitulates a spe- cific model of a strongly coupled oscillator network previously used to model cockroach leg kinematics. When segregated by stride fre- quency, slower cockroach strides exhibit more shared information per stride between control and output states than faster strides, in- dicative of an information bandwidth limitation. However, centraliza- tion remains consistent between the two groups. We then used a robotic model to show that centralization can be affected by mechan- ical coupling independent of neural coupling. The mechanically cou- pled bounding gait is decentralized and becomes more decentralized as mechanical coupling decreases while internal parameters of con- trol remain constant. The results of these systems span a design space of centralization and co-information that can be used to test biological hypotheses and advise the design of robotic control. Motor Control Information Theory Locomotion Animal locomotion, the task of actively moving from one position and orientation to another, is achieved through complex dynamics where control is typically distributed across many actuators. For effective locomotion, coordination of muscles and limbs in space and time is necessary to pro- duce directed forces. Locomotor coordination could either be achieved through strong, global coupling with dense connec- tions between components, or though weak, local coupling with sparse connections (1). The continuum between these extreme coupling paradigms is thought of as the centraliza- tion/decentralization axis of locomotor control, though a quan- tifiable measure that can be applied across various systems is lacking. For example, Brambilla et. al. defines a decentral- ized robotic swarm to consist of autonomous individuals that communicate locally and receive no global information (2). Cruse et. al. define stick insect motor control as decentralized because muscle commands are more modulated by peripheral feedback rather than the central nervous system (3). However, strong mechanical coupling and feedback from global states could also result in a highly centralized control architecture (4). Methods to assess the empirical centralization of locomotor systems, preferably without any assumption of an underlying dynamic model, are necessary to answer questions regarding how a multi-actuated system is coupled through mechanics, feedback, and control as shown in Fig. 1. One unresolved locomotor hypothesis is that for fast move- ments, control via sensory feedback is less effective due to inherent time delays (5). This hypothesis predicts a reliance more on fast decentralized mechanical and neural responses local to each leg where global information decreases with speed faster than local information (4). While there is some evidence that neural feedback is too slow to effectively coordinate con- trol for fast locomotion from experiments in cockroaches (6) and flies (7), some examples of fast local feedback exists (8, 9). An alternative hypothesis is that internal feedforward con- trol may need to be highly centralized to maintain dynamic stability at high speeds (10). There is some evidence that overall coupling increases with speed (11) and that precision in timing of leg movements is coordinated through internal neural coupling(7, 10). However, the challenges of measuring centralization in a system, especially without a specific mod- eling framework, leaves the general questions regarding the varying degree of centralization in control of animal movement largely unresolved. Many potential model-based measures exist for quantifying the centralization of systems. Given a full network model, node centrality can indicate which points in the network most govern information flow (12). Distributions of node central- ity over networks also indicate overall network architecture (13). For locomotion, we are more interested in a centraliza- tion measure that encapsulates the dynamics of how networks coordinate. Coupled-oscillator network models can exhibit co- ordinated or synchronized behavior similar to the coordination of neural networks or the mechanics of limbs in animals (14). The Kuramoto model of many globally coupled oscillators has been well characterized (15) where oscillators transition from endless incoherence to fast synchronization as coupling increases and global influences outweigh local influences (16). The ability to reduce a large network of oscillators to low dimensional coupled phases or a single global phase described by an order parameter while still capturing dynamics of the system, especially under perturbations, could also indicate a highly centralized architecture (17). Coupled oscillators have been used to represent networks of central-pattern genera- tor (CPG) circuits that drive leg movements (14, 18). These coupled-oscillator models have been used to estimate coupling strengths between control of legs in animal systems (11) as I.D.N. and S.S. conceived research. I.D.N. and A.T. performed experiments. I.D.N. and S.S. wrote the manuscript. The authors declare no conflicts of interest. 1To whom correspondence should be addressed. E-mail: ineveln2gmail.com April 17, 2019 1 -- 7 Fig. 1. A) Complex dynamical systems can use numerous possible control signals which cascade information through successive levels of integration to effect a relatively small number of global task related outcomes. Coupling between control can range from weak and local (decentralized) to strong and global (centralized).B) Our centralization measure uses empirical observations of control signals, local states, and global states to infer the coupling of control. We estimate the mutual information between the control signal and both the global and local states. Centralization is IG minus IL, which removes any ICO between the three variables. We expect decentralized systems to contain more IL and centralized systems to contain more IG. well as controllers for robotic systems (19). While increased coupling of CPGs should result in increased centralization, so should increased mechanical coupling and feedback, and all such forms of coupling should be captured in a centralization measure (Fig. 1). Systems can be coordinated solely through mechanical coupling, such as a passive walker (20). Mechanical coupling can also affect feedback circuits that detect changes in force to one leg due to the lift-off of others as has been investigated in stick insects (21). Interlimb coordination, including energy-efficient gait transitions, can be achieved in a quadruped robot solely through local force sensing without any other communication between the leg con- trollers (22). Force changes can affect the ability to coordinate locomotion as seen in flies (23). Our measure of centraliza- tion should reflect how shifts in mechanics can change overall coupling whether through changes in the passive dynamics or feedback circuits that depend on mechanical signals. Quantifying all of the concepts of centralization so far de- scribed rely on a model of the system. Here we develop an empirical measure that can be used to compare the relative centralization of control across different systems and condi- tions. What unifies concepts of centralization is the amount of global information a control signal shares about the state of the system compared to the amount of local information. We use an information theoretic approach which can assess the dependencies among types and contexts of various locomotor signals. This approach also allows us to separately measure how much net information the control signal shares with both local and global states in a quantity called co-information. We validate these measures of centralization and co-information using a coupled oscillator network of locomotion to ensure that it can reconstruct changes in a model where centralization has been previously defined as the coupling strength. We then analyze the centralization of the coupled oscillator model as it relates to cockroach locomotion and apply our measure to test the hypothesis that cockroach control becomes more centralized at faster running speeds. We also examine how a mechanically coordinated robot is decentralized under varying inertial loadings in order to test if the measure can detect shifts in mechanical coupling. We discuss how these various systems map onto an information space containing centralization and co-information that can be used as a tool for comparing bio- logical control strategies as well as designing robotic control strategies. An Information Theoretic Measure of Centralization Locomotor control is spread out among many control signals that affect local subsystems (Fig, 1A). In the cockroach, con- trol of the muscles in the legs drive the movement of that leg locally. The local states of these subsystems, such as the extension of one leg, each contribute to produce global states, such as the average of all leg extensions. The global states drive the system toward reaching the overall goals, such as how the cockroach's six legs together move its body through the environment. This cascade of signals could consist of multiple layers (e.g. the muscles, joints, legs, etc. of the cockroach) and include feedback from any of those layers (for an investigation into different feedback architectures, see (17)). Furthermore, signals could correlate through coupling between subsystems within a layer, such as mechanical coupling between cockroach legs through the body and ground. By measuring the informa- tion shared between a chosen set of control signal, local state, and global state, we aim to infer the structure of the control architecture. Shared information between variables is quantified by mu- tual information I, which can be depicted as the amount of overlapping entropy as shown in Fig. 1B. A background on information theory is presented in Supplementary Information. An intuitive decomposition of the mutual information be- tween the control signal and the joint local and global states (depicted by the combination of the red, blue and gray areas in Fig. 1B and Fig. S1) into four separate positive values is given by I(C;{L, G}) = IU L + IU G + IR + ISY N , [1] where IU L and IU G represents information shared uniquely between the control signal and the local and global states respectively, IR is redundant information shared when either of local or global states are known, and ISY N is synergistic information shared only when both states are known. The axioms that allow for such a decomposition are debated, and estimating these quantities becomes challenging (24). We 2 Neveln et al. potential system architecturesdecentralizedcentralized. . . . . . . . . . . . local statesglobal statescontrol signalslocalfeedbackintegrationsubsysinternal control. . . . . . integrationsubsysinternal controlH(C)control signal entropyH(L)local stateentropyH(G)IUL+ISYNIGICENT = IG - ILICO= IR - ISYNIL IUG+ISYNglobal stateentropyABglobalfeedback Fig. 2. A) We simulated this six phase oscillator network that has been used to model the alternating tripod gait of cockroaches. B) The control signal was a perturbation to a single oscillator with random amplitude. The local state was the integrated absolute deviation from baseline of the phase velocity of the perturbed oscillator, shown by the shaded red region. The global state was similarly calcu- lated from the average phase velocity. C) IL decreased rapidly with increased coupling strength while IG remained constant, resulting in a monotonic increase from negative ICEN T to positive ICEN T . ICO coincides with IG at K = 0 and coincides with IL at high K. where fi is the natural frequency of each oscillator (set to 10 Hz is all cases), aij is 1 if there exists a connection between oscillator i and j and is zero otherwise, φij is the preferred phase difference between oscillator i and j, νi is additive gaussian noise (0 Hz mean, 0.03 Hz standard deviation), and K is the coupling strength between oscillators. We integrate Eqn. 4 using the Euler-Maruyama method (25). We want to characterize how the information present in a perturbation to an oscillator is spread throughout the network. We prescribe a square pulse Pi lasting one half cycle put into just one oscillator as shown in Fig. 2B, which we use as the control signal C, with an amplitude drawn from a random Gaussian distribution (−5 Hz mean, 4 3 Hz standard deviation). We then measure both the local response of that oscillator (the integrated deviation away from the steady state phase velocity) to that input and the average global response (same as local only all phase velocities are averaged together) of all oscillators as shown in Fig. 2B. As shown in Fig. 2C, the system is fully decentralized when K = 0, meaning IL outweighs IG. ICO matches IG indicating that any IG is redundant with local information resulting in no IU G. As the perturbation cannot propagate to the other oscillators, no additional information can be present in the global signal. As coupling is introduced and increases, ICEN T increases, becomes positive, and levels out to a maximal value. At these high coupling strengths the IL is completely redundant, meaning the value for ICEN T equals the amount of IU G. Thus, though IG stays constant with increased coupling strength, IU G must increase from zero to a positive value as a positive value of ICEN T requires there exist IU G. Changes to coupling strength can manifest in physical oscillators through changing the mass of a freely moving platform that holds a number of metronomes (26) or increasing the number of connections between central pattern generating circuits driving locomotion (27). Our centralization measure could empirically determine the relative coupling strengths of these systems, validating it as a useful diagnostic tool. Cockroach Centralization During Running. We ran 9 cock- roaches over flat terrain while recording EMG activity from the femoral extensor muscle 137 in the middle leg and tracking the 2D kinematics of the ends of all 6 legs as shown in Fig. 3A. This muscle has previously been implicated in control even during high speed running (29). See Supplementary Meth- ods section Cockroach Experiments for more details on the experimental protocol. When analyzing 2343 strides from all cockroaches (Fig. 3B- F), we find that ICEN T is positive (Fig. 3G). Positive ICEN T means motor unit spikes are more informative about the global average kinematics than the local kinematics of the limb where April 17, 2019 3 avoid estimating these quantities directly, as we can compute differences of these quantities that are useful measures of the systems we study from our estimates of local MI IL, global MI IG, and total MI IT OT as shown in Fig 1B. For more discussion on the estimation methods used to calculate mutual information, see the Supplementary Material. We define our measure of centralization ICEN T to be ICEN T = IG − IL = IU G − IU L [2] and thus quantifies the balance between the amount of information about C that is uniquely global versus that which is uniquely local. Redundant or synergistic information does not contribute to centralization. ICO = IL + IG − IT OT = IR − ISY N , Co-information Ico, the gray region in Fig. 1B, is given by [3] and it is a measure of net redundancy and does not contain the unique information (24). A negative value indicates that syn- ergistic information outweighs redundant information. ICEN T and ICO are therefore two measures that look at how different parts of the total information balance and can both be poten- tially useful to discriminate different types of neuromechanical control systems. Grounding these measures back into the specific biologi- cal signals, a positive value of ICEN T indicates that control signal from the selected muscle is more informative about the global kinematic state averaged from all limbs than the local kinematic state of the leg in which the muscle resides. Also, positive ICEN T guarantees non-zero IU G, meaning there must be global information not present locally. Therefore, this global information would have to come from some source of coupling (mechanical or neural) within the system. A pos- itive value for ICO indicates some net redundancy between local and global information. As ICO increases, it become less important to know both local and global states to have information about the control signal. Results Centralization of a Coupled Oscillator Model Increases with Coupling Strength. We first test whether our measure of cen- tralization captures the previous model-specific definition of centralization based on the strength of a coupled oscillator network shown in Fig. 2A. Not only are coupled oscillator models used as a tool to understand locomotion (14), but this particular model has been previously used to estimate coupling parameters between the six legs used in insect locomotion (11). The dynamics of each oscillator θi is given by Kaij sin(θj − θi − φij) + 2πνi + 2πPi, [4] 6X j=1 θi = 2πfi + Neveln et al. perturbation00.10.2-0.1050100150200250300-2-10123ICENTILIGICOcoupling strength Kbits / cyclecontrol signal:perturbation (a.u.)phase trajectories (a.u.)global state: averaged phase velocity (a.u.)local state: perturbed phase velocity (a.u.)time (s)BAC the control signal originates. It is surprising that the activation of a muscle in one leg indicates more about the average state of all the limbs than that of the leg in directly activates. This main result is likely because of strong neural and mechanical coupling between the legs (11, 30). When just spike count is considered, IL slightly outweighs IG, though the proportion of overall information is small. Most information, and the positive value for ICEN T , arises only when spike timing is also factored into the analysis. Estimates of mutual information between the control signal and the local and global states were stabilized when just a two-dimensional representation of the states were used (see SI Fig. S2-3 and Supplementary Text). Recent studies show that the timing of individual motor unit spikes has causal effects on motor dynamics down to the millisecond scale(31, 32). Our results show that most of the dependencies between the control signal and the local and global states only manifest when the timing of the spike relative to the phase of the stride is considered. Many analyses of motor neuronal activity in insects use only the rate of activity (33 -- 35). It is possible that much of the encoded information regarding leg coupling is suppressed in such analyses. Muscle 137 (as well as its homologous muscle 178 in the hind leg) is driven by a single fast motor neuron that also drives other extensor muscles 136, 135d, and 135e (179, 177d, and 177e in the hind leg). (36). These muscles can produce varying mechanical work from the same signal (37), including positive work to drive extension or negative work to slow flexion. Therefore, this single motor unit has been implicated in the control of leg flexion and reversal (38) and the start of joint extension and stride length (29). Our results for the middle leg indicate the control signal shares non-redundant information with both the stance and swing portions of the stride, which both corroborates the reported versatility of this motor unit as well as the observation that muscle work depends on the state of the limb (39). The Effect of Speed on ICEN T and ICO. Given that the cock- roaches tested exhibited a wide range of stride frequencies (Fig. 3B), we can test whether faster strides were more central- ized possibly for maintaining dynamic stability (10) or more decentralized possibly due to bandwidth constraints (4). When we segment the cockroach data into two groups according to stride frequency, we observe ICEN T does not change (leftmost column of Fig. 4). However, both IG and IL do change in interesting ways. IG and IL per stride is higher for slower strides than for faster strides (Fig. 4A). This difference is due to a similar trend when looking at timing information. IG and IL when just spike count is considered is slightly lower in the slower group, though again count information contributes much less information overall. When converted to bits per second using the median frequency of each group (Fig. 4B), we actually see that the information per unit time (bit rate) is greater for the faster group. Though the balance of local and global information does not change, perhaps the two states become more redundant with greater speed. Overall, ICO is similar between fast and slow groups. However, the faster group closer to full redundancy as the IT OT is smaller. The slower group has higher ICO in timing and lower ICO in count. Timing is therefore more redundant for the slow group, whereas ICO is actually negative in count indicating some degree of synergy between local and global signals. Therefore, for the slow group, the local and global output together are more informative on the number of spikes in a stride than when taken separately. The amount of information available to be transferred from motor neuron to muscle to leg output and back again through Fig. 3. A) Control, local, and global signals recorded from the cockroach. Strides were separated according to kinematic phase calculated using the Phaser algorithm (28). B) Distribution of stride frequencies across all 2343 strides taken from 9 animals. C) Distribution of the number of spikes in the femoral extensor over a stride. D) Probability density functions of the timing of the first four spikes if present, with time normalized by stride period T . E) Local leg extension trajectories colored by the timing of the first spike (colormap from D). Correlations between the timing of the first spike and the resulting local and global states are visible in trajectories when the first spike occurred early in the stride colored blue are distinguished from those when the first spike occurred late in the cycle colored yellow. F) Global leg extension trajectories colored as in E. G) ICEN T , IG, IL, and ICO for all strides. 4 Neveln et al. 05101500.050.10.150.20.25stride frequency (Hz)probability0123456700.10.20.30.40.5spike countprobability00.050.10.150.20.25probabilitytime/T1st spike2nd spike3rd spike4th spike-10010-20-100102000.20.40.60.81extension (mm)extension (mm)spike countspike timingcount + timingbits / stride00.10.20.30.4time (s)EMGcontrol signal (a.u.)local legextension (a.u.)global average leg extension (a.u.)kinematic phaseto align strides (a.u.)BCDEFG-0.100.10.20.30.40.5ICENTILIGICOA Fig. 4. A) Fast versus slow strides. A) When splitting the strides into slow and fast halves, ICEN T remains unchanged. However, both local and global mutual information decrease equally, and this decrease is attributed to information in timing. ICO remains constant due to competing trends in timing and count. B) Same as (A) only information is converted to a bits/second rate using the median stride frequency of each group. Though the information per stride decreases, the information per unit time increases. feedback is expected to decrease for faster strides(6), which is what we see in our data. However, the decrease per stride is not as much as expected with the assumption of a constant information rate. The predictions for ICEN T are complicated because while this information decrease with speed is expected, internal and mechanical coupling is hypothesized to increase to maintain dynamic stability (40). Spatial coordination (7) and temporal coordination (34, 41) has been shown to increase with speed (7). Spatial coordination degrades when sensory feedback is disrupted (10). When fitting thoracic ganglia burst activity to coupled-oscillator models, no correlation is observed between burst frequency and coupling strength (42) and a very weak positive correlation between running speed and coupling strength was observed when fitting free-running cockroach leg kinematics to such a model (11). Our measure of centralization, which takes into account a neural control signal with local and global kinematics indicates no shifts in overall coupling with speed when considering cockroach running. Coordination Through Mechanical Coupling is Decentralized in Robot Bounding Gate. If neural feedback delays are too long to effectively couple limbs during fast locomotion to properly respond to perturbations, mechanical coupling could potentially compensate. Furthermore, changes to mechanical coupling will alter feedback signals related to the state of the system and its parts. Clearly, mechanics must be considered when analyzing the control architecture of dynamic locomo- tion. The Minitaur robot (Ghost Robotics, Inc. Philadelphia, PA) shown in Fig. 5A demonstrates coordination through mechanical coupling (43). As one pair of legs impacts the ground, the rest of the body translates and rotates, generally resulting in movement of the hips of the alternate leg pair. Therefore, even if the commanded torque to one leg pair does not explicitly depend on the states of the other leg pair (i.e. no internal 'neural' coupling), the two leg pairs will tend towards a bound gait where the front pair of synchronized legs alter- nates with the synchronized back pair, or a pronk gait where Fig. 5. A) Image of Minitaur indicating the variables measured for the centralization calculation. B) The commanded torque at the hip to drive the extension of a single leg was selected as the control signal. The local state was the measured extension of that leg. The global state was the average of all leg extension trajectories. Strides were aligned by the pitch angle of the robot. C) ICEN T and ICO of the three different moment of inertia conditions. The torque to a rear leg was used as the control variable. D) Torque to a front leg was used as a control variable. all legs are synchronized. Transitions between these gaits can occur by changing mechanical coupling through changes to the moment of inertia M around the pitch axis, or by adding phase coupling into the internal control block. We altered M by shifting two weights in opposite directions longitudinally along the robot, thus keeping the center of mass constant. These varying conditions test whether our empiri- cal centralization measure can detect changes to mechanical coupling. We also predict ICO will be positive and close to maximal, because any information transfer through mechan- ical coupling should be redundant if the system is relatively stiff. We ran the bound gait described in (43) over flat terrain, which still produces variability in each stride, and measured local mutual information between the torque command and the actual extension of that leg as shown in Fig. 5B. We compare the local information to the global mutual information between that same torque signal and the average extension trajectories of all four legs. As shown in Fig. 5, we compare low, intermediate, and high values of M. IL is greater than IG for all experimental conditions, result- ing in a negative value of ICEN T . For the more decentralized rear leg pair, ICEN T is minimized for the intermediate M con- dition, confirming the prediction for when mechanical coupling is minimized. ICEN T is greatest for the low M, where IG is fully redundant. For the front leg pair, ICEN T is minimized for both the intermediate and high M conditions. ICEN T of control for signals from the front legs is overall higher when considering the front leg versus the back leg, indi- cating an asymmetry to the mechanical coupling not predicted in the reduced models that only consider bounding in place. As this asymmetry is the same regardless of which of the front or back legs are analyzed, we expect that this difference is partly due the forward movement of the robot. This result is an example of how measuring ICEN T can result in new discoveries that may not be predicted from simplified models. Consideration of mechanics is necessary for understanding locomotor control (30). The virtual leg of running animals all use a similar non-dimensionalized stiffness that also optimizes Neveln et al. April 17, 2019 5 -0.100.10.20.3ICENTILIGICOfullhalvesStride Freq.stride frequency (Hz)bits / stridecounttimingtotal6810-10123681068106810bits / secondAB00.511.522.5lowmedhighICENTILIGICO-1.5-1-0.500.511.52lowmedhightime (s)bits / stride minitaur pitch moment of inertiacontrol signal:commanded torque(a.u.)local state:leg extenstion(a.u.)global state:averaged extension(a.u.)stride alignment:robot pitch angle (a.u.)rear legfront legABCD information. Fig. 6. A) Centralization and co-information for all systems normalized by total B) The centralization/co-information control architecture space. Any system will fall within a diamond region bounded by the solid black diagonal lines. Centralized systems, where a control signal is more informative about global states than local states, will fall into the blue region. Conversely, decentralized systems will be more weighted toward local information and fall in the red region. The purple region represents systems with high redundancy, where local and global states carry overlapping information about the control signal. The yellow region represents synergistic systems, where knowing both local and global states to- gether with give more information about the control signal than separately. robot control or analyzing animal locomotion. While we found no overall change in centralization with speed during running in cockroaches, the constituent informations, IL and IG, did change (Fig. 6). Changing gait is likely to shift location in the architecture space. Slower walking gaits in cockroaches (49), stick insects (14), and robots (50) are thought to be more more decentralized using local or neighbor-based reflex rules (3). Even though central pattern generator circuits are still involved they are distributed and typically weakly coupled (51) predicting a more decentralized information strategy. Dif- ferent environments might also demand different information strategies. Tests in robotic models indicate that some amount of decoupling between legs, rather than a single centralized controlled trajectory for all legs, results in increased robustness over more variable terrain (52). These results would predict a leftward shift along ICEN T on rough terrain. Movement in either direction on the centralization axis could simplify control, such as a highly actuated ribbon fin that only needs to shift trajectories of several of a hundred fin rays to ma- neuver (53) might be more decentralized, while few control signals driving the coordination of many muscles might be more centralized (54). Overall, scenarios where positive ICEN T is beneficial suggest that it is more important sense the global state whereas scenarios where negative ICEN T is beneficial suggest emphasizing local state sensing. Due to the general collapse of high dimensional control inputs to a low dimensional outputs, one might expect the positive ICO indicating net redundancy between global and local information in all the systems (55). In most all examples shown here, ICO is positive except for the slower group of cockroach data when considering only spike count (Fig. 4). Changes along the ICO axis are possible and could give dif- ferent performance benefits. For the example of mechanical coupling, stiffer legs and body coupling would likely result in a highly redundant system in the purple area in Fig. 6B. In terms of maximizing the possible information the control variable could share with both the local and global states, it would be beneficial to have synergistic information rather than redundant information. Such a scenario could be possible if the control receives both global and local feedback, where both types of feedback affect the control signal differently together than they do separately. We expect soft animals and robots could benefit from a synergistic control architecture because local states might be very independent from global states. One example is a robotic slime mold where each actuator on the edge receives feedback relating to its local neighbors as well as the inner protoplasm that globally interacts with all locomotion in robots (18), allowing a six legged robot with correctly tuned mechanics to move with just a single actuator (44). Adding stiff spines to legs (45), flexible joints to the body (46), or streamlined shells (47) allow animals and robots to traverse challenging terrain. The ability to estimate the effects of mechanical feedback such as in these examples could allow for adaptive control (48). Our centralization measurements resolve changes to the mechanics in the robot that would not be evident from kinematics or footfall patterns alone. Discussion We have introduced an empirical measure of centralization of locomotor control that does not rely on any underlying model. Therefore, we can use ICEN T to compare systems. We choose a normalization scheme that compares ICEN T and ICO in proportion to IT OT shown in Fig. 6. The coupled oscillator model, which has been used to de- scribe legged locomotion (14) and the control of robots (19), has been used previously to represent gradations of control along the centralized/decentralized axis (4). When plotted on the ICEN T / - ICO axes in Fig. 6A, the coupled oscilla- tor does vary from decentralized to centralized as coupling strength increases. However, we also see that fast cockroach locomotion is centralized, meaning that IG outweighs IL, and ICEN T increases with stride frequency. The overall ICEN T of the cockroach matches that of the coupled oscillator model with a slightly centralized coupling strength. This result is further validated by a previous study which fit cockroach leg kinematics explicitly to a coupled-oscillator model (11). The coupling strengths estimated from this fitting averaged to 76.1, and when we estimate ICEN T for this model we find that IG does indeed outweigh IL and is very close to the measured ICEN T of the cockroach (Fig. 6). The robotic results contrast that of the cockroach as they fall on the decentralized side of the axis. More importantly, the differences in ICEN T due to changes in mechanical cou- pling show the importance of considering interactions between the limbs, body, and environment when designing control for robotic coordination. Our centralization measure can thus be used as an empirical diagnostic tool to assess how coupling between legs is affected by mechanical changes that could be difficult to model. The animal, robotic, and coupled oscillator systems ex- plicitly explore the centralization axis. Other systems likely populate the rest of the information space and exploring ben- efits of the different regions could be a guide in developing 6 Neveln et al. -11-11ICENT / ITOTICO / ITOTcoupling strength = 75 as reported in Couzin-Fuchs et. al.-11-11coupled-oscillator modelcockroachminitar robot - front legcoupling strength robot pitch Mlowmedhigh075150minitar robot - rear legABICENT / ITOTICO / ITOT motion: adaptive control of centrally coupled pattern generator circuits. Frontiers in Neural Circuits 4:125. 34. Fuchs E, Holmes P, David I, Ayali A (2012) Proprioceptive feedback reinforces centrally gen- erated stepping patterns in the cockroach. Journal of Experimental Biology 215(11):1884 -- 1891. 35. Mantziaris C, et al. (2017) Intra- and intersegmental influences among central pattern gen- Journal of Neurophysiology erating networks in the walking system of the stick insect. 118(4):2296 -- 2310. 36. Pearson KG, Iles JF (1971) Innervation of coxal depressor muscles in the cockroach, Peri- planeta americana. The Journal of experimental biology 54(1):215 -- 232. 37. Ahn AN, Meijer K, Full RJ (2006) In situ muscle power differs without varying in vitro mechan- ical properties in two insect leg muscles innervated by the same motor neuron. The Journal of experimental biology 209(Pt 17):3370 -- 82. 38. Full R, Stokes D, A (1998) Energy absorption during running by leg muscles in a cockroach. The Journal of experimental biology 201 (Pt 7)(7):997 -- 1012. 39. Sponberg S, Libby T, Mullens CH, Full RJ (2011) Shifts in a single muscle's control potential of body dynamics are determined by mechanical feedback. Philosophical transactions of the Royal Society of London. Series B, Biological sciences 366(1570):1606 -- 20. 40. Kukillaya R, Proctor J, Holmes P (2009) Neuromechanical models for insect locomotion: Sta- bility, maneuverability, and proprioceptive feedback. Chaos 19(2):026107. 41. Wosnitza A, Bockemühl T, Dübbert M, Scholz H, Büschges A (2013) Inter-leg coordination in the control of walking speed in Drosophila. The Journal of experimental biology 216(Pt 3):480 -- 91. 42. David I, Holmes P, Ayali A (2016) Endogenous rhythm and pattern-generating circuit interac- tions in cockroach motor centres. Biology Open 5(9). 43. De A, Koditschek DE (2018) Vertical hopper compositions for preflexive and feedback- stabilized quadrupedal bounding, pacing, pronking, and trotting. The International Journal of Robotics Research 37(7):743 -- 778. 44. Hoover AM, Burden S, Xiao-Yu Fu, Shankar Sastry S, Fearing RS (2010) Bio-inspired design and dynamic maneuverability of a minimally actuated six-legged robot in 2010 3rd IEEE RAS & EMBS International Conference on Biomedical Robotics and Biomechatronics. (IEEE), pp. 869 -- 876. 45. Spagna JC, Goldman DI, Lin PC, Koditschek DE, Full RJ (2007) Distributed mechanical feed- back in arthropods and robots simplifies control of rapid running on challenging terrain. Bioin- spiration & Biomimetics 2(1):9 -- 18. 46. Jayaram K, Full RJ (2016) Cockroaches traverse crevices, crawl rapidly in confined spaces, and inspire a soft, legged robot. Proceedings of the National Academy of Sciences of the United States of America 113(8):E950 -- 7. 47. Li C, et al. (2015) Terradynamically streamlined shapes in animals and robots enhance traversability through densely cluttered terrain. Bioinspiration & Biomimetics 10(4):046003. 48. Sastry SS, Bodson M (2011) {A}daptive {C}ontrol: {S}tability, {C}onvergence and {R}obustness. (Courier Corporation). 49. Bender JA, et al. (2011) Kinematic and behavioral evidence for a distinction between trotting and ambling gaits in the cockroach Blaberus discoidalis. The Journal of experimental biology 214(Pt 12):2057 -- 64. 50. Schilling M, Hoinville T, Schmitz J, Cruse H (2013) Walknet, a bio-inspired controller for hexa- pod walking. Biological Cybernetics 107(4). 51. Büschges A, Borgmann A (2013) Network Modularity: Back to the Future in Motor Control. Current Biology 23(20):R936 -- R938. 52. Von Twickel A, Büschges A, Pasemann F (2011) Deriving neural network controllers from neuro-biological data: Implementation of a single-leg stick insect controller. Biological Cyber- netics 104(1-2):95 -- 119. 53. Sefati S, et al. (2013) Mutually opposing forces during locomotion can eliminate the tradeoff between maneuverability and stability. Proceedings of the National Academy of Sciences of the United States of America 110(47):18798 -- 803. 54. Ting LH, Macpherson JM (2005) A Limited Set of Muscle Synergies for Force Control During a Postural Task. Journal of Neurophysiology 93(1):609 -- 613. 55. Full RJ, Koditschek DE, Full RJ (1999) Templates and anchors: neuromechanical hypotheses of legged locomotion on land. The Journal of Experimental Biology 2(12):3 -- 125. 56. Umedachi T, Takeda K, Nakagaki T, Kobayashi R, Ishiguro A (2010) Fully decentralized con- trol of a soft-bodied robot inspired by true slime mold. Biological Cybernetics 102(3):261 -- 269. actuators (56). The controller takes advantage of the different information the global and local state provides, which would suggest synergy and predict a location in the bottom half of the control architecture spaces. ACKNOWLEDGMENTS. We thank Avik De and Dan Koditschek for facilitating experiments with Minitaur. Shai Revzen, Sam Bur- den, Ilya Nemenman, Bob Full, Adrienne Fairhall and Sara Solla provided helpful discussions. This work was supported by NSF CAREER MPS/PoLS 1554790 to S.S. 1. Aoi S, Manoonpong P, Ambe Y, Matsuno F, Wörgötter F (2017) Adaptive Control Strategies for Interlimb Coordination in Legged Robots: A Review. Frontiers in neurorobotics 11:39. 2. Brambilla M, Ferrante E, Birattari M, Dorigo M (2013) Swarm robotics: a review from the swarm engineering perspective. Swarm Intelligence 7(1):1 -- 41. 3. Cruse H, Durr V, Schmitz J (2007) Insect walking is based on a decentralized architecture revealing a simple and robust controller. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 365(1850):221 -- 250. 4. Holmes P, Full RJ, Koditschek D, Guckenheimer J (2006) The Dynamics of Legged Locomo- tion: Models, Analyses, and Challenges. SIAM Review 48(2):207 -- 304. 5. Brown IE, Loeb GE (2000) A Reductionist Approach to Creating and Using Neuromuscu- loskeletal Models in Biomechanics and Neural Control of Posture and Movement. (Springer New York, New York, NY), pp. 148 -- 163. 6. Sponberg S, Full RJ (2008) Neuromechanical response of musculo-skeletal structures in cockroaches during rapid running on rough terrain. The Journal of experimental biology 211(Pt 3):433 -- 46. 7. Mendes CS, Bartos I, Akay T, Márka S, Mann RS (2013) Quantification of gait parameters in freely walking wild type and sensory deprived Drosophila melanogaster. eLife 2:e00231. 8. Fayyazuddin A, Dickinson MH (1996) Haltere afferents provide direct, electrotonic input to a steering motor neuron in the blowfly, Calliphora. The Journal of neuroscience 16(16):5225 -- 32. 9. Höltje M, Hustert R (2003) Rapid mechano-sensory pathways code leg impact and elicit very rapid reflexes in insects. The Journal of experimental biology 206(Pt 16):2715 -- 24. 10. Couzin-Fuchs E, Gal O, Holmes P, Ayali A (2015) Differential control of temporal and spatial aspects of cockroach leg coordination. Journal of Insect Physiology. 11. Couzin-Fuchs E, Kiemel T, Gal O, Ayali A, Holmes P (2015) Intersegmental coupling and recovery from perturbations in freely running cockroaches. Journal of Experimental Biology. 12. Opsahl T, Agneessens F, Skvoretz J (2010) Node centrality in weighted networks: Generaliz- ing degree and shortest paths. Social Networks 32(3):245 -- 251. 13. Watts DJ, Strogatz SH (1998) Collective dynamics of 'small-world' networks. Nature 393(6684):440 -- 442. 14. Ayali A, et al. (2015) The comparative investigation of the stick insect and cockroach models in the study of insect locomotion. Current Opinion in Insect Science 12:1 -- 10. 15. Kuramoto Y, Nishikawa I (1987) Statistical macrodynamics of large dynamical systems. Case of a phase transition in oscillator communities. Journal of Statistical Physics 49(3-4):569 -- 605. 16. Strogatz SH (2000) From Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators. Physica D: Nonlinear Phenomena 143(1-4):1 -- 20. 17. Revzen S, Koditschek DE, Full RJ (2009) Towards testable neuromechanical control architec- tures for running in Progress in Motor Control. (Springer), pp. 25 -- 55. 18. Koditschek DE, Full RJ, Buehler M (2004) Mechanical aspects of legged locomotion control. 19. Arthropod Structure & Development 33(3Portland, OregonSeattle, Washington):251 -- 272. Ijspeert AJ, Crespi A, Ryczko D, Cabelguen JM (2007) From swimming to walking with a salamander robot driven by a spinal cord model. Science 315(5817):1416 -- 20. 20. McGeer T, , et al. (1990) Passive dynamic walking. I. J. Robotic Res. 9(2):62 -- 82. 21. Dallmann CJ, Hoinville T, Dürr V, Schmitz J (2017) A load-based mechanism for inter-leg coordination in insects. Proceedings. Biological sciences 284(1868). 22. Owaki D, Ishiguro A (2017) A Quadruped Robot Exhibiting Spontaneous Gait Transitions from Walking to Trotting to Galloping. Scientific Reports 7(1):277. 23. Mendes CS, Rajendren SV, Bartos I, Márka S, Mann RS (2014) Kinematic Responses to Changes in Walking Orientation and Gravitational Load in Drosophila melanogaster. PLoS ONE 9(10):e109204. Ince R, Ince, A. RA (2017) Measuring Multivariate Redundant Information with Pointwise Common Change in Surprisal. Entropy 19(7):318. 24. 25. Kloeden PE, Platen E (1992) Numerical {S}olution of {S}tochastic {D}ifferential {E}quations. (Springer Berlin Heidelberg), p. 636. 26. Pantaleone J (2002) Synchronization of metronomes. American Journal of Physics 27. 70(10):992 -- 1000. Ijspeert AJ (2008) Central pattern generators for locomotion control in animals and robots: A review. Neural Networks 21(4):642 -- 653. 28. Revzen S, Guckenheimer JM (2008) Estimating the phase of synchronized oscillators. Phys- ical Review E 78(5):051907. 29. Watson JT, Ritzmann RE (1998) Leg kinematics and muscle activity during treadmill running in the cockroach, blaberus discoidalis: Ii. fast running. J. Comp. Physiol. A. 182(1):23 -- 33. 30. Jindrich DL, Full RJ (2002) Dynamic stabilization of rapid hexapedal locomotion. The Journal of experimental biology 205(Pt 18):2803 -- 23. 31. Sponberg S, Daniel TL (2012) Abdicating power for control: a precision timing strategy to modulate function of flight power muscles. Proceedings. Biological sciences 279(1744):3958 -- 66. 32. Srivastava KH, et al. (2017) Motor control by precisely timed spike patterns. Proceedings of the National Academy of Sciences of the United States of America 114(5):1171 -- 1176. 33. Fuchs E, Holmes P, Kiemel T, Ayali A (2011) Intersegmental coordination of cockroach loco- Neveln et al. April 17, 2019 7 Supporting Information Text Background on Information Theory. The discrete Shannon entropy (H) of a signal, given by the equation H(S) =X p(si)log p(si), [1] i quantifies the amount of information present in the signal, where si is each possible state the signal S can take and p is the probability distribution of the states (1). When the base of the logarithm is two, the unit of entropy is bits, where the number of bits represents the expected number of yes or no questions to determine the state of the signal. Entropy can be similarly defined for joint distributions as H(S1; S2) and conditional distributions as H(S1S2). [2] with equality only when the two signals are independent. The level of dependency between two signals is quantified by the mutual information I, which is the difference from equality in Eq. 2 given by the equation H(S1; S2) ≤ H(S1) + H(S2) Mutual information can also be written as I(S1; S2) = H(S1) + H(S2) − H(S1, S2). [3] I = H(S1) − H(S1S2) = H(S2) − H(S2S1). [4] Therefore, the mutual information measures the decrease in entropy in one signal when the state of the other signal is known. These overlapping entropies for our chosen set of signals (hereafter labeled C for the set of possible ci control states, L for the set of local states, and G for the set of global states) are graphically presented in Fig. S1. Estimation of mutual information of continuous variables can have error or bias due to limited sampling (2). We use a bin-less nearest neighbor estimator of I (2) which handles these issues well, as described in the next section. The entropy diagram in Fig. S1 helps build intuition about how the different mutual information quantities contain different parts of the decomposition of the total mutual information between the control signal and the joint local and global states. Local MI outlined by the dashed red line is the red and gray areas together in Fig. S1 and is given by IL = IU L + IR. I(C; LG) = IU L + ISY N [5] This is the mutual information between C and L when G is not known. When G is known, then the red area in Fig. S1 is given by [6] and does not include IR. Therefore, the grey area must have IR and a negative ISY N to balance out the positive ISY N in the red area to have zero overall synergy in IL. Estimating Mutual Information. We used the k-nearest neighbor method for estimating MI (2). In brief, the underlying conditional probability densities are estimated by counting how many samples in the marginal spaces are contained within the distance to the k-nearest neighbor of each sample point in the joint space. The choice of k sets the resolution to which the probability densities are estimated as the method assumes a uniform distribution in the ball smaller than the distance to the k-nearest neighbor. For details of the estimator see (2). We renormalized our variables to have zero mean and unit variance. Such a reparametrization has no impact on the actual MI between two variables, but can produce a better estimate as each variable is scaled equally and outliers have a smaller influence (2). Also, as our spiking variable is discrete, we added a small amount of noise with a standard deviation of 10−4 as otherwise many points in the dataset would have the same coordinates and therefore counting to the k-nearest neighbor becomes impractical. To calculate the co-information, we first estimated the total mutual information between the control variable and the joint local and global variables. Co-information is given by ICO = IG + IL − IT OT [7] We chose a value of k for which the estimates of the different mutual information values remained consistent as k varied. From Fig. S2, values of k between 5 and 10 give the same estimates for count (top plot), and they fall off consistently for timing (bottom plot). These consistent trends mean that the local and global estimates give consistent values for centralization whether or not normalized by the total information. We therefore use a value of k = 7 for calculating centralization and note that conclusions do not depend on changing k between 5 and 10. We followed a procedure similar to that in (3) to determine the error of our estimate of MI. We subsampled the data into m equally sized and independent groups containing N/m samples, calculated the MI for those m groups, and then calculated the standard deviation of those m MI estimates. We repeated this process 10 times and averaged the standard deviations (σ) for each value of m. We fit these mean standard deviations to log σ2 = A + log m relationship and estimated σ for the original full dataset by setting m = 1. The errorbars displayed in Fig. S2 show these measured and extrapolated σ values. We are also able to assess whether there exists sample-sized bias in our MI estimates if the estimates of MI stay within the errorbars as m is increased and the sample size is decreased. As shown in the right column of plots in Fig. S2, estimates for count (top plot) remain consistent and estimates for timing (bottom) fall off with the number of groups at the same rate, resulting in similar estimates of centralization whether or not normalized by total information. Izaak D. Neveln, Amoolya Tirumalai and Simon Sponberg 1 of 5 Cockroach Experiments. Blaberus discoidalis (henceforth cockroaches) were kept in an incubation chamber set to 37◦ C, 60% humidity, and a 12h/12h light cycle with ample supply to food and water. Cockroaches were first cold anesthetized in a refrigerator at 4◦ C for about 30 minutes. We then removed their wings and cut back their pronotum so that their legs would be more visible for our overhead video recordings. To insert the electromyogram (EMG) wires, we first restrained them ventral side up to gain access to their legs. The waxy coating on their abdomen and legs was scored with an insect pin to provide better adhesion for the super glue. We made a pair of small holes a couple millimeters apart through the exoskeleton of their medial coxa on both the left mesothoracic and metathoracic legs to gain access to femoral extensor muscles 137 and 179 respectively. We then inserted insulated silver wire electrodes (0.003 in. wire diameter, A-M Systems, Sequim, WA) into the holes just underneath the exoskeleton and glued them in place. A fifth ground electrode was inserted and glued into the abdomen following the same procedure. The wires were routed along the abdomen and glued on one rostral and one caudal segment. The light tether trailed behind the cockroach and was elevated to a connector above the experimental chamber. These methods are similar to those in (4, 5). Each electrode pair was amplified 100x using a differential amplifier (A-M Systems, Sequim, WA). Amplified signals were recorded through a data acquisition board (National Instruments, Austin, TX) and logged using custom software written in Matlab (Mathworks, Natick, MA). High speed video (Photron, San Diego, CA) was recorded at 500 fps from above. The arena was lighted with an array of infrared LEDs (Larson Electronics LLC, Kemp, TX). We prodded the cockroaches to run through a narrow opening that led to a wider field and recorded only trials in which the cockroach remained at least a centimeter from the walls. After 12 successful trials each lasting less than 2 seconds, videos were downloaded from the camera to a hard drive, and the cockroach rested for around 10 minutes until the next set of 12 trials. Up to 8 sets of 12 trials were collected per individual. EMG data were processed offline using a digital bandpass filter. A simple peak finding method was used to discriminate spikes from the filtered EMG data. The 2D kinematics of the endpoints of all six legs were tracked semi-automatically in the horizontal plane from the high speed video using custom software written in Matlab. Cubic spline interpolation was used to estimate the position of the leg endpoints during occlusions. Interpolated kinematics were manually checked for a subset of videos to insure accuracy. A global phase variable was estimated using the Phaser algorithm (6) and subsequently used to separate both the EMG and leg kinematic data into individual strides. Stride frequency was estimated from the average change in global phase versus change in time over a stride. Dimensionality Reduction of Cockroach Output States*. The output states for the cockroach, as well as all time series data, is too high dimensional to be able to effectively estimate mutual information. As the data is auto-correlated with time, we expect that a low dimensional representation of the states will contain all mutual information. The simplest dimensionality reduction is to take one sample from the trajectory in phase for each stride, which we call a phase slice. We found that adding a second slice increased the estimated information significantly, but not a third or fourth. The phases of the two slices also can result in various estimations of information as shown in Fig. S3. We thus chose two slices that were a half cycle apart that rested on the plateau of both the local and global MI landscape as shown by the black point in Fig. S3. We confirmed that conclusions concerning centralization did not change with as the phase of these slices varied or more slices were added. We also tried other dimensionality reduction methods such as principle component analysis (7) and partial least squares (8). We found that the two phase slice method resulted in higher estimates of mutual information than the first two components of these other methods, although overall conclusions were robust to the different methods of dimensionality reduction. References 1. Burks AW, Shannon CE, Weaver W (1951) The Mathematical Theory of Communication. The Philosophical Review 60(3):398. 2. Kraskov A, Stögbauer H, Grassberger P (2004) Estimating mutual information. Physical Review E 69(6):066138. 3. Srivastava KH, et al. (2017) Motor control by precisely timed spike patterns. Proceedings of the National Academy of Sciences of the United States of America 114(5):1171 -- 1176. 4. Watson JT, Ritzmann RE (1998) Leg kinematics and muscle activity during treadmill running in the cockroach, Blaberus discoidalis: I. Slow running. Journal of Comparative Physiology - A Sensory, Neural, and Behavioral Physiology 182(1):11 -- 22. 5. Sponberg S, Full RJ (2008) Neuromechanical response of musculo-skeletal structures in cockroaches during rapid running on rough terrain. The Journal of experimental biology 211(Pt 3):433 -- 46. 6. Revzen S, Guckenheimer JM (2008) Estimating the phase of synchronized oscillators. Physical Review E 78(5):051907. 7. Pearson K (1901) Principal components analysis. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 6(2):559. 8. Sponberg S, Daniel TL, Fairhall AL (2015) Dual dimensionality reduction reveals independent encoding of motor features in a muscle synergy for insect flight control. PLoS Computational Biology 11(4):e1004168. 2 of 5 Izaak D. Neveln, Amoolya Tirumalai and Simon Sponberg Fig. S1. Representation of overlapping entropies of the control signal, local state, and global state. We calculate IL, the area encapsulated by the dotted red line, IG, the area encapsulated by the dotted blue line, and IT OT , which is the filled in red, blue, and gray areas. ICEN T is negative when there is more red than blue area, and positive when there is more blue than red area. Izaak D. Neveln, Amoolya Tirumalai and Simon Sponberg 3 of 5 decentralizedcentralizedICENT < 0ICENT > 0H(C)control signal entropyH(L)local stateentropyH(G)IUL+ISYNIGICENT = IG - ILICO= IR - ISYNIL IUG+ISYNglobal stateentropy Fig. S2. Estimates of IG and IL while varying k and sample size. Top row is information in count, bottom row is information in timing. The left column shows how the estimates vary with k using all data. The right column shows how estimates vary with the number of subdivided groups. Errorbars show the standard deviation of the estimate as calculated by the procedure given in the text and adapted from (3). 4 of 5 Izaak D. Neveln, Amoolya Tirumalai and Simon Sponberg Mutual Information (bits)Groups0510-0.100.10.2k051000.10.20.30.4local MIglobal MIcoinformation Fig. S3. Effect of phase slicing on mutual information estimates in the cockroach. A) IL (including both count and timing) estimates depend on which two slices of the local state are considered. B) Same as (A) for the IG. We looked for a slice pair offset by a half cycle that resulted in a estimate close to maximal for both IL and IG. We selected the two slices indicated by the black point and verified that conclusions concerning centralization did not change with small variations to the phase of these slices. Izaak D. Neveln, Amoolya Tirumalai and Simon Sponberg 5 of 5 1st slice (phase)0.40.200.62nd slice (phase)0.800.20.40.61IL (bits)1st slice (phase)0.40.200.62nd slice (phase)0.800.20.40.61IG (bits)AB
1201.5428
1
1201
2012-01-26T02:14:16
A point process framework for modeling electrical stimulation of the auditory nerve
[ "q-bio.NC", "q-bio.QM" ]
Model-based studies of auditory nerve responses to electrical stimulation can provide insight into the functioning of cochlear implants. Ideally, these studies can identify limitations in sound processing strategies and lead to improved methods for providing sound information to cochlear implant users. To accomplish this, models must accurately describe auditory nerve spiking while avoiding excessive complexity that would preclude large-scale simulations of populations of auditory nerve fibers and obscure insight into the mechanisms that influence neural encoding of sound information. In this spirit, we develop a point process model of the auditory nerve that provides a compact and accurate description of neural responses to electric stimulation. Inspired by the framework of generalized linear models, the proposed model consists of a cascade of linear and nonlinear stages. We show how each of these stages can be associated with biophysical mechanisms and related to models of neuronal dynamics. Moreover, we derive a semi-analytical procedure that uniquely determines each parameter in the model on the basis of fundamental statistics from recordings of single fiber responses to electric stimulation, including threshold, relative spread, jitter, and chronaxie. The model also accounts for refractory and summation effects that influence the responses of auditory nerve fibers to high pulse rate stimulation. Throughout, we compare model predictions to published physiological data and explain differences in auditory nerve responses to high and low pulse rate stimulation. We close by performing an ideal observer analysis of simulated spike trains in response to sinusoidally amplitude modulated stimuli and find that carrier pulse rate does not affect modulation detection thresholds.
q-bio.NC
q-bio
A point process framework for modeling electrical stimulation of the auditory nerve Joshua H. Goldwyn1,2,3 Jay T. Rubinstein4,5,6 Eric Shea-Brown1,7 August 16, 2018 1 Department of Applied Mathematics, University of Washington 2 Center for Neural Science, New York University 3 Courant Institute of Mathematical Sciences, New York University 4 Department of Biomedical Engineering, University of Washington 5 Department of Otolaryngology, University of Washington 6 Virginia Merrill Bloedel Hearing Research Center, University of Washington 7 Program in Neurobiology and Behavior, University of Washington Abstract Model-based studies of auditory nerve responses to electrical stimulation can provide insight into the func- tioning of cochlear implants. Ideally, these studies can identify limitations in sound processing strategies and lead to improved methods for providing sound information to cochlear implant users. To accomplish this, models must accurately describe auditory nerve spiking while avoiding excessive complexity that would preclude large- scale simulations of populations of auditory nerve fibers and obscure insight into the mechanisms that influence neural encoding of sound information. In this spirit, we develop a point process model of the auditory nerve that provides a compact and accurate description of neural responses to electric stimulation. Inspired by the framework of generalized linear models, the proposed model consists of a cascade of linear and nonlinear stages. We show how each of these stages can be associated with biophysical mechanisms and related to models of neu- ronal dynamics. Moreover, we derive a semi-analytical procedure that uniquely determines each parameter in the model on the basis of fundamental statistics from recordings of single fiber responses to electric stimulation, including threshold, relative spread, jitter, and chronaxie. The model also accounts for refractory and summa- tion effects that influence the responses of auditory nerve fibers to high pulse rate stimulation. Throughout, we compare model predictions to published physiological data and explain differences in auditory nerve responses to high and low pulse rate stimulation. We close by performing an ideal observer analysis of simulated spike trains in response to sinusoidally amplitude modulated stimuli and find that carrier pulse rate does not affect modulation detection thresholds. 2 1 0 2 n a J 6 2 ] . C N o i b - q [ 1 v 8 2 4 5 . 1 0 2 1 : v i X r a 1 Introduction Cochlear implants restore a sense of hearing to individu- als with severe to profound hearing loss by electrically stimulating the primary sensory neurons in the audi- tory pathway. A complete understanding of the trans- formation from electrical stimulation to neural responses would aid the design of improved cochlear implant sound processing strategies. Computational and mathematical models, especially those that are carefully constrained by available neurophysiological data, can play an essen- tial role in exploring this transformation. In addition to providing an efficient platform for testing and evaluat- ing stimulation paradigms, models provide quantitative tools for studying how neural mechanisms influence the transmission of sound information in the auditory sys- tem. In this paper, we develop a point process modeling framework that can be used to simulate the response of the auditory nerve (AN) to cochlear implant stim- ulation. The dynamical and stochastic features of the model are matched to statistics that characterize neural responses to single and paired pulses of electrical cur- rent -- stimuli that are commonly used to characterize the response properties of AN fibers in animal models of electric hearing (Hartmann et al. 1984; van den Honert and Stypulkowski 1984; Dynes 1996; Bruce et al. 1999c; Miller et al. 1999; Shepherd and Javel 1999; Cartee et al. 2000; Miller et al. 2001a,b; Cartee et al. 2006). We go on to show that this model can provide insight into the re- sponses of neurons to extended pulse trains of electrical stimulation. Point process models also have mathemat- ical properties that facilitate analyses of auditory signal detection (Heinz et al. 2001a,b), and we close by relat- ing results from our model to the psychophysical test of amplitude modulation detection. Our approach connects two prior modeling frame- works. The first is the stochastic threshold crossing model (Bruce et al. 1999a,c). This model is a useful phenomenological representation of AN spiking. A num- ber of cochlear implant psychophysics experiments have been studied with this model (Bruce et al. 1999b; Xu and Collins 2007; Goldwyn et al. 2010). Unfortunately, as stated by its creators, it lacks sufficient dynamical details to provide valid predictions of AN responses to high rate stimulation (Bruce et al. 1999a). Studying re- sponses to high pulse rate stimulation is necessary to characterize contemporary cochlear implant stimulation strategies, and for reasons that we discuss in more detail below, is a primary focus of this work. We therefore ex- tend the stochastic threshold crossing model by turning to a more general class of point process models. Point processes describe spiking via an instantaneous firing rate that varies over time (Perkel et al. 1967; John- son 1996; Truccolo et al. 2005). They have been fre- quently applied to model auditory nerve firing (Miller and Mark 1992; Litvak et al. 2003b; Trevino et al. 2010; Plourde et al. 2011). Our approach is largely moti- vated by a specific family of point processes: gener- alized linear models (GLMs) (Paninski 2004; Truccolo et al. 2005). GLMs have emerged as an essential tool for modeling physiological data and investigating the cod- ing and computational properties of neurons (Paninski 2004; Paninski et al. 2007), including auditory neurons (Trevino et al. 2010; Plourde et al. 2011). Moreover, they hold particular promise for applications to sensory neural prostheses because they have useful mathemati- cal properties that permit efficient parameter fitting to spike train data (Paninski 2004), they can be used to optimally decode spike trains (Paninski et al. 2007), and they can be used in connection with real-time optimiza- tion methods to identify stimulation patters that control the timing of evoked spikes (Ahmadi et al. 2011). We apply our point process model to investigate dif- ferences between AN responses to low and high pulse rate electric stimulation. Interest in high pulse rate stim- uli arises from the intuitive notion that higher stimula- tion rates should provide greater temporal information to cochlear implant listeners and thereby improve speech reception. Indeed, computational modeling and neu- rophysiological studies have indicated that stimulating neurons with 5000 pulses per second (pps) pulse trains can desynchronize neural responses (Rubinstein et al. 1999; Litvak et al. 2001, 2003a). This may return AN ac- tivity to a state that more closely resembles normal spon- taneous activity in the healthy cochlea (Rubinstein et al. 1999; Litvak et al. 2003a). In related modeling and ex- perimental studies, high pulse rate stimulation has been shown to produce neural firing rates that more faith- fully reproduce the envelope-s of temporally-modulated stimuli (Litvak et al. 2003c,b; Chen and Zhang 2007; Mino 2007). Psychophysical data from tests of ampli- tude modulation detection, however, have not yielded clear indications that high pulse rate stimulation im- proves the ability of listeners to detect temporal mod- ulation (Galvin III and Fu 2005; Pfingst et al. 2007; Galvin III and Fu 2009). More generally, speech tests have reported mixed results as to whether higher pulse rate stimulation improves speech reception in cochlear implant listeners (Brill et al. 1997; Kiefer et al. 2000; Loizou et al. 2000; Vandali et al. 2000; Holden et al. 2002; Friesen et al. 2005). Computational modeling can help trace the connections between the perceptual bene- fits, if any, of high pulse rate stimulation and the neural responses that it evokes. 2 Method In this section we introduce the point process framework and discuss how it can be parameterized using commonly reported statistics of responses to single pulse and paired pulse stimuli. These include threshold, relative spread, 1 jitter, chronaxie, summation threshold, and two refrac- tory effects. We begin by introducing the response statis- tics that we used to construct the model (Sec. 2.1), fol- lowed by a general discussion of how these statistics can be extracted from a point process model (Sec. 2.2). We then introduce the specific model framework that we will use throughout the study (Sec. 2.3) and explain the pro- cedure by which each parameter in the model can be uniquely associated with a response statistic (Sec. 2.4). 2.1 Response statistics A first-order description of neuronal excitability in re- sponse to electrical stimulation is the firing efficiency curve. This is an input/output function that relates the current level of a single pulse of current to the probability that the stimulus evokes a spike. As shown by Verveen in his pioneering recordings of green frog axons (Verveen 1960), the firing efficiency curve can take the form of a sigmoid function and be approximated by the integral of a Gaussian distribution. The median of the underlying Gaussian distribution represents the stimulus level that elicits a spike with probability one-half. This level is re- ferred to as the threshold of the neuron, which we denote by θ. A measure of the variability of spike initiation is the relative spread (RS), defined as the standard devia- tion of the underlying Gaussian distribution divided by its mean. Firing efficiency curves are frequently used to characterize the response of AN fibers to cochlear implant stimulation in electrophysiological experiments (Dynes 1996; Bruce et al. 1999c; Miller et al. 1999; Shep- herd and Javel 1999; Miller et al. 2001b) and to parame- terize computational models (Bruce et al. 1999c; Stocks et al. 2002; Macherey et al. 2007; Imennov and Rubin- stein 2009; O'Gorman et al. 2009). We will follow this approach and use the firing efficiency curve, in particu- lar the parameters θ and RS, as the starting point for constructing our point process model. The firing efficiency curve summarizes the excitabil- ity of a neuron in response to a single pulse of stimu- lation. For longer pulse durations, the threshold cur- rent level is typically smaller, due to the capacity of the neural membrane to integrate charge over time. This dependence of threshold on pulse duration can be sum- marized by the chronaxie; the pulse duration at which the threshold level is twice what it would be for a much longer pulse duration (how long, and therefore reported values of chronaxie, vary in different experiments). This basic feature of membrane dynamics is absent from ear- lier stochastic threshold crossing models (Bruce et al. 1999a; Litvak et al. 2003b). The firing efficiency curve encapsulates variability in the initiation of spikes (or lack thereof), but an addi- tional source of randomness is in the timing of spikes. The standard deviation of spike times evoked by a sin- gle pulse is known as jitter. It is another statistic that has been widely studied in physiological and modeling studies of electrical stimulation of the auditory nerve (van den Honert and Stypulkowski 1984; Miller et al. 1999; Javel and Shepherd 2000; Miller et al. 2001b; Car- tee et al. 2006). Jitter can depend on pulse duration and pulse level, but in our model we will only use the value of jitter that is most commonly reported in the literature: the value measured for the same pulse duration and level that defines the threshold. In order for the model to generate realistic responses to high pulse rate stimulation, it must also include spike history effects. The first, and most essential, is a refrac- tory effect that reduces the excitability of the model neu- ron in the immediate aftermath of an evoked spike. This effect will be implemented as a transient increase in the threshold value θ following a spike, similar to the method of Bruce and colleagues (Bruce et al. 1999a). The point process will also include a second spike history effect -- a transient increase in RS immediately after a spike (Miller et al. 2001a). Based on simulation results from a bio- physically detailed computational model of an AN fiber, it has been hypothesized that this phenomenon is a sig- nature of the random activity of ion channels in a cell membrane known as channel noise (Matsuoka et al. 2001; Mino and Rubinstein 2006). Modeling channel noise in the auditory nerve in the context of cochlear implant stimulation is important for a number of reasons. First, neurons in the deafened cochlea do not receive the typi- cal synaptic input from inner hair cells, and therefore ion channels may generate a dominant noise source in this stage of auditory processing. Second, channel noise may be a mechanism by which high pulse rate stimulation can desynchronize neural responses, thereby leading to improved encoding of electrical stimuli (Rubinstein et al. 1999; Litvak et al. 2003c,b; Mino 2007). Finally, we incorporate an additional feature into the model that is relevant for high carrier pulse rates -- the capacity of a neuron to integrate consecutive subthresh- old pulses. In other words, the model will account for the phenomenon by which multiple pulses are more likely to evoke a spike than would be expected if each pulse acted independently on the neuron. This effect, known as summation or facilitation, has been observed in re- sponses to pairs of pulses with short interpulse intervals (Cartee et al. 2000, 2006) and to high rate pulse train stimuli (Heffer et al. 2010). As we will see, this form of interpulse interaction creates fundamental differences between responses to low and high rate pulse train stim- ulation. 2.2 Point process theory In this section, we introduce a number of basic ideas from the theory of point processes and illustrate their connec- tions to the response statistics discussed above. A point process model is completely defined by its conditional 2 intensity function (Daley and Vere-Jones 2003), which can be interpreted as the instantaneous probability that a neuron will spike (Truccolo et al. 2005). We denote the conditional intensity function as λ(tI, H), where I is the stimulus applied to the neuron and H represents the past history of the neuron, both of which should be viewed as functions of time. A related and important quantity is the integrated intensity function: (cid:90) t2 Λ(t1, t2I, H) = λ(sI, H)ds. t1 From Λ(t1, t2I, H), one can define the probability that there will be a spike in a time interval [t1, t2, ]. This function is known as the lifetime distribution function and is given by (Daley and Vere-Jones 2003): L(t1, t2I, H) = 1 − e−Λ(t1,t2I,H). (1) The lifetime distribution function is central to our analysis because it forms the mathematical link between the point process model and the statistics that describe the responses of neurons to single and paired pulses. Consider, for instance, responses to a single pulse of cur- rent level I and some duration D. If we let t2 in Eq. 1 become sufficiently large, then the lifetime distribution function defines the probability with which a single pulse of a certain current level I will ever evoke a spike. In other words, when viewed as function of I, it is equiva- lent to the firing efficiency curve. If we now change our perspective, and view Eq. 1 as a function of t2, for a fixed value of I, then this function gives the probability that a spike has been evoked before time t2. In this context, the lifetime distribution function describes the temporal dispersion of spike times, and can thus be used to cal- culate jitter. We now present how the single pulse and paired pulse response statistics discussed in Sec. 2.1 can formally be obtained using point process theory. We begin with a set of measures that are obtained from responses to single pulses, and we therefore assume that they are not influenced by spike history effects. This allows us to omit the term H for now and let I refer the current level of a single stimulating pulse. As introduced above, the threshold value θ is the current level at which the firing efficiency curve has a value of 1/2. From the definition of the lifetime distribution, θ must satisfy = L(0,∞θ). 1 2 Applying Eq. 1, this can be rewritten in terms of the integrated intensity function: Λ(0,∞θ) = log 2 (2) To define RS, we generalize the traditional defini- tion that is based on assuming the firing efficiency curve is shaped like the integral of a Gaussian distribution (Verveen 1960). From Eq. 1, it is apparent that the firing 3 efficiency curve for a point process model is not neces- sarily an integrated Gaussian function. It is, however, the cumulative distribution of some probability density function. To be concrete, we define an associated prob- ability density function as the derivative of the lifetime distribution function with respect to the current level I: lI (0,∞I) = d dI [Λ(0,∞I)] e−Λ(0,∞I) By analogy with the traditional definition, we let RS be the standard deviation of this associated density function normalized by its mean. To define chronaxie in terms of a point process model, we use Eq. 2 and compare threshold current levels for varying pulse durations. Let θ(D∞) denote the threshold for a monophasic pulse of a long pulse of duration D∞. Then, the chronaxie is the phase duration Dc for which the single pulse threshold θ(Dc) is twice the value of θ(D∞). Specifically, Dc must satisfy the following two conditions: θ(Dc) = 2θ(D∞) Λ(0,∞θ(Dc)) = Λ(0,∞θ(D∞)) = log(2) The final single pulse statistic we consider is jitter. As mentioned above, jitter can be obtained from the life- time distribution function when it is viewed a function of time. Following the standard practice, we will consider jitter in spike times evoked by a single pulse presented at the threshold level θ. Then the probability of at least one spike occurring in the interval [0, t], conditioned on the event that the stimulus produces a spike at any time, is twice the lifetime distribution function: 2L(0, tθ). We define the associated probability density function by tak- ing a derivative with respect to time of Eq. 1, and see that jitter is the standard deviation of the following den- sity function for the probability that a spike occurs at time t: lt(0, tθ) = 2λ(0, tθ)e−Λ(0,tθ). (3) We now turn to responses evoked by pairs of pulses. Using a summation pulse paradigm (Cartee et al. 2000, 2006), we can quantify how threshold current levels change for a stimulus consisting of two pulses of equal current levels, separated by an interpulse interval (IPI) of vary- ing duration. Note that in experimental studies, the question asked is whether the pulse pair evokes a spike; the timing of the spike is not relevant. Assuming that the response to the pulse pair is not influenced by any prior spiking activity, the summation threshold satisfies the same threshold condition given by Eq. 2, the only changes are that the stimulus I is a pair of pulses and θ is the current level at which this pair of pulses evokes at least one spike on half of all trials, on average. Lastly, we observe how refractory effects can be in- corporated by allowing the single pulse threshold θ and RS to depend on the time since the last pulse. In the re- fractory pulse paradigm (Cartee et al. 2000; Miller et al. 2001a; Cartee et al. 2006), a strong first pulse forces the neuron to spike and then a firing efficiency curve is measured from the responses to a second pulse pre- sented some time later. The mathematical relationships between that firing efficiency curve and the lifetime dis- tribution function remain unchanged, but they must be formally modified by adding the spike history term H. 2.3 Model formulation We have presented the connection between the point pro- cess model and response statistics in a general manner. In this section, we make these connections more explicit by proposing a specific structure for the conditional in- tensity function. The model consists of a cascade of linear and nonlinear stages followed by a probabilistic spike generator. This structure is inspired by the pop- ular generalized linear model (GLM) class of point pro- cess neuron models. Fig. 1 illustrates the model fea- tures. The model differs from standard GLMs in several ways. These include a nonlinear stage that depends on the time since the last spike, an asymmetry in how the positive and negative phases of charge-balanced pulses are filtered, and a secondary filter that adds variability to simulated spike times. The action of the model, depicted schematically in Fig. 1, can be summarized as follows: an incoming pulse train I(t) is passed through stimulus filters K +(t) and K−(t), the outputs of the filters are recombined and used as input to a nonlinear function f (·). To incorporate re- fractory effects, the stimulus filters and the nonlinearity are all modified depending on the time since the last spike. An additional filter J(t) is included to add vari- ability to simulated spike times. The result of this chain of events is an instantaneous, history-dependent value for the conditional intensity function that defines the point process model: λ(tI, H) = [J ∗ f (K + ∗ I + + K− ∗ I−)](t), (4) where I + and I− represent the positive and negative portions of the input and ∗ represents the convolution operator. A variety of methods exist for fitting GLMs to neural data including reverse correlation to white noise stimuli (Simoncelli et al. 2004) and maximum likelihood meth- ods (Paninski 2004). We pursue a different route and take advantage of the special structure of the proposed model and the mathematical relationships in Sec. 2.2 to develop a semi-analytical procedure that uniquely iden- tifies parameters in the point process model with spe- cific response statistics. For reasons of mathematical tractability and biological relevance, which we articulate more fully below, we define the components of the model as follows: f (v) = K +(t) = K(−t) = J(t) = 0 (cid:26) vα (cid:26) κ (cid:26) βκ (cid:26) 1 τK 0 τK 0 if v ≥ 0 else e−t/τK if t ≥ 0 else e−t/τK if t ≥ 0 e−t/τJ τJ 0 else if t ≥ 0 else (5a) (5b) (5c) . (5d) This model has five parameters: α, κ, τK, β, and τJ . We next show how they are uniquely determined by the five response statistics discussed in Sec. 2.1: relative spread, threshold, chronaxie, jitter, and summation pulse thresh- old. We will also illustrate how the model can account for refractory effects by allowing the parameters α and κ to depend on the time since the previous spike. 2.4 Model parameterization We begin by discussing the response measures that do not depend on spike history, and thus neglect H for now. To simplify our presentation, we introduce some addi- tional notation. Let w(t) be the waveform of the stim- ulus; for instance, the waveform of a monophasic pulse is a rectangular step that reaches a maximum value of 1. Alternatively, the waveform function for a biphasic pulse consists of two rectangular phases, one reaches a value of +1 and the other −1. We next define the filtered waveform function W (t), which represents the action of K + and K− on w(t): (cid:2)K +(s)w+(t − s) + K−(s)w−(t − s)(cid:3) ds, (cid:90) t W (t) = 0 where w+ and w− are the positive and negative parts of the waveform function. For the case that all pulses have identical current level I, for instance in single pulse or summation pulse experiments, the intensity function for the point process model has the compact form: λ(tI) = (κI)α[J ∗ W α](t). The parameter κ and the pulse current level I both fac- tor out due to our choice of a power law nonlinearity. In addition, the jitter filter J(t) can be ignored when con- sidering the integrated intensity function over all time. This follows from Fubini's Theorem (Jones 2001) and the fact that(cid:82) ∞ (cid:90) ∞ 0 J(t)dt = 1: [J ∗ W α](t)dt = W (t)αdt. (cid:90) ∞ 0 If we denote this integral as 0 (cid:90) ∞ 0 4 W (t)αdt. ≡ Wα, (6) Figure 1: Schematic diagram of the point process model. Current input to the model, shown here as a train of biphasic pulses, is passed through a cascade of linear filters and a nonlinear function. This produces a value -- the conditional intensity -- that defines the instantaneous probability of a spike, which is then used to generate random sequences of spike times. Previous spikes provide feedback that modulates the stimulus filters and the nonlinearity. then the lifetime distribution function for the time inter- val [0,∞) can be written as: distribution: L(0,∞I) = 1 − exp (−καI αWα) . (7) θ = 1 κ (cid:114) log 2 α Wα . (9) As noted above, this lifetime distribution function is, in the language of neurophysiology, the firing efficiency curve or input/output function of the neuron in response to a single pulse of applied current. Next, we use this equation to parameterize the model. 2.4.1 Relative spread rameter (cid:2)κ α √Wα The choice of a power law nonlinearity significantly sim- plifies the parameterization process. The lifetime distri- bution function in Eq. 7 is a Weibull cumulative distri- bution function with shape parameter α and scale pa- . We define the relative spread as the associated standard deviation divided by the mean. Using known expressions for these values we find (Rinne 2009): (cid:3)−1 RS = Γ(1 + 2 α ) Γ2(1 + 1 α ) − 1, (8) where Γ(·) is the Gamma function. Remarkably, the re- lationship between RS and α does not depend on any other model or stimulus parameters. There is therefore a one-to-one relationship between the response statistic RS and the model parameter α. To fit the model, we invert this relationship to find α as a function of RS. In practice this must be done numerically. 2.4.2 Threshold The threshold current level θ is obtained by solving Eq. 2, which is equivalent to finding the median of the Weibull (cid:115) We invert this relationship in order to have an expression for the model parameter κ as a function of the response statistic θ: (cid:114) log 2 α Wα κ = 1 θ . (10) Note that κ depends on the model parameters α as well as β, and τκ (through the definition of Wα in Eq. 6). As we will see, the values of these parameters are inde- pendent of θ, and can therefore be viewed as (known) constants in this equation. 2.4.3 Chronaxie In order to fit the model parameter τκ, observe that the waveform function w(t) and thus the associated func- tion Wα both depend on the pulse duration. We denote this dependence as Wα(D). By definition, the chronaxie value is the pulse duration Dc for which the threshold for a monophasic pulse is twice the value of the thresh- old in response to a monophasic pulse of a long duration D∞ (the exact value of D∞ varies across experimental studies). We can use Eq. 9 to define the threshold cur- rent level as a function of pulse duration, and obtain the relationship: (cid:115) α 1 κ (cid:115) α log 2 Wα(Dc) = 2 κ log 2 Wα(D∞) . The factors of κ cancel one another and the equation can be further simplified to: Wα(D∞) Wα(Dc) = 2α. (11) 5 Pulse K+K-Stimulus FiltersNonlinearityI(t)JJitter FilterProbabilistic SpikingSpike Refractory f Conveniently, the solution to this equation depends only on Wα and α. Now, α can be assumed to have a known value since it is uniquely determined by RS (see Eq. 8). Furthermore, Wα does not depend on β since chron- axie is defined with respect to monophasic pulses. As a consequence, the only undetermined model parame- ter in this equation is τκ, which enters in the definition of Wα. There is no analytical way to identify τκ as a function of chronaxie, but we can numerically evaluate Wα(D∞) and Wα(Dc) and then use numerical root find- ing methods to determine the value of τκ that satisfies the relationship in Eq. 11, for given values of chronaxie Dc and reference pulse duration D∞. 2.4.4 Summation threshold To parameterize the temporal summation properties of the model, we replicate the paired pulse experiments of Cartee et al. (2000) using pseudo-monophasic pulses (Cartee et al. 2000; Cartee 2000). In these experiments, two charge-balanced pulses are presented at varying in- terpulse intervals (tIP I ). The threshold current level for the combined response to both pulses, which we denote θ2, is compared to the threshold current level for a single pulse presented in isolation, which we denote θ1. Car- tee et al. (2000) measured thresholds at three interpulse intervals (100, 200 and 300 µs) and summarized the re- lationship between θ1 and θ2 using the function: θ2 θ1 = 1 − 1 2 exp (−tIP I /τsum). (12) The parameter τsum controls the length of time over which the neuron can effectively sum consecutive sub- threshold pulses. To translate this into the language of the point pro- cess model, we use Eq. 2 to obtain a relationship between the single and paired pulse thresholds. Let Wα,1 denote the filtered waveform for a single pulse and Wα,2(tIP I ) be the filtered waveform for a pair of pulses separated by an interpulse interval of length tIP I . Then the ratio of the thresholds can be written as: (cid:115) Wα,1 θ2 θ1 = α Wα,2(tIP I ) (13) The model parameters that enter into this equation are α, τκ, and β (since the pulses in this paradigm have negative phases). We treat α and τκ as constants in Eq. 13 since they do not depend on β and can be de- termined at prior steps in the fitting procedure. We de- termine β by minimizing the mean square error between the two representations of θ2/θ1 on the right hand sides of Eqs. 12 and 13. This must be done using numerical methods for integration and minimization. The parame- ter β will take a value between zero and one, with smaller values leading to greater subthreshold integration of con- secutive pulses. We are are not aware of experimental 6 studies that that indicate how the temporal integration properties of the AN change immediately after a spike, so β is left as a constant that does not depend on spike activity. 2.4.5 Jitter The remaining undetermined parameter is τJ . To de- termine its value, we recall that jitter in the point pro- cess model is defined as the standard deviation of the probability density function in Eq. 3. Once again, we cannot obtain a simple analytical relationship between τJ and jitter, but we can use numerical integration and root finding methods to find the value of τJ for which the standard deviation of the density function in Eq. 3 is equal to a measured value of jitter. In other words, we solve the minimization problem: (cid:18)(cid:90) ∞ (cid:19)2 − γ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) , (cid:115)(cid:90) ∞ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) argmin τJ t2lt(0, tθ)dt − tlt(0, tθ)dt 0 0 where γ is the experimentally observed jitter value. Al- though the density function lt(0, tθ) depends on all of the other parameters in the model, by this step in our parameterization method all other parameter values will have been specified and we can therefore treat them as (known) constants when determining the value of τJ . Available data from cat AN fibers did not provide strong evidence that jitter changes significantly due to spike his- tory effects (Miller et al. 2001a). Thus, the parameter τJ does not depend on spike activity and is held constant throughout our simulations. 2.4.6 Refractory effects Finally, we describe how spike history effects can be in- corporated in the model by allowing the parameters κ and α to depend on the time since the last spike. The re- lationship between κ and θ in Eq. 10 shows that increas- ing κ immediately following a spike allows the model to exhibit a common refractory property, whereby neu- rons are less excitable in the aftermath of a spike. Ad- ditionally, the relationship between α and RS in Eq. 8 shows that increasing α after a spike allows the model to exhibit an increase in RS during the refractory period. We choose the dynamics of these parameters to match the experimentally measured values reported in Miller et al. (2001a). This study used a paired pulse stimula- tion paradigm to probe the changes in threshold and RS due to spike time. Following an initial "masker" pulse that is sufficiently strong to always evoke a spike, there is an interpulse interval before a second pulse is presented. The current level of the second pulse is varied in order to measure the firing efficiency curve of the neuron within the refractory period. Miller et al. (2001a) defined the dependence of θ and RS on the interpulse interval, which we denote as ∆t, with the following equations: θ(∆t) = θ0 1 − e−(∆t−tθ)/τθ . and RS(∆t) = RS0 1 − e−(∆t−tRS )/τRS . (14) (15) Here, θ0 and RS0 are baseline values obtained from sin- gle pulse responses. The parameter tθ represents the absolute refractory period during which no spikes can be produced. We implement the absolute refractory period by holding the intensity function λ(tI, H) at zero un- til the elapsed time since the last spike exceeds tθ. The time constant τθ quantifies how quickly the effects of a previous spike fade, and is therefore a measure of the relative refractory period. The parameters tRS and τRS play similar roles in defining the evolution of RS, as a function of the time since the last spike. To simulate the model with refractory effects, we let ∆t in the above equations represent the time since the last spike. We then update the values of κ and α at the onset of each pulse, based on the time since the last spike. These parameters are then kept at a constant value from the onset of the pulse until the onset of the next pulse in the stimulus, at which time their values are updated again to reflect the changed time since the last spike. This simplification, as opposed to allowing the pa- rameters to vary continuously, allows a straightforward parameterization of κ and α using the same single pulse relationships with θ and RS, respectively, which were derived above. The value of α must be obtained using numerical methods, but a formula for the value of κ can be obtained by combining Eq. 10 and Eq. 14: κ(∆t) = κ0(1 − e−(∆t−tθ)/τθ ), (16) where κ0 is the baseline value that is determined from the single pulse threshold θ0. 2.5 Summary of parameterization method The fitting sequence we have described above provides a semi-analytical method by which model parameters in Eqs. 5 can be uniquely determined based on single and paired pulse response statistics. This sequence can be summarized as follows: 1. Use RS value to determine nonlinearity parameter α in Eq. 5a. 2. Use α and chronaxie value to determine the stim- ulus filter time constant τκ in Eqs. 5b and 5c. 3. Use α, τκ and the summation time constant to de- termine the value β in Eq. 5c for scaling negative phases of the stimuli. 7 4. Use α, τκ, β, and threshold value to determine the scale factor κ in Eqs. 5b and 5c. 5. Use α, τκ, β, κ and jitter value to determine the jitter time constant τJ in Eq. 5d. 6. Use the refractory function for RS in Eq. 15 and the relationship between RS and α to determine the dynamics of α after a spike. 7. Use the refractory function for threshold in Eq. 14 and the relationship between threshold and κ to determine the dynamics of κ after a spike. Several comments are in order regarding the effect of pulse shape and duration on these parameters. Thresh- old should decrease with increasing pulse duration (van den Honert and Stypulkowski 1984), this property is cap- tured for monophasic pulses by the chronaxie statistics and τκ. However, thresholds also depend on pulse shape. In particular, thresholds measured with biphasic pulses are higher than those measured with monophasic pulse (Miller et al. 2001b). The model exhibits this prop- erty, but it is not quantitatively matched to experimental data. In contrast, the unique relationship between RS and α in Eq. 8 shows that, for this model, RS does not depend on pulse shape or pulse duration. This feature is consistent with data reported in (Miller et al. 1999), al- though others have suggested that RS may increase with pulse duration (Bruce et al. 1999b). The parameters β and τJ will both depend on the pulse shape and dura- tions in ways that we do not attempt to fit to physiolog- ical data. Ideally one would define the model with a set of data obtaining using a self-consistent set of stimula- tion parameters, and these same stimulation parameters would then be used to obtain additional physiological recordings or perform cochlear implant psychophysics experiments. We fit model parameters and compared subsequent simulation results to data from a variety of published sources; we will comment on variations in stim- ulation protocols and how they may impact our modeling results where appropriate. 2.6 Numerical methods All simulations were performed using original computer code written in the Fortran programming language. Nu- merical quadrature to evaluate the stimulus and jitter filters was performed using the trapezoid method and a time step of 1 µs. Random numbers were generated using the Mersenne twister algorithm (Woloshyn 1999). The original Fortran code is available from the authors upon request and a more user-friendly version of the code written for Matlab (Mathworks, Inc.) can be down- loaded from the ModelDB website (accession number 143760) or the website of one of the authors (http://www.cns.nyu.edu/∼goldwyn/). model neuron. We can reformulate the action of the stimulus filters K + and K− in terms of a differential equation, and find that the state variable v(t) evolves according to: τκ dv dt = −v + g(I(t)), (17) where the function g(I) is pictured in Fig. 2A and defined as: (cid:26) κ(∆t)I βκ(∆t)I if I ≥ 0 else . g(I) = The function κ(∆t), which depends on the time ∆t since the last spike, is defined in Eq. 16. The form of the differential equation for v(t) in Eq. 17 reveals that its dynamics are similar to the widely-used leaky integrate and fire model (Burkitt 2006) and related spike response models (Gerstner and Kistler 2002), a fact noted in previous presentations of GLMs (Paninski 2004; Paninski et al. 2010). The effect of spike history on g(I), which is determined by the dynamics of κ(∆t) in Eq. 16, is to reduce the slope of this piecewise linear function, as shown in the transition from the blue to green lines in Fig. 2A. This change in g(I) makes the model neuron less excitable immediately after a previous spike. This feature has also been included in spike response models (Gerstner and Kistler 2002) and integrate and fire mod- els (Badel et al. 2008). The asymmetry in g(I) enables the model to exhibit facilitation, or summation of con- secutive charge-balanced pulses. A modeling study using Hodgkin-Huxley type models has shown that summation of pseudomonophasic pulses, and in particular the time constant τsum in Eq. 12, is determined by the dynamics of the m-gating variable (Cartee 2000). The asymmetry in g(I), therefore, can be viewed as a phenomenolgical approximation of the nonlinear process -- mediated by Na+ activation -- by which the excitability of a neuron in response to consecutive charge-balanced pulses is in- creased beyond what would be expected if both pulses were presented in isolation from one another. The parameter τκ appears in Eq. 17 as the time scale of the dynamics of v(t). In the context of an integrate and fire model of a point neuron, this time scale is of- ten interpreted as the membrane time constant and re- lated to the passive integration properties of the neural membrane. In the context of cochlear implants, such a description overlooks the fact that intracochlear electric fields can interact with AN fibers at multiple, anatom- ically diverse regions of the neural membrane. For in- stance, membrane time constants of excitable nodes of Ranvier, somata, and unmyelinated portions of neurons may differ significantly (Rattay and Felix 2001). The parameter τκ, therefore, must be interpreted as a sin- gle time scale that summarizes the combined integrative properties of a spatially extended AN fibers in an exter- nal electric field. 8 Figure 2: Relationship between the point process model and a soft-threshold integrate and fire model. A: Illus- tration of the function g(I(t)) in Eq. 17. The parameter β controls the summation time constant of the model by decreasing the slope of g(I(t)) for negative currents. Refractory effects reduce the excitability of the neuron by decreasing the slope for all current values. B: Illus- tration of the probability density function for the noisy variable η in the stochastic threshold analogy (see text). Smaller values of α generate broader distributions, or equivalently a more variable spike initiation mechanism. 3 Results 3.1 Connections to biophysical mechanisms and models In order to gain intuition into the structure and behav- ior of the point process model, we take a brief detour to explore its key features and parameters. To provide this intuition, we connect aspects of the point process model to biophysical mechanisms and more familiar mathemat- ical models of neurons. 3.1.1 Subthreshold dynamics The result of applying the exponential filters K +(t) and K−(t) to the incoming stimulus I(t) can be equated with a dynamic state variable that we denote by v(t). This notation is meant to suggest, in a loose sense, that this state variable represents the membrane potential of the −0.500.50510ηfη(η) α=5α=15α=25−101 Current (I)g(I)κ Iβ κ Irefractoryκκ/20−κ/2AB 3.1.2 Spike generation In the second stage of the model, the power law nonlin- earity in Eq. 5a is applied to the state variable v(t) to generate the desired probability of spiking. In the GLM and related linear-nonlinear frameworks, this nonlinear- ity is typically static and often interpreted as a reduced description of the spike generating mechanisms of a neu- ron. We have introduced the innovation that the nonlin- earity depends on the time since the last spike in order to account for the observation that RS increases during the refractory period (Miller et al. 2001a). Here, we explain the relationship between the nonlinearity and spike ini- tiation and show why a dynamic nonlinearity produces the desired change in RS. We begin by viewing the state variable v(t) along with the nonlinear function f (v) = vα as a definition of a point process conditional intensity function. The probability of observing a spike in a small time window δt can then be obtained by approximating the lifetime distribution function in Eq. 1. We find that: P(spike; δt) ≈ 1 − e−vαδt. (18) We now consider an alternate point of view in which v(t) represents the subthreshold state of a neuron. We suppose that the probability that a spike occurs in the δt time window is related to the distance between v(t) and some voltage threshold, which we denote Θ. This idea is known as an escape noise model (Plesser and Gerst- ner 2000) and is closely related to stochastic threshold crossing models (Bruce et al. 1999b; Litvak et al. 2003b). From this perspective, we can imagine that the noise free trajectory v(t) is modified by adding a random number η that represents the noise in the spike generation process. The probability that a spike occurs in a δt window can be written as: P(spike; δt) ≈ P(v + η > Θ). (19) If we define the cumulative distribution function Fη for the noise variable η, then we can equate Eqs. 18 and 19 and get: Fη(η) = e−(Θ−η)αδt By taking the derivative of this distribution function, we derive a probability density function for η: fη(η) = αδt(Θ − η)α−1e−(Θ−η)αδt. In Fig. 2B, we illustrate the relationship between α and the probability density function of η by plotting fη(η) for several values of α. We computed these dis- tributions using δt = 1µs and Θ = α δt . The prob- ability density function for η becomes broad for small values of α. This shows how the parameter α character- izes variability in the spike initiation process. Moreover, (cid:113) log 2 this illustrates that allowing α to decrease immediately after a spike creates the desired behavior in the model -- a spike generation mechanism that is more variable within the refractory period. To provide a biophysical interpre- tation for this parameter, we note that the random open and closing of ion channels in the cell membrane (known as channel noise) has been shown to determine the value of RS in simulation studies (Rubinstein 1995). Thus, we view η as an abstract representation of channel noise. 3.1.3 Spike time variability The final stage of the model is to apply the jitter filter J(t) to the output of the nonlinearity. As we will see below, this secondary filter is necessary in order for this point process model to have realistic amounts of spike time variability. J(t) is applied after the nonlinearity, which suggests that it represents a source of timing vari- ability subsequent to the spike generator in the cell. In the context of extracellular electrical stimulation of the AN, one possible source of this variability is spatial in- teraction among multiple possible spatial sites of spike initiation. For instance, in a simulation study of a spa- tially extended neuron with multiple nodes of Ranvier, Mino et al. (2004) showed that stimulating electrodes which are more distant from a model AN fiber produce more variable spike times. The more distant electric fields broaden the distribution of nodes of Ranvier at which spike initiation can occur, thus spatial interactions across the axon may represent one post-spike mechanism that adds jitter to spike times. 3.2 Model parameterization With this understanding of the model features in hand, we proceed to parameterize the model following the steps outlined in Sec. 2.5. We defined parameters in the model using published data from single pulse and paired pulse recordings in cat. Table 1 summarizes the data values we used, their sources in the literature, and the param- eter values that we obtained from the fitting procedure. Whenever possible, we used mean values obtained from experiments that used biphasic pulses: these include the mean threshold, RS, and jitter values reported in (Miller et al. 2001b). For chronaxie, we used the mean value re- ported by van den Honert and Stypulkowski (1984) for intracochlear electrical stimulation of cat AN fibers. The longest phase duration tested was 2000 µs. For the sum- mation time constant (τsum in Eq. 12), we did not find a mean value reported in the published literature, so we assumed a value that fell within the range reported in Cartee et al. (2006) for AN responses to an electrode placed in the scala tympani. The parameter values that describe the refractory effects on threshold and RS were obtained from Miller et al. (2001a). Specifically, we used tθ = 332µs and τθ = 411µs, which were the mean values reported in this study. Mean values were not reported 9 Table 1: Neural response statistics and corresponding model parameter values Response statistic Value Reference Model Parameter Value Threshold Relative Spread Chronaxie Jitter Summation Time 0.852 mA (Miller et al. 2001b) (Miller et al. 2001b) 4.87% 276 µs (van den Honert and Stypulkowski 1984) (Miller et al. 2001b) 85.5 µs 250 µs (Cartee et al. 2006) κ α τκ τJ β 9.342 24.52 325.4 µs 94.3 µs 0.333 stimuli, but it is straightforward to modify our refrac- tory model should data obtained using biphasic pulses become available. Fig. 3 depicts the relationships between the response measures shown on the x-axes and the model parameters shown on the y-axes. Each panel illustrates the unique relationships that arose by following the parameteriza- tion sequence summarized in Sec. 2.5. For instance, in panel A we show the dependence of the nonlinearity pa- rameter α on RS. Once we had fixed the value of α based on the desired RS value, we then computed how the stim- ulus filter time constant τκ depended on chronaxie. This relationship is shown in panel B. One-by-one, we ob- tained values for each parameter, the values of which are marked by the × symbol in Fig. 3 and reported in Table 1. In order to introduce spike history effects on RS, we must recompute α at the onset of every pulse. In principle, we could use a numerical root finding method to invert Eq. 8, but to avoid this and thereby increase the computational speed of our simulations, we observed that the relationship between α and RS could be approx- imated by the power law: α ≈ RS−1.0587. (20) The combination of this equation for α and Eq. 15 pro- vided a simple rule for evolving α according to the time since the last spike. The approximation, which is shown with the red dashed line in the first panel of Fig. 3, is suitably accurate. 3.3 Model predictions: single pulse stim- uli 3.3.1 Firing efficiency curve We first simulated responses of the model to a single pulse of current. To be consistent with the experimental data from which we obtained values for threshold, RS, and jitter, we used a 40 µs per phase biphasic pulse stim- ulus as the input to the model. By varying the current level of the pulse, we could measure the complete firing efficiency curve, as shown in Fig. 4A. The probability of a spike is defined as the proportion of trials, out of a total of 5000 for each current level, on which the model gen- erated a spike. Simulation results are shown with black Figure 3: Relationships between response statistics and model parameters. Following the procedure summarized in Sec. 2.5, we obtain unique relationships between re- sponse statistics (x-axis) and model parameter values (y- axis). At each step, a parameter value is chosen to fit the response statistic (black ×, see Table 1 for exact values), and then this parameter value can be used to obtain a unique relationship between the next response statistic and model parameter value in the fitting hierarchy. The approximation used to determine α (Eq. 20) is shown as a dashed red line in the first panel. for tRS and τRS, so we estimated that tRS = 199µs and τRS = 423µs based on plotted data (Figure 8 in Miller et al. (2001a)). These experiments used monophasic 10 02550050100RS (%)α100250400100300500Chronaxie (μs)τκ (μs)1002504000.10.40.7Summation Time Constant (μs)β0.511.551015Threshold (mA)κ5015025050150250Jitter (μs)τJ (μs)ACBDE Figure 4: Responses of the point process model to sin- gle pulse stimulation. A: Firing efficiency curves for the point process model (blue), an integrated Gaussian func- tion with the same mean and RS (red), and simulation results from the model (black ×). B: Distributions of spike times obtained from 5000 simulations of the model with a jitter filter (blue) and without a jitter filter (red). ×-marks. As expected, they line up with the analyti- cally derived firing efficiency curve obtained from Eq. 1 and plotted with a blue line. In order to compare this input/output function to the more standard functional form that is used in the stochastic threshold crossing model of Bruce et al. (1999c), we show in red the integral of a Gaussian distribution with mean and standard de- viation consistent with the threshold and relative spread values that we used to fit the point process model. At the highest and lowest current levels, the firing efficiency curve for the point process model is slightly greater, but overall both methods produce similar spiking probabili- ties in response to a single pulse. Figure 5: Responses of the point process model to the summation pulse paradigm. A: Simulated spiking prob- abilities (symbols) for three interpulse intervals: 200 µs (green), 500 µs (red), 1000 µs (blue). Solid lines are ana- lytically obtained from the lifetime distribution function in Eq. 1. Single pulse firing efficiency curve is shown in black for reference. B: Thresholds estimated from simulated firing efficiency curves and normalized by the threshold for two independent pulses are shown as ×- marks for the three interpulse intervals. Facilitation oc- curs if this threshold ratio is less than one. Black curve is the analytically predicted result in Eq. 13. of τJ = 94.3µs, as determined by the parameterization method. This produces a distribution of spike times with standard deviation of 86µs, consistent with the mean value in (Miller et al. 2001b) that was used to param- eterize the model. In the absence of a second filtering stage, the model produces an extremely narrow distri- bution of spike times, as shown by the red distribution. The secondary filter is necessary, therefore, for the model to have realistic amounts of spike time jitter. 3.3.2 Jitter Fig. 4B shows the distribution of 10000 spike times ob- tained from the point process model in response to a sin- gle biphasic pulse of phase duration 40 µs presented at the threshold stimulus level. The blue line shows shows results for the model with a jitter filter time constant 3.4 Model predictions: paired pulse stim- uli 3.4.1 Summation pulse paradigm To test how the neuron model responds to pairs of pulses, we first simulated the summation pulse procedure in 11 −1−0.500.5100.20.40.60.81Probability of SpikeStimulus Level (dB re Threshold) Lifetime cdfGaussian cdfSimulationA05010015020000.050.10.150.2Spike Time (s)Proportion of occurences st. dev.=3μsst. dev.=86μsJitter FilterNo Jitter FilterBμ−4−2000.51Current (dB re Single Pulse Threshold)Probability of SpikeA05001000150020000.70.80.91Interpulse Interval (s)Threshold re Independent PulsesμB (Cartee et al. 2006). As inputs, we used two biphasic pulses of equal current level separated by an interpulse interval of either 200µs, 500µs, or 1ms. The probabili- ties of observing at least one spike in response to both pulses, for varying interpulse intervals and current levels, are shown as ×-marks in Fig. 5A. These were computed from 5000 repeated simulations of the model. We also plotted the analytically obtained firing probabilities for pulse pairs, obtained from Eq. 1, as curves in panel A. The effect of summation is visible as a leftward shift of these curves at shorter interpulse intervals. As a ref- erence, the black line reproduces the single pulse firing efficiency curve from Fig. 4A. To summarize these results, we estimated summa- tion thresholds at each interpulse interval by fitting an integrated Gaussian function. We then normalized the estimated threshold values relative to the single pulse threshold and plotted the values in Fig. 5B. The black curve represents the analytical prediction of the point process model, given by Eq. 13, for the threshold of two independent pulses. Facilitation occurs if the threshold ratio is less than one. In these simulations we see facilita- tion for interpulse intervals shorter than approximately 1000 µs. The point process model, therefore, will exhibit facilitation for carrier pulse rates of roughly 1000 pps and above. 3.4.2 Refractory pulse paradigm To probe the effects of spike history in the model, we fol- lowed the experimental procedure in (Miller et al. 2001a) and simulated responses to masker-probe pulse pairs. The current level of the first pulse was set to a very high value so that it always elicited a spike. The level of the second pulse was then varied in order to measure firing efficiency curves. We used pairs of biphasic pulses in or- der to be consistent with our previous simulations. Spike probabilities were defined as the proportion of trials (out of 5000 total) in which the second pulse produced a spike. Firing efficiency curves obtained from responses to the second pulse of current are shown in Fig. 6A for three dif- ferent interpulse intervals (667, 1000, and 1500 µs). As the interpulse interval (and equivalently the time since the last spike) decreases, the spike history effect has a greater influence on the probability that the second pulse will evoke a spike. As discussed in Sec. 2.4.6, there are two types of re- fractory effects in the model. The first is the standard refractory phenomenon whereby the model neuron is less excitable immediately after a spike. This is apparent in the rightward shift in the firing efficiency curves for shorter interpulse intervals in Fig. 6A. The second effect is that RS increases at shorter interpulse intervals. This leads to shallower slopes in the firing efficiency curves at shorter interpulse intervals, but the decibel scale in panel A obscures the change. There are some discrep- ancies between the simulation results (×-marks) and the Figure 6: Responses of the point process model to the refractory pulse paradigm. A: Simulated spiking prob- abilities (symbols) for three interpulse intervals: 667 µs (green), 1000 µs (red), 1500 µs (blue). Solid lines are an- alytically obtained firing efficiency curves obtained from lifetime distribution function in Eq. 1. Single pulse firing efficiency curve is shown in black for reference. B: Thresholds estimated from the simulated firing effi- ciency curves and normalized by the single pulse thresh- old are shown as symbols for the three interpulse in- tervals. Black curve is Eq. 14. C: Relative spread es- timated from the simulated firing efficiency curves and normalized by the single pulse relative spread are shown as symbols for the three interpulse intervals. Black curve is Eq. 15. predicted values from point process theory at the small- est interpulse intervals (green line). The differences arise 12 50010001500200011.52Time since spike (s)Normalized RSμ024600.51Current (dB re Single Pulse Threshold)Probability of SpikeA50010001500200011.52Time since spike (s)Threshold re Independent PulsesμCB from the fact that the simulated spike times in response to the first pulse have a small amount of random vari- ability, whereas the analytical result is computed using the assumption that the time of the first spike is locked to the time of the onset of the first pulse. To summarize these effects, we estimated threshold and RS from the simulated firing efficiency curves by fitting integrated Gaussian functions. These values are normalized with respect to single pulse threshold and RS and shown in panels B and C, respectively. By construc- tion of the model, the values computed from simulations agree with the refractory functions Eq. 14 and Eq. 15, which we plot as black curves. We emphasize that a model with a static nonlinearity, namely one in which α does not depend on spike history, would produce RS values that decrease at shorter interpulse intervals. Al- ternatively, the stochastic threshold crossing model of Bruce et al. (1999a) assumes a constant RS and would produce a flat line in panel C. 3.5 Model predictions: constant pulse train stimuli The results to this point have validated the fitting method and demonstrated that the model reproduces a range of measures that characterize responses to single pulses and pairs of pulses. Of greater relevance to cochlear implant speech processing strategies are the responses of neurons to extended trains of pulsatile stimuli. We first simulated responses to trains of biphasic pulses with phase dura- tion 40 µs and constant current levels, and sought to characterize how carrier pulse rate affects the sequence of evoked spikes. We relate our simulation results to relevant physiological data throughout. 3.5.1 Firing rate and interspike interval distri- butions Fig. 7A shows how firing rates in the model increase with current level for three different pulse rates. The stimuli were one second in duration. Mean and stan- dard deviations (shown with error bars) were estimated from 100 repeated simulations. At the lowest pulse rate (250 pps, blue line), the interpulse interval is longer than the summation time scale as well as the relative refrac- tory periods for threshold and RS. Thus the firing rate curve follows directly from the single pulse firing effi- ciency curve in Fig. 4A. The firing rate saturates at one spike per pulse, in this case 250 spikes per second. The 1000 pps and 5000 pps pulse trains have interpulse in- tervals of 1 ms and 200 µs, respectively, which are short enough for refractory and summation effects begin to im- pact the behavior of the model. Due to summation and the higher pulse rate, the same firing rate can be evoked by higher pulse rate stimuli using less current per pulse. Thus the firing rate curves are shifted leftward as the pulse rate increases. The remaining panels of Fig. 7 show interspike inter- val distributions at the three pulse rates, where the cur- rent levels were set to evoke approximately 100 spikes per second. The distributions were obtained from his- tograms of 10000 spike times in time bins of 100 µs. The distributions obtained from responses to the 250 pps and 1000 pps pulse trains, but not the high rate 5000 pps pulse train, show peaks that are clearly aligned to the interpulse intervals in the stimulus. This feature of the simulated interspike interval distributions is qualitatively similar to distributions recorded from cat AN fibers using the same pulse rates (Miller et al. 2008). For ease of com- parison, we include examples of interspike interval data reported by Miller et al in the insets of each panel. The model is capable of reproducing some characteristics of the interspike interval distribution for this single neuron, although an important caveat is that the experimentally- measured interspike intervals were obtained under dif- ferent stimulus conditions and likely also different firing rates. We do not intend to claim that the model com- pletely reproduces all of the behavior of this or other cells. The results of these simulations suggest that spike time jitter, which was incorporated into the model on the basis of responses to single pulse stimuli, can account for distinctive features of the interspike interval distribu- tions and how they vary with the stimulus pulse rate. On the one hand, temporal variability in spike times is small relative to the interpulse intervals for the lower pulse rate stimuli, which leads to the periodic nature of these dis- tributions in Fig. 7C and D. On the other hand, this small temporal variability is sufficient to spread the sim- ulated spike times across the interpulse interval for the 5000 pps, creating a more smoothly varying interspike interval distribution in panel B, consistent with the ap- pearance of the distribution shown in the inset. 3.5.2 Synchronization of spike times Simulation and physiological studies have generated in- terest in the possibility that high rate pulse trains may desynchronize neural responses to cochlear implant stim- ulation (Rubinstein et al. 1999; Litvak et al. 2003a). We tested whether the model exhibited similar signatures of desynchronization by computing a synchronization in- dex, known as vector strength (Goldberg and Brown 1969), and then compared our results to physiological data reported in Miller et al. (2008). For a sequence of N spike times {ti}N 1 , vector strength is defined with respect to a period T as: (cid:35)2 (cid:34) N(cid:88) (cid:35)2 (cid:118)(cid:117)(cid:117)(cid:116)(cid:34) N(cid:88) VS = 1 N cos(2πti/T ) + i=1 i=1 sin(2πti/T ) . (21) VS takes values between 0 and 1, with higher values be- ing interpreted as a more synchronized spike train. In 13 Figure 7: Firing rate and interspike interval distributions for three stimulus pulse rates. A: Firing rate as a function of current level per pulse. Stimuli were one second long trains of biphasic pulses with constant current level per pulse, the level of which is shown on the x-axis. Error bars indicate the standard deviation of the mean firing rate, estimated from 100 repeated simulations of the model. B-D: Interspike interval distributions estimated from 10000 interspike intervals with the pulse rates indicated in each panel. Current level per pulse was chosen so that the stimuli evoked approximately 100 spikes per second. Inset figures are single fiber recordings from cat AN fibers at the corresponding pulse rates reproduced from Fig. 1 in Miller et al. (2008) with kind permission from Springer Science+Business Media: J. Assoc. Res. Otolaryngol. these simulations, we use VS to measure the strength of phase locking to the period of the stimulus pulse train, and thus T represents the interpulse interval. In all sim- ulations, as well as the experimental data to which we compare our results, the stimulus is a train of biphasic pulses with a constant current level and a 40 µs phase duration. Simulation results for three pulse rates are shown in the left column of Fig. 8, where each circle rep- resents the firing rate and VS values computed from the response to a 10 second long pulse train. Our results can be compared to VS values obtained from recordings of cat AN fibers responding to biphasic pulses trains pre- sented at the same pulse rates and pulse shape. These data, which we reproduce from Miller et al. (2008), are shown in the right column of Fig. 8. For the lowest pulse rate, VS values obtained from simulations exceeded 0.98 for all firing rates. This rep- resents near perfect phase locking to the 250 pps pulse train. VS systematically decreases with pulse rate -- note the change of scale on the y-axes. Comparisons with the data of Miller et al. (2008) show good agreement be- tween simulated and measured VS values. The primary discrepancy between the two sets of VS values is the large variability present in the data values, seen as scat- ter in the vertical direction of these plots. A likely cause of this difference is the fact that we simulated a single model neuron with a fixed set of model parameters. In contrast, the data were obtained from a sampling of 37 AN fibers. Presumably each neuron had different in- trinsic properties that may even change further over the course of the experiment. Additionally, Miller et al sug- gested that some scatter in the VS values measured from responses to the 5000 pps stimulus may have been due to limitations in their ability to resolve the phase of spike times with respect to the 200 µs interpulse interval. On the basis of the lower VS values at higher pulse rates, one is tempted to conclude that these results re- veal the desynchronizing effects of high pulse rate stimu- lation. As noted in Miller et al. (2008), however, higher VS values at higher carrier pulse rates may not indicate desynchronization since VS is computed with a different reference period for each stimulus, depending on the in- 14 −8−6−4−200200400Firing Rate (spike/sec)Current (dB re Single Pulse Threshold) 250 pps1000 pps5000 pps01020051015Interspike Interval (ms)% Occurrences01020024Interspike Interval (ms)% Occurrences0102000.511.5Interspike Interval (ms)% Occurrences5000 pps1000 pps250 ppsBCDA Figure 8: Vector strength measured from responses to three pulse rates. Left column: VS computed from simulated responses to 10 second long pulse trains of biphasic pulses with equal current levels. Current levels were varied to explore a range of firing rates. Note the change of scale on the y-axis; VS values decrease as pulse rate increases. Right column: VS values reported by Miller et al. (2008) based on responses of 37 cat AN fibers stimulated with intracochlear electrodes using monopo- lar, biphasic pulse trains. Reproduced from Fig. 8 in (Miller et al. 2008) with kind permission from Springer Science+Business Media: J. Assoc. Res. Otolaryngol. The black diamonds do not represent median values, see Miller et al. (2008) for details. terpulse interval. In particular, at higher carrier pulse rates the period T in Eq. 21 is smaller, and this can lead to lower VS values for spike trains with equivalent amounts of temporal dispersion. To test whether the low VS values computed in simulations using 5000 pps pulse trains do in fact indicate desynchronizing effects of high pulse rate stimulation, we examined the distribu- tion of spike times relative to the onset time of the pulse immediately preceding the spike. These spike time dis- tributions are shown in Fig. 9 and were estimated from 10000 spike times using a 10µs bin width. Panel A shows distributions for the 250 pps (blue) and 5000 pps (green) pulse trains for a current level set to evoke approximately 100 spikes per second. In panel B, we show distribu- tions of spike times obtained from stimuli that caused the model to spike at 225 spikes per second. The spike time distributions are nearly identical at the lower firing rate indicating that the lower VS values obtained with Figure 9: Further comparison of spike timing variabil- ity in response to low and high rate pulse trains. A: Distribution of times between simulated spikes and the onset of the previous pulse. Stimuli are trains of biphasic pulses of either 250 pps (blue) or 5000 pps (green), with a constant current level per pulse. Current level was chosen so that firing rates of the model were approxi- mately 100 spikes per second for both stimuli. 10000 simulated spike times were used to estimate the distri- butions, which were plotted using 10µs bin window. B: Similar to A, with current levels increased so that firing rates in simulations were approximately 250 spikes per second. C: Similar relationship between jitter and firing efficiency in the point process model. Black curve rep- resents analytical calculation of jitter from lifetime dis- tribution function, blue circles mark the firing efficiency values, interpreted as average probability of a spike per pulse, corresponding to the pulse rates and firing rates in B and C. D: Relationship between and jitter and fir- ing efficiency in measurements of responses of cat AN fibers to monopolar, monophasic (cathodic) stimulation by an intracochlear electrode. Reprinted from Hearing Research (Miller et al. 1999) with permission from Else- vier. the 5000 pps stimulus do not reveal any desynchroniza- tion in this case. At the higher firing rate, however, the distribution of spike times in response to the 250 pps stimulus is considerably narrower than the distribution of spike times measured from the 5000 pps pulse train. One could interpret this as desynchronization, although it may be more accurate to say that the 5000 pps stimu- lus is maintaining a degree of spike time variability that is lost when the current level of the 250 pps stimulus is increased and the evoked firing rate approaches the maximal value of 250 spikes per second. We can explain the narrowing of the spike time dis- 15 Firing rate (spike/s)Vector strengthVector strengthVector strengthData (Miller et al 2008)01252500.960.98101252500.960.98102004000.50.751020040002004000.50.751020040000.5100.51Firing rate (spike/s)Simulations250 pps1000 pps5000 pps010020001020% OccurrencesTime since pulse onset (μs) 100 spike/s010020001020Time since pulse onset (μs)% Occurrences 225 spike/s050100Jitter (μs)050100Firing Efficiency (%)ACB0300200100Jitter (μs)D050100Firing Efficiency (%)250 pps5000 ppsMiller et al (1999)100 spike/s225 spike/s tribution by considering the main source of spike time variability in the model. In Fig. 9C, we show the amount of jitter in simulated spike times, as a function of firing efficiency for single pulse stimulation. We obtained this curve directly from the density function for spike times in Eq. 3. For the 250 pps stimulus, firing rates of 100 and 225 spike per second translate to firing efficiency values per pulse of 40% and 90%, respectively. The blue circles mark the location of these firing efficiency values. Jitter in the model decreases substantially at the higher firing efficiency value, which leads to the narrower spike time distribution in Fig. 9B. In contrast, increasing the fir- ing rate from 100 to 225 spikes per second when using a 5000 pps stimulus translates to a small change in the firing efficiency, here interpreted as the average proba- bility of a spike per pulse, from 2% to 4.5% (green cir- cles). The model predicts, therefore, that high rate pulse trains can maintain temporal variability in spike times even at relatively high firing rates because the probabil- ity of a spike on any single pulse remains low. The key feature of the model that explains these results -- the fact that jitter decreases with firing efficiency -- has also been observed in the responses of cat AN fibers to elec- tric stimulation using monophasic pulses (Miller et al. 1999). These data are reproduced in Fig. 9D. They show a qualitative match with our point process predictions, and underline the conclusion that the desynchronization observed at high pulse rates largely follows from the use of weaker current impulses. 3.5.3 Irregularity of firing responses An additional response statistic that has been used to characterize responses to high rate pulse trains is the Fano factor. This quantity measures irregularity in the number of observed spikes. It is defined as the variance in the number of spikes observed in a time window nor- malized by the mean value. A Fano factor of one, which is the value generated by any Poisson process model of spiking, is considered to signify a highly irregular spike generator and lower values indicate more regularity. To estimate Fano factor from the model, we simulated re- sponses to a one thousand second long train of biphasic pulses (40 µs pulse duration), using 250 pps and 5000 pps pulse trains. Ten estimates of Fano factor were obtained by subdividing these spike trains. Results are shown in Fig. 10A, with error bars that represent standard error of the ten estimates of Fano factor for each stimulus con- dition. In Fig. 10B, we reproduce a figure from (Miller et al. 2008) in order to compare the behavior of the point process model to AN fibers in cat, also stimulated by biphasic pulse trains. These are median values obtained from recordings of 37 AN fibers in 8 deafened cats, as such they do not reveal the considerable variation around the medians present in the experimental data. Fano factors obtained from simulations in response to 250 pps pulse trains are shown by the blue curve in Figure 10: Fano factor measured from responses to low (250 pps) and high (5000 pps) pulse rates. A: Simula- tion results using the point process model. Results for the low pulse rate stimulation (blue line) are predicted by the Fano factor for a binomial distribution (black line, see text for details). The gray line shows the correspond- ing prediction of a binomial distribution for the 5000 pps pulse train. The green line illustrates that simulated re- sponses to the 5000 pps pulse train can have larger Fano factors than responses to the 250 pps pulse train if the refractory period is shortened (τθ = 200 µs). Fano factor values are estimated from 100 intervals of simulated spike trains, where each interval is 100 s long. Error bars rep- resent standard error in the mean of these estimates. B: Medians of Fano factors recorded from cat auditory nerve fibers, reproduced from Fig. 9 in (Miller et al. 2008) with kind permission from Springer Science+Business Media: J. Assoc. Res. Otolaryngol. Fig. 10A and results for the high rate pulse trains are in red. The clear trend is a decrease in Fano factor as firing rate increases, consistent with the physiological data in panel B. As firing rates approach 250 spikes per second, the responses to the low rate stimulus saturate at one spike for every pulse, leading to a sharp decrease in the Fano factor, a trend also seen in the data. One sub- stantial discrepancy between simulation results and the median Fano factors reported in Miller et al. (2008) is that, in the data, Fano factors obtained from responses to high pulse rate stimuli were larger than those obtained from response to the low pulse rate stimuli for most fir- ing rates. The numbers above datapoints in Fig. 10B 16 B Data (Miller et al 2008)A Simulation 0.10.5110100Response Rate (spike/s)Fano Factor 250 pps (binomial)5000 pps (binomial)250 pps5000 pps5000 pps (short refractory)0.10.5110100Response Rate (spike/s)Fano Factor indicate the multiplicative factor by which the high rate Fano factors exceed the low rate values. These experi- mental results supported the notion that high pulse rate stimulation can generate more irregular spiking activity. Can our model predict this phenomenon? As we show in greater detail in the Appendix, the effects of refractori- ness and interpulse interactions tend to decrease Fano factor; see also (Berry II and Meister 1998). When the response of a neuron to any pulse is independent of all previous stimuli and spike history, the activity can be de- scribed as a sequence of Bernoulli trials, where the prob- ability p that any one pulse evokes a spikes is equal to the average firing rate divided by the number of pulses. The total number of spikes, in this case, would be distributed binomially with mean N p and variance N p(1− p), where N is the total number of pulses. If we denote the aver- age firing rate by r and the pulse rate by ρ, then the Fano factor would have a simple dependence on p, and, consequently, the average firing rate r: Fbinomial = 1 − p = 1 − r ρ (22) . Our simulation results for the 250 pps pulse train are completely explained by this analogy to a binomial dis- tribution. Due to the relatively long interpulse interval of 4 ms, past spike and stimulus histories have no ef- fect on the model's response to subsequent pulses. The curve representing Fbinomial in Eq. 22 for the 250 pps pulse train is shown in black in Fig. 10 and follows the simulation results. At 5000 pps, there are strong effects of stimulus and spike history, so the binomial analogy is not expected to approximate the behavior of the model. However, it does provide an upper bound for the pos- sible Fano factors that the model can achieve, which is shown by the gray line. This illustrates that there is room for the model to generate higher Fano factor values in response to high pulse rate stimulation, but different parameter values must be chosen. To test the capacity for the model to produce higher Fano factors, we weak- ened the refractory effect by decreasing the time scale of the relative refractory period, τθ in Eq. 14, to 200 µs. This value is still within the range observed in physio- logical recordings (Miller et al. 2001b). All other model parameters were kept the same. The results from sim- ulations of this model, shown in green, illustrate that it can achieve higher Fano factor values, although the dif- ference between Fano factors at the low and high pulse rates remains larger in the data. We were able to explore the relationship between re- fractory dynamics and Fano factor in the model more fully by developing a discrete-time Markov chain approx- imation to the point process model. The Markov chain framework extends the Bernoulli process analogy (intro- duced above for low pulse rate stimulation) for stimuli in which spike history and interpulse interaction effects are present. We used the Markov chain approximation to obtain estimates of firing rate and Fano factor and explore the full range of possible behaviors in the model. See the Appendix for further details. 3.6 Model predictions: amplitude-modulated pulse train stimuli Modern cochlear implant speech processors provide tem- poral information to cochlear implant listeners by mod- ulating the current levels of pulses over time. It is im- portant, therefore, to explore how the model responds to non-constant pulse trains. In order to relate model re- sults to available physiological and psychoacoustic data, we simulated responses to sinusoidally amplitude-modulated pulse trains. Specifically, as inputs to the model we used trains of biphasic pulses with the current level of the nth pulses defined by the equation: In = ¯I(1 + m sin(2πtnfm)). In is the current level of the pulse with onset at time tn, and the parameters m and fm parameterize the modula- tion depth and modulation frequency, respectively. As in previous sections, these pulses were charge balanced and had phase duration of 40 µs. To compare the model with available physiological data (Litvak et al. 2003b) and a recent modeling study (O'Gorman et al. 2010), we used a 5000 pps carrier pulse rate modulated at a frequency of 417 Hz. We simulated two mean current levels, one that evoked approximately 50 spikes per second when the pulse train was unmodulated and a second that evoked approximately 100 spikes per second. The duration of all pulse trains used in simulations was ten seconds, and the results from ten repeated simulations for each stim- ulus condition were used to compute standard errors in estimated response statistics. In Fig. 11A, we show simulated firing rates, as a func- tion of modulation depth. Firing rates increase with modulation depth for m greater than approximately 5%. We compared these results to cat single fiber data mea- sured by Litvak and colleages (Litvak et al. 2003b), which we have reproduced in Fig. 11C. The simulation results qualitatively match the experimental data, especially when compared to the two fibers that have lower firing rates (circle and filled square in C). The main discrepancy be- tween the simulated and recorded firing rates is that the model does not exhibit increased spiking until the modu- lation depth increases beyond 5%, whereas the recorded firing rates increase at modulation depths as low as 1%. To quantify the sensitivity of spike timing to the pe- riod of the modulated waveform, Litvak et al computed VS as in Eq. 21. The relevant period in this case is the period of modulation, so T =2.4ms in Eq. 21. Simula- tion results using the point process model are shown in Fig. 11C and the experimental measurements of Litvak et al. (2003b) are reproduced in panel D. The model again qualitatively captures the increase in VS with in- creasing modulation depth, although with important quan- 17 Figure 11: Responses to sinusoidally amplitude- modulated pulse trains with 417 Hz modulation fre- quency. A,C: Simulation results obtained from re- sponses to ten second long biphasic pulse trains with 40 µs pulse durations and modulation frequency given on the x-axis. Two mean current levels were used, one that produced a mean firing rate of 50 spikes per sec- ond for unmodulated input (blue) and a higher value that produced a mean firing rate of 100 spikes per sec- ond (red). Error bars represent standard error in the mean of ten repeated simulations. VS is computed with reference to the period of modulation. B,D: Firing rate and VS measures obtained from five cat AN fibers stimu- lated by intracochlear monopolar cochlear implant stim- ulation using biphasic pulse trains with duration of 25 µs per phase. Reprinted with permission from Litvak et al. (2003b), copyright Acoustical Society of America. titative differences. In particular, the experimental data show that AN fibers have exquisite sensitivity to weak modulations, producing VS values of approximately 0.5 for modulation depths of only 0.5%. As shown by O'Gorman et al. (2010), this sensitivity to weak modulations may be a consequence of nonlinear mechanisms in spike gen- erators that are accounted for in Fitzhugh-Nagumo dy- namical models, but appear to be lacking in this point process description. 3.7 Application to amplitude modulation detection In this final set of simulations, we explore how the tem- poral information in spike patterns can be quantified in a way that enables simulation results to be interpreted in the context of psychoacoustic experiments studying the ability of cochlear implant listeners to detect modu- lation at varying carrier pulse rates (Galvin III and Fu 2005; Pfingst et al. 2007; Galvin III and Fu 2009). We begin with a simple observation -- a 250 pps carrier, due to the tight phase-locking of spikes relative to the tim- Figure 12: Histograms of response to 250 pps (left col- umn) and 5000 pps (right column) pulse trains modu- lated at 75 Hz. Modulation depth increases from 1% (top row) to 10% (bottom row). Histograms were esti- mated from 10000 repeated simulations of responses to one cycle of trains of biphasic pulses, and plotted with 50 µs time bins. Current levels were set separately for each pulse rate so that unmodulated pulse trains pre- sented would evoke approximately 50 spikes per second. ing of pulses (see Fig. 8A), does not appear to evoke AN activity that represents a modulated envelope. This point is illustrated in the left column of Fig. 12. Here, we show histograms of spike times obtained from one cycle of a pulse train modulated at 75 Hz. From top to bottom, the modulation depth is increased from 1% to 10%. The current level is fixed so that 50 spikes per second are evoked, on average, in response to unmodu- lated stimuli. Our intuition tells us that this sparse and punctate pattern of spikes will not effectively transmit information about a slowly-varying envelope. Nonethe- less, these responses still carry some information about modulation depth, as evidenced, for instance, by the in- creasing height of the second peak as modulation depth increases. In the right column of Fig. 12, we show responses to a cycle of a 5000 pps pulse train modulated at 75 Hz. These histograms indicate that the high rate carrier ap- 18 SimulationsData (Litvak et al 2003)015100.51015100.51015100200400600 015100200400600Firing RateVector StrengthDBCA Modulation Depth (%)Modulation Depth (%)05100100200300# OccurencesTime (ms)5000ppsm=1%05100100200300# OccurencesTime (ms)5000ppsm=5%05100100200300# OccurencesTime (ms)5000ppsm=10%0510020004000# OccurencesTime (ms)250ppsm=1%0510020004000# OccurencesTime (ms)250ppsm=5%0510020004000# OccurencesTime (ms)250ppsm=10% Truccolo et al. 2005): L({ti}N 1 I) = N(cid:88) i=1 log(λ(tiI(ti), H(ti))− (cid:90) s λ(tI(t), H(t))dt, (23) Percent discrimination of a sinusoidally Figure 13: amplitude-modulated pulse train. Three discrimination measures are were used to discriminate unmodulated pulse train from a modulated pulse train on the basis of simulated spike trains (see text for details). Both stim- uli were one second long trains of biphasic pulses, 40µs per phase and current level was varied for each carrier pulse rate so that the unmodulated pulse train always evoked approximately 50 spikes per second, on average. Modulated pulse trains had 75 Hz modulation frequency and 1% modulation depth. Percent correct values were computed from 1000 repeated simulations of the model, error bars represent standard errors in the mean of ten repeated estimates of percent correct. Chance level is 50% correct. pears to provide a more complete representation of the envelope of the modulated stimulus. In contrast to the low rate carrier, then, we may expect that this carrier would provide greater temporal information regarding the presence of a sinusoidal modulation. In other words, we may expect improved amplitude modulation detec- tion if an observer (for instance higher processing centers in the auditory pathway) had access to this pattern of spikes as opposed to those in the left column. To test these statements quantitatively, we used an ideal observer analysis to simulate modulation detection based on spike trains generated by the point process model. A general result of point process theory allows us to use the conditional intensity function λ(tI(t), H(t)) in Eq. 4 to compute the log-likelihood of observing a sequence of spike times {ti}N i , conditioned on the input stimulus I(t) (Daley and Vere-Jones 2003; Paninski 2004; 0 where s is the total duration of the stimulus. To an- alyze how temporal modulations are encoded in spike trains, we simulated two spike trains: one in response to an unmodulated pulse train and one in response to a modulated pulse train with 75 Hz modulation frequency and 1% modulation depth. Both stimuli were one second long and consisted of biphasic pulses that were 40µs per phase. After computing the likelihood function (Eq. 23) for all possible pairings of spike trains and stimuli, we used a likelihood ratio test to discriminate between the two spike trains (Green and Swets 1966; Pillow et al. 2005; Goldwyn et al. 2010). If the true pairings of stim- ulus and response produce the highest likelihood values, then the ideal ideal observer is said to correctly detect the modulated stimulus. By repeating this procedure, we estimated percent correct detection values. We also varied the carrier pulse rate from a low carrier pulse rate (250 pps) to a high carrier pulse rate (5000 pps) to in- vestigate our previous observation that low pulse rate stimulation produces an (apparently) incomplete repre- sentation of the modulated waveform. The results of this analysis are shown by the red line in Fig. 13, where error bars indicate the standard error in the mean of 10 computed values of percent correct. The x-axis shows the carrier pulse rate used as the input to the point process model. At all pulse rates tested, for this controlled spike rate of 50 spikes per second, this maximum likelihood discrimination procedure produced correct discrimination on approximately 80% of trials. These simulations did not reveal any strong changes due to the carrier pulse rate. Despite the visible differences in Fig. 12, the spike patterns in response to modulated stimuli are equally detectable regardless of the carrier pulse rate. To further characterize the information in these spike patterns, we also computed discrimination measures us- ing decision rules that selected the modulated pulse train on the basis of which spike train response had a higher spike count (blue line) or a higher vector strength with reference to the 75Hz modulation (green line). From Fig. 11, we can see that spike count and vector strength increase with modulation depth, so simulated spike trains can encode the presence of modulation information us- ing either of these cues. For the small modulation depth used in these simulations, the spike count discrimination rule had near chance performance at just over 50% cor- rect. The vector strength discrimination measure pro- duced a higher percent correct, although it was still sub- stantially below the value obtained using the spike train 19 2505001000200050005060708090100Carrier Rate (pulse/s)Percent Correct Spike countVector strengthLikelihood likelihood technique discussed above, indicating that vec- tor strength as a measure of phase locking does not com- pletely describe how temporal modulations are expressed in the precise sequence of spike times. Similar to the likelihood function discrimination measure, neither spike count nor vector strength discrimination predicted per- cent correct values that depended strongly on the carrier pule rate. To summarize these simulations of modulation detec- tion: we found that modulation detection based on three different decision rules did not vary across a range of carrier pulse rates. This ideal observer analysis suggests that modulation detection does not require peripheral spike patterns that fully represent the sinusoidal enve- lope of these pulse trains. 4 Discussion 4.1 Review of main findings In this study, we have used point process theory to derive mathematical expressions for statistical measures of neu- ral excitability that are commonly used to characterize the response properties of AN fibers to electrical stimu- lation. Furthermore, we proposed an explicit model with features that could be related to biophysical mechanisms in a phenomenological sense. These included tempo- ral filtering that represented subthreshold dynamics of the membrane potential, a nonlinearity associated with spike generating processes, and a secondary filter that accounts for variability in spike timing. An essential feature of the proposed modeling frame- work is that all parameters can be determined from re- sponses to single and pairs of electric pulses. The model is minimal in the sense that each model parameter is uniquely identified with a single statistic reported in physiological experiments, as illustrated in Fig. 3. More- over, the point process model incorporated dynamical and stochastic properties that are relevant to high pulse rate stimulation. Specifically, the relative spread of the model depends on the time since the last spike in a man- ner consistent with data reported in (Miller et al. 2001a), and subthreshold pulses presented in rapid succession could combine to increase the excitability of the model neuron, a facilitation phenomenon observed in electrical stimulation of AN fibers (Cartee et al. 2000, 2006; Heffer et al. 2010). The construction of the model ensured that it would reproduce the measures of threshold, relative spread, jit- ter, chronaxie, as well as changes in threshold and rela- tive in the context of response to pairs of pulses separated by an interpulse interval. To further test whether our model generalizes, we simulated responses to extended stimuli of greater relevance to cochlear implant stimu- lation: pulse trains with constant current per pulse and current levels that were sinusoidally amplitude-modulated. The model produced interspike interval distributions that qualitatively agreed with data reported in (Miller et al. 2008) (see Fig. 7), and also provided important insight into how temporal jitter in spike times could account for synchronization to pulse trains presented at multi- ple pulse rates (see Fig. 8). Overall, it appears that the proposed point process framework provides a satisfactory description of AN spike trains, although there are impor- tant shortcomings and potential extensions to consider, which we discuss below. 4.2 Relation to past modeling studies and future directions The point process model represents an extension to pre- vious simplified models of AN spiking. In particular, the model developed in Bruce et al. (1999a,c) and re- lated stochastic threshold crossing models (Litvak et al. 2003b) by including temporal integration, facilitation, spike history dependent relative spread, and a realis- tic amount of spike time jitter. The simplest dynamical models that describe AN responses to electric stimula- tion are integrate and fire models (Stocks et al. 2002; Chen and Zhang 2007). As explained in Sec. 3.1, the point process model can be interpreted as a modified in- tegrate and fire model with escape noise. One significant difference between our approach and standard integrate and fire models is that we have introduced an asymmetry in the otherwise linear subthreshold dynamics, enabling the model to exhibit facilitation in response to charge balanced biphasic pulses. An advantage of the point pro- cess model over standard integrate and fire models is that spiking probabilities can be computed relatively simply using the conditional intensity function and the corre- sponding lifetime distribution function in Eq. 1. One does not need to estimate hazard rates (Plesser and Ger- stner 2000) or solve difficult first passage time problems. We have used this mathematical tractability to derive a parameterization method on the basis of physiological data. We suggest that this feature is an important strength of our model. Cochlear implants stimulate thousands of AN fibers, and response statistics such as threshold, rel- ative spread, and jitter can vary widely across the pop- ulation of AN fibers, see for instance (Miller et al. 1999, 2001b). Constructing populations of AN fibers with a representative distribution of thresholds has been done using a biophysically detailed model (Imennov and Ru- binstein 2009), but the point process framework pro- vides an explicit method for controlling a number of key response properties of model neurons. In addition to facilitating future studies that consider the response properties of populations of AN fibers, this paramet- ric control can also be used to investigate which fea- tures may impact sensory perception in a significant way. For instance, physiological data from rat AN fibers sug- 20 gest that long-term deafness can increase absolute refrac- tory periods and excitability thresholds (Shepherd et al. 2004). Such changes can be incorporated into the model via straightforward modifications of parameter values. Although the model accurately predicted responses to extended pulse trains in several ways, it was unable to quantitatively predict the precise dependence of Fano factor on pulse rate and firing rate (Fig. 10) and the exquisite sensitivity of AN fibers to weak modulations depths as small as 0.5% (Fig. 8). An explanation for these results can be found in O'Gorman et al. (2010), which showed that the Fitzhugh-Nagumo exhibits a dy- namical instability when driven with 5000 pps stimuli, and that this mechanism can account for experimentally measured values of Fano factor and strong phase locking to weak modulations reported by Litvak et al. (2003b). Future work could seek to synthesize these developments by developing models that contain this dynamical mech- anism and can also be both parameterized to additional physiological data and directly used in likelihood-based discrimination studies. One possible approach for im- proving the sensitivity of the model to small modulation is to apply an additional filter to the stimulus that acts as a differentiator on the sinusoidal evelope. Filters could also be included, for instance, to account for resonator dynamics that may affect AN responses to cochlear im- plant stimulation (Macherey et al. 2007). The analogy of subthreshold dynamics with escape noise discussed in Sec. 3.1 can provide useful insight into the effects of ad- ditional filters on the point process model. An alternative approach, that would still be grounded in the point process framework, would be to fit a model directly to a richer set of stimuli that are more relevant to the study of cochlear implant speech processing strate- gies. We have pointed out that the form of the model presented here, a cascade of linear and nonlinear stages, is similar to standard GLMs (Paninski 2004). GLMs have been shown to capture the activity of sensory neu- rons in retina with a high degree of temporal precision and can be fit from sets of spike trains recorded in re- sponse to arbitrary time-varying stimuli (Pillow et al. 2005). They have also been recently applied to auditory nerve recordings for acoustic stimulation (Trevino et al. 2010; Plourde et al. 2011). An important direction for future work, therefore, would be to fit GLMs to single unit recordings of AN fibers using stimulus sets that are clinically relevant to cochlear implant speech processing. One could also pursue a model-based approach and fit GLMs to biophysically detailed models of AN responses to cochlear implant stimluation (Imennov and Rubin- stein 2009; Woo et al. 2010). The resulting point pro- cess descriptions would not necessarily be constrained by data from single pulse and paired pulse stimuli, but the mathematical relationships in Sec. 2.2 could still be used to analyze the resulting models. Incorporating firing rate adaptation is another im- portant direction for future models of AN responses to cochlear implant stimulation. Recordings from AN fibers in deafened cats have shown that adaptation affects neu- ral responses to amplitude-modulated cochlear implant stimuli (Hu et al. 2010) and differentially affects responses to low and high rate pulse trains (Zhang et al. 2007). GLMs can exhibit firing rate adaptation through the ad- dition of longer lasting spike history effects. Compu- tational modeling can provide further insight into the origins and coding consequences of adaptation for stim- ulation strategies (Woo et al. 2010). 4.3 Implications for high pulse rate stim- ulation strategies A motivation for this work was to develop a model that could accurately predict responses to high carrier pulse rate stimulation. We have done this by including dynam- ical and stochastic features that reflect how past stim- ulation and past spiking activity can influence the ex- citability of AN fibers. Our simulation results in Fig. 12, past modeling studies (Rubinstein et al. 1999; Mino et al. 2002; Chen and Zhang 2007), and physiological evidence (Litvak et al. 2003c,b) suggest that responses to high pulse rate stimulation may represent the envelope of tem- porally modulated stimuli with greater fidelity than re- sponses evoked by low pulse rate stimulation. Moreover, one would intuitively expect high pulse rate stimulation to provide greater temporal information to cochlear im- plant listeners and thereby improve speech perception and psychophysical performance on tests of temporal res- olution. The fact that, to date, no psychoacoustic studies have shown clear evidence that high pulse rate stimu- lation improves speech reception or amplitude modula- tion detection poses an intriguing challenge to our under- standing of the connection between AN spiking activity and auditory perception. To connect AN spiking activity to psychoacoustic ex- periments, one can leverage the mathematical theory of point processes which provides access to tools from signal detection and information theory (Heinz et al. 2001a,b; Goldwyn et al. 2010). In Sec. 3.7 we used the likeli- hood function of the point process model (Eq. 23) to simulate modulation detection for a 75 Hz sinusoidally amplitude-modulated pulse train. In human listeners, modulation detection at this frequency is correlated with speech perception (Won et al. 2011), so it is of interest to understand how neural responses transmit the temporal information in these stimuli. The simulated spike trains evoked by low pulse rate stimulation were equally discriminable to those evoked by high pulse rate stimulation in our simulation of am- plitude modulation detection (Fig. 13). These results suggest the possibility that cochlear implant listeners could successfully discriminate modulated and unmod- ulated stimuli in the context of psychoacoustic exper- 21 iments, even when patterns of evoked neural activity in the auditory periphery provide distorted or incom- plete representation of the stimulus envelope. This ob- servation is consistent with psychoacoustic evidence that modulation detection in human listeners does not im- prove with high carrier pulse rate stimulation (Galvin III and Fu 2005; Pfingst et al. 2007; Galvin III and Fu 2009). Moreover, it highlights an essential challenge for improving listening outcomes for cochlear implant users -- in order to identify the relative benefits (or shortcom- ings) of high pulse rate stimulation, research should seek to identify psychoacoustic measures that reveal the per- ceptual consequences of the distinct patterns of neural activity evoked by low and high rate pulsatile stimula- tion. Acknowledgments This work has been supported by NIDCD grants F31 DC010306 (JHG), R01 DC007525 (JTR), and the Bur- roughs Wellcome Fund (ESB). Appendix: Markov chain approxi- mation to the point process model Definition of the Markov chain for con- stant pulse train stimulation In Sec. 3.5.3, we discussed how the firing rate and Fano factor produced by low pulse rate, constant current level stimulation can be estimated for the point process model by making an analogy to a Bernoulli process. In this Ap- pendix, we generalize this type of approximation for the case of high pulse rate, constant current level stimula- tion by using a Markov chain model to account for the effects of past spikes and stimulus history. Results from Markov chain theory allow us to characterize the range of firing rates and firing irregularity (Fano factor) that the model can produce for this class of stimuli. As in the Bernoulli process analogy, we simplify the problem by associating to each pulse a probability that the neuron spikes in response. This probability is given by the lifetime distribution function in Eq. 1, where the integral is evaluated over a duration of one interpulse in- terval. This probability value depends on the history of past pulses and past spikes. If we neglect precise spike timing, and presume that all spike times coincide with the onset of a pulse, then we can approximate the prob- abilities at hand via a sequence of values {pn(I)}, where the subscript n represents the number of pulses that have elapsed since the last occurrence of a spike. In the ab- sence of history effects, for instance for low pulse rate stimulation, pn(I) is identical for all pulses since the in- terpulse interval is long relative to summation and re- fractory effects. To account for history effects, we define a discrete time Markov chain. The states of this chain, which we denote sn, represent the number of pulses that have elapsed since the previous spike. There are two possible transitions away from each state. If a spike occurs, then the chain returns to s1. The transition probability of this event is pn(I). If no spike occurs, then the state of the chain advances by one to sn+1. The probability of this event is 1− pn(I). In practice, there is some number of pulses beyond which the refractory and summation effects no longer alter the probability of a spike, so we can limit the number of states in the chain to some fi- nite number N . A schematic illustration of the resulting Markov chain is shown in Fig. 14A. An example of the sequence of transition probabilities {pn(I)} is shown in Fig. 14B for a 5000 pps stimulus that produces a firing rate of 100 spikes per second. There are 25 total states in this chain (the x-axis), this indicates that history effects in the model do not persist beyond ∼ 5ms. We therefore used a 5 state Markov chain for 1000 pps stimuli and a 2 state Markov chain for 250 pps stimulation. Calculation of the firing rate and Fano fac- tor Results from the theory of discrete-time, discrete-space Markov chains allow us to analytically compute the firing rate and Fano factor of the Markov chain approximation, for arbitrary pulse rates and current levels. To do this, we use the fact that the Markov chain has a stationary distribution π which gives the long time probability that the Markov chain will be in each state. The stationary distribution can be obtained by computing the eigenvec- tor associated with the (unique) eigenvalue that takes the value one (Kemeny and Snell 1960). In other words, the stationary distribution π solves πM = π, where M is the transition matrix for the Markov chain. The first element of π, which we denote by π1, represents the long-term proportion of time steps in which the chain is in state s1. Since the chain only returns to state s1 if there has been a spike in response to the previous pulse, π1 can be used to compute the firing rate. In particular, if we denote the pulse rate of the stimulus by ρ, then the firing rate (in spikes per second) is given by: r = ρπ1 . (24) Fig. 14C compares this Markov chain approximation for firing rate (black lines) to firing rates obtained from sim- ulations of the point process model (colored error bars). The simulated firing rates are the same as shown in Fig. 7A. The fact that the Markov chain approxima- tion accurately predicts firing rates for the point process model at low pulse rates is expected because the spik- ing can be described by a Bernoulli process in this case. 22 Importantly, however, the Markov chain also accurately approximates firing rates for the high pulse rate stimuli. This framework, therefore, adequately captures the in- fluence of spike history and interpulse effects in the point process model and can be used to estimate the firing rate of the point process model via Eq. 24. Next, we show how the Markov chain approximation to the point process model can be used to compute the Fano factor of the point process model for pulse train stimuli with constant current strength. If we let N (T ) be the number of spikes produced over a period of time of length T , then the Fano factor is defined as: F = Var[N (T )] E[N (T )] . The denominator is given via the stationary distribu- tion, as discussed above. In particular, E[N (T )] = π1np, where np is the total number of pulses in the stimulus. To estimate the variance of N (T ), we first define the characteristic matrix (Kemeny and Snell 1960) Z = [Id − M + M∞] −1 where Id is the identity matrix, M is the transition ma- trix for the Markov chain, and M∞ is a matrix whose rows are the vector π. We then appeal to the law of √ large numbers for Markov chains, which states that the np is π1(2Z11 − π1 − 1) (Ke- limiting variance of N (T )/ meny and Snell 1960), where Z11 is the first entry of the characteristic matrix. To compute the Fano factor, we then take the ratio np to its expected value and √ of the variance of N (T )/ find that the Fano factor is: F = 2Z11 − π1 − 1. (25) Fig. 14D shows values of the Fano factor for 250 pps and 5000 pps stimulation computed using the equation for a range of firing rates (black lines). Fano factor values computed for the point process model are shown with colored error bars, and are the same as those in Fig. 7B. Overall, there is close agreement between the Markov chain approximation given by Eq. 25 and the Fano fac- tor values computed from simulations of the point pro- cess model. In sum, the Markov chain approximation accurately captures both the mean firing rate and the normalized variance of the spike count for a range of stimulus levels and pulse rates. Piecewise-linear approximation to the Markov chain As discussed in the text, physiological data suggest that Fano factors measured from responses to 5000 pps pulse trains may be higher than those measured from responses to 250 pps pulse trains (Miller et al. 2008). We were in- terested in identifying whether the point process model Figure 14: Markov chain approximation to the point process model. A: Schematic of the Markov chain, with pn representing the probability that the neuron spikes n pulses after the previous spike. B: Example of pn as a function of sn for 5000 pps stim- ulation set to evoke 100 spikes per second (blue). Corresponding piecewise linear approximation (red), see text for details. C: Fir- ing rates of the point process model predicted by the Markov chain model. D: Fano factor of the point process model predicted by the Markov chain model. E: Dependence of the Fano factor of the piecewise linear approximation to the Markov chain model on the number of initial and recovery stages. Contours show Fano factor values with firing rate set to be 100 spikes per second and pulse rate of 5000 pps. Gray area indicates regions where Fano factor is higher at 5000 pps than the value obtained at 250 pps, and × marks the position of the point process model with parameter values given in Table 1. 23 10100.1.51Response Rate (spike/s)Fano Factor1102000.010.020.03# pulses since spike (pn)Probability of spike (sn) Model (Fano=0.594)Linear (Fano=0.597)−8−6−4−200200400Firing Rate (spike/sec)Current (dB re Single Pulse Threshold)As1s2s3sN. . .1-p1p1p2p31-pN1-p21-p3pNB CDE51015205101520120.330.40.40.50.50.50.50.60.60.60.70.70.80.9# Initial Stages# Recovery Stages0.4250 pps1000 pps5000 pps250 pps5000 ppsInitial stagesRecovery stagesFinal stages could exhibit the same behavior, so we used the Markov chain to systematically explore the space of possible Fano factors that the model could produce. To do this, we first observed that for the 5000 pps pulse train, the pn(I) curves can be roughly caricatured as having three stages: an initial stage where the probability of spiking is near zero, a recovery stage where pn(I) increases in a rela- tively linear manner with n, and a final stage where the pn(I) is nearly constant with n. This piecewise linear approximation is shown in red in Fig. 14B. We therefore characterized the functions pn(I) by two parameters, one that describes the number of initial stages and one that describes the number of recovery stages. We then swept across this space of two parameters to explore the range of possible Fano factor values for a fixed firing rate. An example of this procedure is shown in Fig. 14E, where firing rate is set to be 100 spikes per second and the pulse rate is 5000 pps. We show Fano factor values computed using the Markov chain model with the piece- wise linear pn(I), and indicate in gray the regions of parameter space in which Fano factors are higher when stimulating the neuron with 5000 pps than with 250 pps pulse trains. The black × indicates the location of the model with parameter values given in Table 1. In order to generate Fano factors that are larger for responses to 5000 pps stimuli than 250 pps stimuli, either the num- ber of initial stages or the number of recovery stages must be decreased. One mechanism in the model that controls the number of initial and recovery stages is the time scale of the threshold refractory effect, τθ in Eq. 14. Shortening the refractory period reduces the number of initial and recovery stages, and thus our analysis of the Markov chain model provides a theoretical basis for the observation in Fig. 10 that the point process model with a smaller τθ will have higher Fano factor for 5000 pps stimulation than 250 pps stimulation. The number of initial and recovery stages is also influenced by the time scale of stimulus integration, τκ in the stimulus filters K + and K− in Eqs. 5b and 5c. The Markov chain ap- proximation is therefore a useful tool for connecting be- tween spike train statistics of the point process model to the biophysical properties (refractory effects, membrane time constant, e.g) that are represented by parameters in the model. References Y. Ahmadi, A.M. Packer, R. Yuste, and L. Paninski. Designing optimal stimuli to control neuronal spike timing. J. Neurophysiol., 106:1038 -- 1053, 2011. L. Badel, S. Lefort, T.K. Berger, C.C.H. Petersen, W. Gerstner, and M.J.E. Richardson. Extracting non-linear integrate-and-fire models from experimen- tal data using dynamic I-V curves. Biol. Cybern., 99 (4-5):361 -- 370, 2008. M.J. Berry II and M. Meister. Refractoriness and neural precision. J. Neurosci., 18(6):2200 -- 2211, 1998. S.M. Brill, W. Gstottner, J. Helms, C. Ilberg, W. Baum- gartner, J. Muller, and J. Kiefer. Optimization of channel number and stimulation rate for the fast con- tinuous interleaved sampling strategy in the COMBI 40+. Otol. Neurotol., 18(6):S104, 1997. I.C. Bruce, L.S. Irlicht, M.W. White, S.J. O'Leary, S. Dynes, E. Javel, and G.M. Clark. A stochastic model of the electrically stimulated auditory nerve: Pulse-train response. IEEE Trans. Biomed. Eng., 46 (6):630 -- 637, 1999a. I.C. Bruce, M.W. White, L. Irlicht, S.J. O'Leary, and G.M. Clark. The effects of stochastic neural activity in a model predicting intensity perception with ochlear implants: low-rate stimulation. IEEE Trans. Biomed. Eng., 46(12):1393 -- 1404, 1999b. I.C. Bruce, M.W. White, L.S. Irlicht, S.J. O'Leary, S. Dynes, E. Javel, and G.M. Clark. A stochastic model of the electrically stimulated auditory nerve: Single-pulse response. IEEE Trans. Biomed. Eng., 46 (6):617 -- 629, 1999c. A.N. Burkitt. A review of the integrate-and-fire neuron model: I. Homogeneous synaptic input. Biol. Cybern., 95:1 -- 19, 2006. L.A. Cartee. Evaluation of a model of the cochlear neu- ral membrane. II. Comparison of model and physio- logical measures of membrane properties in response to intrameatal electrical stimulation. Hear. Res., 146: 153 -- 166, 2000. L.A. Cartee, C. van den Honert, C.C. Finley, and R.L. Miller. Evaluation of a model of the cochlear neu- ral membrane. I. Physiological measurement of mem- brane characteristics in response to intrameatal elec- trical stimulation. Hear. Res., 146:143 -- 152, 2000. L.A. Cartee, C.A. Miller, and C. van den Honert. Spiral ganglion cell site of excitation I: Comparison of scala tympani and intrameatal electrode responses. Hear. Res., 215:10 -- 21, 2006. 24 F. Chen and Y.T. Zhang. An integrate-and-fire-based auditory nerve model and its response to high-rate pulse train. Neurocomput., 70(4-6):1051 -- 1055, 2007. D.J. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes. Volume I: Elementary The- ory and Methods. Probability and its Applications. Springer, New York, 2 edition, 2003. S.B.C. Dynes. Discharge characteristics of auditory nerve fibers for electrical stimuli. PhD thesis, Mas- sachusetts Institute of Technology, 1996. L.M. Friesen, R.V. Shannon, and R.J. Cruz. Effect of stimulation rate on speech recognition with cochlear implants. Audiol. Neurotol., 10:169 -- 184, 2005. J.J. Galvin III and Q.-J. Fu. Effects of stimulation rate, mode and level on modulation detection by cochlear implant users. J. Assoc. Res. Otolaryngol., 6:269 -- 279, 2005. J.J. Galvin III and Q.-J. Fu. Influence of stimulation rate and loudness growth on modulation detection and in- tensity discrimination in cochlear implant users. Hear. Res., 250:46 -- 54, 2009. W. Gerstner and W. Kistler. Spiking Neuron Models: Single Neurons, Populations, Plasticity. Cambridge University Press, Cambridge, UK, 2002. J.M. Goldberg and P.B. Brown. Response of binau- ral neurons of dog superior olivary complex to di- chotic tonal stimuli: Some physiological mechanisms of sound localization. J. Neurophysiol., 32(4):613 -- 636, 1969. J.H. Goldwyn, E. Shea-Brown, and J.T. Rubinstein. En- coding and decoding amplitude-modulated cochlear implant stimulia point process analysis. J. Comput. Neurosci., 28(3):405 -- 424, 2010. D.M. Green and J.A. Swets. Signal Detection Theory and Psychophysics. Wiley, New York, 1966. R. Hartmann, G. Topp, and R. Klinke. Discharge pat- terns of cat primary auditory fibers with electrical stimulation of the cochlea. Hear. Res., 13(1):47 -- 62, 1984. L.F. Heffer, D.J. Sly, J.B. Fallon, M.W. White, R.K. Shepherd, and S.J. O'Leary. Examining the audi- tory nerve fiber response to high rate cochlear implant stimulation: chronic sensorineural hearing loss and fa- cilitation. J. Neurophysiol., 104:3124 -- 3135, 2010. M.G. Heinz, H.S. Colburn, and L.H. Carney. Evaluat- ing auditory performance limits: I. one-parameter dis- crimination using a computational model for the au- ditory nerve. Neural Comput., 13:2273 -- 2316, 2001a. M.G. Heinz, H.S. Colburn, and L.H. Carney. Evalu- ating auditory performance limits: Ii. one-parameter discrimination with a random level variation. Neural Comput., 13:2317 -- 2339, 2001b. L.K. Holden, M.W. Skinner, T.A. Holden, and M.E. De- morest. Effects of stimulation rate with Nucleus 24 ACE speech coding strategy. Ear Hear., 23:463 -- 476, 2002. Ning Hu, Charles A. Miller, Paul J. Abbas, Barbara K. Robinson, and Jihwan Woo. Changes in auditory nerve responses across the duration of sinusoidally amplitude-modulated electric pulse-train stimuli. J. Assoc. Res. Otolaryngol., 11:641 -- 656, 2010. N.S. Imennov and J.T. Rubinstein. Stochastic popula- tion model for electrical stimulation of the auditory nerve. IEEE Trans. Biomed. Eng., 56(10):2493 -- 2501, 2009. E. Javel and R.K. Shepherd. Electrical stimulation of the auditory nerve: III. response initiation sites and temporal fine structure. Hear. Res., 140:45 -- 76, 2000. D.H. Johnson. Point process models of single-neuron discharges. J. Comput. Neurosci., 3(4):275 -- 299, 1996. F. Jones. Lebesgue integration on Euclidean space, Re- vised edition. Jones and Bartlett Publishers, Boston, MA, 2001. J.G. Kemeny and J.L. Snell. Finite Markov chains. D. Van Nostrand Company, Inc., Princeton, NJ, 1960. J. Kiefer, C. von Illberg, V. Rupprecht, J. Huber-Egener, and R. Knecht. Optimized speech understanding with the continuous interleaved sampling speech coding strategy in patients with cochlear implants: Effect of variations in stimulation rate and number of channels. Ann. Otol. Rhinol. Laryngol., 109:1009 -- 1020, 2000. L. Litvak, B. Delgutte, and D. Eddington. Auditory nerve fiber responses to electric stimulation: modu- lated and unmodulated pulse trains. J. Acoust. Soc. Amer., 110(1):368 -- 379, 2001. L. Litvak, Z.M. Smith, B. Delgutte, and D. Edding- ton. Desynchronization of electrically evoked auditory- nerve activity by high-frequency pulse trains of long duration. J. Acoust. Soc. Amer., 114(4 pt. 1):2066 -- 2078, 2003a. L.M. Litvak, B. Delgutte, and D.K. Eddington. Im- proved temporal coding of sinusoids in electric stimula- tion of the auditory nerve using desynchronizing pulse trains. J. Acoust. Soc. Amer., 114(4 pt. 1):2079 -- 2098, 2003b. 25 L.M. Litvak, B. Delgutte, and D.K. Eddington. Im- proved neural representation of vowels in electric stim- ulation using desynchronizing. J. Acoust. Soc. Amer., 114:2099 -- 2111, 2003c. H. Mino, J.T. Rubinstein, C.A. Miller, and P.J. Abbas. Effects of electrode-to-fiber distance on temporal neu- ral response with electrical stimulation. IEEE Trans. Biomed. Eng., 51(1):13 -- 20, 2004. P.C. Loizou, O. Poroy, and M. Dorman. The effect of parametric variations of cochlear implant processors on speech understanding. J. Acoust. Soc. Amer., 108: 790 -- 802, 2000. D.E. O'Gorman, J.A. White, and C.A. Shera. Dynamical instability determines the effect of ongoing noise on neural firing. J. Assoc. Res. Otolaryngol., 10(2):251 -- 267, 2009. O. Macherey, R.P. Carlyon, A. van Wieringen, and J. Wouters. A dual-process integrator-resonator model of the electrically stimulated human auditory nerve. J. Assoc. Res. Otolaryngol., 8:84 -- 104, 2007. A.J. Matsuoka, J.T. Rubinstein, P.J. Abbas, and C.A. Miller. The effects of interpulse interval on stochastic properties of electrical stimulation: Models and mea- surements. IEEE Trans. Biomed. Eng., 48(4):416 -- 424, 2001. C.A. Miller, P. Abbas, B. Robinson, J.T. Rubinstein, and A. Matsuoka. Electrically evoked single fiber action potentials from cat: responses to monopolar, monophasic stimulation. Hear. Res., 130:197 -- 218, 1999. C.A. Miller, P. Abbas, and B. Robinson. Response prop- erties of the refractory auditory nerve fiber. J. Assoc. Res. Otolaryngol., 2:216 -- 232, 2001a. C.A. Miller, B.K. Robinson, J.T. Rubinstein, P.J. Ab- bas, and C.L. Runge-Samuelson. Auditory nerve re- sponses to monophasic and biphasic electric stimuli. Hear. Res., 151:79 -- 94, 2001b. D.E. O'Gorman, H.S. Colburn, and C.A. Shera. Au- ditory sensitivity may require dynamically unstable spike generators: Evidence from a model of electri- cal stimulation. J. Acoust. Soc. Am. Express Letters, 128(5):300 -- 305, 2010. L. Paninski. Maximum likelihood estimation of cascade point-process neural encoding models. Network: Com- put Neural Syst, 15:243 -- 262, 2004. L. Paninski, J. Pillow, and J. Lewi. Statistical models for neural encoding, decoding, and optimal stimulus design. In P. Cisek, T. Drew, and J. Kalaska, editors, Computational Neuroscience: Theoretical insights into brain function, Progress in Brain Research, pages 493 -- 507. Elsevier, Amsterdam, 2007. L. Paninski, E.N. Brown, S. Iyengar, and R.E. Kass. Sta- tistical models of spike trains. In C. Laing and G.J. Lord, editors, Stochastic Methods in Neuroscience, pages 272 -- 298. Oxford University Press, 2010. D.H. Perkel, G.L. Gerstein, and G.P. Moore. Neuronal spike trains and stochastic point process: I. The single spike train. Biophys J, 7(4):391 -- 418, 1967. C.A. Miller, N. Hu, F. Zhang, B.K. Robinson, and P.J. Abbas. Changes across time in the temporal response of auditory nerve fibers stimulated by electric pulse trains. J. Assoc. Res. Otolaryngol., 9:122 -- 137, 2008. B.E. Pfingst, L. Xu, and C.S. Thompson. Effects of car- rier pulse rate and stimulation site on modulation de- tection by subjects with cochlear implants. J. Acoust. Soc. Amer., 121(4):2236 -- 2246, 2007. M.I. Miller and K.E. Mark. A statistical study of cochlear nerve discharge patterns in response to com- plex speech stimuli. J. Acoust. Soc. Amer., 92(1):202 -- 209, 1992. J.W. Pillow, L. Paninski, V.J. Uzzel, E.P. Simoncelli, and E.J. Chichilnisky. Prediction and decoding of reti- nal ganglion cell responses with a probabilistic spiking model. J Neurosci, 25(47):11003 -- 11013, 2005. H. Mino. Encoding of information into neural spike trains in an auditory nerve fiber model with electric stimuli in the presence of a pseudospontaneous activ- ity. IEEE Trans. Biomed. Eng., 54(3):360 -- 369, 2007. H. Mino and J.T. Rubinstein. Effects of neural refractori- ness on spatio-temporal variability in spike initiations with electrical stimulation. IEEE T. Neur. Sys. Reh., 14(3):273 -- 280, 2006. H.E. Plesser and W. Gerstner. Noise in integrate-and- fire neurons: From stochastic input to escape rates. Neural Comput, 12:367 -- 384, 2000. E. Plourde, B. Delgutte, and E. Brown. A point process model for auditory neurons considering both their in- trinsic dynamics and the spectrotemporal properties of an extrinsic signal. IEEE Trans. Biomed. Eng., 58: 1507 -- 1510, 2011. H. Mino, J.T. Rubinstein, and J.A. White. Comparison of algorithms for the simulation of action potentials with stochastic sodium channels. Ann. Biomed. Eng., 30(4):578 -- 87, 2002. F.P. Rattay and H. Felix. A model of the electrically excited human cochlear neuron i. contribution of neu- ral substructures to the generation and propagation of spike. Hear. Res., 153:43 -- 63, 2001. 26 R. Woloshyn. Mersenne Twister implemented in Fortran. http://www.math.sci.hiroshima-u.ac.jp/ m- mat/MT/VERSIONS/FORTRAN/mtfort90.f, Sept 1999. J.H. Won, W.R. Drennan, K. Nie, E.M. Jameyson, and J.T. Rubinstein. Acoustic temporal modulation detec- tion and speech perception in cochlear implant listen- ers. J. Acoust. Soc. Amer., 130:376 -- 388, 2011. J. Woo, C.A. Miller, and P.J. Abbas. The dependence of auditory nerve rate adaptation on electric stimulus parameters, electrode position, and fiber diameter: a computer model study. J. Assoc. Res. Otolaryngol., 11(2):283 -- 296, 2010. Y. Xu and L.M. Collins. Predictions of psychophysi- cal measurements for sinusoidal amplitude modulated (SAM) pulse-train stimuli from a stochastic model. IEEE Trans. Biomed. Eng., 54(8):1389 -- 1398, 2007. F. Zhang, C.A. Miller, B.K. Robinson, P.J. Abbas, and N. Hu. Changes across time in spike rate and spike am- plitude of auditory nerve fibers stimulated by electric pulse trains. J. Assoc. Res. Otolaryngol., 8:356 -- 372, 2007. H. Rinne. The Weibull distributrion: A handbook. CRC Press, New York, 2009. J.T. Rubinstein. Threshold fluctuations in an N -- sodium channel model of the node of Ranvier. Biophys. J., 68 (3):779 -- 785, 1995. J.T. Rubinstein, B.S. Wilson, C.C. Finley, and P.J. Ab- bas. Pseudospontaneous activity: stochastic indepen- dence of auditory nerve. Hear. Res., 127:108 -- 118, 1999. R.K. Shepherd and E. Javel. Electrical stimulation of the auditory nerve: II. effect of stimulus waveshape on single fiber response properties. Hear. Res., 130: 171 -- 188, 1999. R.K. Shepherd, L.A. Roberts, and A.G. Paolini. Long- term sensorineural hearing loss induces functional changes in the rate auditory nerve. European Jour- nal of Neuroscience, 20:3131 -- 3140, 2004. E.P. Simoncelli, L. Paninski, J. Pillow, and O. Schwartz. Characterization of neural responses with stochastic stimuli. In M. Gazzaniga, editor, The Cognitive Neuro- sciences, pages 327 -- 338. MIT Press, Cambridge, MA, 3 edition, 2004. N.G. Stocks, D. Allingham, and R.P. Morse. The ap- plication of suprathreshold stochastic resonance to cochlear implant coding. Fluct Noise Lett, 2:L169 -- 181, 2002. A. Trevino, T.P. Coleman, and J. Allen. A dynamical point process model of auditory nerve spiking in re- sponse to complex sounds. J. Comput. Neurosci., 29: 193 -- 201, 2010. W. Truccolo, U.T. Eden, M.R. Fellows, J.P. Donoghue, and E.N. Brown. A point process framework for relat- ing neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects. J. Neuro- physiol., 93:1074 -- 1089, 2005. C. van den Honert and P.H. Stypulkowski. Physiolog- ical properties of the electrically stimulated auditory nerve. II. Single fiber recordings. Hear. Res., 14(3): 225 -- 243, 1984. A.E. Vandali, L.A. Whitford, K.L. Plant, and G.M. Clark. Speech perception as a function of electrical stimulation rate: Using the Nucleus 24 cochlear im- plant system. Ear Hear., 21:608 -- 624, 2000. A.A. Verveen. On the fluctuation of threshold of the nerve fiber. In D.P. Tower and J.P. Schade, editors, Structure and function of the cerebral cortex, pages 282 -- 288. Elsevier, Amsterdam, 1960. 27
1703.03132
1
1703
2017-03-09T04:47:45
From the statistics of connectivity to the statistics of spike times in neuronal networks
[ "q-bio.NC" ]
An essential step toward understanding neural circuits is linking their structure and their dynamics. In general, this relationship can be almost arbitrarily complex. Recent theoretical work has, however, begun to identify some broad principles underlying collective spiking activity in neural circuits. The first is that local features of network connectivity can be surprisingly effective in predicting global statistics of activity across a network. The second is that, for the important case of large networks with excitatory-inhibitory balance, correlated spiking persists or vanishes depending on the spatial scales of recurrent and feedforward connectivity. We close by showing how these ideas, together with plasticity rules, can help to close the loop between network structure and activity statistics.
q-bio.NC
q-bio
From the statistics of connectivity to the statistics of spike times in neuronal networks Gabriel Koch Ocker1*, Yu Hu2*, Michael A. Buice1,3, Brent Doiron4,5, Kresimir Josi´c6,7,8, Robert Rosenbaum9, Eric Shea-Brown3,1,10 1 Allen Institute for Brain Science. 2 Center for Brain Science, Harvard University. 3 Department of Applied Mathematics, University of Washington. 4 Department of Mathematics, University of Pittsburgh. 5 Center for the Neural Basis of Cognition, Pittsburgh. 6 Department of Mathematics, University of Houston. 7 Department of Biology and Biochemistry, University of Houston. 8 Department of BioSciences, Rice University. 9 Department of Mathematics, University of Notre Dame. 10 Department of Physiology and Biophysics, and University of Washington Institute for Neuroengineering. * Equal contribution. Highlights • Remarkable new data on connectivity and activity raise the promise and raise the bar for linking structure and dynamics in neural networks. • Recent theories aim at a statistical approach, in which the enormous complexity of wiring diagrams is reduced to key features of that connectivity that drive coherent, network-wide activity. • We provide a unified view of three branches of this work, tied to a broadly useful "neural response" formula that explicitly relates connectivity to spike train statistics. • This isolates a surprisingly systematic role for the local structure and spatial scale of connectivity in determining spike correlations, and shows how the coevolution of structured connectivity and spiking statistics through synaptic plasticity can be predicted self-consistently. Abstract An essential step toward understanding neural circuits is linking their structure and their dynamics. In general, this relationship can be almost arbitrarily complex. Recent theoretical work has, however, begun to identify some broad principles underlying collective spiking activity in neural circuits. The first is that local features of network connectivity can be surprisingly effective in predicting global statistics of activity across a network. The second is that, for the important case of large networks with excitatory-inhibitory balance, correlated spiking persists or vanishes depending on the spatial scales of recurrent and feedforward connectivity. We close by showing how these ideas, together with plasticity rules, can help to close the loop between network structure and activity statistics. 1 Introduction Here, we focus on relating network connectivity to collective activity at the level of spike times, or correlations in the spike counts of cells on the timescale of typical synapses or membranes (See Box 1). Such correlations are known to have complex but potentially strong relations with coding in single neurons [1] and neural populations [2–5], and can modulate the drive to a downstream population [6]. Moreover, such correlated activity can modulate the evolution of synaptic strengths through spike timing dependent plasticity (STDP)( [7–9], but see [10]). 1 Collective spiking arises from two mechanisms: connections among neurons within a population, and external inputs or modulations affecting the entire population [11–13]. Experiments suggest that both are important. Patterns of correlations in cortical micro-circuits have been related to connection probabilities and strengths [14]. At the same time, latent variable models of dynamics applied to cortical data have revealed a strong impact of global inputs to the population [15–18]. At first, the path to understanding these mechanisms seems extremely complicated. Electron microscopy (EM) and allied reconstruction methods promise connectomes among thousands of nearby cells, tabulating an enormous amount of data [19–25]. This begs the question of what statistics of connectivity matter most – and least – in driving the important activity patterns of neural populations. The answer would give us a set of meaningful "features" of a connectome that link to basic statistical features of the dynamics that such a network produces. Our aim here is to highlight recent theoretical advances toward this goal. 2 Mechanisms and definitions: sources and descriptions of (co)variability in spike trains Neurons often appear to admit spikes stochastically. Such variability can be due to noise from, e.g., synaptic release [28], and can be internally generated via a chaotic "balanced" state [29–31]. As a consequence, the structure of spike trains is best described statistically. The most commonly used statistics are the instantaneous firing rate of each neuron, the autocorrelation function of the spike train (the probability of observing pairs of spikes in a given cell separated by a time lag s), and the cross-correlation function (likewise, for spikes generated by two different cells). As shown in Box 1, even weak correlations yield coherent, population-wide fluctuations in spiking activity that can have a significant impact on cells downstream [6]. Similarly, higher-order correlations are related to the probability of observing triplets, quadruplets or more spikes in a group of neurons, separated by a given collection of time lags. All these quantities correspond to moment densities of the spiking processes (Box 1). 3 Spike train covariability from recurrent connectivity and external input In recent years, neuroscientists have advanced a very general framework for predicting how spike train correlations (more specifically, cumulants; Box 1) depend on the structure of recurrent connectivity and external input. This framework is based on linearizing the response of a neuron around a baseline state of irregular firing. For simplicity we present the result as a matrix of spike train auto- and cross-spectra, C(ω); this is the matrix of the Fourier transforms of the familiar auto- and cross-covariance functions. Under the linear response approximation, this is C(ω) = ∆(ω)C0(ω)∆∗(ω) . (3) (cid:124) (cid:123)(cid:122) + ∆(ω)(cid:0)A(ω)Cext(ω)A∗(ω)(cid:1) ∆∗(ω) (cid:124) (cid:125) (cid:123)(cid:122) (cid:125) Internally generated Externally applied The auto- (cross-) spectra correspond to diagonal (off-diagonal) terms of C(ω). In particular, evaluated at ω = 0, these terms give the variance and co-variance of spike counts over any time window large enough to contain the underlying auto and cross-correlation functions [32, 33]. The matrix C0(ω) is diagonal, and C0 kk(ω) is the power spectrum of neuron k's spike train in the baseline state, without taking into account the effect of the recurrence on its spiking variance. ∆∗ denotes the conjugate transpose of ∆. Here, the ijth entry of the matrix ∆(ω) is called a propagator, and reflects how a spike in neuron j propagates through the network to affect activity in neuron i. When the activity can be linearized 2 Box 1. Spike train statistics. The spike train of neuron i is defined as a sum of delta functions, (cid:1) . Spike train statistics can be obtained from samples of the spike trains of each neuron yi(t) =(cid:80) k δ(cid:0)t − tk i in a population. A joint moment density of n spike trains is defined as a trial-average of products of those spike trains: mi,j,...,n(ti, . . . , tn) = 1 N yi(ti), (1) (cid:88) n(cid:89) trials i=1 (cid:90) T n(cid:89) where N is the number of trials. In practice, time is discretized into increments of size ∆t, and spike trains are binned. Equation (1) is recovered from its discrete counterpart in the limit ∆t → 0. The first spike train moment is the instantaneous firing rate of a neuron i. The second spike train moment, mij(ti, tj), is the correlation function of the two spike trains; if i = j it is an autocorrelation, otherwise a cross-correlation. It is frequently assumed that the spike trains are stationary, so that their statistics do not depend on time. We can then replace averages over trials with averages over time. Additionally, the correlation in this case only depends on the time lag in between spikes. This yields the spike train cross-correlation as, mi,j,...,n(sj, . . . , sn) = 1 T 0 j=i+1 dti yj(ti + sj), (2) where sj = tj − ti for j = i + 1, . . . , n and si=0. The correlation function measures the frequency of spike pairs. Two uncorrelated Poisson processes with rates ri and rj have mij(s) = rirj, independent of the time lag s. The statistics of any linear functional of the spike trains (such as output spike counts, or synaptic outputs or inputs) can be derived from these spike train statistics [16, 26]. Moments mix interactions of different orders. To account for lower-order contributions, we can define cumulants of the spike trains. The first cumulant and the first moment both equal the instantaneous firing rate. The second cumulant is the covariance function of the spike train: Cij(s) = mij(s) − rirj. The third spike train cumulant similarly measures the frequency of triplets of spikes, above what could be expected by composing those triplets of individual spikes and pairwise covariances. Higher order cumulants have similar interpretations [27]. around a stable "mean-field" state, the matrix of propagators obeys: ∆(ω) =(cid:0)I − K(ω)(cid:1)−1 , (4) where K(ω) = A(ω)W is the interaction matrix, which, importantly, encodes weights of synaptic connections Wij between neurons j and i. In general, the connection matrix can have a time dependence corresponding to the filtering and delay of synaptic interactions, and be written as W(ω), but we suppress this ω-dependence for ease of notation. A is a diagonal matrix; Aii(ω) is the linear response (Fourier transform of its impulse response, or PSTH [34]) of neuron i to a perturbation in its synaptic input. 3 The first term of Eq. (3) thus measures how the variability in the spiking of each neuron propagates through the network to give rise to co-fluctuations in pairs of neurons downstream. These correlations thus generated from within the population; the second term in Eq. (3) captures correlations generated by external inputs. The external inputs are described by their power spectrum, Cext ij (ω). A global modulation in the activity of many neurons due to shifts in attention, vigilance state, and/or motor activity, would result in low-rank matrix Cext. In this case the second, external term of Eq. (3) will itself be low rank, since the rank of a matrix product AB is bounded above by the ranks of A and B. Experimentally obtained spike covariance matrices can be decomposed into a low-rank and "residual" terms [35, 36] that correspond to the two terms in the matrix decomposition Eq. (3). In the simplest case of an uncoupled pair of neurons i and j receiving common inputs, Eq. (3) ij (ω)Aj(−ω) [37, 38]. The covariance of the two spike trains is thus given reduces to Cij(ω) = Ai(ω)Cext by the input covariance, multiplied by the gain with which each neuron transfers those common inputs to its output. Eq. (4) can be expanded in powers of the interaction matrix K(ω) as ∆(ω) = Km(ω). (5) ∞(cid:88) m=0 This expansion has a simple interpretation: Km are exactly m synapses long (with synaptic weights W weighted by the postsynaptic response gain A(ω)) [39, 40]. Using this expansion in Eq. (3), without external input, also provides an intuitive description of the spike train cross-spectra in terms of paths through the network, ij (ω) represents paths from a neuron i to neuron j that Cij(ω) ≈(cid:88) X(cid:88) Y(cid:88) k m=0 n=0 (cid:16) (cid:124) (cid:17) (cid:17) (cid:16) (cid:123)(cid:122) ik path terms Km(ω) Kn(−ω) C0 kk(ω). (cid:125) jk (6) This expression explicitly captures contribution to the cross-spectrum, Cij(ω), of all paths of up to X synapses ending at neuron i, and all paths of up to Y synapses ending at neuron j. The index k runs over all neurons. A first step toward the path expansion was taken by Ostojic, Brunel & Hakim, who explored the first-order truncation of Eq. (3) for two cells, thus capturing contributions from direct connections and common inputs [41]. The formulation of Eq. (3) has a rich history. Sejnowski used a similar expression in describing collective fluctuations in firing rate networks [42]. Linder, Doiron, Longtin and colleagues derived this expression in the case of stochastically driven integrate and fire neurons [43, 44], an approach generalized by Trousdale et al. [40]. Hawkes derived an equivalent expression for the case of linearly interacting point processes (now called a multivariate Hawkes process), as pointed out by Pernice et al., who applied and directly related them to neural models [39, 45, 46]. The correspondence of the rate dynamics of the Hawkes model, networks of integrate-and-fire neurons, and binary neuron models, was discussed in detail in [47] and the direct approximation of integrate-and-fire neurons by linear-nonlinear-Poisson models in [48]. Moreover, Buice and colleagues [49] developed a field theoretical method that encompasses the above approach and extends the formulation to correlations of arbitrary order. Importantly, this also allows for nonlinear interactions. An expansion can be derived via this method that describes the coupling of higher order correlations to lower moments: in particular, pairwise correlations can impact the activity predicted from mean field theory [26]. The field theoretic approach has also been applied to models of coupled oscillators related to neural networks, specifically the Kuramoto model [50, 51] and networks of "theta" neurons [52]. Similarly, Rangan developed a motif expansion of the operator governing the stationary dynamics of an integrate-and-fire network [53]. 4 Fig 1. (A) Various motifs are identified throughout the neural network. Their frequency can be measured by counting their occurrence. (B) The probability of a motif (motif moments µm,n, see text) can be decomposed into cumulants of smaller motifs. (C) Comparing the average correlation between excitatory neurons calculated using motif statistics of all orders (X, Y → ∞ in Eq. (6)) and approximations using up to second order (X + Y ≤ 2) for 512 networks (different dots) with various motif structures. Each network is composed of 80 excitatory and 20 inhibitory neurons. The deviation from the dashed line y = x shows that motif moments beyond second order are needed to accurately describe correlations. (D) Same as (C) for a motif cumulant expansion (Eq. (44) in [54], a generalization of the motif cumulant theory for networks with multiple neuron populations) truncated after second order motif terms; both panels adapted from [54]. 4 Network motifs shape collective spiking across populations The relationship of spike train cross-spectra to pathways through the network provides a powerful tool for understanding how network connectivity shapes the power spectrum of the population-wide network activity (Box 1). The population power spectrum is given by the average over the cross-spectral matrix: C(ω) = (cid:104)Cij(ω)(cid:105)i,j (the angle brackets denote averaging over pairs of neurons within a given network). Therefore C(ω) is the average of the left-hand side of Eq. (6). We show that the right-hand side of Eq. (6) can in turn be linked to the motif moments which describe the mean strength of different weighted microcircuits in the network. Assuming cellular response properties are homogeneous, the interaction matrix can be written as K(ω) = A(ω)W, where the scalar kernel A(ω) is the same response kernel for all cells. If the baseline auto-correlations, hence C0 directly proportional to the motif moments of the connectivity matrix W, defined as: k, are also equal across the network, the "path terms" appearing in Eq. (6) are (cid:68) Wm(cid:0)WT(cid:1)n(cid:69) µm,n = /N m+n+1. i,j (7) This measures the average strength of a (m, n)–motif composed of two paths of synapses emanating from a neuron k with one path of m synapses ending at neuron i, the other path of n synapses ending at neuron j. Examples of a (1, 1)-, and (1, 2)-motif are shown in Fig. 1A. For networks where Wij = 0 or 1, µm,n is also the frequency of observing a motif in the network. Eq. (6) thus provides a way to approximate how average correlations depend on the frequency of motifs in the network. An accurate approximation of the population power spectrum typically requires keeping many terms in the series (Fig. 1C). This is a challenge for applications to real neural networks, where the statistics of motifs involving many neurons are much harder to determine. Importantly, however, the contribution of these higher order motifs can often be decomposed into contributions of smaller, component motifs by introducing motif cumulants (Fig. 1 B) [54, 55]. This approach allows us to remove redundancies in motif statistics, and isolate the impact solely due to higher order motif structure. This motif cumulant 5 Truncating the motif cumulant expansion21p(connection) =p(div motif) = 1,1++++p(div motif) = p(1,2-motif) = 11,1131121,2ABCDExcitatory-Excitatory AverageExcitatory-Inhibitory AverageInhibitory-Inhibitory AverageNonlinear Resumming Theory0.1All motifsApprox.0.20.10.200.0500.05All motifsApprox.-0.04-0.02-0.04-0.02All motifsApprox.0.1Excitatory-Excitatory AverageAll motifsApprox.Excitatory-Inhibitory AverageInhibitory-Inhibitory AverageSecond OrderTruncation Theory0.20.10.200.0500.05-0.04-0.02-0.04-0.02All motifsAll motifsApprox.Approx.Truncating the motif moment expansion expansion allows Eq. (6) to be re-arranged in order to only truncate higher-order motif cumulants, rather than moments, providing a much-improved estimation of the spike train covariances (Fig. 1C vs D, [54]) that is still based only on very small motifs. In sum, population-averaged correlation can be efficiently linked to local connectivity structures described by motif cumulants. Significant motif cumulant structure exists in local networks of both cortex [56, 57] and area CA3 of the hippocampus, where they play a crucial role in pattern completion [58]. Moreover, as we see in Sec. 6, correlations in turn can shape a network's motif structure through synaptic plasticity. Higher-order correlations and network structure While we have discussed how network structure gives rise to correlations in pairs of spike trains, joint activity in larger groups of neurons, described by higher-order correlations, can significantly affect population activity [59–61]. Analogous results to Eq. (3) exist for higher-order correlations in stochastically spiking models [26, 49, 62, 63]. Network structure has been linked to the strength of third-order correlations in networks with narrow degree distributions [64], and the allied motif cumulant theory developed [65], advancing the aim of understanding how local connectivity structure impacts higher-order correlations across networks. Finally, a range of work has characterized the statistics of avalanches of neuronal activity: population bursts with power law size distributions [66], which potentially suggest a network operating near an instability [67]. Motifs and stability The results discussed so far rely on an expansion of activity around a baseline state where neurons fire asynchronously. How such states arise is a question addressed in part by the mean-field theory of spiking networks [29, 31, 68–71]. The existence of a stable stationary state depends on the structure of connectivity between neurons. In particular, when connectivity is strong and neurons have heterogeneous in-degrees, the existence of a stable mean-field solution can be lost [72, 73]. One way to rescue a stable activity regime is to introduce correlations between neurons' in and out-degrees [72]; these correspond to chain motifs (κ2 in Fig. 1B). Therefore the motifs that control correlated variability also affect the stability of asynchronous balanced states. Motif structure also affects oscillatory population activity. Roxin showed that in a rate model generating oscillations of the population activity, the variance of in-degrees (related to the strength of convergent motifs) controls the onset of oscillations [74]. Zhao et al. took a complementary approach of examining the stability of completely asynchronous and completely synchronous states, showing that two-synapse chains and convergent pairs of inputs regulate the stability of completely synchronous activity [75]. 5 Spatial scale of connectivity and inputs determines correlations in large networks Cortical neurons receive strong external and recurrent excitatory projections that would, if left unchecked, drive neuronal activity to saturated levels. Fortunately, strong recurrent inhibition balances excitation, acting to stabilize cortical activity and allow moderate firing. These large and balanced inhibitory and excitatory inputs are a major source of synaptic fluctuations, ultimately generating output spiking activity with Poisson-like variability [76, 77]. A central feature of balanced networks is that they produce asynchronous, uncorrelated spiking activity (in the limit of large networks). Original treatments of balanced networks by van Vreeswijk and Sompolinsky [29, 30] and Amit and Brunel [78] explained asynchronous activity by assuming sparse wiring within the network, so that shared inputs between neurons were negligible. While the connection probability between excitatory neurons is small [79–81], there is abundant evidence that connections between excitatory and inhibitory neurons can be quite dense [79, 82]. Renart, de la Rocha, Harris and colleagues showed that homogeneous balanced networks admit an asynchronous solution despite dense wiring [83] (for large networks). This result suggests a much deeper relationship between balance and 6 asynchronous activity than previously realized. Building on this work, Rosenbaum, Doiron and colleagues extended the theory of balanced networks to include spatially dependent connectivity [84, 85]. We review below how the spatial spread of connectivity provides new routes to correlated activity in balanced networks. Consider a two-layer network, with the second layer receiving both feedforward (F ) and recurrent (R) inputs (Fig. 2A). For simplicity we assume that the feedforward and recurrent projections have Gaussian profiles with widths σF and σR, respectively. Each neuron receives the combined input I = F + R, decomposing the input covariance CII (d) to a representative pair of layer two neurons separated by a distance d as: CII (d) = CF F (d) + CRR(d) + 2CRF (d). (8) Here CF F and CRR are the direct covariance contributions from feedforward and recurrent pathways, respectively, while CRF is the indirect contribution to covariance from the recurrent pathway tracking the feedforward pathway. If the network coupling is dense and layer one neurons are uncorrelated with one another then CF F , CRR and CRF are all O(1). This means that feedforward and recurrent projections are potential sources of correlations within the network. The asynchronous state requires that CII ∼ O(1/N ). This can only be true if the feedforward and recurrent correlations are balanced so that the recurrent pathways tracks and cancels the correlations due to the feedfoward pathway. If we take N → ∞ then in the asynchronous state, C(d) → 0 implies: CRF (d) = − 1 2 (CF F (d) + CRR(d)) . (9) This must be true for every distance d, and from Eq. (8) we derive [85] that the various spatial scales must satisfy: σ2 F = σ2 R + σ2 rate. (10) Here σ2 rate is the spatial scale of correlated firing within the network. The intuition here is that for cancellation at every d the spatial scale of feedforward and recurrent correlations must match one another. While the spatial scale of CF F (d) is determined only by σF , the scale of recurrent correlations is calculated from the correlated spiking activity convolved with the recurrent coupling (hence the sum σ2 R + σ2 model output that must be determined. For any solution to make sense we require that σrate > 0. This gives a compact asynchrony condition: σF > σR. In other words for feedforward and recurrent correlations to cancel, the spatial spread of feedforward projections must be larger than the spatial spread of recurrent projections. rate). While σF and σR are architectural parameters of the circuit (and hence fixed), σrate is a To illustrate how the spatial scales of connectivity control the asynchronous solution, we analyze the activity of a balanced spiking network when σF > σR is satisfied (Fig. 2B1, left). As expected, the spiking activity is roughly asynchronous with spike count correlations near zero (Fig. 2B1, middle). When we examine the contributions of the feedforward and recurrent pathways, we see that the relation in Eq. (9) is satisfied (Fig. 2B1, right). We contrast this to the case when σF < σR, violating the asynchrony condition (Fig. 2B2, left). Here, a clear signature of correlations is found: neuron that are nearby one another are positively correlated while more distant neuron pairs are negatively correlated. Indeed, the cancellation condition Eq. (9) is violated at almost all d (Fig. 2B2, right). Layer 2/3 of macaque visual area V1 is expected to have σF < σR, with long range projections within While these arguments give conditions for when the asynchronous solution will exist, it cannot give 2/3 being broader than L4 projections to L2/3 [86, 87]. Smith and Kohn collected population activity over large distances in layer 2/3 of macaque V1 [88]. When a one dimensional source of correlations is removed from the data then the model prediction is supported (Fig. 2C). an estimate of correlated activity. Rather, when either N < ∞ or σF < σR, we require the linear response formulation of Eq. (3) to give a prediction of how network correlations scale with distance (red dashed in Fig. 2B1 and B2, middle). It is interesting to note that a majority of the fluctuations are internally generated within the balanced circuit and subsequently have a rich spectrum of timescales. For spikes counted over long windows, the linear response formalism of Eq. (3) once again predicts the underlying correlations ("theory" curves). 7 Fig 2. Correlated activity in balanced networks with spatially dependent connections. (A) Schematic of the two layer network. The blue/red zones denotes the spatial scale of feedforward/recurrent connectivity. (B1) Broad feedforward and narrow recurrent connectivity (left) produce an asynchronous state (middle). The asynchrony requires a cancellation of CRR(d) and CF F (d) by CRF (d) at all distances (right). (B2) Narrow feedforward and broad recurrent connectivity (left) produce spatially structured correlations (middle) because CRF (d) does not cancel CRR(d) and CF F (d) (right). (C) When a one dimensional latent variable is extracted and removed from the primate V1 array data, the model predictions (left) are validated (right). Panels are from [85]. 6 Joint activity drives plasticity of recurrent connectivity The structure of neuronal networks is plastic, with synapses potentiating or depressing in a way that depends on pre- and postsynaptic spiking activity [89]. When synaptic plasticity is slow compared to the timescales of spiking dynamics, changes in synaptic weights are linked to the statistics of the spike trains [90]: specifically, the joint moment densities of the pre- and postsynaptic spike trains [91] (Box 1). For plasticity rules based on pairs of pre- and postsynaptic spikes, this results in a joint evolution of the weight matrix W, the firing rates, (cid:126)r, and the cross-covariances, C(s) (Fig. 3A-C). As we saw above, spike train cumulants depend on the network structure. In the presence of plasticity mechanisms, the structure of neuronal networks thus controls its own evolution – both directly by generating correlations [8, 9] and indirectly, by filtering the correlations inherited from external sources [92, 93]. Recent work has leveraged this connection to determine how particular structural motifs shape spike train correlations to drive plasticity [9]. A further step is to close the loop on motifs, leveraging approximations of the true spike-train correlations in order to predict the plasticity dynamics of motif cumulants [8, 93]: (cid:126)κ = F(cid:0)m(cid:0)(cid:126)κ, Cext(cid:1)(cid:1) , d dt (11) where (cid:126)κ represents a chosen set of motif cumulants and the form of the function F depends on the plasticity model used. Such analysis reveals that under an additive, pair-based plasticity rule where pre-post pairs cause potentiation and post-pre pairs cause depression, an unstructured weight matrix (with zero motif cumulants) is unstable: motifs will spontaneously potentiate or depress, creating structure in the synaptic weights (Fig. 3D), [8]. So far, such studies have focused on plasticity driven by spike pairs, relying on the linear response theories of [39, 40, 94]. More biologically realistic plasticity models rely on multi-spike interactions and variables measuring postsynaptic voltage or calcium concentrations [7]. Theories describing higher-order spike-train and spike-voltage or spike-calcium correlations provide a new window through which to examine networks endowed with these richer plasticity rules. 8 RecurrentLayerFFwd.Layerαffwdαrecαrecdistance (a.u.)spk. count corr.000.150.3sim.theorymeandistance (a.u)spk. count corr.00.250.50.060sim.theorymean0.010.0200-11-22-3L2/3 residual corr.neuron distance (mm)Model3-44-55+electrode distance (mm)0-11-22-33-400.01Macaque V1L2/3 residual corr.5-10******AB1B2C0.001broad ffwdnarrow reccnarrow ffwdbroad recccurrent covariance0.250.5distance (a.u.)01ffwd-ffwdrec-recffwd-rectotal-10.250.5distance (a.u)01current covariance Fig 3. Spike timing-dependent plasticity gives rise to joint evolution of synaptic weights, spike train covariances, and motif statistics. Panels from Ocker et al., 2015. A) Diagram of network structure. B) Evolution of the synapses highlighted in panel A. B) Evolution of the spike train covariances between the pre- and postsynaptic cells of each synapse, predicted using the first-order truncation of Eq. (6). Shading corresponds to the time points marked with arrows below the time axis of panel B. D) Projection of the joint dynamics of two-synapse motifs into the (divergent, chain) plane under a pair-based, additive Hebbian plasticity rule. 7 Conclusion The next few years could be pivotal for the study of how network structure drives neural dynamics. Spectacular experimental methods are producing vast datasets that unite connectivity and activity data in new ways [14, 19, 21, 23]. Does our field have a theory equal to the data? We've reviewed how mathematical tools can separate the two main mechanisms giving rise to collective spiking activity – recurrent connectivity and common inputs – and how this connects with decomposition methods in data analysis [15–17, 35]. Moreover, we are beginning to understand how local and spatial structures scale up to global activity patterns, and how they drive network plasticity. Daunting challenges remain, from the role of myriad cell types to the impact of nonlinear dynamics. As the bar ratchets up ever faster, the field will be watching to see what theories manage to clear it. Acknowledgements We acknowledge the support of the NSF through grants DMS-1514743 and DMS-1056125 (ESB), DMS-1313225 and DMS-1517082 (B.D.), DMS-1517629 (KJ), DMS-1517828 (R.R.) and NSF/NIGMS-R01GM104974 (KJ), as well as the Simons Fellowships in Mathematics (ESB and KJ). GO, MB, and ESB wish to thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their 9 Synaptic Weight, W ( A/cm2)μTime (min)051015x 10−7051015x 10−7Cross−covareiance, C(s) (sp/ms)2−500500246x 10−7Time Lag, s (ms)BDA00.010.020.010.020.0300.0201000Divergent motif, ( A/cm2)κ1,1μκ2Chain motif, ( A/cm2)-10100.20.40.60.81μx 10-8x 10-7C vision, encouragement and support. 10 8 Highlighted References • ** Pernice et al., PLOS Comp. Biol., 2011. Calculates the full matrix of pairwise correlation functions either in networks of interacting Poisson (Hawkes process) neurons. Shows how network-wide correlation is related to connectivity paths and motifs. • *Trousdale et al., PLOS Comp. Biol, 2012. Calculates the linear response predictions for the full matrix of pairwise correlation functions, and their relationship to connectivity paths, in networks of integrate-and-fire neurons. • ** Jovanovic & Rotter, PLOS CB 2017. Building on their PRE 2015 paper, calculates the linear response predictions for the n-th order cumulant densities of multivariate Hawkes processes (i.e. linear-Poisson neural networks), and shows how these are related to connectivity in regular networks (those with narrow degree distributions). • ** Ocker et al, ArXiV 2016. Extends existing theories for spike correlations to allow for nonlinear interactions among inputs and firing rates, deriving series of "diagrams" that show how connectivity influences activity statistics for quadratic and higher orders firing rate responses. • ** Hu et al., J Stat Mech 2013. Decomposes the effects of higher order motifs on spike correlations into smaller ones, and show how this motif cumulant approach enables one to predict global, population-wide correlations from the statistics of local, two-synapse motifs. • * Hu et al., PRE 2014 . Formulates network motif decompositions via a combinatorial relationship between moments and cumulants, extends the theory to multi-branch motifs related to higher order correlations, and to the impact of heterogeneous connectivity. • ** Renart, de la Rocha et al. Science 2010. Shows that for balanced networks with dense and strong connectivity, an asynchronous state emerges in the large-N limit in which excitatory-inhibitory interactions dynamically cancel excitatory-excitatory and inhibitory-inhibitory correlations. This leads to spike correlations that vanish on average. Also in the large N limit, this arises from population responses that instantaneously and linearly track external inputs. • ** Tetzlaff, Helias et al., PLoS CB 2012. A highly complementary study to the more well-known work of Renart et al. Using linear response theory for finite-size integrate-and-fire networks, Tetzlaff, Helias et al. expose negative feedback loops in the dynamics of both purely inhibitory and excitatory-inhibitory networks, which give rise to the dynamical cancellation of correlations in finite-size networks. • * Helias et al., PLoS CB 2014. Building on the study of Renart, de la Rocha et al. and on their previous work, the authors disentangle correlation cancellation by excitatory-inhibitory interactions (reflected in the suppression of fluctuations in the population activity) from the tracking of external inputs. • ** Rosenbaum et al., Nature Neuroscience 2016. Building on their 2014 work in PRX, the authors extends the theory of correlations in balanced networks to systems with spatially dependent connectivity, showing that mismatches in the spatial scales of feedforward and recurrent inputs can give rise to stable average firing rates but significant spike correlations. • *Rangan, PRL 2009. Develops a diagrammatic expansion for the statistics of stationary integrate-and-fire networks with delta synapses in terms of subnetworks (i.e. motifs) via a functional representation of the full network dynamics. (See also the accompanying PRE article) • ** Zhao et al, Frontiers in Comp. Neuroscience, 2011. A pioneering study of how network motifs impact network-wide synchrony, including simulations, analytical results, and methods of generating useful networks to test theories linking structure to activity. 11 • ** Gilson et al., Biol. Cyb. 2009iv. Building on the authors' previous work on STDP in networks of Poisson neurons (Gilson et al., Biol. Cyb. 2009i-iii), uses a combination of the Hawkes theory and a mean-field calculation of synaptic weights to show that externally-generated correlations can give rise to selective connectivity in recurrent networks of Poisson neurons through spike timing-dependent plasticity. • ** Ocker et al., PLoS CB 2015. Uses the linear response theory for integrate-and-fire neurons to construct self-consistent predictions for internally-generated correlations and synaptic plasticity, and then derive a reduced dynamics for the spike timing-dependent plasticity of 2-synapse motif cumulants. This shows that an initially unstructured (Erdos-R`enyi) network connectivity is unstable under additive, Hebbian STDP unless all synapses potentiate or depress together. References 1. Dettner A, Munzberg S, Tchumatchenko T. Temporal pairwise spike correlations fully capture single-neuron information. Nature Communications. 2016;7:13805. doi:10.1038/ncomms13805. 2. Hu Y, Zylberberg J, Shea-Brown E. The Sign Rule and Beyond: Boundary Effects, Flexibility, and Noise Correlations in Neural Population Codes. PLoS Comput Biol. 2014;10(2):e1003469. doi:10.1371/journal.pcbi.1003469. 3. Moreno-Bote R, Beck J, Kanitscheider I, Pitkow X, Latham P, Pouget A. Information-limiting correlations. Nature Neuroscience. 2014;17(10):1410–1417. doi:10.1038/nn.3807. 4. Zylberberg J, Cafaro J, Turner MH, Shea-Brown E, Rieke F. Direction-Selective Circuits Shape Noise to Ensure a Precise Population Code. Neuron. 2016;89(2):369–383. doi:10.1016/j.neuron.2015.11.019. 5. Franke F, Fiscella M, Sevelev M, Roska B, Hierlemann A, Azeredo da Silveira R. Structures of Neural Correlation and How They Favor Coding. Neuron. 2016;89(2):409–422. doi:10.1016/j.neuron.2015.12.037. 6. Kumar A, Rotter S, Aertsen A. Spiking activity propagation in neuronal networks: reconciling different perspectives on neural coding. Nat Rev Neurosci. 2010;11:615–627. 7. Markram H, Gerstner W, Sjostrom PJ. A history of spike-timing-dependent plasticity. Frontiers in Synaptic Neuroscience. 2011;3:4. doi:10.3389/fnsyn.2011.00004. 8. Ocker GK, Litwin-Kumar A, Doiron B. Self-Organization of Microcircuits in Networks of Spiking Neurons with Plastic Synapses. PLoS Comput Biol. 2015;11(8):e1004458. doi:10.1371/journal.pcbi.1004458. 9. Tannenbaum NR, Burak Y. Shaping Neural Circuits by High Order Synaptic Interactions. PLOS Comput Biol. 2016;12(8):e1005056. doi:10.1371/journal.pcbi.1005056. 10. Graupner M, Wallisch P, Ostojic S. Natural Firing Patterns Imply Low Sensitivity of Synaptic Plasticity to Spike Timing Compared with Firing Rate. Journal of Neuroscience. 2016;36(44):11238–11258. doi:10.1523/JNEUROSCI.0104-16.2016. 11. Cohen MR, Kohn A. Measuring and interpreting neuronal correlations. Nat Neurosci. 2011;14(7):811–9. doi:10.1038/nn.2842. 12. McGinley MJ, Vinck M, Reimer J, Batista-Brito R, Zagha E, Cadwell CR, et al. Waking State: Rapid Variations Modulate Neural and Behavioral Responses. Neuron. 2015;87(6):1143–1161. doi:10.1016/j.neuron.2015.09.012. 12 13. Doiron B, Litwin-Kumar A, Rosenbaum R, Ocker GK, Josi´c K. The mechanics of state-dependent neural correlations. Nature Neuroscience. 2016;19(3):383–393. doi:10.1038/nn.4242. 14. Cossell L, Iacaruso MF, Muir DR, Houlton R, Sader EN, Ko H, et al. Functional organization of excitatory synaptic strength in primary visual cortex. Nature. 2015;518(7539):399–403. doi:10.1038/nature14182. 15. Ecker A, Berens P, Cotton RJ, Subramaniyan M, Denfield G, Cadwell C, et al. State Dependence of Noise Correlations in Macaque Primary Visual Cortex. Neuron. 2014;82(1):235–248. doi:10.1016/j.neuron.2014.02.006. 16. Rosenbaum R, Smith MA, Kohn A, Rubin JE, Doiron B. The spatial structure of correlated neuronal variability. Nature Neuroscience. 2017;20(1):107–114. doi:10.1038/nn.4433. 17. Goris RLT, Movshon JA, Simoncelli EP. Partitioning neuronal variability. Nature Neuroscience. 2014;17(6):858–865. doi:10.1038/nn.3711. 18. Ecker AS, Denfield GH, Bethge M, Tolias AS. On the Structure of Neuronal Population Activity under Fluctuations in Attentional State. Journal of Neuroscience. 2016;36(5):1775–1789. doi:10.1523/JNEUROSCI.2044-15.2016. 19. Lee WCA, Bonin V, Reed M, Graham BJ, Hood G, Glattfelder K, et al. Anatomy and function of an excitatory network in the visual cortex. Nature. 2016;532(7599):370–374. doi:10.1038/nature17192. 20. Kasthuri N, Hayworth KJ, Berger DR, Schalek RL, Conchello JA, Knowles-Barley S, et al. Saturated Reconstruction of a Volume of Neocortex. Cell. 2015;162(3):648–661. doi:10.1016/j.cell.2015.06.054. 21. Bock DD, Lee WCA, Kerlin AM, Andermann ML, Hood G, Wetzel AW, et al. Network anatomy and in vivo physiology of visual cortical neurons. Nature. 2011;471(7337):177–182. doi:10.1038/nature09802. 22. Kleinfeld D, Bharioke A, Blinder P, Bock DD, Briggman KL, Chklovskii DB, et al. Large-Scale Automated Histology in the Pursuit of Connectomes. Journal of Neuroscience. 2011;31(45):16125–16138. doi:10.1523/JNEUROSCI.4077-11.2011. 23. Briggman KL, Helmstaedter M, Denk W. Wiring specificity in the direction-selectivity circuit of the retina. Nature. 2011;471(7337):183–188. doi:10.1038/nature09818. 24. Helmstaedter M, Briggman KL, Turaga SC, Jain V, Seung HS, Denk W. Connectomic reconstruction of the inner plexiform layer in the mouse retina. Nature. 2013;500(7461):168–174. doi:10.1038/nature12346. 25. Mishchenko Y, Hu T, Spacek J, Mendenhall J, Harris KM, Chklovskii DB. Ultrastructural Analysis of Hippocampal Neuropil from the Connectomics Perspective. Neuron. 2010;67(6):1009–1020. doi:10.1016/j.neuron.2010.08.014. 26. Ocker GK, Josi´c K, Shea-Brown E, Buice MA. Linking structure and activity in nonlinear spiking networks. arXiv:161003828 [q-bio]. 2016;. 27. Novak J, LaCroix M. Three lectures on free probability. arXiv:12052097 [math-ph]. 2012;. 28. Faisal AA, Selen LPJ, Wolpert DM. Noise in the nervous system. Nature Reviews Neuroscience. 2008;9(4):292–303. doi:10.1038/nrn2258. 29. van Vreeswijk C, Sompolinsky H. Chaos in neuronal networks with balanced excitatory and inhibitory activity. Science (New York, NY). 1996;274(5293):1724–1726. 13 30. van Vreeswijk C, Sompolinsky H. Chaotic balanced state in a model of cortical circuits. Neural Computation. 1998;10(6):1321–1371. 31. Renart A, Rocha Jdl, Bartho P, Hollender L, Parga N, Reyes A, et al. The Asynchronous State in Cortical Circuits. Science. 2010;327(5965):587–590. doi:10.1126/science.1179850. 32. De La Rocha J, Doiron B, Shea-Brown E, Josi´c K, Reyes A. Correlation between neural spike trains increases with firing rate. Nature. 2007;448(7155):802–806. 33. Bair W, Zohary E, Newsome WT. Correlated Firing in Macaque Visual Area MT: Time Scales and Relationship to Behavior. J Neurosci. 2001;21(5):1676–1697. 34. Gabbiani F, Cox SJ. Mathematics for Neuroscientists. Academic Press; 2010. 35. Yatsenko D, Josi´c K, Ecker AS, Froudarakis E, Cotton RJ, Tolias AS. Improved Estimation and Interpretation of Correlations in Neural Circuits. PLOS Computational Biology. 2015;11(3):e1004083. doi:10.1371/journal.pcbi.1004083. 36. Ecker AS, Berens P, Cotton RJ, Subramaniyan M, Denfield GH, Cadwell CR, et al. State dependence of noise correlations in macaque primary visual cortex. Neuron. 2014;82(1):235–248. 37. de la Rocha J, Doiron B, Shea-Brown E, Josi´c K, Reyes A. Correlation between neural spike trains increases with firing rate. Nature. 2007;448(7155):802–6. doi:10.1038/nature06028. 38. Shea-Brown E, Josi´c K, de la Rocha J, Doiron B. Correlation and synchrony transfer in integrate-and-fire neurons: basic properties and consequences for coding. Phys Rev Let. 2008;100(10). 39. Pernice V, Staude B, Cardanobile S, Rotter S. How structure determines correlations in neuronal networks. PLoS Comput Biol. 2011;7(5):e1002059. doi:10.1371/journal.pcbi.1002059. 40. Trousdale J, Hu Y, Shea-Brown E, Josi´c K. Impact of Network Structure and Cellular Response on Spike Time Correlations. PLoS Computational Biology. 2012;8(3):e1002408. doi:10.1371/journal.pcbi.1002408. 41. Ostojic S, Brunel N, Hakim V. How connectivity, background activity, and synaptic properties shape the cross-correlation between spike trains. J Neurosci. 2009;29(333):10234–10253. 42. Sejnowski TJ. On the stochastic dynamics of neuronal interaction. Biological Cybernetics. 1976;22(4):203–211. doi:10.1007/BF00365086. 43. Doiron B, Lindner B, Longtin A, Maler L, Bastian J. Oscillatory activity in electrosensory neurons increases with the spatial correlation of the stochastic input stimulus. Phys Rev Let. 2004;93(4). 44. Lindner B, Doiron B, Longtin A. Theory of oscillatory firing induced by spatially correlated noise and delayed inhibitory feedback. Phys Rev E. 2005;72. 45. Hawkes AG. Spectra of some self-exciting and mutually exciting point processes. Biometrika. 1971;58(1):83–90. doi:10.1093/biomet/58.1.83. 46. Pernice V, Staude B, Cardanobile S, Rotter S. Recurrent interactions in spiking networks with arbitrary topology. Physical Review E. 2012;85(3):031916. doi:10.1103/PhysRevE.85.031916. 47. Grytskyy D, Tetzlaff T, Diesmann M, Helias M. A unified view on weakly correlated recurrent networks. Frontiers in Computational Neuroscience. 2013;7. doi:10.3389/fncom.2013.00131. 48. Ostojic S, Brunel N. From Spiking Neuron Models to Linear-Nonlinear Models. PLOS Computational Biology. 2011;7(1):e1001056. doi:10.1371/journal.pcbi.1001056. 14 49. Buice MA, Chow CC, Cowan JD. Systematic fluctuation expansion for neural network activity equations. Neural Comp. 2010;22(377-426). 50. Hildebrand EJ, Buice MA, Chow CC. Kinetic Theory of Coupled Oscillators. Physical review letters. 2007;98(5):054101. doi:10.1103/PhysRevLett.98.054101. 51. Buice MA, Chow CC. Correlations, fluctuations, and stability of a finite-size network of coupled oscillators. Physical Review E. 2007;76(3):031118. doi:10.1103/PhysRevE.76.031118. 52. Buice MA, Chow CC. Dynamic Finite Size Effects in Spiking Neural Networks. PLoS Comput Biol. 2013;9(1):e1002872. doi:10.1371/journal.pcbi.1002872. 53. Rangan AV. Diagrammatic expansion of pulse-coupled network dynamics in terms of subnetworks. Physical Review E. 2009;80(3):36101. 54. Hu Y, Trousdale J, Josi´c K, Shea-Brown E. Motif statistics and spike correlations in neuronal networks. Journal of Statistical Mechanics: Theory and Experiment. 2013;2013(03):P03012. doi:10.1088/1742-5468/2013/03/P03012. 55. Hu Y, Trousdale J, Josi´c K, Shea-Brown E. Local paths to global coherence: Cutting networks down to size. Physical Review E. 2014;89(3):032802. doi:10.1103/PhysRevE.89.032802. 56. Song S, Sjostrom PJ, Reigl M, Nelson S, Chklovskii DB. Highly nonrandom features of synaptic connectivity in local cortical circuits. PLoS Biol. 2005;3(3):e68. doi:10.1371/journal.pbio.0030068. 57. Perin R, Berger TK, Markram H. A synaptic organizing principle for cortical neuronal groups. Proceedings of the National Academy of Sciences. 2011;108(13):5419–5424. doi:10.1073/pnas.1016051108. 58. Guzman SJ, Schlogl A, Frotscher M, Jonas P. Synaptic mechanisms of pattern completion in the hippocampal CA3 network. Science. 2016;353(6304):1117–1123. doi:10.1126/science.aaf1836. 59. Ohiorhenuan IE, Mechler F, Purpura KP, Schmid AM, Hu Q, Victor JD. Sparse coding and high-order correlations in fine-scale cortical networks. Nature. 2010;466(7306):617–621. doi:10.1038/nature09178. 60. Shimazaki H, Amari Si, Brown EN, Grun S. State-Space Analysis of Time-Varying Higher-Order Spike Correlation for Multiple Neural Spike Train Data. PLOS Computational Biology. 2012;8(3):e1002385. doi:10.1371/journal.pcbi.1002385. 61. Tkacik G, Marre O, Amodei D, Schneidman E, Bialek W, Ii MJB. Searching for Collective Behavior in a Large Network of Sensory Neurons. PLOS Computational Biology. 2014;10(1):e1003408. doi:10.1371/journal.pcbi.1003408. 62. Buice MA, Chow CC. Beyond mean field theory: statistical field theory for neural networks. Journal of Statistical Mechanics (Online). 2013;2013:P03003. doi:10.1088/1742-5468/2013/03/P03003. 63. Jovanovi´c S, Hertz J, Rotter S. Cumulants of Hawkes point processes. Physical Review E. 2015;91(4):042802. doi:10.1103/PhysRevE.91.042802. 64. Jovanovi´c S, Rotter S. Interplay between Graph Topology and Correlations of Third Order in Spiking Neuronal Networks. PLOS Comput Biol. 2016;12(6):e1004963. doi:10.1371/journal.pcbi.1004963. 65. Hu Y, Josic K, Shea-Brown E, Buice M. From structure to dynamics: origin of higher-order spike correlations in network motifs; COSYNE, 2015. 15 66. Plenz D, Thiagarajan TC. The organizing principles of neuronal avalanches: cell assemblies in the cortex? Trends in Neurosciences. 2007;30(3):101–110. doi:10.1016/j.tins.2007.01.005. 67. Buice M, Cowan J. Field-theoretic approach to fluctuation effects in neural networks. Phys Rev E. 2007;75:051919. 68. Ginzburg I, Sompolinsky H. Theory of correlations in stochastic neural networks. Physical Review E. 1994;50(4):3171–3191. doi:10.1103/PhysRevE.50.3171. 69. Brunel N. Dynamics of Sparsely Connected Networks of Excitatory and Inhibitory Spiking Neurons. Journal of Computational Neuroscience. 2000;8(3):183–208. doi:10.1023/A:1008925309027. 70. Tetzlaff T, Helias M, Einevoll GT, Diesmann M. Decorrelation of neural-network activity by inhibitory feedback. PLoS Comput Biol. 2012;8(8):e1002596. doi:10.1371/journal.pcbi.1002596. 71. Helias M, Tetzlaff T, Diesmann M. The Correlation Structure of Local Neuronal Networks Intrinsically Results from Recurrent Dynamics. PLoS Comput Biol. 2014;10(1):e1003428. doi:10.1371/journal.pcbi.1003428. 72. Pyle R, Rosenbaum R. Highly connected neurons spike less frequently in balanced networks. Physical Review E. 2016;93(4). doi:10.1103/PhysRevE.93.040302. 73. Landau ID, Egger R, Dercksen VJ, Oberlaender M, Sompolinsky H. The Impact of Structural Heterogeneity on Excitation-Inhibition Balance in Cortical Networks. Neuron. 2016;0(0). doi:10.1016/j.neuron.2016.10.027. 74. Roxin A. The role of degree distribution in shaping the dynamics in networks of sparsely connected spiking neurons. Frontiers in Computational Neuroscience. 2011;5:8. doi:10.3389/fncom.2011.00008. 75. Zhao L, Beverlin BI, Netoff T, Nykamp DQ. Synchronization from second order network connectivity statistics. Frontiers in Computational Neuroscience. 2011;5:28. doi:10.3389/fncom.2011.00028. 76. Doiron B, Litwin-Kumar A. Balanced neural architecture and the idling brain. Frontiers in Computational Neuroscience. 2014;8. doi:10.3389/fncom.2014.00056. 77. Den`eve S, Machens CK. Efficient codes and balanced networks. Nature neuroscience. 2016;19(3):375–382. 78. Amit DJ, Brunel N. Model of global spontaneous activity and local structured activity during delay periods in the cerebral cortex. Cerebral Cortex. 1997;7(3):237 –252. doi:10.1093/cercor/7.3.237. 79. Oswald AMM, Doiron B, Rinzel J, Reyes AD. Spatial profile and differential recruitment of GABAB modulate oscillatory activity in auditory cortex. The Journal of Neuroscience. 2009;29(33):10321–10334. 80. Holmgren C, Harkany T, Svennenfors B, Zilberter Y. Pyramidal cell communication within local networks in layer 2/3 of rat neocortex. The Journal of physiology. 2003;551(1):139. 81. Lefort S, Tomm C, Sarria JCF, Petersen CC. The excitatory neuronal network of the C2 barrel column in mouse primary somatosensory cortex. Neuron. 2009;61(2):301–316. 82. Fino E, Yuste R. Dense inhibitory connectivity in neocortex. Neuron. 2011;69(6):1188–203. doi:10.1016/j.neuron.2011.02.025. 16 83. Renart A, de la Rocha J, Bartho P, Hollender L, Parga N, Reyes A, et al. The Asynchronous State in Cortical Circuits. Science. 2010;327(5965):587–590. doi:10.1126/science.1179850. 84. Rosenbaum R, Doiron B. Balanced Networks of Spiking Neurons with Spatially Dependent Recurrent Connections. Physical Review X. 2014;4(2):021039. doi:10.1103/PhysRevX.4.021039. 85. Rosenbaum R, Smith MA, Kohn A, Rubin JE, Doiron B. The spatial structure of correlated variability in cortical circuits. Nature Neuroscience. 2017;20:107–114. 86. Bosking WH, Zhang Y, Schofield B, Fitzpatrick D. Orientation selectivity and the arrangement of horizontal connections in tree shrew striate cortex. The Journal of neuroscience. 1997;17(6):2112–2127. 87. Lund JS, Angelucci A, Bressloff PC. Anatomical Substrates for Functional Columns in Macaque Monkey Primary Visual Cortex. Cerebral Cortex. 2003;13(1):15–24. doi:10.1093/cercor/13.1.15. 88. Smith MA, Kohn A. Spatial and temporal scales of neuronal correlation in primary visual cortex. Journal of Neuroscience. 2008;28(48):12591–12603. 89. Feldman D. The Spike-Timing Dependence of Plasticity. Neuron. 2012;75(4):556–571. doi:10.1016/j.neuron.2012.08.001. 90. Kempter R, Gerstner W, Van Hemmen JL. Hebbian learning and spiking neurons. Physical Review E. 1999;59(4):4498. 91. Gerstner W, Kistler WM. Mathematical formulations of Hebbian learning. Biological Cybernetics. 2002;87(5-6):404–415. doi:10.1007/s00422-002-0353-y. 92. Gilson M, Burkitt AN, Grayden DB, Thomas DA, Hemmen JL. Emergence of network structure due to spike-timing-dependent plasticity in recurrent neuronal networks IV: Structuring synaptic pathways among recurrent connections. Biological Cybernetics. 2009;101(5-6):427–444. doi:10.1007/s00422-009-0346-1. 93. Ocker GK, Doiron B. Training and spontaneous reinforcement of neuronal assemblies by spike timing. arXiv:160800064 [q-bio]. 2016;. 94. Hawkes A. Spectra of some self-exciting and mutually exciting point processes. Biometrika. 1971;58(1):83. 17
1708.09072
1
1708
2017-08-30T01:07:18
Continual One-Shot Learning of Hidden Spike-Patterns with Neural Network Simulation Expansion and STDP Convergence Predictions
[ "q-bio.NC", "cs.CV", "stat.ML" ]
This paper presents a constructive algorithm that achieves successful one-shot learning of hidden spike-patterns in a competitive detection task. It has previously been shown (Masquelier et al., 2008) that spike-timing-dependent plasticity (STDP) and lateral inhibition can result in neurons competitively tuned to repeating spike-patterns concealed in high rates of overall presynaptic activity. One-shot construction of neurons with synapse weights calculated as estimates of converged STDP outcomes results in immediate selective detection of hidden spike-patterns. The capability of continual learning is demonstrated through the successful one-shot detection of new sets of spike-patterns introduced after long intervals in the simulation time. Simulation expansion (Lightheart et al., 2013) has been proposed as an approach to the development of constructive algorithms that are compatible with simulations of biological neural networks. A simulation of a biological neural network may have orders of magnitude fewer neurons and connections than the related biological neural systems; therefore, simulated neural networks can be assumed to be a subset of a larger neural system. The constructive algorithm is developed using simulation expansion concepts to perform an operation equivalent to the exchange of neurons between the simulation and the larger hypothetical neural system. The dynamic selection of neurons to simulate within a larger neural system (hypothetical or stored in memory) may be a starting point for a wide range of developments and applications in machine learning and the simulation of biology.
q-bio.NC
q-bio
Continual One-Shot Learning of Hidden Spike-Patterns with Neural Network Simulation Expansion and STDP Convergence Predictions Toby Lighthearta,∗, Steven Graingera, Tien-Fu Lua aUniversity of Adelaide, School of Mechanical Engineering, Adelaide SA 5005, Australia Abstract This paper presents a constructive algorithm that achieves successful one- shot learning of hidden spike-patterns in a competitive detection task. It has previously been shown (Masquelier et al., 2008) that spike-timing-dependent plasticity (STDP) and lateral inhibition can result in neurons competitively tuned to repeating spike-patterns concealed in high rates of overall presynaptic activity. One-shot construction of neurons with synapse weights calculated as estimates of converged STDP outcomes results in immediate selective detec- tion of hidden spike-patterns. The capability of continual learning is demon- strated through the successful one-shot detection of new sets of spike-patterns introduced after long intervals in the simulation time. Simulation expansion (Lightheart et al., 2013) has been proposed as an approach to the development of constructive algorithms that are compatible with simulations of biological neural networks. A simulation of a biological neural network may have orders of magnitude fewer neurons and connections than the related biological neural systems; therefore, simulated neural networks can be assumed to be a subset of a larger neural system. The constructive algorithm is developed using simu- lation expansion concepts to perform an operation equivalent to the exchange of neurons between the simulation and the larger hypothetical neural system. The dynamic selection of neurons to simulate within a larger neural system (hypothetical or stored in memory) may be a starting point for a wide range of developments and applications in machine learning and the simulation of biology. Keywords: Constructive Neural Network, Spike-Timing-Dependent Plasticity, Spike-Pattern Detection, Continual Learning, Unsupervised Learning, Simulation Expansion ∗Corresponding author Email addresses: [email protected] (Toby Lightheart), [email protected] (Steven Grainger), [email protected] (Tien-Fu Lu) Preprint submitted to arXiv August 31, 2017 1. Introduction Neural network models and simulations are important tools used in ma- chine learning and neuroscience. The selection of the neural network size is an important step in the design and application of these tools in both fields. The development of constructive algorithms for machine learning was initially motivated by a desired to automate the selection of the size of multilayered perceptrons (e.g., Ash, 1989; Fahlman & Lebiere, 1990). Without constructive algorithms, neural network design may rely on manual trial-and-error selection of neural network structures and sizes. Methods for machine learning and models of biological learning in neural networks often focus on the modification of the weight or efficacy of synapses between neurons. The training of deep neural networks often requires significant effort to avoid overfitting training input data due to the large number of network parameters (e.g., Krizhevsky et al., 2012). Trained neural networks are also typically at risk of catastrophic forgetting if subsequently trained on new data (Goodfellow et al., 2013). Constructive algorithms may be used to incrementally increase numbers of neurons and synapses to accommodate new data. Machine learning algorithms that add neurons and synapses to neural net- works have been developed under a range of names including: constructive neu- ral networks and constructive algorithms (for a review see do Carmo Nicoletti et al., 2009), growing neural networks (e.g., Fritzke, 1995; Huang et al., 2005), evolving connectionist systems (Watts, 2009; Schliebs & Kasabov, 2013), struc- tural plasticity (Roy & Basu, 2017), and adaptive structure (Wang et al., 2017). Despite the different names, these algorithms all have two types of processes: 1) processes for performance evaluation with conditions for changing the neu- ral network structure and 2) processes for calculating and assigning parameter values for new neurons or synapses. In this article, algorithms that change the neural network structure will be referred to as constructive algorithms. When a constructive algorithm operates during or between periods of neural network operation, the neural network will be referred to as a constructive neural network. Biological neurons have time-dependent output that is commonly repre- sented using spiking neuron models (Gerstner & Kistler, 2002). A limited number of machine learning algorithms have been developed that change the structure of spiking neural networks (e.g., Wysoski et al., 2010; Roy & Basu, 2017). However, only the past work on spike-timing-dependent construction (Lightheart et al., 2013) presents an argument for the biological plausibility of simulations that include the constructive algorithm. Spike-timing-dependent construction (Lightheart et al., 2013) was presented with the concept of the simulated neurons in a neural network existing in a larger neural system with many neurons that are not simulated. This concept leads to the biologically plausible interpretation of changes in the structure of a simulated neural network as a transfer of synapses or neurons between the simulation and the set of surrounding synapses and neurons (Figure 1). 2 Figure 1: Simulation expansion and contraction may be considered a process of exchanging neurons between the set of simulated neurons and synapses and the set of surrounding neurons and synapses. Surrounding neurons may be hypothetical or may be stored in memory but not presently simulated. The dashed arrow represents the process of transferring a neuron between the simulated neuron set Nsim and the surrounding neuron set Nsur. This paper summarises assumptions and principles for constructive algo- rithms to perform simulation expansion and simulation contraction and remain compatible with simulations of biological neural networks. These principles are applied in a novel constructive algorithm that uses a simulated spiking neuron to trigger construction and calculates synapse weights based on the bimodal convergence of additive spike-timing-dependent plasticity (Song et al., 2000). The one-shot construction process is demonstrated reproducing the capability of additive spike-timing-dependent plasticity (STDP) to tune neurons to com- petitively detect hidden spike-patterns (Masquelier et al., 2008). The construc- tive algorithm also demonstrates the capability to perform continual learning of new spike-patterns introduced after long simulation times where the STDP tuning process alone fails. 2. Spike-Triggered Construction with STDP Convergence The constructive algorithm presented in this article is based on concepts of simulation expansion (Lightheart et al., 2013) to develop automatic processes for changing the structure of a simulated neural network with biologically plausible results. The processes of performance evaluation and parameter calculation of the previous algorithm for spike-timing-dependent construction (Lightheart et al., 2013) have been adapted to a simulation of competitive detection of spike-patterns through STDP (Masquelier et al., 2008). Postsynaptic neuron spike-times are predicted and construction is triggered in response to the spikes of a designated simulated neuron. The synapse weights are calculated assuming a bimodal convergence typical of additive STDP (Song et al., 2000): the most recent active presynaptic neurons up to a given number have synapse weight set to the maximum and the remaining synapses are given the minimum weight. 3 NsurNsimN 2.1. Simulation Expansion and Contraction Simulation expansion re-interprets neuron construction as a transfer of neu- rons from the larger neural system to the simulated neural network that exists within it (Figure 1). A large neural system can be defined as a set of neurons, N . The set of neurons, N , can be divided into the mutually exclusive sets of simulated neurons, Nsim, and surrounding neurons, Nsur. The expansion of the simulation to include neuron n ∈ Nsur can be expressed, Nsim ← Nsim ∪ n Nsur ← Nsur \ n, (1) (2) where ∪ is the union set operation and \ is the relative complement or difference of sets. Simulations may also be contracted through the transfer of neuron, n ∈ Nsim, from the simulated neural network to the larger neural system, Nsim ← Nsim \ n Nsur ← Nsur ∪ n. (3) (4) The prior work on simulation expansion (Lightheart et al., 2013) introduced two assumptions about the neural system and simulation. First, the larger neu- ral system is assumed to have many neurons with the simulated neural network being a small subset. This assumption will be referred to as the large surround- ing network assumption. Second, the neural system is assumed to have existed for a long time prior to the simulation; therefore, observed spike-patterns may have occurred numerous times prior to the start of the simulation and synapse weights may have converged accordingly. This assumption will be referred to as the mature network assumption. These assumptions have been used as a rationale and basis for designing a constructive algorithm that creates neurons and synapses with biologically plausible parameters. Given that the simulated neural network is a part of a larger neural system, a biologically plausible simulation should account for interactions with the sur- rounding neurons. If the simulation activity does not include an approximation of the effects of a given surrounding neuron, then that neuron should have a low probability of existing in the surrounding network. This is an important con- sideration for maintaining biological plausibility when performing construction or pruning as simulation expansion or contraction. The construction or pruning of a neuron that causes a significant change in the activity of other simulated neurons requires justification. The short-term absence of the effects of a neuron can potentially be justified as a period of stochastic failure to activate or external inhibition. In general, however, a constructive algorithm designed to perform simulation expansion and contraction should aim have a plausible effect on the simulation activity. This is referred to as the principle of plausible effects in the remainder of this paper. The large surrounding network and mature network assumptions and the principle of plausible effects are used in the development of the spike-triggered constructive algorithm with assumed STDP convergence. 4 Figure 2: A simulated neuron nproxy represents activity in a subset of the surrounding network Nsur. A spike from the proxy neuron triggers the exchange (or construction) of a neuron into the simulation Nsim. 2.2. Spike-Based Performance Evaluation Previous work in spike-timing-dependent construction (Lightheart et al., 2013) performed unsupervised construction in the event that the number of active presynaptic neurons in a time-window exceeded a threshold. The specifi- cation of a threshold number of active presynaptic neurons and a time-window length may be interpreted as an approximation of the conditions that cause a postsynaptic neuron to spike. The time at which this threshold was exceeded was used as a prediction of the postsynaptic neuron spike-time for parameter calculation processes. This paper introduces a novel unsupervised method for evaluating the neural network performance and controlling construction: the simulation of a neuron as a proxy for surrounding neurons (Figure 2). The activation of this proxy neuron is used to predict the postsynaptic neuron spike-time of a surrounding neuron and triggers the process of calculating new synapse weights. The activity of the proxy neuron, and therefore construction, is controlled by the weights of its input synapses and its threshold. Uneven synapse weights would bias proxy neuron activity toward input from neurons with higher synapse weights. In the simulations presented a single proxy neuron is used to trigger construction with constant uniform synapse weights. The constructive algorithm processes have been developed for simulation conditions presented in a study of the competitive detection of spike-patterns resulting from STDP and lateral inhibition (Masquelier et al., 2008). The proxy neuron is assigned parameter values compatible with those used in this past study. The synapse weights of simulated neurons are constrained to the range w ∈ [0, 1] = [wmin, wmax]. The synapse weights of the proxy neuron are set to the expected value of a uniform distribution in this range, that is, wproxy = 0.5. The proxy neuron threshold is assigned the value provided for all postsynaptic 5 NsurNsimNnproxy neurons in the simulations performed, θproxy = θ = 550. In the neural network simulations performed, the desired result of compet- itive learning is that each neuron tunes to a different patterns of input spikes; therefore, construction should not be triggered if a neuron already responds to a pattern. The activity of the proxy neuron may be controlled to prevent construction being triggered immediately after another simulated postsynaptic neuron spikes. Lateral inhibition was used to reduce the likelihood of postsynap- tic neurons responding to the same spike-pattern in the simulation reproduced in this paper (Masquelier et al., 2008). A preliminary investigation, however, showed that the combination of high average rate of presynaptic spikes (64 Hz), the synapse weights and the threshold made lateral inhibition insufficient to pre- vent proxy neuron activation and redundant construction (results not shown). Instead a period of absolute inhibition of the proxy neuron was used to prevent construction immediately following a simulated postsynaptic neuron spike. The absolute inhibition of the proxy neuron does not prevent neuron con- struction immediately before another simulated postsynaptic neuron spikes. The two neurons will be tuned to approximately the same pattern resulting in an undesired redundancy. In addition to an absolute inhibition period after a postsynaptic neuron spike, neuron construction is cancelled by the spiking of another postsynaptic neuron within a time threshold. A grid search found that 15 ms absolute inhibition time and a 15 ms construction cancellation time threshold effectively prevented the construction of redundant neurons in the sim- ulation conditions tested without preventing neurons from being constructed to detect different sub-samples of extended spike-patterns. Continuous neural network simulations may have periods where the input activity is not associated with any significant spike-pattern. In ideal conditions the majority of the input activity will be associated with spike-patterns of in- terest. Low input neuron activity will indicate the absence of any significant spike-pattern and the parameters of the proxy neuron could be chosen to remain silent for these conditions. If the overall level of presynaptic activity does not adequately distinguish a significant spike-patterns from background noise (such as in Masquelier et al., 2008), it may be necessary to perform construction regu- larly to ensure that concealed spike-patterns are detected. To prevent excessive growth in the neural network size ineffective neurons may be pruned. The aim of construction is to reproduce the competitive learning result: postsynaptic neurons respond to a specific spike-pattern and remain inactive at other times. A neuron constructed in the absence of a regular spike-pattern will be expected to rarely spike. Simulating these neurons uses computational resources and does not provide a clear benefit to the simulation behaviour; therefore, these neurons can be pruned or contracted. It is assumed that a spike-pattern associated with a significant event will initially repeat a substantial number of times. Given this assumption the following condition for pruning has been introduced: if a constructed neuron does not spike five times in the 5 s following construction, the neuron is removed from the simulation. In cases when the background input activity triggers ongoing neuron con- struction of subsequently inactive neurons, the pruning rule counteracts con- 6 struction and can result in a stable number of simulated neurons. Nevertheless, a maximum number of simulated neuron or maximum number of total neurons constructed can be introduced to prevent runaway neural network expansion. In simulations, a maximum number of total neurons constructed nmax = 500 is enforced. Lateral inhibition between postsynaptic neurons is included in the simulated network model (Masquelier et al., 2008) and must be considered when perform- ing simulation expansion and contraction to ensure that implausible changes in network behaviour do not result. All postsynaptic neurons (simulated and surrounding) are assumed to be tuned to different spike-patterns with lateral inhibition largely subsiding between neuron spikes. The construction of a new neuron that responds to a different spike-pattern to the neurons is assumed to have a small, plausible effect on the activity of other simulated neurons. At the time of neuron construction lateral inhibition is not applied to postsynaptic neurons so that are free to spike and cancel the neuron construction. Postsy- naptic neurons that are constructed during a period of spike-noise are expected to have a low spike-rate (<1 Hz) and are able to be added and removed from the simulation with an insignificant effect on the simulation activity. We be- lieve that the proposed processes for neurons construction satisfy the principle of plausible effects. 2.3. Bimodal STDP Convergence Prediction Once it has been determined that construction should be performed, pro- cesses are performed to calculate the parameter values for the new neuron and synapses. The large surrounding network and mature network assumptions can inform the development of processes for calculations. In simulations that include a model of spike-timing-dependent plasticity, the synapses can be assigned val- ues from predictions of converged values of synapse weights. In previous work on spike-timing-dependent construction (Lightheart et al., 2013), Izhikevich neurons were constructed with model parameters set to those for regular spiking neurons (Izhikevich, 2003). Synapse weights were set to the maximum value if the presynaptic neuron spiked in a specified time-window before the predicted postsynaptic neuron spike-time. All other synapse weights were set to the minimum value. This represents an approximation of the con- vergence of additive STDP given that a pattern of presynaptic and postsynaptic neuron spikes was repeated numerous times. This paper introduces modifications of this constructive processes that are compatible with a reproduction of the simulation neurons tuning through STDP to competitively detect spike-patterns (Masquelier et al., 2008). The compet- itive detection simulation uses a spike-response neuron model for postsynaptic neurons and this neuron model is constructed in the simulations presented in this paper (details of the model are provided in Section 3.3). The past simulation of the competitive detection of hidden spike-patterns (Masquelier et al., 2008) demonstrated postsynaptic neurons initially spiking indiscriminately from the high average input neuron spike-rate of 64 Hz. Ad- ditive STDP and lateral inhibition then caused neurons to selectively respond 7 to a one of a number of repeating spike-patterns. Additive STDP has been found to produce bimodal distributions of synapse weights (Song et al., 2000) and the combination of STDP parameters weighted towards depression produce a reduction in the total weight. Assuming the network being simulated is mature, the aim of constructive processes is to produce a possible distributions of converged synapse weights from observed input neuron activity. The selection of presynaptic neurons us- ing a time-window (Lightheart et al., 2013) ensures that only recently active presynaptic neurons are selected but gives indirect control over the total input weight. A modified rule for selecting synapses has been implemented: a pre- defined number of the most recently active presynaptic neurons are selected to have the maximum weight, wmax = 1. The remaining synapses are depressed to the minimum value, wmin = 0. A grid search of total weight values found that 450 was an effective value for new neurons given the simulation conditions. Neurons with this total input weight value respond to a repetition of the hidden spike-pattern but do not respond to other activity with similar overall spike-rates (results from this grid search are provided in Appendix A). Neurons that are added to the network outside of a repeating pattern rarely spike more than one additional time as the set of active input neurons fluctuates randomly and has a low probability of returning to a similar set. The processes for evaluating performance and calculating parameter values must be integrated with the simulation of the neural network. An event-driven spiking neural network simulation has been developed as a reproduction of the past simulated study of competitive spike-pattern detection (Masquelier et al., 2008). The details of the simulation are provided in the next section. The pseudocode for the integration of the constructive algorithm with the neural network simulation is then presented in Section 3.4, Algorithm 1. 3. Neural Network Simulation Two sets of simulations are presented in this paper: a reproduction of the competitive learning and detection of hidden spike-patterns (Masquelier et al., 2008) and a modified simulation with the periodic introduction of new patterns and the elimination of periods of pure background activity. Non-expanding sim- ulations and expanding simulations are performed for both sets of simulation conditions. All simulations have a network with 2000 presynaptic neurons mod- elled as Poisson processes with random fluctuations in spike rate and activity of selected neurons irregularly replaced with noisy spike-patterns. Non-expanding simulations have nine postsynaptic neurons based on a spike-response model and simulated using an event-driven update process. Expanding simulations have the same postsynaptic neuron model but initially only simulate one proxy neuron with all other postsynaptic neurons constructed during the simulation. 8 3.1. Input Activity and Repeating Patterns Input activity is pre-generated in batches of 225 s for the 2000 presynaptic neurons modelled as Poisson processes with time-varying rate (Appendix B provides details on the generation of input spikes). In summary, a 225 s batch of Poisson input activity is generated and divided into 0.05 s segments (4500 segments in total). The base for each repeating spike-pattern is the activity of 1000 neurons from a randomly selected segment. A different segment and selection of neurons is chosen for each pattern. Segments are then selected at random to have the activity of the selected neurons replaced with a copied spike- pattern. Spike-patterns are copied with a 0 ms-mean, 1 ms-standard deviation Gaussian jitter. Finally, each input neuron has an additional 10 Hz Poisson process spike-noise overlaid. The simulation conditions of the past study of competitive detection of spike- patterns (Masquelier et al., 2008) had three spike-patterns (p = 3) that were repeated with one third of the input segments including a pattern, that is, 1/p × 1/3 × 4500 = 500 segments for each pattern in a 225 s batch. After the batch of input activity was generated, the simulation repeated the batch three times for a total simulation time of 675 s. The past study (Masquelier et al., 2008) reported that each presynaptic neuron had an average spike-rate of approximately 64 Hz. This average spike-rate was found in the reproduced presynaptic activity. Figure 3 presents an example of the input neuron activity and the change in postsynaptic spike-latency relative to the start of a pattern from plasticity. This paper presents an additional set of simulations with four spike-patterns in each batch (p = 4) and spike-pattern repetitions in all time segments, that is, each pattern pattern repeats in 1/p × 4500 = 1125 segments in a 225 s batch. Three different batches of input activity and hidden spike-patterns were generated for a single simulation for a total simulation time of 675 s. 3.2. Additive Spike-Timing-Dependent Plasticity Simulations are performed with an implementation of nearest-neighbour ad- ditive spike-timing-dependent plasticity (Brette et al., 2007; Masquelier et al., 2008; Morrison et al., 2008). As the name implies, spike-timing-dependent plas- ticity (STDP) is the plasticity or change of synapse weight that is dependent on the relative spike-timing of the presynaptic and postsynaptic neurons. The on- line nearest-neighbour STDP model implemented only performs synapse weight updates for the first postsynaptic neuron spike after a presynaptic neuron spike and for the first presynaptic neuron spike after a postsynaptic neuron spike. times of the consecutive spikes of postsynaptic neuron i are denoted t(f−1) and t(f +1) denoted t(g−1) neuron j and postsynaptic neuron spike i at time t is denoted wi,j(t). To aid the technical specification of the implemented model of STDP, the , t(f ) , and the times of the consecutive spikes of presynaptic neuron j are . The weight of a synapse between presynaptic and t(g+1) , t(g) j i i i j j A synapse receives a positive weight update for the first postsynaptic spike after a presynaptic neuron spike. If more than one presynaptic neuron spike 9 Figure 3: Example of presynaptic neuron and postsynaptic neuron activity in the reproduced simulation. A: The spike times of 100 presynaptic neurons over 0.5 s with spikes in one of the three repeating hidden patterns overlaid with a ×, (cid:7), or +. B: The spike rate of each of the 100 neurons in this 0.5 s time window. C: The average spike rate of all 2000 presynaptic neurons in 10 ms time-bins. D–F: Spike latencies of three postsynaptic neurons tuned to the same 50 ms spike-pattern over the full simulation time. Spikes outside the 50 ms duration of pattern repetitions are treated as a false positive and shown with a latency of 0 ms. 10 120406080100Presynaptic NeuronsA050100Firing Rate [Hz]BC11.11.21.31.41.5Time [s]020406080Firing Rate [Hz]01000157502550Latency [ms]D010001595Spike NumberE010001538F occurs between postsynaptic spikes, only the most recent presynaptic spike is considered for the synapse weight update. Given consecutive spike-times of presynaptic and postsynaptic neurons that increase t(f−1) i ≤ t(g+1) , then the synapse weight is updated at postsynaptic spike-time t(f ) , j < t(f ) ≤ t(g) j i i wi,j(t(f ) i ) = wi,j(t(g) j ) + A+ · exp([t(g) j − t(f ) i ]/τ+). (5) Parameter values have been replicated from the past study of competitive detec- tion of hidden spike-patterns (Masquelier et al., 2008). The amplitude param- eters for the positive and negative plasticity curves take values A+ = 0.03125 and A− = 0.85 · A+, respectively. The STDP curves have exponential decay is given by the time constants τ+ = 16.8 ms and τ− = 33.7 ms. Synapse weights are restricted to the range [0, 1]. A synapse receives a negative weight update for the first presynaptic neuron spike after the postsynaptic neuron spike. For STDP purposes, a presynaptic spike at the exact time of the postsynaptic spike is considered to occur after the postsynaptic spike. If more than one postsynaptic neuron spike occurs between presynaptic spikes, only the most recent postsynaptic spike is considered for the synapse weight update. Given consecutive spike-times of presynaptic and post- synaptic neurons that increase t(g−1) , then the synapse weight is updated at presynaptic spike-time t(g) i ≤ t(g) j < t(f +1) < t(f ) , j i j wi,j(t(g) j ) = wi,j(t(f ) i ) − A− · exp([t(f ) i − t(g) j ]/τ−). (6) 3.3. Event-Driven Postsynaptic Neuron Simulation The event-driven neural network simulation updates postsynaptic neuron variables at presynaptic spike-times rather than in uniform time-steps. The simulation of 2000 neurons spiking at an average rate of 64 Hz (Masquelier et al., 2008) will have an expected mean time between presynaptic neuron spikes of 7.8125 µs. The expected time between presynaptic neuron spikes is several orders of magnitude smaller the time constants for neuron potential (τm = 10ms and τs = 2.5ms). Therefore in the reproduced simulation, calculating postsynaptic potential updates and spikes at presynaptic spike-times is assumed to be sufficiently accurate. Simulations with fewer presynaptic neurons or less consistently high spike rates may need to approach postsynaptic neuron spike time estimation using a different method. At each presynaptic neuron spike-time, simulation variables are updated in sequence: 1. The change in simulation time is calculated. 2. Postsynaptic neuron potential variables receive exponential decay updates (Equations 7a and 7b). 3. Postsynaptic neurons with potential above the activation threshold spike: (a) For each postsynaptic neuron spike, update the potential variables of other postsynaptic neurons for lateral inhibition transmission (Equa- tions 11a and 11b). 11 (b) Spiking postsynaptic neurons have potential variables set to spike values (Equations 10a and 10b). (c) Synapses to spiking postsynaptic neurons receive a positive weight update according to the STDP model (Equation 5). 4. Synapses from the presynaptic neuron with the current spike-time receive a negative weight update according to the STDP model (Equation 6). 5. Postsynaptic neuron potential variables are increased in proportion to the weight of the synapse from the current presynaptic neuron spike (Equa- tions 11a and 11b). J This sequence is incorporated into the update equations using the infinitesimal time, , to denote times before or after the presynaptic neuron spike times, that is, t(g) are consecutive spikes in the population of presynaptic neurons J. The change in simulation time between presynaptic spikes is calculated as ∆t(g,g−1) J + . The times t(g−1) J −  and t(g) and t(g) J = t(g) J − t(g−1) J The past study (Masquelier et al., 2008) describes the spike-response neuron model as a summation of kernel functions. This neuron model has been imple- mented as two variables for the neuron potential, pm,i(t) and ps,i(t), with the decay time constants τm and τs, respectively. After the initial step of selecting the next presynaptic neuron spike as an event time the potential variables of all postsynaptic neurons i ∈ I undergo exponential decay with time constants τm and τs, J . pm,i(t(g) ps,i(t(g) J − ) = pm,i(t(g−1) J − ) = ps,i(t(g−1) J J ) · exp(−∆t(g,g−1) ) · exp(−∆t(g,g−1) J J /τm), /τs). (7a) (7b) The decay of potential is assumed to have occurred in the interval up to an J − . Note that infinitesimal time prior to the presynaptic neuron spike time, t(g) the potential variables are often different polarities and magnitudes; therefore, the exponential decay may result in an increase or decrease in the overall neuron potential (the sum of the potential variables). After calculating the decay of the potential variables the total potential of each postsynaptic neuron is calculated (8) J −. A postsynaptic neuron spike occurs if the total potential exceeds pi(t) = pm,i(t) + ps,i(t) for t = t(g) the threshold pi(t(g) J − ) > θ. (9) The threshold value θ = 550 has been replicated from the past study (Masquelier et al., 2008). Postsynaptic neuron spikes are implemented in simulation by setting the potential variables to values given by constant scaling factors and the threshold 12 value, pm,i(t(g) ps,i(t(g) J ) = (K1 − K2) · θ, J ) = K2 · θ. (10a) (10b) The scaling factors K1 = 2 and K2 = 4, have values taken from the previous study (Masquelier et al., 2008). These equations reproduce the sharp spike in the total neuron potential, simulating the neuron spike, followed by a rapid decay to below the resting membrane potential. The model in the previous work and the reproduction include an activation refractory period of 5ms that prevents the spike in postsynaptic neuron potential from immediately triggering another spike. If a postsynaptic neuron spikes in the current simulation update cycle, all the synapses to the postsynaptic neuron from presynaptic neurons that satisfy the nearest-neighbour spike condition receive a positive update according to the additive STDP model (Equation 5). Next, all synapses from the presynaptic neuron that spikes at the current simulation time to postsynaptic neurons that satisfy the nearest-neighbour spike condition received a negative weight update according to the additive STDP model (Equation 6). The weight updates take effect prior to updating the potential variables of postsynaptic neurons under the assumption that the change in synapse weight due to STDP occurs before the induced postsynaptic potential reaches its peak value. The postsynaptic potential is modelled as an alpha-function (difference of exponentials) for both the excitatory connections from the input neurons and the inhibitory lateral connections between the simulated output neurons. The transmission is modelled using the two potential variables with different time- constants. An excitatory input from a presynaptic neuron makes a contribution to the slow decaying variable (pm,i(t)) and the fast decaying variable (ps,i(t)), pm,i(t(g) ps,i(t(g) J + ) = pm,i(t(g) J + ) = ps,i(t(g) J ) + K · wi,j∈J (t(g) J ) − K · wi,j∈J (t(g) J + ), J + ). (11a) (11b) The initial value of the alpha-function is zero at the onset of a transmission; this is a result of the equal magnitude but opposite polarity of the contribution (K · wi,j∈J (t(g) J + )) to the variables for postsynaptic potential. The faster decay of the negative update to ps,i(t) results in a smooth rise and peak in the induced potential, then the potential undergoes a slow decay to zero following the decay of the positive update to pm,i(t) . For the given time constants, the scaling constant K = (4(4/3))/3 produces a peak induced potential equal to the synapse weight. The lateral inhibition is performed using the same process of updating post- synaptic neuron variables for potential; however, the inhibition has a negative peak as the lateral inhibition is modelled as a constant negative weight that is a fraction of the activation threshold value, winh = (−1/4) · θ = 137.5. After this last update step the next presynaptic neuron spike time is selected 13 and this update process is repeated. This update process is performed for all generated presynaptic neuron spike times with the simulation concluding when all presynaptic neuron spikes have been processed. 3.4. Constructive Algorithm The constructive algorithm incorporates the processes for evaluating the simulation performance and calculating parameters with the event-driven sim- ulation. The evaluation of simulation performance depends on the simulation of a proxy for surrounding postsynaptic neurons. The simulation conditions for construction can be summarised: 1. A proxy neuron spike triggers postsynaptic neuron construction. 2. A simulated postsynaptic spike inhibits proxy neuron spikes (and therefore neuron construction) for 15 ms. 3. A simulated postsynaptic spike cancels the construction of any neuron that has been constructed in the last 15 ms. 4. A constructed neuron must spike an additional 5 times in the first 5 s after construction or be pruned. 5. The maximum number of neurons that can be constructed is nmax = 500. The proxy neuron does not produce lateral inhibition and neither does a con- structed neuron at the time of construction. The synapse weight calculations for neuron construction are performed at the time of proxy neuron spikes and requires a record of the most recent presynaptic neurons to spike. The most recent 450 presynaptic neurons to spike have their synapses set to the maxi- mum weight. The remaining synapses are set to the minimum. A summary of the integration of the event-driven simulation steps and constructive algorithm processes is given in Algorithm 1. 3.5. Learning Success Criteria The past study of competitive detection of hidden spike-patterns (Masque- lier et al., 2008) considers a neuron to have achieved learning success if the percentage of true positives is greater than 90% and the rate of false positives is less than 1 Hz for a single pattern in the last 75 s of the simulation. A true positive is recorded if the postsynaptic neuron spikes within the 50 ms of the hidden spike-pattern. False positive rates are calculated from the total number of spikes outside the selected hidden spike-pattern times over the test period. In the modified simulation conditions where new patterns are introduced in every 225 s batch of simulated input activity, the true positive and false positive rates are recorded for the last 75 s of each batch in the simulation. 4. Simulations Results Simulation results are divided into sections: the first for the reproduction of a past study of competitive detection of hidden spike-patterns (Masquelier et al., 2008); and the second for modified conditions where all segments include 14 Algorithm 1 Neural network simulation with expansion and contraction update simulation time, t ← t + ∆t (event-driven) if t > tprune for any neuron then 1: while presynaptic spikes remain do 2: 3: 4: 5: 6: end if update postsynaptic neuron potential for ∆t if any simulated postsynaptic neuron spikes then remove that neuron from the simulation 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: 29: if t < tcancel for any other neuron then cancel construction of that neuron end if if postsynaptic neuron has spiked 5 times then update the pruning time of that neuron, tprune ← ∞ end if update synapse weights for postsynaptic spike (STDP) update potential of other postsynaptic neurons for lateral inhibition set spiking postsynaptic neuron potential to spike values extend proxy neuron inhibition time, tinhibit ← t + 15 ms end if if completed neuron constructions < nmax and t > tinhibit then update potential of proxy neuron for ∆t if proxy neuron spikes then calculate new synapse weights (STDP convergence prediction) add the postsynaptic neuron to the simulation set neuron construction cancellation time, tcancel ← t + 15 ms set neuron pruning time, tprune ← t + 5 s set proxy neuron refractory time, tinhibit ← t + 15 ms reset proxy neuron potential else update proxy neuron potential for input spike end if 30: 31: 32: 33: 34: end while end if update synapse weights for presynaptic spikes (STDP) update postsynaptic neuron potential for input spike 15 Figure 4: Times of neuron construction and spikes demonstrating the one-shot learning of hidden spike-patterns. Neurons are numbered in order of construction. The time of construc- tion is shown as the first spike-time (dot) for that number neuron. The grey bands represent simulation times that contain a hidden spike-pattern with the darkness of the tone indicating which of the three patterns is occurring (the lightest band is pattern 1; the darkest band is pattern 3). The spike-latencies of the final simulated neurons over the full simulation are given in Figure 5. a hidden pattern and new pattern sets are introduced over time. The presented simulation results focus on evidence for the capability one-shot learning of hid- den spike-patterns through neuron construction, the improvement in the final learning success of constructed neurons, and the rates of construction and neu- ron simulation resulting from the constructive algorithm, and the capability of continual learning of new spike-pattern sets. 4.1. Intermittent Repetition of Spike-Patterns One-hundred simulations are performed with distinct presynaptic activity and spike-patterns generated. Each set of generated presynaptic activity is applied to a network with constant structure (9 postsynaptic neurons) and a network with dynamic structure operating with the constructive algorithm. Figure 4 show an example of the early construction and activity of con- structed neurons in a simulation. When construction coincides with a hidden spike-pattern, the constructed neuron selectively responds to future repetitions of that spike-pattern with approximately the same relative spike-latency. The spike-latencies of neurons are subsequently tuned through additive STDP and typically settle into a stable sequence of spike-latencies relative to the given pattern (see Pattern 2 in Figure 5). Neurons constructed outside of hidden spike-pattern times cease to respond as the background activity changes and are removed after the 5 s pruning time elapses. 16 00.20.40.60.81Time [s]1510152025303540Constructed Neurons f o s a s e i c n e t a l e k i p S i . g n n u r p d n a n o i t c u r t s n o c r e t f a s n o r u e n c i t p a n y s t s o p d e t a l u m i s l a n fi e h t r o f s e m i t n r e t t a p o t e v i t a l e r s e i c n e t a l e k i p S : 5 e r u g i F d e t n e s e r p e r e r a s e i c n e t a l - e k i p s e h t t o l p h c a e i n h t i W . s w o r n i d e s i n a g r o s n r e t t a p - e k i p s e v i t a l e r e h t h t i w s n m u l o c n i d e s i n a g r o e r a n o r u e n e m a s e h t y c n e t a l e k i p s e h T . ) e u l a v k c i t s i x a e h t y b n e v i g s i s e k i p s f o r e b m u n l a n fi e h t ( s i x a e h t g n o l a t s a l o t t s r fi e h t m o r f g n i s a e r c n i r e b m u n e k i p s h t i w s t o d . e v i t i s o p e s l a f a s a d e d r o c e r d n a s m 0 f o y c n e t a l a n e v i g n o i t a r u d n r e t t a p s m 0 5 e h t e d i s t u o g n i r r u c c o s e k i p s h t i w s m 0 5 o t s m 0 m o r f e g n a r a n e v i g s i 17 02550Pattern 1Latency [ms]Neuron 11139002550Pattern 2Latency [ms]1139002550Pattern 3Latency [ms]11390Neuron 2114821148211482Neuron 3116171161711617Neuron 6114841148411484Neuron 7114781147811478Spike NumberNeuron 8114771147711477Neuron 28191819181918Neuron 29112751127511275Neuron 37113621136211362 The tuning of synapse weights through STDP has also been observed to produce other postsynaptic activity outcomes. In Figure 5, Neuron 3 develops an increasing rate of false positives for Pattern 1 through tuning. This increas- ing rate of false positives is a result of the total input weight increasing and a subsequent increase in the likelihood of the postsynaptic neuron spiking from background activity. Neurons that consistently spike late in the 50 ms hidden spike-pattern may have connections from a large number of neurons with reg- ular activity in the potentiating region of the STDP curve. Lateral inhibition prevents the postsynaptic spike latency from decreasing to a point where more input spikes occur after the postsynaptic spike and weights are depressed. Another uncommon result is two neurons competitively tuning to and shar- ing approximately the same spike-latency. For example, Figure 5 shows neu- rons 28 and 29 tune to Pattern 3 with average spike latencies in the final 75 s test period of 18.8 ms and 23.4 ms, respectively. Neurons that tune to similar spike-latencies for the same pattern rarely spike for the same occurrence of that spike-pattern. Neurons 28 and 29 spike in same occurrence of the pattern in only 5/166 occurrences in the test period. Individually, neurons that have such close spike latencies may have low true positive percentage (neuron 28: 34.34%; neuron 29: 68.67%); however, in the simulations performed the combined activ- ity of these neurons often detects all occurrences of the pattern in test periods. When the activity of neurons 28 and 29 are combined all occurrences of Pattern 3 are detected. Postsynaptic neurons would also sometimes exhibit less frequent activity at a consistent earlier spike-latency. This would typically occur when a postsynaptic neuron with an earlier spike-latency fails to spike. The absence of the earlier spike and associated lateral inhibition and delay of subsequent postsynaptic neurons results in earlier spike-times. There is a significant difference in the early performance (true positives per- centage and false positive rate) of neurons initialised and simulated in a static network structure and the constructed neurons (Figures 6 and 7). The neurons initialised in the static structure simulations have a high initial spike-rate and false positive rate, whereas typical constructed neurons that are not pruned have an initial false positive rate of zero and high or perfect true positive per- centages. In the first 15 s of the simulation all neurons in the static structure simulations have a false positive rate between 6.73 Hz to 9.80 Hz. In the first 15 s after neuron construction, 864/939 neurons have a false positive rate of 0 Hz with a maximum rate of 1.80 Hz. The constructed neurons with the highest false positive rates are a result of the neuron spiking a few milliseconds after the pattern segment ends. Treating spikes up to 5 ms after the pattern as a true positive increases the number of neurons with zero false positives to 895/939 and reduces the maximum false positive rate to 0.067 Hz (that is, the neuron produces a single false positive spike in the 15 s after construction). Applying this less strict definition of a true positive has a negligible effect on the range of false positive rates of neurons in static structure simulations (6.73 Hz to 9.73 Hz). High initial spike-rates of the postsynaptic neurons in static-structure simu- 18 Figure 6: The distribution of true positive percentages in the first 15 s of neuron simulation in non-expanding simulations (Top; 900 neurons) and expanding simulations (Bottom; 939 neurons). A true positive is recorded when the neuron spikes in the 50 ms duration of a pattern. Each postsynaptic neuron has a true positive percentage for each pattern with the maximum percentage for each neuron presented. The expected number of repetitions of each pattern in 15 s is 33.3; therefore, the expected increment in true positive percentage for each missed pattern is greater than the histogram bin width of 0.02. Figure 7: The distribution of false positive rates in the first 15 s of neuron simulation in non- expanding simulations (Top; 900 neurons) and expanding simulations (Bottom; 939 neurons). A false positive is recorded when the neuron spikes outside the 50 ms duration of a pattern. Each postsynaptic neuron has a false positive rate for each pattern with the minimum rate for each neuron presented. 19 100101102103NeuronsStatic Structure00.20.40.60.81True Positives [%]100101102103NeuronsConstructed100101102103NeuronsStatic Structure0246810False Positives [Hz]100101102103NeuronsConstructed lations are a product of the high initial total input weight (2000 synapses with an initial uniform distribution in [0, 1] gives an expected total input weight of 1000). This causes the neurons to spike indiscriminately in the presence of the high overall rate of presynaptic neuron activity. The spike-rate of postsynaptic neurons drops sharply once STDP sufficiently depresses the majority of synapse weights. If the remaining potentiated synapses sufficiently correspond with the active presynaptic neurons in a hidden spike-pattern the postsynaptic neuron will selectively respond to that spike-pattern. With the strict definition of true positives as spikes occurring within the 0.05 s pattern time, simulations with static structure and random weight initial- isation show 7/900 neurons achieve 100% true positives in the first 15 s of the simulation. The early simulation of networks with static structure produced a true positive percentage greater than 90% in 80/900 (8.89%) of the postsynaptic neurons. None of these neurons, however, satisfy the learning success criterion of a false positive rate requirement of less than 1 Hz. The construction of neurons with a lower total input weight eliminates the requirement of a period of synaptic depression before selectively responding to spike-patterns. For the given simulation conditions, detection of spike-patterns is successfully achieved by potentiating synapses from presynaptic neurons with the most recent spikes. The number of neurons with a true positive percentages greater than 90% in the first 15 s of neuron simulation is 819/939 (87.2%) and all of these neurons satisfy the false positive rate criterion. The majority of the unpruned constructed neurons (635/939) have a true positive rate of 100% in the first 15 s of simulated time after construction (Figure 6). The past study of spike-timing-dependent plasticity (Masquelier et al., 2008) evaluated the neuron performance in the last 75 s of the simulation time. The authors reported an average of 5.71/9 neurons (63.4% in 100 simulations) suc- cessfully learning to detect a portion of one of three hidden patterns. The reproduction of this non-expanding simulation produced a similar result (Fig- ure 8) with 100 simulations producing an average of 5.85/9 neurons (65.0%) that achieve learning success. The expanding simulations result in higher numbers of successful neurons with higher frequency. Expanding simulations performed with the same presynaptic neuron activity produced an average of 7.46 success- ful neurons with an average of 9.39 final simulated neurons (79.4% final success rate). The simulations with neuron construction and pruning on average pro- duce a greater number and final proportion of simulated neurons that achieve learning success. Comparisons of the performance of neurons in the first 15 s after neuron construction and in the test period (the final 75 s of the simulation) should be done cautiously. Early performance is evaluated over a shorter simulation time and starts from the time of neuron construction. Early simulation also includes significant numbers of simulated neurons that are yet to be pruned. Never- theless, constructive simulations show a reduction in success rate from early performance to the test period. This was an effect of STDP and competition through lateral inhibition, which was observed to produce unsuccessful postsy- naptic neurons that share approximately the same spike-latency (e.g., Neurons 20 Figure 8: Top: The distributions of successful neuron numbers for constructive and static structure simulations (one-hundred simulations each). Bottom: The distribution of the dif- ference in successful neuron numbers (constructed minus static structure) for the same presy- naptic neuron activity. Criteria for success are provided in Section 3.5. 28 and 29 in Figure 5). The distribution of neuron true positive percentages (Figure 9) and false positive rates (Figure 10) had similar features in simulations with static struc- ture and neuron construction. The most frequent true positive percentage was observed at perfect pattern detection in the test phase of both types of sim- ulations. Neurons with true positive percentage values across the full range, (0, 1), were observed. Constructive simulations, however, had no neurons with zero true positives, while simulations with a static structure had a total of 148 neurons with zero true positives in the 75 s test period. The rates of false positives in simulations with constant structure and with dynamic structure (Figure 10) show some differences in trends and possible outcomes. Simulations with constant structure produced 5 neurons with false positive rate greater than 1 Hz. This occurred exclusively when the neuron tunes to more than one pattern, with one occasion resulting in greater than 90% true positives for both patterns in the test period. The constructive simulations do not produce any neurons with false positive rate above 1 Hz, but do show a higher number of neurons with false positive rates greater than zero and less than 0.55 Hz. The past study (Masquelier et al., 2008) also reported that neurons would typically become inactive if unable to tune to a pattern. Neurons with long inactive periods including the whole test period (the last 75 s of simulation time) 21 01234567891011Successful Neurons010203040SimulationsStatic StructureConstructed-4-3-2-101234567Same Trial Difference0510152025Simulations Figure 9: The number of postsynaptic neurons with true positive percentages in the test period for static-structure simulations (Top; 900 neurons) and constructive simulations (Bottom; 939 neurons). Figure 10: The number of postsynaptic neurons with false positive spike rates in the test period for static-structure simulations (Top; 900 neurons) and constructive simulations (Bottom; 939 neurons). 22 100101102103NeuronsStatic Structure00.20.40.60.81Test True Positives [%]100101102103NeuronsConstructed100101102103NeuronsStatic Structure00.511.522.53Test False Positives [Hz]100101102103NeuronsConstructed Figure 11: The distribution of the final number of active neurons for simulations with static structure and neuron construction. A neuron is deemed to be active if it spiked once in the test phase of the simulation (the last 75 s). were observed in the reproduced simulations with static structure (Figure 11). The construction of neurons demonstrates improvements in the early per- formance of selected neurons and the final learning outcomes; however, the constructive algorithm can result in periods with a substantially larger number of simulated neurons. This is especially true for the reproduced simulations con- ditions: intermittent repetition of spike-patterns hidden in a high overall rate of presynaptic activity. Simulation expansion occurs rapidly with the number of simulated neurons reaching a plateau until the maximum number of constructed neurons (nmax = 500) is reached and simulation contraction reduces the net- work to a final size (Figure 12). The intermittent repetition of spike-patterns means that performance conditions that prevent redundant neuron construc- tion are ineffective; therefore, the simulation size is controlled by the pruning mechanism and the maximum number of neurons. The total number of neurons introduced in these simulation is controlled by the construction limit, nmax = 500. The 100 constructive simulations average 9.39 neurons after simulation contraction (pruning); therefore, on average 490.6 neurons are simulated for 5 s and then pruned. Neurons may also be simulated for up to 15 ms before the construction is cancelled by another postsynaptic spike. Approximately 90% of the neurons (841/939) that remain in the final simulation were in the first 50 neurons introduced (Figure 13). Reducing the maximum number of neurons would reduce the overall compu- tational expense; however, new spike-patterns of significance could occur in later activity. Without the capability of distinguishing a spike-pattern of significance from background activity, the best option may be to perform construction con- tinuously. The reproduced simulation conditions are next modified to remove breaks between hidden spike-patterns and new sets of spike-patterns are intro- duced in each 225 s period of the simulation. 23 567891011121314Active Neurons010203040506070SimulationsStatic StructureConstructed Figure 12: The number of neurons simulated (Top) and the number of neurons constructed (Bottom) for expanding simulations with intermittent generations of hidden spike-patterns. The heavy centre-line indicates the median number of neurons across all simulations at that given simulation time. The interquartile values are represented by the grey shaded area around the median centre-line. The grey lines outside the interquartile range represent the minimum and maximum values. Values are calculated in 50 ms increments and exclude cancelled neuron construction. Figure 13: The distribution of neuron construction numbers that result in a neuron that remains in the simulation. Spike-patterns are typically learned on their first observation; therefore, 841/939 of the final simulated neurons were in the first 50 neurons constructed. 24 050100150200Simulated Neurons051015202530Time [s]0100200300400500Constructed0100200300400500Construction Number100101102Final Neurons 4.2. Dense Repetition of Spike-Patterns Modification of presynaptic neuron activity to remove gaps between hidden spike-patterns allows performance conditions that prevent neuron construction to control the simulation size and detect new spike-patterns as they are intro- duced. The dense repetition of spike-patterns results in more regular simulated postsynaptic neuron activity that inhibits the activation of the proxy neuron or cancels neuron construction. This is sufficient to control the number of neurons simulated and constructed without relying on pruning or the construction limit. The extended absence of a postsynaptic neuron spike indicates the presence of a unrecognised pattern at any stage of the simulation. Neuron construction au- tomatically responds to new patterns, demonstrating the capability to perform continual learning. Expanding simulations demonstrate high rates of pattern detection for each 225 s period that introduces new patterns (Figure 14). Static-structure simula- tions with 9 postsynaptic neurons demonstrate success in the first 225 s period but fail to detect patterns introduced in later periods of activity. The dense repetition of spike-patterns resulted in static-structure simulations producing a higher rate of postsynaptic neurons responding to more than one pattern (219/900 neurons have test true positive percentage above 50% for more than one pattern; 131/900 neurons have test true positive percentages above 90% for more than one pattern). These neurons were treated as unsuccessful in the presented figures due to a high false positive rate that has been interpreted as a lack of specificity. A small scale test showed that the tuning to multiple patterns was reduced by increasing the number of postsynaptic neurons, but increasing the number of postsynaptic neurons up to 24 did not result in any neurons that successfully tune to detect spike-patterns introduced in later 225 s activity periods. In 100 simulations with neuron construction, the first pattern set resulted in a 72.51% success rate (1137/1568 simulated neurons; 1843 total neurons con- structed); the second pattern set resulted in a 74.67% success rate (1288/1725 simulated neurons; 1955 total neurons constructed); and the third pattern set resulted in a 77.14% success rate (1326/1719 simulated neurons; 1951 total neu- rons constructed). The number of neurons simulated and the total number of neurons con- structed (Figure 15) have a trend of rapidly rising in response to the intro- duction of new patterns (at times 0 s, 225 s and 450 s) before stabilising. The median number of simulated neurons after each 225 s is 15 neurons (at 225 s), 33 neurons (at 450 s, and 50 neurons at 675 s. The median number of neurons constructed (including those pruned) is 18 neurons at 225 s, 37 neurons at 450 s, and 57 at 675 s. There is a narrow difference between the first and third quartile values of simulated neurons (a maximum of 4 neurons between first and third quartiles) and total neurons constructed (a maximum of 5 neurons between first and third quartiles). 25 Figure 14: The distributions of successful neuron numbers for each set of hidden spike- patterns. Repeating hidden spike-patterns are selected from pattern set 1 in 0 s to 225 s of the simulation (Top), selected from pattern set 2 in 225 s to 450 s (Middle), and selected from pattern set 3 in 450 s to 675 s (Bottom). Criteria for success (provided in Section 3.5) are applied to the last 75 s of the 225 s of simulation for each set of patterns. Constant structure simulations (dark bars) have zero successful neurons in the second and third sets of patterns (results omitted). Distributions are created from 100 constructive simulations and 100 static-structure simulations. 26 Pattern Set 10102030Pattern Set 20102030SimulationsPattern Set 305101520Successful Neurons0102030Static StructureConstructed Figure 15: The number of neurons simulated (Top) and the number of neurons constructed (Bottom) for simulations with dense repetition of spike-patterns. The shaded area represents the range between the first and third quartile values for 100 simulations over time; the solid black line inside the shaded area represents the median. Grey lines above and below the shaded area represent the minimum and maximum number of neurons in the 100 simulations. Values are calculated in 50 ms increments and exclude cancelled neuron construction. 5. Discussion This article has advanced the concept of simulation expansion (Lightheart et al., 2013) and developed a constructive algorithm for a simulation of bio- logical learning in a spiking neural network. The biological plausibility of the simulations that include the constructive algorithm is discussed with reference to the results obtained. The capacity for continual learning after long inter- vals of simulation time through construction is discussed and interpreted in the context of simulation expansion. Related literature on simulating biological neural network growth and con- structive spiking neural networks is discussed and contrasted with presented developments. The discussion concludes with speculation about the future di- rections of research: investigating applications in machine learning and in neu- roscientific study, and the further development of concepts and methods for the dynamic selection neurons to simulate. 5.1. Biologically Plausible Simulation Expansion A comparison of the results of non-expanding simulations of neurons tuning through additive STDP (reproduced from Masquelier et al., 2008) with the re- sults from expanding simulations shows that individual neuron performance has similar distributions. The neurons constructed with synapse weights calculated as a one-shot prediction of additive STDP convergence immediately reproduce the behaviour of selectively responding to repeating input spike-patterns. The 27 0102030405060Simulated Neurons0100200300400500600675Time [s]020406080Constructed design of the constructive algorithm based on the large surrounding network and mature network assumptions has successfully resulted in the construction of neurons that behave as mature neurons tuned through additive STDP have been found to in non-expanding simulations (reported in Masquelier et al., 2008, and reproduced in this article). This evidence supports the compatibility of this constructive algorithm with simulations of biological neural networks. The principle of plausible effects (stated in Section 2.1) requires that the ad- dition or removal of neurons from the simulation should not have a significant effect on the behaviour of remaining simulated neurons. A model and simu- lation should consider the potential impact of activity in surrounding neurons regardless of whether simulation expansion is performed. The models of the reproduced simulation include lateral inhibition between postsynaptic neurons. Mature neurons spike infrequently and with a regular delay from pattern onset with lateral inhibition producing a small delay in the spike times of neurons that immediately follow. The impact of lateral inhibi- tion being introduced from the addition of neurons to the simulation was not significant enough to reduce the success rate of neurons for the provided criteria (Masquelier et al., 2008). The high average input activity that conceals infrequently repeating spike- patterns produced in the past study (Masquelier et al., 2008) may not be typical of many biological neural networks. Biological neural networks may benefit from sparse coding of signals with increased storage capacity, creating explicit and simplified representations of input, and saving energy (Olshausen & Field, 2004); therefore, spiking neurons are unlikely to have high rates of activity unless associated with an event of some significance. A review of studies (Shoham et al., 2006) has found that electrode recording from a range of brain regions in various animals indicate average spike-rates less than 1 Hz. The performance of simulation expansion in biologically accurate low-activity conditions should receive further investigation. These conditions are expected to be more suitable for spike-triggered construction using a proxy neuron. 5.2. Continual Learning The constructive algorithm presented demonstrates a range of advantages in learning performance over the non-expanding simulations. Simulations showed that the constructive algorithm performs one-shot synapse weight calculations (based on predictions of additive STDP convergence) that are immediately suc- cessful in detecting future repetitions of hidden spike-patterns. Results have shown that, although individual neuron performance is similar, the collection of neurons produced through construction are likely to result in more successful neurons and fewer inactive neurons than simulations with randomly initialised neurons tuned through additive STDP. Simulation results demonstrate that the constructive algorithm provides ca- pabilities for continual learning. Dense generation of patterns with later intro- duction of new pattern sets demonstrated that neuron construction can be stable and immediately respond to and learn new patterns that are introduced after 28 long intervals in the simulation time. As learning is based on neuron construc- tion rather than re-training of connection weights, the constructive algorithm may also reduce the likelihood of catastrophic forgetting from ongoing learning. The non-expanding simulations with dense spike-pattern generation were unable to learn new patterns that were introduced in later intervals of the simulation. The expanding simulation assumed the existence of such neurons and produced a plausible outcome of detection of all hidden spike-patterns. This suggests that the base model of synaptic plasticity tuning neurons with randomly-initialised synapses with high total input weight (Masquelier et al., 2008) requires modification to handle learning of patterns introduced later in long simulations. The intermittent generation of patterns with a consistently high-rate of back- ground activity results in continuous construction. This is a result of the proxy neuron being unable to distinguish between input activity that includes a hid- den pattern and the background activity. The absence of patterns also results in an absence of simulated postsynaptic neuron activity that would prevent or cancel construction. Either condition (lower background activity or continu- ous generation of hidden patterns) could result in the network stabilising with zero construction. Nevertheless, a network that performs continuous construc- tion may stabilise the network size through pruning. The cost of simulating and pruning neurons may be acceptable when patterns of interest may occur at any simulation time and are not distinguishable from the average input activity level. 5.3. Related Work The work presented in this paper is most closely related to simulations of growth and structural plasticity in biological neural networks and constructive algorithms for machine learning. Biological processes of synapse growth (synaptogenesis) and neuron growth (neurogenesis) occur in regions of mature mammalian brains (Holtmaat & Svo- boda, 2009; Aimone et al., 2014). Theoretical and simulated studies have shown that synaptogenesis and neurogenesis increase computational capabilities and memory capacity (Butz et al., 2006; Aimone et al., 2009; Knoblauch & Som- mer, 2016; Spiess et al., 2016). Simulations and software have been developed for modelling the spatial and physical processes of neuron and synapse growth (Zubler & Douglas, 2009; Breit et al., 2016). Models of the growth processes can be used to automatically create or modify the structure of the neural network; however, biologically accurate models of neuron and synapse growth in a mature neural network would be limited in the speed and degree of structural modification. Greater flexibility and speed in automation of simulation size selection can be expected with processes that select neurons from the surrounding network to simulate. Constructive algorithms automatically modify the size of artificial neural networks for machine learning; however, biological plausibility is typically a peripheral concern in machine learning and constructive algorithms are rarely 29 studied for compatibility with simulations of biological neural networks. Many constructive algorithms operate with time-independent neuron models. Spiking neural models are based more closely on the time-dependent function of biolog- ical neurons; therefore, constructive algorithms for networks of spiking neurons are of particular interest. Spiking-timing-dependent construction (Lightheart et al., 2013) was pro- posed as a category of constructive algorithms that use spike-timing in pro- cesses. The constructive algorithm proposed in this article fits in this cat- egory. Other constructive spiking neural networks that fit the category of spike-timing-dependent construction include evolving spiking neural networks (Wysoski et al., 2010; Schliebs & Kasabov, 2013) and its later adaptation and extensions (Kasabov et al., 2013; Dora et al., 2016; Wang et al., 2017) and an algorithm for synaptic structural plasticity (Roy & Basu, 2017). The evolving spiking neural network or eSNN and its extensions have been widely applied to machine learning tasks (Schliebs & Kasabov, 2013; Kasabov et al., 2013; Dora et al., 2016; Wang et al., 2017) and spatio-temporal data classification including the classification of brain activity patterns (Kasabov et al., 2016). The eSNN operates on distinct training samples and calculates a synapse weight vector for each training sample based on the order of input neuron spikes. The network performance is calculated as the similarity (inverse Euclidean distance) of synapse weight vectors of existing neurons to the new synapse weight vector. If the similarity is below a threshold the new neuron is added to the network; otherwise, the construction is cancelled and the calculated synapse weight vector is used to update the weights of the most similar neuron. The eSNN implements rank-order coding (Thorpe et al., 2001), which was proposed as a model of the rapid visual processing of neural systems. New synapse weights descend in the order that the presynaptic neurons spike, with the earliest spiking presynaptic neuron receiving the highest weight. A threshold for new postsynaptic neurons is calculated as a fraction of the maximum poten- tial for the given synapse weights and spike-pattern. The decreasing weights of synapses from neurons that spike closer to the postsynaptic neuron spike-time does not closely fit the standard anti-symmetric Hebbian STDP model (Morri- son et al., 2008). The authors do not provide a case for the biological plausibility of the construction or adaptation processes of eSNNs. The evolving spiking neural network (Wysoski et al., 2010) and its later developments (Kasabov et al., 2013; Dora et al., 2016; Wang et al., 2017) can perform one-shot construction (similar to the constructive algorithm presented in this article). This can be expected to accommodate learning of new pat- tern sets introduced during operation. The one-shot construction performed by eSNNs could also be expected to result in the continuous construction of neurons in response to pure noise training samples as was observed in the presented con- structive algorithm applied to simulations with intermittent repetition of spike- patterns (Section 4.1). The base eSNN algorithm, however, does not include a pruning mechanism comparable to the presented pruning algorithm based on early inactivity. Therefore, neurons constructed in response to noise may more easily remain indefinitely in eSNNs. 30 The calculation of weights from the ranking of the first spikes of input neu- rons is not suitable for indefinitely long sequences of input neuron spikes that could be expected in long duration simulations. Dynamic evolving spiking neural networks (Kasabov et al., 2013) introduce an additional spike-driven plasticity rule to calculate subsequent adaptations of constructed synapse weights. This plasticity rule incorporates later input spike times and improves the capacity of the eSNN to learn long spike-patterns. Nevertheess, the eSNNs and its re- cent extensions (Wysoski et al., 2010; Kasabov et al., 2013; Dora et al., 2016; Wang et al., 2017) rely on distinct training samples or the provision of reference times and intervals of network simulation. The eSNN algorithms found in liter- ature would require modifications to be applicable in continuous neural network simulations that have unknown patterns times and durations. Spike-timing-dependent construction is also performed in the synaptic struc- tural plasticity algorithm (Roy & Basu, 2017) inspired by the biological process of synaptogenesis. This algorithm operates with neurons that have multiple non- linear dendrites and binary synapses. The performance of each binary synapse is calculated using a rule closely related to spike-timing-dependent plasticity. After a training period the poorest performing synapse is selected for replace- ment. Candidate replacement synapses on the same dendrite are simulated for the same input activity and the best performing candidate is selected as the replacement. The synaptic structural plasticity algorithm is applied to a competitive spike- pattern detection problem based upon the same past study reproduced in this paper (Masquelier et al., 2008). The simulations of the synaptic structural plasticity algorithm are conducted with some significant modifications of the original conditions for competitive spike-pattern detection task: • The number of input neurons is 100 rather than 2000; • Input neurons have a stable spike-rate of 20 Hz rather than randomly fluctuating spike-rates with an average of 64 Hz; • The repeated spike-patterns last 0.5 s rather than 0.05 s; • There are no periods of pure noise activity rather than pure noise account- ing for 2/3 of the simulation time; • Most simulation cases for the structural plasticity algorithm had all input neurons repeat the pattern; • One simulation case for the structural plasticity algorithm selected half of the input neurons for each pattern with the remaining neurons made silent; • The original competitive learning study created spike-patterns by ran- domly selecting the activity of half of the neurons with the stochastic activity of the remaining neurons left unaltered. 31 These differences in the simulation conditions prevent a direct comparison of results with the constructive algorithm presented in this paper. Nevertheless, the algorithm features, processes, and performance can be discussed. The structural plasticity algorithm, like the eSNN, performs most processes at the end of each pattern presentation. The algorithm includes a re-simulation of the input neuron activity to determine the performance of new candidate synapses. The requirements to divide activity into partitions and re-simulating activity for performance calculations are undesirable for continuous simulations of neural networks. The growth and pruning of synapses successfully tunes neurons to competi- tively respond to spike sequences in repeating spike-patterns; however, with one synapse replaced per iteration, training requires many iterations. The struc- tural plasticity algorithm is not demonstrated to be capable of rapidly learning or retraining on new patterns without the possibility of overwriting earlier train- ing. The structural plasticity algorithm does not change the overall number of synapses or neurons; therefore; this learning algorithm is not suitable for au- tomating the selection of the neural network size. The earlier communication of the concept of simulation expansion (Light- heart et al., 2013) was accompanied by the development of a constructive al- gorithm. The past constructive algorithm used a threshold for the number of presynaptic neurons that spike in a time-window (the previous 1 ms time- step) to trigger construction and predict the postsynaptic neuron spike-time. A method for triggering construction and predicting postsynaptic spike-times us- ing a proxy neuron has been introduced in this paper. Both of these approaches to triggering construction depends on the overall presynaptic activity and will have reduced effectiveness when presynaptic activity levels are not correlated with the timing or significance of the spike-pattern. Nevertheless, it is logical to include a mechanism that prevents construction when the the level of presy- naptic activity is insufficient to activate a constructed postsynaptic neuron. The past constructive algorithm assigned the maximum weight to synapses from neurons that were active in a time window before the predicted postsynap- tic neuron spike-time. Other synapses were reduced to the minimum synapse weight. This method will produce a variable total input weight for new neu- rons. The modified constructive algorithm selects the most recent presynaptic neurons to spike up to a given number and potentiates those synapses. This pro- duces a fixed total input weight for new neurons. Suitable total input weights will be largely dependent on the simulation conditions and will likely need a custom selection. The inclusion of a process to calculate new neuron potential thresholds (e.g., Wysoski et al., 2010) may allow these constructive algorithms to automatically accommodate wider ranges of simulation conditions. The modified constructive algorithm presented in this article has additional processes for cancelling construction. The algorithm prevents or cancels con- struction if a simulated postsynaptic neuron spikes within 15 ms before or after neuron construction. New neurons were also pruned if they did not spike mini- mum of five times in the first 5 s of simulation. These processes were effective in controlling the size of the simulation. The past constructive algorithm included 32 a condition for the Euclidean distance between the vector of new connection weights and any vector of existing connection weights was within a threshold. The focus of the study of the past constructive algorithm (Lightheart et al., 2013) was a comparison of the synapse weights produced through a simulation of STDP and those produced through construction. This constructive algorithm had yet to be demonstrated producing neurons that successfully detect hidden spike-patterns. The algorithm modifications and simulations studied in this article demonstrate that an algorithm designed from the concept of simulation expansion is capable of successfully detecting hidden spike-patterns and may be compatible with simulations for modelling and predicting biological neural network behaviour. The findings in paper also confirm that neuron construction can result in improved capabilities for continual learning. 5.4. Future Research Directions The directions of future research may be discussed in terms of the broad categories of potential applications, simulating models of biology and machine learning, and in terms of possible directions for further development of the concepts and theory of simulation expansion and contraction. The application of simulation expansion in neural networks used as a model of biology must first and foremost demonstrate that simulation expansion does not introduce biological implausible behaviours or results. The simulation re- sults presented found that neurons produced through construction had similar behaviour to those tuned through STDP. However, the simulation conditions presented do not represent any specific biological neural system. Further devel- opment and study is required to determine whether algorithms for simulation expansion and contraction have an acceptable impact on the accuracy of models of neural systems found in biological brains. Given confirmation that an algo- rithm for simulation expansion and contraction has an acceptable effect on the accuracy of a simulated model of biology, the algorithm may be incorporated into models to predict biological functions and behaviours. Simulation expansion and contraction may also have significant value in stud- ies of the computational functions and capabilities of biological neural networks. The computational capabilities of a model of a biological neural network are con- strained by the number of neurons and connections in the network. Algorithms for simulation expansion and contraction may be implemented to automatically change the simulation size according to the computational requirements and performance of the network. Successful incorporation of large network and ma- ture network assumptions may allow a simulation to be representative of the performance of a larger neural network without the simulation of an equivalent number of neurons and synapses. The application of simulation expansion to machine learning has fewer re- strictions on the features of the constructive algorithm and neural network, but machine learning has a strong demand for qualitative or quantitative improve- ments in learning performance. No other demonstration of one-shot learning of hidden spike-patterns for the given input activity conditions (Masquelier et al., 2008) have been found in literature. The constructive algorithm has also been 33 demonstrated performing learning of new pattern sets introduced after long in- tervals of operation. Performing continual learning through construction can be expected to have a lower risks of catastrophic forgetting (Goodfellow et al., 2013) as the learning is performed by creating new neurons and synapses rather than modifying the trained weights. These results encourage further investigation of the capability of the constructive algorithm as a tool for machine learning. The constructive algorithm is potentially an effective tool in automating as- pects of neural network design in simulations of biology and machine learning. Neural network models used for simulating biology and machine learning typ- ically rely on trial-and-error design and selection of the neural network size. Given a constructive algorithm that produces a neural network size that is pro- portional to characteristics of the input data and network performance, the neural network size resulting from a constructive algorithm may also be treated as indicator of the task complexity. The concepts of simulation expansion and contraction are relatively new and have received little attention from the research community; therefore, there may be additional simple advances on concept and algorithm development that may produce significant improvements in capabilities. An example that emphasises the importance in the distinction between neuron construction and expansion is the potential for selectively simulating neurons stored in memory. That is, at a given period of neural network operation the simulated neurons are a subset of all the neurons in memory, Nsim ⊆ Nmem. In this case, the set of surrounding neurons may be defined as those in memory that are not simulated, Nsur = Nmem \ Nsim, or may still assume a large surrounding network of hypothetical neurons. Neuron drop-out (Srivastava et al., 2014) can be interpreted as performing simulation expansion and contraction by randomly selecting neurons to train (or simulate) for different periods of training. This concept may be extended beyond training to the ongoing operation of the simulation with more intelligent selection of which neurons are simulated. This speculated method of selectively simulating neurons in memory is distinct from methods for pruning and com- pressing deep neural networks (Han et al., 2015) that are typically performed one time after the network has been trained. Selective simulation of neurons may be performed according to a mixture of bottom-up predictions from early processing of input or top-down attention- based mechanisms. Bottom-up prediction assumes that the significance of neu- rons is predictable from early processing of input. In general, selective simula- tion also assumes that computational contributions of neurons to some task is sufficiently separable. Human brains have localised groups of neurons that perform different func- tions, for example, visual cortices have clear, spatially organised structures that reflect the location of details detected by the retinas (Wandell & Winawer, 2011). The speed of visual sensory processing is evidence that human brains use bottom-up mechanisms to predict object categories in sensory input (Thorpe et al., 2001) and top-down mechanisms to selectively attend sensory input (Gilbert & Sigman, 2007). The simulation of large neural networks may be made 34 tractable for relatively low power computers if the behaviour of the simulation is sufficiently accurate when restricted to neuron groups selected by bottom-up and top-down processes. The dynamic selection of neurons from memory may then be suitable for reducing the computational expense of large-scale brain simulations (de Garis et al., 2010). Simulations that select neurons to simulate from memory may still perform neuron construction and add these neurons to memory. Constructive processes (creating new neurons and synapses) and selective processes (choosing sets of neurons from memory to simulate) may each be topics of study that have sig- nificance for the fields of neuroscience and machine learning. Appendix A. Neuron Input Weight Selection A preliminary investigation was conducted to examine the performance of neurons constructed with a range of total input weights for 225 s of the input ac- tivity with intermittent repetition of spike-patterns. For this preliminary inves- tigation simulation expansion is automatically triggered at 10 ms time-intervals starting at 60 ms until 150 neurons have been added to the simulation. At each expansion time neurons are inserted into separate simulation sets for total weights from 300 to 800 in increments of 50. The total weight indicates the number of synapses that are selected for po- tentiation in the new neurons. These synapses are selected from presynaptic neurons that have the most recent spike time. Lateral inhibition is not pro- duced at the time of expansion, but is produced for all subsequent postsynaptic neuron spikes. Lateral inhibition only affects neurons in the same simulated neuron set, i.e., those neurons constructed with the same initial total input weight. The total number of spikes and the maximum number of true positives in the 225 s simulation are presented visually in Figure A.16. The preliminary study of constructed neurons demonstrate immediate pat- tern detection capabilities with low or zero false positives for a range of total input synapse weights. Higher initial total input weights demonstrate longer post-construction tuning times; at the initial total input weight of 800 almost all neurons produced outside of pattern times spike between 140 and 150 times and continue spiking until the end of the 225 s simulation. The transition from neurons produced during non-pattern input activity that remain silent and con- tinuing to spike occurs between initial total input weights of 450 and 500. Initial total input weights of 400 and 450 produced neurons during pattern activity that were immediately tuned and largely rejected background activity. Neurons produced outside of pattern activity at these initial total input weights immediately ceased spiking. The initial total input weight of new neurons was chosen to be 450 for the remaining simulations presented in this article. Appendix B. Input Neuron Activity The reproduced simulation (Masquelier et al., 2008) makes use of fluctuating Poisson processes to model the input or presynaptic neuron activity. Each 35 Figure A.16: The total number of postsynaptic neuron spikes (Left) and maximum number of true positives (Right) for increasing initial total input synapse weight. Different grey-scales are indicated above each plot. New neurons are introduced at 10 ms intervals starting at t = 0.06 s. Three 0.05 s spike-patterns repeat over 225 s and occur inside the 0.06 s to 1.55 s time of simulation expansion: pattern 1 at 1.1 s, 1.2 s and 1.5 s; pattern 2 at 0.2 s, 0.5 s and 1 s; and pattern 3 at 0.05 s, 0.15 s and 1.45 s. 36 300350400450500550600650700750800Initial Input Weight Total1255075100125150Postsynaptic Neuron100200300400500Spike Total300350400450500550600650700750800Initial Input Weight Total0100200300400True Positives presynaptic neuron, j ∈ J, has an independent Poisson spiking rate, rj(t), that is time-variant. The generation of neuron spikes is performed in discrete time- steps, ∆t = 1ms. The probability of a spike in each time step is estimated as the spike-rate at that time multiplied by the time-step length, P r(j, t) = rj(t) · ∆t, where the rate rj(t) is in Hertz and the time-step length ∆t is in seconds. In each time-step a pseudo-random number, uj(t), with uniform probability in [0, 1] is computer-generated for each presynaptic neuron and compared with probability of a spike in that time-step, uj(t) ≤ P r(j, t). Each neuron that returns a true comparison spikes in that time-step with an exact time generated with uniform probability in the 1ms time-step. This method of generating Poisson neuron activity is simulated indepen- dently for 2000 neurons. In each time-step each neuron spike rate has a ran- domly generated acceleration, ∆∆rj(t), with uniform probability in the range [−360, 360] Hz s−1 ms−1. The velocity of the spike rate of each presynap- tic neuron is updated by the full value of the acceleration in each time-step, ∆rj(t) = ∆rj(t − ∆t) + ∆∆rj(t). Before updating the spike rate, ∆rj(t) is restricted to the range [−1800, 1800] Hz s−1. The spike rate for each neuron is then updated, rj(t) = rj(t − ∆t) + ∆rj(t) · ∆t. After this step the neuron spike rates are all restricted to the range 0 Hz to 90 Hz. The authors of the past study (Masquelier et al., 2008) chose this update process to produce smooth random fluctuations in the presynaptic neuron spike rates. This process of varying the spike rate of each neuron and then generat- ing presynaptic neuron spikes is performed for each time-step up to 225 s. All presynaptic neuron activity is generated before the postsynaptic neuron activ- ity is simulated. During presynaptic neuron activity generation, spikes are also generated for any neuron that has not spiked in the past fifty time-steps. This prevents neurons from remaining silent for more than 50 ms and means that all selected neurons will have at least one spike in the repeating pattern. The process for creating the final spike activity for presynaptic neurons op- erates on the results of simulating the Poisson processes. Each pattern is a different randomly selected a 50 ms segment of the Poisson activity. Each pat- tern has a random selection of half the presynaptic neurons (1000) to generate the hidden spike pattern. The activity of those selected neurons is replaced at other pattern times. The process for generating the repeating activity pattern can be summarised as follows: 1. Input neuron spikes are generated with randomly fluctuating Poisson rates and a maximum silent period of 50 time-steps for 225 second of simulation time; 2. The generated activity is divided into sequential 50 ms segments; 3. A different 50 ms segment is selected at random for each hidden spike- pattern; 4. The activity of half the neurons (designated as pattern neurons) in this segment is copied; 5. Random 50 ms segments that are not adjacent to a copy of the same pattern repetition are selected to repeat the hidden-spike pattern; 37 6. In segments selected to repeat a pattern, the pattern neurons have spikes replaced with the copied spike-pattern with 0 ms-mean, 1 ms-standard de- viation Gaussian noise applied to spike times; 7. Equal numbers of segments are selected for each pattern at random with one quarter of all segments having a repetition of a spike pattern (or all segments are selected in the case of continuous pattern generation); 8. An additional 10 Hz Poisson activity is then overlaid on the spike activity all neurons; 9. The resulting 225 s of activity is repeated two times to form one continuous 675 s spike-train (or new 225 s intervals and patterns are generated in the case of continuous pattern generation). This process of input neuron activity and hidden spike pattern generation is performed with different random number generator seed states for each simu- lation trial to provide variable simulation input conditions. An example of the result of this process of generating presynaptic neuron activity is provided in Figure 3. References Aimone, J. B., Li, Y., Lee, S. W., Clemenson, G. D., Deng, W., & Gage, F. H. (2014). Regulation and function of adult neurogenesis: From genes to cognition. Physiological Reviews, 94 , 991–1026. doi:10.1152/physrev. 00004.2014. Aimone, J. B., Wiles, J., & Gage, F. H. (2009). Computational influence of adult neurogenesis on memory encoding. Neuron, 61 , 187–202. doi:10.1016/ j.neuron.2008.11.026. Ash, T. (1989). Dynamic node creation in backpropagation networks. Connec- tion Science, 1 , 365–375. doi:10.1080/09540098908915647. Breit, M., Stepniewski, M., Grein, S., Gottmann, P., Reinhardt, L., & Queisser, G. (2016). Anatomically detailed and large-scale simulations studying synapse loss and synchrony using NeuroBox. Frontiers in Neuroanatomy, 10 , 8. doi:10.3389/fnana.2016.00008. Brette, R., Rudolph, M., Carnevale, T., Hines, M., Beeman, D., Bower, J. M., Diesmann, M., Morrison, A., Goodman, P. H., Harris Jr., F. C., Zirpe, M., Natschlager, T., Pecevski, D., Ermentrout, B., Djurfeldt, M., Lansner, A., Rochel, O., Vieville, T., Muller, E., Davison, A. P., El Boustani, S., & Des- texhe, A. (2007). Simulation of networks of spiking neurons: A review of tools and strategies. Journal of Computational Neuroscience, 23 , 349–398. doi:10.1007/s10827-007-0038-6. Butz, M., Lehmann, K., Dammasch, I. E., & Teuchert-Noodt, G. (2006). A the- oretical network model to analyse neurogenesis and synaptogenesis in the den- tate gyrus. Neural Networks, 19 , 1490–1505. doi:10.1016/j.neunet.2006. 07.007. 38 do Carmo Nicoletti, M., Bertini, J. a., Elizondo, D., Franco, L., & Jerez, J. (2009). Constructive neural network algorithms for feedforward architectures suitable for classification tasks. In L. Franco, D. Elizondo, & J. Jerez (Eds.), Constructive Neural Networks (pp. 1–23). Berlin / Heidelberg, Germany: Springer-Verlag volume 258 of Studies in Computational Intelligence. doi:10. 1007/978-3-642-04512-7_1. Dora, S., Subramanian, K., Suresh, S., & Sundararajan, N. (2016). Development of a self-regulating evolving spiking neural network for classification problem. Neurocomputing, 171 , 1216–1229. doi:10.1016/j.neucom.2015.07.086. Fahlman, S. E., & Lebiere, C. (1990). The Cascade-Correlation learning archi- tecture. In D. S. Touretzky (Ed.), Advances in Neural Information Processing Systems 2 (pp. 524–532). San Francisco, CA: Morgan Kaufmann Publishers Inc. Fritzke, B. (1995). A growing neural gas network learns topologies. In G. Tesauro, D. S. Touretzky, & T. K. Leen (Eds.), Advances in Neural Infor- mation Processing Systems 7 1 (pp. 625–632). Cambridge, MA: MIT Press. de Garis, H., Shuo, C., Goertzel, B., & Ruiting, L. (2010). A world survey of ar- tificial brain projects, part I: Large-scale brain simulations. Neurocomputing, 74 , 3–29. doi:10.1016/j.neucom.2010.08.004. Gerstner, W., & Kistler, W. M. (2002). Spiking Neuron Models: Single Neurons, Populations, Plasticity. Cambridge, United Kingdom: Cambridge University Press. Gilbert, C. D., & Sigman, M. (2007). Brain states: Top-down influences in sensory processing. Neuron, 54 , 677–696. doi:10.1016/j.neuron.2007.05. 019. Goodfellow, I. J., Mirza, M., Xaio, D., Courville, A., & Bengio, Y. (2013). An empirical investigation of catastrophic forgetting in gradient-based neural networks. ArXiv e-prints, 1312 . arXiv:1312.6211. Han, S., Mao, H., & Dally, W. J. (2015). Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. ArXiv e-prints, 1510 . arXiv:1510.00149. Holtmaat, A., & Svoboda, K. (2009). Experience-dependent structural synaptic plasticity in the mammalian brain. Nature Reviews Neuroscience, 10 , 647– 658. doi:10.1038/nrn2699. Huang, G.-B., Saratchandran, P., & Sundararajan, N. (2005). A gener- alized growing and pruning RBF (GGAP-RBF) neural network for func- tion approximation. IEEE Transactions on Neural Networks, 16 , 57–67. doi:10.1109/TNN.2004.836241. 39 Izhikevich, E. M. (2003). Simple model of spiking neurons. IEEE Transaction on Neural Networks, 14 , 1569–1572. doi:10.1109/TNN.2003.820440. Kasabov, N., Dhoble, K., Nuntalid, N., & Indiveri, G. (2013). Dynamic evolv- ing spiking neural networks for on-line spatio- and spectro-temporal pattern recognition. Neural Networks, 41 , 188–201. doi:10.1016/j.neunet.2012. 11.014. Kasabov, N., Scott, N. M., Tu, E., Marks, S., Sengupta, N., Capecci, E., Oth- man, M., Doborjeh, M. G., Murli, N., Hartono, R., Espinosa-Ramos, J. I., Zhou, L., Alvi, F. B., Wang, G., Taylor, D., Feigin, V., Gulyaev, S., Mah- moud, M., Hou, Z.-G., & Yang, J. (2016). Evolving spatio-temporal data machines based on the neucube neuromorphic framework: Design methodol- ogy and selected applications. Neural Networks, 78 , 1–14. doi:10.1016/j. neunet.2015.09.011. Knoblauch, A., & Sommer, F. T. (2016). Structural plasticity, effectual connectivity, and memory in cortex. Frontiers in Neuroanatomy, 10 , 63. doi:10.3389/fnana.2016.00063. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. In F. Pereira, C. Burges, L. Bottou, & K. Weinberger (Eds.), Advances in Neural Information Processing Systems 25 (pp. 1097–1105). Red Hook, NY: Curran Associates, Inc. Lightheart, T., Grainger, S., & Lu, T.-F. (2013). Spike-timing-dependent con- struction. Neural Computation, 25 , 2611–2645. doi:10.1162/NECO_a_00501. Masquelier, T., Guyonneau, R., & Thorpe, S. J. (2008). Competitive STDP- based spike pattern learning. Neural Computation, 21 , 1259–1276. doi:10. 1162/neco.2008.06-08-804. Morrison, A., Diesmann, M., & Gerstner, W. (2008). Phenomenological models of synaptic plasticity based on spike timing. Biological Cybernetics, 98 , 459– 478. doi:10.1007/s00422-008-0233-1. Olshausen, B. A., & Field, D. J. (2004). Sparse coding of sensory inputs. Current Opinion in Neurobiology, 14 , 481–487. doi:10.1016/j.conb.2004.07.007. Roy, S., & Basu, A. (2017). An online unsupervised structural plasticity algo- rithm for spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 28 , 900–910. doi:10.1109/TNNLS.2016.2582517. Schliebs, S., & Kasabov, N. (2013). Evolving spiking neural network–a survey. Evolving Systems, 4 , 87–98. doi:10.1007/s12530-013-9074-9. Shoham, S., O'Connor, D. H., & Segev, R. (2006). How silent is the brain: is there a "dark matter" problem in neuroscience? Journal of Comparative Physiology A, 192 , 777–784. doi:10.1007/s00359-006-0117-6. 40 Song, S., Miller, K. D., & Abbott, L. F. (2000). Competitive Hebbian learning through spike-timing-dependent synaptic plasticity. Nature Neuroscience, 3 , 919–926. doi:10.1038/78829. Spiess, R., George, R., Cook, M., & Diehl, P. U. (2016). Structural plasticity denoises responses and improves learning speed. Frontiers in Computational Neuroscience, 10 , 13. doi:10.3389/fncom.2016.00093. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15 , 1929–1958. Thorpe, S., Delorme, A., & Rullen, R. V. (2001). Spike-based strategies for rapid processing. Neural Networks, 14 , 715–725. doi:10.1016/S0893-6080(01) 00083-1. Wandell, B. A., & Winawer, J. (2011). Imaging retinotopic maps in the human brain. Vision Research, 51 , 718–737. doi:10.1016/j.visres.2010.08.004. Wang, J., Belatreche, A., Maguire, L. P., & McGinnity, T. M. (2017). SpikeTemp: An enhanced rank-order-based learning approach for spiking neu- ral networks with adaptive structure. IEEE Transactions on Neural Networks and Learning Systems, 28 , 30–43. doi:10.1109/TNNLS.2015.2501322. Watts, M. J. (2009). A decade of Kasabov's evolving connectionist systems: A review. IEEE Transactions on Systems, Man, and Cybernetics, Part C , 39 , 253–269. doi:10.1109/TSMCC.2008.2012254. Wysoski, S. G., Benuskova, L., & Kasabov, N. (2010). Evolving spiking neural networks for audiovisual information processing. Neural Networks, 23 , 819– 835. doi:10.1016/j.neunet.2010.04.009. Zubler, F., & Douglas, R. (2009). A framework for modeling the growth and development of neurons and networks. Frontiers in Computational Neuro- science, 3 , 25. doi:10.3389/neuro.10.025.2009. 41
1903.04201
2
1903
2019-10-16T00:45:27
A causal role of sensory cortices in behavioral benefits of 'learning by doing'
[ "q-bio.NC" ]
Despite a rise in the use of "learning by doing" pedagogical methods in praxis, little is known as to how these methods improve learning outcomes. Here we show that visual association cortex causally contributes to performance benefits of a learning by doing method. This finding derives from transcranial magnetic stimulation (TMS) and a gesture-enriched foreign language (L2) vocabulary learning paradigm performed by 22 young adults. Inhibitory TMS of visual motion cortex reduced learning outcomes for abstract and concrete gesture-enriched words in comparison to sham stimulation. There were no TMS effects on words learned with pictures. These results adjudicate between opposing predictions of two neuroscientific learning theories: While reactivation-based theories predict no functional role of visual motion cortex in vocabulary learning outcomes, the current study supports the predictive coding theory view that specialized sensory cortices precipitate sensorimotor-based learning benefits.
q-bio.NC
q-bio
Running head: CAUSAL ROLE OF SENSORY CORTICES A causal role of sensory cortices in behavioral benefits of „learning by doing‟ *†Brian Mathias1,2, †Leona Sureth2, Gesa Hartwigsen3, Manuela Macedonia2,4, Katja M. Mayer5, & Katharina von Kriegstein1,2 1Chair of Cognitive and Clinical Neuroscience, Faculty of Psychology, Technical University Dresden, Dresden, Germany 2Research Group Neural Mechanisms of Human Communication, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany 3Lise Meitner Research Group Cognition and Plasticity, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany 4Linz Center of Mechatronics, Johannes Kepler University Linz, Linz, Austria 5Department for Psychology, University of Münster, Münster, Germany †Joint first authors *Corresponding author: Brian Mathias Technical University Dresden Chair of Cognitive and Clinical Neuroscience Faculty of Psychology Bamberger Str. 7 01187 Dresden Germany Email: [email protected] CAUSAL ROLE OF SENSORY CORTICES 2 Abstract Despite a rise in the use of "learning by doing" pedagogical methods in praxis, little is known as to how these methods improve learning outcomes. Here we show that visual association cortex causally contributes to performance benefits of a learning by doing method. This finding derives from transcranial magnetic stimulation (TMS) and a gesture-enriched foreign language (L2) vocabulary learning paradigm performed by 22 young adults. Inhibitory TMS of visual motion cortex reduced learning outcomes for abstract and concrete gesture-enriched words in comparison to sham stimulation. There were no TMS effects on words learned with pictures. These results adjudicate between opposing predictions of two neuroscientific learning theories: While reactivation-based theories predict no functional role of visual motion cortex in vocabulary learning outcomes, the current study supports the predictive coding theory view that specialized sensory cortices precipitate sensorimotor-based learning benefits. CAUSAL ROLE OF SENSORY CORTICES 3 Foreign language (L2) vocabulary learning by adults is effortful and time-consuming. Words must be relearned often to build up robust memory representations. L2 vocabulary learning typically relies on unisensory material such as written word lists or audio recordings (Choo, Lin, & Pandian, 2012). More recent learning-by-doing-based approaches contrast with these techniques. Though initially viewed as unconventional, principles of learning by doing have shifted from the periphery of educational science toward its center over the past few decades. Such strategies make use of visual and somatosensory information as well as motor information. We will therefore in the following refer to learning by doing strategies as sensorimotor-enriched teaching. Sensorimotor-enriched teaching methods boost test performance in science, engineering, mathematics, and L2 learning compared to other teaching methods (Freeman et al., 2014; for a review see Macedonia, 2014). Underlying human brain mechanisms that support beneficial effects of sensorimotor-enriched teaching methods are currently unknown. A recent behavioral and functional magnetic resonance imaging (fMRI) study investigated sensorimotor-enriched L2 vocabulary learning using four enriched teaching methods and non- enriched control conditions (Mayer, Yildiz, Macedonia, & von Kriegstein, 2015). In the first enrichment condition, learners heard L2 words and their native language (L1) translations while performing iconic gestures that were modelled by an actress in video clips (gesture performance enrichment). In the second enrichment condition they heard a different set of L2 words while viewing videos without performing the gestures (gesture viewing enrichment). Gesture performance enrichment but not gesture viewing enrichment enhanced post-training learning outcomes compared to non-enriched L2 learning. Interestingly, the third enrichment condition, which entailed viewing iconic pictures during learning (picture viewing enrichment), also enhanced post-training outcomes compared to non-enriched L2 learning. However, the fourth enrichment condition, in which pictures were traced by hand (picture performance enrichment) did not benefit post-training performance compared to non-enriched L2 learning. Gesture- performance-enriched and picture-viewing-enriched learning outcomes did not differ from each other immediately after the 5-day training protocol. Gesture performance enrichment, however, led to more favorable learning outcomes than picture viewing enrichment over the long-term (6 months post-training). Mayer and colleagues‟ (2015) fMRI results demonstrated that L2 words presented in the auditory modality following training could be decoded in specialized visual association cortices on the basis of how the words were learned: The biological motion area of the superior temporal sulcus (bmSTS), a region implicated in visual perception of biological movements (Grossman et CAUSAL ROLE OF SENSORY CORTICES 4 al., 2000), was engaged by words that had been learned under both types of gesture enrichment (i.e., gesture performance and gesture viewing). The lateral occipital complex (LOC) was instead engaged following picture-enriched learning (picture performance and picture viewing). These neuroimaging results are consistent with studies showing that the presence of complementary sensory information during learning elicits reactivation of specialized sensory brain regions at test (Danker & Anderson, 2010; Lahav, Saltzman, & Schlaug, 2007). The reactivation of visual brain regions elicited by stimuli presented in the auditory modality may be viewed as epiphenomenal, a view taken by reactivation theories of multisensory learning (Fuster, 2009; Nyberg, Habib, McIntosh, & Tulving, 2000; Wheeler, Petersen, & Buckner, 2000). For example, if the taste of a cake triggered the recall of a visual memory, the recollection might reactivate visual areas; reactivation theories, however, assume that those visual responses do not aid in making the sensory experience of the cake‟s taste more precise. Within a reactivation-based framework, benefits of sensorimotor enrichment on learning could be relegated to arousal-based effects that are not dependent on representations subserved by sensory brain regions. Reactivation theories can be seen as a critical counterpart to the notion that sensorimotor networks formed during real-world experience support perception (for reviews see Barsalou, 2010, von Kriegstein 2012, and Matusz, Wallace, & Murray, 2017). According to this alternate view, sensory brain responses to previously-learned items directly benefit recognition processes by increasing recognition speed and accuracy. The predictive coding theory of multisensory learning (Mayer et al., 2015; von Kriegstein & Giraud 2006; for review see von Kriegstein, 2012) takes this approach by proposing that sensory and motor cortices build up sensorimotor (e.g., visuomotor) forward models during perception that predict or simulate missing input, which functionally benefit behavioral learning outcomes. If that were the case, teaching techniques could be optimized to target specific sensory structures that underlie task performance. The first aim of the current study was thus to adjudicate between tenets of reactivation and predictive coding theories of multisensory learning. The second aim of the study was to evaluate the role of multisensory brain responses not only immediately following learning, but also over extended post-training durations. This aim was based on the finding that gesture-performance enrichment has shown to benefit learning over longer time scales than the initially equally effective viewing of pictures (Mayer et al., 2015), Though functional neuroimaging has contributed much to our understanding of interactions between information arising from distinct sensory modalities (for review see James & Stevenson, 2012), neuroimaging techniques are limited to demonstrations of correlational rather than causal CAUSAL ROLE OF SENSORY CORTICES 5 effects (Ramsey et al., 2010). Conversely, transcranial magnetic stimulation (TMS) is a key method for making causal claims regarding the role of specialized sensory and motor areas in behavioral benefits of sensorimotor-enriched learning. TMS can be used to test whether brain regions that are engaged during sensorimotor-enriched learning causally contribute to its beneficial behavioral effects. During TMS, small magnetic fields are applied non-invasively to the scalp, targeting a specific brain area. The magnetic fields induce electrical currents in the underlying brain tissue, transiently interfering with processing in a specific area. If the stimulated region is causally relevant for an ongoing task, then an observable behavioral effect, usually an increase in response latencies, can be induced (Sack et al., 2007). TMS cannot be substituted by behavioral, fMRI, or other correlational neuroscience measurements. Methods Overview and Hypotheses The study was constructed in close analogy to the design by Mayer et al (2015). Adult learners were trained on 90 novel L2 words and their L1 translations over 4 consecutive days (Fig. 1a). L2 vocabulary (concrete and abstract nouns, supplementary Table S1) was learned in two conditions. In a gesture-performance-enriched learning condition, individuals viewed and performed gestures while L2 words were presented auditorily. In a picture-viewing-enriched learning condition, individuals viewed pictures while L2 words were presented auditorily (Fig. 1b). Gestures and pictures were congruent with word meanings. We selected the gesture performance enrichment and picture viewing enrichment conditions for two main reasons. First, of the four enrichment conditions previously tested in adults (Mayer et al., 2015), only these two conditions benefitted post-training L2 translation compared to auditory-only learning. Second, benefits of gesture performance enrichment and picture viewing enrichment were associated with responses in two different visual cortical areas, i.e., the bmSTS for gesture performance enrichment and the LOC for picture viewing enrichment (Mayer et al., 2015). For succinctness, we hereafter refer to the gesture performance enrichment condition as the "gesture enrichment" learning condition, and the picture viewing enrichment condition as the "picture enrichment" learning condition. We used TMS to target the bmSTS, as the bmSTS is more easily accessible for TMS than the LOC (Siebner & Ziemann, 2007). TMS was applied to the bilateral bmSTS while participants translated auditorily- presented L2 words into L1 at two time points: 5 days and 5 months following the start of L2 training. Participants did not perform gestures or view pictures during the TMS task. A within- CAUSAL ROLE OF SENSORY CORTICES 6 participants control condition was included in each TMS session by applying both effective and sham TMS to the bilateral bmSTS (Fig. 1c). Classification accuracy of neural responses within the bmSTS previously correlated with performance in a multiple choice translation task (Mayer et al., 2015). In this task, participants selected the correct L1 translation of an auditorily-presented L2 word from a list of options presented on a screen. This task was also used for the present TMS design and we refer to it as the "multiple choice task" (Fig. 1d). In addition, we included an exploratory recall task in the present TMS design in which participants pressed a button as soon as the L1 translation came to their mind when hearing each L2 word at the start of each trial (not shown in Fig. 1d, see procedure and supplementary results for more details). Figure 1. Experimental procedure and design. a, Participants learned foreign language (L2) vocabulary over four consecutive days (‟learn‟) in groups to emulate a classroom setting. Free recall and translation tests („test‟) were administered on days 2 through 4. Transcranial magnetic stimulation (TMS) sessions occurred during day 5 and month 5. b, In both gesture and picture learning conditions, participants heard an L2 word, followed by the translation in their native language (L1) and a repetition of the L2 word. Videos of iconic gestures and pictures TMS taskLearning procedureGesture trialPicture trialStudy timeline1 tram2 tent3 thought4 unknown / other wordIntertrial interval TMS320.5 -1L2 word:"umuda"L2 word:"umuda"L1 translation:"tram"L2 word:"umuda"Day 1Day 2Day 3Day 4Day 5Month 5testtesttestTMSTMS learnlearnlearnlearntestlearnPicture enrichment: 2 blocks 33 minsGesture enrichment: 2 blocks 33 minsFree recallL1 translation L2 translation0 0.2 4 4.2 5.7 5.9 9.7 10.7 (s)L2 word:"umuda"L1 translation:"tram"L2 word:"umuda"0 0.2 4 4.2 5.7 5.9 9.7 10.7 (s)dca(s)b CAUSAL ROLE OF SENSORY CORTICES 7 accompanied L2 words in gesture and picture trials, respectively. Participants performed the gesture along with the video during its repetition. c, During both TMS sessions (day 5 and month 5), two TMS coils targeted the bilateral biological motion superior temporal sulcus (bmSTS) using stereotactic neuronavigation based on individual structural brain scans. Two additional coils generated ineffective placebo stimulation (i.e., sham TMS) and were positioned on top of the bmSTS coils at an angle of 90°. d, During each TMS session, participants heard the L2 words that they had learned during the 4-day training and then selected the L1 translation by button press from a list of options presented on a screen (multiple choice task). L1 words were presented in German. Trains of seven TMS pulses at 10 Hz were delivered 50 ms following each L2 word onset. Trials with effective and sham TMS alternated in blocks. Note that participants additionally pressed a button as soon as the L1 translation came to their mind following each auditorily-presented L2 word (recall task; not shown, see methods and supplementary results for more details). We tested three main hypotheses. First, according to the predictive coding theory of multisensory learning (von Kriegstein, 2012), the application of inhibitory stimulation to the bmSTS should slow down the translation of an auditorily-presented L2 word in comparison to sham stimulation if the word has been learned with biological motion, as was the case in our gesture-enrichment condition. There should be no such effect if the word has been learned with pictures. Thus we expected an interaction between learning condition (gesture, picture) and stimulation condition (effective, sham). The interaction should be driven by a simple main effect of stimulation condition on the translation of gesture-enriched words. In contrast, reactivation learning theories (Fuster, 2009; Nyberg et al., 2000; Wheeler et al., 2000), which assume that reactivated areas do not play a functional role in recognition, would predict no differential effects of bmSTS stimulation in contrast to sham bmSTS stimulation on the translation of auditorily- presented L2 words (i.e., there would be no interaction between learning and TMS conditions). Our second hypothesis was that bmSTS integrity would support the auditory translation of gesture-performance-enriched words at the later time point (5 months post-learning) even more than the earlier time point (5 days post-learning). This hypothesis was based on the finding that gesture-performance enrichment is particularly powerful for learning outcome on time-scales of several months post learning in comparison to picture-viewing enrichment (Mayer et al., 2015). In our design, this hypothesis could be tested by examining the interaction between learning condition, stimulation type, and time point variables: We expected greater effects of bmSTS stimulation compared to sham stimulation on translation response times for gesture-enriched CAUSAL ROLE OF SENSORY CORTICES 8 words at the later time point than at the earlier time point, and no effects of bmSTS stimulation compared to sham stimulation for picture-enriched words at either time point. Our third hypothesis was that bmSTS stimulation would yield similar effects on the translation of both concrete and abstract words (see supplementary introduction on influences of the conceptual perceptibility of L2 word referents). This prediction was based on previous results showing that sensorimotor enrichment can benefit the learning of both word types (Macedonia, 2014; Mayer et al., 2015). Response time was used as the dependent variable for testing our three hypotheses, because response time is the standard measure for TMS tasks: TMS typically influences response times rather than accuracy (Ashbridge, Walsh, & Cowey, 1997; Hartwigsen et al., 2017; Pascual-Leone et al., 1996; Sack et al., 2007). In addition to testing our three main hypotheses, the design allowed us to test the reliability of the previously-reported finding that benefits of gesture performance enrichment exceed those of picture viewing enrichment over the long-term (Mayer et al., 2015). To this end, we examined accuracy outcomes in the multiple choice task, and predicted that overall accuracy would be greater for gesture-enriched words compared to picture-enriched words 5 months post-learning. Participants Twenty-two right-handed native German speakers (15 females; M age = 26.6 years, SD = 4.7 years, range 18-35) completed the study. The sample size was based on two previous experiments (n = 22 per experiment) that estimated beneficial effects of gesture and picture enrichment on foreign language (L2) learning outcomes (Mayer et al., 2015, Experiments 1 and 2). None of the participants reported a history of neurological or psychiatric disorders, head injury, or any contraindications for TMS. All participants reported normal hearing and normal or corrected-to-normal vision. None of the participants were raised in bilingual households. Of 32 total participants who registered for the study, one participant experienced an adverse reaction to TMS (convulsive syncope) and did not complete the testing. Syncope is the most common adverse reaction to TMS; the exact prevalence is unknown (Rossi, Hallett, Rossini, & Pascual- Leone, 2009). Another participant completed the training sessions but did not start the initial TMS session for medical reasons. Four participants were excluded because they were unable to return for additional testing sessions that occurred 5 months following the initial testing sessions due to time constraints, and 4 other participants were excluded due to poor localization CAUSAL ROLE OF SENSORY CORTICES 9 of right or left bmSTS coordinates in individual participant space. Written informed consent was obtained from all participants prior to the study. Participants were informed that the goal of the study was to test the effectiveness of different vocabulary learning strategies in adulthood but they were naïve to the specific hypotheses. All participants were evaluated by a medical doctor prior to the study in order to be approved for TMS and magnetic resonance imaging (MRI). The study was approved by the ethics committee of the University of Leipzig. Stimuli Stimuli consisted of 90 pseudowords (see supplementary material, Table S1). The pseudowords were derived from an artificial foreign language corpus referred to as „Vimmi‟, developed by Macedonia and colleagues (Macedonia & Knösche, 2011) and intended for use in experiments on L2 learning. The corpus was created in order to control for participants‟ prior knowledge of foreign languages and for differences between words (e.g., length, frequency) in natural languages. Vimmi words conform to rules of Italian phonotactics (words sound like Italian but do not exist in the Italian language). All Vimmi words used in the current study were composed of three syllables consisting of vowels and consonants. The 90 Vimmi words and 90 German translations used in the current study were previously evaluated by Mayer and colleagues (2015). Half of the 90 words were concrete nouns and the other half were abstract nouns. Lengths of concrete and abstract German words did not differ (concrete M = 2.40 syllables, SD = 0.84 syllables; abstract: M = 2.69 syllables, SD = 0.90 syllables). Frequency of the concrete and abstract words in written German did not differ (concrete frequency score: M = 11.00, SD = 1.18, range 9 to 13; abstract frequency score: M = 10.96, SD = 0.98, range: 9 to 13) (http://wortschatz.uni-leipzig.de/de). Videos, pictures, and audio files. For each of the 90 Vimmi words, a 4 s color video was created using a Canon Legria HF S10 camcorder (Canon Inc., Tokyo, Japan). In each video, an actress performed a gesture that conveyed a word meaning. The actress was always positioned in the center of the video recording. She performed the gestures using head movements, movements of one or both arms or legs, fingers, or combinations of these body parts and maintained a neutral facial expression throughout each video. The word "bottle", for example, was represented by the actress miming drinking from an imaginary bottle, and the word "good deed" was represented by the actress miming laying a donation in the imaginary hat of a homeless individual. The actress began and ended each gesture by standing motionless with her arms at her sides. Large gestures (e.g., steps or jumps) were restricted to a 1 m radius CAUSAL ROLE OF SENSORY CORTICES 10 around the body‟s starting position. Gestures used to convey the meanings of abstract words were agreed upon by 3 independent parties (Mayer et al., 2015). A black-and-white line drawing was created by a professional illustrator for each of the 90 Vimmi words. Pictures conveyed word meanings by portraying humans, objects, or scenes. Pictures illustrating concrete nouns were mostly drawings of single objects, and pictures illustrating abstract nouns were often scenes. The complexity of the illustrations for concrete and abstract words was not matched, since similar differences are expected in natural teaching settings. The same actress that carried out the gestures in the videos spoke the Vimmi and the German words in audio recordings. Words were recorded using a Rode NT55 microphone (Rode Microphones, Silverwater, Australia) in a sound-damped chamber. The actress is an Italian native speaker and recorded the Vimmi words with an Italian accent to highlight the foreign language aspect of the stimuli for German-speaking participants. Vimmi audio stimuli ranged from 654-850 ms in length (M = 819.7 ms, SD = 47.3 ms). For more details on the video, picture and audio files used in the current study, see Mayer and colleagues (2015). Sample stimuli can be found at http://kriegstein.cbs.mpg.de/mayer_etal_stimuli/. Experimental Design The study utilized a 2 × 2 × 2 × 2 repeated-measures design. Within-participant independent factors were learning enrichment condition (gesture, picture), TMS condition (effective stimulation, sham stimulation), testing time point (5 days, 5 months), and L2 vocabulary type (concrete, abstract). Participants received both effective and sham TMS within the same session at each time point. The order in which effective and sham TMS conditions were administered within each session was counterbalanced across participants within each time point and between time points. Procedure L2 vocabulary learning. Participants learned L2 words in two conditions. In the gesture learning condition, individuals viewed and performed gestures while L2 words were presented auditorily. In the picture learning condition, individuals viewed pictures while L2 words were presented auditorily. Each day of learning comprised four 33-minute learning blocks. Blocks alternated between gesture and picture enrichment conditions. Each of the 45 Vimmi words included in a single learning block was repeated 4 times per block, yielding a total of 180 randomly-ordered trials per block. Participants took breaks of 10 minutes between blocks. CAUSAL ROLE OF SENSORY CORTICES 11 During breaks, participants conversed with each other and consumed snacks and drinks that were provided. Enrichment condition orders were counterbalanced across participants and learning days. Participants were instructed prior to the start of learning that the goal was to learn as many Vimmi words as possible over the 4 days of training. Participants received no further instruction during the training except to be informed about which learning condition would occur next (i.e., gesture or picture enrichment). Since the L2 vocabulary learning took place in groups of up to 4 individuals, training sessions occurred in a seminar room with a projector and a sound system. Audio recordings were played via speakers located on each side of the screen. The volume of the playback was adjusted so that all participants could comfortably hear the words. The assignment of the 90 stimuli to learning enrichment conditions was counterbalanced such that half of the participants learned one set of 45 words in the gesture learning condition and the other set of 45 words in the picture learning condition. The other half of the participants received the reverse assignment of stimuli to gesture and picture conditions. This manipulation ensured that each Vimmi word was equally represented in both the gesture enrichment condition and the picture enrichment control condition. In each gesture enrichment trial (Fig. 1b), participants first heard an L2 word accompanied by a video of an actress performing a gesture that conveyed the meaning of the word (shown for 4 s). They then heard the native language (L1) translation paired with a blank screen. Finally, the L2 word was presented a second time, again accompanied by the same video of the actress performing the gesture. Participants were asked to enact the gesture along with the actress during the second showing of each video. They were free to perform the gestures mirror-inverted or they could use their right arm when the actress in the video used her right arm, for example; they were asked to use only one of the two strategies throughout the learning period. In each picture enrichment trial (Fig. 1b), participants first heard an L2 word accompanied by a picture that conveyed the meaning of the word (shown for 4 s). They then heard the L1 translation paired with a blank screen. Finally, the same L2 word was presented a second time, again accompanied by the same picture. A motor task was not included in the picture enrichment condition as the enrichment of picture viewing with motor information (e.g., tracing an outline of presented pictures) has been shown to be less beneficial for learning than simply viewing the pictures without performing a motor task (Mayer et al., 2015). We therefore did not combine picture enrichment with motor performance in the current study. Participants stood during all learning blocks. Standing locations during the training were counterbalanced over the 4 learning days. CAUSAL ROLE OF SENSORY CORTICES 12 On days 2, 3, and 4 of the L2 vocabulary learning, participants completed paper-and- pencil vocabulary tests prior to the training, shown in Fig. 1a. We included these tests in order to maintain the same L2 training procedure used by Mayer et al (2015). Participants completed free recall, L1 translation, and L2 translation tests on each day. More information on the vocabulary tests as well as the test results can be found in the supplementary material. The participant with the highest combined scores on the paper-and-pencil vocabulary tests across days 2, 3, and 4 was rewarded with an additional 21€ beyond the total study compensation of 211€. Participants were informed about the financial incentive on day 1 prior to the start of the learning blocks. Prior to vocabulary learning on day 1, participants completed three psychological tests examining their concentration ability (Concentration test; Brickenkamp, 2002) (M score = 211.6, SD = 51.1), speech repetition ability (Nonword Repetition test; Korkman, Kirk, & Kemp, 1998) (M score = 98.2, SD = 8.8), and verbal working memory (Digit Span test; Neubauer & Horn, 2006) (M score = 18.7, SD = 3.7). None of the participants were outliers (3 SD above or below the group mean) with respect to their scores on any of the three tests, and all participants performed within the norms of the Concentration test for which norms were available. Participants also completed a questionnaire on their prior knowledge of foreign languages and language learning experience. TMS translation tasks. Participants performed recall and multiple choice tasks (Fig. 1d) while undergoing effective and sham TMS in two TMS sessions (5 days and 5 months following the start of L2 vocabulary learning). The two tasks were performed in four 6-minute blocks, each containing 45 words that had been presented on days 1 to 4. Each of the 90 words learned during the learning days was presented twice per TMS session, for a total of 180 test trials per TMS session and task. Effective and sham stimulation alternated across blocks, with half of the participants receiving effective stimulation during the first block and the other half receiving sham stimulation during the first block. Stimuli were ordered randomly within effective and sham stimulation blocks. Each trial began with the written instruction "Press the button as soon as you know the translation" presented for 1.5 s on a screen. This was followed by the auditorily-presented L2 word accompanied by a black screen. A train of seven TMS pulses at 10 Hz delivered to the bilateral bmSTS began 50 ms after the onset of each word. Participants responded as soon as they recalled the L1 translation of the L2 word by pressing a button with their right index finger (recall task, not shown in Fig. 1d). If they did not know the L1 translation, they did not respond. Three seconds following L2 word onsets, a screen with four response options appeared and CAUSAL ROLE OF SENSORY CORTICES 13 participants were given up to 2 s to select the correct L1 translation (Fig. 1d). The fourth response option was always "Unknown / Other word"; participants were told to select this option if they did not know the L1 translation or thought that the correct translation was different from the three options presented. They responded by pressing one of four buttons on the response pad with their index, middle, ring, or little fingers (multiple choice task). Even if participants did not know the translation of the L2 word after hearing it, they were still able to select one of the four options presented. Responses were considered correct if participants pressed the correct button while the response screen was present. Participants were instructed to always respond as quickly and as accurately as possible. Each trial ended with a jittered inter-stimulus interval (0.5 to 1 s) paired with a black screen. Following the first TMS session, participants completed a questionnaire on strategies that they used to learn and remember the L2 words. Several months following the first TMS session, participants were invited to participate in a second TMS session. The second session occurred approximately 18 weeks (M = 18.0 weeks, SD = 1.4 weeks) following the first session. Participants completed the same two tasks as during the first TMS session while again undergoing effective and sham stimulation. Following the second TMS session, participants completed a questionnaire on strategies they used to remember meanings of the L2 words during the second session. Finally, participants returned to complete the pencil-and-paper vocabulary tests (free recall, L1 translation, and L2 translation) 2 to 6 days (M = 4.1 days, SD = 1.3 days) after their second TMS session. Participants had no knowledge of the additional TMS and behavioral sessions until they were contacted a few weeks prior to their 5-month target testing dates. This was done to avoid potential rehearsal of the vocabulary during the 5-month interval between testing time points. Transcranial Magnetic Stimulation Neuronavigation. Stereotactic neuronavigated TMS was performed using Localite software (Localite GmbH, Sankt Augustin, Germany). Neuronavigation based on structural neuroimaging data from individual participants allows precise positioning of TMS coils. T1- weighted MRI scans for each participant were obtained with a 3-Tesla MAGNETOM Prisma-fit (Siemens Healthcare, Erlangen, Germany) using a magnetization-prepared rapid gradient echo (MPRAGE) sequence in a sagittal orientation (repetition time = 2300 ms, echo time = 2.98 ms, inversion time = 900 ms, flip angle = 9°, voxel size = 1x1x1 mm). Structural T1 brain scans used for TMS neuronavigation were obtained from all participants prior to the TMS sessions. During each TMS session, participants were co- CAUSAL ROLE OF SENSORY CORTICES 14 registered to their T1 scans. The two stimulation coils used in the current study were placed over Localite-indicated entry points of the respective target sites on the scalp. Entry points were those coordinates on each participant‟s scalp that were the shortest distance to the target neural coordinates (right and left bmSTS). To stimulate the bmSTS bilaterally, a tangential coil orientation of 135◦ to the sagittal plane was applied with current flow within both stimulation coils reversed, resulting in a posterior to anterior (PA) current flow in the brain. A 135◦ coil orientation with a PA current flow is equivalent to a 45◦ coil orientation with an anterior to posterior (AP) current flow. Coils were secured in position using fixation arms (Manfrotto 244, Cassola, Italy). Mean Montreal Neurological Institute (MNI) coordinates for bilateral bmSTS stimulation were derived from the functional MRI findings of Mayer and colleagues (2015): right bmSTS, x, y, z = 55, −41, 4; left bmSTS, x, y, z = −54, −41, −5. Mayer and colleagues (2015) found that participants translated auditorily-presented L2 words learned previously with gesture enrichment more accurately than L2 words learned without enrichment (auditory-only learning), referred to as a gesture enrichment benefit. Using multivariate pattern analysis (MVPA), they found that a classifier trained to discriminate BOLD responses to gesture-enriched and auditory-only words showed significant classification accuracy in the bmSTS. Classifier accuracy in the bmSTS positively correlated with the gesture enrichment benefit, suggesting a role of this area in improving learning outcomes following multisensory learning. In the current study, we stimulated the mean location across participants that demonstrated maximal classifier accuracy within the bmSTS. To ensure precise individual stimulation of target coordinates, mean MNI coordinates for the two target sites (right and left bmSTS) were transferred into individual subject space using SPM8 (Wellcome Trust Center for Neuroimaging, University College London, UK, http://www.fil.ion.ucl.ac.uk/spm/). TMS parameters. Two MagPro X100 stimulators (MagVenture A/S, Farum, Denmark) and a total of four focal figure-of-eight coils (C-B60; outer diameter = 7.5 cm) were used for stimulation. Signal software version 1.59 (Cambridge Electronic Design Limited, Cambridge, UK) was used to control the TMS pulse sequence. Presentation software (Neurobehavioral Systems Inc., Berkeley, CA, USA) was used for stimulus delivery, response recording, and to trigger TMS pulses. An EIZO 19" LCD monitor approximately 1 m in front of the seated participant displayed task-related text (white letters, font: Arial, font size: 32 pt; black background). Shure SE215 sound isolating in-ear headphones (Shure Europe, Eppingen, Germany) were used to deliver L2 word recordings during the TMS sessions. Sound volume was individually adjusted prior to beginning the TMS task. CAUSAL ROLE OF SENSORY CORTICES 15 During each TMS session, a within-participants control condition was included by applying not only effective TMS to the bilateral bmSTS but also sham TMS. Sham TMS coils for each hemisphere were positioned at a 90◦ angle over each stimulation coil, as shown in Fig. 1c, and therefore did not effectively stimulate the brain. Coil locations were monitored and adjusted for head movements during the TMS sessions. The repetitive TMS protocol used (a seven-pulse train of 10 Hz TMS) was in line with published TMS safety guidelines (Rossi et al., 2009). Prior to the TMS translation task, each participant‟s individual stimulation intensity was determined by measuring their resting motor threshold (RMT). To measure RMT, we stimulated the hand region of the left primary motor cortex (M1) using single-pulse TMS, resulting in the conduction of motor-evoked potentials (MEPs) in the relaxed first dorsal interosseous muscle (FDI) of the right hand. The RMT was defined as the lowest stimulation intensity producing 5 MEPs out of 10 consecutive TMS pulses that exceeded a 50 mV peak-to-trough amplitude. A meta-analysis by Mayka and colleagues (2006) provided mean stereotactic coordinates of the left M1 (x, y, z = −37, −21, 58 mm, MNI space), which were used as a starting point to locate the M1 FDI hotspot. The coil used to elicit MEPs was oriented at 45° to the sagittal plane, inducing a PA current flow in the brain. Effective and sham TMS intensity during the L2 translation task was set to 90% of each participant‟s RMT. The same intensity was used for both TMS sessions for each participant (M = 40.1% of maximum stimulator output, SD = 5.6%). Data Analysis All participants who completed the study (n = 22) were included in the analyses. Analysis of response times in the translation tasks. Participants indicated that they recalled the L1 translation prior to the appearance of the four response options during fewer than half of all trials across the two TMS sessions (M = 41.7% of trials, SE = 4.5%), leaving an insufficient number of trials for analysis of the recall task. An exploratory analysis of these data can be found in the supplementary results. In contrast, in the multiple choice task, participants selected a translation from the multiple choice options presented on the screen during M = 88.6% (SE = 3.6%) of trials across the two TMS sessions. In the following we focus the analyses on the response times for the multiple choice task. Response times in the multiple choice task were computed as the time interval from the appearance of the multiple choice options on the screen until the response. Trials in which participants did not respond following the appearance of the multiple choice options, selected CAUSAL ROLE OF SENSORY CORTICES 16 the incorrect translation, or selected the fourth response options ("Unknown / Other word") were excluded from the response time analyses. To test our first hypothesis (see overview and hypotheses subsection of the methods), we ran a two-way repeated measures analysis of variance (ANOVA) with the factors learning condition (gesture, picture) and stimulation type (effective, sham) on response times in the multiple choice task. To evaluate whether the observed patterns of response times were due to speed-accuracy tradeoffs, we correlated response times in the multiple choice translation task with accuracy (percent correct) for each learning condition, stimulation condition, and time point. To test our second hypothesis (see overview and hypotheses subsection of the methods), we ran a three-way repeated measures ANOVA with factors learning condition (gesture, picture), stimulation type (effective, sham), and time point (day 5, month 5) on response times in the multiple choice task. To test our third hypothesis (see overview and hypotheses subsection of the methods), we ran a four-way repeated measures ANOVA on response times in the multiple choice task with factors learning condition (gesture, picture), stimulation type (effective, sham), testing time point (day 5, month 5), and vocabulary type (concrete, abstract). Pairwise comparisons for all analyses were conducted using two-tailed Tukey‟s honestly significant difference (HSD) post-hoc tests. Linear mixed effects modeling of response times in the multiple choice translation task. To evaluate the robustness of the observed effects using an alternate analysis technique, we also tested our three hypotheses using a linear mixed effects modeling approach. We performed backwards model selection to select the model‟s random effects structure, beginning with a random intercept by subject, a random intercept by auditory stimulus, a random slope by subject for each of the four independent factors (stimulation type, learning condition, time point, and vocabulary type), and a random slope by stimulus for the stimulation type and time point factors. We removed random effects terms that accounted for the least variance one by one until the fitted mixed model was no longer singular, i.e., until variances of one or more linear combinations of random effects were no longer (close to) zero. The final model included three random effects terms: a random intercept by subject, a random intercept by stimulus, and a random slope by subject for the time point factor. Please see the supplementary material for more methodological details. Analysis of response accuracy. Besides testing our main hypotheses, the data also allowed us to test the reliability of the previous finding that benefits of gesture performance enrichment on L2 translation exceeded those of picture viewing enrichment over the long-term CAUSAL ROLE OF SENSORY CORTICES 17 (Mayer et al., 2015). We ran a four-way repeated measures ANOVA on accuracy in the multiple choice task, with the factors learning condition (gesture, picture), stimulation type (effective, sham), testing time point (day 5, month 5), and vocabulary type (concrete, abstract), and examined all interactions involving the learning and time point factors. Results Stimulation of the bmSTS slows the translation of gesture-enriched foreign vocabulary Our first and primary hypothesis was that a brain region specialized in the perception of biological motion, the bmSTS (Grossman et al., 2000), causally contributes to L2 translation following gesture-enriched L2 learning, but not picture-enriched L2 learning. We therefore first tested whether bmSTS stimulation modulated L2 translation, irrespective of testing time point. The results confirmed our hypothesis. A two-way ANOVA on response times in the multiple choice task revealed a stimulation type × learning condition interaction (F 1,21 = 11.82, p = .002, two-tailed, .36) (see Table S3, supplementary material, for the full set of ANOVA results). Tukey‟s HSD post-hoc tests revealed that response times for words that had been learned with gesture enrichment were significantly delayed when TMS was applied to the bmSTS compared to sham stimulation (p = .005, Hedge‟s g = .33). This was not the case for words learned with picture enrichment. This indicates that perturbation of a brain area related to biological motion slowed the translation of L2 words that had been learned with gestures, but not of L2 words learned with pictures (Fig. 2). Figure 2. Effects of bmSTS stimulation on speed of L2 translation. Bilateral bmSTS stimulation slowed the translation of L2 vocabulary learned using gestures compared to sham stimulation in the multiple choice task. There was no such effect for L2 vocabulary learned using pictures. The mean of each condition across time points (5 days and 5 months following 90010001100GesturePictureResponse time (ms)Learning typeTMSSham*n.s.**aMean across time points8009001000GesturePictureResponse time (ms)Learning type95011501350GesturePictureLearning typeTMSSham**Day 5Month 5n.s.*bn.s.n.s.n.s.90010001100GesturePictureResponse time (ms)Learning typeTMSSham*n.s.**aMean across time points8009001000GesturePictureResponse time (ms)Learning type95011501350GesturePictureLearning typeTMSSham**Day 5Month 5n.s.*bn.s.n.s.n.s. CAUSAL ROLE OF SENSORY CORTICES 18 the start of learning) is shown (n = 22 participants). Error bars represent one standard error of the mean. *p < .05, **p < .01. In a control analysis, we tested whether differences in response times under effective stimulation compared to sham stimulation conditions could be due to tradeoffs between translation speed and accuracy. Response times for correct answers in the multiple choice translation task were correlated with accuracy (percent correct) for each learning condition, stimulation condition, and time point. If there were a speed-accuracy tradeoff, one would expect a positive correlation between response times and accuracy (i.e., the longer the response time, the greater the accuracy). Response times, however, did not correlate or correlated negatively with translation accuracy (Table 1). Thus, participants did not trade speed for accuracy. Day 5 TMS Sham r (p) r (p) Gesture -.84 (<.001)* -.89 (<.001)* Picture -.89 (<.001)* -.95 (<.001)* Month 5 TMS Sham r (p) r (p) -.63 (.002)* -.34 (.12) -.48 (.02) -.46 (.03) Table 1. Speed-accuracy relationships in L2 translation. In most tests, slower response times correlated with lower translation accuracy, indicating that there was no speed-accuracy tradeoff. df = 20 for all correlations. *p < .05, Bonferroni corrected. bmSTS supports auditory foreign vocabulary translation 5 months post-learning Our second hypothesis was that bmSTS integrity would support the auditory translation of gesture-enriched words at the later time point (5 months post-learning) even more than the earlier time point (5 days following the start of learning). In agreement with this hypothesis, a three-way ANOVA on response times for the multiple choice task yielded a significant three-way stimulation type × learning condition × time point interaction (F 1, 21 = 7.51, p = .012, two-tailed, .26) (see Table S4, supplementary material, for the full set of ANOVA results). Tukey‟s HSD post-hoc tests CAUSAL ROLE OF SENSORY CORTICES 19 revealed a response benefit (faster responses) for words learned with gesture enrichment compared to words learned with picture enrichment under sham stimulation 5 months following learning (p < .001, Hedge‟s g = .69). The application of TMS to the bmSTS negated this benefit: Response times for gesture- and picture-enriched words did not significantly differ at month 5 under effective stimulation, and responses were significantly slower under effective stimulation compared to sham stimulation for words learned with gesture enrichment (p = .001, Hedge‟s g = .61). In sum, significant effects of bmSTS stimulation on translation were more prominent 5 months following the L2 training period compared to 5 days following the start of learning (Fig. 3). Figure 3. Effects of bmSTS stimulation on speed of L2 translation by time point. Effects of bmSTS stimulation on response times in the multiple choice task occurred 5 months following learning (n = 22 participants). Error bars represent one standard error of the mean. *p < .05, **p < .01. Role of L2 vocabulary concreteness Next, we tested our third hypothesis that the disruptive effects of bmSTS stimulation would occur independent of whether a word was classified as concrete or abstract (see also supplementary introduction). A four-way ANOVA on translation response times in the multiple choice task yielded a significant four-way learning condition × stimulation type × time point × vocabulary type interaction (F 1, 21 = 5.24, p = .033, two-tailed, .20) (see Table S5, supplementary material, for the full set of ANOVA results). Tukey‟s HSD post-hoc tests revealed that concrete nouns paired with gestures during learning were translated significantly more slowly during bmSTS stimulation compared to sham stimulation at day 5 (p = .05, Hedge‟s g = .31; Fig. 4). Contrary to our hypothesis, this comparison was not significant for abstract nouns at day 5. At month 5, however, TMS significantly slowed the translation of both L2 word types following gesture-enriched learning (concrete words: p = .002, Hedge‟s g = .44; abstract words: p < .001, 90010001100GesturePictureResponse time (ms)Learning typeTMSSham*n.s.**aMean across time points8009001000GesturePictureResponse time (ms)Learning type95011501350GesturePictureLearning typeTMSSham**Day 5Month 5n.s.*bn.s.n.s.n.s. CAUSAL ROLE OF SENSORY CORTICES 20 Hedge‟s g = .48). Response times under effective and sham stimulation did not significantly differ for words of either type that were learned in the picture enrichment condition at either time point. In sum, stimulation of the bmSTS modulated the translation of the concrete gesture-enriched L2 vocabulary at the earlier time point, and the translation of both concrete and abstract gesture- enriched vocabulary at the later time point. Figure 4. Effects of bmSTS stimulation on speed of L2 translation by vocabulary type. L2 vocabulary translation response times at the day 5 TMS session (left) and month 5 TMS session (right) by stimulation type, learning type, and vocabulary type (n = 22 participants). Compared to sham stimulation, stimulation of the bmSTS delayed response selection for concrete gesture- enriched nouns at day 5 and for both concrete and abstract gesture-enriched nouns at month 5. Error bars represent one standard error of the mean. *p < .05, **p < .01. Testing of our three hypotheses using linear mixed effects modeling yielded qualitatively similar results as the ANOVA-based approach reported heretofore, with the exception of the four-way interaction, which was significant in the four-way ANOVA but not in the mixed effects model analysis. Please see the supplementary material (Table S2) for the full mixed effects model results. 8009001000GesturePictureResponse time (ms)Learning typeAbstract Nouns8009001000GesturePictureResponse time (ms)Learning type*95011501350GesturePictureLearning typeAbstract Nouns95011501350GesturePictureLearning typeTMSShamConcrete nounsAbstractnouns****n.s.Day 5Month 5*n.s.*n.s.*n.s.n.s.n.s. CAUSAL ROLE OF SENSORY CORTICES 21 Gesture-enriched training facilitates long-term L2 translation accuracy Besides testing our main hypotheses related to effects of TMS on translation, the data also allowed us to test the reliability of the previous finding that benefits of gesture performance enrichment on L2 translation exceed those of picture viewing enrichment over the long-term (Mayer et al., 2015). To test this, we conducted a four-way ANOVA on translation accuracy scores in the multiple choice task (percent correct) with the factors learning condition, stimulation type, time point, and vocabulary type. The ANOVA revealed a significant learning condition × time point interaction (F 1, 21 = 6.86, p = .016, two-tailed, .25) (see Table S6, supplementary material, for the full set of ANOVA results). Tukey‟s HSD post-hoc tests revealed greater response accuracy following gesture-enriched learning compared to picture-enriched learning at month 5 (p = .035, Hedge‟s g = .11), which did not occur at day 5, suggesting that gesture-enrichment-based benefits on response accuracy emerged over a period of several months (Fig. 5). This finding is consistent with the previous report that gesture enrichment outperforms picture enrichment over longer timescales (Mayer et al., 2015). Figure 5. Accuracy of L2 translation following learning. Learning condition and time point variables in the multiple choice task significantly interacted: Participants translated gesture- enriched L2 words more accurately than picture-enriched L2 words at month 5 only (n = 22 participants). We report here, for completeness, further tests and results related to multiple choice task accuracy. As expected, there were no significant effects of stimulation type on accuracy for either vocabulary type or learning condition at either time point (Fig. 6). However, there was an unexpected significant four-way learning condition × stimulation type × time point × vocabulary type interaction (F 1, 21 = 8.23, p = .009, two-tailed, .28) attributable to a significant difference in response accuracy between concrete and abstract gesture-enriched -- but not picture-enriched -- 255075100Day 5Month 5Time point GesturePicture* n.s. * Accuracy (% correct) CAUSAL ROLE OF SENSORY CORTICES 22 words at month 5 under sham stimulation (p < .001, Hedge‟s g = .79). Participants translated concrete words significantly more accurately overall in the multiple choice task than abstract words, a main effect of vocabulary type (F 1, 21 = 35.62, p < .001, two-tailed, = .63). This effect was expected based on previous studies (Macedonia & Klimesch, 2014; Macedonia & Knösche, 2011), Figure 6. Accuracy of L2 translation depending on learning condition, stimulation type, time point, and vocabulary type. As expected, no significant effects of TMS on L2 translation accuracy in the multiple choice task were observed at either time point (n = 22 participants). Participants were also significantly less accurate in the multiple choice task at month 5 compared to day 5, a main effect of time point (F 1, 21 = 124.77, p < .001, two-tailed, .86), suggesting that L2 memory representations decayed over time. To test whether L2 memory decayed less following gesture-enriched learning then picture-enriched learning, we computed changes in multiple choice task accuracy (percent correct) across the two time points (month 5 -- day 5) in the absence of neurostimulation. A two-way ANOVA on changes in translation accuracy 255075100TMSShamAccuracy (% correct)Stimulation typeAbstract Nounsn.s.n.s.n.s.Concrete nounsAbstractnounsDay 5Month 5255075100TMSShamAccuracy (% correct)Stimulation type255075100TMSShamStimulation typeGesturePicture255075100TMSShamStimulation typeAbstract Nounsn.s.n.s.n.s.n.s.n.s.n.s.n.s.n.s.n.s. CAUSAL ROLE OF SENSORY CORTICES 23 with the factors learning condition and vocabulary type yielded a significant learning condition × vocabulary type interaction (F 1, 21 = 13.84, p = .001, two-tailed, = .40). Tukey‟s HSD post-hoc tests revealed a greater decrease in translation accuracy over the 5-month interval following picture-enriched learning compared to gesture-enriched learning for concrete words only (p = .009, Hedge‟s g = .58) (see Table S7, supplementary material, for the full set of ANOVA results). Thus, gesture-enriched representations of concrete L2 words decayed less than picture-enriched L2 representations (Fig. S1a, supplementary results). This result cannot be explained by a tradeoff in accuracy and response time, as the same two-way ANOVA on changes in response times (month 5 -- day 5) with factors learning condition and vocabulary type yielded qualitatively similar results: Response times increased more over the 5-month interval following picture-enriched learning compared to gesture-enriched learning for both vocabulary types (Fig. S1b, supplementary results), a main effect of learning condition (F 1,21 = 11.05, p = .003, two-tailed, = .34) (see Table S8, supplementary material, for the full set of ANOVA results). Thus, gesture-enriched learning benefitted translation response times more than picture-enriched learning over the long-term, further indicating the greater robustness of gesture-enriched L2 representations in memory over a long timescale (see supplementary results for more details). Participants‟ scores on the paper-and-pencil vocabulary tests, which were completed on days 2, 3, and 4 of the L2 training period at 5 months post-training, converged on a similar pattern of results as the multiple choice decay analyses (for details, please see the supplementary results, Fig. S2 and Fig. S3). Discussion This study revealed causal links between responses in specialized sensory cortices and facilitative effects of sensorimotor-enriched learning. There were three main findings. First, behavioral benefits of gesture-enriched learning were caused in part by responses within a specialized visual brain area, the bmSTS; this area was causally engaged in the auditory translation of gesture-enriched but not picture-enriched L2 words. Second, bmSTS integrity supported the auditory translation of gesture-enriched words at 5 months post-learning even more than 5 days post-learning. Third, bmSTS integrity supported the translation of both concrete and abstract L2 words; stimulation effects were observed for concrete nouns at the earlier time point, and for both word types at the later time point. Taken together, these findings show that sensorimotor-enriched teaching constructs strong associations between auditory L2 words and their L1 translations by way of representations arising from specific visual cortices. Robust long- CAUSAL ROLE OF SENSORY CORTICES 24 term memory representations established by sensorimotor-enriched teaching can therefore be supported by task-specific, specialized sensory brain responses. The causal relation observed between bmSTS responses and L2 translation adjudicates between influential reactivation (Fuster, 2009; Nyberg et al., 2000; Wheeler et al., 2000) and predictive coding (Friston, 2012; von Kriegstein, 2012) theories of multisensory learning. The fact that brain responses in one sensory modality (e.g., visual) can improve task performance in another modality (e.g., auditory), depending on associations forged during learning, is expected based on predictive coding theories but not reactivation theories. Reactivation theories do not, in general, consider the idea that reactivated brain regions could serve some kind of role in perception. Given that knowledge of the gesture associated with an L2 word was not critical for achieving success in the multiple choice task, reactivation theories would presume that bmSTS integrity would not contribute to task performance and that inhibitory stimulation of the bmSTS would not interfere with completing the task. Though both gesture- and picture-enriched training involved complementary visual information, disruptive effects of bmSTS stimulation occurred only for the condition that contained stimulus information related to biological motion. Therefore, bmSTS engagement depended on sensorimotor experience. Based on predictive coding theories, we would expect motor and somatosensory stimulation to similarly disrupt the translation of gesture- but not picture-enriched words (for preliminary results, see Mathias et al., 2019), and expect LOC stimulation to disrupt the translation of picture- but not gesture-enriched words. Our "virtual lesion" TMS approach took advantage of the focal spatial resolution of TMS to transiently interfere with processing in a specific cortical target (Sack et al., 2007). If the current results were due to a whole- brain effect of TMS rather than one localized to the STS region, TMS would have lengthened response times not only for gesture-enriched words, but also picture-enriched words. The performance of visually-modeled gestures yielded beneficial long-term effects on L2 translation accuracy and lessened long-term L2 decay compared to picture-enriched learning. Gesture enrichment facilitated learning, in part, by establishing representations of learned information within specific visual cortices. In our experiment, we characterize "learning by doing" as "sensorimotor-enriched learning" rather than "motor-enriched learning" because motor components of gesture-based enrichment can never be fully separated from associated sensory components. Even if learners performed self-created gestures without viewing a model, they would still receive visual feedback from their own and others‟ body movements, as well as other types of movement- associated sensory feedback. Learning by doing inevitably involves the integration of sensory and motor aspects of one‟s experience. CAUSAL ROLE OF SENSORY CORTICES 25 In order to recall the meaning of a newly-acquired L2 word, the brain may internally simulate sensory and motor processes that were involved in learning that word. This view is consistent with the notion that the presence of additional dimensions (e.g., visual) along which stimuli can be evaluated during recognition underlie learning-by-doing-based benefits (MacLeod, Gopie, Hourihan, Neary, & Ozubko, 2010). Mayer and colleagues (2015) found that viewing videos of gestures did not benefit post-training performance compared to auditory-only learning. However, behavioral outcomes following the video-viewing condition also correlated with decoded bmSTS responses measured using fMRI. This could indicate that the bmSTS is also engaged if learning involves viewing gestures and that other regions encode the visually-enriched vocabulary less efficiently, but that the bmSTS is unable to compensate for these deficiencies. Whether bmSTS stimulation would disrupt the translation of vocabulary learned by viewing gestures (and not performing them) is an open question. On the basis of previous fMRI results (Mayer et al., 2015), we reason that stimulation would disrupt performance. A growing literature has reported positive effects of arousal-based interventions such as physical exercise (Hötting, Schickert, Kaiser, Röder, & Schmidt-Kassow, 2016), emotion regulation (Storbeck, & Maswood, 2016), and even music (Schellenberg, Nakata, Hunter, & Tamoto, 2007) on cognitive task performance. Though effective, these approaches do not encode associations between different components of the word acquisition experience in the same way as gesture-enriched learning, as gestures are intrinsically bound to specific stimulus information (Markant, Ruggeri, Gureckis, & Xu, 2016). If behavioral benefits of enrichment were due solely to increased arousal during gesture-enriched learning compared to picture-enriched learning, then stimulation of a specialized visual area would not have disrupted those benefits. Further, any potential differences between gesture- and picture-enriched learning in terms of arousal were not large enough to distinguish these conditions in terms of performance accuracy. Previously, the combination of a motor task with picture viewing (tracing an outline of presented pictures) during L2 learning benefitted learning outcomes less than simply viewing pictures without performing any movements (Mayer et al., 2015), and the performance of semantically-related gestures enhanced learning outcomes compared to the performance of meaningless gestures (Macedonia, Müller, & Friederici, 2011). These outcomes suggest that gesture enrichment benefits cannot be explained simply by the presence of movement during learning. The current results therefore steer away from more general explanations for beneficial effects of sensorimotor-enriched or multisensory-enriched learning such as increased arousal or attention. Hence, teaching strategies may be advanced by establishing links between new information and congruent sensorimotor and multisensory enrichment. CAUSAL ROLE OF SENSORY CORTICES 26 We conclude that sensorimotor-enriched training constructs stronger associations between auditory L2 words and their L1 translations than commonly-practiced sensory-only methods in adults. Beneficial behavioral effects of sensorimotor-enriched training are caused in part by responses within specialized sensory brain regions. Spoken language perception may therefore rely not only on auditory information stored in memory, but also on the sensorimotor context in which words are experienced. The causal relation observed between sensory brain responses and behavioral performance significantly advances our knowledge of neuroscientific mechanisms contributing to benefits of sensorimotor-enriched learning. The current findings also shed new light on the idea that sensorimotor-enriched teaching practices can be used to enhance learning outcomes by linking sensory brain functions with behavioral performance, and may have repercussions for the ways in which current classroom teaching practices are evaluated. Acknowledgements This work was funded by German Research Foundation grant KR 3735/3-1, a Max Planck Research Group grant to K.v.K, and an Erasmus Mundus Postdoctoral Fellowship. B.M. is also supported by the European Research Council Consolidator Grant SENSOCOM 647051 to K.v.K. We thank Frieder Schillinger for comments on a previous version of the manuscript. Author Contributions M.M., K.M.M., and K.v.K. conceived the study. B.M., G.H., K.v.K. developed the study design. B.M. and L.S. acquired and analyzed the data. B.M. and K.v.K. wrote the manuscript. L.S., G.H., M.M., and K.M.M. contributed to writing the manuscript. Open Practices Statement The experiment reported in this article was not formally preregistered. The hypotheses were derived directly from the findings of a previous paper (Mayer et al., 2015). Neither the data nor the materials have been made available on a permanent third-party archive; requests for the data or materials can be sent via email to the lead author at [email protected]. References Ashbridge, E., Walsh, V., & Cowey, A. (1997). Temporal aspects of visual search studied by transcranial magnetic stimulation. Neuropsychologia, 35(8), 1121-1131. Barsalou, L. W. (2010). Grounded cognition: Past, present, and future. Topics in Cognitive Science, 2(4), 716-724. CAUSAL ROLE OF SENSORY CORTICES 27 Butler, A. J., & James, K. H. (2011) Cross-modal versus within-modal recall: Differences in behavioral and brain responses. Behavioral Brain Research, 224, 387-396. Brickenkamp, R. (2002). Test d2: Aufmerksamkeits-Belastungs-Test; 9. überarbeitete und neu normierte Auflage. Göttingen, Germany: Verlag für Psychologie Hogrefe. Choo, L. B., Lin, D. T. A., & Pandian, A. (2012). Language learning approaches: A review of research on explicit and implicit learning in vocabulary acquisition. Procedia-Social and Behavioral Sciences, 55, 852-860. Danker, J. F., & Anderson, J. R. (2010). The ghosts of brain states past: remembering reactivates the brain regions engaged during encoding. Psychological Bulletin, 136(1), 87. Freeman, S., Eddy, S. L., McDonough, M., Smith, M. K., Okoroafor, N., Jordt, H., & Wenderoth, M. P. (2014). Active learning increases student performance in science, engineering, and mathematics. Proceedings of the National Academy of Sciences, 111(23), 8410- 8415. Friston, K. (2012). Prediction, perception, and agency. International Journal of Psychophysiology, 83(2), 248-252. Fuster, J. M. (2009). Cortex and memory: Emergence of a new paradigm. Journal of Cognitive Neuroscience, 21(11), 2047-2072. Grossman, E., Donnelly, M., Price, R., Pickens, D., Morgan, V., Neighbor, G., & Blake, R. (2000). Brain areas involved in perception of biological motion. Journal of Cognitive Neuroscience, 12(5), 711-720. Hartwigsen, G., Bzdok, D., Klein, M., Wawrzyniak, M., Stockert, A., Wrede, K., ... & Saur, D. (2017). Rapid short-term reorganization in the language network. Elife, 6, e25964. Hoffman, P. (2016). The meaning of „life‟ and other abstract words: Insights from neuropsychology. Journal of Neuropsychology, 10(2), 317-343. Hötting, K., Schickert, N., Kaiser, J., Röder, B., & Schmidt-Kassow, M. (2016). The effects of acute physical exercise on memory, peripheral BDNF, and cortisol in young adults. Neural Plasticity, 1-12. James, T. W., & Stevenson, R. A. (2012). The use of fMRI to assess multisensory integration. In M. M. Murray, M. T. Wallace (Eds.), The neural bases of multisensory processes. Boca Raton: CRC Press/Taylor & Francis. Korkman, M., U. Kirk, & Kemp, S. (1998). NEPSY: A Developmental Neuropsychological Assessment. Psychological Corporation. CAUSAL ROLE OF SENSORY CORTICES 28 Lahav, A., Saltzman, E., & Schlaug, G. (2007). Action representation of sound: audiomotor recognition network while listening to newly acquired actions. Journal of Neuroscience, 27(2), 308-314. Macedonia, M. (2014). Bringing back the body into the mind: gestures enhance word learning in foreign language. Frontiers in Psychology, 5, 1467. Macedonia, M., & Klimesch, W. (2014). Long‐term effects of gestures on memory for foreign language words trained in the classroom. Mind, Brain, and Education, 8(2), 74-88. Macedonia, M., & Knösche, T. R. (2011). Body in mind: How gestures empower foreign language learning. Mind, Brain, and Education, 5(4), 196-211. Macedonia, M., & Müller, K. (2016). Exploring the neural representation of novel words learned through enactment in a word recognition task. Frontiers in Psychology, 7, 953. MacLeod, C. M., Gopie, N., Hourihan, K. L., Neary, K. R., & Ozubko, J. D. (2010). The production effect: Delineation of a phenomenon. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36(3), 671-685. Markant, D. B., Ruggeri, A., Gureckis, T. M., & Xu, F. (2016). Enhanced memory as a common effect of active learning. Mind, Brain, and Education, 10(3), 142-152. Mathias, B., Klingebiel, A., Hartwigsen, G., Sureth, L., Macedonia, M., Mayer, K., & von Kriegstein, K. (2019). Sensorimotor cortices casually contribute to auditory foreign language vocabulary translation following multisensory learning. Brain Stimulation: Basic, Translational, and Clinical Research in Neuromodulation, 12(2), 401-402. Matusz, P. J., Wallace, M. T., & Murray, M. M. (2017). A multisensory perspective on object memory. Neuropsychologia, 105, 243-252. Mayer, K. M., Yildiz, I. B., Macedonia, M., & von Kriegstein, K. (2015). Visual and motor cortices differentially support the translation of foreign language words. Current Biology, 25(4), 530-535. Mayka, M. A., Corcos, D. M., Leurgans, S. E., & Vaillancourt, D. E. (2006). Three-dimensional locations and boundaries of motor and premotor cortices as defined by functional brain imaging: a meta-analysis. Neuroimage, 31(4), 1453-1474. Meltzoff, A. N., Kuhl, P. K., Movellan, J., & Sejnowski, T. J. (2009). Foundations for a new science of learning. Science, 325(5938), 284-288. Mitchell, D. B., Brown, A. S., & Murphy, D. R. (1990). Dissociations between procedural and episodic memory: Effects of time and aging. Psychology and Aging, 5(2), 264. CAUSAL ROLE OF SENSORY CORTICES 29 Neubauer, A., & Horn, R. V. (2006). Wechsler Intelligenztest für Erwachsene: WIE. In M. von Aster (Ed.), Übersetzung und Adaption der WAIS-III. Frankfurt, Germany: Harcourt Test Services. Nyberg, L., Habib, R., McIntosh, A. R., & Tulving, E. (2000). Reactivation of encoding-related brain activity during memory retrieval. Proceedings of the National Academy of Sciences, 97(20), 11120-11124. Pascual-Leone, A., Wassermann, E. M., Grafman, J., & Hallett, M. (1996). The role of the dorsolateral prefrontal cortex in implicit procedural learning. Experimental Brain Research, 107(3), 479-485. Pulvermüller, F., Hauk, O., Nikulin, V. V., & Ilmoniemi, R. J. (2005). Functional links between motor and language systems. European Journal of Neuroscience, 21(3), 793-797. Ramsey, J. D., Hanson, S. J., Hanson, C., Halchenko, Y. O., Poldrack, R. A., & Glymour, C. (2010). Six problems for causal inference from fMRI. Neuroimage, 49(2), 1545-1558. Robertson, E. M., Pascual-Leone, A., & Miall, R. C. (2004). Current concepts in procedural consolidation. Nature Reviews Neuroscience, 5(7), 576. Rossi, S., Hallett, M., Rossini, P. M., & Pascual-Leone, A. (2009). Safety, ethical considerations, and application guidelines for the use of transcranial magnetic stimulation in clinical practice and research. Clinical Neurophysiology, 120(12), 2008-2039. Sack, A. T., Kohler, A., Bestmann, S., Linden, D. E., Dechent, P., Goebel, R., & Baudewig, J. (2007). Imaging the brain activity changes underlying impaired visuospatial judgments: simultaneous FMRI, TMS, and behavioral studies. Cerebral Cortex, 17(12), 2841-2852. Schwanenflugel, P. J. (1991). Why are abstract concepts hard to understand?. In P. J. Schwanenflugel (Ed.), The psychology of word meanings (pp. 223-250). Hillsdale, NJ: Erlbaum. Schellenberg, E. G., Nakata, T., Hunter, P. G., & Tamoto, S. (2007). Exposure to music and cognitive performance: Tests of children and adults. Psychology of Music, 35(1), 5-19. Scimeca, J. M., Kiyonaga, A., & D‟Esposito, M. (2018). Reaffirming the sensory recruitment account of working memory. Trends in Cognitive Sciences, 22(3), 190-192. Shams, L., & Seitz, A. R. (2008). Benefits of multisensory learning. Trends in Cognitive Sciences, 12(11), 411-417. Siebner, H. R., & Ziemann, U. (Eds.). (2007). Das TMS-Buch: Handbuch der transkraniellen Magnetstimulation. Springer-Verlag. CAUSAL ROLE OF SENSORY CORTICES 30 Storbeck, J., & Maswood, R. (2016). Happiness increases verbal and spatial working memory capacity where sadness does not: Emotion, working memory and executive control. Cognition and Emotion, 30(5), 925-938. von Kriegstein, K. (2012). A multisensory perspective on human auditory communication. In M. M. Murray, M. T. Wallace (Eds.), The neural bases of multisensory processes (pp. 683- 700). Boca Raton: CRC Press/Taylor & Francis. Wheeler, M. E., Petersen, S. E., & Buckner, R. L. (2000). Memory's echo: vivid remembering reactivates sensory-specific cortex. Proceedings of the National Academy of Sciences, 97(20), 11125-11129. von Kriegstein, K., & Giraud, A. L. (2006). Implicit multisensory associations influence voice recognition. PLoS biology, 4(10), e326. CAUSAL ROLE OF SENSORY CORTICES 31 Supplementary Material Influences of the conceptual perceptibility of L2 word referents Introduction. A potentially limiting factor in the success of sensorimotor-enriched approaches to L2 vocabulary learning may arise from the conceptual perceptibility of word referents. Conceptual perceptibility refers to the extent to which referents can be perceived by the body‟s sensory systems (e.g., tangibility; Hoffman, 2016). The referent of the concrete noun ball, for example, is highly tangible and can be iconically represented by using one‟s arms to throw an imaginary ball. Referents of other words, such as the abstract noun mentality, are less tangible and more difficult to convey using gestures or pictures. Despite differences in terms of intrinsic sensorimotor associations, the learning of both concrete and abstract vocabulary has been shown to benefit from gesture and picture forms of enrichment (Macedonia, 2014; Mayer, Yildiz, Macedonia, & von Kriegstein, 2015), suggesting that sensory brain regions should contribute to learning benefits for concrete and abstract words. Sensorimotor facilitation of the learning of abstract nouns may function by building associations between abstract concepts and perceptible sensory and motor events. The L2 translation of the word innocence, for example, is difficult to learn if paired simply with its native language translation. It becomes easier to learn if paired with the iconic gesture of shrugging of one‟s shoulders, even though innocence is not defined as shrugging (Macedonia & Knösche, 2011). Given that enriched learning may establish sensorimotor associations with both concrete and abstract words, the third aim of the current study was to address whether sensory brain regions differentially contribute to gesture-enriched learning benefits for these word types. Given that gesture-enriched learning previously benefitted concrete and abstract words similarly (Mayer et al., 2015), we hypothesized that specialized visual sensory responses would contribute to the translation of both word types. The concrete and abstract vocabulary used in the study is shown in Table S1. Discussion of findings. We found that bmSTS stimulation inhibited the translation of only concrete words at the earlier time point, and of both gesture-enriched concrete and abstract words at the later time point. We can only speculate on why this might be the case. Concrete concepts may map more easily onto gestures compared to abstract concepts. This may have resulted in learners‟ greater reliance on alternate learning strategies for translation of abstract L2 words at the earlier time point. Nevertheless, bmSTS stimulation inhibited the translation of both gesture- enriched concrete and abstract words at the later time point, consistent with the previous demonstration of greater long-term memory benefits for both word types following gesture enrichment compared to picture enrichment (Mayer et al., 2015). This result suggests that, if CAUSAL ROLE OF SENSORY CORTICES 32 alternate strategies were used for the abstract words, they were used only in the short-term, and learners relied over the long-term instead on sensorimotor representations. Table S1. Vocabulary used in the experiment. 90 Vimmi and German words, and their English translations. Assignment of words to the gesture learning condition and the picture learning control condition was counterbalanced across participants, ensuring that each Vimmi word was represented equally in both learning conditions. Concrete nouns Abstract nouns German Ampel Anhänger Balkon Ball Bett Bildschirm Briefkasten Decke Denkmal English traffic light trailer balcony ball bed monitor letter box ceiling memorial Eintrittskarte Faden Fahrrad Fenster Fernbedienung remote control entrance ticket thread bicycle window Flasche Flugzeug Gemälde Geschenk Gitarre Handtasche Kabel Kamera Kasse Katalog Kleidung Koffer Maschine Maske Papier Reifen Ring bottle airplane painting present guitar purse cable camera till catalog clothes suitcase machine mask paper tire ring Rucksack Sammlung Schlüssel Schublade Sonnenbrille backpack collection key drawer sunglasses Vimmi gelori afugi usito miruwe suneri zelosi abota siroba frinupo edafe kanede sokitu uribo wilbano aroka wobeki bifalu zebalo masoti diwume zutike lamube asemo gebamo wiboda mewima nelosi epota serawo wasute guriwe lofisu etuko abiru lutepa woltume English Cancellation Alternative requirement Arrival German Absage Alternative Anforderung Ankunft Aufmerksamkeit Attention Aufwand Aussicht Befehl Besitz Effort View Command Property Vimmi munopa mofibu utike matilu fradonu muladi gaboki magosa mesako Bestimmung Bitte Disziplin Empfehlung Gedanke wefino Destination pokute Plea Discipline motila recommendation giketa Thought atesi Geduld Gleichgültigkeit Information Korrektur Langeweile Mentalität Patience Indifference Information Correction Boredom Mentality Methode Mut Partnerschaft Rücksicht Sensation Stil Talent Tatsache Teilnahme Tendenz Theorie Therapie Tradition Triumph Übung Unschuld Method Bravery Partnership consideration Sensation Style Talent Fact Participation Tendency Theory Therapy Tradition Triumph Exercise Innocence dotewa frugazi sapezo fapoge elebo gasima efogi wirgonu nabita ukowe boruda lifawo puneri botufe pamagu pefita sigule giwupo uladi gepesa fremeda dafipo CAUSAL ROLE OF SENSORY CORTICES 33 Spiegel Strassenbahn Tageszeitung Telefon Teller Teppich Verband dubeki mirror tram umuda daily newspaper gokasu telephone plate carpet bandage esiwu buliwa batewo magedu Wohltat Veränderung Verständnis Vorgehen Vorwand Warnung Wohlstand Change Sympathy Procedure Excuse Warning Wealth Benefaction zalefa gorefu denalu Pirumo Gubame Bekoni migedu Analysis of long-term decay of L2 translation speed and accuracy To test whether memory decay over time was significantly less severe following gesture- enriched learning compared to picture-enriched learning, we computed changes in multiple choice task accuracy (percent correct) across the two time points (month 5 -- day 5). Only sham condition accuracy was evaluated, in order to assess differences between gesture- and picture-enriched learning in the absence of neurostimulation. If gesture-enriched learning results in more robust L2 representations overall than picture-enriched learning, one might expect those representations to also decay less over time. A two-way ANOVA on changes in multiple choice response time across the two testing time points (month 5 -- day 5) with the factors learning condition and vocabulary type revealed a main effect of learning condition (F 1,21 = 11.05, p = .003, two-tailed, = .34) (see Table S7 for the full set of ANOVA results).. A greater increase in response times occurred over the 5-month interval following picture-enriched learning compared to gesture-enriched learning, for both concrete nouns and abstract nouns (Fig. S1a). There were no other significant main effects or interactions. Thus, gesture-enriched learning benefitted translation response times more than picture-enriched learning over the long-term, indicating the greater robustness of gesture-enriched L2 representations in memory over a long timescale. Figure S1. Long-term change in accuracy and speed of L2 translation following learning. a, Compared to traditional audiovisual (picture-enriched) learning, the performance 204060ConcreteAbstractAccuracy difference (Day 5 − Month 5, % correct)Vocabulary typeGesturePicture**n.s.**100300500ConcreteAbstractVocabulary typeGesturePicture**Response time difference (Month5 − Day 5, ms)ba** CAUSAL ROLE OF SENSORY CORTICES 34 and viewing of iconic gestures during L2 vocabulary learning resulted in less decay of concrete L2 word knowledge over a 5-month period in the absence of neurostimulation compared to picture-enriched learning (n = 22 participants). b, Gesture-enriched learning also resulted in less of an increase in translation response times at 5 months following training in the absence of neurostimulation (n = 22 participants). Error bars represent one standard error of the mean. **p < .01. This result cannot be explained by a tradeoff in accuracy and response time, as a two-way ANOVA on changes in response times (month 5 -- day 5) with factors learning condition and vocabulary type the same two-way analysis of the response time variable (month 5 -- day 5 difference) yielded qualitatively similar results: Response times increased more over the 5-month interval following picture-enriched learning compared to gesture-enriched learning for both vocabulary types (Fig. S1b, supplementary results), a main effect of learning condition (F 1,21 = = .34) (see Table S8, supplementary material, for the full set of 11.05, p = .003, two-tailed, ANOVA results). There were no other significant main effects or interactions. Thus, gesture- enriched learning benefitted translation response times more than picture-enriched learning over the long-term, further indicating the greater robustness of gesture-enriched L2 representations in memory over a long timescale. Analysis of response times in the exploratory recall task In the recall task, response time was defined as the time elapsed between the start of the auditory L2 word presentation and the participant‟s indication by button press (prior to the appearance of the four response options) that they knew the L1 translation of the presented L2 word. Participants indicated that they recalled the L1 translation prior to the appearance of the four response options during fewer than half of all trials across the two TMS sessions (M = 41.7% of trials, SE = 4.5%), leaving an insufficient number of trials for analysis of this exploratory task component. We nevertheless explored the data and analyzed the recall response times for correct response trials (in response to a reviewer‟s request). In order to evaluate recall response times for correct response trials, we analyzed trials in which participants first indicated by button press that they recalled the L1 translation and subsequently selected the correct translation from the list of response options presented on the screen. A four-way ANOVA on recall response times for correct response trials with factors learning condition, stimulation type, time point, and vocabulary type yielded a significant main effect of time point, (F 1, 21 = 86.66, p < .001, two-tailed, .80). Recall response times were CAUSAL ROLE OF SENSORY CORTICES 35 significantly faster at day 5 than month 5. There was, however, no significant main effect of vocabulary type (p = .96), which was one of the most robust effects throughout our other dependent measure, i.e., the multiple choice task reported in the main manuscript. Recall response times for concrete words (M = 1527 ms, SE = 26 ms) did not differ from response times for abstract words (M = 1526 ms, SE = 23 ms). The ANOVA yielded a significant learning condition × vocabulary type interaction (F 1, 21 = 6.52, p = .019, two-tailed, .24), and significant learning condition × time point × vocabulary type interaction (F 1, 21 = 4.38, p = .049, two-tailed, .17). However, Tukey‟s HSD post-hoc tests revealed no significant differences between concrete and abstract noun response times within any time point or learning condition. The predicted two-way interaction between learning condition and stimulation type variables was also not significant (p = .49). Response times did not significantly differ between any conditions (TMS-Gesture: M = 1506 ms, SE = 33 ms; Sham-Gesture: M = 1536 ms, SE = 32 ms; TMS-Picture: M = 1532 ms, SE = 37 ms; Sham-Picture: M = 1533 ms, SE = 36 ms). There were no other significant main effects or interactions. Given that not even the robust difference between concrete and abstract vocabulary types emerged in this analysis of recall response times, we assume that the low response rate yielded an insufficient number of trials for analysis of this task component. An alternative interpretation is that there was no effect of bmSTS stimulation on this specific vocabulary task. Analysis of TMS effects using linear mixed effects modeling To evaluate the robustness of the observed effects using an alternate analysis technique, we also tested our three hypotheses using a linear mixed effects modeling approach. Linear mixed effects models were generated in R version 1.2.1335 using the „lme4‟ package (Bates, Maechler, Bolker, & Walker, 2015). To select the random effects structure, we performed backwards model selection, beginning with a random intercept by subject, a random intercept by auditory stimulus, a random slope by subject for each of the four independent factors (stimulation type, learning condition, time point, and vocabulary type), and a random slope by stimulus for the stimulation type and time point factors. We removed random effects terms that accounted for the least variance one by one until the fitted mixed model was no longer singular, i.e. until variances of one or more linear combinations of random effects were no longer (close to) zero. The final model included three random effects terms: a random intercept by subject, a random intercept by stimulus, and a random slope by subject for the time point factor. CAUSAL ROLE OF SENSORY CORTICES 36 Contrasts were coded using simple coding, i.e. ANOVA-style coding, such that the model coefficient represented the size of the contrast from a given predictor level to the (grand) mean (represented by the intercept). The dependent measure was response times in the multiple choice translation task. Significance testing was performed using Satterthwaite‟s method implemented in the „lmerTest‟ package, with an alpha level of α = 0.05 (Kuznetsova, Brockhoff, & Christensen, 2017). Post-hoc Tukey tests were conducted using the „emmeans‟ package (Lenth, Singmann, Love, Buerkner, & Herve, 2019). The full model results are shown in Table S2. We first examined whether bmSTS stimulation modulated L2 translation. The model revealed a significant interaction of stimulation type and learning condition factors (β = -15.43, t = - 3.34, p < .001, 95% CI [-24.50 -6.37]), confirming our main hypothesis. Tukey‟s HSD post-hoc tests revealed that response times for words that had been learned with gesture enrichment -- but not picture enrichment -- were significantly delayed when TMS was applied to the bmSTS compared to sham stimulation (β = -42.99, p = .006). We next examined whether bmSTS integrity supported the auditory translation of gesture- enriched words at the later time point (5 months post-learning) even more than the earlier time point (5 days following the start of learning). The model revealed a significant three-way interaction of stimulation type, learning condition, and time point variables (β = -12.41, t = -2.69, p < .001, 95% CI [-21.48 -3.35]). Tukey‟s HSD post-hoc tests revealed a response benefit (faster responses) for gesture-enriched learning compared to picture-enriched learning under sham stimulation 5 months following learning (β = -85, p = .001). The application of TMS to bmSTS negated this benefit: Responses were significantly slower for the gesture condition at month 5 during TMS compared to sham stimulation (β = -68, p = .023). Finally, we examined whether the disruptive effects of bmSTS stimulation would occur independent of the conceptual perceptibility of the L2 word referents (i.e., whether a word was concrete or abstract). The four-way stimulation type × learning condition × time point × vocabulary type interaction was not reliable in the fitted model (β = 7.03, t = 1.52, p = .13, 95% CI [-2.02 16.08]), suggesting that effects of bmSTS stimulation did not significantly differ across vocabulary types. In sum, linear mixed modeling yielded the same results as the ANOVA-based approach reported in the main manuscript, with the exception of the four-way interaction, which was significant in the ANOVA but not in the mixed model analysis. Table S2. Linear mixed effects regression testing the effects of stimulation type, learning condition, time point, and vocabulary type on response times in the multiple choice translation task. CAUSAL ROLE OF SENSORY CORTICES 37 Fixed effects Random effects Estimate SE t p CI Variance SD Intercept 1032 32.42 31.82 <.001 965.66, Participant Intercept 21325 146.03 1098.10 Stimulation 6.07 4.62 1.31 .19 -3.00, 15.13 Learning 3.23 4.63 .70 .49 -5.85, condition 12.31 Time point 161.5 1.39 11.65 <.001 132.69, 189.71 Vocabulary -32.79 6.99 -4.69 <.001 -46.52, -18.83 Stimulation × -15.43 4.62 -3.34 <.001 -24.50, Learning -6.37 Stimulation × -.009 4.63 -.002 .99 -9.08, Time 9.07 Learning × 11.51 4.62 2.49 .013 2.44, Time 20.57 Stimulation × -.69 4.61 -.15 .88 -9.73, Vocabulary 8.36 Learning × 7.26 4.63 1.57 .12 -1.82, Vocabulary 16.35 Time × -29.12 4.70 -6.20 <.001 -38.34, Vocabulary -19.91 Stimulation × -12.41 4.62 -2.69 .007 -21.48, Learning × Time -3.35 Stimulation × -4.55 4.62 -.98 .33 -13.60, Learning × Vocabulary 4.51 Time 3674 60.61 Stimulus Intercept 5061 71.14 CAUSAL ROLE OF SENSORY CORTICES Stimulation × -3.38 4.62 -.73 .46 -12.42, Time × Vocabulary 5.67 Learning × 4.94 4.63 1.07 .29 -4.14, Time × Vocabulary 14.01 Stimulation × 7.03 4.62 1.52 .13 -2.02, 38 Learning × Time × Vocabulary 16.08 Analysis of paper-and-pencil test data Methods. On days 2, 3, and 4 of the L2 vocabulary learning, participants completed paper-and-pencil vocabulary tests prior to the training. Participants completed free recall, L1 translation, and L2 translation tests on each day. During the translation tests, participants received a list of either the 90 German words or the 90 Vimmi words and were asked to write the correct translation next to each word. During the free recall test, participants received a blank sheet of paper and were asked to write down as many German words, Vimmi words, or any combination of a Vimmi word with its German translation that occurred during the learning as they could remember. The free recall test was always administered before the translation tests, and the order of the two translation tests was counterbalanced across days and participants. Paper-and-pencil tests were independently scored for accuracy by two raters. L1 and L2 translation tests were scored in terms of the total number of correct translations recalled in each test (one point for each correct translation). A Vimmi word was considered correct if the two independent raters agreed that the word that was written down was valid for the sound pronounced in the audio file according to German sound-letter-mapping. A German word was considered correct if a participant wrote down the German word that was assigned to the Vimmi word during learning or if a participant wrote down a synonym of the German word, according to a standard German synonym database (http://www.duden.de). Free recall were scored in terms of the number of translations (German-Vimmi or Vimmi-German word pairs), German words that were missing corresponding Vimmi words, and Vimmi words that were missing corresponding German words. Three points were given for each correct translation (German-Vimmi or Vimmi- CAUSAL ROLE OF SENSORY CORTICES 39 German word pair). One point was given for each correctly-recalled German word that was missing a corresponding Vimmi translation and vice versa. Effects of enrichment on paper-and-pencil vocabulary test accuracy. Translation tests. To analyze the translation tests, percentages of correctly translated words were averaged across the two tests (as in Mayer et al., 2015) and submitted to a three- way ANOVA with the factors learning condition (gesture, picture), testing time point (day 2, day 3, day 4, month 5), and vocabulary type (concrete, abstract). The ANOVA did not yield any interactions of the learning condition factor with other variables, suggesting similar effects of gesture- and picture-enriched learning on vocabulary test performance. There was a significant main effect of testing time point (F 3, 63 = 94.28, p < .001, two-tailed, .82). Tukey‟s HSD post-hoc tests revealed that overall test scores at each time point differed significantly from test scores at all other time points (all ps < .001, Hedge‟s g range: .62-2.55; Fig. S2a). The ANOVA additionally yielded a significant main effect of vocabulary type (F 1, 21 = 135.17, p < .001, two- tailed, .87) and a significant time point × vocabulary type interaction (F 3, 63 = 5.78, p = .001, two-tailed, .22). Overall, test scores were significantly higher for concrete nouns compared to abstract nouns. There were no other significant main effects or interactions. CAUSAL ROLE OF SENSORY CORTICES 40 Figure S2. Paper-and-pencil translation test scores. a, Performance on paper-and-pencil translation tests significantly improved during days 2 to 4 of gesture- and picture-enriched training. Evidence of decay occurred 5 months following both gesture- and picture-enriched training (n = 22 participants). b, The same pattern of performance was observed for the free recall test (n = 22 participants). Error bars represent one standard error of the mean. *p < .05, **p < .01, ***p < .001. Free recall test. We next examined performance on the free recall paper-and-pencil test. Points for correctly recalled German words, Vimmi words, and German-Vimmi translations were summed for each participant, learning condition, testing time point, and vocabulary type (cf. Mayer et al., 2015). Free recall test scores by condition are shown in Fig. S2b. A three-way ANOVA on free recall scores with factors learning condition (gesture, picture), testing time point n.s050100Day 2Day 3Day 4Month 5Time pointGesture-ConcretePicture-ConcreteGesture-AbstractPicture-AbstractTranslation accuracy (% correct)aTranslation paper-and-penciltestsb02040Day 2Day 3Day 4Month 5Time pointGesture-ConcretePicture-ConcreteGesture-AbstractPicture-AbstractFree recall paper-and-pencil testFree recall total score***************n.s.n.s.n.s.n.s.n.s.n.s.n.s.n.s. CAUSAL ROLE OF SENSORY CORTICES 41 (day 2, day 3, day 4, month 5), and vocabulary type (concrete, abstract) did not yield any significant interactions of the learning condition factor with other variables besides a significant learning condition × vocabulary type interaction (F 1, 21 = 7.11, p = .014, two-tailed, .25). Tukey‟s HSD post-hoc tests revealed higher scores for concrete words compared to abstract words following gesture-enriched learning but not picture-enriched learning (p < .001, Hedge‟s g = .40). There was a signficant main effect of vocabulary type (F 1, 21 = 11.14, p = .003, two- tailed, .35); scores were significantly higher for concrete words than abstract words. There was also a significant main effect of testing time point (F 3, 63 = 66.48, p < .001, two-tailed, .76). Tukey‟s HSD post-hoc tests revealed that overall test scores were significantly higher at day 3 compared to day 2 (p = .0062, Hedge‟s g = 1.33), day 4 compared to day 3 (p = .014, Hedge‟s g = .85), and at month 5 compared to day 4 (p < .001, Hedge‟s g = 1.39). There was also a significant time point × vocabulary type interaction (F 3, 63 = 18.90, p < .001, two-tailed, .47). There were no other significant main effects or interactions. Gesture-enriched training reduces long-term decrease in L2 translation accuracy on paper-and-pencil vocabulary tests. Finally, we tested whether gesture-enriched learning diminished long-term decreases in L2 translation accuracy over time compared to picture-enriched learning on the paper-and-pencil vocabulary tests. Translation tests. In order to evaluate long-term changes in translation test accuracy, we computed the difference in mean performance on the translation tests (percent correct) at day 4 and month 5 for each participant, learning condition, and word type. A two-way ANOVA on difference scores (percent correct) with the factors learning condition (gesture, picture) and vocabulary type (concrete, abstract) yielded a significant main effect of learning condition (F 1, 21 = 5.84, p = .025, two-tailed, .22). Performance decreased significantly less for gesture- enriched vocabulary compared to picture-enriched vocabulary 5 months following training (Fig. S3). There was also a significant main effect of vocabulary type (F 1, 21 = 8.75, p = .007, two- tailed, .29). Performance decreased significantly less for concrete vocabulary compared to abstract vocabulary 5 months following training. The interaction between learning condition and vocabulary type variables was not significant. CAUSAL ROLE OF SENSORY CORTICES 42 Figure S3. Long-term decrease in translation accuracy on paper-and-pencil vocabulary tests. Gesture-enriched L2 vocabulary learning resulted in less of a decrease in performance on paper-and-pencil translation tests 5 months following learning compared to picture-enriched learning. On the free recall paper-and-pencil test, participants demonstrated less long-term decay of concrete vocabulary compared to abstract vocabulary over a 5-month period (n = 22 participants). Error bars represent one standard error of the mean. *p < .05, **p < .01, ***p < .001. Free recall test. In order to evaluate long-term changes in recall accuracy, we computed the difference in free recall paper-and-pencil test scores at day 4 and month 5 for each participant, learning condition, and word type. A two-way ANOVA on difference scores with the factors learning condition (gesture, picture) and vocabulary type (concrete, abstract) yielded only a significant main effect of vocabulary type (F 1, 21 = 18.43, p < .001, two-tailed, .47). Recall accuracy decreased significantly less for concrete vocabulary compared to abstract vocabulary 5 months following training (Fig. S3). The main effect of learning condition and interaction between learning condition and vocabulary type variables were not significant. Summary. Taken together, the paper-and-pencil test scores revealed significant improvement for both gesture- and picture-enriched words from day 2 to day 3 and day 3 to day 4 of the L2 training period, as well as a significant decrease in performance 5 months post- learning compared to day 4. This pattern of performance was consistent across test types (translation and free recall tests). Analyses of L2 memory decay over a 5-month interval (day 4 scores -- month 5 scores) revealed greater decay of memories for picture-enriched words compared to gesture-enriched words over a 5-month period on the translation tests. However, no difference between gesture- and picture-enriched words in terms of amount of decay was 103050GesturePictureLearning typeFree recall paper-and-pencil test01530GesturePictureLearning typeConcrete nounsAbstract nounsScore difference(Day 4 − Month5)****Accuracy difference(Day 4 − Month5, % correct)Translation paper-and-penciltestsn.s.n.s.n.s.n.s. CAUSAL ROLE OF SENSORY CORTICES 43 observed on the free recall tests. Instead, the free recall tests were sensitive to word type: A greater amount of decay was observed for abstract words compared to concrete words based on free recall scores. Analysis of variance (ANOVA) summary tables In this section, we summarize using tables the main effects and interactions tested in all ANOVA analyses reported in the main manuscript. *p < .05, **p < .01, ***p < .001. Table S3. Two-way ANOVA testing effects of stimulation type and learning condition on response times in the multiple choice translation task. Intercept Stimulation Learning df F p 21 927.96 <.001 21 2.49 .13 .11 21 .003 .96 <.001 Stimulation × Learning 21 11.82 .002** .36 Table S4. Three-way ANOVA testing effects of stimulation type, learning condition, and time point on response times in the multiple choice translation task. Intercept Stimulation Learning Time df F p 21 1013.03 <.001 21 3.30 .084 .14 21 1.18 .29 .05 21 18.65 <.001*** .84 Stimulation × Learning 21 106.95 <.001*** .47 Stimulation × Time 21 1.23 .28 .06 Learning × Time 21 6.42 .019* .23 Stimulation × Learning × Time 21 7.51 .012* .26 CAUSAL ROLE OF SENSORY CORTICES 44 Table S5. Four-way ANOVA testing effects of stimulation type, learning condition, time point, and vocabulary type on response times in the multiple choice translation task. Intercept Stimulation Learning Time df F p 21 1055.04 <.001 21 2.07 21 .54 .17 .54 .09 .02 21 131.51 <.001*** .86 Vocabulary 21 25.48 <.001*** .55 Stimulation × Learning 21 9.03 .007** .30 Stimulation × Time 21 .014 .91 <.001 Learning × Time 21 8.29 .009** .28 Stimulation × Vocabulary 21 .25 Learning × Vocabulary 21 .03 .62 .86 .01 .001 Time × Vocabulary 21 3.30 .083 .14 Stimulation × Learning × Time 21 10.97 .003** .34 Stimulation × Learning × Vocabulary Stimulation × Time × Vocabulary 21 2.95 .10 .12 21 .44 .52 .02 Learning × Time × Vocabulary 21 .11 .74 .005 Stimulation × Learning × Time × Vocabulary 21 5.24 .033* .20 Table S6. Four-way ANOVA testing effects of stimulation type, learning condition, time point, and vocabulary type on accuracy in the multiple choice translation task. Intercept df F p 21 361.33 <.001 CAUSAL ROLE OF SENSORY CORTICES 45 Stimulation Learning Time 21 1.13 21 .001 .30 .97 .05 <.001 21 124.77 <.001*** .86 Vocabulary 21 35.62 <.001*** .63 Stimulation × Learning 21 .035 Stimulation × Time 21 .19 .85 .67 .002 .009 Learning × Time 21 6.86 .016* .25 Stimulation × Vocabulary 21 3.86 Learning × Vocabulary 21 1.25 Time × Vocabulary 21 3.93 Stimulation × Learning × Time 21 .016 .06 .28 .90 .90 .15 .06 .16 <.001 Stimulation × Learning × Vocabulary Stimulation × Time × Vocabulary 21 3.54 .074 .14 21 .013 .91 <.001 Learning × Time × Vocabulary 21 2.60 .12 .11 Stimulation × Learning × Time × Vocabulary 21 8.23 .009** .28 Table S7. Two-way ANOVA testing effects of learning condition and vocabulary type on changes in multiple choice task accuracy across time points (month 5 -- day 5). Intercept Learning df F p 21 110.74 <.001 21 2.73 .11 .12 Vocabulary 21 3.41 .079 .14 Learning × Vocabulary 21 13.84 .001*** .40 CAUSAL ROLE OF SENSORY CORTICES 46 Table S8. Two-way ANOVA testing effects of learning condition and vocabulary type on changes in multiple choice task response time across time points (month 5 -- day 5). Intercept Learning df F p 21 88.08 <.001 21 11.05 .003** .34 Vocabulary 21 2.36 .14 .10 Learning × Vocabulary 21 .017 .90 <.001 Supplementary References Bates, D., Maechler, M., & Bolker, B. S. Walker (2015). Fitting Linear Mixed-Effects Models Using lme4. Journal of Statistical Software, 67(1), 1-48. Hoffman, P. (2016). The meaning of „life‟ and other abstract words: Insights from neuropsychology. Journal of Neuropsychology, 10(2), 317-343. Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2017). lmerTest package: tests in linear mixed effects models. Journal of Statistical Software, 82(13). Lenth, R., Singmann, H., Love, J., Buerkner, P., & Herve, M. (2019). Package "emmeans": Estimated Marginal Means, aka Least-Squares Means. The Comprehensive R Archive Network, 1-67. Macedonia, M. (2014). Bringing back the body into the mind: gestures enhance word learning in foreign language. Frontiers in Psychology, 5, 1467. Macedonia, M., & Knösche, T. R. (2011). Body in mind: How gestures empower foreign language learning. Mind, Brain, and Education, 5(4), 196-211. Mayer, K. M., Yildiz, I. B., Macedonia, M., & von Kriegstein, K. (2015). Visual and motor cortices differentially support the translation of foreign language words. Current Biology, 25(4), 530-535.
1709.06950
1
1709
2017-09-20T16:18:17
Spatial features of synaptic adaptation affecting learning performance
[ "q-bio.NC", "cond-mat.dis-nn", "cs.LG", "cs.NE" ]
Recent studies have proposed that the diffusion of messenger molecules, such as monoamines, can mediate the plastic adaptation of synapses in supervised learning of neural networks. Based on these findings we developed a model for neural learning, where the signal for plastic adaptation is assumed to propagate through the extracellular space. We investigate the conditions allowing learning of Boolean rules in a neural network. Even fully excitatory networks show very good learning performances. Moreover, the investigation of the plastic adaptation features optimizing the performance suggests that learning is very sensitive to the extent of the plastic adaptation and the spatial range of synaptic connections.
q-bio.NC
q-bio
Spatial features of synaptic adaptation affecting learning performance Damian L. Berger1,*, Lucilla de Arcangelis2,3, and Hans J. Herrmann1 1ETH Zurich, Computational Physics for Engineering Materials, IfB, Wolfgang-Pauli-Strasse 27, 2Department of Industrial and Information Engineering, University of Campania "Luigi 8093 Zurich, Switzerland Vanvitelli", 81031 Aversa (CE), Italy 3INFN sez. Naples, Gr. Coll. Salerno (Italy) *[email protected] July 9, 2021 Abstract Recent studies have proposed that the diffusion of messenger molecules, such as monoamines, can mediate the plastic adaptation of synapses in supervised learning of neural networks. Based on these findings we developed a model for neural learning, where the signal for plastic adaptation is assumed to propagate through the extracellular space. We investigate the conditions allowing learning of Boolean rules in a neural network. Even fully excitatory networks show very good learning performances. Moreover, the investigation of the plastic adaptation features optimizing the performance suggests that learning is very sensitive to the extent of the plastic adaptation and the spatial range of synaptic connections. Introduction The brain is a large neural network, where sensory information is processed, possibly triggering a specific response. If this response turns out to generate a result, that is different from the expected one, error signals cause changes in the synaptic weights attempting to minimize further mismatches [1, 2, 3]. Models of neural network attempting to contribute to the understanding of learning in the brain usually look at a task where for a specific set of inputs the network has to adapt to generate a desired result. Error back propagation [4], one of the most popular algorithms proposed, can however not serve as a candidate for learning in a biological sense [5]. A teaching signal which transmits information antidromically with memory of all neurons it has passed before, seems unlikely to exist in the brain. Different studies have proposed a variety of models for a neural network able to learn without passing error gradients antidromically. It has been suggested that errors could propagate backwards through a second network [6, 7], where forward and backward connections have to be chosen symmetric. Another interesting approach is to calculate error gradients by using an alternation of two activity phases, where one is determined by a teacher [8]. A similar performance, as the back propagation algorithm, can also be achieved by using exact error gradients only for the last synapses and the synapses in the hidden layers are modified only in the same direction as with back propagation [9]. There exist also interesting models, called liquid state machines or echo state machines, where the weights of the main network remains unchanged and the inputs are connected to a large recurrent neural network, thereby mapping the input to a higher dimensional space [10, 11]. These models are mainly used for spatiotemporal problems like speech recognition, but their capability for recognizing handwritten digits has also been demonstrated [12]. The variety of model reduces if we insist on the assumption that synaptic changes have to be triggered by a specific error signal. Since there is no evidence that a neural network can transfer feedback informa- tion antidromically, such a signal should propagate through the extracellular space. A feedback signal to change the strength of synapses in a larger region could for instance be delivered by monoamine releasing 1 neurons. It is known that these neurons release their transmitters deep into the extracellular space [13]. In particular for dopamine, it has been verified, that its release can mediate plasticity [14, 15]. In several recently proposed models these ideas are implemented by the assumption that spike-timing-dependent plasticity (STDP) and dopamine induced plasticity are directly coupled [16, 17, 18]. These models as- sume that dopamine serves as a rewarding signal. In an other model it has been proposed [19] that the transmitted feedback signal changes the vesicle release probability of previously activated synapses. Opposite to rewarding reinforcement it has been argued that negative feedback signals, which change synapses only if mistakes occurred, are more biologically plausible and preserve the adaptability of the system [20, 21, 22]. Based on such ideas, de Arcangelis et al.[23, 24] proposed a model of spiking integrate- and-fire neurons which is able to learn logical binary functions in a neural network. For wrong responses by the neural network, activated synapses are modified proportionally to the inverse number of synaptic connections on the signal path to the output neuron. This method implies that the feedback signal propagates backwards on the synaptic connections and decreases in strength depending on the number of neurons it has passed before. In this article we extend these ideas by a model, where the strength of the feedback signal does not depend on a network distance, but on the euclidean distance. By this model we study the importance of the localization of a teaching signal, i.e. if a localized learning signal can represent advantages over one acting widely in space. This is the case, for instance, of dopamine whose effect covers a finite range of tens to thousands of synapses [13]. Network model Network topology The network consists of 4 input neurons, 1 output neuron and N neurons in the hidden network, which are placed randomly in a square, with side length L. The area is chosen to scale proportional to N , such that the density of neurons remains constant (Fig. 1). In the further discussions we will use N to characterize the size of the network. Each neuron in the hidden network has ten outward connections. Each connection is established by choosing a distance d from an exponential distribution [25] p(d) = 1 e−d/d0 and searching d0 for a neuron at a distance sufficiently close to d. The input neurons also have ten outward connections with their nearest neighbours whereas the output neuron has no outward but only inward connections from its ten nearest nodes. When inhibitory neurons are considered, a fixed fraction pinh of the neurons will be inhibitory. Neuronal model For the firing dynamics we choose an integrate-and-fire model with discrete time steps. During each time step all neurons with potential exceeding a certain threshold vi ≥ vmax = 1.0 fire. After firing the potential of the neuron i is set back to zero and the voltage of all connected neurons j becomes vj(t + 1) = vj(t) ± ωijηi (0.1) where ωij is the synaptic strength and ηi ∈ [0, 1] stands for the amount of releasable neurotransmitter, which for simplicity is the same for all synapses of neuron i. The plus or minus sign is for excitatory or inhibitory synapses, respectively. After each firing ηi decreases by a fixed amount ∆η = 0.2, which guarantees that the network activity will decay in a finite time. The initial conditions are v = 0 and η = 1 for all neurons. All neurons which have reached a potential higher than the threshold will fire at the next time step. After firing a neuron will be in a refractory state for tref r time step, during which it cannot receive or send any signal. Learning Learning binary input-output relations such as the XOR-rule has been a classical benchmark problem for neural networks. In our case the XOR-rule turned out to be too simple as only very small networks with unfavorable parameters were not always able to learn it. We choose to study a more complex problem 2 Figure 1: Network architecture: Network with one hundred neurons in the network (gray neurons). The four input neurons (green) on the left are labeled 1 to 4 (top to bottom) and are connected to the network. On the right is the output neuron (red) which senses incoming signals from its ten nearest nodes. The strength of the synapses is indicated by their thickness. with up to 15 different Boolean functions of 4 binary inputs (see Table 1). We will study how our learning procedure performs for different parameters and network sizes. Input 1 Input 2 Input 3 Input 4 Output 1 0 0 0 1 0 1 0 0 1 1 1 0 0 0 0 0 1 0 1 0 0 0 1 1 0 0 1 1 0 1 1 1 1 0 1 0 1 0 1 1 1 1 0 0 1 0 0 1 1 0 1 1 0 0 0 1 0 1 1 1 1 0 1 0 1 0 1 1 1 0 1 1 1 0 Table 1: Input-output relations to be learned by the network. Each input bit is applied to one input neuron by making it fire or not fire if the input is one or zero, respectively. The activity then propagates through the network and the binary output value 1 is identified with the output neuron firing at least once and zero if it does not fire during the whole network activity. Whether the activity reaches the output neuron depends solely on the strength of the synapses. If the average strength of the synapses is too low, the propagation of the activity stops soon and only a fraction of the entire network will fire. On the other hand for very strong synapses almost all neurons will fire several times and the output neuron will fire for each input pattern. Therefore an optimal choice for the average synaptic strength is a "critical value" where on average for each input pattern there is an equal chance that the output neuron fires or does not fire. We approximate this "critical point" by initially choosing weak synapses (ωij = 1.0 for the synapses of the input neurons and ωij = 0.1 for all other synapses) and when applying the input patterns, as long as no input leads to firing of the output neuron, all synapses are strengthened by a small amount (0.001 times the actual strength of the synapse). As soon as the first input leads to the output neuron firing, the network is considered to be close to the "critical point" and the learning procedure starts. Learning proceeds as follows: All inputs of the rule are fed into the network one after the other, where the value of the neurotransmitters and the voltages are restored back to η = 1.0 and v = 0 before each input. Whenever the result is wrong the synaptic strengths are adapted. If the correct result is one but the output neuron did not fire, synapses are strengthened, otherwise if the correct result is zero but 3 the output neuron fired, synapses are weakened [20]. Only those synapses activated during the activity propagation are modified. After a neuron fires all synapses which do not lead to neurons which were in the refractory state are considered as activated. The adaptation signal is assumed to be released at the output and decrease exponentially in space ∆ωij = ±αωijnacte−r/r0 (0.2) where α = 0.001 is the adaptation strength and r is the euclidean distance between the output neuron and the postsynaptic neuron of the synapse, neuron j. The postsynaptic neuron is chosen based on the assumption that a synapse is located much closer to the postsynaptic than to the presynaptic neuron. Different values for α have been tested and are found to affect the learning speed but not the qualitatively results. The plus or minus sign is for excitatory or inhibitory synapses, respectively. Additionally the adaptation is proportional to the strength of the synapse itself and to the number of times the synapse was activated (nact). The ωij and the nact dependence are considered since they lead to an increase in learning performance. A maximum synaptic strength (ωmax = 2) is imposed. If the activity is so weak that the voltage of the output neuron remains unchanged, the system is not able to provide an answer and therefore we increase all the weights by a small amount (∆ωij = αωij). XOR gate with only excitatory neurons A biological neural network consists of excitatory (ωij ≥ 0) and inhibitory (ωij ≤ 0) synapses. It has actually been shown that a certain percentage of inhibitory synapses can increase the learning performance [24]. The importance of inhibiting signals can be seen by looking at the XOR-rule (Fig. 2): When one of the two input neurons fires the output neuron should fire as well, but when both input neurons fire the output neuron should not fire. Using only synapses with positive weights we might intuitively think that by the principle of superposition a stronger input should always lead to a stronger output. This, however, does not hold in general. An interesting and important ingredient to generate inhibition is the refractory time. Figure 2 shows an example of a small network that performs the XOR rule with excitatory neurons and a refractory time of one time step. Black and gray arrows are for weights equal to and barely less than one, respectively. When both input neurons (neuron 1 and 2 in the figure) are stimulated neuron 4 fires one time step earlier than in the other two cases, where neuron 4 fires after a second stimulation by neuron 3. This earlier firing of neuron 4 leads to neuron 7 stimulating neuron 6 when it is in the refractory period. Therefore, neuron 6 only fires once which is not sufficient to make the output neuron (neuron 8) fire. The self-organization of the learning mechanism can take advantage of this mechanism to suppress certain signals. Therefore, even if purely excitatory network is not biologically reasonable, the refractory time can induce inhibition which can be of computational relevance. Therefore, we chose to perform most simulations with purely excitatory neurons. Results Our learning procedure runs until either sequentially all input-output relations are learned or a maximal number of learning steps Tmax has been performed, where a learning step occurs whenever the network generates a wrong result for the actual input. For most of the simulations the networks will be trained to learn the first ten patterns from Table 1. The learning performance is evaluated by generating a statistical ensemble of 2000 random networks with the same parameters and the success rate s is defined by the ratio of networks which are able to learn all patterns within Tmax. We first investigate the role of the learning length r0 on the success rate s (see Fig. 3a). For very localized learning signals (r0 < 10−1) the success rates are close to zero. This is not surprising as only the synapses leading to the output neuron are modified by a noticeable amount. For larger r0 the learning adaptation penetrates deeper into the network and the success rate rapidly increases up to s = 1.0 for r0 = 10 and Tmax = 100, 000. Interestingly, when r0 exceeds the system size (r0/L = 1), which implies that all synapses in the network are adapted, the learning performance strongly decreases. Overall, we find that the learning performance can strongly be optimized by the spacial extent of the teaching signal. 4 ! 2 t=1 1 1 2 1 1 2 1 1 4 2 3 2 4 3 3 2 4 3 3 2 5 3 5 4 5 4 7 4 6 3 7 5 6 3,6 7 5 6 3,6 8 8 7 8 7 neuron fired twice neuron fired once neuron did not fire ωij = 1 ωij = 1 - ε Figure 2: XOR gate with only excitatory neurons. The figure above shows a possibility for a XOR gate with only excitatory neurons with a refractory period of one time-step. The blue numbers below each neuron indicate at which time step the neuron fires. Figure 3b shows the learning performance for a different number of patterns to be learned by the network. Unsurprisingly the performance is higher when only the first three patterns from Table 1 have to be learned compared to the case when all the fifteen patterns have to be learned. It is interesting to notice that the value for r0 at which the learning performance increases shifts towards larger values the more patterns the network has to learn. This suggests that the more patterns a network has to learn the deeper the learning signal has to penetrate into the network. Figure 3: Ranges of the learning length optimizing the learning performance. The success rate s as a function of r0/L. (a) Different maximal learning times Tmax show a consistent improvement in performance (N = 1000, d0 = 2,tref r = 1, first ten patterns of Table 1). (b) The performance dependence on the number of patterns the system has to learn (Tmax = 10000, N = 1000,tref r = 1, d0 = 2). For Tmax = 3000 (in Fig. a) and for 12 patterns (in Fig. b) we also performed simulations with inhibitory neurons pinh = 0.2 resulting in an overall performance improvement. The error bar represent a 95% confidence interval. Next we investigate, the role of the network size on the performance. Indeed, we find a monotonic increase in performance (see Fig. 4a). Interestingly the performance exhibits its maximum at a value of r0 increasing proportionally to L, whereas the s-decay shifts towards larger r0 for larger system size. This result confirms that the learning capabilities decrease when r0 exceeds the system size L and the adaptation changes all activated synapses in an almost uniform way. In Fig. 4b we vary the characteristic length for synaptic connections d0 and study its effect on the learning performance. When r0 exceeds the system length, d0 strongly affects the learning performance. For longer synapses (higher d0) the 5 Figure 4: System size dependence and influence of the synaptic length. (a) The success rate s as a function of r0/L for different network sizes N (Tmax = 10000, d0 = 2,tref r = 1, first ten patterns of Table 1). Larger networks show a better performance and the peak of the performance roughly moves proportionally to L. (b) The effect of the network structure is analyzed by plotting s for different synaptic characteristic length d0 (N = 1000, Tmax = 10000,tref r = 1, first ten patterns of Table 1). For N = 100 (in Fig. a) and for d0 = 2 (in Fig. b) we also performed simulations with inhibitory neurons pinh = 0.2 resulting in an overall performance improvement. The error bars represent a 95% confidence interval. performance drops much faster as r0 increases. This result is in agreement with the abrupt decrease of the performance for plastic adaptation acting over a range larger than the system size L. For very large d0, the synaptic network also undergoes quite uniform plastic adaptation over large distances. The maximum performance appears to weakly depend on d0, except for very localized connections. We also ask the question whether the space dependence of the adaptation influences the presented results. Therefore we perform simulations using a plastic adaptation that decrease in space as a Gaus- sian. Results are very similar (data not shown) to those obtained for a synaptic adaptation decaying exponentially, except that for the Gaussian adaptation the peaks are slightly narrower. Figure 5: Role of the refractory time on the performance. (a) The success rate s for different refractory times, ranging from tref r = 0 to tref r = 10 (N = 1000, Tmax = 10000, d0 = 2, first ten patterns out of Table 1). (b) Performance obtained for the neuronal activation function given in Eq. 0.3. For tref r = 0 (in Fig. a and b) we also performed simulations with inhibitory neurons pinh = 0.2 resulting in an overall performance improvement. The error bars represent a 95% confidence interval. Further, we tested the influence of the refractory time on the learning performance (see Fig. 5a). 6 When the refractory time is zero learning is possible, but the performance is worse compared to tref r = 1 and tref r = 2. For tref r = 10 the performance decreases almost to zero. This might be related to the fact that for larger refractory times it becomes less likely that a single neuron fires more than once, therefore preventing the occurrence of activity loops. This result shows that the refractory time is an important ingredient to generate inhibition, but also the activation function (Heaviside step function) causes dissipation which can create inhibition. We could reduce this type of dissipation by using an activation function that is proportional to the voltage of the firing neuron. Equation 0.1 changes then to: vj(t + 1) = vj(t) ± ωijηivi (0.3) To eliminate dissipation completely we would need to set the firing threshold to zero. However, this would make learning impossible as every input will make all successive neurons fire and nothing can stop the output neuron from firing. In Fig. 5b we show the effect of the linear activation function (eq. 0.3) on the learning performance. The result confirms our assumption. For a refractory time equal to zero the learning performance drops to zero. For refractory times larger than zero the overall performance is higher compared to the Heaviside activation function. The performance also does not decrease for larger refractory times. Adding inhibitory neurons makes it possible for neurons with tref r = 0 to learn, but with a low performance. In each case of Fig. 3 - 5 we perform the same simulation, but with a fraction pinh = 0.2 of inhibitory neurons. Inhibitory neurons noticeably increase the maximum learning performance and the performance decay for larger r0 is slightly less pronounced. In Fig. 5b the inhibitory neurons make it possible to learn if the refractory time is zero, which is not possible without inhibitory neurons. Conclusions We show that a neural network trained by a distance dependent plastic adaptation can learn Boolean rules with a very good performance. The spatial extent of the learning signal was found to have a huge impact on the learning performance. A very localized plastic adaptation, which only modifies the synapses directly connected to the output neuron, results, unsurprisingly, in poor performance. For deeper adaptation signal, the performance strongly increases, until the learning length r0 exceeds the system length L, where the performance decreases. When r0 ≫ L the adaptation signal becomes quite uniform in space and the lack of variability might limit the learning capabilities. Similar behaviour is observed for increasing characteristic length of synaptic connections. Indeed, current evidence from functional magnetic resonance imaging (fMRI) experiments [26, 27] and EEG data [28, 29] shows that a greater brain signal variability indicates a more sophisticated neural system, able to explore multiple functional states. Signal variability also reflects a greater coherence between regions and a natural balance between excitation and inhibition originating the inherently variable response in neural functions. Furthermore, the observation that older adults exhibit less variability reflecting less network complexity and integration, suggests that variability can be a proxy for optimal systems. Interestingly, we show that a network with only excitatory neurons can learn non-linearly separable problems with a success rate of up to 100%. In this case, the neural refractory time generates inhibition in the neuronal model. If we include additional inhibitory neurons the performance still increase noticeably. Regarding the network structure we find that the synaptic length also has an impact on the perfor- mance: For networks dominated by local connections the performance is worse compared to networks with long range connections. We conclude that the structure of the network and the locality of the plastic adaptation have an important role in the learning performance. To assess the generality of our results it will be necessary to study more complex forms of learning like, for example, pattern recognition. Further investigations should also try to improve the learning performance to become comparable to machine learning algorithms. 7 References [1] Schultz, W. & Dickinson, A. Neuronal coding of prediction errors. Annual review of neuroscience 23, 473 -- 500 (2000). [2] Keller, G. B. & Hahnloser, R. H. Neural processing of auditory feedback during vocal practice in a songbird. Nature 457, 187 -- 190 (2009). [3] Keller, G. B., Bonhoeffer, T. & Hubener, M. Sensorimotor mismatch signals in primary visual cortex of the behaving mouse. Neuron 74, 809 -- 815 (2012). [4] Williams, D. & Hinton, G. Learning representations by back-propagating errors. Nature 323, 533 -- 536 (1986). [5] Crick, F. The recent excitement about neural networks. Nature 337, 129 -- 132 (1989). [6] Kolen, J. F. & Pollack, J. B. Backpropagation without weight transport. In Neural Networks, 1994. IEEE World Congress on Computational Intelligence., 1994 IEEE International Conference on, vol. 3, 1375 -- 1380 (IEEE, 1994). [7] Stork, D. G. Is backpropagation biologically plausible. In International Joint Conference on Neural Networks, vol. 2, 241 -- 246 (IEEE Washington, DC, 1989). [8] O'Reilly, R. C. Biologically plausible error-driven learning using local activation differences: The generalized recirculation algorithm. Neural computation 8, 895 -- 938 (1996). [9] Lillicrap, T. P., Cownden, D., Tweed, D. B. & Akerman, C. J. Random synaptic feedback weights support error backpropagation for deep learning. Nature Communications 7, 13276 (2016). [10] Maass, W., Natschlager, T. & Markram, H. Real-time computing without stable states: A new framework for neural computation based on perturbations. Neural computation 14, 2531 -- 2560 (2002). [11] Jaeger, H. The A´uecho state A`u approach to analysing and training recurrent neural networks-with ' ' an erratum note. Bonn, Germany: German National Research Center for Information Technology GMD Technical Report 148, 13 (2001). [12] Jalalvand, A., Van Wallendael, G. & Van De Walle, R. Real-time reservoir computing network- In 2015 7th International Conference on based systems for detection tasks on visual contents. Computational Intelligence, Communication Systems and Networks. [13] Fuxe, K. et al. The discovery of central monoamine neurons gave volume transmission to the wired brain. Progress in neurobiology 90, 82 -- 100 (2010). [14] Liu, F. et al. Direct protein -- protein coupling enables cross-talk between dopamine d5 and γ- aminobutyric acid a receptors. Nature 403, 274 -- 280 (2000). [15] Reynolds, J. N. & Wickens, J. R. Dopamine-dependent plasticity of corticostriatal synapses. Neural Networks 15, 507 -- 521 (2002). [16] Izhikevich, E. M. Solving the distal reward problem through linkage of stdp and dopamine signaling. Cerebral cortex 17, 2443 -- 2452 (2007). [17] Florian, R. V. Reinforcement learning through modulation of spike-timing-dependent synaptic plas- ticity. Neural Computation 19, 1468 -- 1502 (2007). [18] Aswolinskiy, W. & Pipa, G. Rm-sorn: a reward-modulated self-organizing recurrent neural network. Frontiers in computational neuroscience 9, 36 (2015). 8 [19] Seung, H. S. Learning in spiking neural networks by reinforcement of stochastic synaptic transmis- sion. Neuron 40, 1063 -- 1073 (2003). [20] Bak, P. & Chialvo, D. R. Adaptive learning by extremal dynamics and negative feedback. Physical Review E 63, 031912 (2001). [21] Chialvo, D. R. & Bak, P. Learning from mistakes. Neuroscience 90, 1137 -- 1148 (1999). [22] Bosman, R., Van Leeuwen, W. & Wemmenhove, B. Combining hebbian and reinforcement learning in a minibrain model. Neural Networks 17, 29 -- 36 (2004). [23] de Arcangelis, L. & Herrmann, H. J. Learning as a phenomenon occurring in a critical state. Proceedings of the National Academy of Sciences 107, 3977 -- 3981 (2010). [24] Capano, V., Herrmann, H. J. & de Arcangelis, L. Optimal percentage of inhibitory synapses in multi-task learning. Scientific Reports 5, 9895 (2015). [25] Roerig, B. & Chen, B. Relationships of local inhibitory and excitatory circuits to orientation pref- erence maps in ferret visual cortex. Cerebral cortex 12, 187 -- 198 (2002). [26] Garrett, D. D., Kovacevic, N., McIntosh, A. R. & Grady, C. L. Blood oxygen level-dependent signal variability is more than just noise. Journal of Neuroscience 30, 4914 -- 4921 (2010). [27] Garrett, D. D., Kovacevic, N., McIntosh, A. R. & Grady, C. L. The importance of being variable. Journal of Neuroscience 31, 4496 -- 4503 (2011). [28] Ghosh, A., Rho, Y., McIntosh, A. R., Kotter, R. & Jirsa, V. K. Noise during rest enables the exploration of the brain's dynamic repertoire. PLoS computational biology 4, e1000196 (2008). [29] McIntosh, A. R., Kovacevic, N. & Itier, R. J. Increased brain signal variability accompanies lower behavioral variability in development. PLoS computational biology 4, e1000106 (2008). Acknowledgements The authors thank L.M. van Kessenich for many interesting discussions. We acknowledge financial support from the European Research Council (ERC) Advanced Grant FP7-319968 FlowCCS. Author contributions statement D.L.B conceived the research and conducted the numerical simulations. L.d.A and H.J.H. gave important advise. All authors contributed to the writing of the manuscript. Additional information Competing financial interests The authors declare no competing financial interests. 9
1611.04758
1
1611
2016-11-15T09:49:01
Structure and Dynamics of Brain Lobes Functional Networks at the Onset of Anesthesia Induced Loss of Consciousness
[ "q-bio.NC" ]
Anesthetic agents are neurotropic drugs able to induce dramatic alterations in the thalamo-cortical system, promoting a drastic reduction in awareness and level of consciousness. There is experimental evidence that general anesthesia impacts large scale functional networks leading to alterations in the brain state. However, the way anesthetics affect the structure assumed by functional connectivity in different brain regions have not been reported yet. Within this context, the present study has sought to characterize the functional brain networks respective to the frontal, parietal, temporal and occipital lobes. In this experiment, electro-physiological neural activity was recorded through the use of a dense ECoG-electrode array positioned directly over the cortical surface of an old world monkey of the species Macaca fuscata. Networks were serially estimated over time at each five seconds, while the animal model was under controlled experimental conditions of an anesthetic induction process. In each one of the four cortical brain lobes, prominent alterations on distinct properties of the networks evidenced a transition in the networks architecture, which occurred within about one and a half minutes after the administration of the anesthetics. The characterization of functional brain networks performed in this study represents important experimental evidence and brings new knowledge towards the understanding of neural correlates of consciousness in terms of the structure and properties of the functional brain networks.
q-bio.NC
q-bio
Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Structure and Dynamics of Brain Lobes Functional Networks at the Onset of Anesthesia Induced Loss of Consciousness Eduardo C. Padovani ∗ Universidade de São Paulo Abstract Anesthetic agents are neurotropic drugs able to induce dramatic alterations in the thalamo-cortical system, promoting a drastic reduction in awareness and level of consciousness. There is experimental evidence that general anesthesia impacts large scale functional networks leading to alterations in the brain state. However, the way anesthetics affect the structure assumed by functional connectivity in different brain regions have not been reported yet. Within this context, the present study has sought to characterize the functional brain networks respective to the frontal, parietal, temporal and occipital lobes. In this experiment, electro-physiological neural activity was recorded through the use of a dense ECoG-electrode array positioned directly over the cortical surface of an old world monkey of the species Macaca fuscata. Networks were serially estimated over time at each five seconds, while the animal model was under controlled experimental conditions of an anesthetic induction process. In each one of the four cortical brain lobes, prominent alterations on distinct properties of the networks evidenced a transition in the networks architecture, which occurred within about one and a half minutes after the administration of the anesthetics. The characterization of functional brain networks performed in this study represents important experimental evidence and brings new knowledge towards the understanding of neural correlates of consciousness in terms of the structure and properties of the functional brain networks. I. Introduction Our understanding on how the brain works has ripened as neuroscience has evolved and consolidated. In general, the knowledge we have about the brain was slowly constructed from the integration of several discoveries and reports of many re- searchers, being accomplished mainly over the last 150 years. Most of the advances in neuro- science were associated with the development of new technologies and methods which pro- vided novel information about the brain and neural elements, permitting scientists to formu- late new hypotheses and theories. In the early years of the 20Th century, a German neurologist Korbinian Brodmann ∗Email: [email protected] conducted a cornerstone research (Brodmann, 1908) which changed the view we had about the brain. Studying the cytoarchitecture organi- zation of the cortex, he discovered that distinct anatomical areas were composed by popula- tions of cells of different morphology and den- sities. This experimental evidence allowed him to make the remarkable assumption that cy- toarchitecture distinctions over specific cortical regions might have functional consequences. Since then, a lot of research has been done aimed at classifying the cortex into sets of anatomical specialized areas and understand- ing the working brain in terms of the activa- tion/deactivation of those specialized regions. This approach has shown fruitful and many experiments demonstrate that different parts 1 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • of the cerebral manto are accounted for special- ized functions and activities. Over the last decades, the availability of recording techniques of higher accuracy and resolution, permitted precise and simultaneous recording of several brain regions and struc- tures. Experiments involving those techniques began to reveal a novel aspect of the working brain. They have brought evidence that many distinct specialized areas functionally interact while the brain perform its activities. Based on this evidence, many neuroscientists are getting into a consensus that, in essence, specialized brain areas do not work in isolation, and re- gard that neural activities are accomplished by networks that involve specific interactions among several distinct anatomical areas. It is believed that the assumed networks repre- sent the precise way how the brain processes and integrates information (Stam and Reijn- eveld, 2007; Sporns, 2011), and is also believed that those networks constitute a substrate ac- counted for the support of different dynamics and neural processes (Bassett and Bullmore, 2006). Up to now, the understanding on the way how cognitive capacities or different phys- iological states of the organism are translated into networks of distinct properties is not com- pletely known, motivating research and exper- iments in this area. Among many relevant topics to be investigated, the present study has sought to verify the influence of general anes- thesia on functional brain networks. General anesthesia is a physiological stable and reversible state induced by drugs that is characterized by analgesia, immobility, amne- sia and loss of consciousness (Schwartz et al., 2010). In addition to the undeniable impor- tance and usefulness for clinical medicine, anes- thetic agents also constitute tools of great im- portance for neuroscience. Once their phar- macological effects involve a drastic and fast reduction in awareness and levels of conscious- ness, they offer an opportunity for the study of consciousness and neural correlates of con- sciousness (Uhrig et al., 2014), providing possi- bilities for the comprehension of fundamental processes and phenomena that happen in the brain (Hameroff et al., 1998). The characterization of the large scale func- tional brain networks during general anesthe- sia has already been performed and reported (Padovani, 2016a). However, once different cor- tical areas are functionally specialized, the ef- fects of general anesthesia on different brain regions are not the same, and this motivates the investigation of the structure assumed by func- tional interactions of distinct brain regions1. Within this context, this study has sought to characterize the functional networks in four cortical brain lobes (frontal, parietal, temporal and occipital). Note: The networks presented in (Padovani, 2016a) are not "the sum" of the networks of this manuscript, once the large scale functional brain networks (Padovani, 2016a) also involve interac- tions established among areas of distinct brain lobes. There is experimental evidence that segregated and highly functionally integrated neural processes that involve large cortical areas and transcend brain lobes anatomical divisions occur most of the time (Padovani, 2016b). II. Methods I. Concepts Animal Model Experimentally, an old world monkey of the species Macaca fuscata was used as animal model. This species of macaque has great anatomical and evolutionary similarity with humans, making them an excellent platform for the study of the human brain (Iriki and Sakura, 2008). Being the Macaca fuscata one of the main primates used as experimental model in neuroscience, specially among Japanese neu- roscientists (Isa et al., 2009). 1Networks of interactions that involve cortical areas of each brain lobe, without considering interactions that involve areas of distinct brain lobes. 2 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Anesthetic Agents A cocktail of Ketamine and Medetomidine was used to induce anesthesia. Ketamine is a drug able to induce an anesthetic state characterized by the dissociation between the thalamocorti- cal and limbic systems (Bergman, 1999). It acts as a non competitive antagonist of the recep- tor N-metil-D-asparthate (Green et al., 2011). Medetomidine (agonist alfa-2 adrenergic recep- tor) was combined with Ketamine to promote muscular relaxation (Young et al., 1999). The antagonist of the Medetomidine, Atipamezole, was used to trigger and promote the recovering process (Young et al., 1999). Recording Technique The technique used to record neural activ- ity was the Multidimensional Recording Elec- trocorticogram (MDR-ECoG). This technology records neural activity using an ECoG elec- trode array implanted chronically at subdu- ral space directly over the cortical surface of the monkey. According to the researchers that developed this technology, the MDR-ECoG is the most advanced and balanced technology available to record neural activity, it is able to sample neural signals at temporal resolutions higher than 1KHz, and a spatial resolution up to 3mm (Yanagawa et al., 2013; Fukushima et al., 2014). Neural Connectivity Estimator As a neural connectivity estimator, a method based on Granger causality (Granger, 1969) was used in order to infer statistical depen- dencies between the time series of the elec- trodes. When applied in neuroscience, Granger causality provides an indication about the infor- mation flow from one cortical area to another (Seth and Edelman, 2007; Seth, 2010). Experimental Procedures Steps - Summary of The modelling of functional neural activities performed in this study followed the steps: 1. Neural registers were recorded by using the MDR-ECoG technique. The matrix of ECoG electrodes continuously covered an entire brain hemisphere and parts of the medial walls. 2. Each one of the electrodes of the array was considered as a vertex of the net- work, and represented the cortical area in which it was positioned. 3. A neural connectivity estimator of Granger causality in the frequency do- main was used to estimate values of asso- ciation between the registers (time series) of the electrodes. 4. For each brain lobe, an adjacency ma- trix was assembled, containing the pair- wise association values between the cor- responding nodes of the respective brain lobe. 5. The characterization of the topology of the estimated networks was performed by some complex networks measures. For each brain lobe, networks were esti- mated serially in time intervals of five seconds, in five physiological frequency bands along the experiment. For each frequency band, all the networks in the sequences were estimated through the same procedures and parameters2. Thus, the alterations observed between the measures of distinct networks in the sequence came only from differences on the records of neural activity. Note: All the experimental procedures involv- ing the animal model and data records (Methods Sections II and III), were performed by researchers from the laboratory of adaptive intelligence at the RIKEN BRAIN SCIENCE INSTITUTE, laboratory under the supervision of PhD. Naotaka Fujii. 2For each frequency band, the same threshold was used on the networks of different cortical lobes. 3 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • II. Subjects and Materials A multi-channel ECoG electrode array (Unique Medical, Japan) was chronically implanted in the subdural space of an adult male macaque (Macaca fuscata). Each electrode was made of a 3 mm platinum disc, coated with insulating silicone, except for a 0.8mm diameter central area of the disc. One hundred and twenty eight ECoG electrodes with an average inter- electrode distance of 5 mm were implanted in the left brain hemisphere, continuously cov- ering the frontal, parietal and occipital lobes. Some electrodes were also disposed in the medial frontal and parietal walls. Rectangular platinum plates were placed in the subdural space between the ECoG array and the du- ramater, to serve as reference electrodes. A ground electrode was placed in the epidural space, see (Nagasaka et al., 2011) for a detailed description. Electromyography (EMG) was performed at 1Khz , by a CerebusTM data acqui- sition system (Blackrock Microsystems, USA). All experimental and surgical procedures were performed according to experimental protocols (No. H24-2-203(4)) approved by the RIKEN ethics committee and the recommendations of the Weatherall report, "The use of non-human primates in research". The database used in this study is available in public domain at (http://neurotycho.org), for further information see (Nagasaka et al., 2011). III. Experimental Procedures The monkey was seated in a proper chair with head and arms restrained. Neural activity started to be recorded with the monkey awake and with open eyes. After that, the eyes were covered with a patch to prevent visual evoked response. After about 10 minutes, a Ketamine- Medetomidine cocktail (5.6mg/Kg of Ketamine + 0.01mg/Kg of Medetomidine) was injected intramuscularly to induce anesthesia. The loss of consciousness point (LOC) was set at the time when the monkey no longer responded to external stimulus (touching the nostrils or 4 opening the hands). After established the LOC, neural activity was recorded for more 25-30 minutes. Heart rate and breathing were moni- tored throughout the entire experiment. IV. Signal Processing and Granger Causality in the Frequency Domain Data Processing 1. A reject-band IIR-notch filter was used to attenuate components of the signal at 50Hz. 2. The signal was down-sampled from 1KHz to 200Hz. 3. The signal was divided into windows of 1000 points (equivalent to a five-second recording of neural activity). 4. For each of the 128 time series, the trend was removed and the average was sub- tracted. 5. To verify the stationary condition of the time series, the tests KPSS (Kwiatkowski et al., 1992) and ADF [Augmented Dickey Fuller](Hamilton, 1989) were applied. Libraries Used For the computation of association values us- ing Granger causality in the frequency do- main, were used with some adaptations the following libraries: MVGC GRANGER TOOLBOX, developed by PhD. Anil Seth (Sussex Uni- versity, UK), described in (Seth, 2010), avail- able at www.anilseth.com, and the library BSMART toolbox (Brain-System for Multivariate AutoRegressive Timeseries toolbox) described in (Cui et al., 2008) and available at www.brain-smart.org. Computation of Causal Interactions 1. Model Order: To find the model order (number of ob- servations to be used in the regression model), the criteria of selection of models from Akaike (AIC) and Bayes/Schwartz Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • (BIC) were used. Both methods returned an order of the model equal to seven. 2. Causal Interactions each window of At 1000 points, Granger causality in the frequency domain interactions were pair-wise computed among the time series re- spective to each brain lobe, by us- ing the function cca_pwcausal() (MVGC GRANGER TOOLBOX). 3. Frequency Bands Granger causality interactions were cal- culated in five physiological frequency bands: Delta (0-4Hz), Theta (4-8Hz), Al- pha (8-12Hz), Beta (13-30Hz) and Gamma (25-100Hz). The interaction values obtained were saved in adjacency matrices. • Gamma (25-100Hz), threshold = 2.0 As discussed in (Bullmore and Sporns, 2009; Sporns, 2011), it is possible for scientists to use different criteria in order to determine the threshold parameter. In the present study, due to experimental conditions, each sequence of networks contained graphs with distinct connectivity. Thresholds were chosen in such a way to prevent graphs with lower connectivity in each networks sequence from presenting many disconnected parts or vertices, which might introduce distortions in the analysis. Networks contained only vertices respective to electrodes positioned over the correspond- ing brain lobe analysed3. The number of vertices of the networks of each brain lobe were: Graphs and Networks 1. Assemble Networks For each sequence of graphs respective to a frequency band, a threshold was cho- sen, and only the interactions with mag- nitude values higher this threshold were considered as edges of the graphs. The same threshold was applied on the dis- tinct brain lobes analysed. • Delta (0-4Hz), threshold = 0.6 • Theta (4-8Hz), threshold = 0.4 • Alpha (8-12Hz), threshold = 0.3 • Beta (13-30Hz), threshold = 0.6 • Frontal lobe = 47 vertices • Parietal lobe = 23 vertices • Temporal lobe = 26 vertices • Occipital lobe = 34 vertices After obtaining non-weighted graphs, the directions of the edges were removed, resulting in undirected and non-weighed networks. Analysis of the Topology Network measures were used in order to char- acterize the topology and properties of the graphs. 3A few electrodes positioned over sulci divisions were included on both brain lobes. 5 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • III. Results I. Measures Related to Connectivity I.1 Average Degree Delta 0-4Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 1: Average Degree. Average degree vertex respective to the networks of the Delta frequency band (0-4Hz). Vertical axis average degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 1: Average Degree. Mean, variance (Var), and standard deviation (SD) of the average degree vertex of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Delta (0-4Hz). Eyes Open Delta Band (0-4Hz) Corresponding Graph Mean Var 37.5 Frontal Lobe 5.5 Parietal Lobe Temporal Lobe 5.80 22.0 Occipital Lobe 9.62 3.91 5.45 7.68 Eyes Closed Anesthesia SD Mean Var 63.4 6.12 16.7 2.35 2.41 13.0 47.4 4.69 11.2 6.25 5.55 13.9 SD Mean Var 48.0 7.96 20.7 4.08 3.61 8.87 25.5 6.89 14.7 7.23 7.86 14.3 SD 6.93 4.55 2.98 5.05 6 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Theta 4-8Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 2: Average Degree. Average degree vertex respective to the networks of the Theta frequency band (4-8Hz). Vertical axis average degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 2: Average Degree. Mean, variance (Var), and standard deviation (SD) of the average degree vertex of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Theta (4-8Hz). Eyes Open Theta Band (4-8Hz) Corresponding Graph Mean Var Frontal Lobe 60.7 6.88 Parietal Lobe 6.27 Temporal Lobe Occipital Lobe 11.3 14.4 5.33 6.62 5.96 Eyes Closed Anesthesia SD Mean Var 7.79 106 16.8 2.62 10.0 2.50 3.35 44.1 22.7 10.8 10.6 11.5 SD Mean Var 10.3 6.7 7.95 4.10 2.57 3.21 6.64 19.1 6.7 5.18 5.61 13.58 SD 2.59 2.82 1.60 4.37 7 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Alpha 8-12Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 3: Average Degree. Average degree vertex respective to the networks of the Alpha frequency band (8-12Hz). Vertical axis average degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 3: Average Degree. Mean, variance (Var), and standard deviation (SD) of the average degree vertex of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Alpha (8-12Hz). Eyes Open Alpha Band (8-12Hz) Corresponding Graph Mean Var Frontal Lobe 85.6 9.57 Parietal Lobe 6.87 Temporal Lobe Occipital Lobe 7.55 21.5 8.10 8.22 6.60 Eyes Closed Anesthesia SD Mean Var 9.25 118 15.0 3.09 10.7 2.62 2.75 28.5 29.7 13.8 12.3 11.4 SD Mean Var 10.9 1.30 2.46 3.87 1.54 3.28 5.34 19.5 3.14 3.30 4.12 12.8 SD 1.14 1.57 1.24 4.42 8 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Beta 13-30Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 4: Average Degree. Average degree vertex respective to the networks of the Beta frequency band (13-30Hz). Vertical axis average degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 4: Average Degree. Mean, variance (Var), and standard deviation (SD) of the average degree vertex of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Beta (13-30Hz). Eyes Open Beta Band (13-30Hz) Corresponding Graph Mean Var Frontal Lobe 89.3 7.27 Parietal Lobe 5.90 Temporal Lobe Occipital Lobe 15.1 32.9 12.7 11.8 13.5 Eyes Closed Anesthesia SD Mean Var 9.45 99.2 9.39 2.70 4.84 2.43 3.88 20.1 34.1 14.7 12.0 13.6 SD Mean Var 9.96 2.48 4.77 3.06 2.90 2.20 4.48 23.2 3.29 4.24 4.97 14.8 SD 1.57 2.18 1.70 4.82 9 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Gamma 25-100Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 5: Average Degree. Average degree vertex respective to the networks of the Gamma frequency band (25-30Hz). Vertical axis average degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 5: Average Degree. Mean, variance (Var), and standard deviation (SD) of the average degree vertex of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Gamma (25-100Hz). Gamma Band (25-100Hz) Corresponding Graph Frontal Lobe Parietal Lobe Temporal Lobe Occipital Lobe Eyes Open Eyes Closed Anesthesia Mean Var 9.28 31.2 24.1 5.18 2.65 4.75 6.47 11.9 SD Mean Var 5.59 36.1 31.5 4.91 8.43 1.63 3.43 8.65 9.64 7.03 5.92 4.37 SD Mean Var 6.01 1.17 2.53 5.62 5.85 2.90 2.94 19.4 4.08 3.94 8.16 15.2 SD 1.08 1.59 2.42 4.40 10 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • I.2 Assortativity Delta 0-4Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 6: Assortativity. Assortativity respective to the networks of the Delta frequency band (0-4Hz). Vertical axis assortativity; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 6: Assortativity. Mean, variance (Var), and standard deviation (SD) of the assortativity of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Delta (0-4Hz). Eyes Open Delta Band (0-4Hz) Corresponding Graph Mean Var 0.03 Frontal Lobe 0.05 Parietal Lobe Temporal Lobe 0.05 0.02 Occipital Lobe -0.22 -0.22 -0.07 -0.04 Eyes Closed Anesthesia SD Mean Var 0.01 0.18 0.02 0.22 0.22 0.05 0.02 0.15 -0.23 -0.22 -0.11 -0.01 SD Mean Var 0.01 0.12 0.03 0.15 0.22 0.03 0.03 0.14 0.06 -0.09 -0.01 0.08 SD 0.12 0.17 0.17 0.16 11 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Theta 4-8Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 7: Assortativity. Assortativity respective to the networks of the Theta frequency band (4-8Hz). Vertical axis assortativity; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 7: Assortativity. Mean, variance (Var), and standard deviation (SD) of the assortativity of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Theta (4-8Hz). Eyes Open Theta Band (4-8Hz) Corresponding Graph Mean Var Frontal Lobe 0.02 0.03 Parietal Lobe 0.03 Temporal Lobe Occipital Lobe 0.02 -0.24 -0.19 -0.01 -0.08 Eyes Closed Anesthesia SD Mean Var 0.15 0.01 0.02 0.17 0.03 0.18 0.15 0.02 -0.17 -0.17 0.02 0.0 SD Mean Var 0.12 0.02 0.04 0.13 0.04 0.17 0.15 0.01 0.12 -0.05 0.03 0.02 SD 0.13 0.19 0.19 0.12 12 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Alpha 8-12Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 8: Assortativity. Assortativity respective to the networks of the Alpha frequency band (8-12Hz). Vertical axis assortativity; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 8: Assortativity. Mean, variance (Var), and standard deviation (SD) of the assortativity of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Alpha (8-12Hz). Eyes Open Alpha Band (8-12Hz) Corresponding Graph Mean Var Frontal Lobe 0.02 0.02 Parietal Lobe 0.03 Temporal Lobe Occipital Lobe 0.03 -0.20 -0.16 -0.02 -0.07 Eyes Closed Anesthesia SD Mean Var 0.15 0.01 0.02 0.14 0.03 0.17 0.17 0.02 -0.11 -0.11 0.10 0.01 SD Mean Var 0.10 0.03 0.05 0.13 0.04 0.17 0.13 0.02 0.06 -0.08 0.00 0.01 SD 0.18 0.23 0.21 0.12 13 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Beta 13-30Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 9: Assortativity. Assortativity respective to the networks of the Beta frequency band (13-30Hz). Vertical axis assortativity; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 9: Assortativity. Mean, variance (Var), and standard deviation (SD) of the assortativity of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Beta (13-30Hz). Eyes Open Beta Band (13-30Hz) Corresponding Graph Mean Var Frontal Lobe 0.01 0.03 Parietal Lobe 0.02 Temporal Lobe Occipital Lobe 0.01 -0.07 -0.07 0.05 -0.02 Eyes Closed Anesthesia SD Mean Var 0.10 0.01 0.02 0.17 0.02 0.12 0.11 0.01 -0.04 -0.02 0.18 0.08 SD Mean Var 0.08 0.03 0.04 0.15 0.05 0.15 0.10 0.01 -0.02 -0.07 -0.03 -0.06 SD 0.18 0.20 0.22 0.12 14 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Gamma 25-100Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 10: Assortativity. Assortativity respective to the networks of the Gamma frequency band (25-100Hz). Vertical axis assortativity; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 10: Assortativity. Mean, variance (Var), and standard deviation (SD) of the assortativity of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Gamma (25-100Hz). Gamma Band (25-100Hz) Corresponding Graph Frontal Lobe Parietal Lobe Temporal Lobe Occipital Lobe Eyes Open Eyes Closed Anesthesia Mean Var -0.27 0.01 0.05 -0.17 0.04 0.07 0.08 0.03 SD Mean Var 0.08 0.01 0.05 0.22 0.04 0.21 0.16 0.04 -0.26 0.07 0.08 0.05 SD Mean 0.09 -0.25 -0.06 0.22 0.03 0.21 0.21 0.00 Var 0.002 0.04 0.02 0.01 SD 0.13 0.20 0.15 0.12 15 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • II. Measures Related to Global Integration II.1 Average Path Length Delta 0-4Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 11: Average Path Length. Average path length respective to the networks of the Delta frequency band (0-4Hz). Vertical axis average path length; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 11: Average Path Length. Mean, variance (Var), and standard deviation (SD) of the average path length of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Delta (0-4Hz). Eyes Open Delta Band (0-4Hz) Corresponding Graph Mean Var Frontal Lobe 0.22 0.20 Parietal Lobe 0.14 Temporal Lobe 0.24 Occipital Lobe 2.16 2.16 2.16 2.19 Eyes Closed Anesthesia SD Mean Var 0.46 0.20 0.21 0.45 0.25 0.38 0.49 0.16 2.03 1.97 2.25 1.74 SD Mean Var 0.45 0.13 0.28 0.45 0.21 0.5 0.4 0.11 1.91 1.99 2.06 1.72 SD 0.37 0.53 0.45 0.33 16 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Theta 4-8Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 12: Average Path Length. Average path length respective to the networks of the Theta frequency band (4-8Hz). Vertical axis average path length; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 12: Average Path Length. Mean, variance (Var), and standard deviation (SD) of the average path length of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Theta (4-8Hz). Eyes Open Theta Band (4-8Hz) Corresponding Graph Mean Var Frontal Lobe 0.13 0.13 Parietal Lobe 0.10 Temporal Lobe Occipital Lobe 0.27 1.83 1.95 2.07 2.39 Eyes Closed Anesthesia SD Mean Var 0.37 0.13 0.09 0.35 0.11 0.32 0.52 0.29 1.57 1.57 1.74 1.95 SD Mean Var 0.36 0.28 0.33 0.31 0.21 0.32 0.54 0.09 2.60 2.26 2.40 1.71 SD 0.47 0.57 0.46 0.30 17 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Alpha 8-12Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 13: Average Path Length. Average path length respective to the networks of the Alpha frequency band (8-12Hz). Vertical axis average path length; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 13: Average Path Length. Mean, variance (Var), and standard deviation (SD) of the average path length of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Alpha (8-12Hz). Eyes Open Alpha Band (8-12Hz) Corresponding Graph Mean Var Frontal Lobe 0.11 0.12 Parietal Lobe 0.10 Temporal Lobe Occipital Lobe 0.31 1.61 1.77 1.95 2.30 Eyes Closed Anesthesia SD Mean Var 0.33 0.12 0.07 0.35 0.12 0.31 0.56 0.29 1.41 1.40 1.64 1.93 SD Mean Var 0.35 0.33 0.27 0.26 0.24 0.34 0.54 0.10 2.95 2.32 2.43 1.75 SD 0.58 0.52 0.49 0.32 18 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Beta 13-30Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 14: Average Path Length. Average path length respective to the networks of the Beta frequency band (13- 30Hz). Vertical axis average path length; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 14: Average Path Length. Mean, variance (Var), and standard deviation (SD) of the average path length of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Beta (13-30Hz). Eyes Open Beta Band (13-30Hz) Corresponding Graph Mean Var Frontal Lobe 0.07 0.03 Parietal Lobe 0.04 Temporal Lobe Occipital Lobe 0.05 1.29 1.45 1.63 1.68 Eyes Closed Anesthesia SD Mean Var 0.26 0.07 0.03 0.17 0.04 0.19 0.22 0.11 1.26 1.35 1.63 1.73 SD Mean Var 0.27 0.42 0.25 0.17 0.18 0.20 0.34 0.06 3.13 2.25 2.33 1.62 SD 0.65 0.50 0.42 0.24 19 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Gamma 25-100Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 15: Average Path Length. Average path length respective to the networks of the Gamma frequency band (25-100Hz). Vertical axis average path length; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 15: Average Path Length. Mean, variance (Var), and standard deviation (SD) of the average path length of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Gamma (25-100Hz). Gamma Band (25-100Hz) Corresponding Graph Frontal Lobe Parietal Lobe Temporal Lobe Occipital Lobe Eyes Open Eyes Closed Anesthesia Mean Var 1.95 0.08 0.33 2.04 0.20 2.61 2.36 0.28 SD Mean Var 0.29 0.09 0.35 0.57 0.31 0.45 0.53 0.51 1.97 1.96 2.45 2.82 SD Mean Var 0.31 0.14 0.28 0.59 0.12 0.56 0.71 0.05 2.63 2.46 1.97 1.60 SD 0.37 0.53 0.34 0.22 20 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • II.2 Diameter Delta 0-4Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 16: Diameter. Diameter respective to the networks of the Delta frequency band (0-4Hz). Vertical axis diameter; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 16: Diameter. Mean, variance (Var), and standard deviation (SD) of the diameter of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Delta (0-4Hz). Eyes Open Delta Band (0-4Hz) Corresponding Graph Mean Var 1.82 Frontal Lobe 1.68 Parietal Lobe Temporal Lobe 1.13 2.20 Occipital Lobe 4.62 4.47 4.67 4.84 Eyes Closed Anesthesia SD Mean Var 2.06 1.35 1.65 1.30 1.06 2.19 1.53 1.48 4.29 4.03 4.98 3.97 SD Mean Var 1.37 1.43 1.97 1.29 1.48 1.60 1.01 1.24 4.19 4.16 4.31 3.65 SD 1.17 1.40 1.27 1.00 21 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Theta 4-8Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 17: Diameter. Diameter respective to the networks of the Theta frequency band (4-8Hz). Vertical axis diameter; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 17: Diameter. Mean, variance (Var), and standard deviation (SD) of the diameter of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Theta (4-8Hz). Eyes Open Theta Band (4-8Hz) Corresponding Graph Mean Var Frontal Lobe 1.29 1.16 Parietal Lobe 0.96 Temporal Lobe Occipital Lobe 2.39 3.69 4.03 4.55 5.28 Eyes Closed Anesthesia SD Mean Var 1.14 1.61 0.94 1.08 1.21 0.98 1.54 2.22 3.11 2.99 3.58 4.24 SD Mean Var 1.27 2.27 2.39 0.97 1.72 1.10 1.49 0.93 6.31 4.90 5.42 3.40 SD 1.51 1.55 1.31 0.96 22 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Alpha 8-12Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 18: Diameter. Diameter respective to the networks of the Alpha frequency band (8-12Hz). Vertical axis diameter; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 18: Diameter. Mean, variance (Var), and standard deviation (SD) of the diameter of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Alpha (8-12Hz). Eyes Open Alpha Band (8-12Hz) Corresponding Graph Mean Var Frontal Lobe 1.39 1.20 Parietal Lobe 0.86 Temporal Lobe Occipital Lobe 2.32 3.24 3.65 4.22 5.00 Eyes Closed Anesthesia SD Mean Var 1.18 1.40 0.84 1.10 1.12 0.93 1.52 2.27 2.82 2.60 3.40 4.15 SD Mean Var 1.18 3.01 2.09 0.91 1.96 1.06 1.51 1.04 7.02 5.02 5.42 3.44 SD 1.74 1.44 1.40 1.02 23 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Beta 13-30Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 19: Diameter. Diameter respective to the networks of the Theta frequency band (13-30Hz). Vertical axis diameter; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 19: Diameter. Mean, variance (Var), and standard deviation (SD) of the diameter of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Beta (13-30Hz). Eyes Open Beta Band (13-30Hz) Corresponding Graph Mean Var Frontal Lobe 0.63 0.45 Parietal Lobe 0.46 Temporal Lobe Occipital Lobe 0.51 2.43 2.80 3.34 3.33 Eyes Closed Anesthesia SD Mean Var 0.80 0.76 0.37 0.67 0.55 0.68 0.71 0.88 2.45 2.54 3.27 3.61 SD Mean Var 0.87 3.75 1.82 0.61 1.61 0.74 0.94 0.66 7.51 4.82 5.08 2.99 SD 1.94 1.35 1.27 0.81 24 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Gamma 25-100Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 20: Diameter. Diameter respective to the networks of the Gamma frequency band (25-100Hz). Vertical axis diameter; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 20: Diameter. Mean, variance (Var), and standard deviation (SD) of the diameter of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Gamma (25-100Hz). Gamma Band (25-100Hz) Corresponding Graph Frontal Lobe Parietal Lobe Temporal Lobe Occipital Lobe Eyes Open Eyes Closed Anesthesia Mean Var 3.81 1.32 2.62 4.23 2.03 6,.7 5.24 2.47 SD Mean Var 1.15 1.44 3.12 1.62 2.81 1.42 1.57 4.43 3.93 4.02 5.46 6.57 SD Mean Var 1.20 1.66 2.34 1.77 1.13 1.68 2.11 0.63 6.09 5.34 4.16 2.98 SD 1.29 1.53 1.06 0.79 25 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • III. Measures Related to Influence and Centrality III.1 Average Betweenness Centrality Degree Delta 0-4Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 21: Average Betweenness Centrality Degree. Average betweenness centrality degree respective to the networks of the Delta frequency band (0-4Hz). Vertical axis average betweenness centrality degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 21: Average Betweenness Centrality Degree. Mean, variance (Var), and standard deviation (SD) of the average betweenness centrality degree of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Delta (0-4Hz). Eyes Open Delta Band (0-4Hz) Corresponding Graph Mean Var 45.4 Frontal Lobe 17.9 Parietal Lobe 21.0 Temporal Lobe Occipital Lobe 36.0 21.6 8.53 11.0 16.0 Eyes Closed Anesthesia SD Mean Var 54.1 6.73 17.7 4.23 3.62 4.59 6.00 33.2 19.8 8.52 12.4 10.9 SD Mean Var 63.4 7.35 19.9 4.21 27.7 5.62 5.76 23.3 20.0 8.77 12.5 11.5 SD 7.96 4.46 5.26 4.82 26 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Theta 4-8Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 22: Average Betweenness Centrality Degree. Average betweenness centrality degree respective to the networks of the Theta frequency band (4-8Hz). Vertical axis average betweenness centrality degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 22: Average Betweenness Centrality Degree. Mean, variance (Var), and standard deviation (SD) of the average betweenness centrality degree of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Theta (4-8Hz). Eyes Open Theta Band (4-8Hz) Corresponding Graph Mean Var 32.4 Frontal Lobe 13.6 Parietal Lobe 11.2 Temporal Lobe Occipital Lobe 42.0 16.4 7.76 11.2 18.1 Eyes Closed Anesthesia SD Mean Var 41.8 5.69 8.79 3.69 11.0 3.35 6.48 47.3 12,0 5.75 8.74 13.5 SD Mean Var 84.0 6.46 27.4 2.96 33.9 3.32 6.88 19.7 28.4 10.8 15.8 11.5 SD 9.17 5.23 5.82 4.44 27 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Alpha 8-12Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 23: Average Betweenness Centrality Degree. Average betweenness centrality degree respective to the networks of the Alpha frequency band (8-12Hz). Vertical axis average betweenness centrality degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 23: Average Betweenness Centrality Degree. Mean, variance (Var), and standard deviation (SD) of the average betweenness centrality degree of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Alpha (8-12Hz). Eyes Open Alpha Band (8-12Hz) Corresponding Graph Mean Var 37.0 Frontal Lobe 10.3 Parietal Lobe 13.1 Temporal Lobe Occipital Lobe 35.5 13.3 7.33 11.1 18.2 Eyes Closed Anesthesia SD Mean Var 52.8 6.09 5.58 3.21 14.4 3.61 5.96 48.4 9.10 4.09 7.73 14.1 SD Mean Var 145 7.26 34.8 2.36 48.0 3.79 6.96 20.5 20.9 9.24 13.2 12.0 SD 12.1 5.90 6.93 4.52 28 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Beta 13-30Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 24: Average Betweenness Centrality Degree. Average betweenness centrality degree respective to the networks of the Beta frequency band (13-30Hz). Vertical axis average betweenness centrality degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 24: Average Betweenness Centrality Degree. Mean, variance (Var), and standard deviation (SD) of the average betweenness centrality degree of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Beta (13-30Hz). Eyes Open Beta Band (13-30Hz) Corresponding Graph Mean Var 35.0 Frontal Lobe 3.51 Parietal Lobe 5.53 Temporal Lobe Occipital Lobe 13.0 6.52 4.89 7.89 11.1 Eyes Closed Anesthesia SD Mean Var 37.1 5.91 3.48 1.87 5.73 2.35 3.60 24.5 5.92 3.81 7.87 11.9 SD Mean Var 179 6.09 23.7 1.87 30.1 2.39 4.95 14.6 26.8 9.82 13.5 10.0 SD 13.4 4.87 5.49 3.82 29 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Gamma 25-100Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 25: Average Betweenness Centrality Degree. Average betweenness centrality degree respective to the networks of the Gamma frequency band (25-100Hz). Vertical axis average betweenness centrality degree; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 25: Average Betweenness Centrality Degree. Mean, variance (Var), and standard deviation (SD) of the average betweenness centrality degree of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Gamma (25-100Hz). Gamma Band 25-100Hz) Corresponding Graph Frontal Lobe Parietal Lobe Temporal Lobe Occipital Lobe Eyes Open Eyes Closed Anesthesia Mean Var 27.2 20.1 21.3 6.42 27.5 17.6 18.9 54.5 SD Mean Var 28.1 5.21 25.3 4.62 44.3 5.25 7.38 100 20.0 7.30 16.1 22.0 SD Mean Var 51.0 5.3 26.9 5.03 16.4 6.66 10.0 13.0 29.7 12.5 11.6 9.84 SD 7.14 5.19 4.05 3.60 30 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Measures Related to Local Integration III.2 Transitivity Delta 0-4Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 26: Transitivity. Transitivity coefficient respective to the networks of the Delta frequency band (0-4Hz). Vertical axis transitivity coefficient; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 26: Transitivity. Mean, variance (Var), and standard deviation (SD) of the transitivity coefficient of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Delta (0-4Hz). Eyes Open Delta Band (0-4Hz) Corresponding Graph Mean Var Frontal Lobe 0.02 0.02 Parietal Lobe 0.02 Temporal Lobe 0.02 Occipital Lobe 0.34 0.26 0.44 0.40 Eyes Closed Anesthesia SD Mean Var 0.16 0.03 0.04 0.14 0.03 0.16 0.15 0.03 0.37 0.36 0.43 0.63 SD Mean Var 0.19 0.02 0.04 0.19 0.01 0.16 0.18 0.02 0.54 0.48 0.53 0.64 SD 0.12 0.20 0.11 0.15 31 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Theta 4-8Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 27: Transitivity. Transitivity coefficient respective to the networks of the Theta frequency band (4-8Hz). Vertical axis transitivity coefficient; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 27: Transitivity. Mean, variance (Var), and standard deviation (SD) of the transitivity coefficient of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Theta (4-8Hz). Eyes Open Theta Band (4-8Hz) Corresponding Graph Mean Var Frontal Lobe 0.03 0.02 Parietal Lobe 0.02 Temporal Lobe Occipital Lobe 0.02 0.44 0.38 0.50 0.34 Eyes Closed Anesthesia SD Mean Var 0.17 0.03 0.03 0.15 0.01 0.13 0.13 0.03 0.62 0.59 0.63 0.56 SD Mean Var 0.19 0.01 0.02 0.16 0.01 0.11 0.18 0.01 0.47 0.43 0.50 0.58 SD 0.07 0.15 0.10 0.10 32 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Alpha 8-12Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 28: Transitivity. Transitivity coefficient respective to the networks of the Alpha frequency band (8-12Hz). Vertical axis transitivity coefficient; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 28: Transitivity. Mean, variance (Var), and standard deviation (SD) of the transitivity coefficient of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Alpha (8-12Hz). Eyes Open Alpha Band (8-12Hz) Corresponding Graph Mean Var Frontal Lobe 0.03 0.02 Parietal Lobe 0.01 Temporal Lobe Occipital Lobe 0.01 0.58 0.52 0.56 0.39 Eyes Closed Anesthesia SD Mean Var 0.18 0.03 0.02 0.14 0.01 0.12 0.10 0.02 0.75 0.72 0.71 0.54 SD Mean Var 0.17 0.01 0.02 0.13 0.01 0.09 0.15 0.01 0.40 0.36 0.46 0.55 SD 0.09 0.15 0.11 0.10 33 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Beta 13-30Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 29: Transitivity. Transitivity coefficient respective to the networks of the Beta frequency band (13-30Hz). Vertical axis transitivity coefficient; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 29: Transitivity. Mean, variance (Var), and standard deviation (SD) of the transitivity coefficient of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Beta (13-30Hz). Eyes Open Beta Band (13-30Hz) Corresponding Graph Mean Var Frontal Lobe 0.02 0.01 Parietal Lobe 0.0 Temporal Lobe Occipital Lobe 0.01 0.81 0.70 0.68 0.58 Eyes Closed Anesthesia SD Mean Var 0.14 0.02 0.01 0.09 0.0 0.07 0.09 0.01 0.83 0.77 0.72 0.61 SD Mean Var 0.16 0.01 0.02 0.10 0.02 0.06 0.11 0.01 0.29 0.38 0.44 0.56 SD 0.09 0.15 0.12 0.11 34 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Gamma 25-100Hz (a) Frontal Lobe (b) Parietal Lobe (c) Temporal Lobe (d) Occipital Lobe Figure 30: Transitivity. Transitivity coefficient respective to the networks of the Gamma frequency band (25-100Hz). Vertical axis transitivity coefficient; Horizontal axis time (minutes). At t=11 minutes, the monkey was blindfolded, the first red line in each sub-figure represents the moment when a patch was placed over the eyes. At t=23 minutes the Ketamine-Medetomidine cocktail was injected, being represented by the second red line. The point of loss of consciousness (LOC) was registered at t=33 minutes, and is indicated by the third red line. Sub-figures: (a) Frontal Lobe; (b) Parietal Lobe; (c) Temporal Lobe; (d) Occipital Lobe. Table 30: Transitivity. Mean, variance (Var), and standard deviation (SD) of the transitivity coefficient of the networks respective to each one of the four cortical lobes, on the three different conditions in which the monkey was exposed during the experiment: awake with eyes open, awake with eyes closed and anesthesia (eyes closed). Frequency band Gamma (25-100Hz). Gamma Band (25-100Hz) Corresponding Graph Frontal Lobe Parietal Lobe Temporal Lobe Occipital Lobe Eyes Open Eyes Closed Anesthesia Mean Var 0.36 0.03 0.04 0.36 0.01 0.44 0.40 0.01 SD Mean Var 0.16 0.03 0.04 0.21 0.01 0.10 0.12 0.02 0.38 0.47 0.49 0.32 SD Mean Var 0.17 0.01 0.02 0.19 0.01 0.11 0.12 0.01 0.25 0.37 0.55 0.57 SD 0.08 0.12 0.08 0.10 35 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • IV. Discussion In this research, the administration of the anes- thetics has led to alterations on distinct prop- erties of functional brain networks on the four brain lobes and on the five frequency bands an- alyzed. Those results constitute experimental evidence that reveal distinctions in the way how functional interactions are established among cortical areas inside each brain lobe (frontal, parietal, temporal and occipital), dur- ing awake and general anesthesia conditions. Within less than two minutes after the ad- ministration of the anesthetic cocktail, expres- sive changes were concomitantly observed on several networks properties. The way how these alterations occurred strongly suggests the occurrence of a phase transition, in which the functional brain networks changed and as- sumed a distinct architecture. The fact that the changes were observed quite after the injection of the anesthetics indicates with a high degree of confidence that those alterations were re- lated to the pharmacological effects of the anes- thetics on the animal model, which involved an expressive reduction in awareness and level of consciousness. As it is believed that functional brain networks constitute a substrate that un- derlies different types of physical dynamics and emergent properties (Bassett and Bullmore, 2006), alterations on many networks properties at the onset of the transition to unconscious- ness suggest that the structure of the functional networks assumed during general anesthesia no more supported neural processes/activities necessary for awareness and conscious experi- ences. Thus, the results of this research indicate potential links between structural properties of brain functional networks to neural correlates of consciousness. I. Average Degree Prominent alterations on the average degree of the networks vertices were observed during the experiment (see Figs. 1-5 ; Tables 1-5). The average degree of networks respective to dis- tinct brain lobes presented different and char- 36 acteristic dynamic behavior in response to the experimental conditions (see Figs. 1-5 ; Tables 1-5). On lower frequency bands (0-4 Hz), the anesthetics induced an increase in the average degree on the frontal, parietal and temporal lobes (compare 0-20 min to 23 -33 min on Fig. 1 (cid:100)a, b, c(cid:101) ; Table 1). In the same brain regions, after established LOC, the average degree de- creased and also presented a smaller variation (see Fig. 1 (cid:100)a, b, c(cid:101)). In the experiment, the most remarkable al- terations in the average degree of the networks vertices were observed on medium frequencies (4-30Hz). In Theta, Alpha and Beta frequency bands, the anesthetics promoted an expressive reduction in the functional connectivity on the frontal, parietal and temporal lobes (see Figs. 2,3,4 (cid:100)a, b, c(cid:101) ; Tables 2,3,4). A different re- sponse to the administration of the drugs cock- tail occurred at the occipital lobe where the mean connectivity substantially increased (see Figs. 2,3,4 (cid:100)d(cid:101) ). In the Gamma frequency band (25-100Hz), the anesthetic induction led to a decrease on the average degree and on its variation over time on the frontal and parietal lobes (see Fig. 5 (cid:100)a, b(cid:101) ; Table 5). In the same frequency band, an increase on the connectivity of the networks was verified after the drug injection on the oc- cipital and temporal lobes (see Fig. 5 (cid:100)c, d(cid:101) ; Table 5). The average degree reflects the average number of connections of the networks ver- tices (Rubinov and Sporns, 2010) and brings information related to the graphs global con- nectivity. Thus, this network measure indicates how interactive the elements of the system are. Regarding the functional brain networks of the present study, this property provides an esti- mative of the functional connectivity on the respective brain lobe and frequency band. The results obtained in this research re- vealed that the administration of the anesthet- ics, led to an expressive reduction in the aver- age degree on networks of the frontal, parietal and temporal lobes (Theta, Alpha, and Beta bands) (see Figs. 2,3,4 ; Table 2,3,4). From Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • this experimental evidence, it is verified that the Ketamine-Medetomidine anesthetic induc- tion led to a drastic reduction on the functional connectivity in frontal, parietal and temporal areas. II. Assortativity Expressive alterations on the assortativity were observed during the experiment (see Figs. 6- 10; Tables 6-10). The networks of distinct brain lobes presented characteristic assortativity; and the distinct controlled experimental conditions in which the animal model was exposed dur- ing the experiment induced different behavior on this network property. The most remarkable alterations on the net- works assortativity occurred on the frontal lobe. During awake conditions (eyes open and closed) on the frequency bands Delta, Theta, Alpha and Beta, the functional networks of the frontal lobe were disassortative. Within less than two minutes after the administration of the anesthetics, a clear transition on this network structural property was verified, the networks once disassortative turned to be as- sortative4 (see Fig 6-9 (cid:100)a(cid:101) ; Tables 6-9). In awake conditions, between (13-30Hz), most of the networks of the temporal and oc- cipital lobes were assortative , after the admin- istration of the Ketamine-Medetomidine cock- tail, most of the time, the networks prevailed disassortative (see Fig 9 (cid:100)c, d(cid:101) ; Table 9). sition that the expressive and consistent alter- ations observed on the frontal lobe (see Fig 6-9 (cid:100)a(cid:101) ; Tables 6-9), may have impacted the networks in such a way that they no longer sup- ported certain types of neural processes and functional dynamics necessary for awareness and conscious experiences. Thus, these experi- mental findings indicate potential associations between the assortativity of the functional net- works of the frontal lobe to neural correlates of consciousness. III. Average Path Length Clear alterations on the average path length of the networks were observed within about one and a half minutes after the administration of the anesthetics (see Figs.11-15 ; Tables 11-15). The most remarkable alterations due to the anesthetic induction were verified on the frontal, parietal and temporal lobes. In the frequency bands Theta, Alpha and Beta a sub- stantial increase in the average path length oc- curred (see Figs. 12-14 (cid:100)a, b, c(cid:101), Tables 12-14). The networks of the occipital lobe presented a quite different response to the same experimen- tal conditions, once during general anesthesia (mainly after established LOC), a tendency of the networks to present shorter geodesic paths was verified (see Figs. 12-14 (cid:100)d(cid:101)). On the Gamma frequency band (25-100Hz) the administration of the anesthetics led to a decrease on the average path length on the tem- poral and occipital lobes networks (see Fig. 15 (cid:100)c, d(cid:101) ; Table 15). The assortativity (Boccaletti et al., 2006) is related to the existence of preferential attach- ment between networks vertices with respect to their connectivity degree (Boccaletti et al., 2006). Alterations in the assortative character reveal structural changes on the graphs, in the way in which the connections are established the among the nodes. As reported by Costa, the assortativity character may have a great influence on the dynamic processes supported by the system (di Bernardo et al., 2005; Brede and Sinha, 2005; Costa et al., 2007). From this observation it is possible to induce a suppo- According to Latora and Marchiori, the capacity and global efficiency of a network in transmitting information is directly related to the average of its minimum paths, be- ing the most efficient, those networks which have the shortest paths (Latora and Marchiori, 2001). The experimental results of this re- search demonstrate that the administration of the anesthetics led to a substantial increase in the average path length on the frontal parietal and temporal lobes. Thus, the results indicate 4Networks were assortative most of the time prevailing this character, there were also some instants that networks were dissassortative (see Figs. 6-9 (cid:100)a(cid:101)). 37 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • that on these brain lobes the capacity and effi- ciency of transmission information is reduced during general anesthesia. IV. Diameter In the experiment, alterations on the diame- ter were observed in response to the different experimental conditions in which the animal model was exposed (see Figs. 16-20 ; Tables 16-20). In the frequency bands Delta, Theta, Alpha and Beta, the diameter of the networks of the frontal parietal and temporal lobes increased after the administration of the anesthetics (see Figs. 16-19 (cid:100)a, b, c(cid:101), Tables 16-19). A differ- ent phenomenon occurred in the occipital lobe, where the diameter decreased during general anesthesia (see Figs. 16-19 (cid:100)d(cid:101), Tables 16-19). In the Gamma frequency band (25-100Hz) the diameter of the networks respective to the temporal and occipital lobes were shorter dur- ing general anesthesia (see Figs. 16-19 (cid:100)c, d(cid:101), Tables 16-19). The diameter of a network (Costa et al., 2007) as being related only to the larger geodesic path of the graph, is a measure less informative than the average path length, once the latter takes into account all the minimum paths of the network. The diameter reflects the length of the largest minimum path, rep- resenting the largest distance existent on the network5. In the present study, the main al- teration was an increase in the length of the diameter of the networks of the frontal pari- etal and temporal lobes on Delta, Theta, Al- pha and Beta frequency bands (see Figs. 16-19 (cid:100)a, b, c(cid:101), Tables 16-19). Such result supports the conclusions obtained analyzing the aver- age path length of the networks, that during general anesthesia, the global transmission of information is reduced in the frontal parietal and temporal lobes. 5Assuming that the graph is connected. 38 V. Average Betweenness Centrality Degree Alterations on the vertices mean betweenness centrality degree were observed during the experiment on all frequency bands and brain lobes analyzed (see Figs. 21-25 ; Tables 21-25). The most remarkable changes occurred on the frontal parietal and temporal lobes; on Al- pha and Beta frequency bands, the administra- tion of the anesthetics led to a substantial in- crease and also a considerable higher variation on the vertices average betweenness centrality degree (see Figs. 23, 24 (cid:100)a, b, c(cid:101) ; Tables 23, 24). VI. Transitivity In this study, the administration of the anes- thetics has led to alterations on the transitivity coefficient of the networks on the distinct brain lobes and frequency bands analyzed (see Figs. 26-30 ; Tables 26-30). The most prominent changes occurred on the frontal, parietal and temporal lobes on Al- pha and Beta frequency bands. In these brain regions, within less than two minutes after the administration of the anesthetics, the transi- tivity coefficient of the networks substantially decreased (see Figs. 28, 29 (cid:100)a, b, c(cid:101) ; Tables 28, 29). According to Latora and Marchiori, the lo- cal efficiency of transmission of information of a network is directly related to its transi- tivity coefficient, the larger the coefficient, the greater the local efficiency of the network (La- tora and Marchiori, 2003). The results obtained in this study reveal that the administration of the anesthetics led to a reduction on the tran- sitivity coefficient (see Figs. 28, 29 (cid:100)a, b, c(cid:101) ; Tables 28, 29). Such decrease observed on the frontal parietal and temporal lobes, indicates that during the induced state of anesthesia, the efficiency of transmission of information is re- duced in these cortical lobes. Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • V. Conclusions Authors Contribution The experimental results of this research re- vealed that the anesthetic agents altered the functional brain networks. Concomitant alter- ations on distinct networks properties observed within less than two minutes after the admin- istration of the drugs, strongly indicate that the loss of consciousness experienced by the animal model during the experiment was asso- ciated with a phase transition on the networks architecture. Such results also corroborate with the supposition that awareness and conscious experiences may be dependent on functional brain networks specifically structured . Networks of the four brain lobes, on the five frequency bands analyzed had character- istic properties; and each lobe also presented a distinct response due to the administration of the anesthetics. This observation reveals that, at local levels, in distinct anatomical ar- eas, functional brain networks are structured specifically. It was observed that functional neural ac- tivities are dynamic, as, considerable changes were verified on the networks measures in short time intervals. A dynamic nature was ver- ified in both awake (eyes open/closed) and gen- eral anesthesia conditions. Those results also reveal that the anesthetic induction far from "shutting down" brain activity, led the brain into a specific complex and dynamic state. Regarding the networks measures, the most remarkable results were: An accentuated de- crease in the average degree on frontal, parietal and temporal lobes 6 which indicates that the anesthetic agents compromised significantly the functional connectivity in these anatomi- cal areas; A prominent change in the assorta- tivity character of the functional networks of the frontal lobe that changed from disassorta- tive (awake) to assortative (general anesthesia). This experimental evidence potentially indi- cates the assortativity character of the frontal lobe functional networks as a neural correlate of consciousness. 6On frequency bands Theta, Alpha and Beta. All the experimental procedures involving the animal model and data records (Methods Sec- tions II and III), were idealized and performed by researchers from the laboratory of adaptive intelligence at the RIKEN BRAIN SCIENCE INSTITUTE, laboratory the under supervision of PhD. Naotaka Fujii. Eduardo C. Padovani worked with neural records database (Methods Section IV and V) . Idealized and performed the procedures, anal- ysed the results, and wrote the manuscript. Finantial Support This research was partially financed by CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior). During part of the development of this study, the author used the structure of the Lab- oratory Vision-eScience at Universidade de São Paulo, laboratory supported by FAPESP grant 2011/50761-2, CNPq, CAPES, NAP-eScience, PRP and USP. References Danielle Smith Bassett and ED Bullmore. Small- world brain networks. The neuroscientist, 12 (6):512 -- 523, 2006. Stewart A Bergman. Ketamine: review of its pharmacology and its use in pediatric anes- thesia. Anesthesia progress, 46(1):10, 1999. Stefano Boccaletti, Vito Latora, Yamir Moreno, Martin Chavez, and D-U Hwang. Complex networks: Structure and dynamics. Physics reports, 424(4):175 -- 308, 2006. Markus Brede and Sitabhra Sinha. Assortative mixing by degree makes a network more un- stable. arXiv preprint cond-mat/0507710, 2005. Korbinian Brodmann. Beiträge zur histologischen Lokalisation der Grosshirnrinde. 1908. 39 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • Ed Bullmore and Olaf Sporns. Complex brain networks: graph theoretical analysis of struc- tural and functional systems. Nature Reviews Neuroscience, 10(3):186 -- 198, 2009. L da F Costa, Francisco A Rodrigues, Gonzalo Travieso, and Paulino Ribeiro Villas Boas. Characterization of complex networks: A survey of measurements. Advances in Physics, 56(1):167 -- 242, 2007. Jie Cui, Lei Xu, Steven L Bressler, Mingzhou Ding, and Hualou Liang. Bsmart: a matlab/c toolbox for analysis of multichannel neural time series. Neural Networks, 21(8):1094 -- 1104, 2008. Mario di Bernardo, Franco Garofalo, and Francesco Sorrentino. Synchronization of degree correlated physical networks. arXiv preprint cond-mat/0506236, 2005. Makoto Fukushima, Richard C Saunders, Matthew Mullarkey, Alexandra M Doyle, Mortimer Mishkin, and Naotaka Fujii. An electrocorticographic electrode array for si- multaneous recording from medial, lateral, and intrasulcal surface of the cortex in Journal of neuroscience macaque monkeys. methods, 233:155 -- 165, 2014. Clive WJ Granger. Investigating causal re- lations by econometric models and cross- spectral methods. Econometrica: Journal of the Econometric Society, pages 424 -- 438, 1969. Steven M Green, Mark G Roback, Robert M Kennedy, and Baruch Krauss. Clinical prac- tice guideline for emergency department ke- tamine dissociative sedation: 2011 update. Annals of emergency medicine, 57(5):449 -- 461, 2011. Stuart R Hameroff, Alfred W Kaszniak, and Alwyn Scott. Toward a science of consciousness II: The second Tucson discussions and debates, volume 2. Mit Press, 1998. and the business cycle. Econometrica: Jour- nal of the Econometric Society, pages 357 -- 384, 1989. Atsushi Iriki and Osamu Sakura. The neu- roscience of primate intellectual evolution: natural selection and passive and intentional niche construction. Philosophical Transactions of the Royal Society B: Biological Sciences, 363 (1500):2229 -- 2241, 2008. Tadashi Isa, Itaru Yamane, Miya Hamai, and Haruhisa Inagaki. Japanese macaques as lab- oratory animals. Experimental Animals, 58(5): 451 -- 457, 2009. Denis Kwiatkowski, Peter CB Phillips, Peter Schmidt, and Yongcheol Shin. Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of econometrics, 54(1):159 -- 178, 1992. Vito Latora and Massimo Marchiori. Efficient behavior of small-world networks. Physical review letters, 87(19):198701, 2001. Vito Latora and Massimo Marchiori. Economic small-world behavior in weighted networks. The European Physical Journal B-Condensed Matter and Complex Systems, 32(2):249 -- 263, 2003. Yasuo Nagasaka, Kentaro Shimoda, and Nao- taka Fujii. Multidimensional recording (mdr) and data sharing: an ecological open re- search and educational platform for neuro- science. PloS one, 6(7):e22561, 2011. Eduardo C Padovani. Characterization of large scale functional brain networks during ketamine-medetomidine anesthetic induc- tion. arXiv preprint arXiv:1604.00002, 2016a. Eduardo C Padovani. Characterization of the community structure of large scale functional brain networks during ketamine- medetomidine anesthetic induction. arXiv preprint arXiv:1606.04719, 2016b. James D Hamilton. A new approach to the eco- nomic analysis of nonstationary time series Mikail Rubinov and Olaf Sporns. Complex net- work measures of brain connectivity: uses 40 Padovani, E. C. - arXiv preprint - Neurons and Cognition • November 2016 • and interpretations. Neuroimage, 52(3):1059 -- 1069, 2010. the brain. Nonlinear biomedical physics, 1(1):3, 2007. Robert S Schwartz, Emery N Brown, Ralph Lydic, and Nicholas D Schiff. General anes- thesia, sleep, and coma. New England Journal of Medicine, 363(27):2638 -- 2650, 2010. Anil K Seth. A matlab toolbox for granger causal connectivity analysis. Journal of neuro- science methods, 186(2):262 -- 273, 2010. Anil K Seth and Gerald M Edelman. Distin- guishing causal interactions in neural pop- ulations. Neural computation, 19(4):910 -- 933, 2007. Olaf Sporns. Networks of the Brain. MIT press, 2011. Cornelis J Stam and Jaap C Reijneveld. Graph theoretical analysis of complex networks in L Uhrig, S Dehaene, and B Jarraya. Cerebral In An- mechanisms of general anesthesia. nales francaises d'anesthesie et de reanimation, volume 33, pages 72 -- 82. Elsevier, 2014. Toru Yanagawa, Zenas C Chao, Naomi Hasegawa, and Naotaka Fujii. Large-scale in- formation flow in conscious and unconscious states: an ecog study in monkeys. PloS one, 8(11):e80845, 2013. SS Young, AM Schilling, S Skeans, and G Ri- tacco. Short duration anaesthesia with medetomidine and ketamine in cynomolgus monkeys. Laboratory animals, 33(2):162 -- 168, 1999. 41
1905.11249
1
1905
2019-05-24T14:07:21
Network properties of healthy and Alzheimer's brains
[ "q-bio.NC", "physics.soc-ph" ]
Small-world structures are often used to describe structural connections in the brain. In this work, we compare the structural connection of cortical areas of a healthy brain and a brain affected by Alzheimer's disease with artificial small-world networks. Based on statistics analysis, we demonstrate that similar small-world networks can be constructed using Newman-Watts procedure. The network quantifiers of both structural matrices are identified inside the probabilistic valley. Despite of similarities between structural connection matrices and sampled small-world networks, increased assortativity can be found in the Alzheimer brain. Our results indicate that network quantifiers can be helpful to identify abnormalities in real structural connection matrices.
q-bio.NC
q-bio
Cognitive Neurodynamics manuscript No. (will be inserted by the editor) Network properties of healthy and Alzheimer's brains Jos´e C. P. Coninck1 C. Iarosz4 · Fabiano A. S. Ferrari2 · Adriane S. Reis3 · Kelly · Antonio M. Batista5 · Ricardo L. Viana3 9 1 0 2 y a M 4 2 ] . C N o i b - q [ 1 v 9 4 2 1 1 . 5 0 9 1 : v i X r a Received: date / Accepted: date Abstract Small-world structures are often used to de- scribe structural connections in the brain. In this work, we compare the strucutural connection of cortical areas of a healthy brain and a brain affected by Alzheimer's disease with artificial small-world networks. Based on statistics analysis, we demonstrate that similar small- world networks can be constructed using Newman-Watts procedure. The network quantifiers of both structural matrices are identified inside the probabilistic valley. Despite of similarities between strcutural connection matrices and sampled small-world networks, increased assortivity can be found in the Alzheimer brain. Our results indicate that network quantifiers can be helpful to identify abnormalities in real structural connection matrices. Keywords Network · human brain · Alzheimer's disease · small-world 1Technolgical University of Paran´a Department of Statistics Curitiba, Brazil 2Federal Univesity of the Valleys of Jequitinhonha and Mu- curi, Institute of Enginnering, Science and Technology Jana´uba, Brazil 3Federal University of Paran´a, Department of Physics Curitiba, Brazil 4University of Sao Paulo, Institute of Physics Sao Paulo, Brazil 5State University of Ponta Grossa, Department of Mathematics and Statistics Ponta Grossa, Brazil 1 Introduction One of the first fully reported neural network was the worm C. elegans (White et al., 1986). The nervous sys- tem of C. elegans consists of 302 neurons connected through 5000 chemical and 600 electrical synapses. Wa- tts and Strogatz showed that the C. elegans brain net- work can be described by a small-world network (Watts and Strogatz, 1998; Varshney et al., 2011). Small-world networks are characterized by high clustering and short average dis- tance between nodes. They have been observed in brain networks of animals and humans (Sporns and Zwi, 2004; Bassett and Bullmore, 2006; Stam, 2014; Medina et al., 2008). Evidences of small-world properties can also be found in ensembles of neurons in vitro (Bettencourt et al., 2007). The small worldness of neuronal networks is hy- pothesized to be a consequence of optimization process associated with minimal wiring cost, robustness and balance between local processing and global integration (Reijneveld et al., 2007; Bullmore and Sporns, 2012). Brain networks can be obtained in different levels, such as microscale, mesoscale, and macroscale (Sporns et al., 2005; Heuvel and Yeo 2017, 2017). Microscale is in the level of the neurons and synapses, macroscale is used to define brain regions and large-scale communication pathways. Mesoscale is an intermediate level between micro and macroscale, where connections between large portions of the neuronal system are defined. A sim- ple example of mesoscale network is the mini-columns (Stoop et al., 2013). Neuronal networks are defined into structural or functional (Bullmore and Bassett, 2011). Functional net- works are based on EEG, MEG or fMRI measures (Stam and Straaten, 2012). Functional networks of Alzheimer's patients present increased path length when compared with healthy sub- jects (Stam et al., 2007). 2 Jos´e C. P. Coninck1 et al. 2 Methodology 2.1 Properties of networks Networks properties provide information about segre- gation, integration and influence (Rubinov and Sporns, 2010; Sporns, 2013). Segregation properties are associ- ated with the presence of clusters or modules and in- tegration properties are related to the network ability to transmit information through its nodes. Segregation and integration are linked with the network features while influence focus on the node features proving in- formation about the relevance of a node inside the net- work. Structural connections can be characterized by dif- fusion weighted magnetic resonance imaging (DW-MRI) and graph theory (Lo et al., 2010). DW-MRI analyzes water diffusion in white matter, and together with fiber tractography it can be used to identify structural con- nections in the brain (Medina et al., 2007). The struc- tural connection matrices of macaque and cats exhibit a complex structure (Hilgetag et al., 2000). The pres- ence of clusters and modular architecture in structural connection matrices are observed by means of corti- cal thickness measurements (Chen et al., 2008). Matri- ces with small-world properties and exponentially trun- cated power law distribution were also reported (Gong et al., 2008). In humans, the structural connection matrix medi- ates several complex cognitive functions (Bressler, 1995). Abnormalities in structural networks were found in pa- tients with psychiatric disorders and neurodegenera- tive diseases (Stam et al., 2007; He et al., 2008, 2009; Yao et al., 2010; Pol and Bullmore, 2013; Stam, 2014). Disconnection between frontal and temporal cortices were observed in patients with Schizophrenia (Friston and Frith, 1995; Zalesky et al., 2011). Hyperconnectivity in the frontal cortex were reported in patients with Autism (Courchesne and Pierce, 2005). Alzheimer's patient showed increased path length and reduced global efficiency (Lo et al., 2010). The alterations in brain networks are good indi- cators that network properties can be used as biomark- ers for clinical applications (Kaiser, 2011). 2.1.1 Eigenvalues of the adjacency matrix The eigenvalues of the adjacency matrix A are obtained by solving the characteristic equation of A, det(A − λI) = 0, (1) where I is the identity matrix and the values of λ that satisfy Eq. (1) are the eigenvalues (Cvetkovic et al. 2008, 2008). If the network is symmetric, Aij = Aji, then all the eigenvalues are real. 2.1.2 Degree and node strength Using diffusion tensor tractography, Lo et al. (Lo et al., 2010) constructed structural connection matrices of the human brain of healthy and Alzheimer's subjects. The network is divided in 78 areas according to the auto- mated anatomic label template (Tzourio-Mazoyer et al., 2002). The connection between the areas are defined in terms of the number of fibers, that were obtained through fiber assignment by continuous tracking algo- rithm (Mori et al., 1999). In this work, we analyze the network properties of one structural connection matrix related to a healthy subject and other related to a subject suffering of spe- cific neurodegenerative disease (Alzheimer). We demon- strate that similar networks to these brain matrices can be constructed using Newman-Watts procedure. In Section 2, we provide a brief discussion about the network representation of the connectome. In Sec- tion 3, we introduce basic quantities that can be used to quantify networks. In Section 4, we discuss the basic quantities of small-world networks in the light of statis- tical analyses. In Section 5, we compare the properties of human brain networks to small-world networks. In Section 6, we present our final remarks. Degree κi is the number of neighbors of a node i, κi = N Xj=1 Aij, (2) where N is the network size. It is considered one of the simplest measures to provide information about the influence of the network. The degree distribution is used to differentiate regular networks from random networks. For weighted networks (Wij ), the use of node streng- th si instead of degree κi may be more appropriated (Opsahl et al. 2010, 2010). Node strength si is defined as the sum of the node connections, si = N Xj=1 Wij. 2.1.3 Transitivity (3) Transitivity T , Also known as clustering, is a measure of the segregation of a network. The Transitivity T is a measure of the amount of clustering between the node i and its ki neighbours, the maximum number of con- nections between i neighbors is Cmax(i) = ki(ki − 1)/2. Network properties of healthy and Alzheimer's brains 3 Ci is defined as the ration between the number of ac- tive connections over the maximum number of connec- tions Cmax(i). The Transitivity T is the average over all nodes of the network. The transitivity shows the effective proportion of the triangulation formed between the sites as a measure of clustering capacity, G(E, V ). Then, T is calculated by the following proportional ratio T = 3δ(G) τ (G) , (4) where δ(G) or the number of triangles in graph G and τ (G) to denote the number of triples in graph G (Schank and Wagner, 2005). One simple method is to use the arithmetic mean (Opsahl and Pazaransa, 2009). If the nodes i, j, and k are connected, forming a triplet, the value of the triplet is the arithmetic mean between Wij and Wjk. A triplet is considered a close tripled when the nodes i, j, and k are all connected to each other. 2.1.4 Characteristic path length Characteristic path length L measures the average of the shortest paths dij between all pairs of nodes in the network, the site j belongs to group 1, then sj = 1, if j belongs to group 2, then sj = −1. Q can be either positive or negative, positive values indicate the possible presence of community structure. 2.1.6 Assortativity Assortativity ASR is a measure of the tendency of high connected nodes to be connected to others of similar degree k (Foster et al., 2010). When high connected nodes are more often connected to low connected nodes, the network exhibits dissortative mixing. To define as- sortivity it is necessary to define the remaining q(k) and p(k). The probability that a random node has a degree k is given by the degree distribution p(k), however, the probability to select a random edge is not proportional to p(k) but to kp(k), because the most connected nodes receive more connections. Considering that node i is connected to node j through a random selected edge, the remaining degree is the number of nodes that leaves the node j, excluding node i. The normalized remaining degree distribution is given by (k + 1)p(k) . q(k) = Pj jp(j) q Xij 1 σ2 The Assortativity ASR is defined as: ij(e(i, j) − q(i)q(j)), (7) (8) L = 2 N (N − 1) N N Xi=1 Xj=1 dij . (5) ASR = This quantity is used for weighted and unweighted net- works, it provides information about the network inte- gration. When dealing with diffusion process and weigh- ted networks, to calculate the shortest paths dij the in- verse of the node strength should be used (Opsahl et al. 2010, 2010). For example, if W12 = 2, then dij = 1/2, this ap- proach considers that the higher is the node strength the faster information can be diffused through it. where σ2 q is the variance of the remaining degree and e(i, j) is the joint probability distribution of the remain- ing degree of two nodes (Newman, 2002). The Assorta- tivity A is defined in the interval −1 ≤ A ≤ 1, when A = 1 the network has perfect assortative mixed pat- terns, A = 0 indicates the network is not assortative and A = −1 means the network is completely dissorta- tive. 2.1.5 Modularity Networks can be divided in two or more modules, the trivial solution is to divide them into two modules, where one module has one node and another module containing all the remaining nodes. Basically, the mod- ular structure is defined for any network and the ques- tion is to know the best method to identify modules in complex networks. An optimized quantity to character- ize the modularity Q was defined by Newman (Newman, 2006), that is given by 2.2 Statistical analysis 2.2.1 Generalized regression analysis A generalized linear model is made up of a combination of linear predictor with link function, 1. Pedictor linear: ηi = β0 + β1x1i + β1x2i · · · + βpxpi 2. Link function: g(µi) = η for exponencial family of distributions f (y, θ, φ) = exp(cid:26) yθ − b(θ) φ − c(y, φ)(cid:27) , Q = 1 4m Xij (cid:18)Aij − kikj 2m (cid:19) (sisj + 1), (6) with where m = 1/2Pi ki, si, and sj are indices that depend on the group. The network is divided in two groups, if E(Yi) = µi = d[b(θi)] dθ V ar(Yi) = φ−1Vi φ−1 > 0. V = dµ dθ   4 Jos´e C. P. Coninck1 et al. 2.2.2 Multivariate data analysis Table 1 Cronbachs alpha. Multivariate analysis is a branch of statistics that deals with the relationship between many variables, includ- ing the reduction of the number of variables observed during an experiment. The main tools for multivariate data analysis are principal component analysis (PCA) (Gray, 2017), factor analysis (Jhonson and Wischern, 1998), classifications (Jhonson and Wischern, 1998), struc- tural equations models (SEM) (Grace, 2016; Maruyama, 1998), among other techniques. In our case, the multi- variate analysis of the data is useful to vary the pos- sible second order relationships between variables not directly correlated, such as transitivity, assortativiness and the modularity of the human network. At the end of this paper we will see how these measures are related using the SEM (Maruyama, 1998). 2.2.3 Development of a questionnaire We apply a questionnaire to a population in the small- world models artificially generated with network size in 3 to 100 sites, from a single connection to the global con- nection. The determination of sample (Cochran, 1977) , α/2 (9) n = P100 N =3PN α/2 + (P100 k=1(cid:0)N N =3PN k(cid:1)pqz2 k=1(cid:0)N pqz2 k(cid:1) − 1)E2 for optimization p = q = 1 2 , with error E ≈ ±3% in N = 2499 population with zα/2 is a z-score distribu- tion with level of significance α ≈ 5%. In this case, the sample is n ≈ 492 small-world models. This ques- tionnaire is composed of 34 variables or questions about graph proprieties and applied to each small-world. Each model randomly generated with a certain probability is measured with these thirty-four variables. alpha Internal consistency 0.9 ≤ α Excellent 0.8 ≤ α ≤ 0.9 Good 0.7 ≤ α ≤ 0.8 Acceptable 0.6 ≤ α ≤ 0.7 Questionable 0.5 ≤ α ≤ 0.6 Poor α < 0.5 Unacceptable means of these measures. The small global templates are generated with 10 sites up to 100 sites. Each gener- ated model has different connections of its neighbor- hood between a single neighbor and the global net- work. This way, 492 samples of small-world models are produced. The measure of sampling adequacy (MSA) through the Kaiser-Meyer-Olkin (KMO) test indicates considered reasonable for value KM O = 0.73 (Kaiser, 1974). The KMO and RMSA measures are given by and KM O = Pk i=1Pk Pk i=1Pk j=1 r2 RM SA = Pk Pk j=1 r2 j=1 rij ij + a2 ij , j=1 rij ij + a2 ij , (11) (12) where rij is the correlation matrix term and aij is the anti-image-correlation matrix term. In this method, the inverse correlation matrix is close to the diagonal ma- trix. To verifies if matrix correlations is statistical equiv- alent to an identity matrix, we use the Bartlett's test. The basic hypothesis is that population's correlation matrix is an identity matrix equivalent. In our variables group, p-value<< 0.05 implies the rejection of the null hypothesis and accepting the factorial analysis. We verify the quality of the questionnaire through Cronbach's alpha (Cronbach 1951, 1951) 3 Results α = K K − 1 (cid:18)1 − Pi σ2 σ2 Xi Yi (cid:19) , where K is the number of components, σ2 is the vari- Xi ance of the observed total test scores, and σ2 is the Yi variance of the current sample of generated small world. The questionnaire quality applied to small-world net- works is equal to 0.89, indicating a good Internal con- sistency (Cronbach 1951, 1951; Maruyama, 1998). The questionnaire is composed of thirty-four ques- tions (or variables) measured directly in each artificial small-world network. Each variable measures an impor- tant network property and the comparison between the human network and the small-world model is given by 3.1 Structural connection matrices (10) In this work, we use two structural connection matri- ces, one for the healthy brain (Fig. 1(a)) and other for the Alzheimer's brain (Fig. 1(b)) (Lo et al., 2010). Both networks are weighted and symmetric, the weight is as- sociated with the intensity of connections and can as- sume five values: 0 (no connections, white region), 1 (low density of connections, indigo circles), 2 (interme- diate density of connections, red circles), and 3 (high density of connections, orange circles). The main re- sults for the networks are shown in Table 3. Figure 1 exhibits two adjacency matrix connection Wij : (a) healthy brain and (b) Alzheimer's brain. The eigenval- ues for these adjacency matrix are evaluated in Fig. 2. The healthy structural connection matrix is in black Network properties of healthy and Alzheimer's brains 5 Table 2 Variables of questionnaire - graph. N : p : Number of nodes Probability connection DE : Density edge L : Rep : E : N : N L : N LI : DL : T ST : T SF R : Cn : ALW : ACT : Cp : T : ASR : Ecc : Average path length Reciprocity Edges Vertices number of links Internal number of links link Density Total System Throughput Total System Flow Rate Conectancie Average Link Weight Average Compartment Throughflow Compartmentalization Transitivity Assortivity Eccenticity D : Diameter Imax : Maximum interweaving Imin : Minimum interweaving Q : Modularity E : R : Eficience Radius Kmax : Max K-Core Kmin : Min K-Core Kmean : Mean K-Core Iso : Auto : Ef cM : Ef cm : λp : Isomorfism Automorphism Number of edge with max efficiency Number of edge with minimal efficiency Principal eigenvalures detM : Determinant matrix and Alzheimer's structural connection matrix is in red. The eigenvalue spectrum for small world with 78 nodes is p = 0.0375 (blue line). The ordinate eigenvalues are very close for three structures matrix. The eigenvalues are equivalent when there is some difference in the dis- persion of adjacency matrix. Table 3 Network Indicators (W:Weighted Un:Unweighted). Transitivity T W 0.578 Assortivity ASR 0.081 Path Length L 2.248 Modularity M1 0.451 23.27 σ Average Degree 8.000 Healthy Un 0.578 0.010 2.248 0.423 0.590 1.383 Alzheimer Un 0.559 0.125 2.281 0.428 0.590 1.383 W 0.560 0.226 2.281 0.483 7.534 17.487 3.2 Small-world networks A network with small-world properties can be gener- ated by means of different methods (Newman, 2000). The most common method was developed by Watts and Strogatz (Newman, 2000), where the regular edges are replaced by random edges. When about 1% of the total edges are replaced, the network exhibits high transitiv- ity and low path length (Watts and Strogatz, 1998). For our analysis, we consider an alternative method, where (a) 80 x i r t a M x e d n I 60 40 20 0 0 80 (b) x i r t a M x e d n I 60 40 20 0 0 20 40 Index Matrix 60 80 20 40 Index Matrix 60 80 Fig. 1 Weighted connection matrix Wij for (a) healthy and (b) Alzheimer's brains. The weight is associated with the in- tensity of connections: 0 (no connections, white region), 1 (low density of connections, indigo circles), 2 (intermediate density of connections, red circles), and 3 (high density of connections, orange circles). instead of randomly replace regular edges by random edges, we only add random edges, known as Newman- Watts procedure (Newman, 2003). We add pN K new random edges, where N is the network size, K is the regular network degree, and p is the probability to add new edges. We vary p and identify the small-world prop- erties comparing T and L with the values of the regular network T (0) and L(0). We find -- Healthy brain DL : link density −→ DL 2 = 13.3333 2 ≈ 7, (13) -- Alzheimer's brain DL : link density −→ DL 2 = 13.38462 2 ≈ 7. (14) 6 Jos´e C. P. Coninck1 et al. l s e u a v n e g E i + + + + + + + + + +++ +++ + ++ ++ ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 0 2 5 1 0 1 5 0 5 − 0 20 40 60 80 Index Matrix Fig. 2 Eigenvalue spectrum for the weighted matrices of Fig. 1. Healthy structural connection matrix is in black and Alzheimer's structural connection matrix is in red. Eigenvalue spectrum for small-world with 78 nodes and connection prob- ability is p = 0.0375 in blue continue line. The number of nodes in the human graph is 78 sites and the average degree is 7 neighbors per node. Due to this fact we create the equivalent connection net- work under these conditions to depend exclusively on the probability of calling. It is generated small-world graph with number of nodes N = 78, 7 neighbours per node, and without the likelihood of connection (p = 0), implying in average length L0 = 3.27273 and transi- tivity C0 = 0.69231. For healthy human structural con- nection matrix, we find Lhealthy human = 2.24875 and Chealthy human = 0.57813. Then, we obtain Chealthy human C0 Lhealthy human L0 CAlzheimer's structural C0 LAlzheimer's structural L0 = 0.8350739, (15) = 0.6871175, ⇒ C C0 > L L0 , = 0.8087265, = 0.6971917, ⇒ C C0 > L L0 . C/C0 is approximately 1.4% more than Alzheimer's disease human cluster standardized, and L/L0 is ap- proximately 3.1% less than Alzheimer's disease human. The propagated information in Alzheimer's disease hu- man presents greater difficulty for diffusion of informa- tion in network, becoming more complex than healthy human matrix. Therefore, there seems to be a relation- ship between the transfer of the network and its group- ing, i.e., relations between assortivity, modularity and transitivity. In correlation matrix, we present a statistical corre- lations rij for variables assortivity (ASR), modularity (Q), and transitivity (T), respectively, for small-world samples classes used in RMSA and KMO analysis. All values are low and indicate the lack of direct correla- tion. 3.2.1 Regression analysis In a convenience sample, for n = 429 small-world type networks, five replicas are executed to create variation within the others. The dispersion of the transitivity ac- cording to the logarithm of the connection probability shows a decay adjusted by generalized model Gaussian family with link identity f (yµ, σ2) = exp(cid:20) 1 σ2 (cid:18)yµ − µ2 2 (cid:19) +(cid:18)− 1 2 ln(2πσ2) − y2 2σ2(cid:19)(cid:21) , resulting in the following regression TSW = 0.300314 − 0.029338 ln p. (16) For instance, when the probability connection is p = 0.00091188196, then log p is equal to −7. The result is approximately 0.50568. This probability value p is in agreement with the probability of small-world connec- tion. When the connection probability increases, the dispersion of the transitivity value increases as well. On the other hand, the decrease in probability link- age causes the dispersion to become smaller and more concentrated, characterizing a good small-world region. TH = TM = TSW is valid for the small-world model. Another feature of Eq. 16 is its rate of transitivity in relation to the log of the probability, dTSW d ln p = −0.029338 ≈ −3%. (17) The ratio of transitivity to ln p is equal to the loss value in the small-world model when we compare the matrix of healthy human adjacency with disease Alzheimer hu- man matrix. The transfer rate and the rate of flow in the network with Alzheimer's exhibit a drop equal to that caused in the transitivity when compared with the human net- work in the normal state, as shown in Table 4. It sug- gests that the rate of transfer and rate of flow for people with Alzheimer's disease declines with 5.4%, possibly due to the fall in transitivity in 3.1%. Network properties of healthy and Alzheimer's brains 7 Table 4 Transitivity loss (weighted). Transitivity Nodes Links total Transfer rate Leak rate Healthy Alzheimer 0.57813 0.5598876 loss 3.1% 78 1040 1438 1438 78 1044 1364 1364 5.4% 5.4% 3.2.2 Assortativity One of the most difficult measure to be statistically analyzed is the assortivity of the network. Due to the fact that the network topology of small-world is very sensitive to the probability of (re)connection. This can be verified in healthy and Alzheimer's human matrices. for T = 0.57, the assortivity shows very different values. However, the assortiveness is 2.5 times higher for the Alzheimer's brain than for healthy (Table 3). The Alzheimer's brain is more assortive than the healthy brain, this means that in the Alzheimer network the nodes with high degree are, in average, connected with other nodes of high degree more intensely than the healthy human network. We calculate the assortativity distribution for small-world networks for n = 492 sam- ples. In a sample, for example the small-world N = 10 and second order connection, the assortiveness presents an sample average of < ASR >= −0.01335938 not being statistically zero according to t-Student test for the hypothesis H0 : µASR = 0. There is no significant evidence to support the null hypothesis for a p-value << 0.05, inclining us to accept the hypothesis that as- sortiveness in the sample question is, in fact, negative. The network is on average disassortative. This does not imply the formation of positive assortiveness as verified in the graph. 3.2.3 Probabilistic valley The probabilistic valley is a region where the small- world structure behaves by sequences of abrupt changes. It is precisely in this region that we identify abrupt behaviors of assortativiness, given the equivalent mod- ularity and transitivity. These three measures of the small-world model that are equivalent to the measure- ments of the human matrix are found in this valley. The probabilistic voucher is developed through the struc- tural equations model (SEM), in which it is related in- directly to assortivity, transitivity and modularity. As assortiveness represents the equivalent of the correla- tion between the links of the sites of a network, we write the assortivity in function of the transitivity and the modularity of the network for determinate proba- bilistic valley. The probabilistic valley region indicates a possible existence of probability as a function of the modularity, transitivity and efficiency, that it is in agreement with the SEM analysis. This indicates that there is a possible dependence on the functions of modularity, assortive- ness, transitivity and efficiency, according to the SEM analysis. In the same region random overflow occurs in assortivity increasing transitivity and modularity. The increase in assortivity and transitivity implies in the de- cay of the connection probability, which confirms that the probability value decreases. A more detailed view of the level curve with the modularity in the abscissa of the assortative at the ordinate reveals a complex struc- ture of the curves. Outside this region the value of the assortiveness is zero or close to zero. The valley has many interesting behavior. The as- sortivity, transitivity and modularity measures exist on- ly because there are valley probabilistic. To generate Fig. 3, we consider a set of 100 independent models of small-world networks starting with 10 sites up to the amount 100 sites. In all models, we vary the probability of linkage between the non-coupling state (p = 10−6) and the overall state (p = 1). The red dot in Fig. 3 is lo- cated in the region where the modularity (≈ 0.45) and transitivity (≈ 0.58) have values equal to the results found through the human matrix of healthy individ- uals. The same point coincides with the result of the assortiveness (≈ −0.02) in the Alzheimer's human ma- trix. This graph located in this point have 78 sites with 7 connections neighbors and probability range 8.10−8 < p < 0.5. 4 Discussion The human networks are located in the probabilistic valley. We locate the healthy and Alzheimer's human networks within the small-world samples. The super- position both red and blue dots in evolutionary aver- age length measures and transitivity graphics in connec- tion probability is p = 0.0375, as shown by the vertical dotted line of Fig. 4. This ordered pair (cid:16) T (0.8350739, 0.6871175) for p = 0.0375 is exhibited by the blue dots in Fig. 4. The same technique is used to find the ordered pair representing Alzheimer's struc- tural connection matrix for T (0) , L L(0)(cid:17) = (cid:18) T T (0) , L L(0)(cid:19) = (0.8087265, 0.6971917). (18) In this case, it is represented by the red dots in same figure. In fact, when we select the region of the transitivity of Fig. 4 in the value of the probability of connection, we have that the difference between the blue and red 8 (a) ASR 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 0.1 0.2 0.3 0.4 0.5 T 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 (b) p 0.6 0.7 0.55 0.5 0.45 0.4 0.35 0.3 Q 0.25 0.2 0.15 0.1 0.5 0.45 0.4 0.35 0.3 Q 0.25 0.2 0.15 0.1 0.1 0.2 0.3 0.4 0.5 T 0.6 0.7 0.55 Fig. 3 Probabilistic valley as functions of modularity and transitivity for: (a) Assortivity and (b) probability of non- local connections. points is 3.1%, which represents the healthy individ- ual and Alzheimer's disease, respectively. The human Alzheimer's network exhibits greater difficulty in the transmission of information due to the fall of transi- tivity in the network. On the other hand, Fig. 4 also shows that the path length (L) is larger than the case of healthy individuals. The individuals with Alzheimer's disease have a decrease in the effectiveness of the tran- sitivity. T/T(0) L/L(0) 1 0.8 0.6 0.001 0.01 Probability P = 0.0375 0.1 Fig. 4 L/L(0) and T /T (0) as a function of the probability. The results for healthy and Alzheimer's structural connection matrix are shown by blue and red balls, respectively. Jos´e C. P. Coninck1 et al. The weighted connection matrix eigenvalues Wij is useful for the comparison between the matrix struc- tures. In Fig. 2, we display the eigenvalue spectrum for both networks of Fig. 4. We verify that the eigen- values of both networks are similar. The eigenvalues of the proposed small-world model are very close to the human adjacency matrices. The approximation of the variation of the transitivity region in Fig. 4 can be seen in Fig. 5. This figure show us two points, one blue and another red that represent the health human and Alzheimer's disease, respectively. The difference be- tween the the healthy and Alzheimer's brains is about 3.1%. 0.9 0.85 0.8 0.025 3.1 % 0.03 Probability 0.035 0.04 P = 0.0375 Fig. 5 Magnification of T /T (0) as a function of the proba- bility of Fig. 4. Table 5 shows that the average path length value of the healthy brain is very close to small-world network. The transitivity , assortivity, eccentricity, and modu- larity are almost identical. In fact, the relationship be- tween a human graph and a small-world structure is pertinent. In Table 6, we see that the Alzheimers's brain and small-world network have similar values, except the assortivity value, ε = 59.62%. Table 5 Values of the healthy brain and small-world net- work. Average path length Density of links Transitivity Assortivity Eccentricity Modularity Health human (real) 2.2487 13.333 0.5781 0.0815 3.6667 0.4515 Small-world (Simulated) 2.1964 14.000 0.5386 0.0882 3.5128 0.4889 error (ε) +2.32% -5.00% +6.79% -8.32% +4.20% -8.27% Network properties of healthy and Alzheimer's brains 9 Table 6 Values of the Alzheimer's brain and small-world network. References Alzheimer's brain (real) Small-world (Simulated) Average path length Density of links Transitivity Assortivity Eccentricity Modularity 2.28172 13.3846 0.55989 0.21846 3.76923 0.49083 2.1964 14.000 0.5386 0.0882 3.5128 0.4889 error (ε) +3.74% -4.59% +3.80% +59.62% +6.80% +0.39% 5 Conclusions In this work, we show that small-world networks can be used to mimic brain networks. Comparing the healthy human matrices and the small-world model with proba- bility of connection about 3.75%. It is evident the prox- imity of the measures indicators of the graphs, such as the average path length that presented 97% of proxim- ity with the result of the human matrix. We find a relation of construction among the vari- ables associated with the transmission of information in the network, such as the transitivity (0.57813 for the healthy brain and 0.5386 for the small-world net- work), the assortivity (0.08151 forthe healthy brain and 0.08829 for the small-world network), the eccentricity (3.66667 for the healthy brain and 3.51282 for the small- world network), and the modularity (0.45157 for the healthy brain and 0.48891 for the small-world network), whose values are very close to each other. In all four measures, we obtained errors (ε) smaller than 10% in the measurements up or down. This characteristic was verified in the variation of the transitivity and the length of the average path as a function of the probability. In both cases, the healthy human network and the hu- man network for the Alzheimer's brain were within the simulated region for a small-world sample, thus indi- cating a close linkage probability of 3.75%. Exactly a small-world model in this region, for an equivalent as- sertiveness value, have very similar graph properties thus demonstrating that the human network (diseased or not) behave as a small-world network. We verify that the healthy brain can be mimicked by networks with small-world properties. The network indicators of the Alzheimer's brain are almost identical with the small-world network, except the assortivity. Therefore, the assortivity could be a diagnostic tool to identify Alzheimer's brain. Acknowledgements We wish to thank the Brazilian gov- ernment agencies: Funda¸cao Arauc´aria, CNPq (420699/2018- 0, 407543/2018-0), FAPESP (2015/50122-0, 2018/03211-6), and CAPES for partial financial support. Bassett DS and Bullmore E (2006) Small-World Brain Networks. The Neuroscientist 12:512-523. Bettencourt LMA, Stephens GJ, Ham MI and Gross GW (2007) Functional structure of cortical neuronal networks grown in vitro. Phys. Rev. E 75:021915. Bressler SL (1995) Large-scale cortical networks and cognition. Brain Res. Rev. 20:288-304. Bullmore ET and Bassett DS (2011) Brain graphs: graphical models of the human brain connectome. Annu. Rev. Clin. Psycho. 7:113-140. Bullmore E and Sporns O (2012) The economy of brain network organization. Nat. Rev. Neurosci. 13:336- 349. Chen Z1, He Y, Rosa-Neto P, Germann J and Evans AC (2008) Revealing modular architecture of human brain structural networks by using cortical thickness from MRI. Cereb. Cortex 18:2374-2381. Cochran WG (1977). Sampling Techniques. John Wiley & Sons. ISBN-10:047116240X. Courchesne E and Pierce K (2005) Why the frontal cortex in autism might be talking only to itself: local over-connectivity but long-distance disconnec- tion. Curr. Opin. Neurobiol. 15:225230. Cronbach LJ (1951) Psychometrika 16:297. https://doi.org/10.1007/BF02310555 Cvetkovi´c D, Rowlinson P and Simi´c S (1997) Eigenspaces of graphs. Cambridge University Press, Cambridge. Ferrari FAS, Viana RL, Reis AS, Iarosz KC, Caldas IL and Batista AM (2018) A network of networks model to study phase synchronization using struc- tural connection matrix of human brain. Physica A 496:162170. Foster JG, Foster DV, Grassberger P and Paczuski M (2010) Edge direction and the structure of networks. P. Natl. Acad. Sci. USA 107:10815-10820. Friston KJ and Frith CD (1995) Schizophrenia: a dis- connection syndrome? Clin. Neurosci. 3:89-97. Gone G, He Y, Conhca L, Lebel C, Grossa DW, Evans AC and Beaulieu C (2009) Mapping anatomical con- nectivity patterns of human cerebral cortex using in vivo diffusion tensor imaging tractography. Cereb. Cortex 19: 524-536. Grace JB (2006) Structural Equation Modeling and Natural Systems. Cambridge, UK: Cambridge Uni- versity Press. Gray V (2017) Principal Component Analysis: Meth- ods, Applications, and Technology. Hauppauge, New York: Nova Science Publishers, Inc (Mathematics Re- search Developments). 10 Jos´e C. P. Coninck1 et al. He Y, Chen Z and Evans A (2008) Structural Insights into Aberrant Topological Patterns of Large-Scale Cortical Networks in Alzheimers Disease. J. Neurosci. 28:4756-4766. He Y, Dagher A, Chen Z, Charil A, Zijdenbos A, Wors- ley K and Evans A (2009) Impaired small-world effi- ciency in structural cortical networks in multiple scle- rosis associated with white matter lesion load. Brain 132:3366-3379. van den Heuvel MP and Yeo BTT (2017) A Spotlight on Bridging Microscale and Macroscale Human Brain Architecture. Neuron 93:1248-1251. Hilgetag CC, Burns GA, O'Neil MA, Scannell JW and Young MP (2000) Anatomical connectivity defines the organization of clusters of cortical areas in the macaque monkey and the cat. Philos. Trans. R. Soc. Lond. B 355: 91-110. Humphries MD and Gurney K (2008) Network Small- World-Ness: A Quantitative Method for Determin- ing Canonical Network Equivalence. Plos One 3: e0002051. Jhonson RA and Wischern DW (1998) Applied Multi- variate Statistical Analysis. Pretice Hall, Upper Sad- dle River, New Jersey 4th ed. ISBN 0-13-834194-X Kaiser M (2011) A tutorial in connectome analysis: Topological and spatial features of brain networks. NeuroImage 57:892-907. Lameu EL, Borges FS, Borges RR, Iarosz KC, Caldas IL, Batista AM, Viana RL and Kurths J (2016) Sup- pression of phase synchronisation in network based on cat's brain. Chaos 26:043107. Lo CY, Wang PN, Chou KH, Wang J, He Y and CP Lin (2010) Diffusion tensor tractography reveals ab- normal topological organization in structural cortical networks in Alzheimers disease. J. Neurosc. 30:16876- 16885. Iturria-Medina Y, Canales-Rodr´ıguez EJ, Melie- Garc´ıa L, Vald´es-Hern´andez PA, Mart´ınez-Montes E, Alem´an-G´omez Y and S´anchez-Bornot JM. (2007) Characterizing brain anatomical connections using diffusion weighted MRI and graph theory. Neuroim- age 36:645-660. Iturria-Medina Y, Sotero RC, Canales-Rodr´ıguez EJ, Alem´an-G´omez Y, Melie-Garc´ıa L (2008) Studying the human brain anatomical network via diffusion- weighted MRI and Graph Theory. Neuroimage 40:1064-1076. Kaiser HF (1974) Psychometrika 39:31. https://doi.org/10.1007/BF02291575 Mori S, Crain BJ, Chacko VP and van Zijl PC (1999) Three-dimensional tracking of axonal projections in the brain by magnetic resonance imaging. Ann. Neu- rol. 45: 265269. Maruyama G (1998) Basics of Structural Equation Modeling. Thousand Oaks, Calif: SAGE Publica- tions, Inc. Newman MEJ (2000) Models of the Small World. J. Stat. Phys. 101:819-841. Newman MEJ (2002) Assortative Mixing in Networks. Phys. Rev. Lett. 89: 208701. Newman MEJ (2003) The Structure and Function of Complex Networks. SIAM Rev. 45:167-256. Newman MEJ and Girvan M (2003) Finding and eval- uating community structure in networks. Phys. Rev. E 026113:1-15. Newman MEJ (2006) Modularity and community structure in networks. P. Natl. Acad. Sci. USA 103:8577-8582. Opsahl T and Panzarasa P (2009) Clustering in weighted networks. Soc. Networks 31:155-163. Opsahl T, Agneessens F and Skvoretz J (2010) Node centrality in weighted networks: Generalizing degree and shortest paths. Soc. Networks 32:245-251. Pol HH and Bullmore E (2013) Neural networks in psy- chiatry. Eur. Neuropsychopharmacol. 23:1-6. R Core Team (2018) R: A Language and Envi- ronment for Statistical Computing. R Foundation for Statistical Computing. Vienna, Austria. 2008. URL:¡http://www.R-project.org/¿ Reijneveld JC, Ponten SC, Berendse HW and Stam CJ (2007) The application of graph theoretical analysis to complex networks in the brain. Clin. Neurophysiol. 118:2317-2331. Rubinov M and Sporns O (2010) Complex network measures of brain connectivity: uses and interpreta- tions. Neuroimage 52:1059-1069. Sporns O and Zwi JD (2004) The small world of the cerebral cortex. Neuroinformatics 2:145-162. Sporns O, Tononi G and Kotter R (2005) The Human Connectome: A Structural Description of the Human Brain. PLOS Comput. Biol. 1:e42. Sporns O (2013) Structure and function of complex brain networks. Dialogues Clin. Neurosci. 15:247-262. Schank T and Wagner D (2005) Approximating Clustering-Coefficient and Transitivity. J. Graph Al- gorithms Appl. 9:265-275. Stam CJ, Jones BF, Nolte G, Breakspear M and Schel- tens P (2007) Small-World Networks and Functional Connectivity in Alzheimers Disease. Cereb. Cortex 17:92-99. Stam CJ and Reijneveld JC (2007) Graph theoretical analysis of complex networks in the brain. Nonlinear Biomed. Phys. 1:3. Stam CJ and van Straaten EC (2012) The organization of physiological brain networks. Clin. Neurophysiol. 123: 1067-1087. Network properties of healthy and Alzheimer's brains 11 Stam CJ (2014) Modern network science of neurological disorders. Nat. Rev. Neurosci. 15:683-695. Stoop R, Saase V, Wagner C, Stoop B and Stoop R (2013) Beyond Scale-Free Small-World Networks: Cortical Columns for Quick Brains. Phys. Rev. Lett. 110:108105. Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O,Delcroix N, Mazoyer B and Joliot M (2002) Automated anatomical labeling of activa- tions in SPM using a macroscopic anatomical par- cellation of the MNI MRI single-subject brain. Neu- roimage 15:273289. Varshney LR, Chen BL, Paniagua E, Hall DH and Chklovskii DB (2011) Structural Properties of the Caenorhabditis Elegans Neuronal Network. PLOS Comp. Biol. 7:e1001066. Watts DJ and Strogatz SH (1998) Collective dynamics of small-world networks. Nature 393:440-442. Weisstein EW (2019) Adjacency Matrix. From MathWorld -- A Wolfram Web Resource. http:// mathworld.wolfram.com/AdjacencyMatrix.html White JG, Southgate E, Thomson JN and Brenner S (1986) The structure of the nervous system of the ne- matode Caenorhabditis Elegans. Philos. Trans. Royal Trans. R. Soc. Lond. B 314:1-340. Yao Z, Zhang Y, Lin L, Zhou Y, Xu C and Jiang T (2010) Abnormal Cortical Networks in Mild Cog- nitive Impairment and Alzheimers Disease. PLOS Comp. Bio. 6:e1001006. Zalesky A, Fornito A, Seal ML, Cocchi L, Westin CF, Bullmore ET, Egan GF and Pantelis C (2011) Dis- rupted Axonal Fiber Connectivity in Schizophrenia. Biol. Psychiatry 69:80-89.
1909.09004
1
1909
2019-09-19T14:02:38
Classification of Open and Closed Convex Codes on Five Neurons
[ "q-bio.NC", "math.CO" ]
Neural codes, represented as collections of binary strings, encode neural activity and show relationships among stimuli. Certain neurons, called place cells, have been shown experimentally to fire in convex regions in space. A natural question to ask is: Which neural codes can arise as intersection patterns of convex sets? While past research has established several criteria, complete conditions for convexity are not yet known for codes with more than four neurons. We classify all neural codes with five neurons as convex/non-convex codes. Furthermore, we investigate which of these codes can be represented by open versus closed convex sets. Interestingly, we find a code which is an open but not closed convex code and demonstrate a minimal example for this phenomenon.
q-bio.NC
q-bio
Classification of Open and Closed Convex Codes on Five Neurons Sarah Ayman Goldrup∗1 and Kaitlyn Phillipson†1 1Department of Mathematics, St. Edward's University September 20, 2019 Abstract Neural codes, represented as collections of binary strings, encode neural activity and show relationships among stimuli. Certain neurons, called place cells, have been shown experimen- tally to fire in convex regions in space. A natural question to ask is: Which neural codes can arise as intersection patterns of convex sets? While past research has established several criteria, complete conditions for convexity are not yet known for codes with more than four neurons. We classify all neural codes with five neurons as convex/non-convex codes. Further- more, we investigate which of these codes can be represented by open versus closed convex sets. Interestingly, we find a code which is an open but not closed convex code and demonstrate a minimal example for this phenomenon. 1 Introduction Understanding neural firings patterns and studying what they represent is an important problem in neuroscience. Neural activity can be modeled via neural codes, which are binary patterns repre- senting the recorded neural activity. Neural codes show relationships between stimuli, like distance between locations in an environment. Through neural codes, the brain is able to characterize and map the physical world. In 1971, O'Keefe discovered place cells in the hippocampus, which is a part of the brain that processes and stores memories and is involved in navigation. Place cells are a special type of neuron that form internal maps of the external world. O'Keefe found that place cells exhibited high firing rate when the rat was in a specific area in space, called the neuron's place field. Through experimental results, it was shown that place fields were approximately convex regions of the space. O'Keefe was awarded a shared Nobel Prize in Medicine in 2014 for this work (see [6] for more detail). Some mathematical questions that arise are: Given a neural code, can it arise from a collection of convex open sets? Can we find criteria to determine whether a neural code is convex by its combinatorial structure alone? Furthermore, if a neural code is a convex code, what is the minimal ∗[email protected] †[email protected] 1 dimension needed to represent the code geometrically? While past results gave necessary and sufficient criteria for a code to be convex, these conditions are incomplete for codes with five or more neurons. In this paper, we completely classify all neural codes on five neurons which are open convex, and give partial results for closed convex codes. In Section 2, we review past results on neural codes and provide concise definitions. In Section 3, we give a catalog of all open convex codes on five neurons which were not classified by prior results. In Section 4, we give a new definition for codes which are open convex but not closed convex (called unstable codes), and prove that three codes on five neurons are unstable. We summarize and give open questions in Section 5. 2 Background and previous results We review some notation and definitions pertaining to this problem; see [4] for more detail. A codeword on n neurons is a string of 0's and 1's of length n, where 1 denotes neural activity and 0 denotes silence. We can also write the codeword σ as a subset of active neurons σ ⊂ [n] := {1, 2, . . . , n}. Both notations will be used interchangeably. A neural code on n neurons is a collection of codewords C ⊂ 2[n]. For computational convenience, we will always assume that the "silent" codeword 000··· 0(= ∅) is in C. Given open sets U = {U1, ..., Un} ⊂ Rd, the code of the cover is the neural code defined as: (cid:91) (cid:92) i∈σ C(U) = {σ ⊆ [n] : Ui \ Uj (cid:54)= ∅} j∈[n]\σ Each codeword in C(U) corresponds to intersections of open sets in U which are not covered by other sets in U. If a neural code C = C(U), and U = {U1, ..., Un} is a cover with each Ui a convex subset of Rd, then C is a convex code and U = {U1, ..., Un} is a geometric realization of the code C. For example, the code C1 = {∅, 1, 2, 12} is a convex code because it can be realized as a collection of convex open sets (see Figure 1). The minimal embedding dimension is the smallest dimension d = d(C) such that the neural code C is realizable as a convex code in Rd. Note that though C1 is drawn in R2, d(C1) = 1. Figure 1: An open convex realization of code C1 For a code C, we can investigate its intersection structure by constructing its simplicial complex : ∆(C) := {σ ⊆ [n] : σ ⊆ c for some c ∈ C}. This is the smallest abstract simplicial complex1 which contains all elements of C. Elements of 1A collection K of subsets of a finite set X is an abstract simplicial complex if, for every α ∈ K and β ⊂ α, then β ∈ K. 2 U1U2X ∆(C) that are maximal under inclusion are maximal codewords (also called facets). A code C is max intersection-complete if it contains all intersections of maximal codewords in C. For a face (or codeword) σ ∈ ∆, the link of σ in ∆ is the simplicial complex: Lk∆(σ) = {w ∈ ∆ : σ ∩ w = ∅, σ ∪ w ∈ ∆}. One approach to deciding if a code is not convex is to determine whether a code has an obstruc- tion to convexity due to a topological inconsistency in the intersections of its codewords. This kind of obstruction is called a local obstruction, and the process to find these obstructions is given in [3]. We summarize their findings here. A simplicial complex is contractible if its geometric realization is contractible2. For a given simplicial complex ∆, we let Cmin(∆) = {σ ∈ ∆ : Lk∆(σ) is non-contractible} ∪ {∅}. It was shown in [3] that if C is a any code with simplicial complex ∆, C has no local obstructions if and only if Cmin(∆) ⊆ C. Moreover, they showed that every nonempty element of Cmin(∆) is an intersection of facets of ∆. Thus, we can consider the non-empty elements of Cmin(∆) as a collection of the maximal codewords (or facets) of ∆ and non-maximal codewords σ such that Lk∆(σ) is non-contractible. We will call these non-maximal codewords the mandatory codewords for ∆, since any convex code with simplicial complex ∆ must contain these codewords. A complete condition for the convexity of a neural code is still unknown; we summarize here the known results. Proposition 2.1. For a neural code C: 1. If C is max intersection-complete, then C is convex. 2. If C is convex, then C has no local obstructions. Part 1 of Proposition 2.1 is due to [2], while Part 2 is due to [3]. Note that Part 2 implies that if C is convex, Cmin(∆) ⊆ C. The converses of Part 1 and Part 2 of Proposition 2.1 hold for n ≤ 4 (see [3]); however, these statements fail for n = 5. An example of a convex code that is not max intersection-complete is the code C1 in Table 1. An example of a non-convex code that has no local obstructions was found in [5], which is code C4 in Table 1. Thus, the classification of which max intersection-incomplete codes with no local obstructions are actually convex remains an open problem. As a step in this direction, we investigate all codes with five neurons that are max intersection-incomplete with no local obstructions. This problem has also been investigated independently in [7]. 3 Classification of open convex codes on five neurons Given a simplicial complex ∆, let Cmin := Cmin(∆) be the minimal code for ∆ defined in Section 2. It was shown in [2] that open convex codes exhibit monotonicity in the following sense: if C is an open convex code, and D is a code such that C ⊆ D ⊆ ∆(C), then D is also open convex. Therefore, we define a simplicial complex ∆ to be convex minimal if the corresponding minimal code of ∆, Cmin(∆) is open convex. If a simplicial complex ∆ is convex minimal, then all codes C with ∆(C) = ∆ and no local obstructions are convex. 2A set is contractible if it is homotopy equivalent to a point. 3 In [5], all unique simplicial complexes on 5 vertices were computed, as well as the maximal facets and the corresponding mandatory codewords for each simplicical complex. It was found that of the 157 unique simplicial complexes, for 22 of these codes, the set of mandatory codewords did not contain all possible intersections of facets. Thus, the minimal code of each of these simplicial com- plexes is a max intersection-incomplete code with no local obstructions, which cannot be classified by Proposition 2.1. In Table 1, we classify these 22 max intersection-incomplete codes with no local obstructions. For each simplicial complex ∆, we list the maximal codewords, mandatory codewords, and non- mandatory intersections of maximal codewords, and describe the convexity of Cmin(∆). We enu- merate the list of minimal codes using the designations Cm, where m = 1, . . . , 22. Columns 2, 3, and 4, were computed in [5]. Of the 22 simplicial complexes with this property, only one minimal code does not have a convex realization: C4 was proved to be non-convex in [5]. Convex realizations for each code, except C4, are given in B. Interestingly, only one code on the list has minimal dimension 3, which is C22, the construction of which is given in A. The remainder of the codes have minimal dimension 1 or 2. Combining our findings in Table 1 with the property of monotonicity for open convex codes, we immediately obtain the following theorem. Theorem 3.1. Let C be a code on five neurons for which ∆(C) is not isomorphic to ∆(C4). Then C is open convex if and only if C has no local obstructions. These results can also be summarized by noting the sparsity of the codes. A code C is k-sparse if σ ≤ k for all σ ∈ C. The catalog of codes in Table 1 gives the following result: Theorem 3.2. All 3-sparse codes with no local obstructions on five neurons are open convex. This gives affirmative evidence to the question introduced in [1]: Is every 3-sparse code with no local obstructions convex? If a counterexample exists, it must have at least 6 neurons. 4 Closed convex codes and unstable codes 4.1 Background on closed codes Section 3 gave a complete classification for open convex codes on five neurons. However, the question of classifying closed convex codes, which can be realized as a collection of closed convex sets in Rd, is not fully answered. In [2], it was shown that the results of Proposition 2.1 can be extended to closed convex codes as well: Max intersection-complete codes are both open and closed convex, and if a code is closed convex, then it has no local obstructions. However, it is unknown whether the monotonicity condition holds for closed convex codes. Note that though C4 is not open convex, it has been shown to be closed convex (see [2]). Initially, it was conjectured that every open convex code is also a closed convex code; however, this is not the case for neural codes with n > 4. The authors of [2] showed that the neural code D = {123, 126, 156, 456, 345, 234, 12, 16, 56, 45, 34, 23,∅} (with facets given in bold) is an open convex code but not a closed convex code with six neurons. Intuitively, unstable codes arise when the neural code of the open convex realization would change drastically if the open sets were perturbed slightly. It was unknown whether D was the minimal example of an unstable code with respect to the number of neurons. Here, we show that five neurons is the minimal case where unstable codes occur. We define unstable codes as codes that are open convex but not closed convex. 4 Maximal Mandatory Cmin(∆) codewords (facets) codewords Non- mandatory intersections of facets 12, 14 13, 34, 1, 2 34, 35, 1, 2 12, 14, 23, 24, 45, 2, 4 12, 14, 23, 24, 2 12, 15, 23, 4 24, 45, 1, 2, 3 1 3 3 1 1, 4 1, 2 4 Classification (see B) convex minimal convex minimal convex minimal Not open con- vex (see [5]) convex minimal convex minimal convex minimal C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C18 C19 C20 C21 C22 124, 123, 123, 124, 145 134, 135, 234, 12 134, 235, 345, 12 2345, 145 123, 124, 145, 234 123, 125, 145, 234 145, 234, 245, 12, 13 135, 145, 235, 12, 34 134, 135, 145, 234, 12 134, 135, 234, 245, 12 134, 135, 245, 345, 12 123, 124, 125, 134, 345 123, 124, 135, 145, 234 123, 124, 125, 145, 234 123, 125, 145, 234, 345 145, 235, 345, 12, 13, 24 145, 234, 235, 245, 12, 13 134, 135, 145, 234, 235, 12 134, 135, 234, 245, 345, 12 123, 124, 125, 135, 145, 234 123, 124, 125, 145, 234, 345 123, 124, 125, 135, 145, 234, 235 15, 35, 1, 2, 3, 4 5 convex minimal 13, 14, 15, 34, 1, 2 3, 4 convex minimal 13, 24, 34, 1, 2, 5 3, 4 convex minimal 13, 34, 35, 45, 1, 2, 3 12, 13, 14, 34, 1, 5 12, 13, 14, 15, 23, 24, 1, 2 12, 14, 15, 23, 24, 1, 2 12, 15, 23, 34, 45 4, 5 3, 4 3, 4 4 convex minimal convex minimal convex minimal convex minimal 1, 2, 3, 4, 5 convex minimal 35, 45, 1, 2, 3, 4 5 convex minimal 23, 24, 25, 45, 1, 2, 3 13, 14, 15, 23, 34, 35, 1, 2, 3 13, 24, 34, 35, 45, 1, 2, 3, 4 12, 13, 14, 15, 23, 24, 1, 2 12, 14, 15, 23, 24, 45, 45, 1, 2, 4 12, 13, 14, 15, 23, 24, 25, 35, 1, 2, 3, 5 4, 5 4, 5 5 3, 4 3, 5 4 convex minimal convex minimal convex minimal convex minimal convex minimal convex minimal with d(C22) = 3 Table 1: Classification of max intersection-incomplete codes with no local obstructions on five neurons 5 4.2 Minimal case of unstable codes Through inspection, we can see that all the open convex realizations for the codes in Table 1 can be viewed as closed convex realizations except for three cases: C6, C10, and C15. The following result shows that these three codes are indeed not closed convex. Theorem 4.1. The following codes are open convex but not closed convex: C6 = {125, 234, 145, 123, 4, 23, 15, 12,∅} C10 = {134, 245, 234, 135, 12, 1, 5, 34, 13, 2, 24,∅} C15 = {145, 125, 123, 234, 345, 23, 15, 45, 34, 12,∅} Proof. Codes C6, C10, and C15 have open convex realizations given in B. We will employ the technique used in [2] to show that C6 is not closed convex. Suppose there is a closed convex cover U = {Ui}5 i=1 in Rd for C6. For σ ⊆ [5], we let Uσ = ∩i∈σUi. We can pick distinct points x234 ∈ U234, and x145 ∈ U145. Let M be the line segment connecting x145 to x234. Pick x123 ∈ U123 so that for every a ∈ U123, we have dist(a, M ) ≥ dist(x123, M ), i.e., dist(x123, M ) is the minimal distance to M . Let L1 = x123x145. L1 ⊂ U1 since U1 is convex, and U1 ⊂ U2 ∪ U5, since whenever neuron 1 appears in a codeword, it will also appear with either 2 or 5. Together, these imply L1 ⊂ U2 ∪ U5. Since L1 is connected and the sets U2 ∩ L1 and U5 ∩ L1 are closed and nonempty, U2 ∩ U5 ∩ L1 ⊂ U125 is nonempty and there is a point x125 ∈ U125 ∩ L1 that is on the line segment L1 (see Figure 2). Figure 2: Proof of Theorem 4.1 Let L2 = x123x234. U23 is convex so L2 ⊂ U23. Since U23 ∩ L2 ⊂ U23 is nonempty, there is a point x23 ∈ U23 ∩ L2 that is on line segment L2. Let L3 = x23x125. By convexity, L3 ⊂ U2 and since U2 ⊂ U1 ∪ U3, then L3 ⊂ U1 ∪ U3. Since L3 is connected and the sets U1 ∩ L3 and U3 ∩ L3 are closed and nonempty, U1 ∩ U3 ∩ L3 ⊂ U123 is nonempty, and there is a point y123 ∈ U123 ∩ L3 that is on the line segment L3. We see that the point y123 lies in the interior of the closed triangle ∆(x123, x145, x234). Therefore, dist(y123, M ) < dist(x123, M ) which is a contradiction. This implies that C6 cannot be realized as a collection of closed convex sets, and thus it is not a closed convex code. The proof for C10 is similar, using x245 and x135 as starting points, and building x234 of minimal distance to x245x135. For C15, we can use x145 and x345 as starting points, and choose x123 of minimal distance to x145x345. The details are left to the reader. Since the condition of being max intersection-complete is equivalent to being closed and open convex for n ≤ 4 neurons (see [2]), we have the following result: Corollary 4.2. If C is an unstable neural code on n neurons, then n ≥ 5. This bound is tight. 6 x234x145x123x23x125y123L2L1L3M 4.3 Discussion on unstable codes We end this section by discussing some observations of the previous examples of unstable codes. We note that D, C6, C10, and C15 are all examples of 3-sparse codes with at least four maximal codewords. When computing the simplicial complexes of D, C6, C10, and C15, we see that they all have at least two distinct non-mandatory intersections of facets. In the proof of Theorem 4.1, we used the absence of these codewords in the code in our construction: for a missing non-mandatory intersection of facets σ, Uσ is a set completely contained in the union of other sets in the cover, which was used to construct a new point on each line. Note that in the in the open convex realizations of each code D, C6, C10, and C15, there appears to be lines slicing the plane and meeting in a common point. Moreover, the boundary points of the open convex sets overlap. Some differences to note are that D has 6 neurons and 6 maximal codewords, C6 has 5 neurons and 4 maximal codewords, and C10 and C15 both have 5 neurons and 5 maximal codewords. C15 and ∆(D) have more than two non-mandatory intersections of facets. From these observations, and from the similarities in the proof technique used in [2] and 4.2, we make the following conjecture: Conjecture 4.3. Let C be a max intersection-incomplete open convex code, where ∆(C) has at least two non-mandatory intersections of facets not contained in C. Suppose C has at least 3 maximal codewords M 1, M 2, M 3, and there is σ ⊂ M 1 with σ ∈ C such that σ ∩ M 2 (cid:54)∈ C. Then C is not a closed convex code. 5 Conclusion and future research In this paper, we showed that all 3-sparse neural codes with no local obstructions on five neurons are open convex. This result also shows that every simplicial complex on five neurons is convex minimal except ∆(C4). Furthermore, we showed that unstable neural codes can only occur for binary codes with no fewer than 5 neurons, and thus showed which of the minimal codes are both open and closed convex code and which are unstable codes. Besides Conjecture 4.3, some future problems that can be investigated are: Does the monotonicity condition hold for closed convex codes (in particular, for the closed convex codes from Table 1)? Which codewords need to be added to C6, C10, or C15 to make each code closed convex? Are there examples of unstable codes which are not 3-sparse? 6 Acknowledgments Over summer 2017, SAG was supported by the Dr. M. Jean McKemie Endowed Student/Faculty Fund for Innovative Mathematics Summer Scholarship, and KP was supported by the Presidential Excellence Grant of St. Edward's University. We thank Zvi Rosen for his contribution to the realization of C10, Luis Garc´ıa Puente for discussion on this topic, and Anne Shiu for helpful comments on previous drafts. References [1] Aaron Chen, Florian Frick, and Anne Shiu. Neural codes, decidability, and a new local obstruc- tion to convexity. SIAM Journal on Applied Algebra and Geometry, 3:44 -- 66, 2019. 7 [2] Joshua Cruz, Chad Giusti, Vladimir Itskov, and Bill Kronholm. On open and closed convex codes. Discrete and Combinatorial Geometry, 61:247 -- 270, 2019. [3] Carina Curto, Elizabeth Gross, Jack Jeffries, Katie Morrison, Mohamed Omar, Zvi Rosen, Anne Shiu, and Nora Youngs. What makes a neural code convex? SIAM Journal on Applied Algebra and Geometry, 1:222 -- 238, 2017. [4] Carina Curto, Vladimir Itskov, Alan Veliz-Cuba, and Nora Youngs. The neural ring: an algebraic tool for analyzing the intrinic structure of neural codes. Bulletin of Mathematical Biology, 75:1571 -- 1611, 2013. [5] Caitlin Lienkaemper, Anne Shiu, and Zev Woodstock. Obstructions to convexity in neural codes. Advances in Applied Mathematics, 85:31 -- 59, 2017. [6] John O'Keefe and Jonathan Dostrovsky. The hippocampus as a spatial map. preliminary evi- dence from unit activity in the freely-moving rat. Brain Research, 34:171 -- 175, 1971. [7] Rutger Yager. On the open convexity of neural codes with five neurons. Master's thesis, Sam Houston State University, 2018. A Construction of C22 Proposition A.1. C22 = {145, 124, 135, 235, 125, 123, 234, 12, 13, 14, 15, 23, 24, 25, 35, 1, 2, 3, 5,∅} is convex with minimal embedding dimension 3. Proof. C22 has Helly Dimension 3 (see [3]) by the hollow simplex formed by 135, 235, 125, 123, which implies that the minimal embedding dimension is at least 3. This code is indeed realizable as a convex code in three dimensions via the following construction: Construct a tetrahedron by taking the convex hull of the points {(0, 0, 0), (4, 0, 0), (0, 4, 0), (0, 0, 4)}. Create open sets in R3 by extending each face into the interior of the tetrahedron with the following thicknesses: U1 is the face on the xz-plane with thickness 1. U2 is the face on the yz-plane of thickness 3/4. U3 is the face on the xy-plane of thickness 1/2. U5 is the face on the diagonal of thickness 1/4. This construction is shown in Figure 4. These sets give all codewords involving 1, 2, 3, 5 as desired. Consider the points p3 = (1, 0, 3) and p1 = (0, 2, 0). Under a slight perturbation, p3 ∈ U1 ∪ U5 \ (U2 ∪ U3) and p1 ∈ U2 ∪ U 3 \ (U1 ∪ U5). The goal is to construct the set U4 using the line segment p1p3. We claim the line segment L = p1p3 is completely contained in U1 ∪ U2. Parametrize the line segment L = L(t) as: x = t y = 2 − 2t z = 3t 8 Figure 3: Covering of L(t) Note that L(0) = p1 and L(1) = p3. It is a straightforward exercise to check that the rest of the line segment is covered as shown in Figure 3. For  sufficiently small, the open cylinder C(, L(t)) with radius  and center L(t) will be com- pletely contained in U1 ∪ U2. Thus, if we let U4 = C(, L(t)), we have our desired realization of the neural code C22. Figure 4: Convex Realization of code C22 B Geometric realizations of n = 5 open convex codes 9 t=0t=1U3U2U1U5xyz C1 C3 C6 C2 C5 C7 10 1451412412123∅1122135131343423434345351342351212∅145141241224223234123234414515125123231212313234145122454524∅∅ C8 C10 C12 C9 C11 C13 11 112223533513515145434∅∅3423421211341314115145135∅12121324245135534134234∅2452145345351341333434135312∅1212515345341341413124121231∅14513515114131231241222323424 C14 C15 C16 C17 C18 C19 12 121512312123125124141453423423434345451451512512123231211∅∅∅1451333454542422353523235252452234242323∅21312114545∅1451513513141343433523523234343521213∅11352453545345341324341342341212∅ C20 C21 C22 13 1351312323234241241414515125151112212515145453451141241242421232323434xyz
1902.08599
2
1902
2019-03-08T12:48:01
Slow Waves Analysis Pipeline for extracting the Features of the Bi-Modality from the Cerebral Cortex of Anesthetized Mice
[ "q-bio.NC" ]
Cortical slow oscillations are an emergent property of the cortical network, a hallmark of low complexity brain states like sleep, and represent a default activity pattern. Here, we present a methodological approach for quantifying the spatial and temporal properties of this emergent activity. We improved and enriched a robust analysis procedure that has already been successfully applied to both in vitro and in vivo data acquisitions. We tested the new tools of the methodology by analyzing the electrocorticography (ECoG) traces recorded from a custom 32-channel multi-electrode array in wild-type isoflurane-anesthetized mice. The enhanced analysis pipeline, named SWAP (Slow Waves Analysis Pipeline), detects Up and Down states, enables the characterization of the spatial dependency of their statistical properties, and supports the comparison of different subjects. The SWAP is implemented in a data-independent way, allowing its application to other data sets (acquired from different subjects, or with different recording tools), as well as to the outcome of numerical simulations. By using SWAP, we report statistically significant differences in the observed slow oscillations (SO) across cortical areas and cortical sites. Computing cortical maps by interpolating the features of SO acquired at the electrode positions, we give evidence of gradients at the global scale along an oblique axis directed from fronto-lateral towards occipito-medial regions, further highlighting some heterogeneity within cortical areas. The results obtained on spatial characterization of slow oscillations will be essential for producing data-driven brain simulations and for triggering a discussion on the role of, and the interplay between, the different regions in the cortex, improving our understanding of the mechanisms of generation and propagation of delta rhythms and, more generally, of cortical properties.
q-bio.NC
q-bio
Article Slow Waves Analysis Pipeline for extracting the Features of the Bi-Modality from the Cerebral Cortex of Anesthetized Mice Giulia De Bonis 1, Miguel Dasilva 2, Antonio Pazienti 3 Maria V. Sanchez-Vives 2, Maurizio Mattia 3 and Pier Stanislao Paolucci 1 1 2 3 * Correspondence: [email protected] Istituto Nazionale di Fisica Nucleare (INFN), Sezione di Roma, Rome, Italy Institut d'Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Barcelona, Spain Istituto Superiore di Sanità (ISS), Rome, Italy February 2019 Abstract: Cortical slow oscillations ((cid:46) 1 Hz) are an emergent property of the cortical network that integrate connectivity and physiological features. This rhythm, highly revealing of the characteristics of the underlying dynamics, is a hallmark of low complexity brain states like sleep, and represents a default activity pattern. Here, we present a methodological approach for quantifying the spatial and temporal properties of this emergent activity. We improved and enriched a robust analysis procedure that has already been successfully applied to both in vitro and in vivo data acquisitions. We tested the new tools of the methodology by analyzing the electrocorticography (ECoG) traces recorded from a custom 32-channel multi-electrode array in wild-type isoflurane-anesthetized mice. The enhanced analysis pipeline, named SWAP (Slow Waves Analysis Pipeline), detects Up and Down states, enables the characterization of the spatial dependency of their statistical properties, and supports the comparison of different subjects. The SWAP is implemented in a data-independent way, allowing its application to other data sets (acquired from different subjects, or with different recording tools), as well as to the outcome of numerical simulations. By using SWAP, we report statistically significant differences in the observed slow oscillations (SO) across cortical areas and cortical sites. Computing cortical maps by interpolating the features of SO acquired at the electrode positions, we give evidence of gradients at the global scale along an oblique axis directed from fronto-lateral towards occipito-medial regions, further highlighting some heterogeneity within cortical areas. The results obtained using SWAP will be essential for producing data-driven brain simulations. A spatial characterization of slow oscillations will also trigger a discussion on the role of, and the interplay between, the different regions in the cortex, improving our understanding of the mechanisms of generation and propagation of delta rhythms and, more generally, of cortical properties. Keywords: Slow-Wave Activity; Analysis Pipeline; Software Tools; Cortical Area; Multi-Electrode Arrays (MEAs); Multi-Unit Activity (MUA); Anesthetized Mice 1. Introduction In a brain manifesting slow-wave activity (SWA), expressed in the cerebral cortex under NREM sleep and deep anesthesia [1], the spiking activity -- both single and multi-unit activity (SUA and MUA respectively) -- appears as a regular sequence of Up (high-rate) and Down (almost quiescent) states. In the past 10 years, an accurate procedure for the analysis of electro-neurophysiological data has been developed, aimed at extracting the MUA from raw recordings, identifying alternating Up and Down states associated with SWA and investigating the features of such rhythmic spatio-temporal patterns of activity propagating along the cortex. The analysis pipeline, implemented in MATLAB1, has been 1 MATLAB R(cid:13), The MathWorks, Inc., Natick, Massachusetts, United States. 9 1 0 2 r a M 8 ] . C N o i b - q [ 2 v 9 9 5 8 0 . 2 0 9 1 : v i X r a 2 of 18 properly refined over time, and applied to several experimental data sets, acquired with different setups both in vitro and in vivo from rodents, ferrets, and monkeys [2 -- 5]. The software procedure has been extensively revised and improved, including some new features implemented in Python2, to be used with a new set of data collected in vivo using a 32-electrode array from the cerebral cortex of 11 deeply isoflurane-anesthetized wild-type mice. The data used in this study was obtained in accordance with the Spanish regulatory laws (BOE-A-2013-6271) which comply with the European Union guidelines on protection of vertebrates used for experimentation (Directive 2010/63/EU of the European Parliament and the Council of 22 September 2010), and the protocol was approved by the Animal Ethics Committee of the University of Barcelona. The multi-electrode array (MEA) employed for acquiring the data is described in Figure 1 and covers several cortical areas ranging from sensory (V1, S1), motor (M1) and association (PtA) cortices [6]. The unfiltered field potential (UFP) is sampled from each electrode at a frequency of 5 kHz and each acquisition session lasts at least 300 s (see Table 1), thus ensuring a fine inspection of the signal in both space and time. The 11 recording sessions are each from a different animal, i.e. 11 independent experiments that collectively represent an extensive data sample covering a wide range of biological and unavoidable physiological variability. lateral ←→ medial Figure 1. Representation of the multi-electrode array (MEA) used for the data acquisition [6]. On the top left, a scheme that indicates the scale of the experiment; the reported dimensions have been adopted to define the reference frame in the SWAP. On the bottom left, the color legend that identifies the cortical areas on the array surface. On the right, the grid of electrodes, with the numbering introduced in the SWAP. In the center, an illustration of the MEA positioned on the mouse cortex; the inspected surface is of the order of 10 mm2. The accurate time-and-space sampling together with the richness of the experimental data have driven the development of new analytical tools that aim to characterize the differences between cortical areas when expressing SWA. In addition, given the unavoidable and physiological variability of the data set, particular attention has been also given to the best strategies to adopt in order to perform a thorough comparison of recordings obtained from the 11 different subjects. The guiding principle when combining data was to keep and use the largest amount of signal, avoiding arbitrary removal of noisy channels or the creation of a subset of "golden" cases. Descriptive statements are accompanied by the assessment of statistical significance of the results, taking into account the multiplicity of the hypotheses under testing; the obtained claims can give hints as to the mechanisms underlying SWA in mammals. 2 Python Software Foundation. Python Language Reference, version 2.7.5. Available at http://www.python.org. Data File Hemisphere Duration of the DAQ Session [s] 01 02 03 07 09 10 14 15 16 17 20 R R L L L L R R R R R 304.102 300.912 427.223 351.701 321.942 309.19 313.714 329.605 312.157 305.967 319.942 3 of 18 Notes • Excluded: ch. 8 (SD-outlier) • Weak Bimodality: ch. 1, 2, 3, 4, 7, 9, 10, 11, 12, 14, 15, 16, • Excluded: ch. 30, 31, 32 (failure in the DAQ) 17, 18, 19, 20, 22, 23, 24, 25, 26, 27, 28, 29 • Discontinuity in ch. 2, 4, 8, 12, 15, 19, 20, 23, 28, 31, 32 • Excluded: ch. 3, 13 (SD-outlier) -- -- -- • Discontinuity in ch. 23 • Weak Bimodality: ch. 2, 3, 17, 20 • Excluded: ch. 3, 9 (SD-outlier) • Discontinuity in ch. 12 • Excluded: ch. 12 (failure in the DAQ) • Excluded: ch. 25 (SD-outlier) • Discontinuity in ch. 14 • Excluded: ch. 4 (SD-outlier); 31 (failure in the DAQ) • Excluded: ch. 1 (negative asymmetry); 12, 25 (SD-outlier) Table 1. Summary of the data acquisition (DAQ) sessions. The name of the data file reflects the date of acquisition. R for right hemisphere, L for left hemisphere. In general, discontinuities are not critical, since the failure corresponds to an interruption of the DAQ for a limited time interval (usually, a couple of discontinuities per channel, lasting from a few hundreds of millisecond to a few seconds); therefore, discontinuities in the DAQ are managed in the SWAP by identifying, for each problematic channel, the time interval at which the data acquisition fails, and removing it from the time sequence of the signal. By contrast, excluded channels are those for which the signal presents several irregularities, usually resulting in a number of identified upward transitions well below the median computed over the full channel set (tagged as "failure in the DAQ"); in addition, SD-outlier channels are also excluded; the tag "negative asymmetry" corresponds to the case of having a strong negative skew. Furthermore, the outcome of the data analysis can be employed to feed the input of a dedicated spiking neural network simulation (as [7 -- 9]), in a data-driven approach of in silico studies of the brain, aimed at computing a less stereotypical and a more accurate reproduction of cerebral rhythms. The analysis pipeline itself, hereafter named Slow Waves Analysis Pipeline, or SWAP, can be adopted to study the output of the simulation, and to define a set of benchmark observables for confronting models and experiments, with the goal of having a reliable and flexible set of tools available for the characterization of the slow-wave signal in a wide set of cases. 4 of 18 The material in this paper is structured as follows. Section 2 offers an overview of the SWAP analysis pipeline, with a collection of results obtained from test-bench data used for illustrating the steps of the procedure; the focus is on the methods implemented for the identification of the Up/Down state alternation in the ECoG traces, on the techniques for "stacking" and comparing different subjects, and on the statistical treatment of data with hypothesis testing. Section 3 is dedicated to the discussion of methods and results, with concluding remarks and suggestions for future research. 2. The Slow Waves Analysis Pipeline (SWAP) The study of SWA expressed by the cortex can be tackled starting from a description of the bimodality (i.e. the alternation of Up and Down states), by defining a comprehensive set of observables suited to illustrating the phenomenon of SO. Once this local information is acquired, the second step (not illustrated here) is the characterization of the activity propagation across the brain surface as a wave with delta rhythm. Focusing on the features of Up and Down states, differences between cortical sites can be emphasized. The analysis procedure consists of two phases. First, each data file is examined separately (Single Experiment), yielding a detailed inspection of the single subject. The results from the different experiments are then combined (inter-session data) to produce Summary Results . Finally, conclusive claims are given with the assessment of statistical significance of the results, taking into account the numerousness of the sample and the multiplicity of the hypotheses under testing. Three different levels of description can be enabled: (I) the channel level, providing information at each recording site; (II) the area level, stacking up the information of all the electrodes located in the same cortical area; and (III) the full-set level, computing average properties and summary statements for the entire cortex portion under study. The levels of description can be applied at the single experiment, or at the inter-session data; outcomes obtained at the different levels of description are complementary and not strictly separated, and can be superimposed on graphical representations. Figure 2 details, in terms of phases of action and levels of description, the logic of the SWAP when facing the issue of investigating the features of Up and Down states. Figure 2. The logic of the SWAP when facing the issue of investigating the features of Up and Down states, in terms of phases of action (the sequence of 3 actions listed on the left) and levels of description (in the upper box). The lower box contains information on the statistical treatment presently implemented. 2.1. The channel-level description of the single experiment: from the raw data to the estimates of MUA and of transition times The channel-level description of the single experiment is obtained with a set of scripts, coded in MATLAB R(cid:13), carrying out a loop over the electrodes in the array to extract the MUA from recordings of the raw signal (UFP). For each channel, the Power Spectral Density (PSD) of the signal is computed 5 of 18 and the MUA is used as an estimate of the firing rate of the neurons around the electrode tip [2,10]. The logarithm of the MUA is evaluated, and the shape of the log(MUA) distribution can be described as a peak, at low-MUA values corresponding to Down states, and a tail, at high-MUA values corresponding to the Up states. The log(MUA) peak is fitted with a Gaussian function, and the parameters of the fit are used to single out Down-state periods from Up-state periods in the MUA time series. Once the MUA time series is tagged as "Up" and "Down" (binary MUA), a detailed study of the features of such states and of the transitions among them -- upward (Down-to-Up) and downward (Up-to-Down) -- can be achieved. The workflow is illustrated in Figure 3. Figure 3. The channel-level description of the single experiment. Purple boxes point out the initial data format and the intermediate outcomes of the workflow (as successive elaborations of the raw data). The black box encompasses the key steps of the process. The procedure is coded in MATLAB R(cid:13), the sequence of actions is illustrated in Figure 4. The figure illustrates the main loop (as discussed in Section 2.1), at the end of which further checks are carried out on average and median values (full-set level description), in order to identify further anomalies or outliers. In more detail, the initialization phase is carried out with the steering file setParamsAndOptions, which takes into account the specificity of the data acquisition (DAQ) sessions, accommodating the frame of reference (the positions of the electrodes in the array) and loading information about the recordings. Some of the settings concern the capability of the algorithm to identify the state transitions, and are fine-tuned following a heuristic approach. The steering file informs what to check in order to perform the analysis pipeline on the given input data. Once the initialization phase is completed, the main loop starts and the flow in the pipeline is as follows (see Figure 4): 1. Read the analog data (raw signal). For the ECoG data used as a benchmark in this study, input-related actions (open/close the input file and read the input data) are performed by the SONLibrary [11], whose functions provide the access to the stored neurophysiological data. 2. Analyze the raw signal in the frequency domain. A moving window of samples is defined for the calculation of the spectrogram, i.e. the spectral content of the raw signal as a function of time; the extent of the moving widow, together with the sampling frequency, determines the frequency band to be examined. Since the frequency band of interest for the estimate of the MUA 6 of 18 is (200− 1500) Hz3, a moving window of 5 ms is adopted4. FFT (Fast Fourier Transform) and PSD (Power Spectral Density) are computed using MATLAB R(cid:13)functions. Then, for each frequency in the band, the median of the PSD is computed, considering the full set of spectrograms, i.e. the collection of moving windows (time steps) that constitute the entire acquisition session. The obtained vector of medians, one value for each frequency, is used as a baseline to normalize the PSD. 3. Evaluate the MUA for each time step as the mean amplitude of the normalized PSD. The MUA is intended as an estimate of the collective firing rate r(t) of neurons located at the electrode position. The natural logarithm of the MUA is computed, since logarithmic mapping is adopted to emphasize the bimodality of the distribution; because of the normalization to the median of the PSD, negative values and positive values of log(MUA) identify a spectral content smaller or larger than the median, respectively [2,10]. 4. Fit the distribution of log(MUA), isolating with a Gaussian function the peak corresponding to the (dominant) regime of low-rate states (Figure 4.A). A simplified description of the bimodality of the SWA in the cortex would require a bimodal fit, and initially the sum of two Gaussian profiles was adopted to describe the shape of the distribution. Conversely, what was discovered after a thorough inspection of a huge number of channels from several different animals (both physiological and pathological subjects at different anesthesia levels) is that, while the Down-state peak is highly stable despite the large variability in the subjects, the content of the log(MUA) histogram corresponding to the high-rate regime (obtained from the total distribution after subtracting the Down-state peak) expresses over a large span of "shapes", and rather than as a "second peak" (with a definite Gaussian profile) it can be generically appointed as a "tail" (Figure 4.A, inset). 5. Set the UD_THRESHOLD, i.e. the level of log(MUA) that defines the separation of the two regimes of the bimodality, Up and Down (UD). This is a crucial step in the pipeline, and several options can be adopted to find the optimal criterion, which can depend on the specificity of the data set and on the scope of the analysis. In general, the choice takes into account general settings of the DAQ, the log(MUA) distribution, and the results of the Gaussian fit on the Down-state peak (of the given channel, or considering average properties of the recording session). At the channel level in the main loop, checks are introduced to monitor the convergence of the Gaussian fit, the content of the histogram of log(MUA) values, and the shape of the distribution (Figure 4.A); alerts are activated to signalize: Weak Bimodality (if the area of the tail is smaller than a given threshold -- currently set at 10% of the total); Positive Skewness (if gamma -- the coefficient of skewness -- of the tail is above 1, γ > 1); Negative Skewness (if γ < −1); Right Peak (if the peak, i.e. the dominant component of log(MUA), is centered at large values, on the right segment of the range, with a tail on the left); Large Threshold (if the selected threshold is larger than the mean of the tail); Few Transitions (if the number of transitions obtained with the selected threshold is below 3, the minimum requisite to isolate at least an Up state and a Down state). Some of these alerts may help in defining the threshold level, others provide an indication of "problematic" channels to be inspected (that may or may not be tagged as outliers and excluded from further processing at the end of the main loop). Finally, some conditions are "blocking", in the sense they require a break in the workflow, as is the case with noisy acquisition channels or spoiled recordings. If no blocking conditions are encountered, meaning that a threshold can be set and Down states and Up states are distinguishable, the following operations are performed: 3 4 The experience suggests that frequencies in the raw signal outside this range cannot be associated with the electro-physiological signal of the MUA; in particular, the highest frequency components reflect the presence of electrical noise introduced by the acquisition system. For the benchmark data acquired at 5 kHz, the time window contains 25 samples. 7 of 18 (a) Convert the MUA into a binary sequence (BinaryMUA) that is 1 or 0 depending on the value of log(MUA) above or below the threshold, resulting in the MUA time sequence being tagged as "Up" or "Down" (Figure 4.B). (b) Label the transitions as upward or downward, and assign the time of transition (Trigger Time) with a cubic interpolation of the waveform to locate the time at which the level of MUA crosses the threshold that separates the two regimes of low and high firing rate activity. (c) Study the features of states and transitions, to check the robustness of the algorithm and the stability of the outcome for the different channels. In particular: i. Histograms of the duration of Down states dDOWN, Up states dUP, UD-cycles dUD. The study of these observables is one of the focuses of the analysis; at the channel-level description of the SWAP, the distributions are superimposed for comparison (Figure 4.C). ii. Raster plots of states and transitions; each transition (Trigger Event) is centered at the trigger time; events can be sorted by their time of occurrence, or by their duration (Figure 4.D-E). iii. Waveform plots, for comparing states and transitions, and plots of the average waveform, obtained considering the full set of transitions. A refined algorithm isolates the transition front, to better investigate the transition dynamics (Figure 4.F-G). As anticipated above, once the main loop execution has yielded a full description at the channel level, a key parameter to be monitored for the validation of the procedure is the stability of the conditions used for the identification of the two states (low-MUA and high-MUA). More precisely, since a requirement for the separability of Up and Down states is the successful fit of the Down-state peak of the log(MUA) with a Gaussian function (Figure 4.A), a similar value for the standard deviation (SD or σ) of the peak across channels is requested, ensuring comparable SO dynamics of the probed cell assemblies. Indeed, Down states are almost quiescent, and the variability of the MUA is mainly due to the acquisition chain, which has to be the same across channels. A stable σ allows the application of the same MUA threshold at each recording channel, meaning a unique definition of the Up states, and thus more reliable profiles of traveling wave-fronts and a more sound description of the SWA as a collective phenomenon. Therefore, the choice of a fixed UD_THRESHOLD is a valid option when the goal of the analysis is to study the dynamics of the propagation of the activity as a slow wave across the cortex. On the other end, this means to decide a key parameter a priori, with the burden to be too conservative (increasing the false negative rate) or too tolerant (admitting a larger number of false positives)5. Conversely, the option of linking the choice of threshold to "internal agents" (e.g. parameters evaluated during the workflow of the pipeline) can be convenient, for reducing the number of free variables, or for anchoring the false positive rate per channel. A dedicated study of the standard deviation σ has been carried out (Section 2.2) on the test-bench data. Finally, channels not fulfilling the stability requirements are excluded from the analysis and marked as outliers. Here again, the decision on the stability requirements (which parameters to focus on, which threshold levels, which weight to assign at the different instances, how to define the outliers) is a key-element of the pipeline and may largely depend on specific features of the recording sessions and on the goals of the data analysis. As discussed above, the SWAP pipeline has set in place alerts, warnings and counters based on parameters considered relevant for the test-bench data, but the elements to be monitored can change if conditions vary. Also configurable are the criteria that define outliers; for results presented here (Table 1), excluded channels are those with the Right Peak and with 5 In general, less conservative settings are preferable, since the control of false positives can be addressed by checking the spatio-temporal correlations of the propagating signal across the electrodes grid. 8 of 18 Time (s) Time (s) Figure 4. (A) Distribution of log(MUA) as normalized histogram of values shifted at the position of the first peak, in order to set at level 0 the average firing rate that corresponds to a Down state, thus associated with a null firing rate. The blue curve is the Gaussian fit of the first peak, the µ position is marked by the blue vertical dashed line. The red dotted line is the "tail", obtained after subtracting the Gaussian fit from the total distribution (see inset; vertical dashed lines: mode in magenta, median in cyan, mean in black). In this example channel (Channel 1, file 01), the tail distribution is far from being Gaussian; the area of the tail is ∼ 27%, therefore expressing a clear bimodality, coherently with the value of the skewness of the distribution, γ = 1.97. (B) log(MUA) (in black), BinaryMUA (in red) and raw signal (UFP, in blue). (C) Histograms of the duration of Up states (red), of Down states (blue) and of the full cycles (UD cycle, in green); dashed lines, mean values. (D,E) Raster plot of upward (Down-to-Up) and downward (Up-to-Down) transitions, sorted by state duration. (F,G) Waveforms (WFs) of upward and downward transitions, respectively. For each plot, the first 5 transitions are shown, centered around the transition time (Trigger Time); the average transition, computed considering the full set of n waveforms, is superimposed (black); the shaded area identifies the profile of the waveform at ±1 standard deviation (SD) off the mean. The inset shows the average transition front; the shaded area identifies the profile of the waveform at ±1 standard error (SEM) of the mean (i.e. smaller error in the proximity of the trigger time and larger far from it, since moving away from the transition front the fluctuations increase and the number of events contributing to the mean decreases). Few Transitions, together with the requirement related to the stability of the Gaussian peak, that leaves out channels with σ above Q3 + 1.5 × IQR (IQR is inter-quartile range Q3 − Q1, with Q1 first quartile and Q3 third quartile). As indicated in the caption of Table 1, the presence of discontinuity in the recording sequence -- corresponding to a failure in the data acquisition and a drop in the log(MUA) -- is not a blocking element, since once the discontinuity is removed, the log(MUA) can fulfil the stability requirements. 9 of 18 2.2. Stability of the data sample and channel selection/rejection A key assumption for comparing acquisitions taken at different sites and with different electrodes (or from different animals) is the comparability of the log(MUA) profile in the low firing rate regime. A quantitative evaluation of such a requirement is obtained by monitoring the width of the peak fitted with a Gaussian function, i.e. the σ parameter of the Gaussian (Figure 4.A). The purpose of this analysis step is to evaluate the selection strategy of the UD_THRESHOLD, which can be fixed or channel-dependent. A dedicated routine has been set in place, operating on a given collection of data (inter-session data from different subjects), aiming at stacking the full set of σ parameters estimated from fitting procedures. In Figure 5 we report the statistical distribution of σ values, and we observe large stability across channels and experiments. In more detail, the plot offers an overview of the range of variability, showing that the behavior of the parameter is pretty stable inside each experiment, and stable in the entire collection as well6. Therefore, we adopted a channel-dependent UD_THRESHOLD set at 2σ, corresponding to a fixed false positive rate per channel of about 2.25%. In addition, the results of the Stability Study enable channel selection or rejection based on the entire data sample, giving rise to a further list of outliers, to be added to the one already filled out for each DAQ session at the end of the main loop. Considering both lists, a total of 27 outliers were identified and removed from the analysis when inter-session data were taken into account for producing summary results. 2.3. Description at the cortical area-level 2.3.1. A thumbnail for the single experiment Once the raw signals have been analyzed for each channel, area-related statistics can be obtained as in [2]. More specifically, SWAP provides the following observables: of the UD-cycle is an estimate of the SO period; 1. durations of the Down states, dDown[s]; 2. durations of the Up states, dUp[s]; 3. durations of the UD-cycles (i.e. a pair of consecutive Down and Up states), dUD[s]. The duration 4. upward transition slope, sUp[s−1]; 5. downward transition slope sDown[s−1]; 6. maximum MUA in the Up state ("peak"), p[a.u.] 7. frequency, f [Hz], defined as 1 dUD ; Concerning dDown, dUp and dUD, as already evident from the histograms (Figure 4.C) and in agreement with the box plots represented in Figure 6.A, the distributions at the channel level are skewed and far from being Gaussian. Therefore, for each channel, the median (and not the mean) has been selected as the representative parameter for the observables. The skewness of the distribution is also noticeable at the area level (Figure 6.B), where the statistics for each distribution are increased since values of electrodes belonging to the same cortical area are grouped together. Therefore, the median is assumed as the representative parameter also at the area level. Interestingly, significant differences are apparent when comparing median values of different channels and areas. Differences among areas are consistently observed in all animals, despite the large span of values that each observable expresses considering the total of 11 experiments. Indeed, evaluating each observable for each experiment at the full-set level of description, the large variability of the data sample is clearly seen. This is in agreement with the expected biological variability of the subjects, regardless of the identical surgical treatment they have been undergone, the comparable drug delivery, and the uniform monitoring conditions during the data taking. In other words, each data session is characterized by its own central values for all the observables of interest, with no 6 Experimental outliers, discussed in Section 2.1 and listed in Table 1, are excluded from the representation. 10 of 18 01 02 03 07 09 10 14 15 16 17 20 Figure 5. Stability of the MUA estimate across channels and animals. (A) Stem plot of the SD values obtained as the σ parameter of the Gaussian fit to the Down peak in the log(MUA) distribution (the blue curve in Figure 4.A); marker colors identify the different experiments. The distribution of the SD values is presented as histogram (B) and as box plot (C), the presence of outliers is evident; descriptive parameters are reported in the box; the vertical lines in the histograms indicate the positions displaced 2 or 3 standard deviations of the mean. Outlier values are removed, obtaining the symmetric distribution represented in the histogram in (E); descriptive parameters are listed in the box, the sample of channels is reduced by 12 units. The grid in (D) is a synthetical representation of the outlier channels for the test bench data: in red, the experiment outliers; in blue, the ones excluded after the SD Stability Study. clear correlation with the animal's phenotype, or with any other parameter measurable during the data acquisition [12,13]. The spanning of the frequency values across the experiments (Figure 7) is in particular revealing, because frequency is a property immediately linkable with the onset of the bimodality and with the propagation of slow waves along the cerebral cortex. 2.3.2. Assessment of statistical significance at the area-level description for inter-session normalized data The statistical significance of the effect of differentiation by area, visible for all the observables in the list of interest and for each experiment in the data set, is quantitatively assessed for summary results, i.e. when the outcome of the single experiments are combined to drive comprehensive claims on the phenomena. However, the first obstacle met when trying to compare and aggregate results is the large variability exhibited by the different DAQ sessions, which dominates over any other effect and disguises any similarity or common footprint among subjects. Therefore, to confront the area-level descriptions, the median values of the observables for each area are normalized for each experiment, M1 S1 PtA RSC V1 11 of 18 Motor (nCh=5) Somatosensory (nCh=6) Parietal (nCh=3) Retrosplenial (nCh=4) Visual (nCh=13) Figure 6. Statistics of the measured UD-cycle (dUD) for an example recording (file 01). (A) Channel-level description; channels are grouped by area, channel numbering is as illustrated in Figure1; channel 8 is missing because it is tagged as an outlier (Table 1). (B) Area-level description, represented with notches indicating the confidence interval for the median (Q2 = 50%, orange line). The box plot is delimited by quartiles Q1 = 25% and Q3 = 75%; IQR = Q3 − Q1 is the Inter Quartile Range; the lower whisker is at Q1 − 1.5 × IQR; the upper whisker is at Q3 + 1.5 × IQR; values outside the whiskers are marked as outliers; the blue marker indicates the mean of the distribution. 01 02 03 07 09 10 14 15 16 17 20 Figure 7. Mean frequency across channels for each experiment; error bars, SEM (standard error of the mean). For each channel, the frequency f [Hz] is computed as <dUD> , with <> denoting the arithmetic mean of the dUD[s] values in the channel (< dUD[s] > is an estimate of the SO period). The text box gives information on the distribution of the observable across the 11 experiments. 1 computing for each observable the arithmetic mean across the cortical areas, and using the obtained mean as a normalization factor for that observable. This procedure highlights any trend expressed at different cortical areas, enabling us to check if different experiments express the same trend. Figure 8.A shows the summary result at the area-level description obtained with normalized data. The outcome of the hypothesis tests executed to assess the statistical significance of the differences observed between each couple of cortical areas is represented through a matrix of p-values, with a graded scale in a given range of confidence (Figure 8.B). Since no assumptions are made on the model, and because of the already-discussed evidence of non-Gaussianity of the distributions, non-parametric approaches are followed when analyzing the test bench data, so the Wilcoxon rank-sum test is A B 12 of 18 Figure 8. (A) Box plot representation of the summary results for the observable dUD at the area-level description, obtained with normalized data. Having 11 subjects, for each cortical area 11 median values are available (the orange line in Figure 6); the arithmetic mean of the 5 medians is the normalization factor for the given experiment. (B) Outcome of the Wilcoxon hypothesis tests on the same data; the test results are represented via a matrix of p-values (with Benjamini-Hochberg correction enabled). applied7. To take into account multiple comparisons (the so-called "look-elsewhere effect"8) and reduce the likelihood of incorrectly rejecting a null hypothesis (type I error) when evaluating a family of simultaneous tests, the robustness of the analysis and the control of the false discovery rate (FDR) is obtained by correcting the p-values with the Benjamini-Hochberg (BH) procedure [15]. Analogously to what was done for state duration, the study of a possible differentiation at the area-level has also been carried out for slopes and maximum MUA (sUp, sDown, p). Slopes are computed considering the average transition (Figure 4.F -- G), obtained pooling together the detected Up-to-Down and Down-to-Up transitions. The transition front of the average transition is isolated, considering a 35 ms interval around the transition time t0 ([t0 − 0.025, t0 + 0.010] for downward transitions; [t0 − 0.010, t0 + 0.025] for upward transitions); the profile is fitted with a cubic and the slope is obtained as the derivative of the polynomial at t0. The average upward transition is also used for the estimation of the maximum MUA of the average waveform in the first 250 ms after the transition. Since slopes and maxima are calculated from the average waveform, for each experiment the observables are represented at the channel level by a single value (and not by a distribution). The area-level description for the single experiment is given by the mean and median of values obtained from all the channels belonging to the specific area. In coherence with the analysis carried out on states duration and in order to apply the same non-parametric Wilcoxon tests, the median is taken as the representative parameter. Summary results are produced after the same normalization adopted for states duration, yielding a similar illustration. A similar approach is followed for frequency f . 2.4. Interpolation of the array map The results obtained with the high spatial resolution probe used for the acquisition of the test bench data can still be improved by interpolating the spaces between the electrodes; the accuracy of 7 8 The computation of the Wilcoxon rank-sum statistics is carried out with the statistical function scipy.stats.ranksums of the scipy Python module [14] https://xkcd.com/882/. 13 of 18 the interpolation is assured by the large surface density of sensors offered by the employed MEA. For each experiment and channel, the median (dDown, dUp, dUD) or the mean (sUp, sDown, p, f ) of the observable is taken. This set of values is used to compute an interpolator9 that is applied to points (xy coordinates) on a mesh-grid, made of 50 steps along x and 90 step along y, with a Grid Step ∼ 0.05 mm, i.e. 1/10 of the Array Step = 0.550 mm (Figure 9.A for the single experiment and for a given observable). The same illustration is applied when representing inter-session normalized data (Figure 9.B); here, the size of the marker is inversely proportional to the standard deviation of the distribution of values at the given electrode position, thus measuring the amount of variability registered at that position across the experiments. A r o i r e t s o p → ← r o i r e t n a medial ←→ lateral Figure 9. (A) Contour plot for the single experiment (here, file 01) illustrating the observable dUD (UDcycleLen) interpolated on the mesh-grid, based on values recorded with the MEA. The array map reflects the DAQ schema of Figure 1, with black dots marking the electrode positions and black lines identifying cortical areas; the white circle identifies the outlier channel as resulting from Figure 5 and Table 1. The color bar gives the range of the interpolated variable across the array. (B) Summary results for the observable dUD, represented as a contour plot obtained from normalized data. The interpolation is based on the mean values computed across the experiments, the marker size is inversely proportional to the standard deviation of the distribution of values at each position (the smaller the marker, the less variability measured across the experiments). Since normalized data are used for the plot, the color bar gives indications of trends with respect to the average: for instance, regions colored in blue are characterized by having a duration of the oscillation cycle that is up to 30% less than the average. Comparing (A) (single experiment) and (B) (inter-session data), a similar tendency is visible for the observable: in particular, the contours evidence gradients along the same dominant antero-posterior direction, from fronto-lateral towards occipito-medial regions. These gradients prevail over the borders of cortical areas, suggesting further differentiation within areas and connections among areas. 9 The interpolation is carried out with the scipy Python module [14] using the scipy.interpolate.Rbf class for radial basis function (Rbf) interpolation. More on https://docs.scipy.org/doc/scipy/reference/generated/scipy.interpolate.Rbf.html 14 of 18 2.5. From the array map to the assessment of statistical significance at the channel level for the normalized inter-session data The contour plots traced on the interpolated array maps give qualitative hints on differentiation among inspected cortical sites, enlightening further details within cortical areas. A more quantitative evaluation can be obtained by performing the statistical analysis of the normalized values resulting from inter-session data, considering each pair of electrodes in the MEA and inspecting if there is a statistically significant difference between the distribution of values measured therein. The idea is to compare the normalized values located at two different electrode positions in the map and to test the validity of the null hypothesis, that is the samples are extracted from the same statistical distribution10. We use the Wilcoxon rank-sum statistics with B-H correction on p-values; in this case, the FDR correction has a larger impact, since for MEAs the number of tests in the family is of the order of hundreds11. Given the large number of hypotheses, electrodes are sorted to emphasize those expressing a significant difference with the others. More specifically, the top-three electrodes in this sorted list are highlighted as "core nodes", indicating the positions in the cortex displaying the largest significant difference with respect to other positions in the cortex (Figure 10.A). As a result, the somatosensory cortex emerges as the one having the shortest mean Up/Down cycle, in agreement with what shown in Figure 8.A. In Figure 10.B, the differences in the cycle duration at the channel level show a rather homogeneous rhythm within single areas. It is also apparent that the other primary sensory area, i.e. the visual cortex, is markedly different from frontal and parietal regions, displaying a significantly longer oscillation period. Inspecting the changes in the Up and Down state duration, we found that the modulation of the oscillation cycle is mainly due to a reduction of the dDown, as shown in Figure 10.C. Indeed, only few channels display a significant difference in dUp (not shown) and the p-value matrix for the dDown mirrors the one shown in Figure 10.B. If the local excitability of the cell assemblies underneath the MEA contacts is well represented by the ratio dUp/dDown [2,16], here the leading role of somatosensory cortex is apparent. Such enhanced excitability is further confirmed by inspecting the slope of the Down-to-Up transition in the MUA [17], which is significantly steeper in the same cortical locations (Figure 10.D). 3. Discussion Biological data are characterized by richness in details and large variability. The efforts of the data analysis should aim at extracting tendencies and regularities, producing a concise description without hiding or neglecting complexity and details that could convey informative content. This is the guideline followed when developing the Slow Waves Analysis Pipeline (SWAP), starting from a solid backbone that has been deeply revised and enriched with new features. In particular, the opportunity of using the MEA data as a test bench has allowed us to focus on the spatial differentiation of the observables, with the aim of uncovering hints as to the local excitability of the cortical assemblies. The developed methodology is robust and easily re-configurable, flexible and adaptable at different acquisition conditions, and also suitable to be applied to the output of simulations. In this framework, SWAP can be employed in bridge theory, simulations and experiments, providing a set of general tools that allow an effective comparison between heterogeneous data sets. The adoption of a unique analysis procedure is also useful for comparing different simulation engines; the SWAP can be applied to define benchmarks and evaluate the performance of numerical models and implementations. Several studies are ongoing for the application of SWAP to a large variety of datasets (knock-out mice, subjects in 10 The number of samples at each location depends on how many experiments have channels rejected as outliers at that location. For the test bench data set, a maximum of 11 samples can be present at each electrode position, details for each electrode position can be extracted from Figure 5.D. If the number of electrodes in the MEA is n = 32, the number of hypotheses to be checked is N = n(n−1) 11 2 = 496. A M S P R V C 15 of 18 B D r o i r e t s o p → ← r o i r e t n a medial ←→ lateral Figure 10. Excitability maps of the probed cortices. (A) Normalized dUD (as in Figure 9.B) with the "core-nodes" (i.e the three electrodes with the largest number of significant differences with the other channels). Lines are traced to connect the "core nodes" with their partners in the electrode pair being significantly different (p < 0.05). (B) Matrix of p-values of the statistical test on the dUD differences between channels. The matrix construction is analogous to what is described for Figure 8. The numeric sequence of electrodes in the matrix is as in Figure 6, i.e. channels are grouped by area (M = motor, S = somatosensory, P = parietal, R = retrosplenial, V = visual). (C,D) Normalized Down state duration dDown (DownStateLen) and Down-to-Up transition slope sUp (SlopeUp), averaged across experiments, as in Panel (A). different brain states, data collected with optical techniques), and the stability and reliability of the analysis procedure has been so far confirmed. The introduced new features in the analysis pipeline have been coded in Python with the aim of realizing open software tools for the scientific community. The complete transition towards open software is in the list of objectives to fulfill in the near future. 16 of 18 Concerning the interpretation of the results, the analysis of the large set of data collected from 11 wild-type anesthetized mice with the 32-channel MEA allows us to claim a statistically significant differentiation of cortical areas for several parameters that characterize the onset of SWA along the cerebral cortex. Starting from these observables, the excitability of the cortical tissue expressing SWA can be investigated. Indeed, larger excitability is expected to be associated with faster transition fronts (in particular, upward slopes), shorter duty-cycles (i.e. smaller dUD, dominated by the duration of the Down states) and accordingly larger frequencies [17]. For instance, a smaller dDown reveals the case in which excitability translates in faster Down-to-Up transitions. These features are particularly apparent in the somatosensory area, likely the most excitable cortical region we observed; conversely, the occipital lobe (retrosplenial and visual areas) act as the least excitable. Activation waves traveling across the cortex during SWA are expected to be sensitive to the diverse cortical excitability, as more reactive (i.e. more excitable) areas are expected to display a smaller "inertia" in recruiting neurons involved in the collective phenomena associated with the high-firing Up states. As a consequence, waves might be originating from highly excitable regions such that preferential propagation pathways are expected, as previously highlighted both in humans [18] and rodents [2,19,20]. In turn, heterogeneous excitability might also underlie the non-global nature of the SWA phenomenon observed when wakefulness is approaching [21,22] or under pathological conditions [23]. It must be pointed out that the study here presented is "static", in the sense that the absolute time sequence of the events is not taken into account: states and transitions are time-squashed and stacked regardless of their time of occurrence in the DAQ session, and thus any time-dependent effect, like fatigue and recovery intervals of the neurons, that could affect the excitability and alter the responsiveness of cortical regions, is excluded from this analysis. In other words, the figures presented in this paper can be seen as the average photography of the phenomena investing the monitored cortex during the acquisition period. Related to this, excitability and responsiveness can also be altered by wave propagation dynamics, in particular if different propagation patterns coexist and overlap in the same time interval, for instance when two slow waves originating at distinct sites travel along the cortex at different speeds and each one follows its own path [4,24]. Again, not using the information on the absolute timing of the events at the electrode positions and intending the results presented here to be the average photography of the SWA in the cortex, the included results reflect the dominant propagation pattern, namely the antero-posterior direction [2,18,19,25], in particular along an oblique axis directed from fronto-lateral towards occipito-medial regions, as suggested by values depicted in the contour plots. As in this step of the SWAP we have not focused on time-dependent effects, their impact on the area-differentiation of the bistability modes will be evaluated in further studies. Finally, although in all the animals involved in this study the level of administered anesthesia was the same, the observed inter-subject variability (Figure 7) could in principle be explained by animals being in different brain states. This suggests the possibility of exploiting the characterization of slow oscillations in the cerebral cortex as a new tool for effective classification of the brain states. Several techniques are currently under study, based for instance on the principal components analysis (PCA) to identify and single out the different sources of variability in the experimental data set. Indeed, a more reliable classification of brain states (i.e. of the DAQ sessions that constitute the statistical sample) would provide a more robust comparison of the experiments, allowing us to overcome the limits derived from the need to use normalized data. 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. Author Contributions MVSV, MM, MD, PSP and GDB conceived and designed the study. MD performed the data acquisition. MM provided the backbone of the analysis pipeline. GDB revised the analysis pipeline, 17 of 18 improved it with new functions, applied it to the test bench data, performed the statistical analysis, wrote the first draft of the manuscript. PSP and MM contributed to the data analysis and to the interpretation of the results. AP revised sections of the manuscript. All authors contributed to manuscript revision, and read and approved the submitted version. Funding This work has received funding from the European Union Horizon 2020 Research and Innovation Programme under Specific Grant Agreements No. 785907 (HBP SGA2) and No. 720270 (HBP SGA1). Acknowledgments This study was carried out in the framework of the Human Brain Project (HBP12), funded under Specific Grant Agreements No. 785907 (HBP SGA2) and No. 720270 (HBP SGA1), in particular within activities of sub-project SP3 ("Systems and Cognitive Neuroscience"). Data Availability Statement Data will be available on reasonable request. 1. 2. 3. 4. 5. 6. 7. 8. 9. J Neurosci. Steriade, M.; Nunez, A.; Amzica, F. A novel slow (< 1 Hz) oscillation of neocortical neurons in vivo: depolarizing and hyperpolarizing components. J Neurosci. 1993, 13, 3252 -- 3265. Ruiz-Mejias, M.; Ciria-Suarez, L.; Mattia, M.; Sanchez-Vives, M.V. Slow and fast rhythms generated in the cerebral cortex of the anesthetized mouse. J Neurophysiol. 2011, 106, 2910 -- 2921. doi:10.1152/jn.00440.2011. Mattia, M.; Pani, P.; Mirabella, G.; Costa, S.; Del Giudice, P.; Ferraina, S. Heterogeneous Attractor 2013, 33, 11155 -- 11168. Cell Assemblies for Motor Planning in Premotor Cortex. doi:10.1523/JNEUROSCI.4664-12.2013. Capone, C.; Rebollo, B.; Muñoz, A.; Illa, X.; Del Giudice, P.; Sanchez-Vives, M.V.; Mattia, M. Slow waves in cortical slices: How spontaneous activity is shaped by laminar structure. Cereb. Cortex 2019, 29, 319 -- 335. doi:10.1093/cercor/bhx326. Ruiz-Mejias, M.; Perez-Mendez, L.; Ciria-Suarez, L.; Martinez de Lagran, M.; Gener, T.; Sancristobal, B.; Mattia, M.; Castano-Prat, P.; García-Ojalvo, J.; Gruart, A.; Delgado-García, J.M.; Sanchez-Vives, M.V.; Dierssen, M. Over-expression of Dyrk1A, a Down syndrome candidate, decreases excitability 2016, 36, 3648 -- 3659. and impairs gamma oscillations in the pre-frontal Cortex. doi:10.1523/JNEUROSCI.2517-15.2016. Pazzini, L.; Polese, D.; Weinert, J.F.; Maiolo, L.; Maita, F.; Marrani, M.; Pecora, A.; Sanchez-Vives, M.V.; Fortunato, G. An ultra-compact integrated system for brain activity recording and stimulation validated over cortical slow oscillations in vivo and in vitro. Scientific reports 2018, 8, 16717. Paolucci, P.S.; Ammendola, R.; Biagioni, A.; Frezza, O.; Lo Cicero, F.; Lonardo, A.; Pastorelli, E.; Simula, F.; Tosoratto, L.; Vicini, P. Distributed simulation of polychronous and plastic spiking neural networks: strong and weak scaling of a representative mini-application benchmark executed on a small-scale commodity cluster. arXiv:1310.8478 [cs.DC] 2013. Pastorelli, E.; Paolucci, P.S.; Ammendola, R.; Biagioni, A.; Frezza, O.; Lo Cicero, F.; Lonardo, A.; Martinelli, M.; Simula, F.; Vicini, P. Impact of exponential long range and Gaussian short range lateral connectivity on the distributed simulation of neural networks including up to 30 billion synapses. arXiv:1512.05264 [cs.DC] 2015. Pastorelli, E.; Paolucci, P.S.; Simula, F.; Biagioni, A.; Capuani, F.; Cretaro, P.; Bonis, G.D.; Cicero, F.L.; Lonardo, A.; Martinelli, M.; Pontisso, L.; Vicini, P.; Ammendola, R. Gaussian and Exponential Lateral Connectivity on Distributed Spiking Neural Network Simulation. 2018 26th Euromicro J Neurosci. 12 https://www.humanbrainproject.eu/en/ 18 of 18 International Conference on Parallel, Distributed and Network-based Processing (PDP), 2018, pp. 658 -- 665. doi:10.1109/PDP2018.2018.00110. 10. Mattia, M.; Ferraina, S.; Del Giudice, P. Dissociated multi-unit activity and local field potentials: a theory 11. 12. 13. 14. 15. inspired analysis of a motor decision task. Neuroimage 2010, 52, 812 -- 823. Lidierth, M. sigTOOL: A MATLAB-based environment for sharing laboratory-developed software to analyze biological signals. J Neurosci Methods 2009, 178, 188 -- 196. doi:10.1016/j.jneumeth.2008.11.004. Brown, E.N.; Lydic, R.; Schiff, R. General anesthesia, sleep, and coma. N Engl J Med. 2010, 363, 2638 -- 2650. doi:10.1056/NEJMra0808281. Akeju, O.; Brown, E.N. Neural oscillations demonstrate that general anesthesia and sedative 2017, 44, 178 -- 185. states are neurophysiologically distinct from sleep. doi:10.1016/j.conb.2017.04.011. Jones, E.; Oliphant, T.; Peterson, P.; others. SciPy: Open source scientific tools for Python, http://www. scipy.org/ 2001 -- . Seabold, S.; Perktold, J. Statsmodels: Econometric and statistical modeling with Python. 9th Python in Science Conference, 2010. Curr Opin Neurobiol. 16. Mattia, M.; Sanchez-Vives, M.V. Exploring the spectrum of dynamical regimes and timescales in 17. 18. 19. spontaneous cortical activity. Cogn. Neurodyn. 2012, 6, 239 -- 50. doi:10.1007/s11571-011-9179-4. Sanchez-Vives, M.V.; Mattia, M.; Compte, A.; Perez-Zabalza, M.; Winograd, M.; Descalzo, V.F.; Reig, R. Inhibitory modulation of cortical up states. J. Neurophysiol. 2010, 104, 1314 -- 1324. doi:10.1152/jn.00178.2010. Massimini, M.; Huber, R.; Ferrarelli, F.; Hill, S.L.; Tononi, G. The sleep slow oscillation as a traveling wave. J. Neurosci. 2004, 24, 6862 -- 70. doi:10.1523/JNEUROSCI.1318-04.2004. Stroh, A.; Adelsberger, H.; Groh, A.; Rühlmann, C.; Fischer, S.; Schierloh, A.; Deisseroth, K.; Konnerth, A. Making Waves: Initiation and Propagation of Corticothalamic Ca2+ Waves In Vivo. Neuron 2013, 77, 1136 -- 1150. doi:10.1016/j.neuron.2013.01.031. 21. 22. 20. Mohajerani, M.H.; Chan, A.W.; Mohsenvand, M.; LeDue, J.; Liu, R.; McVea, D.a.; Boyd, J.D.; Wang, Y.T.; Reimers, M.; Murphy, T.H. Spontaneous cortical activity alternates between motifs defined by regional axonal projections. Nat. Neurosci 2013, 16, 1426 -- 35. doi:10.1038/nn.3499. Vyazovskiy, V.V.; Olcese, U.; Hanlon, E.C.; Nir, Y.; Cirelli, C.; Tononi, G. Local sleep in awake rats. Nature 2011, 472, 443 -- 7. doi:10.1038/nature10009. Nir, Y.; Staba, R.J.; Andrillon, T.; Vyazovskiy, V.V.; Cirelli, C.; Fried, I.; Tononi, G. Regional slow waves and spindles in human sleep. Neuron 2011, 70, 153 -- 69. doi:10.1016/j.neuron.2011.02.043. Sanchez-Vives, M.V.; Massimini, M.; Mattia, M. Shaping the default activity pattern of the cortical network. Neuron 2017, 94, 993 -- 1001. doi:10.1016/j.neuron.2017.05.015. Sancristóbal, B.; Rebollo, B.; Boada, P.; Sanchez-Vives, M.V.; Garcia-Ojalvo, J. Collective stochastic coherence in recurrent neuronal networks. Nat. Phys. 2016, 12, 1 -- 8. doi:10.1038/nphys3739. Chan, A.W.; Mohajerani, M.H.; LeDue, J.M.; Wang, Y.T.; Murphy, T.H. Mesoscale infraslow spontaneous membrane potential fluctuations recapitulate high-frequency activity cortical motifs. Nat. Comm. 2015, 6, 1 -- 12. doi:10.1038/ncomms8738. 23. 24. 25.
1806.04703
1
1806
2018-05-28T18:10:04
Electronic schematic for bio-plausible dopamine neuromodulation of eSTDP and iSTDP
[ "q-bio.NC" ]
In this technical report we present novel results of the dopamine bio-plausible neuromodulation excitatory (eSTDP) and inhibitory (iSTDP) learning. We present the principal schematic for the neuromodulation of D1 and D2 receptors of dopamine, wiring schematic for both cases as well as the simulatory experiments results done in LTSpice.
q-bio.NC
q-bio
Electronic schematic for bio-plausible dopamine neuromodulation of eSTDP and iSTDP Max Talanova, Yuriy Gerasimova, Victor Erokhina,b aKFU, Russia. bCNR-IMEM, Italy. Abstract In this technical report we present novel results of the dopamine bio- plausible neuromodulation excitatory (eSTDP) and inhibitory (iSTDP) learn- ing. We present the principal schematic for the neuromodulation of D1 and D2 receptors of dopamine, wiring schematic for both cases as well as the simulatory experiments results done in LTSpice. Keywords: neuromodulation, dopamine, neuromorphic computing, affective computing, simulation 1. The experimental set-up In our previous technical report we presented the electronic schematic to generate complex learning impulses for memristive devices to implement iSTDP and eSTDP learning functions (Talanov et al., 2017). We have also presented the physical implementations and experimental results for dopamine modulation implemented as an amplification of learning impulses. Based on works of Hennequin et al. (2017); Vogels et al. (2013); Gurney et al. (2015) we have extended our previous model with electronic schematic for bio-plausible neuromodulation for D1 and D2 dopamine receptors presented in the Fig. 1 (left graphs). 1.1. Block diagram We have decided to use mixture of two functions that represents extreme positions in the dopamine modulation STDP "Sombrero" and Hebbian in- Email addresses: [email protected] (Max Talanov), [email protected] (Yuriy Gerasimov), [email protected] (Victor Erokhin) Preprint submitted to arxiv.org July 31, 2018 Fig. 1: The block diagram of the neuromodulatory feedback with inputs for: "Sombrero" impulses, Hebbian impulses and DA input that identifies the level of the dopamine. For the complete description of input functions please see our previous technical report (Talanov et al., 2017) dicated as S and H triangles in Fig. 1. Input signals of "Sombrero" and Hebbian learning functions are balanced via dopamine level represented as DA green triangle in Fig. 1. The dopamine level setts up gain for each input "Sombrero" - GS and Hebbian - GH respectively in corresponding amplifier indicated as Amplif ier GH and Amplif ier GS. The output of amplifiers is processed by the Adder that implements the mixture of two modulated via dopamine input learning functions. Two cases of the dopamine modulation is represented as two panels in the Fig. 1 with the main difference in ini- tial learning function for the low level of the dopamine: for D1 inverted -- "Sombrero" and for D2 -- "Sombrero". Thus the Inverter that implements the inverted "Sombrero" for the low level of the dopamine is included in the block diagram for the D1 schematic. 2 1.2. Wiring schematic Fig. 2: The electronic circuit schematic of the D1 modulation. The U 18 is inverter to implement the D1 extreme learning function the inverted "Sombrero". U 15 and U 16 are learning impulses inverting amplifiers with gain set by dopamine level indicated as mod2 that modulates the resistance of R55 and R58. The output weighted inverting adder is represented as U 17, where "Sombrero" signal amplitude is twice lower than Hebbian. The wiring schematic of the dopamine modulation via receptor D1 is depicted in Fig. 2. The op-amp U 18 plays the role of inverter for the "Som- brero" input signals to implement the extreme state of the learning function for D1 receptor the inverted "Sombrero" (see Fig.1). Two op-amps U 15 and U 16 are oppositely balanced amplifiers via resistances R55 and R58. The value of the R55 is set via digital potentiometer: R = ((1 − V (mod2)) ∗ 105 + 10−3) (1) The value of the R58 is set via digital potentiometer opposed to R55: R = (V (mod2) ∗ 105 + 10−3) (2) When V (mod) = 1V , according to eq.1, R55 = 10−3Ohm, which makes gain almost equal zero. At the same time, from eq. 2 R58 = 100K, combined with R57 makes gain equals to one. If V (mod) = 0 gains are opposite. The outbound signal of both amplifiers U 15 and U 16 is processed by the weighted adder implemented via U 17. The "Sombrero" amplitude is twice lower than 3 Fig. 3: The electronic circuit schematic of the D2 modulation. U 12 and U 13 are learning impulses inverting amplifiers with gain set by dopamine level indicated as mod that mod- ulates the resistance of R34 and R37. The output weighted inverting adder is represented as U 14, where "Sombrero" signal amplitude is twice lower than Hebbian. Hebbian, to match global amplitude change in process of dopamine modula- tion. The modulation schematic for D2 receptor is done in similar way except for the inverter of the "Sombrero" input signals that in case of D2 receptor is not used. The neuromodulatory schematic is represented in Fig. 3. 1.3. Results For the simulatory validation we have used LTSpice software framework. Inbound signals are presented in Fig. 4 where green graph is the "Sombrero" signals and blue is the Hebbian signals for both the D1 and D2 dopamine receptors. The results of D1 modulation are indicated in Fig. 5 where the proportion of the influence of the inbound learning function over outbound signal is defined by the level of the dopamine or set up of resistors R55 and R58. The level of the dopamine is represented as the red graph. Learning impulses that indicate the mixture of two learning functions inverted "Sombrero" and Hebbian are represented as blue graph. The results of D2 modulation are presented in the Fig. 6 with the same color coding, where the level of the dopamine is identified via resistors R34 and R37. 4 Fig. 4: "Sombrero" (green) and Hebbian (blue) learning impulses. Fig. 5: Simulation of the D1 circuit in Fig.2. "Dopamine" level (red) and D1 modulation result (blue). 5 Fig. 6: Simulation of the D2 circuit in Fig.3. "Dopamine" level (red) and D2 modulation result (blue). The initial state of both graphs is low level of the dopamine and the learning function is inverted "Sombrero" for D1 and "Sombrero" for D2 re- spectively. The learning function for high level of the dopamine is Hebbian or 1 x. The gradual mixture of initial learning functions with final Hebbian is represented during simulatory experiments and is depicted on both figures starting from 4 seconds till 40 seconds of the simulation. References Gurney, K. N., Humphries, M. D., and Redgrave, P. (2015). A New Framework for Cortico-Striatal Plasticity: Behavioural Theory Meets In Vitro Data at the Reinforcement-Action Interface. PLoS Biology, 13(1):e1002034. Hennequin, G., Agnes, E. J., and Vogels, T. P. (2017). Inhibitory Plasticity: Balance, Control, and Codependence. Annual Review of Neuroscience, 40(1). Talanov, M., Zykov, E., Gerasimov, Y., Toschev, A., and Erokhin, V. (2017). Dopamine modulation via memristive schematic. CoRR, abs/1709.06325. 6 Vogels, T. P., Froemke, R. C., Doyon, N., Gilson, M., Haas, J. S., Liu, R., Maffei, A., Miller, P., Wierenga, C., Woodin, M. A., Zenke, F., and Sprekeler, H. (2013). Inhibitory synaptic plasticity: spike timing- dependence and putative network function. Frontiers in Neural Circuits, 7:119. 7
1605.01381
1
1605
2016-03-04T12:02:40
Review of analytical instruments for EEG analysis
[ "q-bio.NC" ]
Since it was first used in 1926, EEG has been one of the most useful instruments of neuroscience. In order to start using EEG data we need not only EEG apparatus, but also some analytical tools and skills to understand what our data mean. This article describes several classical analytical tools and also new one which appeared only several years ago. We hope it will be useful for those researchers who have only started working in the field of cognitive EEG.
q-bio.NC
q-bio
Review of analytical instruments for EEG analysis S.N. Agapov1, V.A. Bulanov2, A.V. Zakharov3, M.S. Sergeeva4 [email protected] [email protected] IT Universe LLC, Department of BCI. Samara, Russia. [email protected] [email protected] Samara State Medical University, Department of BCI and Applied Neurophysiology. Samara, Russia. Abstract: Since it was first used in 1926, EEG has been one of the most useful instruments of neuroscience. In order to start using EEG data we need not only EEG apparatus, but also some analytical tools and skills to understand what our data mean. This article describes several classical analytical tools and also new one which appeared only several years ago. We hope it will be useful for those researchers who have only started working in the field of cognitive EEG. Copyright: 2016 IT Universe LLC. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are credited. Competing interests: The authors have declared that no competing interests exist. Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Introduction page 2 of 14 Over the past few decades statistic have grown considerably, because new research in abstract and applied mathematics has been conducted. Outspread of computers and the creation of WWW have played a crucial role for statistical analysis. Many powerful methods such as Bayesian analysis, Monte-Carlo simulation, FFT, wavelet analysis, spline interpolation, generalized linear/additive models, machine learning can be applied because we have cheap and fast computers. Some methods such as Bayes function were discovered in the 19th or early 20th century, but they have been out of use due to hard practical implementation. Over the last decade we have seen blow up grow of machine learning methods in science and engineering, and during the past several years the most modern and impressive computer science technology is Big Data. Every day the world science community generate incredible amount of information. It is not only data but also new methods, theories, concepts and application. In order to orientate ourselves inside this "ocean of data" we need new metastructure for analysis procedures and best practices. For this purpose, we want provide an overview of some of the modern methods and show how these groups of methods are interrelated. We suggest metastructure for the data mining procedure and its application for EEG data analysis. Layer 1: Subject area - Basic cognitive EEG concepts Human brain produces complex electric fields patterns. In order to understand the basic EEG technique there are good books about electroactivity and neuron oscillations. The following book is good starting point for learning about biophysics that underlie the EEG signals (Nunez & Srinivasan, 2006). The authors cover small- and large-scale electrogenerators, passive volume conduction, recording strategies, spectral analysis, source localization. If you have some background knowledge in physics (electricity and magnetism), this reference book will prove useful to you. The introductory course of neurobiological mechanisms of oscillations can be found in the book by Buzsaki (Buzsáki, 2006). For general information about organization and design experiments in cognitive EEG you can read the book by Luck (Luck, 2005). You can find comprehensive introduction in EEG analysis in the book by Cohen (Cohen, 2014). The author touches upon many practical questions, from preparing equipment for experiments to statistical inference and overview of the future of cognitive electrophysiology. Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Layer 2: Preprocessing page 3 of 14 Time-series analysis has a huge set of methods, solid mathematics background and a long history in science and engineering. We try to understand the character of our data when we see their visual representation. The investigation of graphic of the data can give keys to the nature and character of function. Even simple graphics observation (peaks, valleys, periods, linear trends) can help us in our research. There is no substitution for raw data investigation by open eye. Before starting to use the data we often need certain preprocessing: cleaning noise, removing artifacts and malfunctioning channels, filtering low and high frequencies, applying baseline assignment. Details about data prepossessing can be found in books by Luck and Cohen (Luck, 2005), (Cohen, 2014). After preprocessing we start using all other methods of analysis. Layer 3: Analytical decomposition. Spectral analysis and TDA Fourier analysis: Fourier analysis is rather old and the most elaborated method of analysis. It is a mathematical technique for transforming the signal from time-based to frequency-based domain. Any function f(t) we can be decomposed in linear combination of sin(x) and cos(x): !" =$%+ 1 '23 $'cos+,"- +.'sin+,"- 10 5 5 6 4 2 2 4 6 10 Figure 1. Function 3sin2& + 5cos3& + 7sin (4&) and its spectrum. Pairs (a, n), (b, n) form a spectrum of signal. Figure 1 is an example of a function and its spectral decomposition. The techniques of Fourier analysis widely known and have a long history. We can analyze spectrum by using both classical statistical methods, i.e. Bayesian and modern techniques, i.e. a topological data page 4 of 14 Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia analysis. However, there are several limitations for Fourier transformation to be used with EEG signals. Fourier analysis is suitable for stationary and linear signals because sin waves do not have time localization, but EEG signals have non-stationary nature. Presentation of frequency dynamics in time domain is not an easy task and it is hard to understand its meaning intuitively. Due to these limitations it is better to use wavelets in EEG analysis. Orthogonal polynomials: The next step in the process of choosing a basis function for decomposition is Chebyshev polynomials. As we know from the theory of partial differential equations, Chebyshev polynomials will be a better choice for spectrum decomposition almost in all cases unless the solution is spatially periodical. They are defined by the equation: !"#$%& ≡cos (-&) !"# = 12"(!*"*#" #+-1 " For non-periodic cases Legendre polynomials can also be used: The theory of orthogonal polynomials is still being developed. And this theory supported by fast computational methods to use it in daily practice. We can try to make use of some of polynomials in EEG analysis for a spectral decomposition, because, as we know from physics, sometimes a change of basis in a system can provide a simpler picture of researched phenomena. For the overview of the modern state of the theory of orthogonal polynomials see, for example, in the lectures notes by Gautschi (Gautschi, 2006). The author provides a library of Matlab functions for computing many modern polynomials. Wavelets: Sin(x) and cos(x) waves are not localized in time domain. By contrast, wavelets are "small" waves which have good time localization. For example, Morlet wavelet is a composition of two functions: Gauss and sin(x) (Figure 2). Figure 2. Example of Morlet wavelet. Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia page 5 of 14 1.0 0.8 0.6 0.4 0.2 1.0 0.5 0.5 1.0 1.5 2.0 Haar wavelet Mexican Hat wavelet Biorthogonal spline wavelet Figure 3. Examples of wavelets. Splines: After De Boor had introduced the B-splines numerical computation method (De Boor, 1972), splines spread widely in many areas of science and engineering. Fast computation and easy construction procedures for splines make them a useful tool in many numerical experiments. As a first step in the construction of a spline we divide an interval of function we wish to approximate into N subintervals. Then we allocate polynomial of degree for each subinterval n, n = 1..N. Thirdly, we join all the adjacent polynomials together to form a piecewise polynomial curve. The points which join polynomials are called knots. Polynomials at knots should satisfy some continuity constraint: they have to be differentiable at these points. Splines play several roles in modern analytical methods. First, they are widely used as auxiliary entities: for approximation and interpolation of non-periodic functions, in wavelets construction (see biorthogonal spline wavelet in Figure 3). In functional data analysis (FDA) B-splines are used as basis functions, in Hilbert-Huang transformation they are used for the implementation of IMF. Periodic splines can be applied for Spline Harmonic Analysis which combines computation speed from FFT and approximation abilities of splines (Zheludev, 1998). Blind source separation and ICA: For the last two decades blind source separation has applied widely in several areas as telecommunications, digital signal processing, biomedical engineering, financial data analysis, astronomical imaging. The goal of BSS acquire the several unknown sources of signals from sensor arrays. Classical problem is known as separate signals from two sources by two microphones. Independent component analysis (ICA) one of the first techniques of BSS revealing hidden parameters of observed signals. But as we know in EEG analysis extensively used only ICA and as part Hilbert – Huang transformations empirical mode decompositions. ICA method assumes that the original sources are statistical independent and uses several good known algorithms as Infomax, maximum likelihood estimation, the maximum a posterior and Fast ICA. In EEG analysis ICA method often uses for removing artifacts from raw signals. page 6 of 14 We think that the modern methods of BSS paired with machine learning Modern set of BSS tools is very rich and consist of such methods as empirical Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia mode decompositions, compressed sensing, factor analysis, dictionary learning. algorithms can give powerful tool for researchers and engineers. Hilbert-Huang transformation: Any set of predefined functions cannot give good results in the cases when are signals nonstationary and nonlinear. In this case we need an adaptive basis for our analysis. It means that basis has to be relevant to the data or data-dependent. Hilbert - Huang transformation resolves some problems for that type of complex data (Huang et al., 1998). HHT consists of two parts – Hilbert spectral analysis and empirical mode decomposition (EMD). EMD is an adaptive signal decomposition into a sum of natural, intrinsic blocks (intrinsic mode functions IMF) which describe complex waves. EMD still lacks rigorous theoretical foundation, it is simple but "empirical" algorithm. HHT is widely used in applications in spite of some problems in theory and methodology of this method. TDA: Topological data analysis appeared only several years ago. Its basis is algebraic topology – a branch of mathematics which is connected with dynamics and computer science. Persistent homology is the main concept of topological data analysis. Any data have a structure, and topology can give us some keys to their nature. We replace data with simplicial complexes using persistent homology, then we transform them into the parameterized form of Betty numbers called barcode. If digital EEG data a presented in time or frequency domain, we receive a point cloud. Then the 'shape' of this cloud provides us with information about the nature of underlying neuron activity. Algorithms for computing have implementations in Matlab code, and in CRAN repository of R. For introduction in some concepts of TDA you can see the book by Edelsbrunner and Harer (Edelsbrunner & Harer, 2008). Layer 4: Synthetic and smooth methods. Statistical methods and FDA Classical statistics: Classical statistics provide powerful tools, both basic and advanced, for any type of analysis. Simple descriptive statistics such as mean, standard deviation, quartiles can be used almost in all kind of data investigations. Hypothesis testing, p-values, confidence intervals are easy but powerful instruments of statistical inference. We use them for avoiding pitfalls in the early stages of data analysis. Analysis of variance (ANOVA), statistical independence test like chi- squared test and correlations give us opportunity to find relations in our data. To understand modern methods such independent component analysis (ICA) we need some background in information theory. Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Bayesian statistics: Bayesian statistics use only one tool – the Bayes' theorem: !"#$%$ = !"×)#$%$" ʃ !"×)#$%$"#" page 7 of 14 This method has several differences from classical statistics: • Classical or "frequentist" statistics do not mean any prior knowledge about the data which we analyze. Bayesian statistics use the prior information about the process and the information about the process is contained in data. • Bayesian method supplies a direct tool for computation of parameters probabilities. • Bayesian statistics can marginalize nuisance parameters effectively. Good practical modern introduction in Bayesian analysis can be found in Bayesian inference can be imagined as a statistic "Occam's razor". The iterative use of Bayesian theorem eliminates all impossible hypotheses from our prior knowledge. In the end, we will receive only most probable outcomes. (Kruschke, 2011). Generalized linear/additive models: The dependence of current values on the previous values can explain the correlation between adjacent points. Time-series analysis uses such methods as autocorrelation and cross-correlation functions. Their modern modification such as ARIMA was introduced in 1970 by Box and Jenkins, and after that linear use generalized models grow significantly. Linear models are statistics models in which responses are presented as linear combinations of predictors and a sum of random error component. Generalized linear models (GLM) weaken the linearity of linear models. They allow a expected value of response to depend on differentiable function of the predictor. Generalized additive models (GAM) are the next step from linearity in time-series analysis. They allow predictors to be sum of smooth functions, which do not have straight parametric form, and may have only limited differentiability. This evolution from linearity to non-linearity is a balance between flexibility and overfitting from one hand, and model interpretability on the other hand. For introduction in GLM/GAM with implementations in R you can read book by Wood (Wood, 2006). FDA: Functional data analysis is a generalization of multivariate data in infinite dimension. The resulting data can be curves, surfaces оr other continuous complex objects. The term functional data analysis was first introduced in the article by (Ramsay and Dalzell, 1991). This framework includes several methods, concepts and ideas from functional analysis, calculus of variations and statistics. The concept of FDA is characterized by the following criteria: Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia page 8 of 14 functional data is continuous individual unit of data is a function • • • smoothness of data is one of the key aspects of this analysis • the derivatives of the functions often play an important role • principal component analysis is one of central concepts For comprehensive introduction in FDA techniques see (Ramsay and Silverman, 2005). Layer 5: Machine learning Over the past few decades the use machine learning methods has increased dramatically. They pull together the best from expert systems, neuron networks, Bayesian statistics and pattern recognition. The concept of machine learning is based on the four fundamental principles: Data, Abstraction, Generalization, Evaluation. Data: Before we start learning something we need data, which must be collected in advance. In case of EEG there must be some EEG device, input interface and data storage for collecting data. It is important to preprocess by cleaning them from any noises, artifacts and mistakes. However, gathering and storing data do not provide us with knowledge, that we should take next step. Abstraction: At the abstraction stage we try to isolate any kinds of structures from data. They may be connected or disjoint graphs, pictures, mathematical equations, clusters, logical rules etc. During this phase we try to look at the data and identify something meaningful or obvious for us. Like ancient people who sow star constellations in the sky, we also seek any close connections in data. For example, for EEG we decompose signals in a spectrum and try to find better statistical distribution. But still it is not full knowledge and in some cases not knowledge at all. Generalization: If we apply the gained abstracted knowledge only for one set of data, it will not be of much use. Good models can be used in many areas of science and engineering. Evaluation: During the last stage we apply our algorithm for new dataset. It is obvious that perfect generalization of model to unseen data is very rare. In many cases failure occurs due to noise in data. If we include the noise in our model, the system may grow redundant and may start recognize of true patterns in data in worse way. This is known as "overfitting" problem. page 9 of 14 Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia All methods of machine learning can be divided into two groups: supervised and unsupervised. Supervised methods involve creating a predictive statistical model. In unsupervised learning there is no model at all, but we can identify connections and structures inside the input data. Machine learning algorithms solve several tasks: clustering, classification, prediction. Clustering is a first task in exploratory data analysis; raw data should be divided in any groups for investigation. After clusterization we can use classification for categorical data, and prediction for nominal datasets. For introduction to machine learning with software implementation in R see (James, Witten, Hastie, and Tibshirani, 2013) and (Lantz, 2013), in Weka (Witten, Frank, and Hall, 2011), in Matlab (Theodoridis & Koutroumbas, 2008). In EEG methods of machine learning have been widely used over the last decade, see for example (AlZoubi, Calvo, and Stevens, 2009) (Sohaib, Qureshi, Hagelbäck, Hilborn, 2013) (Höhne, Bartz, Hebart, Müller, and Blankertz, 2015). Layer 6: Big Data During past several years Big Data has emerged as a new keyword in data analysis. EEG data do not fit into the formal scheme of Five V of Big Data as it is given in (Demchenko, Grosso, De Laat, and Membrey, 2013). The Five V are: • volume • velocity • variety • value • veracity EEG data are obviously not "too" big and have structured nature. Big Data provide several interesting tools for concurrent computations. For instance Hadoop cluster gives opportunities to create systems for processing massive data. One of the most interesting components of Hadoop environment is Apache Spark. This is an open-source, fast, general-purpose engine for large-scale data processing. Apache Spark extends the Map-Reduce models and makes it possible to combine different processing types of workloads. Spark has APIs to Python, Java, Scala, R and SQL. It consists of several integrated parts, which form the following structure: Mlib, Machine learning SQL engine Spark Streaming Standalone Scheduler Spark Core YARN GraphX, graph processing Mesos Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia The integration of SQL engine, machine learning, library, graph computing tool, and scheduler of distributed processes in one framework can offer new opportunities for EEG analysis even for relatively small sets of data. page 10 of 14 Software tools Matlab: Modern data analysis is almost impossible without computer programs. In neuroscience data arrays too big to handle them manually. In DSP (digital signal processing) MathWork Matlab is standard de facto. Matlab have fast linear transform procedures, so its main power in use for applications of linear algebra. Matlab have script program language and mechanism for creation and modernization of toolboxes (user application packages). For general introduction to Matlab you can read (Sizemore and Mueller, 2014). For purposes of EEG analyzes science community created a lot of applications (toolboxes), and new projects still start every year. We list several of them: EEGLab, FieldTrip, BrainStorm. EEGlab good starting point to try. It have elaborate structure, powerful methods for time-series and independent component analyzes, see (Delorme and Makeig, 2004), (Delorme, Makeig, Debener, and Onton, 2004), (Makeig and Onton, 2011). It can read a data from many modern datafile formats. Friendly, easy tutorial you can find on http://sccn.ucsd.edu/wiki/EEGLAB_Wiki and video materials from workshops make it possible to start use it in practice in several days. Also EEGlab have over 20 specific plugins like ERPLab for event-related potentials (Lopez-Calderon and Luck, 2014), BCILab for brain computer interface design and analyzes, NFT – 3D-Head and source location modeling, MobiLab - Mobile brain/body imaging (MoBI) and others (Delorme et al., 2011) R: is a free software environment for statistical computing and graphics. R open source project and freely available for several operational systems on site https://www.r-project.org. It is including linear, nonlinear methods, statistical tests, classical and Bayesian statistics methods, classification, clustering, machine learning libraries. Packages for R develop and maintain many groups and individual researchers. You can find them on https://cran.r-project.org/web/packages, each package usually contains code, documentation and some example data arrays. Several packages were developed for EEG study - eegkit, icaOcularCorrection, eegAnalysis. For modern methods of data analysis good introduction in R (James et al., 2013). Book accompanied on-line MOOC course "Statistical Learning" at Lagunita, Stanford, https://lagunita.stanford.edu. Authors tell about linear and polynomial regression, logistic regression and linear discriminant analysis, cross-validation and the bootstrap, model selection and regularization methods, nonlinear models, splines and generalized additive models, tree-based methods, random forests and boosting, support-vector machines and other methods. For introduction and reference in programming language of R see (Matloff, 2011), (Wickham, 2014). page 11 of 14 libraries (http://www.enthought.com), Python(x,y) Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Python: is a script programming language known as one of scientific research tools. It is free open-source project that has many implementations and huge community of developers. Many Linux distributives have it as installed part. Python central coordination site https://www.python.org, on-line tutorials are available there, it also has big library depository. Python scientific stack SciPy (http://www.scipy.org) you can get with distributions like Anaconda (http://store.continuum.io/cshop/anaconda), (http://python- Enthought Canopy xy.github.io), they also have reach collection of popular libraries for data analyses and complex computations. For introduction to data analysis with Python see (McKinney, 2012). It covers Ipython and NymPy libraries, Pandas data structures, data manipulations and aggregation, visualization, some beginning time-series analyzing techniques. For EEG analyze Python have like MNE-Python (http://martinos.org). Some of its features: raw data visualization, epoching, averaging, removing artifacts with SSP projections, ICA, forward modeling, linear inverse solving, sparse inverse solvers. For example of use MNE-Python in EEG analysis area see (Gramfort et al., 2013). Another library is PyEEG (http://pyeeg.org) have some interesting functions: power spectrum density, petrosian fractal dimension, higuchi fractal dimension, hjorth mobility and complexity, hurst exponent, detrended fluctuation analysis and others. Example of use that packet in practice in paper (Bao, Liu, and Zhang, 2011). Weka: is a stack of over 100 machine learning algorithms for data mining. It is a free open-source Java application developed at the University of Waikato, New Zealand (http://www.cs.waikato.ac.nz/ml/weka). It can solve typical for machine learning groups of task: clustering, classification, regression, attribute selection, finding association rules and others. A companion book for the Weka software is (Witten et al., 2011), University of Waikato also has a MOOC about using Weka for practical tasks (https://weka.waikato.ac.nz/dataminingwithweka). Java cross-platform object- oriented language can be powerful tool for implementation user defined algorithms. Weka have 3 different GUI which can help you design configurations for data processing, change parameters of algorithm and handle statistical tests. All its features give the Weka a low bar for starting practical data mining. For examples of use Weka for EEG analysis see (AlZoubi et al., 2009), (Sohaib, Qureshi, Hagelbäck, Hilborn, 2013). Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Resume page 12 of 14 Modern statistical methods and its computer implementations confirm ecosystems of concepts, algorithms and tools which form a several layers structure. Each level do nоt have strict borders and interrelate between themselves. This system always grows with new researches in mathematics, computer science and statistics. For EEG analysis we may need an orienteer to navigate in this growing and complex structure. In our paper we overview and classify modern branches of analysis. We hope it can help avoid a "analytic's paralyze" for researchers who is a new in this field of science. References A.-T. Sohaib, S. Qureshi, J. Hagelbäck, O. Hilborn, P. J. (2013). Evaluating Classifiers for Emotion Recognition Using EEG. In HCII 2013 (pp. 492–501). AlZoubi, O., Calvo, R. A., & Stevens, R. H. (2009). Classification of EEG for affect recognition: An adaptive approach. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 5866 LNAI, pp. 52–61). http://doi.org/10.1007/978-3-642- 10439-8_6 Bao, F. S., Liu, X., & Zhang, C. (2011). PyEEG: An Open Source Python Module for EEG/MEG Feature Extraction. Computational Intelligence and Neuroscience, 2011, e406391. http://doi.org/10.1155/2011/406391 Buzsáki, G. (2006). Rhythms of the Brain. Rhythms of the Brain (Vol. 1). http://doi.org/10.1093/acprof:oso/9780195301069.001.0001 Cohen, M. (2014). Granger Prediction. In Analyzing Neural Time Series Data: Theory and Practice (pp. 371–389). De Boor, C. (1972). On calculating with B-splines. Journal of Approximation Theory, 6(1), 50–62. http://doi.org/10.1016/0021-9045(72)90080-9 Delorme, A., & Makeig, S. (2004). EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of Neuroscience 9–21. http://doi.org/10.1016/j.jneumeth.2003.10.009 Methods, 134(1), Delorme, A., Makeig, S., Debener, S., & Onton, J. (2004). Mining event-related brain 204–210. Cognitive Sciences, 8(5), dynamics. http://doi.org/10.1016/j.tics.2004.03.008 Trends in Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Delorme, A., Mullen, T., Kothe, C., Akalin Acar, Z., Bigdely-Shamlo, N., Vankov, A., & Makeig, S. (2011). EEGLAB, SIFT, NFT, BCILAB, and ERICA: New tools Intelligence and Neuroscience, 2011. http://doi.org/10.1155/2011/130714 for advanced EEG processing. Computational page 13 of 14 Demchenko, Y., Grosso, P., De Laat, C., & Membrey, P. (2013). Addressing big data issues in Scientific Data Infrastructure. Proceedings of the 2013 International Conference on Collaboration Technologies and Systems, CTS 2013, 48–55. http://doi.org/10.1109/CTS.2013.6567203 Edelsbrunner, H., & Harer, J. (2008). Persistent Homology a Survey. Contemporary Mathematics, 0000, 1–26. http://doi.org/10.1090/conm/453/08802 Gautschi, W. (2006). Orthogonal polynomials, quadrature, and approximation: Computational methods and software (in matlab). Lecture Notes in Mathematics, 1883, 1–77. http://doi.org/10.1007/978-3-540-36716-1_1 Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, D., Brodbeck, C., … Hämäläinen, M. (2013). MEG and EEG data analysis with MNE-Python. Frontiers in Neuroscience, 7. http://doi.org/10.3389/fnins.2013.00267 Höhne, J., Bartz, D., Hebart, M., Müller, K.-R., & Blankertz, B. (2015). Analyzing Neuroimaging Data with Subclasses: a Shrinkage Approach. NeuroImage. http://doi.org/10.1016/j.neuroimage.2015.09.031 Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., … Liu, H. H. (1998). The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proceedings of the Royal Society A: Mathematical, Sciences. http://doi.org/10.1098/rspa.1998.0193 Engineering Physical and James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning. Springer. Kruschke, J. K. (2011). Doing Bayesian Data Analysis: A Tutorial with R and BUGS. Europe's Journal of Psychology (Vol. 7). http://doi.org/10.5964/ejop.v7i4.163 Lantz, B. (2013). Machine learning with R. Packt Publishing Ltd. Lopez-Calderon, J., & Luck, S. J. (2014). ERPLAB: an open-source toolbox for the analysis of event-related potentials. Frontiers in Human Neuroscience, 8(April), 213. http://doi.org/10.3389/fnhum.2014.00213 Luck, S. J. (2005). An Introduction to the Event-Related Potential Technique. Monographs of the Society for Research in Child Development (Vol. 78). http://doi.org/10.1118/1.4736938 Makeig, S., & Onton, J. (2011). ERP Features and EEG Dynamics. In The Oxford 1–37). Handbook http://doi.org/10.1093/oxfordhb/9780195374148.013.0035 Potential Components Event-Related of (pp. Review of analytical instruments for EEG analysis, Agapov et al. February 2016 © IT Universe LLC, Samara, Russia Matloff, N. (2011). The art of R programming: A tour of statistical software design. page 14 of 14 No Starch Press. McKinney, W. (2012). Python for data analysis: Data wrangling with Pandas, NumPy, and IPython. O'Reilly Media, Inc. Nunez, P. L., & Srinivasan, R. (2006). Electric fields of the brain: the neurophysics of EEG. Oxford university press. Ramsay, J. O., & Dalzell, C. J. (1991). Some Tools for Functional Data Analysis. Journal of the Royal Statistical Society. Series B (Methodological), 53(3), pp. 539–572. Retrieved from http://www.jstor.org/stable/2345586 Ramsay, J. O., & Silverman, B. W. W. (2005). Functional Data Analysis (Springer Sciences. Encyclopedia Statistical of Series http://doi.org/10.1007/b98888 Statistics). in Sizemore, J., & Mueller, J. P. (2014). MATLAB for Dummies. John Wiley & Sons. Theodoridis, S., & Koutroumbas, K. (2008). Pattern Recognition, Fourth Edition. Publishing Research Quarterly (Vol. 11). http://doi.org/10.1007/BF02680460 Wickham, H. (2014). Advanced R. CRC Press. Witten, I. H., Frank, E., & Hall, M. A. (2011). Data Mining: Practical Machine Learning Tools and Techniques (Google eBook). Complementary literature None. from http://books.google.com/books?id=bDtLM8CODsQC&pgis=1 Retrieved Wood, S. N. (2006). Generalized additive models: an introduction with R. Wiley Interdisciplinary Reviews: Computational …, 16 (July). Retrieved from http://onlinelibrary.wiley.com/doi/10.1002/wics.10/full\nhttp://books.google.com /books?hl=en&lr=&id=GbzXe- L8uFgC&oi=fnd&pg=PP1&dq=Generalized+Additive+Models:+An+Introductio n+with+R&ots=BLOErV3m12&sig=fy8ER03EKVqUrVRWRdrhQkliAbA Zheludev, V. A. (1998). Periodic splines, harmonic analysis, and wavelets. Wavelet Analysis and Its Applications, 7(C), 477–509. http://doi.org/10.1016/S1874- 608X(98)80019-8
1305.4319
1
1305
2013-05-19T03:12:09
Multi-command Tactile Brain Computer Interface: A Feasibility Study
[ "q-bio.NC", "cs.HC" ]
The study presented explores the extent to which tactile stimuli delivered to the ten digits of a BCI-naive subject can serve as a platform for a brain computer interface (BCI) that could be used in an interactive application such as robotic vehicle operation. The ten fingertips are used to evoke somatosensory brain responses, thus defining a tactile brain computer interface (tBCI). Experimental results on subjects performing online (real-time) tBCI, using stimuli with a moderately fast inter-stimulus-interval (ISI), provide a validation of the tBCI prototype, while the feasibility of the concept is illuminated through information-transfer rates obtained through the case study.
q-bio.NC
q-bio
1 Multi-command Tactile Brain Computer Interface: A Feasibility Study Hiromu Mori1, Yoshihiro Matsumoto1, Victor Kryssanov2, Eric Cooper2, Hitoshi Ogawa2, Shoji Makino1, Zbigniew R. Struzik3,4, and Tomasz M. Rutkowski1,4,∗,† 1 Life Science Center of TARA, University of Tsukuba, Tsukuba, Japan 2 Ritsumeikan University, Shiga, Japan 3 University of Tokyo, Tokyo, Japan 4 RIKEN Brain Science Institute, Wako-shi, Japan ∗ E-mail: [email protected] Abstract The study presented explores the extent to which tactile stimuli delivered to the ten digits of a BCI-naive subject can serve as a platform for a brain computer interface (BCI) that could be used in an interactive application such as robotic vehicle operation. The ten fingertips are used to evoke somatosensory brain responses, thus defining a tactile brain computer interface (tBCI). Experimental results on subjects performing online (real-time) tBCI, using stimuli with a moderately fast inter-stimulus-interval (ISI), provide a validation of the tBCI prototype, while the feasibility of the concept is illuminated through information-transfer rates obtained through the case study. Keywords: tactile BCI, P300, robotic vehicle interface 1 Introduction Contemporary BCIs are typically based on mental visual and motor imagery prototypes, which require extensive user training and non-impaired vision of the users [13]. Recently alternative solutions have been proposed to make use of spatial auditory [1, 10] or tactile (somatosensory) modalities [5, 6, 12] to enhance brain-computer interface comfort and increase the information transfer rate (ITR) achieved by users. The concept proposed in this paper uses the brain somatosensory (tactile) channel to allow targeting of the tactile sensory domain for the operation of robotic equipment such as personal vehicles, life support ∗The corresponding author. †The final publication is available at link.springer.com 2 systems, etc. The rationale behind the use of the tactile channel is that it is normally far less loaded than visual or even auditory channels in such applications. One of the first reports [6] of the successful employment of steady-state somatosensory responses to create a BCI targeted a low frequency vibrotactile stimulus in the range of 20 − 31 Hz to evoke the subjects' attentional modulation, which was then used to define interfacing commands. A recent report [12] proposed using a Braille stimulator with 100 ms static push stimuli delivered to each of six fingers to evoke a somatosensory response potential (SEP) related P300. The P300 response is a positive electroencephalogram event-related potential (ERP) deflection starting at around 300 ms and lasting for 200 − 300 ms after an expected stimulus in a random series of distractors (the so-called oddball EEG experimental paradigm) [7]. P300 responses are commonly used in BCI approaches and are considered to be the most reliable ERPs [9, 13] with untrained subjects. The results indicated that the experiments achieved information transfer rates of 7.8 bit/min on average and 27 bit/min for the best subject. This paper proposes a novel tactile brain-computer interface based on P300 responses evoked by tactile stimuli delivered via vibrotactile exciters attached to the ten fingertips of the subject's hands. The rest of the paper is organized as follows. The next section introduces the materials and methods used in the study and also outlines the experiments conducted. The results obtained in psychophysical and electroencephalogram experiments with eleven BCI-naive subjects are then discussed. Finally, conclusions are formulated and directions for future research are outlined. 2 Materials and Methods Eleven paid BCI-naive subjects (ten males and one female) participated in the experiments. The subjects' mean age was 21.82, with a standard deviation of 0.87. All the experiments were performed at the Life Science Center of TARA, University of Tsukuba, Japan. The psychophysical and online (real-time) EEG tBCI prototype experiments were conducted in accordance with the WMA Declaration of Helsinki - Ethical Principles for Medical Research Involving Human Subjects. 2.1 Tactile Stimuli The tactile stimuli were delivered as sinusoidal waves generated by a portable computer with MAX/MSP software [2]. The stimuli were generated via ten channel outputs (one for each fingertip of the subject) of 3 an external digital-to-analog signal converter MOTU UltraLite-mk3 Hybrid coupled with three YAMAHA P4050 power amplifiers (four acoustic frequency channels each). The stimuli were delivered to the subjects' fingertips via the tactile exciters HiWave HIAX25C10-8/HS working in the range of 100 − 20, 000 Hz, as depicted in Figure 1. Each exciter in the experiments was set to emit a sinusoidal wave at 200 Hz to match the exciter's resonance frequency and to stimulate the Pacini endings (fast-adapting type II afferent type tactile sensory hand innervation receptors) [3]. The subjects placed their fingertips on the exciters (see Figure 1) and attended only to the instructed locations (with a button-press response in the psychophysical experiments, or a mental count of the targets in the EEG experiments). The training instructions were presented visually by means of the MAX/MSP program, as depicted in Figure 2. 2.2 Psychophysical Experiment Protocol The psychophysical experiment was conducted to investigate the stimulus carrier frequency influence on the subjects' response time and accuracy. The behavioral responses were collected using a small numeric keypad and the MAX/MSP program. Each subject was instructed which stimulus to attend to by a cross above the target fingertip (T ARGET ), shown by the program (Figure 2). Then the subject pressed the response button with the dominant foot. In the experiment, each subject was presented with 50 T ARGET S and 450 non − T ARGET S as stimuli. Each trial was composed of a randomized order of 100 ms tactile bursts delivered to each fingertip separately with an inter-stimulus-interval (ISI) of 600 ms. Every random sequence thus contained a single T ARGET and nine non − T ARGET S. A single session included five trials for each T ARGET fingertip (resulting in 50 T ARGET S and 450 non− T ARGET S). The choice of the relatively long ISI is justified by slow behavioral responses in comparison to EEG evoked potentials [7], as described in the next section. The response time delays were registered with the same MAX/MSP program, also used for the stimulus generation and instruction presentation. 2.3 EEG tBCI Experiment The exciters were attached to the fingertips in the same manner (see Figure 1). During the experiment, EEG signals were captured with a portable wireless EEG amplifier system g.MOBllab+ and g.SAHARA by g.tec, using eight dry electrodes. The electrodes were attached to the head locations: Cz, CPz, P3, 4 P4, C3, C4, CP5, and CP6, as in the 10/10 extended international system [4] (see the topographic plot in the top panel of Figure 3). The ground and reference electrodes were attached behind the left and right ears respectively. In order to limit electromagnetic interference, the subjects' hands were additionally grounded with armbands connected to the amplifier ground. No electromagnetic interference was observed from the exciters. Details of the EEG experimental protocol are summarized in Table 1. The recorded EEG signals were processed by a BCI2000-based application [9], using a stepwise linear discriminant analysis (SWLDA) classifier [8] with features drawn from the 0− 700 ms ERP interval. The sampling rate was set at 256 Hz, the high pass filter at 0.1 Hz, and the low pass filter at 40 Hz. The ISI was 400 ms, and each stimulus duration was 100 ms. The subjects were instructed to spell out the number sequences (corresponding to the interactive robotic application commands shown in Table 2) communicated by the exciters in each session. Each T ARGET was presented five times in a single command trial, and the averages of the five ERPs were later used for the classification. Each subject performed three experimental sessions (randomized 50 T ARGET S and 450 non − T ARGET S each), which were later averaged as discussed in Section 3. 3 Results This section discusses the results obtained in the psychophysical experiment and in the EEG experiment. 3.1 Psychophysical Experiment Results The psychophysical experiment results are summarized in Figure 4, where the median response time and the range are depicted for each fingertip as boxplots (see also Figure 2 for the fingertip numbering). The Wilcoxon rank sum tests for pairwise comparisons revealed no differences (at the 0.05 level) among the median values for all the fingertip pairs of each subject. This result confirms the stimulus similarity since the behavioral responses for all the fingers were basically the same. This finding validates the design of the EEG experiment. 3.2 EEG Experiment Results The results of the EEG experiment are summarized in Table 3 and Figure 3. All eleven BCI-naive subjects scored well above the chance level of 10%, reaching an ITR in the range from 0.19 bit/min 5 to 4.46 bit/min, which may be considered to be a good outcome for experiments with beginners (naive subjects). The ITR was calculated as follows [10]: IT R = V · R R = log2N + P · log2P + (1 − P ) · log2 (cid:19) (cid:18) 1 − P N − 1 , (1) (2) where R stands for the number of bits/selection; N is the number of classes (10 in this study); P is the classifier accuracy (see Table 3); and V is the classification speed in selections/minute (3 selections/minute for this study). The maximum ITR it was possible for the BCI-naive subjects to achieve in the settings presented was 9.96 bit/min. 4 Conclusions This paper reports results obtained with a novel ten-command tBCI prototype developed and used in experiments with eleven BCI-naive subjects. The proposed interface could be used for real-time operation of robotic vehicles. The experiment results obtained in this study confirmed the general validity of the tBCI for interactive applications. In the psychophysical experiment, it is shown that all the tested fingertip zones are equally sensitive to the stimuli and all can be used with the tBCI prototype. The EEG experiment with the prototype has confirmed that tactile stimuli can be used to operate robotic devices with up to ten commands and with the command interfacing rate ranging from 0.19 bit/min to 4.46 bit/min for untrained users. The results presented offer a step forward in the development of neurotechnology applications. Due to the still not very practical interfacing rate achieved, allowing for only about three commands per minute, the current prototype would obviously need improvements and modifications. These needs determine the major lines of study for future research. However, even in its current form, the proposed tBCI can be regarded as a practical solution for totally-locked-in (TLS) patients, who cannot use vision or auditory based interfaces due to sensory or other disabilities. 6 Acknowledgments. This research was supported in part by the Strategic Information and Communications R&D Promotion Program no. 121803027 of The Ministry of Internal Affairs and Communication in Japan, and by KAK- ENHI, the Japan Society for the Promotion of Science, grant no. 12010738. We also acknowledge the technical support of YAMAHA Sound & IT Development Division in Hamamatsu, Japan. References 1. Halder, S., Rea, M., Andreoni, R., Nijboer, F., Hammer, E., Kleih, S., Birbaumer, N., Kubler, A.: An auditory oddball brain-computer interface for binary choices. Clinical Neurophysiology 121(4) (2010) 516 -- 523 2. http://cycling74.com/: Max 6 (2012) 3. Johansson, R.S., Flanagan, J.R.: Coding and use of tactile signals from the fingertips in object manipulation tasks. Nature Reviews Neuroscience 10(5) (2009) 345 -- 359 4. Jurcak, V., Tsuzuki, D., Dan, I.: 10/20, 10/10, and 10/5 systems revisited: Their validity as relative head-surface-based positioning systems. NeuroImage 34(4) (2007) 1600 -- 1611 5. Mori, H., Matsumoto, Y., Makino, S., Kryssanov, V., Rutkowski, T.M.: Vibrotactile stimulus frequency optimization for the haptic BCI prototype. In: Proceedings of The 6th International Conference on Soft Computing and Intelligent Systems, and The 13th International Symposium on Advanced Intelligent Systems, Kobe, Japan (November 20-24, 2012) 2150 -- 2153 6. Muller-Putz, G., Scherer, R., Neuper, C., Pfurtscheller, G.: Steady-state somatosensory evoked potentials: suitable brain signals for brain-computer interfaces? Neural Systems and Rehabilitation Engineering, IEEE Transactions on 14(1) (March 2006) 30 -- 37 7. Niedermeyer, E., Da Silva, F.L., eds.: Electroencephalography: Basic Principles, Clinical Applica- tions, and Related Fields. 5th ed. Lippincott Williams & Wilkins (2004) 8. Potes, C.M.: P300 Classifier (2009) 7 9. Schalk, G., Mellinger, J.: A Practical Guide to Brain-Computer Interfacing with BCI2000. Springer-Verlag London Limited (2010) 10. Schreuder, M., Blankertz, B., Tangermann, M.: A new auditory multi-class brain-computer inter- face paradigm: Spatial hearing as an informative cue. PLoS ONE 5(4) (04 2010) e9813 11. Theodoridis, S., Koutroumbas, K.: Pattern Recognition. Fourth ed. Academic Press (2009) 12. van der Waal, M., Severens, M., Geuze, J., Desain, P.: Introducing the tactile speller: an ERP- based brain-computer interface for communication. Journal of Neural Engineering 9(4) (2012) 045002 13. Wolpaw, J.R., Wolpaw, E.W., eds.: Brain-Computer Interfaces: Principles and Practice. Oxford University Press (2012) 8 Figure Legends Figure 1 The experimental set-up. Tactile stimuli are delivered to the fingertips by ten HIAX25C10-8 exciters. The set-up was used in both the psychophysical and the EEG experiments. Figure 2 The visual instruction screen presented to the subjects during the psychophysical experiment. Each fingertip was assigned a number encoding a command in the interactive application. The controls on the right side were used by the subject to adjust the stimulus intensity (with a fader), and also by the experimenter to start the experiment (the button with the speaker launches the digital-to-analog signal converter MOTU UltraLite-mk3 Hybrid) and to save/clear the results. Figure 3 Grand mean averaged results of the fingertip stimulation EEG experiment for all 11 subjects. The top panel presents the head topographic plot of the T ARGET versus non − T ARGET area under the curve (AUC), a measure commonly used in machine learning intra-class discriminative analysis. (AU C > 0.5 is usually assumed to be confirmation of feature separability [11]). The top panel presents the largest difference as obtained from the data displayed in the bottom panel. The topographic plot also depicts the electrode positions. The fact that all the electrodes received similar AUC values (red) supports the initial electrode placement. The second panel from the top presents averaged SEP responses to the T ARGET stimuli (note the clear P300 response in the range of 450 − 700 ms). The third panel presents averaged SEP responses to the non − T ARGET stimuli (no P300 observed). Finally, the bottom panel presents the AUC of T ARGET versus non − T ARGET responses (again, P300 could easily be identified). Figure 4 Results in boxplots (note the overlapping quartiles visualizing no significant differences between medians) of the grand mean averages (11 subjects) of the psychophysical experiment response time delays. Each number represents a finger, as depicted in Figure 1. The dots represent outliers. Tables 9 Table 1. Conditions of the EEG experiment. Number of subjects Tactile stimulus length Stimulus frequency Inter-stimulus-interval (ISI) 11 100 ms 200 Hz 400 ms EEG recording system g.SAHARA & g.MOBIlab+ active dry EEG electrodes system. Number of the EEG channels 8 EEG electrode positions Cz, CPz, P3, P4, C3, C4, CP5, CP6. Reference and ground electrodes Behind both of the subject's ears Stimulus generation 10 HIAX25C10-8 exciters Number of trials for each subject 5 Table 2. Interactive application commands encoded with the finger numbers (see also Figure 1). 10 Finger number Command 1 2 3 4 5 6 7 8 9 10 low speed medium speed high speed stop slow speed reverse go left (−90◦) go straight -- left (−45◦) go straight (0◦) go straight -- right (45◦) go left (90◦) Table 3. The fingertip stimulation EEG experiment accuracy and ITR scores. The theoretical chance level was 10%. For the classifier, features were derived from the averages of the five ERPs of all the subjects. Subject number Maximum accuracy ITR 11 1 2 3 4 5 6 7 8 9 10 11 20% 40% 50% 30% 30% 40% 70% 40% 20% 30% 20% 0.19 bit/min 1.34 bit/min 2.21 bit/min 0.66 bit/min 0.66 bit/min 1.34 bit/min 4.46 bit/min 1.34 bit/min 0.19 bit/min 0.66 bit/min 0.19 bit/min Figures 12 Figure 1. The experimental set-up. Tactile stimuli are delivered to the fingertips by ten HIAX25C10-8 exciters. The set-up was used in both the psychophysical and the EEG experiments. 13 Figure 2. The visual instruction screen presented to the subjects during the psychophysical experiment. Each fingertip was assigned a number encoding a command in the interactive application. The controls on the right side were used by the subject to adjust the stimulus intensity (with a fader), and also by the experimenter to start the experiment (the button with the speaker launches the digital-to-analog signal converter MOTU UltraLite-mk3 Hybrid) and to save/clear the results. 14 Figure 3. Grand mean averaged results of the fingertip stimulation EEG experiment for all 11 subjects. The top panel presents the head topographic plot of the T ARGET versus non − T ARGET area under the curve (AUC), a measure commonly used in machine learning intra-class discriminative analysis. (AU C > 0.5 is usually assumed to be confirmation of feature separability [11]). The top panel presents the largest difference as obtained from the data displayed in the bottom panel. The topographic plot also depicts the electrode positions. The fact that all the electrodes received similar AUC values (red) supports the initial electrode placement. The second panel from the top presents averaged SEP responses to the T ARGET stimuli (note the clear P300 response in the range of 450 − 700 ms). The third panel presents averaged SEP responses to the non − T ARGET stimuli (no P300 observed). Finally, the bottom panel presents the AUC of T ARGET versus non − T ARGET responses (again, P300 could easily be identified). Topographical plot of AUC scores at 500 ms [µV]Averaged targets 0 100 200 300 400 500 600 700 800 CzCPzP3P4C3C4CP5CP6[µV]Averaged non−targets 0 100 200 300 400 500 600 700 800 CzCPzP3P4C3C4CP5CP6time [ms]AUCTargets vs. non−targets AUC scores 0 100 200 300 400 500 600 700 800 CzCPzP3P4C3C4CP5CP60.480.50.520.54−2−1012−2−10120.480.50.520.54 15 Figure 4. Results in boxplots (note the overlapping quartiles visualizing no significant differences between medians) of the grand mean averages (11 subjects) of the psychophysical experiment response time delays. Each number represents a finger, as depicted in Figure 1. The dots represent outliers.
1805.06336
1
1805
2018-04-25T12:52:50
Characterizing Information Propagation in Plants
[ "q-bio.NC", "cs.IT", "cs.IT", "q-bio.TO" ]
This paper considers an electro-chemical based communication model for intercellular communication in plants. Many plants, such as Mimosa pudica (the "sensitive plant"), employ electrochemical signals known as action potentials (APs) for communication purposes. In this paper we present a simple model for action potential generation. We make use of the concepts from molecular communication to explain the underlying process of information transfer in a plant. Using the information-theoretic analysis, we compute the mutual information between the input and output in this work. The key aim is to study the variations in the information propagation speed for varying number of plant cells for one simple case. Furthermore we study the impact of the AP signal on the mutual information and information propagation speed. We aim to explore further that how the growth rate in plants can impact the information transfer rate and vice versa.
q-bio.NC
q-bio
Characterizing Information Propagation in Plants Hamdan Awan:, Raviraj S. Adve♦, Nigel Wallbridge;, Carrol Plummer;, and Andrew W. Eckford: :Dept. of EECS, York University, Toronto, Ontario, Canada ♦The Edward S. Rogers Sr. Dept. of ECE, University of Toronto, Toronto, Ontario, Canada ;Vivent SaRL, Crans-pr`es-C´eligny, Switzerland Corresponding author, email: [email protected] 8 1 0 2 r p A 5 2 ] . C N o i b - q [ 1 v 6 3 3 6 0 . 5 0 8 1 : v i X r a Abstract -- This paper considers an electro-chemical based communication model for intercellular communication in plants. Many plants, such as Mimosa pudica (the "sensitive plant"), employ electrochemical signals known as action potentials (APs) for communication purposes. In this paper we present a simple model for action potential generation. We make use of the con- cepts from molecular communication to explain the underlying process of information transfer in a plant. Using the information- theoretic analysis, we compute the mutual information between the input and output in this work. The key aim is to study the variations in the information propagation speed for varying number of plant cells for one simple case. Furthermore we study the impact of the AP signal on the mutual information and information propagation speed. We aim to explore further that how the growth rate in plants can impact the information transfer rate and vice versa. I. INTRODUCTION Recent work in biological literature suggests that electrical and electromagnetic communication in higher organisms is worth investigating. Action potentials (APs) are electrochem- ical signals in biological communication systems. Though commonly associated with the firing of neurons, APs also play a significant role in plants. For example, Mimosa pudica, the "sensitive plant", closes its leaves when touched: the signal to close the leaves is carried by an AP, as proposed by Bose over a century ago [1]. AP signals in plant can be defined as a sudden change or increase in the resting potential of the cell as a result of an external stimulus [2]. Some mathematical models for AP generation are presented in literature such as [2], [3]. In this work we focus on the electrical AP signal generation in plants and its impact on information propagation through chemical molecules. In this work we present a simple general model of an AP generation in plants. It is clear from the models [2], [3], [4] that understanding of APs is informed by molecular communication [5], [6], a communication paradigm inspired by the communication between living cells [7], [8], [9]. In this paper, we consider diffusion-based molecular communication of signalling molecules (as a result of AP signal) through a fluid medium. We will focus on the mutual information where the receiver is based on chemical reactions i.e. ligand-receptor binding. Furthermore in this paper we compute the mutual information between the input (number of signaling molecules) and the output number of molecules produced by a number of receiver cells in series. We make two main contributions. First we study the impact of growth rate on the information propagation speed in the system. The information propagation speed can be defined as a measure of how fast the information propagates from transmitting to receiving cells. To the best of our knowledge there is a limited study about the impact of increasing length of the chain of cells on the information propagation speed. In this paper we use the mutual information for different number of receiver cells in series and compute the information propagation speed by selecting a suitable threshold. We show that, in general, an increase in the number of cells in the chain results in an increase in information propagation speed. Secondly we study the impact of AP signal on the mutual information and information propagation speed by comparing with the case when we have no AP signal. This paper is organized as follows. We describe the system model in Section II. We present transmitter, action potential generation and voxel model for propagation in subsection II-A. Next we present the diffusion only system in subection II-B. This is followed by the modelling of the receiver in subsection II-C. Section III presents a model for the complete system. The expressions for mutual information and the information propagation speed are derived in Section III-A. Next in Section IV we present the results for mutual information, the infor- mation propagation speed for varying number of receivers in series and the impact of AP signal on mutual information and propagation speed. Section V presents the conclusion. II. SYSTEM MODEL In this work we consider a communication link which consists of a sensing/transmitter cell and a number of receiver cells in one series as shown in Figure 1. In the transmitter cell the AP signal is generated due to an external stimulus such as a change in temperature or electrical signal. As a result of this stimulus the transmitter cell emits an increased number of signalling molecules (as compared to the absence of AP signal) which diffuse freely in the propagation medium. The increased number of signalling molecules emitted by the transmitter cell is proportional to the action potential signal strength. The signalling molecules propagate through the medium to the receiver cells where they react with the receptors to produce output molecules. The number of output molecules in the receiver over time is the output signal of the communication link. Our aim is to use the mutual information of this commu- nication link and use this to obtain the information propagation speed. Fig. 1. System Model A. Transmitter, Action Potential and Voxel Model 1) Sensing/Transmitter Cell: In a typical plant cell, there is a potential difference across the cell membrane known as resting potential. The generation of AP is associated with passive fluxes of ions in the cell. As as a result of external factors such as an electrical signal or change in temperature, the resting potential increases to certain threshold causing ion- channels in the outer cell membrane to open up and resulting in a flow of ions into the cell. The typical ions are calcium (Ca`2), chlorine (Cl´) and potassium (K`). This results in increasing the resting potential value of the cell membrane. The AP signal is generated when this resting potential crosses a certain threshold value. Once this AP signal is generated in the system it triggers the release of additional signalling molecules from the transmitting cell. 2) Action Potential Generation: Simple Model: Let ER represent the resting potential of the cell. The new membrane potential Em as a result of external stimulus causing the change in ion-concentrations is given in [2], [3] as: Em " gkEk ` gclEcl ` gcaEca gk ` gcl ` gca gi " F hi ER ´ Ei where, (1) (2) Note that for simplicity we use this equation to represent the change in resting potential Em (of the transmitting/sensing cell) when the ion channels are open. A detailed discussion about the AP generation and its mathematical model will be presented in a journal version of this work. gi represents the electrical conductivity and Ei represents the resting potential value for ion channel i, i.e. Ek and gk for potassium (K) channel etc. Furthermore the term F represents Faraday's constant and hi, the ion flow across the membrane which is given as: hi " zµPmpo φiηo ´ φoηipexpp´zµqq 1 ´ expp´zµq (3) where z is ion charge. The terms φi and φo (respectively ηi and ηo) represent the probability that ion is (respectively not) linked to the channel inside and outside. Pm represents Fig. 2. Propagation Medium the maximum permeability of the cell membrane. The term µ denotes the normalized resting potential and is given as: µ " EmF RT (4) where R is the gas constant, and T is temperature. The term po represents the ion-channel opening state probability. For k1 (channel opening) and k2 (channel closing) reaction rate constants we obtain this as: " k1p1 ´ poq ´ k2ppoq dpo dt (5) The values of all these parameters are presented in Table II. The input of the system Uptq i.e. the number of signalling molecules emitted by the transmitter/sensing cell is given as: Uptq 9 Em (6) Where Em is given by Equation (1). Note that this Uptq acts as the system input in Section III. This relation means that in the event of an AP signal generation the transmitter emitts higher number of molecules as compared to no AP signal. To explain this we refer to Figure 1. Let us first assume the case when no AP signal is generated. In this case the number of molecules released at the cell membrane (to the neighbour cells) will depend only on the resting potential. We assume this as 5 molecules per second. However as the external stimulus generates an AP signal, the number of molecules released can increase upto 20 per second as the resting potential increases. 3) Voxel Model for Propagation: In this section we explain the voxel model for propagation of the signalling molecules from the sensing/transmitting cell to the receiver cell(s). The propagation occurs through the vascular bundles connecting different cells. We assume the medium of propagation is a three dimensional space of dimension (cid:96)X(cid:96)Y (cid:96)Z where each dimension is an integral multiple of length ∆ i.e. (cid:96)X " Mx∆, (cid:96)Y " My∆ and (cid:96)Z " Mz∆. The medium is divided into Mx My Mz cubic voxels where the volume of each voxel is ∆3 and it represents a single cell. Figure 2 shows an example with Mx = 4, My =1 and Mz = 1. For the ease of presentation we describe this 1-dimensional example. The transmitter and each receiver cell occupy a single voxel. The transmitter and receiver are assumed to be located, respectively, at the voxels with indices Tl " 2 and Rl " 4. The empty circles represent signaling or input molecules whereas the filled circles represent the output molecules. Diffusion is modelled as molecules moving from one voxel to a neighbouring voxel as shown by arrows in Figure 2. For this example we have assumed that the molecules released by the transmitter cell at voxel 2 towards voxel 3 cannot re-enter the transmitter. The diffusion takes place at a rate of d " D ∆2 where D is diffusion coefficient. This means that within an infinitesimal time δt, the probability that a molecule diffuses to a neighbouring voxel is dδt. For further details see [10]. B. Diffusion-Only SubSystem In this work we take the approach of dividing the sys- tem into two sub-systems, i.e., diffusion-only subsystem and reaction-only subsystem as shown in Figure 1. This section explains how to model the diffusion only system. Let nL,i denote the number of signaling molecules in the voxel i. In the absence of chemical reactions, the state of the system consists of only the number of signal molecules in each voxel i.e. nLptq " rnL,1ptq, nL,2ptq, nL,3ptq, nL,4ptqsT (7) where the superscript T in Eq. (7) denotes matrix transpose. The state of the system changes when a molecule diffuses from voxel 1 to a neighboring voxel 2 at a diffusion rate dnL,1. This event causes nL,1 to decrease by 1 and nL,2 to increase by 1. We can indicate this change by using the jump vector qd,1ptq " r´1, 1, 0, 0, 0sT . The state of system will be nLptq ` qd,1 after the occurrence of this diffusion event. We also specify the corresponding jump function Wd,1pnLptqq " dnL,1 which specifies the event rate. The molecules can escape at rate e. Let Jd be the total number of diffusion events, then we have Jd jump vectors qd,j and jump events Wd,jpnLptqq where j " 1, ..., Jd. Combining the jump vectors and jump rate functions of all the diffusion and escape events we obtain a matrix H for the medium as follows: » -- -- -- ´d qd,jWd,jpnLptqq ` Jdÿ 0 0 d ´d b The diffusion events are stochastic and hence modeled by using stochastic differential equation (SDE) [11] as follows: Wd,jpnLptqqγj 9nLptq " Jdÿ d d ´2d 0 0 0 0 ´d ´ e fiffiffifl H " 0 d d qd,j (8) j"1 ` 1T Uptq j"1 (9) Note this is a form of chemical Langevin equation and is similar to our previous works see [10]. There are three terms on the right-hand side of Eq. (9). The first term describes the deterministic dynamics. Since all the jump rates of all the diffusion events are linear, this term can be written as a product of a matrix H and the state vector nLptq. qd,jWd,jpnLptqq (10) HnLptq " Jdÿ j"1 Rv MATRIX FOR DIFFERENT RECEIVER CIRCUITS TABLE I Receiver Receiver Reactions  „ Rv Matrix ´k` k´ ´k´ k` stochastic noise in the system due to diffusion. For a more detailed explanation, the reader can refer to [10], [12]. The third term in Eq. (9) models the input from transmitter. Uptq from Eq. (6) denotes the transmitter emission rate at time t. This means, in the time interval rt, t`δtq the transmitter emits Uptqδt signalling molecules. C. Reaction Only Subsystem In this section, we present the stochastic differential equa- tion (SDE) governing the dynamics of a reaction-only sub- system. This subsystem includes the reactions of incoming signaling molecules (ligands) L from the transmitter with the receptors V in receiver to produce output molecules X as shown in Figure 1. The count of these output molecules over time is the output signal of the system. The reactions in the series of cells continue in the same way. However due to lack of space we present the reactions at first receiver cell only. Note that nL,R and nX denote, respectively, the number of signalling molecules in the receiver voxel and the output molecules. The scalar term nL,R differs from the vector nL which refers to the number of signalling molecules in all the voxels as shown in Equation (7). We use a simple receiver model (based on lineraized form of ligand receptor binding) which consists of following two linear chemical reactions: L Ñ X X Ñ L ´1 1 1 ´1 , k`nL,R , k´nX (11) (12) ‰ ‰ T T " " Each reaction is described by its chemical formula (on the left-hand side), and jump vector and jump rate (on the right-hand side). The symbols k` and k´ denote the reaction rate constants. In reaction (11) the signaling molecules react at rate k`nL,R to produce output molecules. The change in number of signaling and output molecules is indicated by jump vectors. Similarly we can understand the jump vector entries for reaction (12). We can model the reaction only system using stochastic differential equations for different receiver reactions. Note that is nL,R i.e the number of signaling molecules in receiver. The output of this subsystem is the number of output molecules nX. The state vector and SDE for the reaction only system are given as: T the input ‰ nL,Rptq nXptq " 9nRptq " Rv nRptq ` Jd`Jrÿ nRptq " b Wr,jpxnRptqyqγj (13) (14) qr,j j"Jd`1 The second term of Eq. (9) describes the stochastic dynamics. The term γj is continuous-time Gaussian white noise with unit power spectral density and it is needed to correctly model the Like the modeling of the diffusion only module we use jump vectors qr,j and jump rates Wr,j to model the reactions in this module. γj represents the continuous white noise. The where rqd,jsR denote the R-th element of the vector qd,j. The dynamics of the number of signaling molecules in the receiver voxel due to the reactions in the receiver is given by the first element of Eq. (14), which is: b ` Jd`Jrÿ 9nL,Rptq " R11nL,Rptq ` R12nXptq loooooooooooooooooomoooooooooooooooooon Wr,jpnRptqqγj rqr,js1 j"Jd`1 (16) ξrptq where rqr,js1 denotes the first element of the vector qr,j. For the complete system the dynamics of nL,Rptq is obtained by combining Eqs. (15) and (16) as follows: 9nL,Rptq "dnL,3ptq ´ dnL,Rptq ‰ ` R11nL,Rptq ` R12nXptq ` ξtotalptq (17) where ξtotalptq " ξdptq` ξrptq. We are now ready to describe the complete model. Let nptq be the state of the complete system and it is given by: nptq " " 9nptq " Anptq ` Jÿ „ (18) We use qj and Wjpnptqq to denote the jump vectors and jump rates of the combined model. The complete system SDE is: T nLptqT nXptq b Wjpnptqqγj ` 1T Uptq qj i"1 (19) ř where J " Jd ` Jr, and the matrix A is defined by Anptq " i"1 qjWjpnptqq. The input Uptq depends on Em as shown in Equation (6). The matrix A has the block structure: J  A " H ` 1T R R211T R 1RR11 1RR12 R22 (20) where H comes from the diffusion only subsystem. Similarly the Rii terms come from the reaction only subsystem (Rv matrix). The vector 1R is a unit vector with an 1 at the R-th RnLptq " nL,Rptq which is position; in particular, note that 1T the number of signalling molecules in the receiver voxel. Note that, the coupling between the diffusion-only subsystem and the output module, as exemplified by Eq. (17), takes place at the R-th row of A. Next we compute the mutual information using the Laplace transform of expression in Eq. (19). A. Mutual Information and Capacity The input and output signals for the complete system are, respectively, the production rate Uptq of the signalling molecules in the transmitter voxel (dependent on AP signal) and the number of output molecules nXptq in the receiver voxel. In this section, we will derive an expression for the mutual information between the input Uptq and output nXptq. We begin by stating a result in [13] which states that, for two Gaussian distribution random processes aptq and bptq, their mutual information Ipa, bq is: ż 8 1 ´ Φabpωq2 ΦaapωqΦbbpωq dω (21) Ipa, bq " ´1 4π log ´8 Fig. 3. Effect of AP Signal reactions are indexed from Jd ` 1 to Jd ` Jr where Jd is for the diffusion only module and Jr represents the reactions in the receiver. We define the matrix Rv as a 22 matrix and its entries depend on the reactions of signaling molecules in the receiver as given in Table I. In the next section we combine the diffusion only and reaction only modules to obtain a diffusion- reaction combined system. III. DIFFUSION-REACTION COMBINED SYSTEM In this section, we will combine the SDE models for the diffusion-only subsystem and the reaction-only subsystem to form the complete system model. Note that the number of signaling molecules nL,Rptq appears in the state vectors nLptq and nRptq of both diffusion only and receiver only modules respectively. Therefore, the interconnection between the diffusion-only subsystem and the output module is the number of signaling molecules in the receiver voxel, which is common. To illustrate our approach in this section, as an example consider the case when we have a single transmitter cell and three receiver cells; see Figure 3. We compare two cases here (a) diffusion of molecules from a transmitter to receivers in the presence of an AP signal and (b) diffusion of molecules in the absence of an AP signal. As shown in the Figure the presence of an AP signal increases the input signalling molecules and hence the output number of molecules. In later sections we present the simulation results to compare the mutual information and information propagation speed for both these cases. We will use the example in Figure 2 to explain how the diffusion-only subsystem and the reaction only system can be combined together. We consider the dynamics of the diffusion- only subsystem for the example, when the receiver voxel has the index R " 4. The evolution of the number of signaling molecules in the receiver voxel nL,Rptq is given by the R-th (i.e. fourth) row of Eq. (9) i.e.: b ` Jdÿ 9nL,Rptq " dnL,3ptq ´ dnL,Rptq loooooooooooooooomoooooooooooooooon Wd,jpnLptqqγj rqd,jsR j"1 ξdptq (15) 12341234AP Signal(b) Without AP Signal(a) With AP Signal qj i"1 (22) where Φaapωq (resp. Φbbpωq) is the power spectral density of aptq (bptq), and Φabpωq is the cross spectral density of aptq and bptq. In order to apply the above results to the communication link given in Eq. (19), we need a result from [14] on the power spectral density of systems consisting only of chemical reactions with linear reaction rates. Following from [14] if all the jump rates Wjpnptqq in Eq. (19) are linear in nptq, then the power spectral density of nptq is obtained by using following: 9nptq " Anptq ` Jÿ b Wjpxnp8qyqγj ` 1T Uptq where xnptqy denotes the mean of nptq and is the solution to the following ordinary differential equation: 9xnptqy " Axnptqy ` 1T c (23) where c is the mean of input Uptq. s a result, the dynamics of the complete system in Eq. (22) are described by a set of linear SDE with Uptq as the input and nXptq (which is the last element of the state vector nptq) as the output. The input Uptq has the form Uptq " c ` wptq where c (mean of input) depends on Em and wptq is a zero-mean Gaussian random process. The noise in the output nXptq is caused by the Gaussian white noise γj's in Eq. (22). Therefore, Eq. (22) models a continuous-time linear time-invariant (LTI) stochastic system subject to Gaussian input and Gaussian noise. The power spectral density ΦXpωq of the signal nXptq can be obtained from standard results on the output response of a LTI system to a stationary input and is given by: ΦXpωq " Ψpωq2Φepωq ` Φηpωq looooooooomooooooooon xNXpsqy " 1XxNpsqy " 1XpsI ´ Aq´11T (24) where Φepωq is the power spectral density of Eptq and Ψpωq2 is the channel gain with Ψpωq " Ψpsqs"iω defined by: Upsq (25) Ψpsq Note that Eq. (25) can be obtained from Eq. (22) after taking the mean and applying Laplace transform The transfer function Ψpsq takes into account the consumption of signaling molecules, the interaction between output molecules and the signaling molecules, as well as the possibility that a signaling molecule may leave or return in the receiver. For details see [10]. The term Φηpωq denotes the stationary noise spectrum and is given by: 1XpiωI ´ Aq´1qj2Wjpxnp8qyq (26) Φηpωq " Jd`Jrÿ j"1 where nptq denotes the state of the complete system in Eq. (18) and xnp8qy is the mean state of system at time 8 due to constant input c. Similarly, by using standard results on the LTI system, the cross spectral density Ψxepωq is: Ψxupωq2 " Ψpωq2Φepωq2 (27) PARAMETERS AND THEIR DEFAULT VALUES. TABLE II Symbols ER F C Pm µ γ φi φo ηi ηo cin and cout z po Notation and Value Resting Potential = -150-170 mV Faraday's constant = 9.65 104C{mol Membrane capacity = 10´6F cm´2 Permeability per unit area = 10´6 M cm s´1 ER F/ RT where T= Temperature ratio of rate constants = 9.9 10´5M Probability ion link - inside = cin / (cin + γ) Probability ion link - outside = cout / (cout + γ) Probability ion not linked-inside = 1- φi Probability ion not linked-outside = 1- φo 1.28 and 1.15 respectively ion charge e.g. for calcium = +2. ion channel open-state probability By substituting Eq. (24) and Eq. (27) into the mutual informa- tion expression in Eq. (21), we arrive at the mutual information IpnX , Uq between Uptq and nXptq is: 1 ` Ψpωq2 Φηpωq Φepωq IpnX , Uq " 1 2 (28) ż log dω The capacity or maximum mutual information of the com- munication link can be determined by applying the water- filling solution to Eq. (28) subject to a power constraint on input Uptq [15]. B. Information rate vs length of cell chain In this section we discuss how we use of the mutual information to obtain the relationship between information propagation speed and number of cells in the chain. First we obtain the mutual information when we have number of receiver cells in series. The next step is to choose a suitable threshold value so that we can calculate the time difference at which the mutual information curve for each case crosses the threshold value. Next we use the following equation for calculating the information propagation speed V (cells/sec): V " 1 Er∆ti,i`1s (29) Where ∆ti,i`1 represents the time difference at which the mutual information for each case (i.e. increasing receivers) crosses the threshold value. E denotes the expectation opera- tor. This technique is used to compute the propagation speed for an increasing number of receiver cells in the chain in series. We present the results for this approach in numerical examples section. IV. NUMERICAL EXAMPLES- SIMULATIONS In this section we discuss the numerical results related to this work. The parameters used for the generation of AP signal are given in Table II. The magnitude of the generated AP signal can be typically in the range of 20-80 mV. The AP signal generation results are not included due to limited space. We present these results in journal extension of this paper. For this system we obtain an action potential signal of about 60 millivolts which will trigger the release of signaling tion propagation speed for the parallel receiver case will be discussed in the journal extension of this paper. V. CONCLUSION In this paper we presented a simple model for the generation of action potential signal in plants. We realize the information transfer from a transmitter cell to a number of receiver cells in series and computed the mutual information between the input signal from transmitting cell and output signal of the receiver cell(s). By using the values of mutual information and selecting a threshold we obtained the information propagation speed as a function of the number of cells in the chain. We realize that the information propagation speed tends to increase with an increasing number of receiver cells in series. We further show that the presence of an AP signal leads to an increase in mutual information and information propagation speed for an increasing number of receiver cells. REFERENCES [1] J. C. Bose, "An automatic method for the investigation of velocity of transmission of excitation in Mimosa," Phil. Trans. B, vol. 204, pp. 63 -- 97, 1914. [2] V. Sukhov and V. Vodeneev, "A mathematical model of action potential in cells of vascular plants," Journal of Membrane Biology, vol. 232, no. 1-3, p. 59, 2009. [3] V. Sukhov, V. Nerush, L. Orlova, and V. Vodeneev, "Simulation of action potential propagation in plants," Journal of Theoretical Biology, vol. 291, pp. 47 -- 55, 2011. [4] E. Novikova, V. Vodeneev, and V. Sukhov, "Mathematical model of action potential in higher plants with account for the involvement of vacuole in the electrical signal generation," Biochemistry (Moscow), Supplement Series A: Membrane and Cell Biology, vol. 11, no. 2, pp. 151 -- 167, 2017. [5] T. Nakano, A. W. Eckford, and T. Haraguchi, Molecular communication. Cambridge University Press, 2013. [6] N. Farsad, H. B. Yilmaz, A. Eckford, C.-B. Chae, and W. Guo, "A comprehensive survey of recent advancements in molecular communi- cation," IEEE Communications Surveys and Tutorials, vol. 18, no. 3, pp. 1887 -- 1919, 2016. [7] I. Akyildiz, F. Brunetti, and C. Bl´azquez, "Nanonetworks: A new communication paradigm," Computer Networks, vol. 52, pp. 2260 -- 2279, 2008. [8] S. Hiyama and Y. Moritani, "Molecular communication: Harnessing biochemical materials to engineer biomimetic communication systems," Nano Communication Networks, vol. 1, pp. 20 -- 30, May 2010. [9] T. Nakano, T. Suda, Y. Okaie, M. J. Moore, and A. V. Vasilakos, "Molec- ular Communication Among Biological Nanomachines: A Layered Architecture and Research Issues," NanoBioscience, IEEE Transactions on, vol. 13, no. 3, pp. 169 -- 197, 2014. [10] H. Awan and C. T. Chou, "Improving the capacity of molecular communication using enzymatic reaction cycles," IEEE transactions on nanobioscience, 2017. [11] C. Gardiner, Stochastic methods. Springer Berlin, 2009. [12] D. J. Higham, "Modeling and Simulating Chemical Reactions," SIAM Review, vol. 50, no. 2, p. 347, 2008. [13] F. Tostevin and P. R. Ten Wolde, "Mutual information in time-varying biochemical systems," Physical Review E, vol. 81, no. 6, p. 061917, 2010. [14] P. B. Warren, S. Tanase-Nicola, and P. R. ten Wolde, "Exact results for noise power spectra in linear biochemical reaction networks," The Journal of chemical physics, vol. 125, no. 14, p. 144904, 2006. [15] R. G. Gallager, Information theory and reliable communication, vol. 2. Springer, 1968. Fig. 4. Mutual Information With and Without AP Fig. 5. Information Propagation Speed With and Without AP-Series Case molecules from the transmitting cell to the receptor cell(s). For propagation medium we assume a voxel size of ( 1 3 µm)3 3 µm), creating an array of 4 1 1 voxels (i.e., ∆ " 1 for the series configuration. The transmitter and each receiver occupy one voxel each as mentioned in the system model in Section II-A. We assume the diffusion coefficient D of the medium is 1 µm2s´1. The mean emission rate c is dependent on the AP signal which triggers release of molecules. The aim is to compute the mutual information between the input and output number of molecules of the complete system for number of receivers in series and use that to study information propagation speed for increasing number of cells. For this paper we show the result for the case with single sensing/transmitting cell and three receiver cells in series. In this work we present the result where we show the impact information and information of AP signal on the mutual propagation speed. In Figure 4 we show that the mutual information increases in the presence of AP signal for the system with single transmitter and three receiver cells as shown in Figure 3. Next by using the mutual information and selecting a threshold value we show that the information propagation speed increases in the presence of an AP signal as shown in Figure 5. Another important result from Figure 5 is that the information propagation speed increases with the increase in the number of receiver cells in series. The results for the mutual information as well as informa- 12345678910Time00.10.20.30.40.50.6Maximum Mutual InformationMutual Information With and Without APWithout AP r=3With AP r=311.21.41.61.822.22.42.62.83Number of Cells0.60.70.80.911.11.21.3Propagation Speed (cells/sec)Information Propagation Speed With/Without APWithout APWith AP
1007.2787
1
1007
2010-07-16T14:57:07
Fast, scalable, Bayesian spike identification for multi-electrode arrays
[ "q-bio.NC" ]
We present an algorithm to identify individual neural spikes observed on high-density multi-electrode arrays (MEAs). Our method can distinguish large numbers of distinct neural units, even when spikes overlap, and accounts for intrinsic variability of spikes from each unit. As MEAs grow larger, it is important to find spike-identification methods that are scalable, that is, the computational cost of spike fitting should scale well with the number of units observed. Our algorithm accomplishes this goal, and is fast, because it exploits the spatial locality of each unit and the basic biophysics of extracellular signal propagation. Human intervention is minimized and streamlined via a graphical interface. We illustrate our method on data from a mammalian retina preparation and document its performance on simulated data consisting of spikes added to experimentally measured background noise. The algorithm is highly accurate.
q-bio.NC
q-bio
Fast, scalable, Bayesian spike identification for multi-electrode arrays Jason S. Prentice1, Jan Homann1, Kristina D. Simmons2, Gasper Tkacik1, Vijay Balasubramanian1,2, and Philip C. Nelson1 1 Department of Physics and Astronomy, 2 Department of Neuroscience, University of Pennsylvania, Philadelphia, PA 19104, USA Abstract We present an algorithm to identify individual neural spikes observed on high-density multi-electrode arrays (MEAs). Our method can distinguish large numbers of distinct neu- ral units, even when spikes overlap, and accounts for intrinsic variability of spikes from each unit. As MEAs grow larger, it is important to find spike-identification methods that are scal- able, that is, the computational cost of spike fitting should scale well with the number of units observed. Our algorithm accomplishes this goal, and is fast, because it exploits the spatial locality of each unit and the basic biophysics of extracellular signal propagation. Human inter- vention is minimized and streamlined via a graphical interface. We illustrate our method on data from a mammalian retina preparation and document its performance on simulated data consisting of spikes added to experimentally measured background noise. The algorithm is highly accurate. Author summary Single neurons transmit messages in the form of electrical pulses called "spikes." Networked populations of neurons in the brain can use patterns of spikes to encode information or perform computations. To decipher this neural code, we must record simultaneously from many cells. The primary tools for such measurements are grids of closely spaced electrodes called multi-electrode arrays. The challenge in using these probes is that signals from single neurons typically appear on several electrodes, while each electrode records signals from multiple neurons. Disentangling these signals to determine which neurons fired, and when, is a principal bottleneck in understand- ing the collective behavior of neural circuits. Here we present an efficient and accurate Bayesian approach to this "spike sorting" problem that can scale to arrays of hundreds of electrodes. Our techniques accommodate variability in spike waveforms and identify spikes correctly even when nearby neuronal responses obscure one another. The key is judicious use of theory: we systemati- cally model and exploit the spatial characteristics of signal propagation and the structure of noise within a simple Bayesian model of neural activity. 1 1 Introduction The vertebrate retina is an important model system in neuroscience because it is amenable to detailed study despite having a complex structural and functional architecture [1]. Population coding and collective behavior in the retinal output is studied by use of multi-electrode arrays (MEAs) to record extracellularly from many retinal ganglion cells (RGCs) simultaneously [2, 3]. Similar recordings can now also be made in other brain areas [4]. MEAs offer unprecedented possibilities to obtain both single neuron and single action potential resolution from large tissue samples. However, recordings obtained in this way are useful only if every spike can be correctly assigned to the neuron that generated it. Even if each neuron spikes with a unique waveform sig- nature, we must still determine all those "template" waveforms present in a dataset, separating them from each other and from noise. Moreover, in practice there can be wide variation in the spike waveforms from a given unit (for instance in amplitude), complicating the task of determin- ing from data which units fired and when. Solving this "spike sorting problem" is the principal bottleneck in the use of high density arrays with hundreds or thousands of electrodes. Methods that were manageable with tetrodes [5] do not generally scale up to the massive datasets that large arrays generate. For example, some standard methods cluster data by manually examining two-dimensional projections in a feature space of a few tens of dimensions. This approach is infeasible when the feature space contains thousands of dimensions. Another challenge with large arrays is that the chance of seeing a single isolated spike becomes negligible, simply because there is so much activity. Thus we must find template waveforms corresponding to the activity of single neural units without ever seeing a pristine example of one, and we must be prepared to decompose temporally overlapping spikes in essentially every recorded event. Overlaps in both space and time are less frequent, but they nevertheless must be resolved if we wish to unravel the patterns of collective neural activity. Resolution methods that rely on exhaustively checking all possible combinations suffer a combinatorial explosion for large arrays. Further, any spike decomposition method must stop before every spike has been found, because there will be some units whose intrinsic amplitude is not larger than recording noise. We need a principled approach to terminating each fit and to deciding later which units' activities have been reliably captured. Thus, to be useful for large arrays, a spike identification algorithm must both scale well and also be able to decompose overlapping events. This article outlines a method that accomplishes these goals. It first clusters a small subset of a larger dataset, using an automated ordering technique combined with rapid human cluster-cutting. This manual step is efficient, and scalable, because (i) the ordering arranges event data by similarity along a single dimension, (ii) the ordered data display band-like features that are visually very salient for human operators, making cluster cut- ting unambiguous, and (iii) the algorithm is robust to variations and outliers in the cluster-cutting procedure. The algorithm then fits the full dataset to the spike templates thus obtained, using a modified Bayesian approach. In our data (from guinea pig retina) most of the intrinsic variabil- ity of spikes from a given unit consists of amplitude variation only, whereas other variability can be summarized as a universal (spike-independent) noise process. By carefully modeling these circumstances we greatly reduce our computational burden. After characterizing the spatiotemporal character of the noise, our algorithm identifies spikes iteratively in a matching-pursuit (or "greedy") scheme [6, 7]. Fitting terminates when addition of 2 Figure 1: (A) Typical MEA apparatus. A tissue sample was mounted in an inverted microscope, with images projected onto it via a small video monitor at the camera port (not visible). Clockwise from left, 1: suction; 2: tissue hold-down ring; 3: perfusion inflow, with temperature control; 4: preamplifier; 5: location of the multi-electrode array. (B) Example of a single-spike event. Each subpanel shows the time course of electrical potential (µV) on a particular electrode in the 5 × 6 array. The electrodes are separated by 30 µm (similar to RGC spacing). A spike from one unit is visible in the lower right corner and an axonal spike can be seen running vertically in the second column of electrodes. Data were acquired at 10 kHz. After baseline subtraction and high-pass filtering (see supplement), a spatial whitening filter was applied (see Sect. 3.1). another spike does not improve the likelihood score of a fit; no ad hoc complexity penalty is needed. No assumptions are made about spike firing times or cross-correlations; in particular, we do not require a priori any refractory "hole" in the spike time autocorrelation functions. Nevertheless, all of the inferred spike trains corresponding to otherwise acceptable spike types do exhibit such a hole, which serves as a check on our results. Any automated clustering algorithm requires human proofreading; we structure our methods (and develop tools) to make this as easy as possible. Our approach combines successful elements from previous techniques: the empirical charac- terization of the noise [8]; separation of clustering and fitting steps and the iterative subtraction scheme for handling overlaps [7]; and division of the clustering task by leader electrode address [9]. Novel features of our approach include systematic exploitation of the spatial organization of the signals, the use of an ordering algorithm to greatly simplify clustering, the observation that the noise temporal correlation is well represented by a simple function, the characterization and use of a prior distribution on spike amplitude variation, and the introduction of a principled Bayesian likelihood criterion for terminating spike fitting. Each of these innovations adds a critical element to the success of our spike sorting method. Although we focused on data taken on vertebrate retina, the methods should be equally applicable to other kinds of MEA data, for example in other brain areas [4]. 3 2 Results To illustrate our method, we tested our spike sorting algorithm on 120 min of recordings from guinea pig retinal ganglion cells (RGC), acquired with a 30-electrode, dense MEA covering about 0.018 mm2 of tissue (Fig. 1A). The analysis described in this paper identified 1 260 475 spikes in the dataset. A typical firing event took the form Fig. 1B, where each panel shows 3ms of the electrical potential recorded by each electrode (or "channel"). We identified spiking events as voltages crossing a threshold of −40 µV, taking into account the fact that simultaneous threshold crossings on neighboring channels represent the same spike event (see Methods for details). The duration of each spike event was taken to be 3.2 ms centered on the event's peak. In addition to identifiable spikes, each electrode had background activity with a standard de- viation of ∼10 µV that we will collectively refer to as "noise." Potential sources for this activity include true (Johnson) noise in the electrode and electronics, electrical pickup from the environ- ment, as well as a hash of background activity from weak or distant neurons [10]. A challenge for spike identification is that in general there is no way to separate these three classes of "noise" cleanly from each other, nor from the spikes of interest to us. Nevertheless, we will propose a technique for identifying spikes that is very accurate for firing events with intrinsic amplitude at least 4 times the standard deviation of the noise. Fig. 1B illustrates that each single firing unit will be "heard" on multiple electrodes, and that those electrodes form a spatially localized group. Our method is scalable because it systematically exploits this simple observation: even on a large elec- trode array, most firing units will involve only a handful of electrodes. (Some of our signals were not local, and stretched over the entire electrode array in a line (e.g., Fig. 1B). We ignored such axonal firing events, which were also distinguished by their low amplitude and triphasic shape.) 2.1 Preliminary visualization of our data We first attempted a "geographical clustering": from each event we found the minimum of the potential on each channel and the channel containing the deepest minimum ("leader channel"). We then used the absolute values of the minima as weights in a weighted average of the loca- tions of the 9 electrodes neighboring the leader channel. This weighted average gave a particu- larly salient two-dimensional feature, the event's barycenter ( ¯x, ¯y). Taking the third feature z to be minus the absolute peak potential gave a scatterplot that clearly showed many well-separated clusters (Fig. 2A), without any attempt to deduce the best selection of three features by principal component analysis (Fig. 2B). This extension of the "triangulation" method developed for tetrode recordings [5] already shows key aspects of the data: (a) many clusters are highly dispersed in am- plitude, and (b) some cluster pairs appear at nearly the same spatial locations but are nevertheless well separated by amplitude. The first circumstance means that we must allow for variable ampli- tude when fitting spikes to templates representing the clusters. The second warns us that a simple least-squares fit to amplitude could confound two distinct units. For this reason our spike-fitting method creates a Bayesian prior for each cluster's amplitude variation, allowing us to make such discriminations. Although the simple clustering based on spatial location in Fig. 2A looks promising, it can be misleading. Indeed, the restriction of the weighted average to the 9 electrodes around the leader can artificially separate clusters by biasing the barycenter to be located near a particular electrode. This problem could be alleviated by using a larger neighborhood, but on large arrays there will 4 Figure 2: (A) 22 234 firing events cluster well in terms of their barycenter (voltage weighted average spatial location) and absolute peak voltage (see text), despite wide amplitude dispersion in some groups; each combination of color and marker size corresponds to one spiking unit identified by the clustering method developed in the text. Grey points were unassigned to any cluster. A total of 107 clusters are marked. (B) Events cluster poorly when projected onto the three principal features uncovered by principal compo- nent analysis (PCA). Coloring as in (A). (C) Schematic of our spike sorting method. Dashed lines involve a small subset of a full dataset. The backwards arrow describes the introduction of new spike templates found after the first pass of fitting (Sect. 3.3); a total of two passes are performed. (D) The OPTICS algorithm orders all firing events into a linear sequence based on a distance measure (see text). Events are lined up in this order (x-axis), and represented in terms of the 960 voltage samples recorded by all the electrodes during a 3.2 ms firing event (y-axis; from top to bottom, 32 consecutive time samples from one channel, then 32 time samples from the second channel, and so on). The human operator highlights bands of events (typically very clear to an observer) that appear to constitute a single cluster; one such band is shown. Later automated diagnostics refine and check these assessments. inevitably be temporal collisions of spikes from distinct units. The barycentric features in Fig. 2A will register such collisions as a haze of seemingly random spots. Thus, at a minimum the MEA voltage traces must be segmented by exploiting the spatial locality of recorded responses. Despite 5 these shortcomings, Fig. 2A points out why other, more sophisticated, methods can succeed: the "geographic" information encoded by the MEA is a powerful intrinsic clue to each unit's identity. 2.2 Summary of our method Our sorting method is outlined in Fig. 2C (details in Methods). From a subset of the raw data, we made a preliminary classification of spike events in terms of the electrode on which they achieved their peak voltage. All events sharing a given leader channel were cropped to the 9 electrodes neighboring the leader, then ordered with the OPTICS algorithm [11] into a linear sequence. The OPTICS algorithm places similar events nearby in the sequence; distance was measured by a nor- malized Euclidean distance between event voltage traces (see Methods). The linear sequence of events was displayed to the user along with all the recorded voltage samples for each event (Fig. 2D), and manually clustered. Although the ordering was based on events cropped to 9 chan- nels, the full waveforms were displayed to the user (Fig. 2D). Because the data are ordered in one dimension, and because precision is not required in view of later automated refinement, this man- ual step remains rapid. An automated method for cluster cutting could be implemented, but in view of the inevitable need for human proofreading we preferred to simply carry out this step by hand. From each preliminary cluster, we estimated a template waveform representing the corre- sponding neural unit and then fit the templates to the remaining data. Fitting was accomplished by a Bayesian algorithm based on a probabilistic model capturing the dominant sources of vari- ability we observed in our data: background noise, spike amplitude, and overlapping spikes from distinct units. After finding, for each event, the most probable template which accounts for the event, we subtracted it and then iterated. Finally, the fit results were used in a post-hoc validation of the initial clustering, and we repeated the procedure in a second pass if necessary. Details of each step are presented in Methods. 2.3 Tests of our method OPTICS-based clustering of our dataset led to 107 potential templates for events from distinct neu- ral units. Many of these templates had low amplitudes; such low-amplitude templates were some- times mistakenly fit to noise by our algorithm. We therefore rejected units that were likely to con- tain substantial noise fits because they were of amplitude less than or comparable to the noise (details in Sect. 3.4 and Fig. 6D). This left fifty potentially reliable units that were accepted in our dataset. Comparison with geographical clusters: Our OPTICS-based procedure identified 107 potential clusters of events in a subset of the data. To check that the procedure gave reasonable results we plotted each event in the barycentric coordinates of Fig. 2A, and colored the events according to the cluster label. The clusters were spatially localized and separated in peak amplitude, as they should be if they were produced by distinct single neurons. Gray dots in Fig. 2A were not assigned to any cluster. Some of these events contained overlaps of spikes that were not resolved by the initial spatial segmentation of data during the preprocessing step. The subsequent spike fitting step in our algorithm attempted to resolve such ambiguities. To validate our algorithm we tested its performance on synthetic Error rates on synthetic data: data created by adding spikes to experimentally measured background noise clips, then fitting 6 templates to each clip. We took noise clips to be 3.2 ms segments of time during which no spikes were recorded on any channel; we identified 14 000 such clips. For each clip, we randomly chose a fixed number (1, 3, 5, 7, or 9) of templates from the initial set of 107, with uniform probabil- ity and without replacement. We then added these templates to the noise clip at random times, leaving a margin of 0.6 ms on either side of the clip to prevent waveforms from being cut off. (Our typical template waveforms extended approximately 0.5 ms to either side of the peak.) We gave each spike an amplitude drawn from a Gaussian distribution with mean equal to its tem- plate amplitude and standard deviation 10% of the template amplitude (this was similar to the observed distribution). The template fitting algorithm was then run over this synthetic dataset and analyzed for false positive and false negative rates (Fig. 3A,B). We counted a false negative for a template every time that template was present in an event but not fit correctly to within 1 ms; we counted a false positive every time a template was fit to the data without actually being present. The error rates increased with the number of template overlaps; thus, for the fifty tem- plates with amplitudes that exceed the noise, we separately plotted error rate histograms for each degree of overlap. Error rates were robustly low - even within extremely complex events with 9 overlapping spikes (very rare in the data), the majority of spike templates had an error rate of a few percent or less. To gain perspective on these values, we measured the number of templates fit to each event in our recorded data: 60% of events contained 1 spike, 25.5% 2 spikes, 8.5% 3 spikes, 3% 4 spikes, 1% 5 spikes, and 2% had more than 5 spikes. Most of the errors were made on lower amplitude templates for which amplitude variations can lead to confusion with noise. Refractory violations: When sorting spikes recorded extracellularly, ground truth can be as- sessed if simultaneous intracellular recordings are available, e.g., [12]. Since we do not have such recordings, in order to validate our algorithm on real data we examined the rate of refractory vio- lations - i.e., the fraction of interspike intervals of duration less that 1.5 ms. Refractory violations can appear in our sorted data if spikes from distinct neural units are mis-assigned to same unit, or if noise fluctuations are mistaken for spikes. Of the 107 templates constructed from the initial clus- tering 84% had less than 0.5% refractory violations and all had less than 2.5%, providing evidence that the templates produced by the initial clustering rarely merge distinct neural units (Fig. 3C). All fifty templates describing units that rose reliably over the noise level had less than 0.5% refrac- tory violations. Futhermore, 96% of these had less than 0.1% refractory violations (Fig. 3C). This is strong evidence that our algorithm makes few fitting mistakes on the units otherwise identified as reliable. Coverage: While the absence of refractory violations gives evidence that our algorithm does not merge different neural units together, it might still split spikes from the same unit into two distinct clusters if, e.g. there was substantial amplitude variation. To test for this, for each unit that was above the noise level we measured the linear receptive field by taking the spike triggered average (STA) of the flickering checkerboard stimulus. We expect that such receptive fields will be connected regions of the visual field, roughly elliptical in shape, and that no two units will have identical receptive fields. 31 of the 50 reliably identified units had enough spikes to give reliable estimates of the spatial receptive field. Of these, examination of the temporal kernel showed that 19 were OFF cells (responding to light decrement) and 12 were ON cells (responding to light increment), consistent with the expected excess of OFF ganglion cells [13]. None of these receptive 7 Figure 3: (A) The percentage of templates with different false negative fractions in fits to synthetic data (fraction of times a fit was not reported for a template when it was actually present). (B) The percentage of templates with different false positive fractions in fits to synthetic data (fraction of times a fit was reported for a template when it was not actually present). Results in (A) and (B) reported separately for events with different numbers of template overlaps (inset colors). (C) In fitting real data, refractory violations are rare (see text). (D) The centers of 19 OFF cell receptive fields recorded from a single piece of tissue. To map a neuron's receptive field center, we first find the peak (in space and time) of the spike-triggered average stimulus. Restricting to the peak time, we apply cubic spline interpolation in space and then draw contour lines at 75% of the peak value. fields were identical, giving evidence that our algorithm did not split single units into multiple clusters. Further, all of the receptive fields were connected, suggesting that none of our clusters are mixtures of different RGC. In addition, essentially all of the recorded area was covered by at least one receptive field (coverage of OFF cells shown in Fig. 3D). The density of RGCs in guinea 8 Figure 4: (A) Example of a single-spike event. Each subpanel shows the time course of electrical potential (in µV, black curves), on a particular electrode in the 5 × 6 array. After baseline subtraction and high-pass filtering, a spatial whitening filter was applied (see Methods). Blue curves show the result of our fitting algorithm, in this case a single template waveform representing an individual neural unit. (B) Detail of a more complex event and its fit, in which a single unit fires a burst of 9 spikes of varying amplitudes (upper left channel), while a different unit fires 5 other spikes (upper right channel). (C) Example of an overlap event and its fit, which now is a linear superposition of 7 templates. (D) Detail of (C), showing signals on four of the electrodes. This time individual fit spikes are displayed. The red and green traces show fit templates that, although similar, differ significantly in their overall strength, and in the relative strengths of their features. The olive trace shows a fit to a low-amplitude template that was later classified as unusable, and hence was discarded, by the procedure in Sect. 3.4. pig varies from 250 mm−2 to1500 mm−2 [14]. We receive signals from a region slightly larger than the electrode array, roughly 0.065 mm2. Thus the expected number of RGC is 16–97, comparable to our total of 31 receptive fields, keeping in mind that many of the sluggish cell types would not have enough spikes to yield a good spike triggered average. Complex events: A major challenge for a spike sorting algorithm is dealing with variability in spikes produced by individual neural units. An even greater challenge arises from spatio- 9 temporal overlaps between spikes from different neural units. Our low error rate in analysis of synthetic data containing both of these complexities (Fig. 3) provides evidence that our algorithm is effective at resolving overlaps and identifying variable spikes from given units. To test this fur- ther, we manually examined many events in the real data which a human observer could identify as representing overlaps or neural variability; and the algorithm typically did an excellent job of dealing with variable-amplitude bursts (Fig. 4B), as well as events that overlap in space and time (Fig. 4C). Speed: Currently the main fitting code, written in Matlab, requires about 10 ms of real computer time per fit spike on a commercial 2.5 GHz computer, times 2 for the two passes. This is fast enough for our purposes; considerable further improvement is possible with existing software (Mex) and hardware (GPU) techniques. 3 Materials and methods Ethics statement This study was carried out in accordance with recommendations from the Na- tional Institutes of Health and the guidelines of the American Veterinary Medical Association. The protocol was approved by the Animal Care and Use Committee of the University of Pennsylvania (No. 803091). All surgery was performed under ketamine/xylazine and pentobarbital anesthesia, and all efforts were made to minimize suffering. Experimental procedure Our methods were developed and tested on retinal response data, from albino guinea pig, recorded with a dense 30-electrode array (30 µm spacing, Multi Channel Systems MCS GmbH, Reutlingen, Germany). After anesthesia with ketamine/xylazine (100/20 mg/kg) and pentobarbital (100 mg/kg), the eyeball was enucleated and the animal was killed by pentobarbital overdose in keeping with the AVMA guidelines on euthanasia. The eyeball was hemisected and the retina was allowed to dark adapt. A small piece was cut out, separated from the pigment epithelium, mounted (ganglion cells up) onto a piece of filter paper, and placed gan- glion cells down onto the MEA. A 15 × 15 flickering checkerboard consisting of binary noise, updated at 30 Hz, was projected onto the tissue. We alternated between uncorrelated and expo- nentially correlated (space constant 50 µm; time constant 33 ms) stimuli. Our procedure for identifying spikes in the recorded data had four steps: (1) Preprocessing (Sect. 3.1), where spatial locality was exploited to segment the data, (2) Clustering and template building (Sect. 3.2), where a subset of the data was clustered to separate the responses of likely neural units, template waveforms for each neural unit were built, and their variability charac- terized, (3) Spike fitting (Sect. 3.3), where every firing event was separated into a superposition of responses from different neural units, and (4) Validation of templates (Sect. 3.4), where each template and the spikes identified with it were tested for reliability. 3.1 STEP 1: Data preparation and segmentation The first step in our procedure was to prepare the data for clustering of events from different neural units, by separating firing events from noise, and segmenting spatio-temporally distinct regions of spiking activity on the electrode array. 10 Data from the array were sampled at 10KHz, high-pass filtered below 200 Hz with a finite impulse response filter to remove low frequency baseline fluctuations, and then packaged into 3.2 ms clips: (a) "noise clips" in which the potential never fell below −30 µV, and (b) "spike events" surrounding moments at which the potential crossed −40 µV. Clips with potentials be- tween −30 µV and −40 µV were neither used to characterize noise (since they might contain small spikes) nor used to identify spikes (since they were very noisy). The threshold for spikes was set to ∼4 times the standard deviation of the potential in the noise clips. Each spike event thus consisted of N = 3.2 ms × 10 kHz × 5 × 6 = 960 numbers, the potentials on a 32 × 5 × 6 grid of space-time pixels ("stixels"). Spike events sometimes overlapped each other, for example if a burst of spikes lasted longer than 3.2 ms. Cluster identification and spike template building (Sect. 3.2) used four 30-second segments sampled from different times, but subsequent spike fit- ting and sorting (Sect. 3.3) used all the data. Electrodes can share signals because of instrumental cross-talk and because the activity of neu- rons spreads passively to nearby electrodes. Both effects can be captured by a linear filter that spatially blurs signals, and also applies to noise. Thus, we measured the spatial covariance of noise clips - it was spatially isotropic and had a roughly exponential falloff, with a correlation length of ∼ 30 µm. We applied the square root of the inverse of this covariance matrix to all data, and used the resulting "spatially whitened" data for all analysis. In some datasets this transfor- mation sharpened the individual spikes spatially, improving our ability to distinguish them in the clustering stage. (In other datasets the transformation had little effect.) Our data also exhibited temporal correlations, but these have a different physical origin from the essentially instantaneous passive spatial spread. We found that temporal whitening prior to clustering [8] worsened our sig- nal/noise ratio and impeded cluster determination. Thus we incorporated temporal correlations later, during the spike fitting. Each spike event is a superposition of spikes from an unknown number of dis- Segmentation: tinct neural units with stereotyped waveforms that we sought to identify. We first spatially seg- mented the data to isolate waveforms from individual units and their immediate neighbors. To this end, we identified all stixels at which the potential was more negative than the threshold of −40 µV and divided this set into connected components (two stixels were considered connected if they were nearest neighbors in either time or space). Within each connected patch we identified the absolute peak electrode and time, then extracted a 3.2 ms region centered temporally on the peak time and cropped spatially to a neighborhood of nine channels surrounding (and including) the leader electrode. Thus each spike event was segmented into one or more cropped events; each of which was then classified according to its leader electrode. 1 In subsequent clustering, only those events having the same leader electrode were directly compared to each other [9]. Some cropped events might be composites of two spike types corresponding to neighboring, but distinct, neural units. However, this step at least decomposes composite events whose compo- nents are well separated in space or in time, and hence reduces the combinatorial burden inherent in large arrays; later steps handle composites missed at this stage. The method also ensures that, if spikes from two well-separated units frequently co-occur, the two units will nevertheless be correctly handled as separate. 1A similar segmentation method has recently been applied to the spike identification problem by J. Schulman (un- published); see http://caton.googlecode.com. 11 3.2 STEP 2: Cluster identification and template building The second step in our procedure was to cluster spiking events in a subset of the data (four 30- second segments) into groups that had similar waveforms and thus probably came from the same neural unit. For each cluster, we produced a template waveform describing the typical spike, and determined the distribution of amplitude rescalings that best matched spikes to this template. In order to group events into clusters based on the similarity of their Cluster identification: waveforms, some previous approaches have sought a low-dimensional set of discriminable wave- form "features," and have assumed that variability between events in the same cluster arises only from additive noise. In practice, systematic variation in the shape of spikes from single units is often observed that is not due to additive noise. Furthermore, identifying the correct set of salient waveform features that discriminate between units is challenging ([15]; see Fig. 2B). Thus, seeking a technique that did not require feature extraction, we adapted the OPTICS algorithm [11]. Briefly, OPTICS computes distances between all pairs of waveforms, then orders the waveforms such that similar ones are placed close together in a single linear sequence. OPTICS makes no assumption that clusters have a Gaussian distribution in feature space, nor does it set any threshold density in that space to trigger cluster identification. The linear ordering allows for easy visualization and cutting of clusters. We applied this algorithm to cropped and segmented spike events which were upsampled by a factor of 5 (using Matlab's cubic spline interpolation) and then temporally aligned to place the absolute peak of the waveform at a common position before downsampling again. The interpo- lation was necessary to compensate for apparent variations in spike waveforms due to discrete sampling [16]. To reduce the fuzziness of the clusters, we masked spike events by setting voltage samples to zero if they were less negative than −15µV. As a distance metric between V and V′, the masked potentials of spike events, we chose d(V, V′) = (cid:16)∑N i=1[(Vi − V′ i )2/(kqVi + V′ i )](cid:17)1/2 where i indexes the potentials at each channel and timepoint, and k is the total number of nonzero potentials after masking of either V or V′. Division by k normalized for the effective dimension- ality (given by the number of dimensions containing nonzero entries). We observed that higher i partially compensated for voltage traces tended to have a higher variance; the factorqVi + V′ this, leading to more homogeneous clusters. We constructed a graphical user interface (GUI) that allowed a human operator to visualize each spike event in the OPTICS sequence as a vertical column of pixels color-coded by voltage (see Fig. 2D). Transitions between distinct spike types were usually obvious to the operator, who could quickly find and select bands corresponding to each spike type. (For the data in this paper, the operator found over 100 such clusters in about 30 minutes.) The software then wrote the corresponding cropped events to a set of data files. The ease of separation likely occurred because clusters could already be fairly well delineated with just the "geographical" features in Fig. 2A. Up to this point, the events being clustered were still segregated into batches according to their leader channel x0, y0. Thus it was possible for a single unit to be multiply identified: If it stim- ulated two neighboring electrodes nearly equally, the unit could generate events in both of the corresponding batches. We tested for duplicates by manually examining pairs of clusters whose medians had a large cross-correlation and merged the clusters if necessary. There was also a pos- sibility that the initial clustering would assign multiple units to one cluster. In these cases, visual 12 examination of the superposed waveforms of the cluster often showed it to be a composite of mul- tiple units. This was resolved by doing a principal components analysis on the waveforms in the cluster: if the cluster was composite, at least one of the first few principal component weights had a multimodal histogram. The cluster was split by thresholding at the valleys of the histogram; we then tested whether any of the split components ought to be merged with an existing cluster. We developed a graphical user interface to assist the operator in performing these merging and splitting steps. Generally it was clear to the human operator when a band in the GUI output was clean enough and wide enough (contains enough events) to generate a good cluster; thus there was no need to specify a priori the desired number of clusters, an advantage over many automated clustering procedures. Marginally significant clusters were either eliminated during template building (see below), or else generated fits that were themselves discarded during spike fitting (Sect. 3.3) and evaluation of template reliability (Sect. 3.4). Any significant clusters missed at this stage, for exam- ple because of the small fraction of the data used in this step, were found and reincorporated later during spike fitting (Sect. 3.3). Template building: Next we created a consensus waveform ("template") summarizing each cluster of cropped, upsampled events, and characterized meaningful deviations from that consen- sus. We created a draft template by finding the pointwise median over all events in a cluster, then aligned each event to the draft template by maximizing their cross-correlation over time shifts, which we found to be more accurate than aligning to each event's peak time. Finally, we found the pointwise median (to suppress the effects of outliers) of the aligned events; this waveform was our template (Fig. 5B). Our code displays 40 events in each cluster together, so that a human operator can spot any mixed (decomposable) cluster inadvertently missed by earlier stages of the analysis (Fig. 5A). Gen- erally such clusters can safely be discarded, because each "parent" neural unit has also generated its own "pure" cluster; if not, the operator can either revisit the OPTICS code specifically to find the missed events, or else wait for them to show up during spike fitting (Sect. 3.3). A key step was to realize that, in our data, the most significant sources of variation of indi- vidual spikes from the template were (a) additive noise, and (b) overall multiplicative rescaling of the spike's amplitude (Fig. 5C). To quantify (b), we found the rescaling factor that optimized the overlap of each spike with its template, then stored the mean and variance of those factors in a lookup table for later use as a prior probability for amplitude variation. We also logged the number of events associated to each template, converted to an approximate firing rate, and saved those rates, again for later use as a prior. 3.3 STEP 3: Spike fitting The third step in our procedure was to fit the spike templates constructed in Step 2 to each firing event in the data in order to determine which neural units were responsible for the activity. To this end, we constructed a simple generative model of firing events, and included prior probabilities for firing rate and for amplitude variation of each template. The fitting procedure then iteratively identified and subtracted the most likely templates in each firing event. The cluster templates were produced using an upsampled 50 kHz sample rate, but for fitting to data we downsampled back to actual 10 kHz, in each of 5 "reading frames"; that is, we created 13 Figure 5: (A) Detail of 40 of the aligned events used to compute a template, upsampled and shifted into alignment as described in the text. Some outlier traces reflect events in which this unit fired together with some other unit; the unwanted peaks occur at random times relative to the one of interest, and thus do not affect the template. (B) Blue, detail of a template waveform, showing the potential on 12 neighboring electrodes. Time in ms runs horizontally; the vertical axis is potential in µV. Red, for comparison, the pointwise mean of the 430 waveforms used to find this template. (C) Detail of (A), showing only the leader channel. In addition, each trace has been rescaled by a constant to emphasize their similarity apart from variation in overall amplitude. five versions of each template corresponding to subsample shifts. Let Fµ;x,y(t) be the potential of template µ, on the electrode with address x, y, at time t, with time measured in units of the sampling time δ = 0.1 ms, and the template peak at the central point t = 16 within the 3.2 ms template frame. We use the vector notation Fµiti for the template µi shifted to time ti, i.e. its x, y, t component is Fµi,x,y(t − ti). The goal of spike fitting is to identify, for each spike event, all the units {µi} Generative model: which contribute to the event and their firing times {ti} irrespective of their amplitudes {Ai}. Thus we assumed a probabilistic generative model of the data [17, 16, 18, 8] and computed the posterior probability of {(µi, ti)} given the observed data. We assumed that a spike event V could be explained by a linear combination of templates Fµiti with variable amplitudes Ai and correlated, zero-mean Gaussian noise δV: V = ∑ i AiFµiti + δV. (1) Given this model, to obtain the posterior probability that a firing event consists of a particular set of templates, we need to specify the prior probability of µ, t, and A. We chose a Gaussian prior for the amplitude A, a Poisson prior for µ, and a uniform prior for t. Our model assumes that spike waveforms from a given neural unit are stereotyped, apart from their amplitude. We did observe considerable variation in spike amplitude (Fig. 2A), in part due to bursting [10, 12], and thus included it in the model as a distribution of amplitude rescaling factors. Allowing for the possibility of slight variations in spike width (Supplementary Information) also slightly improved our results. But there was little additional variability to be 14 modelled (Fig. 5C). Our model also assumes that signals from different units combine linearly, as does the noise. This is reasonable, because the biophysics of extracellular recording is governed by the equations of electrodynamics, Ohm's law, and other linear relations. A third assumption is that noise and the variability of spike amplitude are well described by Gaussian distributions. Assuming Gaussianity (well-confirmed in some settings [8], but not others [19]) allows for a fast, partially analytic approach to fitting. We validate this assumption quantitatively below. Our generative model has a Poisson prior probability for firing by each neural unit, i.e. a prior that is as simple as possible while being consistent with the mean firing rate. The prior probability could be made somewhat more accurate by including refractory periods, the likelihood of burst- ing, and correlations between neural units. But this would significantly increase the complexity of the model, and inferring the prior would require much more data [20]. Finally, we assumed that all statistical distributions that enter into the model are stationary and independent of the stimulus. While our retinal preparation does not suffer electrode drift (as might implanted electrodes), there are occasionally shifts in spike amplitudes and firing rates over the course of a lengthy experiment. Although in principle our fixed priors could lead to biased estimates, these biases are small when spike identification is robust, i.e. when the likelihood function dominates the prior probability in the posterior probability of a neural unit [21]. In the context of our generative model, in order to assess the probability Noise characterization: that the residual after subtracting a putative spike is indeed noise, we first need to measure the distribution of noise. After applying the spatial whitening filter (Sect. 3.1), our noise clips are decorrelated in space, but not in time (Fig. 6A). Assuming that the noise has a correlated Gaussian distribution, we need the inverse of the noise covariance matrix, C−1. One approach to finding C−1 is literally to invert the empirical covariance matrix C of a large set of noise clips. Besides being intractable for larger arrays, this approach has the disadvantage that a numerically stable evaluation requires a very large noise sample. For these reasons we instead took a parametric approach. After evaluating the covariance C(x, y, t; x′, y′, t′) we noted that it was approximately diagonal and translation-invariant in space (i.e. proportional to δx,x′δy,y′ and independent of x and y). It was also approximately stationary, i.e. invariant under time shifts t → t + ∆t, t′ → t′ + ∆t, and thus only depended on t − t′. Finally, we observed that the time dependence of C was roughly an exponential, ≈ ηe−t−t′/τ (Fig. 6A). This gave: C−1(x, y, t; x′, y′, t′) = η−1 × δx,x′δy,y′  (1 + ξ2)/(1 − ξ2) , if t = t′ −ξ/(1 − ξ2) , if t = t′ ± δt 0 , otherwise. (2) Here δt = 0.1 ms is the sample time, and δx,x′ is the Kronecker symbol. η and ξ = e−δt/τ are obtained from the noise covariance. The dataset used in Results yields noise strength η ≈ 57 µV2 and ξ ≈ 0.58. By construction, our noise model reproduces the 2-point correlations in the noise clips. How- ever, real noise may not be Gaussian distributed. One check on this is to construct the transformed quantities U = C−1/2V empirically, find their full distribution as V ranges over noise clips (the "one-point marginal" distribution), and compare to a normal distribution. Fig. 6B shows this com- parison, lumping together every element of U. The empirical noise deviates from a Gaussian only in the far tails that contain very little weight. 15 Figure 6: After fitting spikes, only noise remains. (A) Noise covariance after spatial whitening. Subpanels: spacetime covariance C(x, y, t; x∗, y∗, t + ∆t) between the central channel and its neighbors as a function of ∆t, for various fixed t (colored curves). Central panel (dotted line): the function (57 µV2) exp(−∆t/(0.18 ms)). (The various t lines and the dotted line are too similar to discriminate vi- sually.) Horizontal axes: ∆t in ms; Vertical axes: C in µV2. (B) Blue curve, Semilog plot of the one point marginal probability density function of decorrelated noise samples. Red curve, same quantity, evaluated on residuals after spikes have been removed from spike events. Dotted curve, The Gaussian chosen to rep- resent this distribution. (C) Green, detail of the same template waveform shown in Fig. 5. Red, pointwise mean of the residuals after the fit spike is subtracted from 4906 one-spike events of this type is nearly flat. This validates our assumption that spikes vary only in overall amplitude, and that noise is independent of spiking. Blue, pointwise standard deviation of the residuals, again evidence that only noise remains after fitting and subtracting spikes. (D, TOP) Histogram of fit values of the scale factor A for a template with peak amplitude −168 µV (well above noise) obtained without a prior on A, superposed with a Gaussian of the same mean and variance. (D, BOTTOM) Similar histogram for a low amplitude template. A secondary bump appears, due to noise-fits, but is well separated from the main peak; a cutoff is shown as a dashed green line. The superposed Gaussian has mean and variance computed from the part of the empirical distribution lying above the cutoff. Fitting algorithm for single spikes: Given the above characterization of the noise distribution, and our Gaussian prior for spike amplitude variation, the generative model Eqn. 1 defined the posterior probability P({µi, ti, Ai}V) for templates {µi} to be present at times {ti} with ampli- 16 tude scale factors {Ai}, given the recorded potentials V. We ideally would have marginalized P({µi, ti, Ai}V) over the nuisance parameters {Ai} and then maximized with respect to {(µi, ti)} to identify the most probable set of units and spike times. In practice, this maximization is com- putationally expensive to perform on many templates simultaneously. Instead, we used a greedy approximation which fit one template at a time. We first assumed that the event contained exactly one spike and identified the spike's type µ1 and time of occurrence t1. Bayes' formula gives for the posterior probability: P(µ1, t1, A1V)dt1dA1 ∝ P(Vµ1, t1, A1)P(µ1, t1, A1)dt1dA1 , (3) up to a constant independent of µ1, t1, and A1. Here P(µ1, t1, A1) is the prior probability of the template µ1 appearing at time t1 with an amplitude A1: P(µ1, t1, A1)dt1dA1 = rµ1dt1(2πσµ1 2)−1/2 exp(cid:16)−(A1 − γµ1)2/2σ2 µ1(cid:17) dA1 , (4) where γµ1 is the mean and σ2 µ1 the variance of the scale factor for cluster µ1; rµ1 is the estimated overall rate of firing for this cluster. The generative model gave the probability of the observed potential V given µ1, t1, A1 (the likelihood) as P(Vµ1, t1, A1) = Pnoise(V − A1Fµ1,t1) , where Pnoise is a Gaussian distribution with zero mean and covariance C (Eqn. 2). Combining the likelihood and prior, then integrating out A1, gave the formula we ultimately used in our fitting algorithm: P(µ1, t1 V) ∝ q1 + σ2 rµ1 µ1 Ft µ1,t1 exp 1 2 C−1Fµ1,t1 (γµ1 + σ2 µ1 Ft 1 + σ2 µ1 VtC−1Fµ1,t1)2 C−1Fµ1,t1 µ1,t1 − γ2 µ1 2σ2 µ1! . (5) The (unwritten) constant of proportionality in (5) is the probability that no templates are present in the event; this quantity cancelled in a subsequent step. Finally, we maximized (5) over µ1 and t1 to identify the template and its firing time. We improved scalability by a slight approximation. Starting from a spike event, we first identified the time and electrode address of its absolute peak and restricted the matrix products in expression (5) to only sum over a spatiotemporal neighbor- hood surrounding this peak. The size of the neighborhood was chosen to match the typical spatial extent and temporal duration of the templates. In principle, we could have extended the single template procedure described Multiple spikes: above to compare the probabilities of all possible combinations of two or more spikes. Such an exhaustive approach, however, would quickly have become impractical. We instead noted that, even if an event contains multiple spikes, the single-spike fit described above still identified that template whose removal would lead to the largest increase in the probability that the remaining waveform is noise. Thus we adopted an iterative (matching-pursuit or "greedy") approach: start- ing with a spike event, we found the absolute peak, fit it, subtracted the fit, and then repeated the process [7]. The single-spike procedure found the most probable spike type µ∗; we then needed the scale factor A that would allow the fit spike to be subtracted as fully as possible. We thus held µ fixed to µ∗ and minimized the ordinary norm kV − AFµ∗,t1k2 over A and t1. The scaled and shifted template obtained in this way was subtracted before repeating the fitting procedure. To determine when to stop fitting spikes, we adopted a likelihood ratio test. At each step of the fitting loop, we marginalized Eqn. 5 over t1, obtaining the probability that an additional spike of type µ1 is present. We then divided by a similar expression for the probability that no additional 17 spike was present. In this ratio, the proportionality constant from Eqn. 5 cancels. We can then say that fitting an additional spike is justified if the ratio exceeds unity for some µ∗. The fitting loop terminates when the significance test fails. Fig. 4C,D shows an example of the successful decomposition of a multiple-spike event using our method. The spike fitting algorithm might exit prematurely if a spike is present that does Second pass: not appear in the list of templates initially extracted from the small subset of data. In this case, the fit will terminate, even though other identifiable spikes of lower amplitude may remain. To check this, if the residual exceeds V∗ trust =−44 µV after termination, the code declares an "incomplete fit" and writes the residual to a file; the small set of resulting waveforms were then reintroduced into our clustering code and used for a second round of fitting. In this way we can be assured of finding even rare spike types, without having to perform clustering on the complete dataset. Using this method, only 0.02% of fits in the second pass were classified as incomplete. It can also happen that the small data sample used for clustering gives a poor estimate of some firing rates and amplitude distributions that enter our priors for spike fitting. Thus, before the second pass of fitting the priors are updated based on the outcome of the first pass. 3.4 STEP 4: Evaluation of template reliability After spike identification, we performed a final evaluation to test whether templates and their sorted spike trains were trustworthy. The primary criteria were: (1) residuals after spike removal should resemble noise, (2) the histogram of amplitude scale factors should be unimodal, (3) the inter-spike interval (ISI) distributions should display "refractory holes"; and the cross-correlation functions should not. Additional criteria are described in the Supplement. Reliable templates were taken to be those that passed all these tests. Most unreliable templates failed multiple tests. For single-spike events, the residual signal after subtracting the fit should resem- Residuals: ble pure noise; in particular it should be stationary in time and translation-invariant in space. Fig. 6C shows that these expectations were met, validating our assumptions. For example, if the unit in Fig. 6C had significant variations other than amplitude rescaling, or if there had been an amplitude-dependent noise process, then we would expect significant non-stationarity in the residual curves [16]. To test that, after termination, the residual of a spike event consists only of noise, we computed the one-point marginal distribution of waveforms after all known spikes had been removed. Fig. 6B shows that this distribution closely resembled that of noise clips, indicating that our code indeed found the significant spikes. Of particular note, the standard deviation of the residuals matched closely that of the noise: For each template, we found the standard deviation of the residuals of events to which only that template was fit. This value ranged from 6.92 µV to 8.63 µV, while the noise standard deviation was 7.57 µV. For large amplitude templates, the distribution of amplitude scale factors (A) ob- Amplitude: tained during spike fitting was typically close to Gaussian (Fig. 6D). On the other hand, for low amplitude templates, accidental noise fits can sometimes lead to a histogram of A values with a secondary, low-amplitude peak well separated from the expected peak near A ≈ 1 (Fig. 6D). Examining the A histograms allowed us to quickly set an individual threshold for each reliable 18 template. Fit spikes with A value below this threshold were discarded. If two peaks were dis- cernible but overlapped significantly, the entire cluster was deemed unreliable and its spikes were not used in further analysis. In addition, our trigger rejected any spike event that did not cross −40 µV; thus any cluster whose A histogram extended closer to zero than this was probably miss- ing some true spikes, and was not used. Interspike interval distributions for single units are ex- ISI distribution and cross-correlation: pected to have a refractory hole; our analysis of these distributions was described in 2.3. Two dis- tinct neural units need not respect any mutual refractory period. Their spike-time cross-correlation function is therefore not expected to display any hole. We looked for such unexpected behavior and, when found, reexamined the corresponding templates. If the templates appeared to be dupli- cates, we merged the corresponding spike trains [9, 7]. Another diagnostic for duplicate templates is a coincident receptive field. 4 Discussion A review of early work on spike identification, prior to the widespread use of MEAs, can be found in [22]. Like some earlier work, our method separates spike identification into distinct steps of clustering and fitting. The clustering step uses all the waveform features, and makes no assumption about the cluster structure (e.g., that it is a mixture of Gaussian distributions). The fitting step acknowledges that each neural unit's signals are subject to intrinsic, multiplicative variation as well as additive noise, and uses a Bayesian approach to infer the identity of the most likely firing unit. Our approach is intentionally not fully automatic, since human proof-reading of the results of automated clustering is generally essential. However, we have been careful to use human judge- ment only where it is indispensable. Further, both the human and machine steps are organized so as to scale well with array area (or number of units monitored). For example, cluster cutting was greatly simplified by representing spikes in an ordered one-dimensional array. This feature, along with systematic exploitation of the spatio-temporal locality of spikes (Fig. 2A), and the use of a simple but powerful generative model, make our method scalable to large arrays. Furthermore, we observed that our more ambiguous templates tended to be located on the boundary of the ar- ray due to recording of units located some distance from the electrodes. These "boundary effects" should become less important for larger arrays; we thus anticipate that the methods described in this paper will yield more accurate spike sorting for large, dense arrays. Our method can be extended in many ways. For example, it would be straightforward to update the priors continually as fitting proceeds, allowing non-stationarity and stimulus depen- dence to be handled more gracefully. In some applications it may be preferable to report spike identification probabilistically, rather than just listing the most-likely spike events; our formulas already provide this information. The method can also be extended to non-planar arrangements of electrodes and neural tissues, for, e.g., cortical applications. Finally, the generative model in the present paper does not take into account correlations within and between spike trains, or the receptive field structure and stimulus dependence of responses. While our algorithm is already very accurate, performance could be improved on complex overlapping spike events via a boot- strapping procedure. We could first use the simple independent, Poisson generative model of this article to produce an accurate preliminary assignment of spikes to units. From this assignment we 19 could construct a more detailed model of correlated activity with pairwise interactions (e.g. [20] or the stimulus-dependent models [23, 24]). This more complex generative model could then be used to further refine spike assignments for applications requiring a very high degree of accuracy. Acknowledgments We thank Michael Berry, Michael Freed, Mike Jarvis, Olivier Marre, Liam Paninski and Joerg Sander for helpful comments. This work was supported by NSF grants IBN-0344678, EF-0928048, NIH grant RO1 EY08124, NIH training grant T32-07035, and NIH training grant 5T90DA022763- 04. References [1] Sterlng P (2004) How retinal circuits optimize the transfer of visual informatio. In: Chalupa L, Werner J, editors, The Visual Neurosciences. [2] Meister M, Pine J, Baylor DA (1994) Multi-neuronal signals from the retina: Acquisition and analysis. J Neurosci Methods 51: 95–106. [3] Devries S, Baylor D (1997) Mosaic arrangement of ganglion cell receptive fields in rabbit retina. J Neurophysiol 78: 2048. [4] Buzs´aki G (2004) Large-scale recording of neuronal ensembles. Nature Neuroscience 7: 446– 51. [5] Gray CM, Maldonado PE, Wilson M, McNaughton B (1995) Tetrodes markedly improve the reliability and yield of multiple single-unit isolation from multi-unit recordings in cat striate cortex. J Neurosci Methods 63: 43–54. [6] Mallat SG, Zhang ZF (1993) Matching pursuits with time-frequency dictionaries. IEEE Trans- actions on Signal Processing 41: 3397–3415. [7] Segev R, Goodhouse J, Puchalla J, Berry MJ (2004) Recording spikes from a large fraction of the ganglion cells in a retinal patch. Nature Neuroscience 7: 1154–61. [8] Pouzat C, Mazor O, Laurent G (2002) Using noise signature to optimize spike-sorting and to assess neuronal classification quality. J Neurosci Methods 122: 43–57. [9] Litke AM, Bezayiff N, Chichilnisky EJ, Cunningham W, Dabrowski W, et al. (2004) What does the eye tell the brain?: Development of a system for the large scale recording of retinal output activity. IEEE Trans Nucl Sci 51: 1434–1439. [10] Fee MS, Mitra P, Kleinfeld D (1996) Automatic sorting of multiple unit neuronal signals in the presence of anisotropic and non-gaussian variability. J Neurosci Methods 69: 175–188. [11] Ankerst M, Breunig MM, Kriegel HP, Sander J (1999) OPTICS: Ordering points to identify the clustering structure. SIGMOD Rec 28: 49–60. 20 [12] Harris KD, Hirase H, Leinekugel X, Henze DA, Buzs´aki G (2001) Temporal interaction be- tween single spikes and complex spike bursts in hippocampal pyramidal cells. Neuron 32: 141–9. [13] Ratliff C, Borghuis B, Kao YH, Sterling P, Balasubramanian V (2010) Retina is structured to process an excess of darkness in natural scenes. PNAS In review. [14] Do-Nascimento JLM, Do-Nascimento RSV, Damasceno BA, Silveira LCL (1991) The neurons of the retinal ganglion cell layer of the guinea pig: Quantitative analysis of their distribution and size. Brazilian J Med Biol Res 24: 199–214. [15] Quiroga RQ (2007) Spike sorting. Scholarpedia 2: 3583. [16] Lewicki MS (1994) Bayesian modeling and classification of neural signals. Neural computa- tion 6: 1005–1030. [17] Atiya A (1992) Recognition of multiunit neural signals. IEEE Transactions on Biomedical Engineering 39: 723–729. [18] Sahani M (1999) Latent variable models for neural data analysis. Ph.D. thesis, California Institute of Technology, Pasadena CA. [19] Shoham S, Fellows MR, Normann RA (2003) Robust, automatic spike sorting using mixtures of multivariate t-distributions. J Neurosci Methods 127: 111–22. [20] Schneidman E, Berry MJ, Segev R, Bialek W (2006) Weak pairwise correlations imply strongly correlated network states in a neural population. Nature 440: 1007–12. [21] Ventura V (2009) Traditional waveform based spike sorting yields biased rate code estimates. Proc Natl Acad Sci USA 106: 6921–6. [22] Lewicki MS (1998) A review of methods for spike sorting: The detection and classification of neural action potentials. Network (Bristol, England) 9: R53–78. [23] Pillow JW, Shlens J, Paninski L, Sher A, Litke AM, et al. (2008) Spatio-temporal correlations and visual signalling in a complete neuronal population. Nature 454: 995–999. [24] Tkacik G, Prentice J, Balasubramanian V, Schneidman E (2010) Optimal population coding by noisy spiking neurons. PNAS To appear. 21
1801.05219
1
1801
2018-01-16T12:08:03
Eligibility Traces and Plasticity on Behavioral Time Scales: Experimental Support of neoHebbian Three-Factor Learning Rules
[ "q-bio.NC" ]
Most elementary behaviors such as moving the arm to grasp an object or walking into the next room to explore a museum evolve on the time scale of seconds; in contrast, neuronal action potentials occur on the time scale of a few milliseconds. Learning rules of the brain must therefore bridge the gap between these two different time scales. Modern theories of synaptic plasticity have postulated that the co-activation of pre- and postsynaptic neurons sets a flag at the synapse, called an eligibility trace, that leads to a weight change only if an additional factor is present while the flag is set. This third factor, signaling reward, punishment, surprise, or novelty, could be implemented by the phasic activity of neuromodulators or specific neuronal inputs signaling special events. While the theoretical framework has been developed over the last decades, experimental evidence in support of eligibility traces on the time scale of seconds has been collected only during the last few years. Here we review, in the context of three-factor rules of synaptic plasticity, four key experiments that support the role of synaptic eligibility traces in combination with a third factor as a biological implementation of neoHebbian three-factor learning rules.
q-bio.NC
q-bio
Eligibility Traces and Plasticity on Behavioral Time Scales: Experimental Support of neoHebbian Three-Factor Learning Rules Wulfram Gerstner∗, Marco Lehmann, Vasiliki Liakoni, Dane Corneil, and Johanni Brea School of Computer Science & School of Life Sciences ´Ecole Polytechnique F´ed´erale de Lausanne, 1015 Lausanne EPFL, Switzerland (*) corresponding author January 17, 2018 Abstract Most elementary behaviors such as moving the arm to grasp an object or walking into the next room to explore a museum evolve on the time scale of seconds; in contrast, neu- ronal action potentials occur on the time scale of a few milliseconds. Learning rules of the brain must therefore bridge the gap between these two different time scales. Modern the- ories of synaptic plasticity have postulated that the co-activation of pre- and postsynaptic neurons sets a flag at the synapse, called an eligibility trace, that leads to a weight change only if an additional factor is present while the flag is set. This third factor, signaling reward, punishment, surprise, or novelty, could be implemented by the phasic activity of neuromodulators or specific neuronal inputs signaling special events. While the theoretical framework has been developed over the last decades, experimental evidence in support of eligibility traces on the time scale of seconds has been collected only during the last few years. Here we review, in the context of three-factor rules of synaptic plasticity, four key experiments that support the role of synaptic eligibility traces in combination with a third factor as a biological implementation of neoHebbian three-factor learning rules. Keywords Elibility trace, Hebb Rule, Reinforcement Learning, Neuromodulators, Surprise, synaptic tagging, Synaptic Plasticity, Behavioral Learning 1 Introduction Humans are able to learn novel behaviors such as pressing a button, swinging a tennis racket, or breaking at a red traffic light; they are also able to form memories of salient events, learn to dis- tinguish flowers, and to establish a mental map when exploring a novel environment. Memory formation and behavioral learning is linked to changes of synaptic connections (Martin et al., 1 2000). Long-lasting synaptic changes, necessary for memory, can be induced by Hebbian pro- tocols that combine the activation of presynaptic terminals with a manipulation of the voltage or the firing state of the postsynaptic neuron (Lisman, 2003). Traditional experimental protocols of long-term potentiation (LTP) (Bliss and Lømo, 1973; Bliss and Collingridge, 1993), long-term depression (LTD) (Levy and Stewart, 1983; Artola and Singer, 1993) and spike-timing depen- dent plasticity (STDP) (Markram et al., 1997; Zhang et al., 1998; Sjostrom et al., 2001) neglect that additional factors such as neuromodulators or other gating signals might be necessary to permit synaptic changes (Gu, 2002; Hasselmo, 2006; Reynolds and Wickens, 2002). Early STDP experiments that involved neuromodulators mainly focused on tonic bath application of modulatory factors (Pawlak et al., 2010). However, from the perspective of formal learning theories, to be reviewed below, the timing of modulatory factors is just as crucial (Schultz and Dickinson, 2000; Schultz, 2002). From the theoretical perspective, STDP under the control of neuromodulators leads to the framework of three-factor learning rules (Xie and Seung, 2004; Legenstein et al., 2008; Vasilaki et al., 2009) where an eligibility trace represents the Hebbian idea of co-activation of pre- and postsynaptic neurons (Hebb, 1949) while modulation of plastic- ity by additional gating signals is represented generically by a 'third factor' (Crow, 1968; Barto, 1985; Legenstein et al., 2008). Such a third factor could represent variables such as 'reward mi- nus expected reward' (Williams, 1992; Schultz, 1998; Sutton and Barto, 1998) or the saliency of an unexpected event (Ljunberg and amd W. Schultz, 1992; Redgrave and Gurney, 2006). In an earlier paper (Fr´emaux and Gerstner, 2016) we reviewed the theoretical literature of, and experimental support for, three-factor rules available by the end of 2013. During recent years, however, the experimental procedures advanced significantly and provided direct physio- logical evidence of eligibility traces and three-factor learning rules for the first time, making an updated review of three-factor rules necessary. In the following, we – a group of theoreticians – review five experimental papers indicating support of eligibility traces in striatum (Yagishita et al., 2014), cortex (He et al., 2015), and hippocampus (Brzosko et al., 2015, 2017; Bittner et al., 2017). We will close with a few remarks on the paradoxical nature of theoretical predictions in the field of computational neuroscience. 2 Hebbian rules versus three-factor rules Learning rules describe the change of the strength of a synaptic contact between a presynaptic neuron j and a postsynaptic neuron i. The strength of an excitatory synaptic contact can be defined by the amplitude of the postsynaptic potential which is closely related to the spine volume and the number of AMPA receptors (Matsuzaki et al., 2001). Synapses contain complex molecular machineries (Lisman, 2003; Redondo and Morris, 2011; Huganir and Nicoll, 2013; Lisman, 2017), but for the sake of transparency of the arguments, we will keep the mathematical notation as simple as possible and characterize the synapse by two variables only: the first one is the synaptic strength wij, measured as spine volume or amplitude of postsynaptic potential, and the second one is a synapse-internal variable eij which is not directly visible in standard electrophysiological experiments. In our view, the internal variable eij represents a metastable 2 transient state of interacting molecules in the spine head or a multi-molecular substructure in the postsynaptic density which serves as a synaptic flag indicating that the synapse is ready for an increase or decrease of its spine volume (Bosch et al., 2014). The precise biological nature of eij is not important to understand the theories and experiments that are reviewed below. We refer to eij as the 'synaptic flag' or the 'eligibility trace' and to wij as the 'synaptic weight' or 'strength' of the synaptic contact. A change of the synaptic flag indicates a 'candidate weight change' (Fr´emaux et al., 2010) whereas a change of wij indicates an actual, measurable, change of the synaptic weight. Before we turn to three-factor rules, let us discuss conventional models of Hebbian learning. 2.1 Hebbian learning rules Hebbian learning rules are the mathematical summary of the outcome of experimental protocols inducing long-term potentiation (LTP) or long-term depression (LTD) of synapses. Suitable experimental protocols include strong extracellular stimulation of presynaptic fibers (Bliss and Lømo, 1973; Levy and Stewart, 1983), manipulation of postsynaptic voltage in the presence of presynaptic spike arrivals (Artola and Singer, 1993), or spike-timing dependent plasticity (STDP) (Markram et al., 1997; Sjostrom et al., 2001). In all mathematical formulations of Heb- bian learning, the synaptic flag variable eij is sensitive to the combination of presynaptic spike arrival and a postsynaptic variable, such as the voltage at the location of the synapse. Under a Hebbian learning rule, repeated presynaptic spike arrivals at a synapse of a neuron at rest do not cause a change of the synaptic variable. Similarly, an elevated postsynaptic potential in the absence of a presynaptic spike does not cause a change of the synaptic variable. Thus Hebbian learning always needs two factors for a synaptic change: a factor caused by a presynaptic signal such as glutamate; and a factor that depends on the state of the postsynaptic neuron. What are these factors? We can think of the presynaptic factor as the time course of gluta- mate available in the synaptic cleft or bound to the postsynaptic membrane. Note that the term 'presynaptic factor' that we will use in the following does not imply that the physical location of the presynaptic factor is inside the presynaptic terminal - the factor could very well be located in the postsynaptic membrane as long as it only depends on the amount of available neurotransmit- ters. The postsynaptic factor might be the calcium in the synaptic spine (Shouval et al., 2002; Rubin et al., 2005), a calcium-related second messenger molecule (Graupner and Brunel, 2007), or simply the voltage at the site of the synapse (Brader et al., 2007; Clopath et al., 2010). We remind the reader that we always use the index j to refer to the presynaptic neuron and the index i to refer to the postsynaptic one. For the sake of simplicity, let us call the presynaptic factor xj (representing the activity of the presynaptic neuron or the amount of glutamate in the synaptic cleft) and the postsynaptic factor yi (representing the state of the postsynaptic neuron). In a Hebbian learning rule, changes of the synaptic flag eij need both xj and yi eij = η xj g(yi) − eij/τe d dt (1) 3 where η is the constant learning rate, τe is a decay time constant and g(yj) is some arbitrary, potentially nonlinear, function of the postsynaptic variable yi. Thus, the synaptic flag eij acts as a correlation detector between presynaptic activity xj and the state of the postsynaptic neuron yi. In some models, there is no decay or the decay is considered negligible on the time scale of one experiment (τe → ∞). Let us discuss two examples. In the Bienenstock-Cooper Munro (BCM) model of develop- mental cortical plasticity (Bienenstock et al., 1982) the presynaptic factor xj is the firing rate of the presynaptic neuron and g(yi) = (yi − θ) yi is a quadratic function with yi the postsynaptic firing rate and θ a threshold rate. Thus, if both pre- and postsynaptic neurons fire together at a high rate xj = yi > θ then the synaptic flag eij increases. In the BCM model, just like in most other conventional models, a change of the synaptic flag (i.e., an internal state of the synapse) leads instantaneously to a change of the weight eij −→ wij so that an experimental protocol results immediately in a measurable weight change. With the BCM rule and other similar rules (Oja, 1982; Miller and MacKay, 1994), the synaptic weight increases if both presynaptic and postsynaptic neuron are highly active, implementing the slogan 'fire together, wire together' (Lowel and Singer, 1992; Shatz, 1992); cf. Fig. 1A(i). As a second example, we consider the Clopath model (Clopath et al., 2010). In this model, ij and e− there are two correlation detectors implemented as synaptic flags e+ ij for LTP and LTD, respectively. The synaptic flag e+ ij for LTP uses a presynaptic factor x+ j (related to the amount of glutamate available in the synaptic cleft) which increases with each presynaptic spike and decays back to zero over the time of a few milliseconds (Clopath et al., 2010). The postsynaptic factor for LTP depends on the postsynaptic voltage yi via a function g(yi) = a+[yi−θ+] ¯yi where a+ is a positive constant, θ+ a voltage threshold, square brackets denote the rectifying piecewise linear function, and ¯yi a running average of the voltage with a time constant of tens of milliseconds. An analogous, but simpler, combination of presynaptic spikes and postsynaptic voltage defines the second synaptic flag e− ij for LTD (Clopath et al., 2010). The total change of the synaptic weight is the combination of the two synaptic flags for LTP and LTD: dwij/dt = de+ ij/dt. Note that, since both synaptic flags e+ ij depend on the postsynaptic voltage, postsynaptic spikes are not a necessary condition for changes, in agreement with voltage-dependent protocols (Artola and Singer, 1993; Ngezahayo et al., 2000). Thus, in voltage-dependent protocols, and similarly in voltage-dependent models, 'wiring together' is possible without 'firing together' - indicating that the theoretical framework sketched above goes beyond a narrow view of Hebbian learning; cf. Fig. 1A(ii). ij/dt − de− ij and e− If we restrict the discussion of the postsynaptic variable to super-threshold spikes, then the Clopath model becomes identical to the triplet STDP model (Pfister et al., 2006) which is in turn closely related to other nonlinear STDP models (Senn et al., 2001; Froemke and Dan, 2002; Izhikevich and Desai, 2003) as well as to the BCM model discussed above (Pfister et al., 2006; Gjorjieva et al., 2011). Classic pair-based STDP models (Gerstner et al., 1996; Kempter et al., 1999; Song et al., 2000; van Rossum et al., 2000; Rubin et al., 2001) are further examples of the general theoretical framework of Eq. (1) and so are some models of structural plasticity (Helias et al., 2008; Deger et al., 2012; Fauth et al., 2015). Hebbian models of synaptic consolidation 4 A B Figure 1: A. Two Hebbian protocols and one three-factor learning protocol. (i) Hebbian STDP protocol with presynaptic spikes (presynaptic factor) followed by a burst of postsynaptic spikes (postsynaptic factor). Synapses in the stimulated pathway (green) will typically show LTP while an unstimulated synapse (red) will not change its weight (Markram et al., 1997). (ii) Hebbian voltage pairing protocol of presynaptic spikes (presynaptic factor) with a depolarization of the postsynaptic neuron (postsynaptic factor). Depending on the amount of depolarization the stim- ulated pathway (green) will show LTP or LTD while an unstimulated synapse (red) does not change its weight (Artola and Singer, 1993; Ngezahayo et al., 2000). (iii) Results of a Hebbian induction protocol are influenced by a third factor (blue) even if it is given after a delay d. The third factor could be a neuromodulator such as dopamine, acetylcholine, noreprinephrine, or serotonin (Pawlak et al., 2010; Yagishita et al., 2014; He et al., 2015; Brzosko et al., 2015, 2017; Bittner et al., 2017). B. Specificity of three-factor learning rules. (i) Presynaptic input spikes (green) arrive at two different neurons, but only one of these also shows postsynaptic activity (orange spikes). (ii) A synaptic flag is set only at the synapse with a Hebbian co-activation of pre- and postsynaptic factors; the synapse become then eligible to interact with the third factor (blue). Spontaneous spikes of other neurons do not interfere. (iii) The interaction of the synaptic flag with the third factor leads to a strengthening of the synapse (green). have several hidden flag variables (Fusi et al., 2005; Barrett et al., 2009; Benna and Fusi, 2016) but can also be situated as examples within the general framework of Hebbian rules. Note that in most of the examples so far the measured synaptic weight is a linear function of the synaptic flag variable(s). However, this does not need to be the case. For example, in some voltage-based (Brader et al., 2007) or calcium-based models (Shouval et al., 2002; Rubin et al., 2005), the synaptic flag is transformed into a weight change only if eij is above or below some threshold, or only after some further filtering. To summarize, in the theoretical literature the class of Hebbian models is a rather general framework encompassing all those models that are driven by a combination of presynaptic ac- tivity and the state of the postsynaptic neuron. In this view, Hebbian models depend on two factors related to the activity of the presynaptic and the state of the postsynaptic neuron. The correlations between the two factors can be extracted on different time scales using one or, if necessary, several flag variables. The flag variables trigger a change of the measured synaptic weight. In the following we build on Hebbian learning, but extend the theoretical framework to include a third factor. 5 d(i)(ii)(iii)(i)(ii)(iii)first factor (pre)second factor (post)third factor (mod)synaptic flag 2.2 Three-factor learning rules We are interested in a framework where a Hebbian co-activation of two neurons leaves one or several flags (eligibility trace) at the synapse connecting these neurons. The flag is not directly visible and does not automatically trigger a change of the synaptic weight. An actual weight change is implemented only if a third signal, e.g., a phasic increase of neuromodulator activity or an additional input (signaling the occurrence of a special event) is present at the same time or in the near future. Theoreticians refer to such a plasticity model as a three-factor learning rule (Xie and Seung, 2004; Legenstein et al., 2008; Vasilaki et al., 2009; Fr´emaux et al., 2013; Fr´emaux and Gerstner, 2016). Three-factor rules have also been called 'neoHebbian' (Lisman et al., 2011; Lisman, 2017) or 'heterosynaptic (modulatory-input dependent)' (Bailey et al., 2000) and can be traced back to the 1960s (Crow, 1968), if not earlier. To our knowledge the wording 'three factors' was first used by (Barto, 1985). The terms eligibility and eligibility traces have been used in (Klopf, 1972; Sutton and Barto, 1981; Barto et al., 1983; Barto, 1985; Williams, 1992; Schultz, 1998; Sutton and Barto, 1998) but in some of the early studies it remained unclear whether eligibility traces can be set by presynaptic activity alone (Klopf, 1972; Sutton and Barto, 1981) or only by Hebbian co-activation of pre- and postsynaptic neurons (Barto et al., 1983; Barto, 1985; Williams, 1992; Schultz, 1998; Sutton and Barto, 1998). The basic idea of a modern eligibility trace is that a synaptic flag variable eij is set according to Eq. (1) by coincidences between presynaptic activity xj and a postsynaptic factor yi. The update of the synaptic weight wij, as measured via the spine volume or the amplitude of the excitatory postsynaptic potential (EPSP), is given by d dt wij = eij M3rd(t) (2) where M3rd(t) refers to the third factor (Izhikevich, 2007; Legenstein et al., 2008; Fr´emaux et al., 2013). Thus, a third factor is needed to transform the eligibility trace into a weight change; cf. Fig. 1A(iii). Note that the weight change is proportional to M3rd(t). Thus, the third factor influences the speed of learning. In the absence of the third factor (M3rd(t) = 0), the synaptic weight is not changed. We emphasize that a positive value of the synaptic flag in combination with a negative value of M3rd(t) leads to a decrease of the weight. Therefore, the third factor also influences the direction of change. What could be the source of such a third factor? The third factor could be triggered by attentional processes, surprising events, or reward. Phasic signals of neuromodulators such as dopamine, serotonin, acetylcholine, or noradrenaline are obvious candidates of a third factor, but potentially not the only ones. Note that axonal branches of most dopaminergic, seroton- ergic, cholinergic, or adrenergic neurons project broadly onto large regions of cortex so that a phasic neuromodulator signal arrives at many neurons and synapses in parallel (Schultz, 1998). Since neuromodulatory information is shared by many neurons, the variable M3rd(t) of the third factor has no neuron-specific index (neither i nor j) in our mathematical formulation. Because of its unspecific nature, the theory literature sometimes refers to the third factor as a 'global' 6 broadcasting signal, even though in practice not every brain region and every synapse is reached by each neuromodulator. Note that we mathematically define a phasic signal as the deviation from the running aver- age so that M3rd(t) in Eq. (2) can take positive and negative values. However, the third factor could also be biologically implemented by positive excursions of the activity using two different neuromodulators with very low baseline activity. The activity of the first modulator could indi- cate positive values of the third factor and that of the second modulator negative ones - similar to ON and OFF cells in the retina. Similarly, the framework of neoHebbian three-factor rules is general enough to enable biological implementations with separate eligibility traces for LTP and LTD as discussed above in the context of the Clopath model (Clopath et al., 2010). 2.3 Examples and theoretical predictions There are several known examples in the theoretical literature of neoHebbian three-factor rules. We briefly present three of these and formulate expectations derived from the theoretical frame- work which we would like to compare to experimental results in the next section. As a first example, we consider the relation of neoHebbian three-factor rules to reward- based learning. Temporal Difference (TD) algorithms such as SARSA(λ) or TD(λ) from the theory of reinforcement learning (Sutton and Barto, 1998) as well as learning rules derived from policy gradient theories (Williams, 1992) can be interpreted in neuronal networks in terms of neoHebbian three-factor learning rules. The resulting plasticity rules are applied to synapses connecting 'state-neurons' (e.g., place cells coding for the current location of an animal) to 'action neurons' (e.g., cells initiating an action program such as 'turn left') (Brown and Sharp, 1995; Suri and Schultz, 1999; Arleo and Gerstner, 2000; Foster et al., 2000; Xie and Seung, 2004; Loewenstein and Seung, 2006; Florian, 2007; Izhikevich, 2007; Legenstein et al., 2008; Vasilaki et al., 2009; Fr´emaux et al., 2013); for a review, see (Fr´emaux and Gerstner, 2016). The eligibility trace is increased during the joint activation of 'state-neurons' and 'action-neurons' and decays exponentially thereafter consistent with the framework of Eq. (1). The third factor is defined as reward minus expected reward where the exact definition of expected reward depends on the implementation details. A long line of research by Wolfram Schultz and colleagues (Schultz et al., 1997; Schultz, 1998; Schultz and Dickinson, 2000; Schultz, 2002) indicates that phasic increases of the neuromodulator dopamine have the necessary properties required for a third factor in the theoretical framework of reinforcement learning. However, despite the rich literature on dopamine and reward-based learning accumulated during the last 25 years, measurements of the decay time constant τe of the eligibility trace eij in Eq. (1) have, to our knowledge, not become available before 2015. From the mathematical framework of neoHebbian three-factor rules it is clear that, in the context of action learning, the time constant of the eligibility trace (i.e., the duration of the synaptic flag) should roughly match the time span from the initiation of an action to the delivery of reward. As an illustration, let us imagine a baby that attempts to grasp her bottle of milk. The typical duration of one grasping movement is in the range of a second, but potentially only the third grasping attempt might be 7 successful. Let us suppose that each grasping movement corresponds to the co-activation of some neurons in the brain. If the duration of the synaptic flag is much less than a second, the co- activation of pre- and postsynaptic neurons that sets the synaptic flag (eligibility trace) cannot be linked to the reward one second later and synapses do not change. If the duration of the synaptic flag is much longer than a second, then the two 'wrong' grasping attempts are reinforced nearly as strongly as the third, successful one which mixes learning of 'wrong' co-activations with the correct ones. Hence, the existing theory of three-factor learning rules predicts that the synaptic flag (eligibility trace for action learning) should be in the range of a typical elementary action, about 200ms to 2s; see, for example, p. 15 of (Schultz, 1998)1, p.3 of (Izhikevich, 2007), 2, p.3 of (Legenstein et al., 2008)3, p. 13327 of (Fr´emaux et al., 2010)4, or p. 13 of (Fr´emaux et al., 2013)5. An eligibility trace of 100ms or 20 seconds would be less useful for learning a typical elementary action or delayed reward task than an eligibility trace in the range of 200ms to 2s. The expected time scale of the synaptic eligibility trace should roughly match the maximal delay of reinforcers in conditioning experiments (Thorndike, 1911; Pavlov, 1927; J.Blac et al., 1985), linking synaptic processes to behavior. For human behavior, delaying a reinforcer by 10 seconds during ongoing actions decreases learning compared to immediate reinforcement (Okouchi, 2009). As a second example, we consider situations that go beyond standard reward-based learning. Even in the absence of reward, a surprising event might trigger a combination of neuromodula- tors such as noradrenaline, acetylcholine and dopamine that may act as third factor for synaptic plasticity. Imagine a small baby lying in the cradle with an attractive colorful object swing- ing above him. He spontaneously makes several arm movements until finally he succeeds, by chance, to grasp the object. There is no food reward for this action. However, the fact that he can now turn the object, look at it from different sides, or put it in his mouth is satisfying because it leads to many novel (and exciting!) stimuli. The basic idea is that, in such situations, novelty or surprise acts as a reinforcer even in complete absence of food rewards (Schmidhuber, 1991; Singh et al., 2004; Oudeyer et al., 2007). Theoreticians have studied these ideas in the context of curiosity (Schmidhuber, 2010), information gain during active exploration (Storck et al., 1995; Sun et al., 2011; Schmidhuber, 2006; Sun et al., 2011; Little and Sommer, 2013; Friston et al., 2016), and via formal definitions of surprise (Storck et al., 1995; Itti and Baldi, 2009; Schmid- huber, 2010; Shannon, 1948; Friston, 2010; Faraji et al., 2018). Note that surprise is not always linked to active exploration but can also occur in a passive situation, e.g. listening to tone beeps or viewing simple stimuli (Squires et al., 1976; Kolossa et al., 2013, 2015; Meyniel et al., 2016). Measurable physiological responses to surprise include pupil dilation (Hess and Polt, 1960) and the P300 component of the electroencephalogram (Squires et al., 1976). 1'Learning ... (with dopamine) on striatal synapses ... requires hypothetical traces of synaptic activity that last until reinforcement occurs and makes those synapses eligible for modification ...' 2'(the eligibility trace c ) decays to c = 0 exponentially with the time constant τc = 1s' 3'the time scale of the eligibility trace is assumed in this article to be on the order of seconds' 4'candidate weight changes eij decay to zero with a time constant τe = 500ms. The candidate weight changes eij are known as the eligibility trace in reinforcement learning' 5'the time scales of the eligibility traces we propose, (are) on the order of hundreds of milliseconds, .. Direct experimental evidence of eligibility traces still lacks, ...' 8 If surprise can play a role similar to reward, then surprise-transmitting broadcast signals should speed-up plasticity. Indeed, theories of surprise as well as hierarchical Bayesian models predict a faster change of model parameters for surprising stimuli than for known ones (Yu and Dayan, 2005; Nassar et al., 2010; Mathys et al., 2011, 2014; Faraji et al., 2018) similar to, but more general than, the well-known Kalman filters (Kalman, 1960). Since the translation of these abstract models into networks of spiking neurons is still missing, precise predictions for surprise modulation of plasticity in the form of three-factor rules are not yet available. However, if we consider noradrenaline, acetylcholine, and/or dopamine as candidate neuromodulators signaling novelty and surprise, we expect that these neuromodulators should have a strong effect on plasticity so as to boost learning of surprising stimuli. The influence of tonic applications of various neuromodulators on synaptic plasticity has been shown in numerous studies (Gu, 2002; Hasselmo, 2006; Reynolds and Wickens, 2002; Pawlak et al., 2010). However, in the context of the above examples, we are interested in phasic neuromodulatory signals. Phasic signals conveying moments of surprise are most useful for learning if they are either synchronous with the stimulus to be learned (e.g., passive listening or viewing) or arise with a delay corresponding to one exploratory movement (e.g. grasping). Hence, we predict from these considerations a decay constant τe of the synaptic flag in the range of 1 second, but with a pronounced effect for synchronous or near-synchronous events. As our final example, we would like to comment on synaptic consolidation. The synaptic tagging-and-capture hypothesis (Frey and Morris, 1997; Reymann and Frey, 2007; Redondo and Morris, 2011) perfectly fits in the framework of three-factor learning rules: The joint pre- and postsynaptic activity sets the synaptic flag (called 'tag' in the context of consolidation) which decays back to zero over the time of one hour. To stabilize synaptic weights beyond one hour an additional factor is needed to trigger protein synthesis required for long-term maintenance of synaptic weights (Redondo and Morris, 2011; Reymann and Frey, 2007). Neuromodulators such as dopamine have been identified as the necessary third factor for consolidation (Bailey et al., 2000; Reymann and Frey, 2007; Redondo and Morris, 2011; Lisman, 2017). Indeed, modern computational models of synaptic consolidation take into account the effect of neuromodulators (Clopath et al., 2008; Ziegler et al., 2015) in a framework reminiscent of the three-factor rule defined by Eqs. (1) and (2) above. However, there are two noteworthy differences. First, in contrast to reward-based learning, the decay time τe of the synaptic tag eij is in the range of one hour rather than one second, consistent with slice experiments (Frey and Morris, 1997) as well as with behavioral experiments (Moncada and Viola, 2007). Second, in slices, the measured synaptic weights wij are increased a few minutes after the end of the induction protocol and decay back with the time course of the synaptic tag whereas in the simplest implementation of the three-factor rule framework as formulated in Eqs. (1) and (2) the visible weight is only updated at the moment when the third factor is present. However, slightly more involved models where the visible weight depends on both the synaptic tag variable and the long-term stable weight (Clopath et al., 2008; Ziegler et al., 2015) correctly account for the time course of the measured synaptic weights in consolidation experiments (Frey and Morris, 1997; Reymann and Frey, 2007; Redondo and Morris, 2011). 9 In summary, the neoHebbian three-factor rule framework has a wide range of applicability. It is well established in the context of synaptic consolidation where the duration of the flag ('synaptic tag') extracted from slice experiments (Frey and Morris, 1997) is in the range of one hour, consistent with fear conditioning experiments (Moncada and Viola, 2007). This time scale is significantly longer than what is needed for behavioral learning of elementary actions or for memorizing surprising events. Theoreticians therefore hypothesized that a process analogous to setting a tag ('eligibility trace') must also exist on the time scale of one second. The next section discusses some recent experimental evidence supporting this theoretical prediction. 3 Experimental evidence for eligibility traces Recent experimental evidence for eligibility traces in striatum (Yagishita et al., 2014), cortex (He et al., 2015), and hippocampus (Brzosko et al., 2015, 2017; Bittner et al., 2017) is reviewed in the following three subsections. 3.1 Eligibility traces in dendritic spines of medial spiny striatal neurons in nu- cleus accumbens In their elegant imaging experiment of dendritic spines of nucleus accumbens neurons, Yag- ishita et al. (2014) mimicked presynaptic spike arrival by glutamate uncaging (presynaptic factor), paired it with three postsynaptic spikes immediately afterward (postsynaptic factor), re- peated this STDP-like pre-before-post sequence ten times, and combined it with optogenetic stimulation of dopamine fibers (3rd factor) at various delays (Yagishita et al., 2014). The ten repetitions of the pre-before-post sequence at 10 Hz took about one second while stimulation of dopaminergic fibers (10 dopamine pulses at 30Hz) projecting from the ventral tegmental area (VTA ) to nucleus accumbens took about 0.3s. In their paper, dopamine was counted as delayed by one second if the dopamine stimulation started immediately after the end of the one-second long induction period (delay = difference in switch-on time of STDP and dopamine), but for consistency with other data we define the delay d here as the time passed since the end of the STDP protocol. After 15 complete trials the spine volume, an indicator of synaptic strength (Matsuzaki et al., 2001), was measured and compared with the spine volume before the in- duction protocol. The authors found that dopamine promoted spine enlargement only if phasic dopamine was given in a narrow time window during or immediately after the 1s-long STDP protocol; cf. Fig. 2A. The maximum enlargement of spines occurred if the dopamine signal started during the STDP protocol (d = −0.4s), but even at a delay of d = 1s LTP was still visible. Giving dopamine too early (d = −2s) or too late (d = +4s) had no effect. Spine enlargement corre- sponded to an increase in the amplitude of excitatory postsynaptic currents indicating that the synaptic weight was indeed strengthened after the protocol (Yagishita et al., 2014). Thus, we can summarize that we have in the striatum a three-factor learning rule for the induction of LTP where the decay of the eligibility trace occurs on a time scale of 1s; cf. Fig. 2A. 10 A C B D Figure 2: Experimental support for synaptic eligibility traces. Fractional weight change (verti- cal axis) as a function of delay d of third factor (horizontal axis) for various protocols (schemat- ically indicated at the bottom of each panel). A. In striatum medial spiny cells, stimulation of presynaptic glutamatergic fibers (green) followed by three postsynaptic action potentials (STDP with pre-post-post-post at +10ms) repeated 10 times at 10Hz yields LTP if dopamine fibers are stimulated during the presentation (d < 0) or shortly afterward (d = 0s or d = 1s) but not if dopamine is given with a delay d = 4s; redrawn after Fig. 1 of (Yagishita et al., 2014), with delay d defined as time since end of STDP protocol. B. In cortical layer 2/3 pyramidal cells, stimulation of two independent presynaptic pathways (green and red) from layer 4 to layer 2/3 by a single pulse combined with a burst of four postsynaptic spikes (orange). If the pre-before-post stimulation was combined with a pulse of norepinephrine (NE) receptor agonist isoproterenol with a delay of 0 or 5s, the protocol gave LTP (blue trace). If the post-before-pre stimulation was combined with a pulse of srotonin (5-HT) of a delay of 0 or 2.5s, the protocol gave LTD (red trace); redrawn after Fig. 6 of (He et al., 2015). C. In hippocampus CA1, a post-before-pre (∆t=-20ms) induction protocol yields LTP if dopamine is present during induction or given with a delay d of zero or one minute, but yields LTD if dopamine is absent or given with a delay of 30min; redrawn after Figs 1F, 3C and 2B (square data point at delay of 1min) of (Brzosko et al., 2015). D. In hippocampus CA1, 10 extracellular stimuli of presynaptic fibers at 20Hz cause depolarization of the postsynaptic potential. The timing of a complex spike (calcium plateau potential) triggered by current injection (during 300ms) after a delay d, is crucial for the amount of LTP. If we interpret presynaptic spike arrival as the first, and postsynaptic depolarization as the second factor, the complex spike could be associated with a third factor; redrawn after Fig.3 of (Bittner et al., 2017). Hight of boxes gives a very rough estimate of standard deviation - see original papers and figures for details. 11 d>01s … … d[s]-1 0 1 2 3 4Dwij[%]DA500d… … d[s]0 1 2 3 4 5 6 7 8 9 10… ∆wij[%]0-20+20NE 5HTorNE 5HTd[min]-10 0 10 30Dwij[%]050d>08min DA… … 10 min -50d<0d[s]-4 -2 0 2 4Dwij[%]02000.5s To arrive at these results, Yagishita et al. (2014) concentrated on medial spinal neurons in the nucleus accumbens core, a part of the ventral striatum of the basal ganglia. Functionally, striatum is a particularly interesting candidate for reinforcement learning (Brown and Sharp, 1995; Schultz, 1998; Doya, 2000a; Arleo and Gerstner, 2000; Daw et al., 2005) for several reasons. First, striatum receives highly processed sensory information from neocortex and hip- pocampus through glutamatergic synapses (Mink, 1996; Middleton and Strick, 2000; Haber et al., 2006). Second, striatum also receives dopamine input associated with reward processing (Schultz, 1998). Third, striatum is, together with frontal cortex, involved in selection of motor action programs (Mink, 1996; Seo et al., 2012). On the molecular level, the striatal three-factor plasticity depended on NMDA, CaMKII, protein synthesis, and dopamine D1 receptors (Yagishita et al., 2014). CaMKII increases were found to be localized in the spine and to have roughly the same time course as the critical window for phasic dopamine suggesting that CaMKII could be involved in the 'synaptic flag' triggered by the STDP-like induction protocol while protein kinase A (PKA) was found to have an unspecific cell-wide distribution suggesting an interpretation of PKA as a molecule linked to the dopamine-triggered third factor (Yagishita et al., 2014). 3.2 Two distinct eligibility traces for LTP and LTD in cortical synapses In a recent experiment of He et al. (2015), layer 2/3 pyramidal cells in slices from prefrontal or visual cortex were stimulated by an STDP protocol, either pre-before-post for LTP induction or post-before-pre for LTD induction. A neuromodulator was applied with a delay after a single STDP sequence before the whole protocol was repeated; cf. Fig. 2B. Neuromodulators, either norepinephrine (NE), serotonin (5-HT), dopamine (DA), or acetylcholine (ACh) were ejected from a pipette for 10 seconds or from endogenous fibers (using optogenetics) for one second (He et al., 2015). It was found that NE was necessary for LTP whereas 5-HT was necessary for LTD. DA or ACh agonists had no effect in visual cortex but DA had a positive effect on LTP induction in frontal cortex (He et al., 2015). For the STDP protocol, He et al. (2015) used extracellular stimulation of two presynaptic pathways from layer 4 to layer 2/3 (presynaptic factor) combined with a burst of 4 postsynaptic action potentials (postsynaptic factor), either pre-before-post or post-before-pre. In a first vari- ant of the experiment, the STDP stimulation was repeated 200 times at 10 Hz corresponding to a total stimulation time of 20s before the NE or 5-HT was given. In a second variant, instead of an STDP protocol, they paired presynaptic stimulation (first factor) with postsynaptic depolariza- tion (second factor) to -10mV to induce LTP, or to -40mV to induce LTD. With both protocols it was found that LTP can be induced if the neuromodulator NE (third factor) arrived with a delay of 5 seconds or less after the LTP protocol, but not 10 seconds. LTD could be induced if 5-HT (third factor) arrives with a delay of 2.5 seconds or less after the LTD protocol, but not 5 seconds (He et al., 2015). A third variant of the experiment involved optogenetic stimulation of the noradrenaline, dopamine, or serotonin pathway by repeated light pulses during 1s applied immediately, or a 12 few seconds, after a minimal STDP protocol consisting of a single presynaptic and four postsy- naptic pulses (either pre-before-post or post-before-pre), a protocol that is physiologically more plausible. The minimal sequence of STDP pairing and neuromodulation was repeated 40 times at intervals of 20s. Results with optogenetic stimulation were consistent with those mentioned above and showed in addition that application of NE or 5-HT immediately before the STDP stimulus did not induce LTP or LTD. Overall these results indicate that in visual and frontal cortex, pre-before-post pairing leaves an eligibility trace that decays over 5-10 seconds and that can be converted into LTP by the neuromodulator noradrenaline. Similarly, post-before- pre pairing leaves a shorter eligibility trace that decays over 3 seconds and can be converted into LTD by the neuromodulator serotonin; cf. Fig. 2B. Functionally, a theoretical model in the same paper (He et al., 2015) showed that the mea- sured three-factor learning rules with two separate eligibility traces stabilized and prolonged net- work activity so as to allow 'event prediction'. The authors hypothesized that these three-factor rules were related to reward-based learning in cortex such as perceptual learning in monkey (Schoups et al., 2001) or mice (Poort et al., 2015) or reward prediction (Shuler and Bear, 2006). The relation to surprise was not discussed but might be a direction for further explorations. Molecularly, the transformation of the Hebbian pre-before-post eligibility trace into LTP involves beta adrenergic receptors and intracellular cyclic AMP whereas the transformation of the post-pre eligibility trace into LTD involves the 5-HT2c receptor (He et al., 2015). Both receptors are anchored at the postsynaptic density consistent with a role in the transformation of an eligibility trace into actual weight changes (He et al., 2015). 3.3 Eligibility traces in hippocampus Two experimental groups studied eligibility traces in CA1 hippocampal neurons using comple- mentary approaches. In the studies of Brzosko et al. (2015,2017), CA1 neurons in hippocampal slices were stimulated during about 8 minutes in an STDP protocol involving 100 repetitions (at 0.2Hz) of pairs of one extracellularly delivered presynaptic stimulation pulse (presynap- tic factor) and one postsynaptic action potential (postsynaptic factor) (Brzosko et al., 2015). Repeated pre-before-post with a relative timing +10 ms gave LTP (in the presence of natural endogenous dopamine) whereas post-before-pre (-20ms) gave LTD. However, with additional dopamine (third factor) in the bathing solution, post-before-pre at -20ms gave LTP (Zhang et al., 2009). Similarly, an STDP protocol with post-before-pre at -10ms resulted in LTP when endoge- nous dopamine was present, but in LTD when dopamine was blocked (Brzosko et al., 2015). Thus dopamine broadens the STDP window for LTP into the post-before-pre regime (Zhang et al., 2009; Pawlak et al., 2010). Moreover, in the presence of ACh during the STDP stim- ulation protocol, pre-before-post at +10ms also gave LTD (Brzosko et al., 2017). Thus ACh broadens the LTD window. The crucial experiment of Brzosko et al. (2015) involved a delay in the dopamine (Brzosko et al., 2015). Brzosko et al. started to perfuse dopamine either immediately after the end of the post-before-pre (-20ms) induction protocol or with a delay. Since the dopamine was given 13 for about 10 minutes, it cannot be considered as a phasic signal – but at least the start of the dopamine perfusion was delayed. Brzosko et al. found that the stimulus that would normally have given LTD turned into LTP if the delay of dopamine was in the range of one minute or less, but not if dopamine started 10 minutes after the end of the STDP protocol (Brzosko et al., 2015). Note that for the conversion of LTD into LTP, it was important that the synapses were weakly stimulated at low rate while dopamine was present. Similarly, a prolonged pre-before- post protocol at +10ms in the presence of ACh gave rise to LTD, but with dopamine given with a delay of less than 1 minute the same protocol gave LTP (Brzosko et al., 2017). To summarize, in the hippocampus a prolonged post-before-pre protocol (or a pre-before-post protocol in the presence of ACh) yields visible LTD, but also sets an invisible synaptic flag for LTP. If dopamine is applied with a delay of less than one minute, the synaptic flag is converted into a positive weight change under continued weak presynaptic stimulation; cf. Fig. 2C. Molecularly, the conversion of LTD into LTP after repeated stimulation of pre-before-post pulse pairings depended on NMDA receptors and on the cyclic adenosine monophosphate (cAMP) - PKA signaling cascade (Brzosko et al., 2015). The source of dopamine could be in the Lo- cus Coeruleus which would make a link to arousal and novelty (Takeuchi et al., 2016) or from other dopamine nuclei linked to reward (Schultz, 1998). Since the time scale of the synaptic flag reported in (Brzosko et al., 2015, 2017) was in the range of minutes, the process studied by Brzosko et al. could be related to synaptic consolidation (Frey and Morris, 1997; Reymann and Frey, 2007; Redondo and Morris, 2011; Lisman, 2017) rather than eligibility traces in rein- forcement learning where shorter time constants are needed (Izhikevich, 2007; Legenstein et al., 2008; Fr´emaux et al., 2010, 2013). The computational study in Brzosko et al. (2017) used an eligibility trace with a time constant of 2s and showed that dopamine as a reward signal induced learning of reward location while ACh during exploration enabled a fast relearning after a shift of the reward location (Brzosko et al., 2017). The second study combined in vivo with in vitro data (Bittner et al., 2017). From in vivo studies it has been known that CA1 neurons in mouse hippocampus can develop a novel, reliable, and rather broadly tuned, place field in a single trial under the influence of a 'calcium plateau potential' (Bittner et al., 2015), visible as a complex spike at the soma. Moreover, an artificially induced complex spike was sufficient to induce such a novel place field in vivo (Bittner et al., 2015, 2017). In additional slice experiments, several input fibers from CA3 to CA1 neurons were stim- ulated by 10 pulses from an extracellular electrode during one second. The resulting nearly synchronous inputs at, probably, multiple synapses caused a total EPSP that was about 10mV above baseline at the soma, and potentially somewhat larger in the dendrite, but did not cause somatic spiking of the CA1 neuron. The stimulated synapses showed LTP if the presynaptic stimulation was paired with a calcium plateau potential (complex spike) in the postsynaptic neuron. LTP occurred, even if the presynaptic stimulation stopped one or two seconds before the start of the plateau potential or if the plateau potential started before the presynaptic stim- ulation (Bittner et al., 2017). The protocol has a remarkable efficiency since potentiation was around 200% after only 5 pairings. Thus, the joint activation of many synapses sets a flag at the 14 activated synapses which is translated into LTP if a calcium plateau potential (complex spike) occurs a few seconds before or after the synaptic activation; cf. Fig. 2D. Molecularly, the plasticity processes implied NMDA receptors and calcium channels (Bittner et al., 2017). Functionally, synaptic plasticity in hippocampus is particularly important because of the role of hippocampus in spatial memory (O'keefe and Nadel, 1978). CA1 neurons get input from CA3 neurons which have a narrow place field. The emergence of a broad place field in CA1 has therefore been interpreted as linking several CA3 neurons (that cover for example the 50 cm of the spatial trajectory traversed by the rat before the current location) to a single CA1 cell that codes for the current location (Bittner et al., 2017). Note that at the typical running speed of rodents, 50 cm correspond to several seconds of running. The broad activity of CA1 cells has therefore been interpreted as a predictive representation of upcoming events or places (Bittner et al., 2017). What could such an upcoming event be? For a rodent exploring a T- maze it might for example be important to develop a more precise spatial representation at the T-junction than inside one of the long corridors. With a broad CA1 place field located at the T-junction, information about the upcoming bifurcation could become available several seconds before the animal reaches the junction. Bittner et al. interpreted their findings as the signature of an unusual form of STDP with a particularly long coincidence window on the behavioral time scale (Bittner et al., 2017). Given that the time span of several seconds between presynaptic stimulation and postsynaptic complex spike is outside the range of a potential causal relation between input and output, they classified the plasticity rule as non-Hebbian because the presynaptic neurons do not participate in firing the postsynaptic one (Bittner et al., 2017). As an alternative view, we propose to classify the findings of Bittner et al. as the signature of an eligibility trace that was left by the joint occurrence of a presynaptic spike arriving from CA3 (presynaptic factor) and a subthreshold depolarization at the location of the synapse in the postsynaptic CA1 neuron (postsynaptic factor); cf. Fig. 2D. In this view, the setting of the synaptic flag is caused by a 'Hebbian'-type induction, except that on the postsynaptic side there are no spikes but just depolarization, consistent with the role of depolarization as a postsynaptic factor (Artola and Singer, 1993; Ngezahayo et al., 2000; Sjostrom et al., 2001; Clopath et al., 2010). In this view, the findings of Bittner et al. suggest that the synaptic flag set by the induction protocol leaves an eligibility trace which decays over 2s. If a plateau potential (related to the third factor) is generated during these two seconds, the eligibility trace caused by the induction protocol is transformed into a measurable change of the synaptic weight. The third factor M3rd(t) in Eq. (2) could correspond to the complex spike, filtered with a time constant of about one second. Importantly, plateau potentials can be considered as neuron-wide signals (Bittner et al., 2015) triggered by surprising novel or rewarding events (Bittner et al., 2017). In this view, the results of Bittner et al. are consistent with the framework of neoHebbian three-factor learning rules. If the plateau potentials are indeed linked to surprising events, the three-factor rule framework predicts that in vivo many neurons in CA1 receive such a third input as a broadcast-like signal. However, only those neurons that also get, at the same time, sufficiently strong input from CA3 might develop the visible plateau potential (Bittner et al., 2015). 15 The main difference between the two alternative views is that, in the model discussed in Bittner et al. (2017), each activated synapse is marked by an eligibility trace (which is inde- pendent of the state of the postsynaptic neuron) whereas in the view of the three-factor rule, the eligibility trace is set only if the presynaptic activation coincides with a strong depolariza- tion of the postsynaptic membrane. Thus, in the model of Bittner et al. the eligibility trace is set by the presynaptic factor alone whereas in the three-factor rule description it is set by the combination of pre- and postsynaptic factors. The two models can be distinguished in future experiments where either the postsynaptic voltage is controlled during presynaptic stimulation or where the number of simultaneously stimulated input fibers is minimized. The prediction of the three-factor rule is that spike arrival at a single synapse, or spike arrival in conjunction with a very small depolarization of less than 2 mV above rest, is not sufficient to set an eligibility trace. Therefore, LTP will not occur in these cases even if a calcium plateau potential occurs one second later. 4 Discussion and Conclusion 4.1 Policy gradient versus TD-learning Algorithmic models of TD-learning with discrete states and in discrete time do not need eligi- bility traces that extend beyond one time step (Sutton and Barto, 1998). In a scenario where the only reward is given in a target state that is several action steps away from the initial state, reward information shifts, over multiple trials, from the target state backwards, even if the one- step eligibility trace connects only one state to the next (Sutton and Barto, 1998). Nevertheless, extended eligibility traces across multiple time steps are considered convenient heuristic tools to speed up learning in temporal difference algorithms such as T D(λ) or SARSA(λ) (Singh and Sutton, 1996; Sutton and Barto, 1998). In policy gradient methods (Williams, 1992) as well as in continuous space-time TD-learning (Doya, 2000b; Fr´emaux et al., 2013) eligibility traces appear naturally in the formulation of the problem of reward maximization. Importantly, a large class of TD-learning and policy gradi- ent methods can be formulated as three-factor rules for spiking neurons where the third factor is defined as reward minus expected reward (Fr´emaux and Gerstner, 2016). In policy gradient methods and related three-factor rules, expected reward is calculated as a running average of the reward (Fr´emaux et al., 2010) or fixed to zero by choice of reward schedule (Florian, 2007; Leg- enstein et al., 2008). In TD-learning the expected reward in a given time step is defined as the difference of the value of the current state and that of the next state (Sutton and Barto, 1998). In the most recent large-scale applications of reinforcement learning the expected immediate reward in policy gradient is calculated by a TD-algorithm for state-dependent value estimation (Greensmith et al., 2004; Mnih et al., 2016). An excellent modern summary of Reinforcement Learning Algorithms and their historical predecessors can be found in (Sutton and Barto, 2018). 16 4.2 Specificity If phasic neuromodulator signals are broadcasted over large areas of the brain, the question arises whether synaptic plasticity can still be selective. In the framework of three-factor rules, specificity is inherited from the synaptic flags which are set by the combination of presynaptic spike arrival and an elevated postsynaptic voltage at the location of the synapses. The require- ment is met only for a small subset of synapses, because presynaptic alone or postsynaptic activity alone are not sufficient; cf. Fig. 1B. Furthermore, among all the flagged synapses only those that show, over many trials, a correlation with the reward signal will be consistently re- inforced (Legenstein et al., 2008; Loewenstein and Seung, 2006). Specificity can further be enhanced by an attentional feedback mechanism (Roelfsema and van Ooyen, 2005; Roelfsema et al., 2010) that restricts the number of eligible synapses to the 'interesting' ones, likely to be involved in the task. Such an attentional gating signal acts as an additional factor and turns the three-factor into a four-factor learning rule (Rombouts et al., 2015). 4.3 Mapping to Neuromodulators The third factor is likely to be related to neuromodulators, but from the perspective of a theo- retician there is no need to assign one neuromodulator to surprise and another one to reward. Indeed, the theoretical framework also works if each neuromodulator codes for a different com- bination of variables such as surprise, novelty or reward, just as we can use different coordinate systems to describe the same physical system (Fr´emaux and Gerstner, 2016). Thus, whether dopamine is purely reward related or also novelty related (Ljunberg and amd W. Schultz, 1992; Schultz, 1998; Redgrave and Gurney, 2006) is not critical for the development of three-factor learning rules as long as dimensions relating to novelty, surprise, and reward are all covered by the set of neuromodulators. Complexity in biology is increased by the fact that dopamine neurons projecting from the VTA to the striatum can have separate circuits and functions changing from reward in ventral striatum to novelty in the the tail of striatum (Menegas et al., 2017). Similarly, dopaminergic fibers starting in the VTA can have a different function than those starting in Locus Coeruleus (Takeuchi et al., 2016). The framework of three-factor rules is general enough to allow for these, and many other, variations. 4.4 Alternatives to eligibility traces for bridging the gap between the behavioral and neuronal timescales From a theoretical point of view, there is nothing – apart conceptual elegance – to favor eli- gibility traces over alternative neuronal mechanisms to associate events that are separated by a second or more. For example, memory traces hidden in the rich firing activity patterns of a recurrent network (Maass et al., 2002; Jaeger and Haas, 2004; Buonomano and Maass, 2009; Sussillo, 2009) or short-term synaptic plasticity in recurrent networks (Mongillo et al., 2008) could be involved in learning behavioral tasks with delayed feedback. In some models, neuronal, 17 rather than synaptic, activity traces have been involved in learning a delayed paired-associate task (Brea et al., 2016) and a combination of synaptic eligibity traces with prolonged single- neuron activity has been used for learning on behavioral time scales (Rombouts et al., 2015). The empirical studies reviewed here support the idea that the brain makes use of the elegant solution with synaptic eligibility traces and three-factor learning rules, but do not exclude that other mechanisms work in parallel. 4.5 The paradoxical nature of predictions in computational neuroscience If a neuroscientist thinks of a theoretical model, he often imagines a couple of assumptions at the beginning, a set of results derived from simulations or mathematical analysis, and ideally a few novel predictions - but is this the way modeling works? There are at least two types of pre- dictions in computational neuroscience, detailed predictions and conceptual predictions. Well- known examples of detailed predictions have been generated from variants of multi-channel biophysical Hodgkin-Huxley type (Hodgkin and Huxley, 1952) models such as: 'if channel X is blocked then we predict that ... ' where X is a channel with known dynamics and predictions include depolarization, hyperpolarization, action potential firing, action potential backpropaga- tion or failure thereof. All of these are useful predictions readily translated to and tested in experiments. Conceptual predictions derived from abstract conceptual models are potentially more inter- esting, but more difficult to formulate. Conceptual models develop ideas and form our thinking of how a specific neuronal system could work to solve a behavioral task such as working mem- ory (Mongillo et al., 2008), action selection and decision making (Sutton and Barto, 1998), long-term stability of memories (Lisman, 1985; Crick, 1984; Fusi et al., 2005), memory for- mation and memory recall (Willshaw et al., 1969; Hopfield, 1982). Paradoxically these models often make no detailed predictions in the sense indicated above. Rather, in these and other con- ceptual theories, the most relevant model features are formulated as assumptions which may be considered, in a loose sense, as playing the role of conceptual predictions. To formulate it as a short slogan: Assumptions are predictions. Let us return to the conceptual framework of three-factor rules: the purification of rough ideas into the role of three factors is the important conceptual work - and part of the assumptions. Moreover, the specific choice of time constant in the range of one second for the eligibility trace has been formulated by theoreticians as one of the model assumptions, rather than as a prediction; cf. the footnotes in section 'Examples and theoretical predictions'. Why is this the case? Most theoreticians shy away from calling their conceptual modeling work a 'prediction', because there is no logical necessity that the brain must work the way they assume in their model - the brain could have found a less elegant, different, but nevertheless functional solution to the problem under consideration; see the examples in the previous subsection. What a good conceptual model in computational neuroscience shows is that there exists a (nice) solution that should ideally not be in obvious contradiction with too many known facts. Importantly, conceptual models necessarily rely on assumptions which in many cases have not (yet) been 18 shown to be true. The response of referees to modeling work in experimental journals therefore often is: 'but this has never been shown'. Indeed, some assumptions may look far-fetched or even in contradiction with known facts: for example, to come back to eligibility traces, experiments on synaptic tagging-and-capture have shown in the 1990s that the time scale of a synaptic flag is in the range of one hour (Frey and Morris, 1997; Reymann and Frey, 2007; Redondo and Morris, 2011; Lisman, 2017), whereas the theory of eligibility traces for action learning needs a synaptic flag on the time scale of one second. Did synaptic tagging results imply that three-factor rules for action learning were wrong, because they used the wrong time scale? Or, on the contrary, did these experimental results rather imply that a biological machinery for three-factor rules was indeed in place which could therefore, for other neuron types and brain areas, be used and re-tuned to a different time scale (Fr´emaux et al., 2013)? As mentioned earlier, the concepts of eligibility traces and three-factor rules can be traced back to the 1960ies, from models formulated in words (Crow, 1968), to firing rate models for- mulated in discrete time and discrete states (Klopf, 1972; Sutton and Barto, 1981; Barto et al., 1983; Barto, 1985; Williams, 1992; Schultz, 1998; Sutton and Barto, 1998; Bartlett and Baxter, 1999), to models with spikes in a continuous state space and an explicit time scale for eligibility traces (Xie and Seung, 2004; Loewenstein and Seung, 2006; Florian, 2007; Izhikevich, 2007; Legenstein et al., 2008; Vasilaki et al., 2009; Fr´emaux et al., 2013). Despite the mismatch with the known time scale of synaptic tagging in hippocampus (and lack of experimental support in other brain areas), theoreticians persisted, polished their theories, talked at conferences about these models, until eventually the experimental techniques and the scientific interests of exper- imentalists were aligned to directly test the assumptions of these theories. In view of the long history of three-factor learning rules, the recent elegant experiments (Yagishita et al., 2014; He et al., 2015; Brzosko et al., 2015, 2017; Bittner et al., 2017) provide an instructive example of how conceptual theories can influence experimental neuroscience. Funding and Acknowledgments This project has been funded by the European Research Council (grant agreement no. 268 689, "MultiRules"), by the European Union Horizon 2020 Framework Program under grant agreement no. 720270 (Human Brain Project), and by the Swiss National Science Foundation (no. 200020 165538). References Arleo, A. and W. Gerstner (2000). Spatial cognition and neuro-mimetic navigation: a model of hippocampal place cell activity. Biological Cybernetics 83(3), 287–299. Artola, A. and W. Singer (1993). Long-term depression of excitatory synaptic transmission and its relationship to long-term potentiation. Trends Neurosci. 16(11), 480–487. Bailey, C., M. Giustetto, Y.-Y. Huang, R. Hawkins, and E. Kandel (2000). Is heterosynaptic modulation essential for stabilizing hebbian plasiticity and memory. Nat. Rev. Neurosci. 1, 11–20. 19 Barrett, A., G. Billings, R. Morris, and M. van Rossum (2009). State based model of long-term potentiation and synaptic tagging and capture. PLOS Comput. Biol. 5, e1000259. Bartlett, P. L. and J. Baxter (1999). Hebbian synaptic modification in spiking neurons that learn. Technical report, Australian National University. Barto, A., R. Sutton, and C. Anderson (1983). Neuronlike adaptive elements that can solve difficult learning and control problems. IEEE transactions on systems, man, and cybernetics 13, 835–846. Barto, A.G. (1985). Learning by statistical ccoperation of self-interested neuron-like computing elements. Human Neurobiol. 4, 229-256. Benna, M. and S. Fusi (2016). Computational principles of synaptic memory consolidation. Nat. Neurosci. 19, 1697–1706. Bienenstock, E., L. Cooper, and P. Munroe (1982). Theory of the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex. J. Neurosci. 2, 32–48. Bittner, K., C. Grienberger, S. Vaidya, A. Milstein, J. Macklin, J. Suh, S. Tonegawa, and J. Magee (2015). Con- junctive input processing drives feature selectivity in hippocampal ca1 neurons. Nat. Neurosci. 357, 1133–1142. doi:10.1038/nn.4062. Bittner, K., A. Milstein, C. Grienberger, S. Romani, and J. Magee (2017). Behavioral time scale synaptic plasticity underlies ca1 place fields. Nature 357, 1033–1036. Bliss, T. and T. Lømo (1973). Long-lasting potentation of synaptic transmission in the dendate area of anaesthetized rabbit following stimulation of the perforant path. J. Physiol. 232, 351–356. Bliss, T. V. P. and G. L. Collingridge (1993). A synaptic model of memory: long-term potentiation in the hippocam- pus. Nature 361, 31–39. Bosch, M., J. Castro, T. Saneyoshi, H. Matsuno, M. Sur, and Y. Hayashi (2014). Structural and molecular remodeling of dendritic spine substructures during long-term-potentation. Neuraon 82, 444–459. Brader, J., W. Senn, and S. Fusi (2007). Learning real-world stimuli in a neural network with spike-driven synaptic dynamics. Neural Computation 19, 2881–2912. Brea, J., A. Gaal, R. Urbancik, and W. Senn (2016). Prospective coding by spiking neurons. PLOS Comput. Biol. 12, e100500. doi:10.1371/journal.pcbi.100500. Brown, M. and P. Sharp (1995). Simulation of spatial-learning in the Morris water maze by a neural network model of the hippocampal-formation and nucleus accumbens. Hippocampus 5, 171–188. Brzosko, Z., W. Schultz, and O. Paulsen (2015). Retroactive modulation of spike timingdependent plasticity by dopamine. eLife 4, e09685. DOI: 10.7554/eLife.09685. Brzosko, Z., S. Zannone, W. Schultz, , C. Clopath, and O. Paulsen (2017). Sequential neuromodulation of hebbian plasticity offers mechanism for effective reward-based navigation. eLife 6, e27756. DOI: 10.7554/eLife.27756. Buonomano, D. and W. Maass (2009). State-dependent computations: spatiotemporal processing in cortical net- works. Nat. Rev. Neuroscience 10, 113–125. Clopath, C., L. Busing, E. Vasilaki, and W. Gerstner (2010). Connectivity reflects coding: A model of voltage-based spike-timing-dependent-plasticity with homeostasis. Nature Neuroscience 13, 344–352. Clopath, C., L. Ziegler, E. Vasilaki, L. Busing, and W. Gerstner (2008). Tag-trigger-consolidation: A model of early and late long-term-potentiation and depression. PLOS Comput. Biol. 4, e1000248. 20 Crick, F. (1984). Neurobiology - memory and molecular turnover. Nature 312, 101. Crow, T. (1968). Cortical synapses and reinforcement: a hypothesis. Nature 219, 736–737. Daw, N., Y. Niv, and P. Dayan (2005). Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nature Neuroscience 8, 1704–1711. Deger, M., M. Helias, S. Rotter, and M. Diesmann (2012). Spike-timing dependence of structural plasticity explains cooperative synapse formation in the neocortex. PLOS Comput. Biol. 8, e1002689. Doya, K. (2000a). Complementary roles of basal ganglia and cerebellum in learning and motor control. Current opinion in neurobiology 10(6), 732–739. Doya, K. (2000b). Temporal difference learning in continuous time and space. Neural Comput. 12, 219–245. Faraji, M., K. Preuschoff, and W. Gerstner (2018). Balancing new against old information: The role of puzzlement surprise in learning. Neural Computation 30, xx–xx. Fauth, M., F. Worgotter, and C. Tetzlaff (2015). The formation of multi-synaptic connections by the interaction of synaptic and structural plasticity and their functional consequences. PLOS Comput. Biol. 11, e1004031. Florian, R. V. (2007). Reinforcement learning through modulation of spike-timing-dependent synaptic plasticity. Neural Computation 19, 1468–1502. Foster, D., R. Morris, and P. Dayan (2000). Models of hippocampally dependent navigation using the temporal difference learning rule. Hippocampus 10, 1–16. Fr´emaux, N. and W. Gerstner (2016). Neuromodulated spike-timing dependent plasticity and theory of three-factor learning rules. Front. Neural Circuits 9, 85. Fr´emaux, N., H. Sprekeler, and W. Gerstner (2010). Functional requirements for reward-modulated spike-timing- dependent plasticity,. J. Neurosci. 40, 13326–13337. Fr´emaux, N., H. Sprekeler, and W. Gerstner (2013). Reinforcement learning using continuous time actor-critic framework with spiking neurons. PLOS Comput. Biol. 9, 4. Frey, U. and R. Morris (1997). Synaptic tagging and long-term potentiation. Nature 385, 533 – 536. Friston, K. (2010). The free-energy principle: a unified brain theory? Nat. Rev. Neurosci. 11, 127–138. Friston, K., T. Fitzgerald, F. Rigoli, P. Schwartenbeck, J. O'Doherty, and G. Pezzulo (2016). Active inference and learning. Neurosci. and Behav. Rev. 68, 862–879. Froemke, R. and Y. Dan (2002). Spike-timing dependent plasticity induced by natural spike trains. Nature 416, 433–438. Fusi, S., P. Drew, and L. Abbott (2005). Cascade models of synaptically stored memories. Neuron 45, 599–611. Gerstner, W., R. Kempter, J. van Hemmen, and H. Wagner (1996). A neuronal learning rule for sub-millisecond temporal coding. Nature 383(6595), 76–78. Gjorjieva, J., C. Clopath, J. Audet, and J. Pfister (2011). A triplet spike-timingdependent plasticity model generalizes the bienenstockcoopermunro rule to higher-order spatiotemporal correlations. Proc. Natl. Sci. Acad. USA 108, 19383–19388. Graupner, M. and N. Brunel (2007). Stdp in a bistable synapse model based on CaMKII and associate signaling pathways. PLOS Comput. Biol. 3, e221. doi:10.1371/journal.pcbi.0030221. 21 Greensmith, E., P. Bartlett, and J. Baxter (2004). Variance reduction technniques for gradient estimates in reinforce- ment learning. J. Machine Learn. Res/ 5, 1471–1530. Gu, Q. (2002). Neuromodulatory transmitter systems in the cortex and their role in cortical plasticity. Neuro- science 111, 815–835. Haber, S., K.-S. Kim, P. Mailly, and R. Calzavara (2006). Reward-related cortical inputs define a large striatal region in primates that interface with associative cortical connections, providing a substrate for incentive-based learning. J. Neurosci. 26, 8368–8376. Hasselmo, M. (2006). The role of acetylcholine in learning and memory. Curr. Opinion Neurobiol. 16, 710–715. He, K., M. Huertas, S. Z. Hong, X. T. abnd Johannes W Hell, H. Shouval, and A. Kirkwood Neuron 88, 528–538. (2015). http://dx.doi.org/10.1016/j.neuron.2015.09.037. Distinct eligibility traces for ltp and ltd in cortical synapses. Hebb, D. O. (1949). The Organization of Behavior. New York: Wiley. Helias, M., S. Rotter, M.-O. Gewaltig, and M. Diesmann (2008). Structural plasticity controlled by calcium based correlation detection. Front. Comput. Neurosci. 2, 7. Hess, E. and J. Polt (1960). Pupil size as related to interest value of visual stimuli. Science 132, 349–350. Hodgkin, A. L. and A. F. Huxley (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. J Physiol 117(4), 500–544. Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. USA 79, 2554–2558. Huganir, R. and R. Nicoll (2013). AMPARs and synaptic plasticity: the last 25 years. Neuron 80, 704–717. Itti, L. and P. Baldi (2009). Bayesian surprise attracts human attention. Vision Research 49, 1295–1306. Izhikevich, E. (2007). Solving the distal reward problem through linkage of STDP and dopamine signaling. Cerebral Cortex 17, 2443–2452. Izhikevich, E. and N. Desai (2003). Relating stdp to bcm. Neural Computation 15, 1511–1523. Jaeger, H. and H. Haas (2004). Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication. Science 304, 78–80. J.Blac, J. Belluzzi, and L. Stein (1985). Reinforcement delay of one second severley impairs acquisition of brain self-stimulation. Brain Res. 359, 113–119. Kalman, R. (1960). A new approach to linear filtering and prediciton problems. J. Basic Eng. 82, 35–45. Kempter, R., W. Gerstner, and J. L. van Hemmen (1999). Hebbian learning and spiking neurons. Phys. Rev. E 59, 4498–4514. Klopf, A. (1972). Brain function and adaptive systems - a heterostatic theory. Air Force Cambridge research laboratories, special reports 133, 1–70. Kolossa, A., T. Fingscheidt, K. Wessel, and B. Kopp (2013). A model-based approach to trial-by-trial p300 amplitude fluctuations. Front. Hum. Neurosci. 6, 359. Kolossa, A., B. Kopp, and T. Finscheidt (2015). A computational analysis of the neural bases of bayesian inference. NeuroImage 106(222-237). 22 Legenstein, R., D. Pecevski, and W. Maass (2008). A learning theory for reward-modulated spike-timing-dependent plasticity with application to biofeedback. PLOS Comput. Biol. 4, e1000180. Levy, W. B. and D. Stewart (1983). Temporal contiguity requirements for long-term associative potentia- tion/depression in hippocampus. Neurosci, 8, 791–797. Lisman, J. (1985). A mechanism for memory storage insensitive to molecular turnover: a bistable autophosphory- lating kinase. Proc. Natl. Acad. Sci. USA 82, 3055–3057. Lisman, J. (2003). Long-term potentiation: outstanding questions and attempted synthesis. Phil. Trans. R. Soc. Lond B: Biological Sciences 358, 829 – 842. Lisman, J. (2017). Glutamatergic synapses are structurally and biochemically complex because of multiple plasticity processes: long-term potentiation, long-term depression, short-term potentiation and scaling. Phil. Trans. Roy. Soc. B 372, 20160260. Lisman, J., A. Grace, and E. Duzel (2011). A neoHebbian framework for episodic memory; role of dopamine- dependent late ltp. Trends Neurosci. 34, 536–547. Little, D. and F. Sommer (2013). Learning and exploration in action-perception loops. Front. Neural Circuits 7, 37. Ljunberg, T. and P. A. amd W. Schultz (1992). Responses of monkey dopamine neurons during learning of behavioral interactions. J. Neurophysiol. 67, 145–163. Loewenstein, Y. and H. Seung (2006). Operant matching is a generic outcome of synaptic plasticity based on the covariance between reward and neural activity. Proc. Natl. Acad. Sci. USA 103, 15224–15229. Lowel, S. and W. Singer (1992). Selection of intrinsic horizontal connections in the visual cortex by correlated neuronal activity. Science 255, 209–212. Maass, W., T. Natschlager, and H. Markram (2002). Real-time computing without stable states: a new framework for neural computation based on perturbations. Neural Computation, 2531–2560. Markram, H., J. Lubke, M. Frotscher, and B. Sakmann (1997). Regulation of synaptic efficacy by coincidence of postysnaptic AP and EPSP. Science 275, 213–215. Martin, S., P. Grimwood, and R. Morris (2000). Synaptic plasticity and memory: an evaluation of the hypothesis. Ann. Rev. Neurosci. 23, 649–711. Mathys, C., J. Daunizeau, K. Friston, and K. Stephan (2011). A bayesian foundation for individual learning under uncertainty. Front. Hum. Neurosci. 5, 39. Mathys, C. D., E. I. Lomakina, K. Daunizeau, S. Iglesias, K. Brodersen, K. Friston, and K. Stephan (2014). Uncer- tainty in perception and the hierarchical gaussian filter. Front. Hum. Neurosci. 8, 825. Matsuzaki, M., G. Ellis-Davies, T. Nemoto, M. Iione, Y. Miyashita, and H. Kasai (2001). Dendritic spine geometry is critical for ampa receptor expression in hippocampal ca1 pyramidal neurons. Nat. Neurosci. 4, 1086–1092. Menegas, W., B. Babayan, N. Uchida, and M. Watabe-Uchida (2017). Opposite initialization to novel cues in dopamine signaling in ventral and posterior striatum in mice. eLife 6, e21886. doi: 10.7554/eLife.21886. Meyniel, F., M. Maheu, and S. Dehaene (2016). Human inferences about sequences: A minimal transition probability model. PLOS Comput. Biol. 12, e1005260. Middleton, F. A. and P. Strick (2000). Basal ganglia and cerebellar loops: motor and cognitive circuites. Brain Res. Rev. 31, 236–250. 23 Miller, K. D. and D. J. C. MacKay (1994). The role of constraints in hebbian learning. Neural Computation 6, 100–126. Mink, J. (1996). The basal ganglia: focused selection and inhibition of competing motor programs. Progr. Neuro- biol. 50, 381–425. Mnih, V., A. Badia, M. Mirza, A. Graves, T. Harley, T. Lillicrap, D. Silver, and K. Kavukcuoglu (2016). Asyn- chronous methods for deep reinforcement learning. arXiv 1602.01783v2, 1–19. Moncada, D. and H. Viola (2007). Induction of long-term memory by exposure to novelty requires protein synthesis: Evidence for a behavioral tagging. J. Neurosci 27, 7476–7481. Mongillo, G., O. Barak, and M. Tsodyks (2008). Synaptic theory of working memory. Science 319, 1543–1546. Nassar, M., R. Wilson, B. Heasly, and J. Gold (2010). An approximately bayesian delta-rule model explains the dynamics of belief updating in a changing environment. J. Neurosci. 30, 12366–12378. Ngezahayo, A., M. Schachner, and A. Artola (2000). Synaptic activation modulates the induction of bidirectional synaptic changes in adult mouse hippocamus. J. Neuroscience 20, 2451–2458. Oja, E. (1982). A simplified neuron model as a principal component analyzer. J. Mathematical Biology 15, 267–273. O'keefe, J. and L. Nadel (1978). The hippocampus as a cognitive map, Volume 3. Clarendon Press Oxford. Okouchi, H. (2009). Response acquisition by humans with delayed reinforcement. J. Exp. Anal. of Beh. 91, 377–390. Oudeyer, P., F. Kaplan, and V. Hafner (2007). Intrinsic motivation systems for autonomous mental development. IEEE Trans. Evol. Comput. 11, 265–286. Pavlov, I. P. (1927). Conditioned reflexes. Oxford Univ. Press. Pawlak, V., J. Wickens, A. Kirkwood, and J. Kerr (2010). Timing is not everything: neuromodulation opens the STDP gate. Front. Synaptic Neurosci. 2, 146. Pfister, J.-P., T. Toyoizumi, D. Barber, and W. Gerstner (2006). Optimal spike-timing dependent plasticity for precise action potential firing in supervised learning. Neural Computation 18, 1318–1348. Poort, H., A. Khan, M. Pachitariu, A. Nemri, I. Orsolic, J. Krupic, M. Bauza, M. Sahani, G. Keller, T. Mrsic-Flogel, and S. Hofer (2015). Learning enhances sensory and multiple non-sensory representations in primary visual cortex. Neuron 86, 1478–1490. Redgrave, P. and K. Gurney (2006). The short-latency dopamine signal: a role in discovering novel actions? Nat. Rev. Neurosci. 7, 967–975. Redondo, R. L. and R. G. M. Morris (2011). Making memories last: the synaptic tagging and capture hypothesis. Nat Rev Neurosci 12, 17–30. Reymann, K. and J. Frey (2007). The late maintenance of hippocampal ltp: requirements, phases,synaptic tagging, late associativity and implications. Neuropharmacology 52, 24–40. Reynolds, J. and J. Wickens (2002). Dopamine-dependent plasticity of corticostriatal synapses. Neural Networks 15, 507–521. Roelfsema, P. and A. van Ooyen (2005). Attention-gated reinforcement learning of internal representations for classification. Neural Computation 17, 2176–2214. 24 Roelfsema, P., A. van Ooyen, and T. Watanabe (2010). Perceptual learning rules based on reinforcers and attention. Trends in Cogn. Sci. 14, 64–71. doi:10.1016/j.tics.2009.11.00. Rombouts, J., S. Bothe, and R. Roelfsema (2015). How attention can create synaptic tags for the learning of working memories in sequential tasks. PLOS Comput. Biol. 11, e1004060. Rubin, J., R. Gerkin, G.-Q. Bi, and C. Chow (2005). Calcium time course as a signal for spike-timing-dependent plasticity. J. Neurophysiology 93, 2600–2613. Rubin, J., D. D. Lee, and H. Sompolinsky (2001). Equilibrium properties of temporally asymmetric Hebbian plas- ticity. Physical Review Letters 86, 364–367. Schmidhuber, J. (1991). Curious model-building control systems. In Proceedings of the International Joint Confer- ence on Neural Networks, Singapore, Volume 2, pp. 1458–1463. IEEE press. Schmidhuber, J. (2006). Developmental robotics, optimal artificial curiosity, creativity, music, and the fine arts. Connection Science 18, 173–187. Schmidhuber, J. (2010). Formal theory of creativity, fun, and intrinsic motivation (1990–2010). IEEE Trans. Au- tonom. Mental Developm. 2, 230–247. Schoups, A., R. Vogels, N. Quian, and G. A. Orban (2001). Practising orientation identification improves orientation coding in v1 neurons. Nature 412, 549–553. Schultz, W. (1998). Predictive reward signal of dopamine neurons. J. Neurophysiooogy 80, 1–27. Schultz, W. (2002, 10). Getting formal with dopamine and reward. Neuron 36(2), 241–263. Schultz, W., P. Dayan, and R. Montague (1997). A neural substrate for prediction and reward. Science 275, 1593– 1599. Schultz, W. and A. Dickinson (2000). Neuronal coding of prediction errors. Annual Reviews of Neuroscience 23, 472–500. Senn, W., M. Tsodyks, and H. Markram (2001). An algorithm for modifying neurotransmitter release probability based on pre- and postsynaptic spike timing. Neural Computation 13, 35–67. Seo, M., E. Lee, and B. Averbeck (2012). Action selection and action value in frontal-striatal circuits. Neuron 74, 947–960. Shannon, C. (1948). A mathematical theory of communication. Bell System Technical Journal 27, 37–423. Shatz, C. (1992). The developing brain. Sci. Am. 267, 60–67. Shouval, H. Z., M. F. Bear, and L. N. Cooper (2002). A unified model of nmda receptor dependent bidirectional synaptic plasticity. Proc. Natl. Acad. Sci. USA 99, 10831–10836. Shuler, M. and M. Bear (2006). Reward timing in the primary visual cortex. Science 311, 1606–1609. Singh, S., A. Barto, and N. Chentanez (2004). Intrinsically motivated reinforcement learning. Advances of Neural Information Processing Systems, NIPS 17, 1281–1288. Singh, S. and R. Sutton (1996). Reinforcement learning with replacing eligibility traces. Machine Learning 22, 123–158. Sjostrom, P., G. Turrigiano, and S. Nelson (2001). Rate, timing, and cooperativity jointly determine cortical synaptic plasticity. Neuron 32, 1149–1164. 25 Song, S., K. Miller, and L. Abbott (2000). Competitive Hebbian learning through spike-time-dependent synaptic plasticity. Nature Neuroscience 3, 919–926. Squires, K., C. Wickens, N. Squires, and E. Donchin (1976). The effect of stimulus sequence on the waveform of the cortical event-related potential. Science 193, 1141–1146. Storck, J., S. Hochreiter, and J. Schmidhuber (1995). Reinforcement-driven information acquisition in non- deterministic environments. In Proc. ICANN'95, vol.2, pp. 159–164. EC2-CIE, Paris. Sun, Y., F. Gomez, and J. Schmidhuber (2011). Planning to be surprised: Optimal bayesian exploration in dynamic environments. In Artificial General Intelligence, pp. 41–51. Springer. Suri, R. E. and W. Schultz (1999). A neural network with dopamine-like reinforcement signal that learns a spatial delayed response task. Neuroscience 91, 871890. Sussillo, D. (2009). Generating coherent patterns of activity from chaotic neural networks. Neuron 63, 544–557. Sutton, R.S and A.G. Barto (1998). Reinforcement learning. MIT Press, Cambridge. Sutton, R.S. and A.G. Barto (2018). Reinforcement learning: an introduction. MIT Press, Cambridge, 2nd edition Sutton, R. S. and A. G. Barto (1981). Towards a modern theory of adaptive networks: expectation and prediction. Psychol. Review 88, 135–171. Takeuchi, T., A. Duszkiewicz, A. Sonneborn, P. Spooner, M. Yamasaki, M. Watanabe, C. Smith, G. Fernandez, K. Deisseroth, R. Greene, and R. Morris (2016). Locus coeruleus and dopaminergic consolidation of everyday memory. Nature 537, 357362. Thorndike, E. (1911). Animal Intelligence. Darien, CT: Hafner. van Rossum, M. C. W., G. Q. Bi, and G. G. Turrigiano (2000). Stable Hebbian learning from spike timing-dependent plasticity. J. Neuroscience 20, 8812–8821. Vasilaki, E., N. Fr´emaux, R. Urbanczik, W. Senn, and W. Gerstner (2009). Spike-based reinforcement learning in continuous state and action space: When policy gradient methods fail. PLOS Comput.Biol. 5, e1000586. Williams, R. (1992). Simple statistical gradient-following methods for connectionist reinforcement learning. Ma- chine Learning 8, 229–256. Willshaw, D. J., O. P. Bunemann, and H. C. Longuet-Higgins (1969). Non-holographic associative memory. Na- ture 222, 960–962. Xie, X. and S. Seung (2004). Learning in neural networks by reinforcement of irregular spiking. Phys. Rev. E 69, 41909. Yagishita, S., A. Hayashi-Takagi, G. Ellis-Davies, H. Urakubo, S. Ishii, and H. Kasai (2014). A critical time window for dopamine actions on the structural plasticity of dendritic spines. Science 345, 1616–1620. Yu, A. and P. Dayan (2005). Uncertainty, neuromodulation, and attention. Neuron 46, 681–692. Zhang, J., P. Lau, and G. Bi (2009). Gain in sensitivity and loss in temporal contrast of stdp by dopaminergic modulation at hippocampal synapses. Proc. Natl. Aca. Sci. USA 106, 13028–13033. Zhang, L., H. Tao, C. Holt, W.A.Harris, and M.-M. Poo (1998). A critical window for cooperation and competition among developing retinotectal synapses. Nature 395, 37–44. Ziegler, L., F. Zenke, D. Kastner, and W. Gerstner (2015). Synaptic consolidation: from synapses to behavioral modeling. J. Neuroscience 35, 1319–1334. 26
1910.00003
1
1910
2019-09-28T09:02:30
A memory theoretic approach for investigating the roles of language and intuition in mathematical thinking activities
[ "q-bio.NC" ]
Questions concerning origin of mathematical knowledge and roles of language and intuition (imagery) in mathematical thoughts are long standing and widely debated. By introspection, mathematicians usually have some beliefs regarding these questions. But these beliefs are usually in a big contrast with the recent cognitive theoretic findings concerning mathematics. Contemporary cognitive science opens new approaches to reformulate the fundamental questions concerning mathematics and helps mathematicians break through the Platonic beliefs about the essence and sources of mathematical knowledge. In this article, we introduce and discuss mathematical thinking activities and fundamental processes such as symbolic/formal and visual/spatial ones. Two different aspects of mathematics should be separated in mathematical cognition. One aspect considers mathematics as an explicit crystallized knowledge. The other aspect considers mathematics as an ongoing and transient mental processing. The cognitive processes and corresponding tasks involved in these aspects are different. Ongoing mathematical activities both elementary and advanced, demand working memory resources. Using dual-task techniques, we design some pilot experiments to differentiate the symbolic/formal and visual/spatial processes. Using this memory theoretic approach, we explain the crucial roles of language-based processes such as verbal articulation and instructive speech and also visuo-spatial intuition such as spatial imagery and mental movement in various aspects of mathematics.
q-bio.NC
q-bio
A memory theoretic approach for investigating the roles of language and intuition in mathematical thinking activities Manouchehr Zaker∗ Department of Mathematics, Institute for Advanced Studies in Basic Sciences, Zanjan, Iran Abstract Questions concerning origin of mathematical knowledge and roles of language and intuition (imagery) in mathematical thoughts are long standing and widely debated. By introspection, mathematicians usually have some beliefs regarding these questions. But these beliefs are usually in a big contrast with the recent cognitive theoretic find- ings concerning mathematics. Contemporary cognitive science opens new approaches to reformulate the fundamental questions concerning mathematics and helps mathe- maticians break through the Platonic beliefs about the essence and sources of math- ematical knowledge. In this article, we introduce and discuss mathematical thinking activities and fundamental processes such as symbolic/formal and visual/spatial ones. Two different aspects of mathematics should be separated in mathematical cogni- tion. One aspect considers mathematics as an explicit crystallized knowledge. The other aspect considers mathematics as an ongoing and transient mental processing. The cognitive processes and corresponding tasks involved in these aspects are differ- ent. Ongoing mathematical activities both elementary and advanced, demand working memory resources. Using dual-task techniques, we design some pilot experiments to differentiate the symbolic/formal and visual/spatial processes. Using this memory theoretic approach, we explain the crucial roles of language-based processes such as verbal articulation and instructive speech and also visuo-spatial intuition such as spa- tial imagery and mental movement in various aspects of mathematics. AMS Classification: 00A30; 00A35; 97C30; 97C50 Keywords: Mathematical cognition; symbolic/formal mental processing; visual/spatial mental processing; working memory; visuo-spatial rehearsal; imagery; intuition 1 Introduction What are the roles of language and intuition for mathematical knowledge? Individual opinions and thought schools in philosophy of mathematics are often distinguished by their answer to this controversial question. Logicians such as Bertrand Russell believes (in 1903) that "The fact that all Mathematics is Symbolic Logic is one of the greatest ∗ [email protected] 1 discoveries of our age; and when this fact has been established, the remainder of the prin- ciples of mathematics consists in the analysis of Symbolic Logic itself." [28]. Russell gives no credit to intuition in mathematical thoughts. Formalists such as David Hilbert believe that mathematics is essentially a production of language and symbolic processing and manipulations of mathematical objects. In contrast to formalists, are intuitionists and propounders of mathematical intuition which emphasize on faculties such as imagery, vi- sualization and creativity. Philosophers such as Kant [19] and many mathematicians from Henri Poincar´e [27] to Solomon Feferman [16] insisted on the vital and irreplaceable role of intuition in mathematical knowledge. The dispute between language and intuition has been remained unresolved in philosophy and epistemology of mathematics. Contemporary cognitive science opens new approaches to reformulate the fundamental questions concern- ing mathematics and helps mathematicians break through the Platonic beliefs about the essence of mathematical thoughts. It is then illuminating and instructive for mathemati- cians to learn about the main findings of cognitive scientific methodologies addressing the fundamental questions of mathematics. The investigation of roles of language and imagery (intuition) is the research subject of a few disciplines in contemporary cognitive science and the results are striking. Developmental psychology such as Piagetian tradition, inves- tigates the roles of cognitive growth of children in the development of their mathematical abilities (e.g. [25]). The working memory approach to mathematics, studies the roles of memory components in mental mathematical activities such as mental arithmetic and rea- soning. Advanced brain scanning and imaging techniques provide available methodologies to explore associations between various mathematical activities and brain structural and functional areas. There is also some neural network modeling for mathematical cognition based on connectionist approach such as Parallel-Distributed Processing (e.g. McClelland et al. [23]). For more related works on mathematical cognition we refer the readers to the book [10]. If we plan to explore the relations between mathematical thoughts and abilities in one side, and cognitive faculties and processes on the other side, we should notice the follow- ing distinction between two different aspects of mathematics. The first aspect considers mathematics as a crystallized formulated knowledge, consisting of all mathematical pos- tulates, definitions and proved facts in the form of formalized topics or theories. This crystallized knowledge is clearly accumulated increasingly for each individual scholar of mathematics. But surprisingly it has not been initiated from vacuum. Humans have in- nate brain substrates by which are able to quantify the objects of environment, perceive differences in number, see numeric correspondences between different sets of objects; and perform elementary operations with numbers such as estimation of magnitude, counting and simple arithmetic. This competence known as Approximate Number System (ANS) (also known as Number Sense) is prior to any verbal education and exploits a visuo-spatial encoding for numbers [26, 12]. In addition to this competence, the brain is able to recog- nize and distinguish the fundamental patterns of Euclidean geometry such as congruency, connectedness, parallel lines and some other elementary metric and topology notions. This ability known as intuitive or core knowledge of geometry is innate and prior to education and culture [13, 20]. Hence, the mathematical knowledge is based on basic mathemati- cal substrates such as ANS and core knowledge of geometry. Although educated brains use purely symbolic concepts of numbers and operations as well as sophisticated theories 2 of geometry but the basic substrates of mathematics are still used while mathematical thinking activities. Figure 1: A diagram representing different aspects of mathematical thinking activities The second aspect, considers mathematics as an on-line mental processing. This second aspect focuses on all mental processes which are involved while doing ongoing mathemati- cal tasks such as calculation, proving a new result and mental logical inference. Sometimes further knowledge is obtained by reformulation or direct inference of previous knowledge. But basically, the formalized knowledge of mathematics is evolved through on-line math- ematical activities. By LTMK (long term mathematical knowledge) we mean all previous knowledge of mathematical objects, theorems, techniques and skills stored in the brain. On-line thinking activities initiates either with some input from external challenges (e.g. someone asks you to do a calculation) or by activating some data and queries from LTMK. Figure 1 depicts the relations between these components. It illustrates the framework of thought in the present paper. If necessary, the on-line component repeatedly retrieves information such as known facts and techniques from LTMK. In this case some informa- tion of LTMK is activated in the working level. In fact after processing in the on-line component, the obtained results transform into portions of LTMK in the form of final propositions. This stage is specified by the huge arrow in Figure 1. We make two comments concerning the above-mentioned distinctions between the two aspects of mathematics. The first is that when mathematicians evaluate their mathemat- ical achievements and the question how their mind are led to obtain these findings, they usually judge by the final formalized form of the obtained results and simply overlook the vital role of the on-line part of mental activity in obtaining these results. It is not easy to judge what exactly is going on in the brain circuits during on-line processes and usually the mechanism of this stage is covert for mathematicians. It is instructive to mention a main instance in this regard. After achieving his mathematical theory for mechanics, Joseph-Louis Lagrange in a famous quotation declares: "I have set myself the problem of reducing mechanics and the art of solving the problems appertaining to it, to general formulas, whose simple development gives all the equations necessary for the solution each problem ... No diagram will be found in this work. The methods which I expound in it, demand neither construction nor geometrical or mechanical reasoning, but solely algebraic operations subjected to a uniform and regular procedure." (quoted in [8]). Now we explain some differences between the two aspects of mathematics. Corresponding to each of these aspects we have different levels of mathematical processing. When we want to evaluate 3 mathematical knowledge of any individual we usually ask quick or yes/no questions. For example, what is the sum of angles in any convex quadrilateral? Is it true that any contin- uous function on a compact set is uniformly continuous? Assume that in a hypothetical task a mathematician is required to judge the validity of a mathematical statement P . In such a situation, it is not possible to perform the whole process of proving the statement P . Instead, he/she tries to recollect the required and related information from the cat- egorized knowledge which logically leads to the validity of P . In other words, the brain tries to obtain some partial propositions related to the target proposition P and relate them logically to P and finally infer the validity of P . The questions corresponding to the second aspect is the ones which require on-line manipulating and processing of mathemat- ical data. Moreover, the cognitive processes and functions involving in these aspects are different. For the first aspect, the normal tasks are "semantic judgment" tasks, in which participants judge the validity of a mathematical statement. But for the second aspect, the normal tasks are on-line tasks in which on-line and transient data are manipulated and processed in the brain. An important point is that although direct inferences from LTMK and corollaries of previous theorems have some contributions in advancement of mathematics but the process of research in mathematics is mainly advanced by on-line mathematical activities. The cognitive systems corresponding to the second aspect is dif- ferent from that of the first aspect in that the systems responsible for the on-line tasks (i.e. the second aspect of mathematics) is working memory components. Our approach in this paper for investigating the involvements of verbal faculties in various mathematical processes, is to analysis the mathematical activities corresponding to the second aspect and investigate some related tasks. Some neuro-psychological theories sug- gest that they are founded upon evolutionarily ancient brain circuits for number and space, and others that they are grounded in language competence. The authors of [1] evaluate which brain systems underlie higher mathematics. They scanned professional mathemati- cians and mathematically naive subjects of equal academic standing as they (quickly) evaluated the truth of advanced mathematical and non-mathematical statements. They conclude that high-level mathematical expertise and basic number sense share common roots in a nonlinguistic brain circuit. According to [1] their results suggest that high-level mathematical thinking makes minimal use of language areas and instead recruits circuits initially involved in space and number. According to the distinctions we explained con- cerning the first and second aspects of mathematics, the type of tasks which have been used in the study of [1] corresponds to the first aspect of mathematics. We comment once again that research in mathematics is advanced by on-line mathematical activities. Hence the results of [1] are limited to some parts of mathematical thinking and it should not be concluded from their experimental results that general high-level mathematical thinking has almost nothing to do with language, even if the subjects are expert mathematicians. The article is organized as follow. The next section is devoted to review some related results and works. In Section 3 we introduce some key terminology such as mathematical thinking activity, data configuration, transform function, symbolic/formal processing, vi- sual/spatial processing. We show how these mathematical processes can be distinguished using some mathematical tasks and dual-task methods. Section 4 discusses more top- ics about mathematical activities and visual/spatial processing. In Section 5 we itemize some spatial rehearsal mechanisms with related tasks and also explain the specific roles 4 of language-based processes such as verbal articulation and inner instructions in various processes of mathematics. 2 A memory theoretic introduction to mathematical cogni- tion Complete description of the cognitive mechanisms involving in various mathematical tasks is very difficult. But many successful methodologies and consequent experimental achieve- ments have been resulted. In this section we first briefly introduce the multi-component theory of working memory (M-WM). Initiated by the paper of Hitch in 1978 [18], many researchers have investigated involvement of the components of M-WM in various math- ematical tasks such as mental arithmetic. Also, some authors study the roles of work- ing memory capacities in academic achievement and math performance, for children and adults. In this section we review some related works in this direction and also some papers which have used brain scanning methodologies. The working memory is a cognitive brain system which is responsible for active and on-line retention and rehearsal of information for mental activities. In addition to maintaining and recollection of information, the working memory is capable of processing information in a highly accessible level. Information may be entered to working memory either from external data or by activating information from long term memory (see Figure 1), but while being processed in working memory, these information are highly accessible and naturally transient. When a chunk of information is processed by the working memory, we say that the information is processed at the working level. The working memory includes short term memory in the sense that any mental processing of information which has short term nature can be described in terms of working memory. A theory for investigating working memory was presented by Baddeley and Hitch in 1974 [6] and has been completed twice by Baddeley. This completed theory is referred as the multi-component theory of working memory (hereafter M-WM). We refer the readers to the books [4, 31] for comprehensive explanations of the theory. It should be mentioned that the multi-component theory of working memory is not the only theory for working memory. There are other theories of working memory which emphasis on attention and capacity. But we employ the multi- component theory because this theory emphasis on the role of sub-systems and investigates specific modalities separately. Now, we introduce briefly three components of M-WM. The phonological or articulatory loop component (PL) is thought to consist of two subcomponents. Phonological store is responsible for short term storage of verbal information in the form of a sound-based (phonological) code. Articulatory rehearsal is responsible for actively refreshing and elab- orating information in the phonological store. In fact the rehearsal process in PL consists of subvocal articulation and is identified with verbal rote rehearsal. The contents of the passive store are thought to be maintained by subvocal rehearsal. PL has other functions to be explained later. The visuo-spatial sketchpad (VSSP) is the second component and is responsible for maintaining visuo-spatial information. The VSSP consists of two subcom- ponents: visual cache for storing the visual features and complexities of the objects and 5 an active visuo-spatial rehearsal system, the so called inner scribe. It is generally believed that visuo-spatial working memory is also responsible for more active visuo-spatial tasks such as imagery. According to [32] visuo-spatial WM is responsible for tasks such as visual imagery, spatial rotation and spatial reasoning. The paper [3] explains more involvement of visuo-spatial rehearsal in working memory. The visuo-spatial rehearsal and imagery are very important cognitive processes and significantly involved in mathematical thoughts. In the present paper we provide some mathematical tasks which are behaviorally associated to motor imagery. M-WM also contains a regulating attentional system the so-called central executive (CE). The CE is an attentional control system which coordinates the processes of phonological and visuo-spatial information. The CE allocates attention to activities of PL and VSSP and is responsible for retrieval of information from long term memory. We explain in the next sections that task switching in mathematical activities and careful execution of procedures and algorithms demand executive control via inner speech in the component PL. It was also discovered that stress and anxiety use executive resources. For this reason anxiety reduces math performance of students [2]. To the best of the author's knowledge, the first study with the aim of investigation of the role of working memory in mental arithmetic was done by Hitch [18]. Hitch suggested that working memory made a crucial contribution to mental arithmetic. Logie and Baddeley in [21] investigated the effects of articulatory suppression and unattended speech on performance in simple counting tasks. Results were consistent in showing substantial disruption of counting performance by con- current articulatory suppression. Much stronger results were obtained in [22] concerning the significant role of phonological loop and subvocal articulation in counting and mental arithmetic. The experimental data of Logie et al. in 1994 [22] support the view that the subvocal rehearsal component of working memory provides a means of maintaining accuracy in mental arithmetic. The authors of [22] concluded that there was a minor involvement of the VSSP, which, furthermore, was restricted to the cases in which the cal- culations were presented visually (and could thus be restricted to a pre-calculation stage, such as in the encoding of the visual problem). The strong involvement of phonological loop in arithmetic was also verified in [24] in 2001, where the authors showed that the phonological loop plays a major role in mental addition, whereas the VSSP does not seem to be particularly involved. Thus, in mental computation, the storing space is verbal, and if calculations are presented visually in digits, a phonological recoding process first takes place. As we mentioned in the introduction, some basic approximate counting and simple arith- metic use visuo-spatial representations of numbers and operations with numbers. How- ever, when operations are more complex or required task switching and or precision, the visuo-spatial representation is not usable and hence the operation should be performed via symbolic/formal processing. In addition to the strong involvement of working memory in mathematical processing, critical dependence of mathematics performance on working memory has shown by many authors (e.g. [2]). Low mathematics performance is not only because of low speed or quality of mathematical processing of a person. Instead, some preoccupations such as fear and anxiety work like a resource-demanding secondary task during accomplishment of main mathematical tasks and decrease significantly the per- formance. The authors of [2] examined the cognitive consequences of math anxiety and 6 showed how performance on a standardized math achievement test varies as a function of math anxiety, and that math anxiety compromises the functioning of working memory. High performance of executive functions and lack of interruptive preoccupations are two effective factors in mathematical performance. 3 Fundamental Mathematical Processes The aim of this section is to introduce three mathematical processes. These processes are extensively used in on-line mathematical activities. Their relationships with the working memory components will be shown. For this reason they are interpreted as fundamental mathematical processes. To begin, we need to introduce the following regulating con- cepts: mathematical thinking activity, data configuration and transform function. These concepts are explained with more details in Section 4. By a mathematical activity we mean any on-line mathematical thinking processing in which the goal is to accomplish a mathematical task such as calculation, numeric or symbolic computation, proof of an assertion, checking the validity of a computation or a proof and or exploration to discover a new idea. A mathematical activity is either run purely mental and completely executed in the brain or worked out on paper and executed via the interaction of the brain and paper. In the second case tools such as paper or blackboard play a supplementary memory workplace for the brain. We mentioned earlier that the capacity to maintain information in working memory stores is very limited. Working out mathematical activities on paper overcomes noticeably the capacity limitations of brain memory components. But as will be explained, since the mathematical activities have by definition on-line and transient nature, manipulation, rehearsal and rumination of information during these tasks require working memory resources and hence are limited to the working memory characteristics of the brain. By a data configuration, we mean any concatenation of mathematical objects and their mathematical inter-relationships which are registered in memory stores during execution of a mathematical activity. A mathematical activity P begins from an initial data con- figuration Sinitial and during the execution of the activity, a present data configuration is transformed into a next one. Hence, to initiate a mathematical activity one is required to create and set up an appropriate data configuration in his/her mind. Even this initial step is set up by inner instructive speech such as "Suppose this ..." or "Assume that ...". By a transform function we mean any kind of mathematical process corresponding to the activity P which transforms one configuration of P into another one. Data configurations concern the representation of existing data and their relations but transform functions ma- nipulate the data and perform mathematical operations with these data as input. Hence, a mathematical activity is specified by its data configurations and intermediate transform functions. Note that at any step of execution of an activity, once a transform function F is specified, then the further step of the task is taken only by operating F on the underlying data configuration. But how these transform functions are determined is a completely different issue. We are now ready to introduce the promised processes. 7 3.1 Symbolic/Formal Processing In symbolic/formal processing, data configurations are represented as phonological and verbal format. The process of information in symbolic/formal processing is done by ver- bal elaboration, inner speech and symbolic manipulation. A symbolic/formal object is encoded by its phonological name. The meaning of a symbolic object is decoded by its phonological content. The meaning of a formal expression is determined by verbal elab- oration or articulation of the expression. This will be explained more latter. In terms of mathematics, meaning of any object is specified by its formal relationships with other objects. This meaning is specified by learning how it is used in formal relations. When we repeat - in the form of speech - the way of using a symbolic object in a formalism, the sense and meaning of the object becomes definite. The meaning of previously known symbols and formal relations as well as computing techniques are stored in LTMK. But in addition to these known notations and relations, some auxiliary notations are defined and used while proving mathematical assertions. For example consider the following generic phrases. Let X be the set consisting of all rational numbers satisfying a property R; or "let r be the supremum value of all real numbers satisfying a property R". In these examples, the phrases "the set consisting of all rational numbers satisfying a property R" and "the supremum value of all real numbers satisfying a property R" are the defining expressions of the set X and the real number r, respectively. In these definitions the notations are syncopated notations. The meaning of X or r is only specified by the vocal/subvocal articulation of their defining expressions. Any time a syncopated notation is required to be decoded and used in the proof, the subject has to decode its definition by full verbal articulation of its defining expression. The "symbolic syncopation" is used throughout mathematics. In particular, in symbolic/formal processing, data configurations use synco- pated symbolic notations. In symbolic/formal processing, the inter-relationships between the objects are formal mathematical relations. Here, the technique of "formal syncopa- tion" is used in order for abbreviation of formal relations. Formal syncopation is a very common methodology in whole mathematics. For example the sum of finitely many sum- (cid:88) (cid:90) mands and infinitely many summands are denoted by and , respectively. Hence, lengthy and complicated computations with infinite sums and series are replaced by formal computations with integrals. Method of generating functions is completely formal synco- pation technique used in combinatorial analysis. Functional transforms such as Laplace transform is a kind of formal syncopation. As we mentioned earlier, some mathematical activities involving numbers use visuo-spatial representation of numbers and operations. But in many other operations with numbers, visuo-spatial representations are not applicable and instead phonological encoding and verbal elaboration are required to perform such tasks. In fact, since in exact counting and precise arithmetic we use verbal encoding of numbers and operations, then they are included in symbolic/formal processing. The following task is a generic example of sym- bolic/formal process. It is presented as a pilot study. Task A. Define a sequence a1, a2, . . . of integer numbers as follows. The first two elements of the sequence are a1 = 1 and a2 = 2. For each n ≥ 3 define an = 2an−1−an−2. Determine in your mind the value of a6. 8 The author has performed Task A under different conditions. In ordinary version, par- ticipants perform Task A without doing any other action. In version A1, the participants perform Task A and simultaneously repeat an irrelevant word or speech. In version A1, the irrelevant inner or external speech is a secondary task. In version A2, the participant performs Task A and simultaneously moves his/her hand regularly left and right (with uniform speed). In version A3, the participant performs Task A on paper with an ir- relevant inner speech or external articulatory suppression as a secondary irrelevant task. Experimental results show that the performance in Tasks A1 and A3 is significantly worse than the performance in Task A2. The main point is that although the brain is able to use visual representation of numbers to help phonological perform of computations, but any external speech or irrelevant subvocal articulation disrupts the whole task because the exe- cution of task needs careful execution of the procedure and task switching, i.e. alternately multiply and subtract. Task A and similar examples suggests that task switching and accuracy in symbolic/formal activities demand intensely phonological resources such as verbal self-instruction. Task A has also an algorithmic aspect, because the subject should follow step by step an arithmetic procedure to obtain the final solution. The author has tested other algorithmic tasks. Even if you perform such tasks using pen and paper you use verbal encoding of numbers and operations and specially inner-speech instructions because you have to perform and control the steps of the task. The roles of phonological loop and inner-speech in general cognitive tasks requiring task switching and action con- trol have been studied in Baddeley et al. [5] in 2001 and in [15] in 2003. Interestingly, mathematical activities provide a big area of tasks which make strong relationships be- tween verbal self-instruction and task switching as well as procedure control. Execution of algorithms, following flowcharts and computational procedures are some instances of these mathematical activities. 3.2 Visual/Spatial Processing Data configurations in visual/spatial processing consist of pictures, shapes, volumes and other objects which are represented in visuo-spatial form such as knots and combinatorial graphs. This processing concerns the visual features and spatial relations of objects. The relations between objects are spatial such as geometric relations. Visual/spatial process in- cludes visuo-spatial manipulations such as spatial movements, rotations, navigation etc. In addition to the above-mentioned realm, the visual/spatial processing has a minimum but efficient use in symbolic/formal processes. Symbols, equations and expressions are firstly represented visually and then their phonological contents will be used in symbolic/formal processes. This means that in terms of memory there should be a link between visual encoding of symbols such as algebraic variables and Arabic numerals and their phono- logical format. Another important point is that correct syntax of formal relations needs appropriate spatial relations between the symbols. Visual/spatial processing has another application in symbolic/formal operations which will be explained in the next section. It should be mentioned that when symbols and all kinds of relations are learned for the first time, then switch of encoding from visuo-spatial appearance of notations toward seman- tic and verbal contents, is necessary to decode the phonological meaning of syncopated information. 9 For mathematical visual/spatial processing the situation is completely different. The main difference is that spatial rehearsal and visuo-spatial imagery involving mathematical items such as geometric ones are major processes in this type of process. In some branches of mathematics such as projective and synthetic geometries we observe the striking role of visuo-spatial rehearsal and imagery, because these fields study figures and geometric ob- jects as such, without recourse to coordinates and formulas. Importance of visuo-spatial imagery in branches of mathematics such as geometry has been insisted by many profes- sional mathematicians (e.g. Conway et al. [11]). Also the role of visuo-spatial sketchpad in academic achievement in geometry has been studied in some papers (e.g. [17]). Although the mathematical descriptions of symbolic/formal and visual/spatial processes clearly dif- ferentiate their distinct natures, but cognitive distinction of these processes is not easily understood. Because even in simple symbolic/formal tasks such as mental computation the brain may use visual format of information. Using the so called dual-task technique, we propose a cognitive methodology to clarify the distinction between symbolic/formal and visual/spatial processing. In the following we present our tasks and experimental results. We note that the mathematical content of these tasks are very common in mathematical visual/spatial processes. Task B. In this task, first the graph displayed in Figure 2 is exposed to each individual subject so that its memory is fixed completely in his/her mind. Then in absence of the figure the experimenter is required to explore all paths from A to B which goes through each node exactly once and say the number of such paths. Figure 2: A picture related to Task B The author has performed Task B under different conditions. In the ordinary perform of Task B the participant performs Task B without any extra action. In version B1, participants perform Task B and at the same time repeat an unrelated word or sentence. In version B2, the participant performs Task B and at the same time moves his/her hand regularly left and right (with uniform speed). The experimental results show that irrelevant inner speech or external speech does not reduce performance of the participants, but irrelevant hand movement do. In fact the performance in B2 is significantly reduced in contrast to the ordinary Task B. The results show that visual/spatial processes have close associations with visuo-spatial sketchpad and mental motor system but no dissociation with verbal component. Study of mental navigation as cognitive task goes back to 1968 by 10 Brooks [9]. The relation between mental animation and visuo-spatial sketchpad has been studied in [30]. Imagery and spatial representation have significant roles in learning [29]. The author has performed other "path construction mental tasks". In the following task which we call "cube navigation task", experimenters are required to construct mentally a closed path. Task C. Imagine a cube in the three dimensional space. Find a closed path along the edges of the cube which goes exactly once through each corner. Similar to Task B, the author has performed Task C under different conditions. In the ordinary version Task C is performed without any other action. In version C1, irrelevant speech is a secondary task and in C2, irrelevant hand movement is a secondary task. The results are all consistent with the results of Task B. The irrelevant inner speech or external speech does not reduce meaningfully performance of the participants, but irrelevant hand movement crucially decreases the performance. It should be mentioned that if we print 2-dimensional version of a cube on paper and perform Task C on this figure in paper, we obtain similar results. It suggests that even when mathematical visual/spatial processes are performed using pen and paper, there exists close association with visuo- spatial scratchpad and motor imagery. 3.3 Combinatory Processing Usually in mathematical tasks symbolic/formal and visual/spatial processes are combined in various ways. For example, assume that in a geometric picture including a number of lines and angles, the size of some angles are given. The task is to obtain the other angles. In this task we need to do some arithmetic but the corresponding quantities come from some visual information (i.e. the size of angles). The initial configuration consists of visuo-spatial and numerical information. In such tasks the processes are combined but the combination are separable. In general, the symbolic/formal and visual/spatial processes can be combined in separable ways in mathematical tasks. But as explained be- low, there exist many mathematical objects whose identity cannot be grasped by single or combination of the previous processes; they are neither symbolic/formal nor visual/spatial objects. Also, the relationships between the new objects cannot be reduced to single or combination of the two previous processes. In fact, the new objects are amalgamation of some visuo-spatial and symbolic information. We call such items "bound objects" because some symbolic and visuo-spatial information bind together to form the bound objects. The identification of mathematical bound objects is determined by decoding the symbolic parts and visual/spatial relations between them, simultaneously. We present some examples of bound objects. A permutation on 5 symbols such as 43521 is a bound object. The simple task that counts the number of permutations on 5 symbols is neither a symbolic/formal nor a visual/spatial processing. Permutations are composed of spatial arrangement of numbers from 1 to n. The other example is labelled graphs. A graph is composed of a number of vertices on the plane and some edges between them. A (vertex) labelled graph is a graph whose vertices are labelled by some symbols such as numbers or letters. In on-line working with labelled graphs, the brain has to process the visual information of the graph and symbolic information of labels, simultaneously. There are 11 also complex bound objects such as magic squares. Figure 3 displays an array of numbers which is called "magic square". In the array, the numeric sum of entries in each row, each column and principal diagonals of the array are a same number, i.e. 175 (in this example). These conditions define the concept of magic squares. To specify the identity of a magic square, one needs to determine the sum of some numbers which are arranged in certain spatial locations. The arrangements along a row, a column or a diagonal are these spatial arrangements. Note that the precise summation of a few integers is in general a verbal operation. Moreover, only the sum of numbers which are placed in some certain spatial arrangements, are required to make the identity of magic squares. Hence, magic squares are bound objects. Figure 3 illustrates a magic square. 22 47 16 41 10 35 4 5 23 48 17 42 11 29 30 6 24 49 18 36 12 13 31 7 25 43 19 37 38 14 32 1 26 44 20 21 39 8 33 2 27 45 46 15 40 9 34 3 28 Figure 3: A magic square The other complex bound objects are very well-known in mathematics. By an n× n Latin square L we mean any square array consisting of the entries 1, 2, . . . , n such that each number is appeared exactly once in each row and each column of L. In fact, in bound objects we find visuo-spatial patterns consisting of symbols in which the symbols form some spatial arrangements. Combinatorial analysis is a branch of mathematics which deals with objects such as permutations, Latin squares, labelled graphs and other similar objects which according to our definition are bound or amalgamated objects. For this reason we term the new process as combinatory processing. A combinatory processing is a mathematical on-line processing which works with bound objects. The mathematical relationships between bound objects are not reduced to any combination of visual/spatial and symbolic/formal processes. Of course it is possible to construct a formalism between appropriate bound objects. In this case, the objects are bound objects but the relations are formal, i.e. the combination of symbolic/formal and combinatory processing. To the best of author's knowledge, cognitive study of tasks concerning mathematical bound objects has not been done yet. This is an unexplored area. We know almost nothing about the ways the brain employs to process the combinatory objects. 12 4 Mathematical activities and processes Recall that by mathematical activity we mean any on-line mathematical thinking process- ing in which a mathematical task is accomplished. Mathematical activities consist of data configurations and corresponding transform functions. Data configurations are stimulated by external stimuli such as questions or through activation of some information from long term mathematical knowledge (LTMK). During mathematical activities, in addition to the initial phase or configuration, some new data and information - which form the fur- ther data configurations of the activity - are produced in mind which are maintained and represented by memory stores. In many mathematical activities which we related to the second aspect, the information under process should be highly active and accessible, and hence transient. Moreover, the exploration of ideas and partial solutions during solving mathematical problems, needs processing of transient but active mathematical informa- tion in mind. We conclude that the mathematical activities are mainly accomplished in working level. Storage of data configurations requires working memory stores and pro- cess of transform functions requires working memory mechanisms such as rehearsal and articulation. In fact, working memory serves like a scratchboard in the brain for on-line mathematical activities. Recall Figure 1 which depicts connections between these compo- nents. Of course there are mathematical tasks which are done by routine recollection of previous knowledge from LTMK. We related such tasks to the first aspect of mathematics and they happen when we are asked to judge or recollect quickly the validity of a math- ematical statement. There is no need to make an exact demarcation between the tasks related to the first and second aspects of mathematics. But there are salient differences between cognitive resources and processes regarding to the aspects which we explained. Note that at any step of execution of an activity, once a transform function say F is de- termined, then the further step of the task is taken only by operating F on the underlying configuration. But how these transform functions are determined is a completely different issue. Here we confront with a dichotomy. Checking the validity of a computation or a proof and in general, operating a definite function on an input data configuration, are in one side of the dichotomy. Production of an algorithm or a proof and in general, search and determination of a transform function are in the other side of the dichotomy. Therefore, corresponding to any mathematical activity there exits an accompanying mathematical processing which explores and determines all transform functions required for the accom- plishment of the activity. This accompanying process is in fact a logical reasoning which guides the whole thinking activity toward the final step of the task. This inner reasoning is done by inner speech which directs and controls the thought. The following helpful fig- ure illustrates the fact that the execution and discovery of transform functions belong to different categories. In Figure 4 the process begins by an initial data configuration Sinitial and is progressed via intermediate configurations S1, S2, . . . to attain the final situation. At each step Si, an appropriate transform function Fi+1 is determined and the next step Si+1 is obtained by operating Fi+1 on the configuration Si. Of course in general, evolution of configurations in mathematical activities is a continuum of mental mathematical configurations. We don't need to discretize it. In Figure 4 they are presented in sequential form just for simplicity. It is possible that in an activity, some 13 Figure 4: Configurations and transform functions transform functions are symbolic/formal and some visual/spatial. Visual/spatial process- ing has at least two essential roles for symbolic/formal thinking. We have mentioned before the first role. Correct logical and mathematical syntaxes of all formal relations re- quire appropriate spatial concatenation of the numeric and functional symbols and logical connectors. Hence, checking the validity of syntaxes is based on correct spatial placements of symbols with respect to each other. The second role is effective and much more non- trivial than the first role. Assume that in a mathematical activity we have a sequence of symbolic/formal configurations. The search for transform functions sometimes needs suitable rearrangements of symbols and expressions. New arrangements may help mind find new and effective formal expressions and relations between symbols. Consider a task in which the sum of numbers from 1 to 100 is required with the minimum possible num- ber of arithmetic operations. One representation is 1 + 2 + ··· + 100, for which 99 sum operations should be done. Now consider the following arrangement 1 2 100 99 3 ··· 4 4 ··· 3 99 100 2 1 By this configuration, the number of operations is only 3. Because all pairs in each column has a same sum 101. Hence the total sum is 100 × 101. But we need the half of it, i.e. 5050. An equivalent task is to determine the following sum 1 + 2 + ··· + 100 + 1 + 2 + ··· + 100. The point is that the suitable pairing of this list is achieved by visuo-spatial scanning of whole data. Some other examples of this sort, are putting parentheses in suitable places in long expressions involving arithmetic or algebraic operations in order for simplify the whole expression or cancel some terms. In general, spatial rearrangements and visual scanning of formal expressions are some key roles of visuo-spatial process in symbolic/formal activities. Sometimes mathematicians overlook these key roles because in some cases such visuo-spatial processes are done completely mentally and transient in the mind of the mathematician. 14 4.1 On language and intuition in mathematics What are the roles of language and intuition in mathematical cognition and which one has the most significant role? This question attracted many mathematicians and philosophers from long time ago. For example disputes between intuitionists and formalists on foun- dation of mathematics have roots in this question. This question in this form seems too simplified, because we have many aspects of language and intuition involved in mathemat- ical thoughts and abilities. In this section we describe the roles of language and intuition in a memory theoretic framework. By visuo-spatial intuition (or simply intuition) we mean the mental faculty which exploits the wide abilities of visual/spatial processing in mathematical activities including spatial representations and transform functions. Spa- tial rehearsal, in general speaking, is interpreted as active and dynamic processing of visuo-spatial information. Hence, visuo-spatial imagery and spatial rehearsal are the main cognitive characteristics of visuo-spatial intuition. We have already insisted the use of visual/spatial process even in symbolic/formal activities in terms of visual scanning and spatial rearrangement of formal expressions. Even in advanced levels of mathematics, on-line activities, in particular of visual/spatial type, recruit working memory resources such as visuo-spatial representations, spatial rehearsal, mental movements and imagery. In Section 3 we presented some tasks which are typically visuo-spatial. Two similar tasks of type "mental path construction" are presented in the following. Note that these tasks like the previous ones are not artificial tasks but they are extracted from real mathematical problems. Various spatial rehearsal and imagery techniques are required to accomplish these tasks. Task 1. Take a 3× 3× 3 array of dots in space, and connect them by edges up-and-down, left-and-right, and forward-and-back. The task is to try and answer does there exists a closed path which visits every dot exactly once? Task 2. Do Task 1 for a 4 × 4 × 4 array of dots, finding a closed path that visits every dot exactly once. The following tasks need different visuo-spatial rehearsal techniques. Task 3. A tetrahedron is a pyramid with a triangular base. Rest a tetrahedron so that it is balanced on one edge. What shape is the topmost part of the tetrahedron in this situation? Now, in your mind slice the tetrahedron horizontally halfway between its lowest edge and its topmost part. What shape is the slice? Task 4. An octahedron is the shape formed by gluing together equilateral triangles four to a vertex. Rest an octahedron on a face. What shape is the topmost part of it? Now, slice the tetrahedron halfway up. What shape is the slice? Task 5. How many colors are required to color the faces of an octahedron so that faces which share an edge have different colors? The following visuo-spatial task is an example of dissection problems in geometry. It is an elementary form of higher geometric operations on surfaces the so called geometric surgery. Dissection and surgery are mainly mental and imagery processes. The following has been chosen from plenty of problems. 15 Figure 5: The figure of Task 6 Task 6. Cut the figure displayed in Figure 5 into four pieces that will fit together and form a square. Finally, we itemize some of the known visuo-spatial rehearsal mechanisms in mathematical activities. • Focus and attention at some part or some features in a data configuration (e.g. Tasks 3, 4 and plenty of other examples) • Grasp of a panoramic or particular perspective (e.g. Task 5) • Navigation, routing, path construction and in general mental movements (e.g. Tasks B, C and Tasks 1, 2) • Cutting and pasting geometric pieces, dissection and surgery of geometric figures (e.g. Task 6) • Determination or estimation of distances and magnitudes In the rest of the article by summarizing the previous explains, we specify the key roles of various aspects of language for mathematics. In Dehaene et al. 1999 [14] using behavioral tasks and fMRI analysis, the roles of language and intuition has been investigated. It was shown that approximate arithmetic shows language independence and relies on visuo- spatial processing. In contrast exact arithmetic is acquired in a language-specific format and employs word-association processes. From the viewpoint of the present article, tasks consisting of many partial sub-tasks (e.g. Task A) which require task switching and control of action, recruit language-specific representation and operation of numbers as well as inner speech to direct the execution and control of the procedure. In fact, the result of [14] is generalized as follows. Not only exact arithmetic but also all algorithmic tasks including computational ones are acquired by language-specific processes. Overall, we summary that in on-line symbolic/formal mathematical activities even using pen and paper, and even 16 in advanced levels, phonological and verbal resources are intensively demanding. We list them as follow. • Verbal encoding and representation of numbers or abstract symbols • Decoding symbolic syncopations and formal syncopations using verbal articulation • Task switching and control of accuracy via inner speech • Correct execution and control of procedures according to flowcharts and algorithms • Mental logical reasoning and inference References [1] M. Amarlic, S. Dehaene, Origins of the brain networks for advanced mathematics in expert mathematicians, Proceedings of the National Academy of Sciences, 113 (2016) 4909 -- 4917. [2] M. H. Ashcraft, J. A. Krause, Working memory, math performance, and math anxiety, Psychonomic Bulletin & Review, 14 (2) (2007) 243 -- 248. [3] E. Awe, J. Jonides, P. A. Reuter-Lorentz, Rehearsal in spatial working memory, Journal of Experimental Psychology: Human Perception and Performance, 24 (1998) 780 -- 790. [4] A. D. Baddeley, Working Memory, Thought and Action, Oxford University Press (2007). [5] A. D. Baddeley, D. Chincotta, A. Adlam, Working memory and the control of action: evidence from task switching, Journal of Experimental Psychology: General, 130 (2001) 641 -- 657. [6] A.D. Baddeley, G.J. Hitch, Working memory, In: The Psychology of Learning and Motivation: Advances in Research and Theory, ed. G. A. Bower, pp 47 -- 89, New York: Academic. [7] A. D. Baddeley, V. Lewis, G. Vallar, Exploring the articulatory loop, Q. J. Exp. Psychol. A 36 (1984) 233 -- 252. [8] E. T. Bell, The development of mathematics, Dover (1992). [9] L. R. Brooks, Spatial and verbal components of the act of recall, Canadian Journal of Psychology, 22 (1968) 349 -- 368. [10] J. I. Campbell (Ed.), Handbook of mathematical cognition, Psychology Press (2005). [11] J. Conway, P. Doyle, J. Gilman, B. Thurston, Geometry and the imagination, Based on materials from the course taught at the University of Minnesota Geometry Center in June 1991, Version 0.941 (2010). 17 [12] S. Dehaene, The number sense: How the mind creates mathematics, Oxford Univer- sity Press (2011). [13] S. Dehaene, V. Izard, P. Pica, E. Spelke, Core knowledge of geometry in an Amazonian indigene group, Science 311 (2006) 381 -- 384. [14] S. Dehaene, E. Spelke, P. Pinel, R. Stanescu, S. Tsivkin, Sources of Mathematical Thinking: Behavioral and Brain-Imaging Evidence, Science 284 (1999) 970 -- 974. [15] M. J. Emerson, A. Miyake, The role of inner speech in task switching: a dual-task investigation, J. Mem. Lang. 48 (2003) 148 -- 168. [16] S. Feferman, Mathematical intuition vs. mathematical monsters, Synthese 125 (2000) 317 -- 332. [17] D. Giofre, I. C. Mammarella, L. Ronconi, C. Cornoldi, Visuospatial working mem- ory in intuitive geometry, and in academic achievement in geometry, Learning and Individual Differences, 23 (2013) 114 -- 122. [18] G. J. Hitch, The role of short-term working memory in mental arithmetic, Cognitive Psychology 10 (1978) 302 -- 323. [19] I. Kant, Critique of Pure Reason, Cambridge University Press (1998) (translated and edited by Paul Guyer and Allen W. Wood). [20] V. Izard, P. Pica, S. Dehaene, D. Hinchey, E. Spelke, Geometry as a universal mental construction, In: Space, Time and Number in the Brain, ed. S. Dehaene and E. Brannon, pp. 319 -- 332, Academic Press (2011). [21] R. H. Logie, A. D. Baddeley, Cognitive Processes in Counting, Journal of Experimen- tal Psychology: Learning, Memory, and Cognition, 1987, Vol. 13, No. 2, 310 -- 326. [22] R. H. Logie, K. J. Gilhooly, V. Wynn, Counting on Working Memory in Arithmetic Problem Solving, Memory and Cognition, 22 (1994) 395 -- 410. [23] J. L. McClelland, K. Mickey, S. Hansen, A. Yuan, Q. Lu, A Parallel-Distributed Processing Approach to Mathematical Cognition (2016). [24] M-P. Noel, M. D´esert, A. Auburn, X. Seron, Involvement of short-term memory in complex mental calculation, Memory & Cognition 29 (2001) 34 -- 42. [25] B. Ojose, Applying Piaget's theory of cognitive development to mathematics instruc- tion, The Mathematics Educator, 18 (1) (2008). [26] M. Piazza, V. Izard, P. Pinel, D. Le Bihan, S. Dehaene, Tuning curves for approximate numerosity in the human intraparietal sulcus, Neuron, 44 (3) (2004) 547 -- 555. [27] H. Poincar´e, The value of science, New York, Dover (1958) (translated by George Bruce Halsted). [28] B. Russell, The principles of mathematics, Cambridge Univ. Press (1903). 18 [29] D. L. Schwartz, J. Heiser, Spatial representations and imagery in learning (pp. 283- 298), Cambridge Univ. Press (2006). [30] V. K. Sims, M. Hegarty, Mental animation in the visuospatial sketchpad: Evidence from dual-task studies. Memory & Cognition, 25 (1997) 321 -- 333. [31] E. E. Smith, S.M. Kosslyn, Cognitive Psychology, Prentice-Hall, Inc. New Jersey, USA (2007). [32] H. D. Zimmer, Visual and spatial working memory: From boxes to networks, Nuero- sciences and Biobehavioral Reviews, 32 (2008) 1373 -- 1395. 19
1608.03461
2
1608
2019-05-14T02:05:45
Black-boxing and cause-effect power
[ "q-bio.NC" ]
Reductionism assumes that causation in the physical world occurs at the micro level, excluding the emergence of macro-level causation. We challenge this reductionist assumption by employing a principled, well-defined measure of intrinsic cause-effect power - integrated information ({\Phi}), and showing that, according to this measure, it is possible for a macro level to "beat" the micro level. Simple systems were evaluated for {\Phi} across different spatial and temporal scales by systematically considering all possible black boxes. These are macro elements that consist of one or more micro elements over one or more micro updates. Cause-effect power was evaluated based on the inputs and outputs of the black boxes, ignoring the internal micro elements that support their input-output function. We show how black-box elements can have more common inputs and outputs than the corresponding micro elements, revealing the emergence of high-order mechanisms and joint constraints that are not apparent at the micro level. As a consequence, a macro, black-box system can have higher {\Phi} than its micro constituents by having more mechanisms (higher composition) that are more interconnected (higher integration). We also show that, for a given micro system, one can identify local maxima of {\Phi} across several spatiotemporal scales. The framework is demonstrated on a simple biological system, the Boolean network model of the fission-yeast cell-cycle, for which we identify stable local maxima during the course of its simulated biological function. These local maxima correspond to macro levels of organization at which emergent cause-effect properties of physical systems come into focus, and provide a natural vantage point for scientific inquiries.
q-bio.NC
q-bio
Black-boxing and cause-effect power William Marshall1, Larissa Albantakis1, Giulio Tononi1, * 1Department of Psychiatry, Center for Sleep and Consciousness, University of Wisconsin, Madison, WI, USA *Corresponding author: [email protected] Abstract Reductionism assumes that causation in the physical world occurs at the micro level, excluding the emergence of macro-level causation. We challenge this reductionist assumption by employing a principled, well-defined measure of intrinsic cause-effect power -- integrated information (Φ), and showing that, according to this measure, it is possible for a macro level to "beat" the micro level. Simple systems were evaluated for Φ across different spatial and temporal scales by systematically considering all possible black boxes. These are macro elements that consist of one or more micro elements over one or more micro updates. Cause-effect power was evaluated based on the inputs and outputs of the black boxes, ignoring the internal micro elements that support their input-output function. We show how black-box elements can have more common inputs and outputs than the corresponding micro elements, revealing the emergence of high-order mechanisms and joint constraints that are not apparent at the micro level. As a consequence, a macro, black-box system can have higher Φ than its micro constituents by having more mechanisms (higher composition) that are more interconnected (higher integration). We also show that, for a given micro system, one can identify local maxima of Φ across several spatiotemporal scales. The framework is demonstrated on a simple biological system, the Boolean network model of the fission-yeast cell-cycle, for which we identify stable local maxima during the course of its simulated biological function. These local maxima correspond to macro levels of organization at which emergent cause-effect properties of physical systems come into focus, and provide a natural vantage point for scientific inquiries. Author Summary 1 2 We challenge the reductionist assumption by studying causal properties of physical systems across different spatiotemporal scales. The result is that -- contrary to reductionist views -- causal power can emerge at macro scales. Rather than relying on the traditional notion of coarse-grains (averages), we introduce the notion of functional black boxes that are defined based on their input-output relationship. Using a sequence of examples, our work demonstrates that black boxes are particularly well suited to capture the heterogeneous and specialized nature of components in biological systems. While the emergence of coarse- grained systems relies on increased specificity, black-boxing reveals the importance of structure and integration. Our framework is mathematically rigorous and fully general, hence applicable across many disciplines; it is particularly useful in objectively identifying informative perspectives on complex systems in the physical sciences. Introduction Reductionist approaches in science usually assume that the optimal causal model of a physical system is at the finest possible scale. Coarser causal models are seen as convenient approximations due to limitations in measurement accuracy or computational power (Kim, 2000; Nagel and Hawkins, 1961). The reductionist view is based on the conjecture that the micro level of causal interaction is causally complete, leaving no room for additional causation at a macro level (Kim, 1993). The reductionist assumption is most obvious in fields such as particle physics (Nakamura et al., 2010), neuroscience (Markram, 2012), and nanotechnology (Bhushan and Marti, 2010), but it can also be found in the social sciences (Imai et al., 2011), where researchers endeavor to 'look inside the black box'. A case has been made for the occurrence of genuine emergence at various macro levels (Ellis, 2011; Fodor, 1974), such as the emergence of mind above and beyond the individual neurons (or atoms) that constitute the brain (Tononi, 2008), and for the autonomy of the special sciences such as chemistry (Scerri and McIntyre, 1997), and biology (Dupré, 2009; Walker and Davies, 2013), above and beyond the underlying physics. However, arguments 3 in favor of emergence have often been vague, or they have focused on the possibility that macro variables may have greater descriptive power than micro variables, rather than greater causal power (List and Menzies, 2009; Pfante et al., 2014; Wolpert et al., 2014). Inspired by statistical physics, macro-level descriptions of a system are typically taken to be coarse-grainings, i.e. averages over micro elements and micro time steps. The reductionist assumption has been challenged by the introduction of explicit measures of cause-effect power, which were used to show that such coarse-grainings can indeed have greater cause-effect power at the macro level (Hoel et al., 2013, 2016). In simulated examples of simple logic gate systems, we coarse-grained (nearly) identical elements ('neurons') into groups ('neuronal groups') and averaged over their states. We demonstrated that, under certain conditions involving degeneracy and/or indeterminism, a macro-level system of coarse-grained elements can "beat" the micro-level system in terms of cause-effect power (Hoel et al., 2013, 2016). However, moving beyond statistical physics to biology, the macro elements of interest cannot be obtained by coarse-graining, because they are constituted of heterogeneous micro elements that are often compartmentalized and have highly specific functions, which would be muddled by averaging (see Box 1). For example, take the neuron, considered as the fundamental unit in much of neuroscience. Clearly, a neuron cannot be represented by a coarse-grained macro element, because it is constituted of a great diversity of specific molecules, organized in highly specific and hierarchical ways, performing highly specific functions. Indeed, it is the very specificity of the internal micro elements that makes the reductionist assumption seem inevitable in these cases: while we can treat a neuron as a black box for ease of understanding and for convenience, it would seem that its full causal power can only be captured by considering all the molecules that constitute the black box, in exquisite and specific detail (Markram, 2006). Here we further challenge the reductionist assumption by generalizing the causal analysis employed for coarse- graining to black-boxing (Tononi, 2010): we first analyze a system of heterogeneous, specific micro elements at the micro level; then we repeat the analysis at the macro level by grouping subsets of those micro elements inside black boxes (macro elements). Black boxes are characterized exclusively by their overall input-output function (Ashby and others, 1956; Bunge, 1963). The heterogeneous micro elements inside the black box are hidden inside a macro 4 element, rather than averaged as with coarse-graining. As an example of a black box, Fig. 1 shows, on the left, a simple, schematic neuron constituted of a number of specific micro elements (synapses S, cell body C, and axon hillock A) that interact internally in specific ways. On the right, the neuron is treated as a single macro element, a black box, that receives inputs (spike or no spike for each input), produces a single output (spike or no spike), and conceals its micro elements inside. In what follows, we assume that the causal power of a system is quantified by its intrinsic cause-effect power as previously defined (Oizumi et al., 2014; Tononi, 2015). While reductionism assumes implicitly that causal power resides exclusively with micro elements, we assess causal power explicitly -- as intrinsic cause-effect power -- and determine the spatiotemporal levels at which new cause-effect properties emerge. Such emergent cause-effect properties may include an increase in the overall intrinsic cause-effect power of the system, but also specific relationships between elements within the system ("mechanisms") that only become apparent at the macro level. To quantify intrinsic cause-effect power and system mechanisms at the micro level and all possible black-boxed macro levels, we use the interventional and counterfactual causal framework of integrated information theory (Oizumi et al., 2014; Tononi, 2015). As a measure of intrinsic cause-effect power, integrated information (Φ) captures several aspects that are often overlooked in causal accounts (Oizumi et al., 2014): the dependence of cause- effect power on the specific state the system is in (state-dependency); how cause-effect power of the system is structured (composition); whether the whole system is causally irreducible to its parts (integration); and what defines the system's borders and grain (exclusion). These features make Φ particularly suited for assessing the cause-effect power intrinsic to a system, independent of external observers. As demonstrated through several examples, including the Boolean network model of the fission yeast cell cycle, the Φ value of systems of black-box macro elements can increase when going from finer to coarser spatiotemporal grains and lead to emergent cause- effect properties at macro scales. 5 Figure 1: A schematic neuron considered as a number of `micro' elements (left), or as a black box (right). At the micro scale, the neuron receives inputs at its synapses (S), which are passed on to the cell body (C) and then to the axon hillock (A), which outputs to other neurons. Cause-effect power is assessed by perturbing each element (small hands) and observing the effects, while irreducibility is assessed by partitioning the elements (dashed red line). At the macro scale, there is only the black-box element (neuron) which receives three inputs and generates an output. Cause-effect power is assessed by perturbing the output of the black box (big hand) and observing its effects without constraining the constituent micro elements, however its irreducibility is still assessed by partitioning between micro elements (dashed red line). Box 1 -- Black-boxing and coarse-graining A discrete, finite physical system can be considered at various spatiotemporal levels. At the most fine- grained scale, it is constituted of a set 𝑆𝑚 of micro elements, each having at least two states. Supervening, physical macro-level systems 𝑆𝑀 can be obtained by a mapping Μ: 𝑆𝑚 → 𝑆𝑀 that groups disjoint subsets of 𝑆𝑚 into non- overlapping macro elements. A physical macro element is thus constituted of one or more micro elements, operating over one or more micro time steps and can be manipulated, observed, and partitioned. For each macro element, Μ defines how the states of its constituting micro elements are mapped onto the possible states of the macro element. In previous work (Hoel et al., 2013, 2016), we demonstrated the emergence of cause-effect power in 'coarse-grained' macro-level systems with average-based state mappings. Here, we extend these results to 'black-box' macro elements with an output-based state mapping (Box Figure). Coarse-graining: Coarse-graining corresponds to the notion of a macro state in statistical physics. In coarse-graining, the state mapping is a function that depends only on the average of the micro states of the micro 6 elements constituting the macro element, without reference to the identity of individual micro elements (Hoel et al., 2013, 2016). This means that all micro states with the same average have to be mapped onto the same macro state, e.g., 𝑠𝑚 = {00, 10,01} → 𝑠𝑀 = 'OFF' and 𝑠𝑚 = {11} → 𝑠𝑀 = 'ON', while 𝑠𝑚 = {00, 10} → 𝑠𝑀 = 'OFF' and 𝑠𝑚 = {01, 11} → 𝑠𝑀 = 'ON' would not be a proper coarse-grain mapping. Black-boxing: Black boxes correspond to the typical notion of macro elements in the special sciences, such as cells or organisms in biology. In black-boxing, the state of a macro element is determined by the state of its output (micro) elements at a specific (micro) time step, without reference to the states of its internal micro elements. A possible mapping for the schematic system shown in Box Figure (left) in which 5 micro elements form a black box is, e.g., 𝑠𝑚(𝑡3) = {𝑋𝑋𝑋𝑋0} → 𝑠𝑀 = 'OFF' and 𝑠𝑚(𝑡3) = {𝑋𝑋𝑋𝑋1} → 𝑠𝑀 = 'ON'. This means that, given an input at time 𝑡0, the macro state of the black box corresponds to the micro state of the output element at time 𝑡3, while the states of the hidden elements are ignored. Box Figure: Prototypical examples of black-box and coarse-grain elements that can lead to the emergence of macro-level cause-effect power. Left: A black-box element conceals many micro elements with specific functions. Right: A coarse grain macro element averages together many homogenous micro elements that share a global function. Increasing intrinsic cause-effect power: In recent work, we showed that coarse-grained physical systems can, under certain conditions, 'beat' the corresponding micro-level system in terms of measures of effectiveness (Hoel et al., 2013) and intrinsic cause-effect power (Φ) (Hoel et al., 2016). As done in this study, we simulated simple physical systems constituted, at the micro level, of collections of logic gates. The main factor enabling higher intrinsic cause-effect power through coarse-graining is a reduction in indeterminism and degeneracy at the macro level (Hoel et al., 2013, 2016). Determinism and degeneracy affect the selectivity of a system in its current state. In a non-degenerate and deterministic system, the current system state constrains with maximum selectivity both the cause repertoire (only one past state is possible - no degeneracy) and the effect repertoire 7 (only one future state is possible -- no indeterminism). In a degenerate system, multiple past states could lead to the current state of the system. In a non-deterministic system, multiple future states could follow the current state. Grouping noisy or degenerate micro elements into less degenerate and more deterministic macro elements may lead to a gain in the selectivity of the system's mechanisms. Everything else being equal, more selective mechanisms have higher intrinsic cause-effect power  (see Theory), which translates to higher Φ at the system level and thus may lead to emergence of macro-level cause-effect power in coarse-grained systems (Albantakis and Tononi, 2015; Hoel et al., 2016). In general, coarse-graining micro systems, in the sense of averaging over subsets of them, may increase intrinsic cause-effect power when the constituting micro elements are all roughly of the same kind and all their inputs and outputs can be treated as equivalent. However, in system architectures constituted of heterogeneous micro elements, with highly specific functions, which are typical for biological and electronic systems, averaging across micro states may blur rather than enhance cause-effect power. It is these types of modular system architectures for which black-boxing is particularly suited to bring about emergent cause-effect properties at the macro level: in the results section, we demonstrate that black-boxing may reveal high-order macro mechanisms that are not present at a micro scale. In turn, these support a more integrated cause-effect structure and higher Φ values at the macro level. Taken together, mapping a finer-grained system into a coarser, macro-level system may increase intrinsic cause-effect power both through coarse-graining (possible increase in selectivity) or black-boxing (possible increase in integration). Which mapping is more suited to bring about emergent cause-effect properties depends on the type of system architecture. Ultimately, we can consider a continuum of possible macro elements combining the two complementary approaches as the general case, where black boxes with one output for all micro elements of a box at a particular micro time-step and coarse grains with an output for each micro element are the extremes. 8 Theory Integrated information (Φ) measures the intrinsic cause-effect power of a physical system (Albantakis and Tononi, 2015; Oizumi et al., 2014) by evaluating five requirements: the system's capacity to make a difference to itself (intrinsicality), composition, information, integration, and exclusion. Loosely defined, Φ quantifies to what extent a system's cause-effect structure, which specifies how all the system's parts constrain each other's past and future states, is integrated, that is, irreducible to subsystems (more below). The measure Φ, which was developed as part of integrated information theory (IIT) (Oizumi et al., 2014), builds on interventionist, counterfactual accounts of causality (Lewis, 1974; Pearl, 2009) and can also be employed as a general measure of complexity that captures to what extent a system is both integrated and differentiated (Albantakis and Tononi, 2015; Albantakis et al., 2014; Hoel et al., 2016). We formally define a physical system as a set of elements, for example neurons in the brain or logic gates in a computer, such that each element has at least two states, inputs that can influence these states, and outputs that in turn are influenced by these states. Furthermore, it must be possible to manipulate, observe, and partition among elements, in order to evaluate their cause-effect power. To fully characterize the cause-effect properties of a physical system, we first randomly perturb its elements into all possible states according to a maximum entropy distribution and observe their subsequent state transitions. Through this process, one obtains the transition probability matrix (TPM) for the physical system. During the perturbations, elements outside the physical system under consideration are held fixed; the states of these elements are considered "background conditions" (Oizumi et al., 2014). By fixing the background conditions we control external influences and use the system's TPM to calculate its intrinsic cause- effect properties, including Φ (see Supp. S1 Text). Given the TPM of a system, the next step is to identify all its mechanisms -- the subsets of the system which, in their current state, have irreducible cause-effect power within the system itself (intrinsicality). To this end, we test the entire power-set of system elements as candidate mechanisms (composition). To have irreducible cause- effect power, a set of elements in its current state must selectively constrain the potential past and future states of the system (information). This is evaluated using the conditional probability distribution of past or future states 9 given the current state of the set of elements. A mechanism can be composed of one or more elements, as long as it constrains the past and future states of the system above and beyond its parts (integration). The degree to which a mechanism in its current state is irreducible is measured by , which quantifies the irreducible cause-effect power of the mechanism within the system (Albantakis and Tononi, 2015; Albantakis et al., 2014; Marshall et al., 2016; Oizumi et al., 2014; Tononi, 2015). In the following, we distinguish between mechanisms consisting of a single element (first-order mechanisms) and those composed of multiple elements (high-order mechanisms), which play an essential role in integrating the whole system. Note that a set of elements that fails to irreducibly constrain the system's past state does not have any potential causes within the system, and a set of elements that fails to constrain the system's future state irreducibly does not have any potential effects within the system; in both cases  = 0 and neither is an intrinsic mechanism of the system. The set of all mechanisms within a system defines its cause-effect structure. If a candidate mechanism in its current state has a value of  = 0, then it is reducible, and does not contribute to the cause-effect structure of the system (see Fig. 4). The intrinsic cause-effect power of the system is quantified by its integrated information Φ (Albantakis and Tononi, 2015; Albantakis et al., 2014; Marshall et al., 2016; Oizumi et al., 2014; Tononi, 2015), which captures the irreducibility of the cause-effect structure: the degree to which the system's cause-effect structure is changed by partitioning the system (eliminating constraints among parts). For Φ to be high, every possible partition must affect many mechanisms that constrain the system in a highly selective, irreducible manner (having high ). If Φ = 0, then there is at least one part of the system that remains unconstrained by the mechanisms of the rest: from the intrinsic perspective, there is no unified system, even though an external observer can treat it as one. Finally, from the intrinsic perspective, the set of elements that form a system must be definite. In other words, it must have a self-defined causal border with its environment that identifies the elements within the border as part of the system, while elements outside the border belong to the system's environment. Even though many subsets and supersets of elements may have Φ > 0, only sets of elements that specify a local maximum of Φ have well defined borders from the intrinsic perspective (exclusion). A system's border is thus defined by the intrinsic cause- 10 effect structure of its elements, such that adding or removing a single element will result in a decrease of cause- effect power. This exclusion principle also applies across spatiotemporal scales: from the intrinsic perspective, the set of elements that form a system must have a definite spatiotemporal grain. As with the system's borders, it is the intrinsic cause-effect structure that self-defines its spatiotemporal scale, which is one that is a local maximum of Φ. Local maxima of Φ identify those scales at which cause-effect properties emerge -- any finer or coarser grains necessarily result in a reduction of cause-effect power and a blurring of intrinsic cause-effect properties. To evaluate intrinsic cause-effect power at macro scales and identify the definite scales at which new cause-effect properties emerge, micro elements can be grouped either by coarse-graining as in Hoel et al., (2013, 2016) or, more generally, by black-boxing, as will be demonstrated here. Black-boxing In typical usage, a black box is an object into which inputs impinge and from which outputs emerge, but its internal workings are not available for inspection (Ashby, 1956; Bunge, 1963). For our purposes, a 'black-box element' is a physical macro element that can be manipulated, observed, and partitioned, which is constituted of several micro elements (spatial), operating over several micro time steps (temporal). To qualify as a black box, it must satisfy the following conditions: (i) It must have at least one input, one output, and two or more (macro) states that can be read from its output (element) (ii) The micro elements and micro updates within the black box are hidden (black box condition) (iii) The micro elements contribute causally to the black box's output (integration) (iv) There cannot be any overlap between the micro elements of multiple black boxes (exclusion) Specifically: 11 (i) The inputs and outputs of a black box are defined in terms of the internal micro elements that receive direct input from other elements/black boxes (e.g., synapses S in Fig. 1) and directly output to other elements/black boxes (e.g., the axon hillock A in Fig. 1). For this work we allow for inputs to arrive at multiple micro elements, but restrict outputs to leave from only a single micro element within the black box. Furthermore, the inputs are taken to arrive at the beginning of the macro time step, while the outputs are taken to depart at the end of the macro time step. In principle, this framework could be extended to multiple output elements and to a more general treatment of time steps by allowing macro elements with different temporal grains. (ii) The state of a black-box element is taken to be the state of its (micro) output element at its (micro) output time step. The transition probabilities associated with a black-box element are determined as usual by causal analysis, perturbing the inputs of the black box into all possible states according to a maximum entropy distribution. At the end of the macro update, the state of the black box is observed from its output element (see Fig. 2). In this way, one can determine the cause-effect power that the inputs (i.e., outputs from other black- box elements) have on the state of the black-box element over the respective macro update. In line with the notion of black boxes, the micro elements within the black box are "hidden" from other black boxes within the system, meaning they do not directly contribute to the intrinsic cause-effect power of the system, but only indirectly through their black box's output. Any other direct micro interactions are not considered intrinsic to the macro-level system and therefore do not contribute to its cause-effect power at all (see Supp. S3 Text). Crucially, for the duration of the macro update, the internal elements are allowed to evolve unperturbed; however, to discount the cause-effect power of micro elements when evaluating Φ, the initial states of micro elements and any micro connections leaving the black box, other than its designated output element at the designated output time step, are noised during the perturbation analysis. A consequence of this perturbation procedure is that potential causes and effects must be direct (i.e. between two black boxes), and that potential causes and effects that are mediated by a third black box are 'screened off' and do not contribute to cause- effect power (Supp. S3 Text, Fig. S4). 12 (iii) The requirement that every constituent micro element must causally contribute to the output of its black box is mandated by the integration principle that cause-effect power must be irreducible. Even at the macro level, a system can only be integrated if its micro level is integrated. Moreover, it is not meaningful to consider a black-box element as a single physical element if it is reducible to two or more unrelated elements. The requirement of micro integration is satisfied implicitly when assessing models using integrated information; any physical system that violates it will be found to be reducible and thus have Φ = 0, as even for macro systems, Φ is evaluated by partitioning between micro elements. This implies that it is not possible to take a non-integrated system of micro elements and to black-box it in such a way as to create an integrated system of macro elements (Supp. S3 Text, Fig S5). (iv) The requirement for no overlap among the constituents of different black boxes (or equivalently that a micro element cannot be a constituent of more than one black-box element) is a consequence of causal exclusion. A physical (macro) element must be definite, meaning that it has a well-defined border which separates it from other macro elements. The importance of the exclusion condition has been independently recognized in the theory of computation: it is only meaningful to say that a physical system implements a computation if the system is constituted of distinct, non-overlapping elements (Chalmers, 1996). If black-box elements were permitted to overlap, then every open physical system could be said to implement any computation (Chalmers, 1996; Putnam and Putnam, 1988). Together, the above requirements allow to specify inputs and outputs of each black-box element, to define its macro state, to include within each black box only micro elements that are integrated and contribute to its input- output function, and to draw 'borders' around each black-box element that exclude any overlap with other black Local Maxima of cause-effect power boxes (Fig. S6). 13 Only systems that support local maxima of Φ, both in terms of constitution and spatiotemporal grain, are definite and have intrinsic cause-effect power. A system of elements is a local maximum if there are no 'neighboring' systems with a higher value of Φ. When only micro elements are considered, such as in (Marshall et al., 2017), it is natural to define a neighbor as any system that differs in constitution by only a single micro element, that is, any system that can be made by either adding or removing a single element. However, to determine whether two systems at different spatiotemporal grains are neighbors, several distance measures have to be taken into account. For the present purposes, we consider three different distances between systems to establish whether two systems are neighbors in this general context. The first is the constitutional distance between two systems, which is the number of micro elements that must be added / removed from one system to transform it into the other. Next is the temporal distance between two systems, which is the difference in the number of micro updates that make up the corresponding macro updates. Finally, the spatial distance between two systems is the distance between the partitions that group micro elements into macro elements. In the current work we use the maximum matching distance between partitions (Almudevar and Field, 1999), which is essentially the number of micro elements that must be moved from one grouping to another. If the sum of the constitutional, temporal and spatial distances between two systems is equal to 1 then those systems are neighbors, i.e., two systems are neighbors if they differ by a single step in exactly one of the three distances. Given a set of micro elements, we evaluate all possible systems (sets of black-box elements) to determine which systems have intrinsic cause-effect power, at which spatiotemporal grain (the set of black-box elements that define the system), and what their borders are (the set of micro elements that constitute the system). Evaluating all possible sets of black-box elements includes all possible groupings of micro elements into macro elements. Then, for every grouping all possible elements of each black box are considered as its output element. Finally, cause- effect power is evaluated over all possible macro time steps of each black-box system. Note that not all micro elements must be grouped into black boxes when searching for maxima of intrinsic cause-effect power. It may be that adding a specific micro element to any black-box element within the system would in fact reduce cause-effect power. In this case, such micro elements are held fixed as background conditions of the macro system (see Supp. S3 Text). 14 Results In the following, we demonstrate black-boxing and its importance for revealing macro-level cause-effect properties based on a set of simple proof-of-principle examples before we apply the framework to a biological model of the fission-yeast cell-cycle. Crucially, we demonstrate that systems of black-box macro elements can have higher intrinsic cause-effect power than their corresponding micro systems, and support local maxima of Φ that reveal emergent functional properties. For the purposes of this work, we shall consider collections of elements that are binary micro elements which cannot be further reduced or split, and the time scale of state transitions to be a micro time step. Time is implicit in the TPM, as micro elements are synchronously updated at discrete micro time steps. In principle, integrated information is defined for any discrete system of elements. The full mathematical details of the Φ calculation are described elsewhere, we recommend (Oizumi et al., 2014) but details are also available in (Hoel et al., 2016; Marshall et al., 2016; Tononi, 2015); full example analyses are presented in Supp. S1. All calculations in this work were performed using the PyPhi software package in Python (Mayner et al., 2016), which includes a documented example for a black-box analysis. How macro beats micro: composition and integration An intuitive example in which black-boxing may be appropriate is propagation delay -- the amount of time between the output of one element and its effect on another element. Such delays are largely ignored in functional analyses and are taken to be an implicit aspect of the element of interest, i.e., they are black-boxed. In the context of logic gates, for example, NOR logic is commonly described as a "universal" in the sense that any other logic can be built strictly from NOR gates. However, building, say, an XOR gate from NOR gates requires in fact a propagation delay as an implicit part of the circuit. 15 In the following example, we explicitly model such propagation delays as (one or more) COPY elements that take a single input and then output the same value. Fig. 2 shows the micro structure of an XOR element with a one- step propagation delay, along with the corresponding macro element, a black box with XOR logic. Figure 2: An XOR logic gate with one time-step propagation delay. Left: Two COPY elements which each take a single input and relay the values to an XOR element. Middle: A black-box element with two inputs and a single output element (dashed red outline). Right: By perturbing the inputs to the black-box element in all possible ways, it is determined that it implements XOR logic over two time steps. Consider a system of three interconnected XOR elements with a one-step propagation delay. At the micro level, this system is constituted of nine micro elements -- six COPY and three XOR, which can be black-boxed over two time steps into a macro system of three interconnected XOR elements (see Fig. 3). The current state of all elements is OFF. 16 Figure 3: A system of three interconnected XOR elements with one-step propagation delay and all elements in the OFF state. Top: At the micro level, the propagation delay is modeled by COPY elements between the XOR elements (left). The micro level has only first-order mechanisms specified by individual XOR elements (right), and Φ = 0.25. Bottom: When the elements are black-boxed over two time steps, the system is comprised of three interconnected macro elements implementing XOR logic. The macro-level system has second-order mechanisms specified by pairs of XOR elements (right) and Φ = 1.875. Assessing the cause-effect structure of the micro system, we find that there are only three first-order mechanisms and no high-order mechanisms. The three XOR elements each specify a mechanism with  = 0.5: by being in the OFF state, each XOR specifies that its two inputs must have been either (OFF, OFF) or (ON, ON) and that its outputs, the COPY elements, must be OFF in the future (Fig. 3, top-right). All other sets of elements do not have cause-effect power, or are reducible, so  = 0 (see Fig. 4). Recall that from the intrinsic perspective, a set of elements must constrain both the system's past and future irreducibly to be a mechanism for the system (see Theory). The six COPY elements, taken individually, lack any potential effect within the system: by being in the OFF state, a COPY by itself does not constrain the future state of its XOR output, which is still equally likely to be ON or OFF depending on the state of its other input (Fig. 4, top). On the other hand, two COPY elements in the 17 state (OFF, OFF) that input to the same XOR element do irreducibly constrain the system's future states, since together they specify that the XOR element they output to will be OFF. Nonetheless, these pairs of COPY elements do not form a second-order mechanism in the system since their constraint on the system's past state is reducible: in the OFF state, the two COPY elements taken individually already specify that their inputs must have been OFF, leaving no room for additional second-order constraints (Fig. 4, bottom). The lack of either irreducible past or future constraints thus prevents the COPY elements from specifying first- or high-order mechanisms in the system. The integrated information of the micro physical system is Φ = 0.25 (see Supp. S1 Text). The macro-level physical system with black-box elements also has three mechanisms with  = 0.5, but they are second-order mechanisms specified by pairs of XOR elements. By being in the state (OFF, OFF), each pair of XOR elements specifies that the past state of the entire model must have been either (OFF, OFF, OFF) or (ON, ON, ON), and that the future state of their common output must be OFF (Fig. 3, bottom-right). Neither of the XOR elements in this high-order mechanism can specify these constraints on its own. Individual XOR elements lack potential effects in the system for the same reason as the individual micro COPY gates above. At the macro level, the collection of mechanisms (cause-effect structure) is more integrated than that of the micro level, with a value of Φ = 1.875. Although the system has the same number of mechanisms and the same  values at both the micro and the macro level, the black-boxed system has higher Φ because a system partition impacts the macro level cause- effect structure more than the micro level cause-effect structure. The black-box system "wins" by having more overlap in its mechanisms, both in terms of the elements they are composed of and the constraints they impose. The high-order mechanisms of the black-box system have overlapping constraints, with each mechanism constraining all elements within the system, whereas the first-order mechanisms of the micro system only constrain their respective COPY inputs and outputs, without overlap. A system partition at the micro level thus only affects a single micro mechanism, whereas a system partition at the black-box level affects all of the mechanisms in the system, resulting in higher integration (see Supp. S1 Text). Consequently, there is irreducible cause-effect power that emerges at this macro level of the physical system. Concealing the COPY elements inside the black boxes reveals the high-order interactions between the XOR gates over two time steps. Note also, that, while the causal analysis is 18 state-dependent, in this example the irreducibility of micro and black-box cause-effect structures (their Φ values) and thus the relationship between levels, is equivalent for all possible system states. Figure 4: Two potential mechanisms from the propagation delay network in Fig. 3 which end up being reducible. On top is a COPY element that does not specify a mechanism. By being OFF in the current state, the COPY element constrains its input to be OFF in the previous state, but it does not constrain the future state of its output element, because the state of the XOR element still completely depends on the unknown state of its other input (shown here in grey). The bottom panel is a set of COPY elements which do not specify a high-order mechanism because they do not have an irreducible cause (the red line partitions the cause in two with no loss of information). Taking each COPY element independently fully constrains the past state of its input to be OFF. Finding local maxima of intrinsic cause-effect power In a second example, we consider a larger micro system constituted of 55 elements that all implement NOR logic. By testing all possible black-boxings, we establish three local maxima of cause-effect power which reveal the organizational hierarchy of the system. Fig. 5, demonstrates how a group of 11 elements implementing NOR 19 logic can be connected in such a way to produce AND/OR logic, or MAJORITY logic at coarser spatiotemporal scales. The 55-element system is arranged into five interconnected groups of 11 elements, with each group organized according to Fig. 5 so that the system exhibits different functions at different spatiotemporal scales. Each group of 11 elements receives inputs from three other groups and has a single element that outputs to three other groups (Fig. 6, top left). We consider the system state in which each of the 55 NOR micro elements is ON. In the following, we focus on the cause-effect structures of the system levels shown in Fig. 5: the micro physical system of NOR elements, a black-boxed system of AND/OR elements, and a black-boxed system of MAJORITY elements. These systems are shown in Fig. 6 (top row) ordered according to the average spatial grain of their elements. Many other possible black-boxing schemes were also evaluated; however none of these other black-box macro systems had Φ > 0. Two representative examples of macro systems with Φ = 0 are included in Fig. 6 (bottom). Figure 5: Left: A collection of 11 NOR elements at the micro scale. Middle: A macro scale black-boxing of these elements into four black boxes. Perturbing the inputs of these black boxes reveals that three implement AND logic and the final one OR logic, each at a time scale of two time steps. Right: A macro scale black-boxing with one element, implementing MAJORITY logic over its three inputs at a time scale of four time steps. Note that there are only three inputs, but each input arrives at two different micro elements. 20 At the micro level, the system's cause-effect structure consists of 55 first order mechanisms, one for each micro element with  = 0.239 on average, and no high-order mechanisms. The integrated information of this micro physical system is Φ = 0.453 (see Supp. S1 Text). Figure 6: A system of 55 interconnected NOR micro elements viewed at several different grain sizes, with all elements in the ON state. The micro elements form 5 interconnected groups, with each group arranged according to Fig. 5. The systems are arranged with micro elements on the far left and black-box elements of increasing spatial grain to the right. The legend on the right specifies the input-output function of each element. Each of the three systems on the top row is a local maximum of cause- effect power, corresponding to the three spatiotemporal grains shown in Fig. 5. Shown in the bottom row are two representative examples of the many systems with Φ = 0 at spatial grains between the local maxima. The macro-level AND/OR black-boxed system with an average spatial grain of 2.75 (Fig. 6, top, middle) has 20 macro elements, 15 implementing AND logic and 5 implementing OR logic, operating over two time steps. Similar to the micro level, its cause-effect structure is composed of 20 first order mechanisms (one for each black- 21 box element) but no high-order mechanisms, with  = 0.112 on average. This black-boxing reduces the number of first-order mechanisms, but does not reveal high-order mechanisms or overlapping constraints, thus the macro system is no more integrated than the micro system. Moreover, this black-boxing in fact reduces the integrated information of the first-order mechanisms in the system compared to the micro level ( values are 0.127 lower on average), leading to lower integrated information for the system (Φ = 0.080). The macro-level black-boxed system with an average spatial grain of 11 (Fig. 6, top right) is defined by considering black-box elements implementing MAJORITY logic over four time steps. Compared to the macro level with an average spatial grain of 2.75, this additional black-boxing step further reduces the number of elements, but increases the average  to 0.216 ( values are still 0.023 lower than the micro level on average). However, this macro system is endowed not only with first-order mechanisms, but with all possible second, third and fourth-order mechanisms. In total, its cause-effect structure includes 30 of 31 possible mechanisms from the power set of black-box elements, resulting in high integration, with Φ = 2.333, more than the micro level. Fig. 6 also shows additional black-box systems with Φ = 0. One of these black-box systems with an average spatial grain of 1.57 has 20 black-box OR elements over two time steps and 15 micro NOR elements. A second black-box system with average spatial grain of 3.66 has 10 black-box AND elements over two time steps and 5 black-box AND elements over four time steps. For both of these systems (and many others not shown), the integrated information is Φ = 0, because there is no common temporal scale over which all the elements in the system have effects on other elements within the system. For any specific temporal scale, there will be elements that do not causally contribute, thus the system is not integrated. In summary, this example demonstrates how evaluating cause-effect power over many different spatial and temporal scales of black boxes identifies local maxima of cause-effect power and reveals emergent cause-effect properties. For this example, the analysis reveals functional relationships between elements; local maxima of cause-effect power occur specifically at the micro level of NOR elements (average spatial grain size of 1, Φ = 0.453), at an intermediate macro level of AND/OR elements (average spatial grain size 22 of 2.75, Φ = 0.080) and at a coarser macro level of MAJORITY elements (average spatial grain of 11, Φ = 2.333). While these spatial grains reveal emergent levels of organization at which the system exhibits intrinsic cause-effect power, which shed light on its cause-effect properties, the vast majority of systems of black-box elements, on the other hand, yield Φ = 0. Boolean network model of the fission yeast cell cycle As a demonstration of black-boxing in biological systems, we apply the framework to the Boolean network model of the fission-yeast cell-cycle (Davidich and Bornholdt, 2008). The model consists of nine Boolean ("micro") elements representing the state of crucial proteins expressed during cell division. Each element implements linear threshold logic, and the connections between elements are weighted, with each connection being either excitatory (+1) or inhibitory (-1) in nature (see Fig 7A). One element, "SK" only inputs to the system, receiving no feedback. This element acts as a catalyst for cell division: when it is activated while the network is in its biological attractor state, the remaining eight elements cycle through a sequence of 9 states, eventually returning to the initial attractor state (see Fig 7B). This cycle of states is called the 'biological sequence' of the model, and captures the specific sequence of protein expressions that occur during the cell-division cycle. Since the element SK receives no feedback from the rest of the cell-cycle network, any system that includes SK will necessarily be reducible (Φ = 0). Only when SK is fixed as a background condition can we potentially identify systems with Φ > 0. Furthermore, if we consider the remaining eight elements (excluding SK) as a system, one of the states of the biological sequence (t2, see Fig 7B) has no cause (potential past state) within the system (it is caused by the catalyst element SK which initializes cell division from outside the system). For this reason, the cause-effect structure of this system is undefined in state t2. In what follows, we refer to the cell-cycle network as the eight strongly connected elements that contain both inputs and outputs (not including SK) and its biological sequence as the eight states (t1, t3-t9) with well- defined cause-effect structures. 23 Previous work analyzing the cause-effect structure of the cell-cycle model demonstrated that the cell- cycle network constitutes a stable local maximum of integrated information across all states of the biological sequence (Marshall et al., 2017). However, this previous work only analyzed the cell-cycle model at the micro level, considering all possible subsets of micro elements. In the current work, we extend this analysis by considering the cell-cycle network at macro spatiotemporal scales. Specifically, we consider all possible groupings of the cell-cycle network into black-box macro elements, at time scales of 2, 3 and 4 micro updates (greater time scales may reveal additional local maxima and emergent cause-effect properties). Figure 7: Boolean network model of the fission-yeast cell-cycle. (A) A network of 9 linear threshold elements connected by excitatory (black) and inhibitory (red) connections. The element, SK receives no input from the other elements. For this reason, we do not consider it as a part of the cell-cycle network, but rather as an external input that serves as a catalyst to initiate cell division. At the micro level, the system comprised of eight elements (in blue) is a stable local maximum of Φ for the duration of the biological sequence. (B) A sequence of 9 states that represent the cell division cycle, called the networks biological sequence. There are 4140 ways to group the eight micro elements in the cell-cycle network into any number of black-box elements, and for each grouping there are on average 10 different ways to define the output 24 elements of the black boxes. Considering three different time scales for each set of black-box elements, results in a total of 124,176 macro systems to analyze. Across all states of the biological sequence, there are 2224 macro systems with Φ > 0, an average of 278 per state, or roughly 0.22% of all possible systems. Among the 2224 macro system with Φ > 0, we identify 33 unique local maxima (some others are duplicates due to symmetries in the network). The majority of these local maxima are transient, occurring in an average of 2.5 out of 8 states in the biological sequence. However, 5 of the local maxima are stable over all states of the biological sequence. The micro system is one example of a stable local maximum, confirming that the results of (Marshall et al., 2017) hold even when considering macro systems. The remaining four local maxima occur at macro spatiotemporal scales, one at a time scale of 3 micro updates, and the others at a time scale of four micro updates (see Fig 8). Note that the intrinsic cause-effect power of a system is state dependent, and stability across subsequent time steps is not assumed at any point in the analysis. That the cell-cycle supports stable local maxima of macro cause-effect power is a feature of this biological system that is revealed by the causal analysis, rather than a requirement imposed by the framework. Our analysis moreover reveals that one element in particular (Slp1) serves as a black box's output in every stable local maximum. This indicates that Slp1 may play a crucial role in stabilizing and integrating the network over longer time scales during the process of cell division -- a property that could not be identified from its micro level interactions (Kim et al., 2013; Marshall et al., 2017). 25 Figure 8: All stable local maxima of macro cause-effect power for the cell-cycle network over the course of its biological sequence. Stable local maxima are identified at two different time scales (over 3 or 4 micro updates) and with groupings of the eight micro elements into either two or three macro elements. The output element for each black box is marked by a green outline; one common feature among all of the stable maxima is that element Slp1 acts as an output element of one black box. Note that connections between black boxes that do not originate from output elements are not shown in the figure because they do not contribute to the cause-effect structure (see Supp. S3 Text). Discussion In this work we expand the framework for evaluating the cause-effect power of physical systems at multiple spatiotemporal scales, to include biologically motivated black-box macro elements defined by their input-output function. We then use this framework to explore the cause-effect power of simple systems 26 of elements considered both at the micro level and after black-boxing, at a macro level. The cause-effect power of these systems was assessed using integrated information (Φ), a measure of the cause-effect power that is intrinsic to a physical system. To properly capture cause-effect power from the intrinsic perspective of the system itself, Φ considers composition, specificity, irreducibility, and exclusion (Albantakis and Tononi, 2015; Oizumi et al., 2014). We show how macro systems based on black boxes can have higher intrinsic cause-effect power than any neighboring systems (including in some cases their micro element counterparts). This result complements and extends previous work that showed how intrinsic cause-effect power can increase when macro elements are defined by coarse-graining micro elements (Hoel et al., 2016). While coarse-graining may reduce degeneracy and/or indeterminism in a system, black-boxing may increase a system's intrinsic cause-effect power by increasing its integration. Reductionist accounts of causation assume that all causal power resides with micro elements and time steps, excluding all macro levels (Kim, 2000). We argue that reductionist accounts of causation conflate the necessity of micro elements as constituents with their cause-effect power within the system. As shown in Fig. 4, a single micro element within a system may completely lack the power to constrain the system's future states -- taken individually, it does not make any difference to the system. Yet, the high-order mechanism with irreducible cause-effect power shown in Fig. 3 would not exist without the individual micro elements to support it. Thus micro elements may play a role as a constituent of a high-order mechanism or a macro element with cause-effect power. The current work reveals the possibility that causal power may emerge at macro spatiotemporal scales, requiring only that a system is definite, with self-defined borders and spatiotemporal grain (by being a local maximum of Φ). In such a case, the micro elements support the macro level as constituents, the macro level still supervenes upon the micro level, yet there are cause-effect properties that are only revealed at this particular macro level. Limitations and Future Work In the current work, we use intrinsic cause-effect power as a quantification of causal power, and demonstrate several examples of systems of black-box macro elements with higher intrinsic cause-effect power than the corresponding micro systems. To the extent that the notion of causal power is appropriately 27 captured and quantified by intrinsic cause-effect power, our results refute the reductionist assumption that causal power resides exclusively at the micro level. The value of our characterization of cause-effect power had been previously demonstrated in a number of contexts (Albantakis and Tononi, 2015; Albantakis et al., 2014; Marshall et al., 2017), and will continue to be evaluated in the future. A limitation on the practical application of this framework is the computational demands for exhaustively evaluating intrinsic cause-effect power. Currently, cause-effect properties can only be fully explored for very small systems (< 10 micro elements; propagation delay example, cell-cycle example) or by exploiting symmetries in the system (local maxima example). Future work will extend the PyPhi software for evaluating intrinsic cause-effect power (Mayner et al., 2016) by including, for example, approximations based on the connectivity matrix. However, practical applications inevitably will have to use a targeted approach and only assess the intrinsic cause-effect power of a predetermined set of macro- level systems instead of evaluating all possible black-box systems. Theoretical investigations like the current work (see below) as well as previous exploration of coarse-grained macro elements (Balduzzi, 2011; Hoel et al., 2016) will be crucial to define the criteria that will guide such a targeted approach. Black-boxing reveals high-order mechanisms and joint constraints The two main requirements for high Φ are that a physical system is differentiated (many specific mechanisms) and integrated (mechanisms with overlapping constraints). Typically, whenever a lower level system is mapped into a higher macro system, there is reduced state differentiation, i.e., the macro system has fewer elements and a smaller state space. This decrease in differentiation means fewer potential mechanisms and thus less potential integrated information (Marshall et al., 2016). In order for a macro level system to have higher cause-effect power (Φ) than a finer grained system over the same elements, the macro system must increase cause-effect power either by having more specific mechanisms, or a more integrated set of mechanisms. Degeneracy and indeterminism are two factors that influence the specificity of a mechanism. Everything else being equal, decreasing degeneracy and indeterminism leads to an increase in the cause- 28 effect power of mechanisms within the system. In (Hoel et al., 2013, 2016) we demonstrated that coarse- graining (averaging) micro elements into macro elements can lead to an increase in intrinsic cause-effect power that can overcome the inherent loss of differentiation in macro systems. An increase in intrinsic cause-effect power through reduction of degeneracy is also possible through black-boxing, as shown in Supp. S2 Text. The particular asset of black-boxing is that it may reveal high-order mechanisms and joint constraints between mechanisms at macro spatiotemporal scales. As demonstrated by the propagation delay example, the macro can even beat the micro level through increased integration. This may occur when elements with few potential effects are concealed within black-box elements, and micro elements with many potential effects serve as the outputs of black-box elements, resulting in a more densely interconnected set of macro elements, where groups of macro elements share common inputs and common outputs. If creating common inputs and common outputs among elements leads to additional, joint constraints on the possible past and future system states, elements may form high-order mechanisms, resulting in a more integrated cause-effect structure and higher Φ. Being a part of high-order mechanisms, or being constrained by multiple mechanisms, gives an element additional ways to contribute to the cause-effect structure; when an element contributes in multiple ways, cutting that element has a greater effect on the cause-effect structure, making the system more irreducible. Being more irreducible means having higher intrinsic cause-effect power (Φ) and may thus lead to a causally emerging macro level. This suggests that black-boxing is most beneficial when there are "causal bottlenecks" in the micro system, that is, when a micro element with a single or few outputs connects to a micro element with a single or few inputs. In such cases, it is impossible for these micro elements to contribute to high-order mechanisms, and such elements represent a "weak link" in the integration of the system. More generally, black-boxing should be particularly appropriate in systems with local modular interactions whose results are distributed across the system, such as molecular interactions within neurons in the brain, or electrical interactions within computer networks. 29 Local maxima of intrinsic cause-effect power Evaluating cause-effect power of black-box systems across many spatiotemporal scales shows that, in general, there can be several local maxima of macro cause-effect power, between which integrated information decreases or falls to zero. In Fig. 6, the local maxima capture emergent functional roles of black-box macro elements, corresponding to the different descriptions of the system as sets of NOR, OR/AND, or MAJORITY elements. Importantly, even within a given spatiotemporal grain, there will generally be several local maxima corresponding to overlapping subsets of elements, such that adding or subtracting an element reduces integrated information (Hoel et al., 2016; Oizumi et al., 2014). These local maxima of intrinsic cause-effect power across and within levels correspond to organizational macro levels and systems having emergent cause-effect properties. These are natural levels and systems for the special sciences to investigate. A prime example is biological systems, since they contain many highly specialized components which are required to perform their function. In biology we can study the molecules within an individual cell, the interactions between networks of cells (nervous system), individual organs (liver, kidneys), whole organism (animals, humans), and communities of organisms (swarms, societies). The Boolean network model of the fission yeast cell cycle is one example of a simulated biological system which contains many heterogeneous micro elements that perform specific functions in order to accomplish cell division. Applying the black- boxing framework reveals several macro local maxima that are stable throughout the biological sequence of the network model, and highlights the role of element Slp1 in stabilizing the cycle. Note that the typical approach of studying biological systems at a particular (macro) spatiotemporal scale is precisely to treat its next-lower level components as black boxes. Here we have proposed a theoretical framework to evaluate cause-effect power and the cause-effect properties of such a black-box system. If an organizational level corresponds to a local maximum of integrated information, then there will be cause-effect properties that emerge at that level, and there is knowledge to be gained by studying the system accordingly. 30 Finally, while local maxima reveal cause-effect properties to an investigator studying the system, the global maximum specifies the set of elements and spatiotemporal grain at which the system has most cause- effect power upon itself -- from its own intrinsic perspective. According to integrated information theory, a set of elements at the spatial-temporal grain that defines the global maximum of intrinsic cause-effect power corresponds to a physical substrate of consciousness (Oizumi et al., 2014; Tononi, 2015). ______________________________ References Albantakis, L., and Tononi, G. (2015). The Intrinsic Cause-Effect Power of Discrete Dynamical Systems -- From Elementary Cellular Automata to Adapting Animats. Entropy 17, 5472 -- 5502. Albantakis, L., Hintze, A., Koch, C., Adami, C., and Tononi, G. (2014). Evolution of integrated causal structures in animats exposed to environments of increasing complexity. PLoS Comput Biol 10, e1003966. Almudevar, A., and Field, C. (1999). Estimation of Single-Generation Sibling Relationships Based on DNA Markers. J. Agric. Biol. Environ. Stat. 4, 136. Ashby, W.R., and others (1956). An introduction to cybernetics. Introd. Cybern. Balduzzi, D. (2011). Detecting emergent processes in cellular automata with excess information. ArXiv11050158 Cs Math Nlin Q-Bio. Bhushan, B., and Marti, O. (2010). Scanning probe microscopy -- principle of operation, instrumentation, and probes. In Springer Handbook of Nanotechnology, (Springer), pp. 573 -- 617. Bunge, M. (1963). A general black box theory. Philos. Sci. 346 -- 358. Chalmers, D.J. (1996). Does a rock implement every finite-state automaton? Synthese 108, 309 -- 333. Davidich, M.I., and Bornholdt, S. (2008). Boolean Network Model Predicts Cell Cycle Sequence of Fission Yeast. PLoS ONE 3, e1672. Dupré, J. (2009). It is not possible to reduce biological explanations to explanations in chemistry and/or physics. Contemp. Debates Philos. Biol. 32 -- 47. 31 Edlund, J.A., Chaumont, N., Hintze, A., Koch, C., Tononi, G., and Adami, C. (2011). Integrated information increases with fitness in the evolution of animats. PLoS Comput Biol 7, e1002236. Ellis, G.F.R. (2011). Top-down causation and emergence: some comments on mechanisms. Interface Focus rsfs20110062. Fodor, J.A. (1974). Special sciences (or: the disunity of science as a working hypothesis). Synthese 28, 97 -- 115. Hoel, E., Albantakis, L., Marshall, W., and Tononi, G. (2016). Can the macro beat the micro? Integrated information across spatiotemporal scales. J. Conscious. Sci. Hoel, E.P., Albantakis, L., and Tononi, G. (2013). Quantifying causal emergence shows that macro can beat micro. Proc. Natl. Acad. Sci. 110, 19790 -- 19795. Imai, K., Keele, L., Tingley, D., and Yamamoto, T. (2011). Unpacking the black box of causality: Learning about causal mechanisms from experimental and observational studies. Am. Polit. Sci. Rev. 105, 765 -- 789. Kim, J. (1993). Supervenience and mind: Selected philosophical essays (Cambridge University Press). Kim, J. (2000). Mind in a physical world: An essay on the mind-body problem and mental causation (MIT press). Kim, J., Park, S.-M., and Cho, K.-H. (2013). Discovery of a kernel for controlling biomolecular regulatory networks. Sci. Rep. 3. Lewis, D. (1974). Causation. J. Philos. 70, 556 -- 567. List, C., and Menzies, P. (2009). Nonreductive physicalism and the limits of the exclusion principle. J. Philos. 106, 475 -- 502. Markram, H. (2006). The blue brain project. Nat. Rev. Neurosci. 7, 153 -- 160. Markram, H. (2012). The human brain project. Sci. Am. 306, 50 -- 55. Marshall, W., Gomez-Ramirez, J., and Tononi, G. (2016). Integrated Information and State Differentiation. Front. Psychol. 7. Marshall, W., Kim, H., Walker, S.I., Tononi, G., and Albantakis, L. (2017). How causal analysis can reveal autonomy in models of biological systems. Phil Trans R Soc A 375, 20160358. Mayner, W., Marshall, W., and Marchman, B. (2016). pyphi: 0.8.1. Nagel, E., and Hawkins, D. (1961). The structure of science. Am. J. Phys. 29, 716 -- 716. 32 Nakamura, K., Group, P.D., and others (2010). Review of particle physics. J. Phys. G Nucl. Part. Phys. 37, 075021. Oizumi, M., Albantakis, L., and Tononi, G. (2014). From the phenomenology to the mechanisms of consciousness: integrated information theory 3.0. PLoS Comput Biol 10, e1003588. Pearl, J. (2009). Causality (Cambridge university press). Pfante, O., Bertschinger, N., Olbrich, E., Ay, N., and Jost, J. (2014). Comparison between different methods of level identification. Adv. Complex Syst. 17, 1450007. Putnam, H., and Putnam, H. (1988). Representation and reality (Cambridge Univ Press). Scerri, E.R., and McIntyre, L. (1997). The case for the philosophy of chemistry. Synthese 111, 213 -- 232. Tononi, G. (2008). Consciousness as integrated information: a provisional manifesto. Biol. Bull. 215, 216 -- 242. Tononi, G. (2010). Information integration: its relevance to brain function and consciousness. Arch. Ital. Biol. 148, 299 -- 322. Tononi, G. (2015). Integrated information theory. Scholarpedia 10, 4164. Walker, S.I., and Davies, P.C.W. (2013). The algorithmic origins of life. J. R. Soc. Interface 10, 20120869. Wolpert, D.H., Grochow, J.A., Libby, E., and DeDeo, S. (2014). Optimal high-level descriptions of dynamical systems. ArXiv14097403 Cond-Mat Q-Bio. Supporting information Legends S1 Text. Full analysis of cause-effect power. Detailed cause-effect structures for the examples presented in the main text. S2 Text. Black-boxing of degenerate or indeterministic systems. Additional examples exploring the effects of degeneracy and indeterminism. S3 Text. The intrinsic perspective. Additional discussion on analyzing cause-effect power from the intrinsic perspective. Black-boxing and cause-effect power: Supplementary Information William Marshall1, Larissa Albantakis1, Giulio Tononi1, * 33 1Department of Psychiatry, Center for Sleep and Consciousness, University of Wisconsin, Madison, WI, USA *Corresponding author: [email protected] S1 Text -- Full analysis of cause-effect power Here we present a more detailed account of the integrated information analysis of the example systems discussed in the main text. All calculations were performed using the PyPhi software package in Python (Mayner et al., 2016). For a given system in a specific state, the first step involves identifying its cause-effect structure, the set of mechanisms in the system. A mechanism is a set of elements that irreducibly constrains the past and future states of the system. Each member of the power set of system elements is tested as a potential mechanism. The set of system elements whose past states are most irreducibly constrained by the mechanism are its past purview (evaluated by the cause integrated information of the mechanism cause). The set of system elements whose future states are most irreducibly constrained by the mechanism are the mechanism's future purview (evaluated by the effect integrated information effect). The way that a mechanism, by being in its current state, constrains its purview elements is captured by its cause-effect repertoire, a pair of probability distributions over the past and future states of the purview elements (e.g. Fig. 3, main text). Note that these probabilities are obtained from the system's transition probability matrix (TPM) assuming a maximum entropy distribution for the marginal distribution of all possible past states. This corresponds to setting the system into all possible states with equal likelihood performing an interventionist causal analysis. cause and effect quantify the difference between the cause-effect repertoire, and the cause-effect repertoire under a partition of the mechanism as the earth-mover's distance between the two probability distributions (Oizumi et al., 2014). The overall integrated information of a mechanism is then the minimum of its cause and effect. In sum, the complete specification of a mechanism thus includes its cause and effect purviews (the elements over which it has maximally irreducible power to constrain the past and future states), the cause-effect repertoires that specify those constraints, and its integrated information value (). The set of all mechanisms constitutes the system's cause-effect structure. For a micro-level system, mechanisms are composed of micro elements and cannot include macro elements. Conversely, if a macro-level system is analyzed, only compositions of macro elements are tested and the potential causes and effects of individual micro elements within the black boxes are ignored. In other words, each level has a particular TPM, obtained from perturbing the system into all possible states with equal likelihood at that micro or macro level, which then determines the system's mechanisms at this particular level. Next, to obtain the integrated information of the system, all possible directed partitions of the system are considered, to find the one that least affects the system's cause-effect structure. After each partition, the cause-effect structure is recalculated, and the result is compared to the cause-effect structure of the whole system. The partition that makes the least difference to the cause-effect structure is the minimum information partition (MIP), and the difference it makes, as measured using an extended earth movers distance (Oizumi et al., 2014), defines the integrated information of the system (Φ). Note that, for Φ, i.e. integrated information at the system level, all possible directed partitions of constituent micro elements are evaluated, regardless of whether the system is defined at a micro or macro level. This excludes the possibility to 'hide' micro elements without cause-effect power inside a black box, which would trivially increase the system's integration. In sum, the cause-effect structure (the set of all mechanisms) of a macro 34 level system is evaluated purely at the macro level; its irreducibility, however, is evaluated by the partition between micro elements that makes the least difference to the macro cause-effect structure. Figure A: Systems of micro elements with labels to facilitate description of the mechanisms. Left: Micro system with elements labeled 0-8. Right: Macro system with elements labeled A-C. All elements are in the 'OFF' state. Example 1 -- Composition and Integration To describe the cause-effect structure of the micro system, we first assign labels to each of the micro elements in the system, as shown in Fig. A. There are three mechanisms in the cause-effect structure, they are all first order mechanisms and each one corresponds to an element implementing XOR logic. 0-8 unpartitioned Mechanism Past Purview Future Purview (0) (3) (6) (2, 7) (1, 5) (4, 8) (1, 8) (2, 4) (5, 7)  0.5 0.5 0.5 To assess the integrated information Φ of the micro system, we identify the unidirectional system partition that makes the least difference to the cause-effect structure, termed the minimum information partition (MIP). For this system, the MIP is to cut all connections from (0, 1, 2, 3, 4, 5, 6, 7) to (8). Under this partition, the mechanism specified by element 0 is altered, its future purview is reduced from (1, 6) to only (1). Note that numbers in bold refer to mechanisms for which the partitioned cause-effect structure which is different from the unpartitioned cause-effect structure. Assessing the difference between the unpartitioned and partitioned cause-effect structure of the micro system using an extended earth-mover's distance (Oizumi et al., 2014), the resulting integrated information value of the micro system is Φ = 0.25. 0-8 partitioned Mechanism Past Purview Future Purview (0) (3) (6) (2, 7) (1, 5) (4, 8) (1) (2, 4) (5, 7)  0.5 0.5 0.5 Next, we describe the cause-effect structure of the macro system displayed in Fig. A. The three black- box macro elements are constituted of micro elements A = (0, 2, 7), B = (1, 3, 4) and C = (5, 6, 8) with corresponding output elements (0), (4) and (6). This black-box system has three high-order mechanisms. Mechanism Past Purview Future Purview (A, B) (A, C) (B, C) (A, B, C) (A, B, C) (A, B, C) (C) (B) (A)  0.5 0.5 0.5 ABC unpartitioned 35 The MIP for this macro system cuts connections from (1, 3) to (0, 2, 4, 5, 6, 7, 8). After the partition, all of the mechanisms have been destroyed. Mechanisms (A, B) and (B, C) no longer have irreducible causes or effects, while the set of elements (A, C) has an effect but no cause. The integrated information of this system is Φ = 1.875. Note that the MIP is a partition of micro elements; yet the black-box system has higher Φ because the partition affects the macro cause-effect structure more than it would affect the cause- effect structure of the corresponding micro system. This is because the mechanisms at the macro level are high-order mechanisms that constrain larger parts of the system (have larger purviews). These macro constraints are completely lost even under the micro partition that makes the least difference. ABC partitioned Mechanism Past Purview Future Purview (A, B) (A, C) (B, C) () () () () (B) ()  0 0 0 Example 2 -- Local Maxima In this example with 55 elements we will not assign numbers to the elements. Instead, we will simply refer to each element based on the number of inputs and outputs it has, e.g., NOR(3, 1) for a NOR element with three inputs and one output. Each of the 55 elements specifies a first order mechanism, summarized in the table below. Multiplicity 30 15 5 5 unpartitioned -- micro Mechanism NOR(1, 1) NOR(2, 1) NOR(3, 1) NOR(1, 6) Past Purview NOR(1, 6) 2*NOR(1, 1) 3*NOR(2, 1) NOR(3, 1) Future Purview NOR(2, 1) NOR(3, 1) NOR(1, 6) 6*NOR(1, 1)  0.25 0.125 0.5 0.25 The minimum information partition (MIP) for this system is to cut the connections from a NOR(1, 1) element to the rest of the system. It doesn't matter which NOR(1, 1) element as they all have the same effect on their respective future purview. The result of the MIP is that one mechanism is destroyed, and another is altered. The integrated information of this system is Φ = 0.453. Multiplicity 29 1 14 1 5 5 Mechanism NOR(1, 1) NOR(1, 1) NOR(2, 1) NOR(2, 1) NOR(3, 1) NOR(1, 6) partitioned -- micro Past Purview NOR(1, 6) NOR(1, 6) 2*NOR(1, 1) NOR(1, 1) 3*NOR(2, 1) NOR(3, 1) Future Purview NOR(2, 1) () NOR(3, 1) NOR(3, 1) NOR(1, 6) 6*NOR(1, 1)  0.25 0 0.125 0.125 0.5 0.25 One option for a macro system is to define black-box elements that implement AND and OR logic (Fig. 5 and 6, main text). This system has an average spatial grain size of 2.75. There is a symmetry in the system, so that the mechanisms specified by each OR gate are the same, and the mechanisms specified by each AND gate are also the same (the OR elements output to six AND elements and take inputs from three AND elements, while the AND elements all take inputs from two OR elements and output to one OR element). At this macro scale, the system has 15 black-box elements implementing AND logic and 5 black- box elements implementing OR logic, over two time steps, and each specifies a first order mechanism. 36 unpartitioned -- AND/OR black boxes Multiplicity Mechanism Past Purview Future Purview 5 15 OR AND 3*AND 2*OR 6*AND OR  0.071 0.125 The MIP for this system is to cut the outputs of a NOR(2, 1) element that is one of the input elements of an AND black-box element. In this case, the mechanisms have the same cause-effect power, but one of the AND mechanisms and one of the OR mechanisms have reduced purviews, constraining less elements. Multiplicity Mechanism Past Purview Future Purview partitioned -- AND/OR black boxes 4 1 14 1 OR OR AND AND 3*AND 3*AND 2*OR OR 6*AND 5*AND OR OR  0.071 0.071 0.125 0.125 Another black-box system at a coarser macro scale has five black-box elements {A, B, C, D, E} over 4 time steps. Each black box implements a MAJORITY function over its three inputs, with a specialized connectivity pattern shown in Fig. 6, main text. Of the 31 (2N-1) possible mechanisms from the power set of 5 elements, 30 specify irreducible past and future constraints: Mechanism unpartitioned -- MAJORITY black box Past Purview Future Purview (C, D, E) (A, D, E) (A, B, E) (A, B, C) (B, C, D) (A, C, E) (A, B, C, D, E) (A, B, C, D, E) (B, D, E) (B, D, E) (A, B, C, D, E) (A, B, C, D, E) (B, C, E) (B, C, D) (C, D, E) (A, D, E) (A, B, E) (A, B, C) (C, D) (A, B, C, D, E) (A, B, C, D, E) (B, C) (D, E) (A, B, C, D, E) (A, B, C, D, E) (A, E) (A, B, C, D, E) (A, B, C, D, E) (A, C, D) (A, B, C, D) (A, C, E) (A, B, C, E) (B, C, E) (B, D, E) (A, B, C, D) (B, C, D, E) (B, D, E) (A, C, D) (A, C, D, E) (A, B) (A, B, C, D, E) (B, C, D, E) (A, B, C, D, E) (A, B, D, E) (A, B, C, D) (A, B, C, D, E) (A, B, C, D, E) (A, C, D, E) (A, B, C, E) (A, B, C, D, E) (A) (B) (C) (D) (E) (A, B) (A, C) (A, D) (A, E) (B, C) (B, D) (B, E) (C, D) (C, E) (D, E) (A, B, C) (A, B, D) (A, B, E) (A, C, D) (A, C, E) (A, D, E) (B, C, D) (B, C, E) (B, D, E) (C, D, E)  0.25 0.25 0.25 0.25 0.25 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.257143 0.2 0.257143 0.257143 0.2 0.2 0.257143 0.257143 0.2 37 (A, B, C, D) (A, B, C, E) (A, B, D, E) (A, C, D, E) (B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) 0.185709 0.185709 0.185709 0.185709 0.185709 The minimum information partition of this network is to cut the outputs of one of the NOR(1, 1) micro elements. By the symmetry in the system, there is an equivalent MIP in each of the black-box elements; however, due to the specialized connectivity structure, not all NOR(1, 1) elements are equivalent. One MIP option is to cut the hidden NOR(1, 1) micro element in black-box element A that receives input from D and outputs to the NOR(2, 1) micro elements along with the NOR(1, 1) micro element that receives input from C. As a result of the MIP, two of the mechanisms are destroyed (BCD and ABCD) and 15 others are modified (shown in bold), resulting in a Φ value of 2.333. Mechanism Past Purview Future Purview partitioned -- MAJORITY black box (A) (B) (C) (D) (E) (A, B) (A, C) (A, D) (A, E) (B, C) (B, D) (B, E) (C, D) (C, E) (D, E) (A, B, C) (A, B, D) (A, B, E) (A, C, D) (A, C, E) (A, D, E) (B, C, E) (B, D, E) (C, D, E) (A, B, C, E) (A, B, D, E) (A, C, D, E) (B, C, D, E) (C, E) (A, D, E) (A, B, E) (A, B, C) (B, C, D) (D, E) (A, B, C, E) (A, B, C, E) (B, C, E) (B, D, E) (A, B, C, D, E) (A, B, C, D, E) (B, C, E) (A, B, C, D, E) (A, C, D) (A, B, C, D) (A, C, E) (A, B, C, E) (B, C, E) (B, C, E) (B, C, D) (C, D, E) (A, D, E) (B, E) (A, B, C) (C, D) (A, B, C, D, E) (B, C, D, E) (B, C) (D, E) (B, C, D, E) (A, B, C, D, E) (A, B, D, E) (A, B, C, D, E) (A, B) (A, B, C, D, E) (B, C, D, E) (A, B, C, D, E) (B, D, E) (A, B, C, D) (A, B, C, D, E) (A, B, C, D, E) (B, D, E) (A, C, D) (A, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, C, D, E) (A, B, C, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E) (A, B, C, D, E)  0.25 0.25 0.25 0.25 0.25 0.227273 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.181816 0.257143 0.2 0.257143 0.257143 0.142158 0.257143 0.257143 0.2 0.207691 0.25 0.161903 0.185709 Black-boxing and cause-effect power: Supplementary Information William Marshall1, Larissa Albantakis1, Giulio Tononi1, * 38 1Department of Psychiatry, Center for Sleep and Consciousness, University of Wisconsin, Madison, WI, USA *Corresponding author: [email protected] S2 Text -- Black-boxing of degenerate or indeterministic systems As demonstrated in (Hoel et al., 2013, 2016), the main factor enabling an increase in intrinsic cause- effect power through coarse-graining is a reduction in indeterminism and degeneracy at the macro level. An increase in intrinsic cause-effect power by reducing degeneracy is also possible through black-boxing, as shown below ("Degeneracy example"). However, black-boxing may increase indeterminism (see below "Propagation delay through noisy channel"). Propagation delay through noisy channels In Example 1 (main text), we demonstrated that black-boxing a system with deterministic propagation delay may lead to an increase in intrinsic cause-effect power, as it can increase system integration through emergent high-order mechanisms. However, indeterminism in the system may influence whether a black- boxed macro level can "beat" the micro level in terms of cause-effect power. Here we explore the effect of noise on black-boxing by considering a scenario in which propagation delays occur over noisy channels. In this case, the COPY elements take a single input and then output the same value with probability p in [0.5, 1]. The original results of Fig. 2 (main text) refer to a noiseless channel (p = 1). A completely noisy channel (p = 0.5) would have no cause-effect power at any level (macro or micro). The integrated information analysis is performed on both the micro and black-box physical system for several different values of p (see Table 1). The number and orders of mechanisms is the same for all values of p > 0.5, but the  value of each mechanism decreases as p decreases, and it does so more steeply for the black-box system than for the micro system. Along with , the overall integrated information Φ of both the black-box and micro systems decrease as the amount of noise increases. Since Φ of the black- box system declines faster with increasing noise than Φ of the micro system, the macro system "beats" the micro for p > 0.6, whereas the micro level system "wins" when p ≤ 0.6. The reduction in cause-effect power due to indeterminism can outweigh the increase due to high-order mechanisms, since indeterminism (noise) disproportionately affects high-order mechanisms. Therefore, several interrelated aspects of cause-effect power have to be considered when assessing if the "macro beats the micro," including the presence of high- order mechanisms, indeterminism, and degeneracy. Table 1: Integrated information for various noise levels (p) in the propagation delay network (Fig. 3 main text) p Micro Φ Micro  Black box Φ Black box  1.0 0.250 0.5 1.875 0.500 0.9 0.160 0.4 1.046 0.320 0.8 0.090 0.3 0.363 0.180 0.7 0.040 0.2 0.070 0.080 0.6 0.010 0.1 0.004 0.020 0.5 0.000 0 0.000 0.000 39 Degeneracy example Reducing degeneracy, the convergence of multiple past system states onto the same current system state, is one way in which a coarse-grained macro level can achieve higher cause-effect power than its corresponding micro level (Hoel et al., 2013, 2016). Here we show that black-boxing micro elements into macro elements can also be exploited to counteract degeneracy at the micro level. Fig. A shows a system of six micro elements -- four COPY gates and two AND gates that are black- boxed into two macro elements, which correspond to macro COPY gates. As illustrated by the TPM in Fig. A (left, bottom), the micro level has a high degree of degeneracy (i.e. multiple rows leading to the same column in the TPM). In the example, two COPY micro elements input to a single AND micro element, whose current state is OFF. This implies degeneracy in the system since three states of the COPY elements (OFF, OFF), (ON, OFF) and (OFF, ON) all lead to the same state of the AND element (OFF). The micro system in state 'all OFF' specifies six first order mechanisms; the AND elements specify mechanisms with  = 0.167; and the COPY elements specify mechanisms with  = 0.25. There are no high-order mechanisms. The integrated information for the micro system is Φ = 0.215. We then consider the system at the macro level, after black-boxing each AND micro element with its two inputting COPY micro elements. Each black-box macro element turns out to implement, over two time steps, a macro COPY logic, specifying a first-order mechanism with  = 0.5. Again, there are no high- order mechanisms. However, the macro system has no degeneracy, since no two past states converge onto the same current state (see macro TPM in Fig. A). For this reason, the integrated information of a COPY element in the macro system is higher ( = 0.5) than in the micro system ( = 0.25), and so is the overall integrated information for the black-box macro system (Φ = 0.639). Thus, appropriately black-boxing micro into macro elements reduces degeneracy and in doing so increases the cause-effect power of the system, as measured by integrated information Φ. Figure A: Left: Two AND elements that each receive inputs from two COPY elements and send output to two other COPY elements, along with the TPM calculated from systematic perturbation of the elements. The current state of all micro elements is OFF. The integrated information of the micro system is Φ = 0.215. Right: Black-box elements consisting of AND elements as outputs, and the corresponding COPY elements hidden within. The current macro state of the black-box elements is OFF, corresponding to the current micro state of the micro elements that define their output. The TPM for this system is found by perturbing the inputs to the black-box element in all possible ways, it is determined that it implements COPY logic over two time steps. The black-box macro system has Φ = 0.639. 40 To describe the cause-effect structure of the micro system in detail, we use the labels shown in Fig. A. Mechanism Past Purview Future Purview o r c i m d e n o i t i t r a p n u (0) (1) (2) (3) (4) (5) (5) (5) (0, 1) (2) (2) (3, 4) (2) (2) (3, 4) (5) (5) (0, 1)  0.25 0.25 0.167 0.25 0.25 0.167 For this system, the MIP is to cut all connections from (1) to (0, 2, 3, 4, 5). Note that there are other equivalent MIPs; we focus on this specific cut without loss of generality. Under the partition, element 1 no longer has an effect and thus does not specify a mechanism, and the mechanism specified by element 2 is altered, as its past purview is reduced from (0, 1) to only (0). Comparing the unpartitioned and partitioned cause-effect structure, we find that the resulting integrated information value for the micro system is Φ = 0.215. Entries in bold highlight mechanisms in the partitioned cause-effect structure that are different from those in the unpartitioned cause-effect structure. d e n o i t i t r a p o r c i m Mechanism Past Purview Future Purview (0) (1) (2) (3) (4) (5) (5) (5) (0) (2) (2) (3, 4) (2) () (3, 4) (5) (5) (0, 1)  0.25 0 0.167 0.25 0.25 0.167 We then consider a black-box system of two COPY elements over two time steps. The black-box elements are constituted of micro elements A = (0, 1, 2) and B = (3, 4, 5), with corresponding output elements 2 and 5. The cause-effect structure of the black-box system is: black box unpartitioned Mechanism Past Purview Future Purview A B B A B A  0.5 0.5 The MIP for this system is to cut all connections from (0) to (1, 2, 3, 4, 5), that is, cut the outputs of micro element 0 (by symmetry, cutting the outputs of elements 1, 3 or 4 would be equivalent). After the partition, both mechanisms have had their irreducible cause-effect power diminished, and the result is Φ = 0.639. black box partitioned Mechanism Past Purview Future Purview A B B A B A  0.167 0.25 Black-boxing and cause-effect power: Supplementary Information William Marshall1, Larissa Albantakis1, Giulio Tononi1, * 41 1Department of Psychiatry, Center for Sleep and Consciousness, University of Wisconsin, Madison, WI, USA *Corresponding author: [email protected] S3 Text -- The intrinsic perspective The procedure for calculating the intrinsic cause-effect power of systems at a micro level is documented in detail in previous work (Oizumi et al., 2014, Mayner et al., 2016). Here we discuss aspects that deserve further comments when the analysis of intrinsic cause-effect power is applied to macro systems. Intrinsic cause-effect power To perform the intrinsic causal analysis of a system, its elements are perturbed with equal probability into all possible states and the resulting state transitions are observed. As with other interventional accounts of causality, by setting the elements into different states rather than merely observing them, true causal relationships can be distinguished from correlations. Additionally, perturbing the system into all possible states allows all counterfactuals to be evaluated, to assess the specificity of the causal relationship between elements and quantify cause-effect power. The goal of this exhaustive perturbational analysis is to unfold the full intrinsic cause-effect power of the system, that is, how the current state of system elements constrains their potential past and future states. Importantly, when analyzing the system at the micro level the micro elements of the system are set into all their possible micro states with equal probability. The same is done when analyzing the system at the macro level, except that this time it is the macro elements that are set into all their possible macro states with equal probability. In general, equal probability for macro states does not correspond to a uniform distribution of micro states. When performing intrinsic causal analysis on a black-box system, some specific considerations have to be taken into account. Most importantly, one must capture the constraints due to the macro state of the black box itself (its output) but not any constraints due to the state of the micro elements within the black box. In this regard, it is useful to consider micro elements that fall into three different categories; micro elements that impose constraints within their corresponding macro element (Fig A-A), micro elements that impose constraints on other black boxes (Fig A-B; connection from F to D), and micro elements that are outside the system (Fig A-C; D is outside the system). A single macro state may be consistent with multiple micro states of its micro elements, which nevertheless may impose different constraints on the past/future states of the system. Such micro constraints that do not originate from a black box's output element at the relevant macro update must be discounted throughout the macro causal analysis and are treated as noise. For example, consider the black-box element γ in Fig. A-A, which has micro elements C and D as inputs (receiving from the output elements of black boxes α and β respectively) and E as its output. There is feedback within this black box (bidirectional connection between micro elements C and D), hence the state of micro elements C and D matters when specifying the input-output relationship of the black box (see Table 2). However, the state of constituent micro elements should not contribute to the cause-effect power of a macro system. Therefore, when perturbing the system into a particular macro state, the relevant micro elements are perturbed with equal probability into all possible micro states that are consistent with the macro state. In practice, this procedure amounts to averaging across the input-output relation for all possible states of the hidden micro elements 42 within the black box, resulting in the truth table shown in Fig A-A, which is the average of the four truth tables resulting from the four possible states of CD (00, 10, 01, 11) (Table 2). Another instance (not dealt with in the current work) is when the output of a generalized macro element (black box) is defined as a coarse-graining of micro elements. In this case, the specific identity of these micro constituents should not contribute to the intrinsic cause-effect power of the macro system. However, in certain situations, such as when a single micro element from one coarse-grained macro element provides output to two different macro elements, constraints due to specific micro constituents can manifest in the macro TPM as "instantaneous causation" or "conditional dependence" (i.e., the current state of a macro element constrains the current state of other macro elements) and must be discounted in the causal analysis. Another type of micro constraint that needs to be discounted while evaluating the intrinsic cause- effect power of a macro-level system is 'lateral' connections between black boxes, i.e. when a micro element hidden within a black box constrains a micro element within a different black box. To ensure that only constraints due to the state of macro elements are captured in the analysis, all connections originating from micro elements within a black box and terminating outside of it are injected with noise. This includes not only connections belonging to micro elements hidden within the black box, but also the micro output element of the black box at times other than the macro time step being considered. For example, in Fig. A- B, the connection from micro element F to micro element D must be noised, since F is not an output element of a black box. Similarly, when assessing whether black-box elements α and β constrain the future state of δ over 4 time steps, the output element E of γ must be noised throughout the macro update. This is to ensure that constraints due to the state of micro elements within γ are not counted towards the cause-effect power of α or β, which here do not constrain δ directly. The black-box element γ can be thought as "screening off" the indirect effect of α and β on δ, because α and β can only effect δ via the intermediate element γ. 43 Figure A: Examples of partial systems of micro elements (A, B, C, D, E, F, G) with several different black-boxings into macro elements and the corresponding input-output function. (A): Four black-box elements α, β, γ, δ. Elements α and β have a joint constraint on γ, but it is not fully specific (there is indeterminism). All outputs of hidden micro elements are internal to their corresponding black box. (B): Similar to the (A) except constituent element F has an output that leaves its black box. The effect of F on D is not intrinsic to the macro system and must be noised during the entire causal analysis. As a result, the effects on gamma are less deterministic. (C): A similar black-boxing as in (B), except micro element D is outside the system, rather than within γ. The element D in its current state 0 is thus taken as a background condition. In this situation, the effect of α and β on γ is more specific. (D): A potential black- boxing with only three elements. In panels (A), (B) and (C), the effect of α and β on δ is screened-off by γ. In this case, the micro connection from E to F is within a black box rather than between so α and β have a direct effect on δ (it is no longer screened-off). 44 Table 2: Input-output relation for different initial states of hidden micro elements (C, D) of black box γ in panel (A) of Fig. A. If (C, D)t = (0, 0) If (C, D)t = (1, 0) (A, B)t (C, D)t+1 Et+2 (A, B)t (C, D)t+1 Et+2 0, 0 1, 0 0, 1 1, 1 0, 0 0, 0 0, 0 1, 0 0 0 0 1 0, 0 1, 0 0, 1 1, 1 0, 0 0, 0 0, 1 1, 1 0 0 1 1 If (C, D)t = (0, 1) If (C, D)t = (1, 1) (A, B)t (C, D)t+1 Et+2 (A, B)t (C, D)t+1 Et+2 0, 0 1, 0 0, 1 1, 1 0, 0 1, 0 1, 0 1, 0 0 1 1 1 0, 0 1, 0 0, 1 1, 1 0, 0 1, 0 1, 1 1, 1 0 1 1 1 In Fig. A-B, both micro elements C and D constrain the output element E, hence C and D would naturally seem to belong inside the black box γ. Nevertheless, the search for local maxima of intrinsic cause- effect power must consider all alternatives, including one in which D is taken to be a background condition rather than a hidden element within γ (or part of any other black box). In this alternate system (Fig. A-C), the state of D (OFF) is fixed as a background condition. In this case, the input-output relation for black box γ changes to the one shown in Fig A-C. From the figure alone, one cannot determine which of the two systems (top or middle panel) should qualify as the local maximum. However, it is apparent that the effect of α and β on γ is not fully deterministic when D is hidden inside γ, but it becomes deterministic when D is treated as a background condition. Furthermore, the repertoire of possible past states of γ is more degenerate when D is hidden inside γ as compared to when D is treated as a background condition (in this case there is only one possible past state of γ = 1, while with D inside the black box all four states of α and β could have led to γ = 1 with some probability). Thus, everything else being equal, the system with D as a background condition is more deterministic and less degenerate than the system with D hidden inside γ and should therefore have higher Φ. This result suggests that even if, from an extrinsic perspective, a set of micro elements may appear to constitute a macro element, from the intrinsic perspective only the set of micro elements that contribute to maximizing cause-effect power (i.e., the "skeleton" mediating the strongest constraints) actually constitutes the macro element. Integration and Exclusion When evaluating cause-effect power at macro scales, we have to consider the micro elements that constitute the macro-level system. As stated in the main text (section "Black-boxing" parts iii and iv), the integration and exclusion postulates both apply to the constituents of black-box systems. By the integration requirement, the set of constituents must be irreducible; two unrelated systems cannot be black-boxed together since their constituents are not integrated (Fig. B-A, B-B). Moreover, micro input (or output) elements cannot be black-boxed (Fig. B-C) because they lack causes (or effects) within the system. Only when the set of constituents is integrated (Fig. B-D) can a macro system be integrated. 45 Figure B: Four examples of potential black-box systems and their constituent micro elements. The constituents of a macro system must be irreducible (integration). Among the examples in the figure only (D) has a properly integrated set of constituents, while (A), (B) and (C) are reducible, with the corresponding cut drawn as a dashed red line. (A, B) Two systems that are not integrated at the micro level cannot constitute a macro, black-boxed system; (C) Elements that provide only inputs to (or only outputs from) the system cannot be constituents of a black box. By the exclusion postulate, a (macro) element must be definite. Thus, a micro constituent cannot contribute to multiple black boxes within a system (Fig. C). Figure C: Two examples of micro constituents having their cause-effect power double counted. (A) The micro element A is contributing to two different black-box elements (A, B) and (A, C, D). The exclusion postulate rules out this potential black-box system.
1808.03875
1
1808
2018-08-11T22:50:52
A single coordinate framework for optic flow and binocular disparity
[ "q-bio.NC" ]
Optic flow is two dimensional, but no special qualities are attached to one or other of these dimensions. For binocular disparity, on the other hand, the terms 'horizontal' and 'vertical' disparities are commonly used. This is odd, since binocular disparity and optic flow describe essentially the same thing. The difference is that, generally, people tend to fixate relatively close to the direction of heading as they move, meaning that fixation is close to the optic flow epipole, whereas, for binocular vision, fixation is close to the head-centric midline, i.e. approximately 90 degrees from the binocular epipole. For fixating animals, some separations of flow may lead to simple algorithms for the judgement of surface structure and the control of action. We consider the following canonical flow patterns that sum to produce overall flow: (i) 'towards' flow, the component of translational flow produced by approaching (or retreating from) the fixated object, which produces pure radial flow on the retina; (ii) 'sideways' flow, the remaining component of translational flow, which is produced by translation of the optic centre orthogonal to the cyclopean line of sight and (iii) 'vergence' flow, rotational flow produced by a counter-rotation of the eye in order to maintain fixation. A general flow pattern could also include (iv) 'cyclovergence' flow, produced by rotation of one eye relative to the other about the line of sight. We consider some practical advantages of dividing up flow in this way when an observer fixates as they move. As in some previous treatments, we suggest that there are certain tasks for which it is sensible to consider 'towards' flow as one component and 'sideways' + 'vergence' flow as another.
q-bio.NC
q-bio
A coordinate framework for optic flow and disparity Glennerster and Read A single coordinate framework for optic flow and binocular disparity Andrew Glennerster1¶* and Jenny C.A. Read2¶ 1 School of Psychology and Clinical Language Sciences, University of Reading, Reading RG6 6AL 2 Institute of Neuroscience, Newcastle University, NE2 4HH ¶ These authors contributed equally to this work. * Corresponding author. Email [email protected] (AG) A coordinate framework for optic flow and disparity Glennerster and Read Abstract Optic flow is two dimensional, but no special qualities are attached to one or other of these dimensions. For binocular disparity, on the other hand, the terms 'horizontal' and 'vertical' disparities are commonly used. This is odd, since binocular disparity and optic flow describe essentially the same thing. The difference is that, generally, people tend to fixate relatively close to the direction of heading as they move, meaning that fixation is close to the optic flow epipole, whereas, for binocular vision, fixation is close to the head-centric midline, i.e. approximately 90 degrees from the binocular epipole. For fixating animals, some separations of flow may lead to simple algorithms for the judgement of surface structure and the control of action. We consider the following canonical flow patterns that sum to produce overall flow: (i) 'towards' flow, the component of translational flow produced by approaching (or retreating from) the fixated object, which produces pure radial flow on the retina; (ii) 'sideways' flow, the remaining component of translational flow, which is produced by translation of the optic centre orthogonal to the cyclopean line of sight and (iii) 'vergence' flow, rotational flow produced by a counter-rotation of the eye in order to maintain fixation. A general flow pattern could also include (iv) 'cyclovergence' flow, produced by rotation of one eye relative to the other about the line of sight. We consider some practical advantages of dividing up flow in this way when an observer fixates as they move. As in some previous treatments, we suggest that there are certain tasks for which it is sensible to consider 'towards' flow as one component and 'sideways' + 'vergence' flow as another. A coordinate framework for optic flow and disparity Glennerster and Read Author Summary "Optic flow" refers to changes in the visual images we receive as we move through a scene. For example, as we drive along a street, the buildings flow past us and distant objects expand in our field of view. This information can tell us both about how we are moving and about the 3D structure of the scene. Conversely, "binocular disparity" refers to the differences between the views seen by our two eyes due to their slightly different positions in their head. Binocular disparity enables us to perceive the distance to objects, even while stationary. Both flow and disparity are rich sources of information, both for humans and other animals, and for robot systems such as autonomous drones and cars. Generally, they have been studied separately, and a separate set of vocabulary has been developed for each. Yet mathematically the two are fundamentally related: optic flow compares views seen by a single, moving eye at two different points in time, while binocular disparity compares views seen simultaneously by two eyes at different positions. Here, we develop a common language for describing both. It is particularly appropriate for eyes that fixate as they move, a universal feature of biological visual systems. A coordinate framework for optic flow and disparity Glennerster and Read 1. Introduction According to the ancient Indian parable, a number of blind men examined an elephant and, because each came into contact with a different part, such as the tusk or the tail, they came up with entirely different conclusions about the nature of the beast (Saxe, 1881). The same could be said about vision researchers' approaches to optic flow and binocular disparity, where the 'elephant' is the flow field generated on the retina by movement of the eye through space (or between the left and right eyes). As an illustration, imagine a train travelling along a track (Figure 1). The (monocular) train driver, looking through the windscreen, sees the scene expanding from a single point in the image (Figure 1, right) while a (monocular) passenger, looking out through a window, sees the landscape moving sideways (Figure 1, left). Yet these two very different flow fields are in fact both samples from the same spherical flow field (Figure 1, centre). The train driver is looking straight ahead at the 'direction of heading' or 'focus of expansion'. This is known as the epipole, i.e. the part of the image pierced by the translation vector. The passenger samples a quite different part of the flow field and, for him or her, the epipole is obscured from view. In neuroscience, those studying optic flow concentrate on the driver's view and those studying binocular stereopsis on the passenger's. In the case of binocular vision, the translation of the optic centre that generates flow is the interocular separation and the epipole is always obscured from view. In this case, no rays can reach the eye along the interocular axis, just like the passenger in the train who cannot see the place that his or her carriage is heading. Different nomenclatures have arisen in the two communities; for example, 'flow' is known as 'disparity' in binocular vision. However, there is no good reason to use different A coordinate framework for optic flow and disparity Glennerster and Read terms and it would be helpful to have a general framework that encompasses both. That is the goal of this paper. Figure 1. The pattern of flow produced as a train travels in a straight line. The orange arrows mark the flow lines on the visual sphere. For a passenger looking out of a side window, the flow is approximately horizontal ('sideways' flow); for the driver, the flowfield is radial ('towards' flow). One apparent point of difference between binocular disparity and optic flow is the length of the baseline. The eyes are separated by a 'baseline' distance of about 6.5cm, whereas for optic flow the baseline separating the two locations of the optic centre at time 1 and time 2 is typically considered to be much smaller, even infinitesimal. However, flow in general (i.e. optic flow or binocular disparity) is determined by the distance of the object relative to the length of the baseline, so any statements that are applicable to a wide range of object distances are applicable for both a large and a small baseline. Also, the 3D structure of the world could change between time 1 and time 2, whereas for binocular stereopsis the views are always simultaneous. However, in this paper we consider only a moving observer (or eye) A coordinate framework for optic flow and disparity Glennerster and Read in a static 3D world so that binocular vision and optic flow are equivalent in this regard. Note that we are not considering the integration of disparity and motion information as a binocular observer moves (Cormack et al, 2017). We are considering a unified framework for describing flow across the whole sphere, both near the epipole (traditionally, the part of the sphere that is considered by optic flow researchers) and around 90 degrees from the epipole (traditionally, the part of the sphere that is considered by binocular vision researchers). Examples of terms that are used in the binocular vision literature with no direct analogue in the optic flow literature are: horizontal disparity, vertical disparity, polar angle disparity and horizontal or vertical size ratios. Optic flow uses some terms that are not used in binocular vision such as 'div', 'def', and 'curl' components (Koenderink and van Doorn, 1976) or the 'centre of the expanding flow pattern' (Regan and Beverley, 1982) although these are less commonly used now. By contrast, the computer vision photogrammetry literature simply defines two basis vectors, usually based on the 2D pixel grid, to describe the change in image location of features (Hartley and Zisserman, 2003). There is no special or different treatment of flow in a 'horizontal' or 'vertical' direction -- all flow is treated equally. The question we explore in this paper is whether there is any compelling theoretical rationale for dividing retinal flow up in particular ways and attaching special meanings to different components (such as 'horizontal' or 'vertical'). An important factor will be the constraints that are introduced by the eye movements humans make, especially the fact that observers maintain fixation on a point as they move or when they view a scene binocularly. We begin by giving an intuitive overview of the main arguments in the paper, followed by a more formal description (Section 2.2). A coordinate framework for optic flow and disparity Glennerster and Read 2. Results 2.1. Overview: decomposing flow in a fixating system We begin by describing in outline the key contributions of the paper. Figure 2A illustrates two eyes (or a single eye at time 1 and time 2) fixating a point F. In biology, most animals fixate as they move (Land, 1999; Land and Nilsson, 2012). This means that as the eye translates from O1 to O2, it also rotates so as to keep the visual axis fixated on F. For simplicity, we also assume that cycloversion of the eye is minimized with respect to the scene, which would mean that the horizon projects to the same retinal meridian before and after the translation (for a review of torsional eye movements during head translation, see Angelaki et al, 2003). This yoking of rotation to translation reduces the six degrees of freedom of the camera/eye (3 translation and 3 rotation) to just three. So, in the simple case illustrated here, optic flow is the sum of just two components: translational flow caused by the displacement of the optic centre (Figure 2B) and rotational flow caused by the rotation of the visual axis necessary to maintain fixation (Figure 2C). We refer to this rotational flow as 'vergence flow', using the terminology from binocular vision. As shown in Figure 2, translational flowlines are lines of longitude on the retina, with a pole corresponding to the direction of heading (the 'epipole'). In binocular vision, the 'direction of heading' is along the interocular axis. Rotational flowlines are lines of latitude on the retina, with a pole corresponding to the axis of rotation. For a fixating eye, this is orthogonal to the plane of regard (the plane O1O2F shown in Figure 2A), so the plane of regard corresponds to the 'equator' of the rotational flowlines. A coordinate framework for optic flow and disparity Glennerster and Read Figure 2. Translational and rotational components of flow. For an observer maintaining fixation on a point F, translation from optic centre O1 to O2 has to be accompanied by a vergence rotation about an axis normal to plane O1O2F. Here, we assume cyclovergence is zero. (A): top-down view of the eyes and the fixation point. Oc is the optic centre of the cyclopean eye, defined to be midway in between Eye 1 and Eye 2. The plane containing O1, O2 and F is called the 'plane of regard'. (B): Flowlines caused by a pure translation along the epipolar vector, e. These are lines of longitude with poles defined by e. (C): Flowlines caused by a pure vergence rotation, about an axis d normal to the plane containing the optic centres and the fixation point. These are lines of latitude with poles defined by d. In this paper, we will argue that it is advantageous to further divide the translation into two: a vector t generated by translation of the optic centre towards the fixation point, F, and a vector s generated by translation of the optic centre in an orthogonal direction, i.e. sideways (Figure 3A). These components generate flow patterns we refer to as 'towards' and 'sideways' flow. For example, approaching the fixated object results in 'towards' flow (Figure 3C) while side- to-side 'bobbing' head movements to obtain distance estimates from motion parallax would result in pure 'sideways' flow (Figure 3D) if the observer were fixating a distant point so that there was no rotational flow. General translational flow (Figure 3B) can be expressed as the sum of these two flow patterns. A coordinate framework for optic flow and disparity Glennerster and Read Figure 3. Decomposition of translational flow. A) The epipolar vector, e, can be decomposed into two components, s ('sideways') and t ('towards'), where t is parallel to the line OcF and Oc is midway between O1 and O2. The factor of 2 is included for consistency with later notation. The same is true for flow: general translational flow (B) can be decomposed into 'towards' flow (C) and 'sideways' flow (D). This type of decomposition is familiar in the optic flow literature for small patches of the image. For example, in a static scene the 'divergence' component of flow for a local surface patch is generated by translation of the optic centre towards the surface (Koenderink and van Doorn, 1976; Koenderink, 1986), i.e. 'towards' flow. However, exactly the same geometry also applies to binocular vision (Figure 4). When the fixated object is on the head-centric midline (Figure 4A) the 'towards' component of translation is zero and the 'sideways' component is equal to the interocular separation. The disparity vectors are all along horizontal lines of longitude, as in the 'sideways' flowfield of Figure 3D. When an observer views an object that is 20 degrees to the left of the headcentric midline (Figure 4B), the sideways component of translation is now slightly smaller than the interocular separation and there is an additional 'towards' component of translation, leading to an expansion of the A coordinate framework for optic flow and disparity Glennerster and Read image in the left eye relative to the right. Consequently, disparities now also have components of the radial 'towards' flowfield shown in Figure 3C. If the observer now rotates his or her head about a vertical axis while maintaining fixation, so that the object is 20 degrees to the right of their midline, the 'sideways' component of translation is unchanged but the 'towards' component changes sign, so that this component now results in a relative expansion of the fixated object in the right eye (Figure 4C) compared to the left eye. In general, therefore, the 'sideways' component of disparity provides a useful signal about the shape of the surface while the 'towards' component provides a signal relating to the position of the surface relative to the midline (Backus et al (1999), Rogers and Bradshaw (1993) and Section 2.4). Figure 4. Head rotation and 'towards' flow. In binocular vision, 'towards' translation indicates fixation off the head-centric midline. (A): The observer binocularly fixates an object on the head-centric midline. The translation component is purely sideways. (B, C): Fixation 20o to the left (B) or right (C) of the midline. Now there is a 'towards' component expanding the image in the eye that is closer to fixation. Figure 5 is similar to Figure 3 but it zooms in on a small patch of the flow close to the fovea in order to illustrate in detail the behaviour of the 'towards' and 'sideways' components of flow. The illustration here is for points on a rough surface, i.e. one with random depth modulations within the surface, but the depth modulations are small compared to the mean A coordinate framework for optic flow and disparity Glennerster and Read distance to the surface. The top row (Figure 5A) shows flow that is generated when the surface fixated is 90 degrees from the epipole. Here, there is no 'towards' component but there are large 'sideways' and 'vergence' components. These two components give rise to optic flow or disparity along lines that can be considered parallel (given that we are considering a small patch). At the fixation point, these components cancel one another out exactly, so there is zero flow here. Away from the fixated point, the cancellation is not exact as shown in the first column (total flow). Flow is in one direction for points that are nearer than the fixation point and in the opposite direction for points that are further away. Of course, this is very familiar as a description of binocular disparities generated by a foveated surface, with positive values of 'horizontal' disparity indicating points that are further than the fixation point and negative values indicating nearer points. However, whenever the surface is not exactly 90 degrees from the epipole (Figure 5B) there is a component of 'towards' flow and this means that flow (including disparity) is no longer 1-dimensional. The sum of the 'vergence' and 'sideways' components still carries useful information about the surface structure whereas the 'towards' component is predominantly useful as a source of information about the relative distance of the surface from the two optic centres. We discuss this division of labour in more detail in Section 3.1. A coordinate framework for optic flow and disparity Glennerster and Read Figure 5. Flow near the fovea. (A) The fixation point is 90ofrom the epipole (i.e. on the headcentric midline for binocular viewing, hence there is no 'towards' component). (B) When fixation is at any other angle from the epipole there is a 'towards' component. The left- most column shows a top-down view of the eyes and a cloud of points around fixation. The 5 sets of axes in each row show different components of flow. In each case, lines link the projection of each point in Eye 1 with its projection in Eye 2 (the latter marked with a dot). The fovea is at the centre and grid lines mark ±5o and 10o from the fovea. Total flow (green) includes all flow components. Vergence flow (cyan) includes only rotational flow due to vergence (cyclovergence is zero in this example). Translational flow (pink) is made up of 'sideways' flow (blue) and 'towards' flow (red). The latter is zero in (A). Finally, Figure 6 illustrates the same decomposition into 'vergence', 'towards' and 'sideways' flow but now applied to features across the whole retina. Away from the fovea, the 'sideways' component of flow is no longer a set of parallel lines, nor it is parallel to the vergence flow as it was in the foveal case (Figure 5). As a result, the directions of flow or disparity of points can vary over a wide range, as shown in Figure 6A (right hand column). The bottom row (Figure 6B) shows the consequences of fixating a point closer to the epipole, hence increasing the'towards'component of the flow while reducing the 'sideways' component. This means that the total flow is close to the canonical 'towards' pattern of flow i.e. expansion outwards from the fovea (Figure 3C). A coordinate framework for optic flow and disparity Glennerster and Read Figure 6. Flow in the periphery. As for Figure 5, but now showing flow across a whole hemisphere of the visual field. As before, the top row (A) shows flow for fixation 90 deg from the epipole ('straight ahead' in the binocular case) and the bottom row (B) shows fixation (much) closer to the epipole. Total flow is again shown in green, made up of components from vergence (cyan), sideways (blue) and towards (red) flow. The final plot in each row shows a zoomed-in view of part of the eyeball, showing how the green total flow is made up of these different components. On each eyeball, we have also marked the lines of latitude for vergence flow (light cyan circles) and the lines of longitude for translational flow for the specified epipole (light pink circles). 2.2. In more detail In this section, we describe 'sideways', 'towards' and 'vergence' flow in more detail and derive the statements that we have asserted above. Precise definitions and mathematical details are presented in the Methods. We also examine in detail the case of a small surface patch and how 'sideways' and 'towards' components of flow relate to surface slant in this case. Then, in the Discussion, we will compare the division of flow suggested here with other A coordinate framework for optic flow and disparity Glennerster and Read divisions of disparity or flow proposed in the literature. We will discuss whether there is a logic for choosing any one of these divisions over others, particularly when we consider flow in general, i.e. both binocular disparity and optic flow. 2.2.1. Coordinate systems and notation Figure 7. Eye and world coordinate frames. (A): An eye-centered Cartesian coordinate system (Xe, Ye, Ze). Ze is the visual axis; Xe is 'horizontal' on the retina, and Ye is 'vertical'. (B): world-centered Cartesian system (Xw, Yw, Zw), related to (Xe, Ye, Ze) by a translation, Ne w, and a rotation (Eq 1). P is an example scene point. To help make clear the orientation of the eyeball, we have marked the pupil and iris on the front, and drawn an azimuth- latitude/elevation-longitude angular coordinate system on the retina (Read et al 2009). We use superscripts to indicate the coordinate system. Figure 7A shows an eye-based coordinate frame (Xe, Ye, Ze). By definition, the origin of this system is the nodal point of the eye. The Ze-axis is the visual axis, which in this idealised eye runs from the fovea through the nodal point and then out through the centre of the pupil; Xe and Ye define the horizontal and A coordinate framework for optic flow and disparity Glennerster and Read vertical meridians on the retina. Figure 7B shows this eye within a world-based coordinate frame (Xw, Yw, Zw). A point P can be defined in either coordinate frame (Pe or Pw). The eye- centered coordinates of a point, Pe, specify the projection of that point onto that eye. Note that the location in the eye to which the point projects is defined solely by the direction of the vector Pe, not by its length. The relationship between the world-centered coordinates Pw of the point P and its eye- centered coordinates, Pe, is: Pe = Rwe (Pw - New) Eq 1 or conversely: Pw = Rew Pe + New where the vector New specifies the location of the eye's nodal point in world-centered coordinates. The matrix Rew is the rotation matrix specifying the eye's orientation in world- centered coordinates, and Rwe is its inverse. For most of this paper, we will work in the coordinate frame of the cyclopean eye, whose location we define to be exactly halfway between Eye 1 and Eye 2 (Figure 2A). We define the visual axis of the cyclopean eye to point at the fixation point (or, more generally, the pseudofixation point, the point midway between the visual axes at their point of closest approach, for non-fixating eye postures; see Methods for details). This fixation constraint specifies two of the three degrees of freedom for the cyclopean eye orientation. Finally, we define the cyclopean eye as having zero torsion in its own coordinate system. This means that if the two eyes have any cyclovergence, they have equal and opposite cyclotorsion in cyclopean coordinates. Note that the cyclopean eye may still have non-zero torsion in world- centered or (more pertinently for binocular vision) head-centered coordinates. A coordinate framework for optic flow and disparity Glennerster and Read We will consider a scene point that projects to location Pc in the cyclopean eye. To find the flow of this point, we need to find its projection in Eye 1 and Eye 2. The flowlines move from the projection P1 in Eye 1, through the projection Pc in the cyclopean eye, to the projection P2 in Eye 2. Assuming a static world, as we do in this paper, flow is generated by the translation and rotation of the eyes. In the cyclopean frame, Eye 2 is at the epipolar vector, N2c = ec, while Eye 1 is at N1c = -ec. The rotation matrices R1c and R2c describe the orientation of each eye in the cyclopean frame. In order to plot the flowlines, we can imagine that the cyclopean eye stays constant and the world moves around it. The coordinate system of Eye 2 is related to cyclopean coordinates by a rotation R2c and a translation +ec. Thus, to find the projection of Pc into Eye 2, we apply a translation of -ec and the inverse rotation Rc1 (Figure 8). Thus, a scene point with cyclopean coordinates Pc projects into Eyes 1 and 2 as follows: P1 = Rc1 (Pc + ec); P2 = Rc2 (Pc - ec ) Eq 2 This is consistent with Eq 1, replacing e with 1,2 for the two eyes, world coordinates w with cyclopean coordinates, c, and setting N1c = -ec, N2c = ec. A coordinate framework for optic flow and disparity Glennerster and Read Figure 8. Relating the projection of a scene point, P, in the cyclopean eye and Eye 2. Eye 2 is at O2, offset from the cyclopean eye, at Oc, by the vector ec. In (A), the eyes have the same orientation, so the scene point Pc projects to Eye 2 at the same location as the point (Pc-ec), shown in red, projects to the cyclopean eye. In (B), Eye 2 is also rotated relative to the cyclopean eye as described by the rotation matrix R2 c. Hence, the inverse rotation Rc have to be applied to the vector (Pc-ec) if it were to project to the same location in the 2(Pc-ec) describes where the cyclopean eye as Pc projects to in Eye 2. Therefore, the vector Rc 1(Pc+ec) would describe where the scene point scene point Pc projects to in Eye 2. Similarly, Rc Pc projects to in Eye 1. The 'flow' generated by the rotation and translation of the eye is the curve joining these two projections. 2 would In the next section, we will decompose the flow described by Eq 1 into 4 components. First, we give a more formal account of how to decompose translation into sideways and towards A coordinate framework for optic flow and disparity Glennerster and Read components, as described above. Second, we describe decomposing rotation into vergence and cyclovergence components. 2.2.2. Decomposing translation into sideways+towards components Pure translational flow is obtained by ignoring the rotational component. From Eq 1 with no rotation, we see that flow in this case is: P1Tr = Pc + ec ; P2Tr = Pc - ec Eq 3 As is clear from Figure 8A, the flowlines here are along a great circle through the image point in question, Pc, and the epipole, ec . Some examples are shown below in Figure 10 (pink circles). In Figure 3, we introduced the idea of dividing the translational component of flow into 'sideways' and 'towards' components that are caused by orthogonal components of translation (Figure 3A): e = s + t e is the translation between the cyclopean eye and Eye 1 or Eye 2; t is the component of translation towards the fixated point, defined as being parallel to the visual axis of the cylopean eye; s is the component of translation orthogonal to t. In cyclopean coordinates, tc lies along Zc while sc is in the XcYc plane. Put another way, for pure 'towards' flow, the epipole is at the cyclopean fovea (Figure 3C). For pure 'sideways' flow, the epipole is at an A coordinate framework for optic flow and disparity Glennerster and Read eccentricity of 90o on the cyclopean retina, i.e. on the white/gold boundary in our figures (Figure 3D). In general, as we have seen, translational flow will be a mixture of 'towards' flow and 'sideways' flow (Figure 3B). From Eq 3, pure towards flow is: P1To = Pc + tc; P2To = Pc - tc Since by definition t points along the cyclopean visual axis, these flowlines are along a great circle through Pc, and the cyclopean fovea. This is shown in red in Figure 10. Similarly, pure 'sideways' flow is: P1Si = Pc + sc; P2Si = Pc - sc To find where this is on the retina, we draw a great circle through the fovea and the epipole, and see where this intersects the 90o-eccentricity meridian. This is the vector sc. Sideways flow is along the great circle through Pc and sc, as shown in blue in Figure 10. 2.2.3. Decomposing rotation into vergence+cyclovergence components A coordinate framework for optic flow and disparity Glennerster and Read Figure 9. Decomposing rotational flow into vergence and cyclovergence components. (A) For a fixating observer, the axis for vergence (dc) is at 90o from Zc . (B)For cyclovergence, the axis is Zc. From Eq 1 with no translation, we see that pure rotational flow is: P1Ro = Rc1 Pc Eq 4 P2Ro = Rc2 Pc In Figure 2, we considered the case of a fixating eye with zero cyclovergence. The rotational component of flow was then about an axis orthogonal to the plane of regard. In general, e.g. A coordinate framework for optic flow and disparity Glennerster and Read including non-biological eye movements, this may not be true. However, it is still helpful to decompose the rotational component of flow into vergence and cyclovergence components. We define the vergence axis d to be orthogonal to the plane of regard defined by the visual axes of the two eyes (plane O1O2F in Figure 2A; see Methods and Figure A1 for more details). Since the cyclopean eye is defined such that its visual axis, Zc, lies in this plane of regard, this means that the vergence axis must be orthogonal to the cyclopean visual axis, at 90o eccentricity in the cyclopean eye (Figure 9A): d .Zc = 0 The axis for cyclovergence rotations is the cyclopean visual axis, Zc (Figure 9B). As we show in the Appendix (Eq A6 - Eq A8), the rotation matrices for the two eyes can be decomposed into vergence and cyclovergence components: Rc1 = T Vc1 ; Rc2 = TT Vc2 Thus, the rotational component of flow can be further decomposed into vergence flow: P1Ve = Vc1 Pc ; P2Ve = Vc2 Pc and cyclovergence flow: P1Cy = T Pc ; P2Cy = TT Pc Eq 5 i.e. when there is non-zero cyclovergence, the torsion of Eye 1 and Eye 2 is equal and A coordinate framework for optic flow and disparity Glennerster and Read opposite with respect to the cyclopean eye. Figure 9 shows flowlines for vergence (cyan) and cyclovergence (gold) components. Note that we do not show the direction of the net rotational flow. This is because the net rotation is, in general, different in the two eyes. Because we chose to define the cyclopean eye to point to the pseudofixation point, the rotation angle from the cyclopean eye to Eye 1 is in general different from that to Eye 2. This is apparent in the example drawn in Figure 2A. This makes vergence different from all the other components, where we have defined the cyclopean eye to be exactly midway between Eye 1 and Eye 2. Because of this symmetry, cyclovergence flow, and sideways and towards flow (and therefore also net translational flow) all have the same magnitude and opposite direction for the two eyes. However, vergence flow is not necessarily equal in magnitude in the two eyes. This means that if there is any cyclovergence, the net rotational flow resulting from vergence+cyclovergence will be about a different axis in the two eyes. Thus, there is no single direction of "total rotational flow" for us to show. Full details are given in the Methods (Figure A2). Like other flow components, we have chosen to define cyclovergence relative to the cyclopean eye. We are only concerned here with flow, i.e. flow caused by changes in torsional state between the cyclopean eye, Eye 1 and Eye2 so issues relating to the torsion of the eye relative to the head in different eye positions do not concern us here (Tweed, 1997; Schreiber et al, 2001; Banks et al, 2015). Thus, in general, flowlines at a given point in the visual field are made up of contributions from four components, two translational components, 'sideways' and 'towards', along great A coordinate framework for optic flow and disparity Glennerster and Read circles on the retina, and two rotational components, 'vergence' and 'cyclovergence', along circles of latitude. 2.3. Biological constraints: The consequences of fixation In this section, we introduce constraints that apply to most animals, as we discussed in the Section 2.1. We now treat these constraints more formally and consider differences between optic flow and binocular viewing. In the previous section, we considered general flow, with the eye free to move with 6 degrees of freedom. The vast majority of animals, from insects through to humans, do not do this; instead they stabilise their gaze on an object as they move, followed by a very rapid saccade. Binocular fixation on a point is one example of this yoked pattern of eye translation and eye rotation. If the eyes are fixating on a single point in space, then the epipolar vector must lie in the same plane as the visual axes, the plane of regard (see Methods, Figure A1). Since the vergence axis, d, is defined to be orthogonal to this plane, this means that the vergence axis is also now orthogonal to the epipolar vector, as well as to the cyclopean visual axis: d .e = 0 for fixation. Relating this to the 'sideways' and 'towards' directions, for fixation, if we write: A coordinate framework for optic flow and disparity Glennerster and Read ec = (sx, sy, t) then d c = (sy , -sx , 0). Figure 10. The components of flow for two different eye postures and image locations. Top row (A,B): A general eye posture (not fixating). Bottom row (C,D): A fixating eye posture. Left column (A,C): eccentric image location. Right column (B,D): parafoveal image location. The coloured circles show the direction of flow components at the scene point Pc. A coordinate framework for optic flow and disparity Glennerster and Read Pink lines show net translational flow which is the sum of 'towards' flow (red) and 'sideways' flow (blue). ec is the epipole for translational flow. Cyan lines show vergence flow, gold lines show cyclovergence flow which together make up rotational flow (see Figure 2). For the fixating eye, the vergence axis (dc ) is orthogonal to the epipolar vector. Thus, near the fovea, vergence flow is roughly parallel to net translational flow (D). The 'sideways' flow epipole is always on the gold/white border, 90o from the fovea. The 'towards' flow epipole is at the fovea. In the case of fixation with pure 'towards' motion, d c =0, meaning the vergence axis is undefined. This is because if a fixating eye is translating directly along the line of sight, there can be no vergence movements, since any such rotation would move the fovea off the fixated object. For a general translation consisting of both 'sideways' and 'towards' components, the vergence axis, dc, will be defined and the angle of rotation about dc will vary depending on the distance to the fixation point. When the fixation point is at infinity, the amplitude of the rotational motion becomes zero. The direction of flow in this case is along the translational lines of longitude, i.e. in the same direction as epipolar lines. Figure 10 illustrates the consequences of applying the fixation constraint as the eye moves. In each panel, we have marked the direction of flow components at a particular image point Pc. The left column is for an eccentric image point, Pc, whereas the right column is for a point near the fovea. The top row shows flow when the eye translates and rotates in a general way (6 degrees of freedom) whereas the bottom row show flow when the eye fixates a point as it translates. The clearest consequence of fixation is near the fovea (Figure 10D) where the direction of translational flow (pink), 'sideways' (blue) and 'vergence' flow (cyan) are all A coordinate framework for optic flow and disparity Glennerster and Read approximately parallel at the point Pc (Figure 10B) whereas, for a general rotation and translation from Eye 1 to Eye 2, 'vergence' flow is in quite a different direction. In Section 3.2, we explore reasons why, for an animal that is moving in relation to a fixation point and making judgements about the depth of points relative to the fixation point, the division of flow described here may have some practical advantages in relation to the control of movement. But first, in the next section, we consider how this division relates to heuristics that have been proposed to recover information about the slant of a small surface patch viewed binocularly or by a moving observer. 2.4. Flow for a small surface patch In the field of binocular vision, many previous workers have pointed out that information about the slant of a surface can be derived from information about its 'horizontal' and 'vertical' size ratios (Gillam and Lawergren, 1983; Rogers and Bradshaw, 1993; Backus et al, 1999; Kaneko and Howard, 1996). Now that we are considering optic flow in a moving observer, some of the approximations that have been used in the binocular case do not apply, so it is worth revisiting this issue and being clear about the information that is present in the flow field in relation to surface slant. We will see that the important coordinates can be described in terms of the epipolar geometry alone, i.e. they are determined by the location of the optic centres in the scene and not the orientation of the eyes. The directions specified by the relevant coordinate frame are 'epipolar' (in the epipolar plane so, for a fixated object, in the plane of regard) and 'ortho-epipolar' (perpendicular to this). To simplify the description, we consider a viewer fixating a point on a surface. This allows us to refer to 'towards' and A coordinate framework for optic flow and disparity Glennerster and Read 'sideways' flow, which we have defined in relation to a fixated point, but the geometry we describe and Eqn 6 do not depend on the orientation of the eyes. The surface is small so it can be considered to be locally planar. We assume zero cyclovergence, so that flow has a 'towards' component, a 'sideways' component and a 'vergence' component which ensures that the point remains fixated (similar to 'affine' divisions of flow in a small region of the visual field, Koenderink and van Doorn (1976, 1991); Shapiro, Zisserman and Brady (1995)). At the fovea, as we have seen, sideways and vergence flowlines are parallel (e.g. Figure 10D). Now consider 4 image-points arranged on a cross around the fovea (Figure 11). Two points shown in pink in Figure 11B, are offset relative to the centre of the cross in a direction parallel to the epipolar vector, ec, hence we will call them 'E-points'. These points lie along the sideways/vergence flowline through the fovea. The scene-points corresponding to the E-points in the image lie in the plane of regard. The cyan points in Figure 11 are offset in the orthogonal direction and so we call those 'O'- points. On the retina, this direction is orthogonal to the plane of regard and hence parallel to the vergence vector, d. We define the 'E-size' and 'O-size' of the cross as the angular separation between the E- points and O-points respectively. In the cyclopean image, by construction, the cross's arms are of equal length, so E-size = O-size, but we now consider what happens in the two eyes' images of the scene points corresponding to the 4 points of this cross. We will consider the ratio of the sizes in each eye, in other words the 'E-size ratio' (ESR) and the 'O-size-ratio' (OSR). A coordinate framework for optic flow and disparity Glennerster and Read The vergence flow simply slides all points by the same amount along the sideways/vergence flowlines, shown in blue in Fig 11. The displacement will be in opposite directions in the two eyes. Thus, vergence flow changes neither E-size nor O-size; it simply re-positions the image on the retina, re-centering it on the fovea after it has been displaced by sideways flow. Figure 11. Sideways flow for a small patch. (A): Cyclopean eye showing directions of towards (red), sideways (blue) and vergence (cyan) flow. Four points are shown arranged on a cross about fixation (size exaggerated for clarity). (B,C,D): Projections of these points on the retina. B: Cyclopean image locations for no surface slant. CD: How sideways flow shifts points away from the cyclopean images, in Eye 1 (red) and Eye 2 (blue), when the cross is slanted about an axis, (C), perpendicular to the plane of regard or, (D), in the plane of regard. In (C), the left-hand end of the cross is further from the viewer so, considering A coordinate framework for optic flow and disparity Glennerster and Read 'sideways' flow alone as shown here, the images in the two eyes lie closer together than for the right-hand end. The effect is to compress the image in Eye 1 and expand it in Eye 2 along the epipolar axis, with no change along the orthogonal axis. In D, the top of the cross is further than the bottom; this shears the images without compression or expansion. Vergence flow would then shift the images equally and opposite in both eyes, such that the center of the cross moves back to the fovea. Towards flow would shift image-points along the red, radial flowlines. In C and D, the images for the left and right eyes have been given a slight vertical shift for clarity. 'Towards' flow shifts image-points radially along the red flowlines. This expands the image in one eye and contracts it in the other. The expansion/contraction is isotropic, so if there is only 'towards' flow, the E-size ratio and O-size ratio will be equal to each other: ESR = OSR. The size ratio is simply due to the relative distance of the patch to the two eyes, i.e. ESR = OSR = (size in Eye 1) / (size in Eye 2) = (distance from P to Eye 2) / (distance from P to Eye 1). For 'sideways flow', a crucial difference from vergence flow is that the magnitude of the displacement for each image-point will depend on the cyclopean distance to that image point. Thus, when the two E-ends of the cross are at the same cyclopean distance, the displacement will be the same and there will be no change in E-size in either eye. However, when the E- points are at different cyclopean distances, the E-size will be expanded in one eye and contracted in the other (Figure 11C). Thus, the E-size ratio will be different from 1. This occurs when the surface is slanted about an axis perpendicular to the plane of regard. Slant about any axis lying in the plane of regard cannot change the distance of the E-points, and so A coordinate framework for optic flow and disparity Glennerster and Read will not alter the E-size. It will, of course, change the distance of the O-points, causing the image of the cross to shear in opposite directions in the two eyes (Figure 11D). The resulting disparity gradient is a retinal cue to the surface slant about an axis in the plane of regard. Putting these elements together, if the fixated surface is face-on to the cyclopean gaze or slanted only about an axis lying in the plane of regard, the E-size ratio will equal the O-size ratio. Slant about an axis orthogonal to the plane of regard will alter the E-size ratio (because of the unequal sideways-flow) but not the O-size ratio. Thus, "ESR:OSR", the ratio of E-size ratio to O-size ratio, contains information about slant about an axis orthogonal to the plane of regard (the vergence axis, in our notation). If ESR=OSR, then the surface is not slanted about an axis orthogonal to the plane of regard (although it may be slanted about an axis lying in this plane). OSR on its own carries information about the amount of towards flow (relative to fixation distance). If OSR=1, then there is no 'towards' flow, i.e. the observer is moving past the object without approaching it. The only situation where ESR and OSR fail to give any information about surface orientation are when an observer is approaching the fixated object directly. Then, the plane of regard is undefined and ESR=OSR regardless of surface slant. When the surface is not facing the cyclopean point, the E-size in the two eyes will be different. The difference indicates the direction of slant and, in one interesting case, the magnitude. When the surface is parallel to the epipolar vector, ec, OSR and ESR have a simple relationship: ESR = OSR2. Essentially, two effects combine to influence ESR and for this particular slant the effects are of the same magnitude. One effect is caused by the ratio of distances to the two eyes, as we have seen, while the other occurs because a slanted surface is foreshortened by different amounts in the two eyes. Figure 12 illustrates this. A coordinate framework for optic flow and disparity Glennerster and Read Figure 12A shows the cyclopean eye along with the direction of the epipole, ec, and a scene point, Pc. Both the epipole and the scene point are chosen to be arbitrary, with no special relationships to each other or to the fovea. The great circle through the epipole and the scene point is shown in teal; this gives the epipolar direction through Pc. This same great circle is shown in the plane of the page in Figure 12B. In Figure 12C and D, we zoom in on the patch in question. Figure 12C shows the situation where the surface patch lies at right angles to the line of sight and hence the surface patch is tangent to the great circle shown in Figure 12B. To calculate the E/OSR in this case, we note that although the patch is shown large for illustration, it is really intended to be infinitesimal, so the angle it subtends in each eye is inversely proportional to its distance from that eye, de. Therefore, in this situation, ESR=OSR= d2/d1 = sinz1/sinz2. Even when the patch is not fixated, the same geometry applies. Figure 12D shows how the situation changes when the patch rotates about an axis orthogonal to the plane of regard, so that it is now parallel to the epipolar vector. The patch is now foreshortened in each eye. Its effective size in Eye 1, measured along the epipolar direction (i.e. in the plane of regard), is shown by the red line. Because the patch is infinitesimal, the lines of sight from O1 to either end of the patch are effectively parallel, and therefore all three shaded angles marked in Figure 12D are approximately equal to z1. Foreshortening therefore reduces the effective E- size of the patch in Eye 1 by a factor of sinz1 and in Eye 2 by a factor of sinz2. The ESR therefore acquires an additional factor of sinz1/sinz2. Thus, for any patch slanted so as to be parallel with the epipolar direction: ESR=OSR2. Eqn 6 A coordinate framework for optic flow and disparity Glennerster and Read Figure 12. The effect of surface slant on epipolar size ratio (ESR). (A) shows the cyclopean eye, the epipolar planes (pink) and a surface patch and the epipolar plane through it shown in teal. That epipolar plane is shown in (B) including the angle between the epipolar vector, ec, and the direction of the patch, Pc. The distance to the patch differs between Eye 1 and Eye 2, as shown in (C), where d2/d1= sinz1/sinz2. When the patch is facing the cyclopean point, as shown here, d2/d1 is the only factor that affects ESR. (D) When the patch is slanted so that it is parallel to the epipolar vector, ec, there is an additional foreshortening factor as illustrated by the red line. The red line is foreshortened by sinz relative to the slanted surface. So, the ratio of foreshortening effects in the two eyes is sinz1/sinz2, just like the distance effect in (C). See text for details. A coordinate framework for optic flow and disparity Glennerster and Read The literature on this relationship can be quite confusing. It is often claimed that a small fronto-parallel patch obeys the rule that its 'horizontal' size ratio (HSR) and 'vertical' size ratio (VSR) are related in the same way as Eqn 6, i.e. that HSR = VSR2. However, if 'horizontal' and 'vertical' refer to a retinal coordinate system (e.g. Howard and Rogers, 1995) then it is not always the case that HSR is the same as ESR nor that VSR is the same as OSR and, hence, it is not true in general that HSR = VSR2 for frontoparallel surfaces. We illustrate this point by considering a very extreme case where the eyes fixate a patch close to the epipole (something that is quite natural for optic flow but impossible in binocular vision as it would mean looking along the interocular axis). In this extreme case, 'horizontal' can be almost orthogonal to the epipolar direction so the difference between ESR and HSR is dramatic. Figure 13 shows how, near the epipole, the direction of epipolar planes changes very rapidly over small retinal distances. This exposes very clearly any differences between retinal and epipolar coordinate frames. The situation shown in Figure 13 corresponds to a monocular observer walking towards a point and fixating on it. As they walk, his or her head bobs up and down and from side to side, adding small lateral components to a translation vector whose main component is towards the fixation point. Figure 13(i) shows the location of the cyclopean eye (with coordinate frame attached), a small square showing the fixated surface and, in (A), Eye1 and Eye2 moving up slightly (like the head bobbing up) while in (B) the change from Eye1 to Eye2 is a slight sideways movement. The consequence of these translations of the eye is that the epipole is just below the cyclopean fovea in (A) and slightly to the left of the fovea in (B), as shown in column (ii). Columns (iii), (iv) and (v) show what happens when the surface is slanted. In each case, the patch is fixated, as shown in (i) and (ii) but the view is now zoomed in on the patch. In (iii), the patch faces the cyclopean point. As we have discussed above, this results in an overall A coordinate framework for optic flow and disparity Glennerster and Read expansion in the eye that is closer to the patch and a contraction in the other eye (pure 'towards' flow). In column (iv), the surface is slanted about a vertical axis. This results in a change in aspect ratio of the image projected into the cyclopean eye (black) and plus an overall expansion in Eye1 and a compression in Eye2, just as in column (iii). But, in addition to this 'towards' flow, there is a shear in the image in (A), in a direction parallel to epipolar lines (shown in pink) while in (B) there is an expansion of the image in Eye1, also in a direction parallel to epipolar lines, and a compression of the image in Eye2 in the same direction. When the axis of slant is rotated through 90 degrees in column (v) the cyclopean compression is rotated by 90 degrees and the type of flow is reversed between (A) and (B): now there is compression/expansion in (A) and shear in (B). Of course, all that has happened is that the direction of the epipolar lines has changed by 90 degrees between (A) and (B). The 'sideways' flow that gives rise to the shear or to the additional (slant-related) compression/expansion is always parallel to epipolar lines. A coordinate framework for optic flow and disparity Glennerster and Read Figure 13: The disparity or flow generated by a small planar patch near the epipole. (i) Two sets of three eyes are shown on the left, arranged in a line that defines the epipolar direction -- this is slightly different in the top and bottom row. All three eyes fixate the surface patch, shown as a black square that has been exaggerated in size by a factor of 30 to make it visible. In the top row, (A), the epipole is just below the fovea whereas in the bottom row, (B), the epipole is just to the left of the fovea. The columns on the left show zoomed-in pictures of the retinal projection of this patch in Eye 1, the cyclopean eye and Eye 2 (red, black, blue) when the patch is facing the cyclopean eye (iii), slanted about a vertical axis (iv), or slanted about a horizontal axis (v). The magnitude of the slant is such that in B(iv) and A(v) the patch is parallel to the line O1O2. Incidentally, the magnitude of the slant applied to the patch in column (iv) is such that it lies parallel to the epipolar vector in (B), ditto for column (v) in (A), i.e. in binocular terms, these are 'frontoparallel' patches (although, as we have said, it is clearly impossible for a binocular observer to see these patches through the side of the head). Numerically, the OSR for these patches is 1.477 (which is the same as the ESR and OSR in column (iii)) and the ESR is 2.182, confirming that ESR = OSR2 in these cases. For the patches shown in A(iv) and B(v), OSR = ESR as expected. This is because the slant axis is in the plane of regard and so 'sideways' flow only produces a shear which does not affect ESR. If we used a retinal frame to describe the patches (which stays constant with respect to the page, as shown by the grey cross) then we would have HSR = VSR2 for Figure 12 B(iv) but VSR = HSR2 for Figure 13A(v). It is important to note how extreme these conditions are compared to the situation that is usually considered for binocular vision. Read et al (2009) examined the relationship between A coordinate framework for optic flow and disparity Glennerster and Read horizontal size ratios and vertical size ratios using a retinal coordinate frame (with two alternative definitions of 'vertical'). They included caveats about the range of eccentric gaze (up to 15 degrees) and the range of retinal eccentricity (up to 15 degrees 'parafoveal' region) over which the relationships that they described would apply and the example shown in Figure 13 is very far outside these ranges. When the eyes of a binocular observer are in primary position, i.e. fixating 90 degrees away from the epipole, a longitude-longitude retinal coordinate frame (Figure 14A and B) is coincident with epipolar planes and in this case, by definition, HSR is the same as ESR and VSR is the same as OSR across the entire retina. The eyes can move a small amount away from primary position, either in a vertical or a horizontal direction, and the directions of the longitude-longitude retinal frame will not depart too much from epipolar and ortho-epipolar directions, at least if we restrict consideration to regions close to the fovea. Nevertheless, the situation in Figure 13 is relevant because we are considering optic flow as well as binocular disparity and, for optic flow, the pattern shown in Figure 13 is entirely typical. Also, it is important to establish the underlying geometry before considering approximations or implementations in a retinal frame. In relation to the range of retinal eccentricities or angles of gaze over which HSR = VSR2 might be a good approximation, Howard and Rogers (1995) made the following claim: "VSRs are unaffected by either the vergence state of the eyes or their state of eccentric gaze within the plane of regard when elevation is measured according to the gun turret coordinate system" (p282) They chose an elevation-latitude retinal coordinate frame to describe vertical position and size, so this statement is true even for large changes in vergence or gaze eccentricity. The A coordinate framework for optic flow and disparity Glennerster and Read axis of rotation for changing gaze or vergence is perpendicular to any plane of equal latitude, hence rotating about this axis does not change vertical position on the retina. However, they then go on to say: "VSRs depend only on the relative distances of the points from the two eyes." (p282) This is not the case (although it is true of OSR). The direction Howard and Rogers define as 'vertical' can differ by as much as 90 degrees from the ortho-epipolar direction, as we have seen from Figure 13, which means that 'sideways' flow can contribute to 'vertical' size. For Eqn 6 to hold, the trick is to extract a pure indicator of relative distance from the two eyes, such as 'towards' flow, and, in relation to coordinate frames, only OSR does that. 3. Discussion Having set out a framework for dividing up flow into various components, we now consider how this relates to other divisions of flow (mainly binocular disparity) that have been proposed in the past. We also discuss how dividing up flow into separate components might be useful in practice. 3.1. Relationship to other divisions of disparity and optic flow Table 1 lists a number of the coordinate systems that have been proposed to describe binocular disparity and how these relate to 'sideways', 'towards' and 'vergence' flow. 'Coordinate 1' and 'coordinate 2' correspond to common definitions of azimuth and A coordinate framework for optic flow and disparity Glennerster and Read elevation respectively, except for row 3 which describes a polar coordinate frame centred on the fovea. Coordinate 1 Coordinate 2 Mayhew et al Mayhew et al: X Mayhew et al: Y Notes 'Sideways-flow' ↔ 1982, Read et al Read et al: a Read et al: h constant value of 2009; (elevation- longitude) Coordinate 2 (This coordinate frame is a longitude-longitude Read et al 2009 Read et al: a Read et al: k system) 'Vergence-flow' ↔ Howard and Rogers: Howard and constant value of a Rogers: b Coordinate 2 (elevation- latitude), Howard and Rogers (1995) 'Towards-flow' ↔ Weinshall All: eccentricity All: Polar angle, (1990), Gårding Read et al: x Weinshall: n constant value of et al (1995), Weinshall: R Gårding et al, and Coordinate 2 Glennerster et al Gårding et al, and Glennerster et al: (2001) Glennerster et al: r q A coordinate framework for optic flow and disparity Glennerster and Read Helmholtz Read et al: z Read et al: l These coordinates relate to frame or 'pencil Change in this Change in this epipolar geometry and are of epipolar planes' coordinate is in 'E' coordinate is in independent of the direction in this paper 'O' direction in orientation of the eyes. For this paper a surface patch at the fovea, 'towards flow' is the sole contributor to change in 'O' direction Table 1. Relationship between 'sideways', 'towards' and 'vergence' flow and some divisions of disparity in the literature. In these cases, cyclovergence is assumed to be zero. The symbol ↔ denotes 'in the same direction as'. Row 1 and 2 of Table 1 both describe retinal coordinate frames with the same longitudinal system of defining eccentricity but different definitions of the vertical component (see Figure 14B and C). In Row 1, the vertical component is also defined by a longitudinal system (h). This corresponds to an (x,y) coordinate frame on a planar image, as shown in Figure 14A. It is also the direction of 'sideways' flow. On the other hand, the direction of 'vergence' flow (row 2 in Table 1) corresponds to the lines of latitude in Figure 14C (constant k). A coordinate framework for optic flow and disparity Glennerster and Read Figure 14. Retinal coordinate frames and a Helmholtz frame. This figure illustrates different possible retinal coordinate frames with the labels applied to horizontal and vertical coordinates. Horizontal is given by a in (B) and (C) or b in (D) and (E). Vertical is given by h in (B) and(D) or k in (C) and (E). Panel (F) shows a polar coordinate retinal frame (eccentricity, x, and meridional angle, q). Panel (G) shows a coordinate frame (l, z) that is dependent only on the location of the optic centres of the eyes, and so is not a retinal A coordinate framework for optic flow and disparity Glennerster and Read coordinate frame. The planes shown in (G) are the same as those indicated by the flow lines in Figure 1. Symbols for coordinates correspond to those in Read et al, (2009). Row 3 shows a different retinal coordinate frame and, in many ways, a much more familiar one with retinal eccentricity (x) and meridional angle (q) defining a polar coordinate frame centred on the fovea. Of course, 'towards' flow corresponds to lines of constant meridional angle (e.g. expanding outwards from the fovea). The only flow that contributes to changes in meridional angle are 'sideways' flow and 'vergence' flow. We discuss some possible advantages of this coordinate frame in the next section. The coordinate frame in row 4 comes into a different category because it is not a retinal frame. It is based simply on the epipolar geometry, i.e. the location of the two optic centres define a line at which a pencil of planes coincide. The elevation of the plane is the vertical coordinate (z) while azimuth in the plane defines the horizontal coordinate (l), see Figure 14G. When the two eyes are in primary position (pointing 90 degrees from the epipole), this Helmholtz coordinate system is the same as the (a,h) or (x,y) retinal coordinate frames (Figure 14A and B) but that is no longer the case as soon as the eyes move away from primary position. The l and z coordinates measure change along and orthogonal to epipolar lines respectively. We will explore the importance of measuring flow in these directions in relation to judgements of slant in Section Flow for a small surface patch. The literature on optic flow does not generally concern itself with the pros and cons of different retinal coordinate frames, unlike the debate in that exists for binocular disparity. Instead, optic flow in a local region has been divided into orthogonal affine components, or differential invariants, 'divergence', 'deformation' and 'curl'. These relate to the 1st order A coordinate framework for optic flow and disparity Glennerster and Read structure of the surface, i.e. local slant ('div', 'def', curl', see Koenderink (1986)). For a fixated surface, 'div' is in the same direction as 'towards' flow. 'Sideways' flow has contributions from both 'div' and 'def'. 'Curl' relates to cyclovergence flow. Koenderink points out that flow generated by rotation of the eye ('cyclovergence' and 'vergence' flow) are not influenced by surface structure. These descriptions of the components of flow do not assume any particular retinal coordinate frame. Koenderink's analysis of optic flow applies only to a small region of the visual field. Our proposed division of translational flow into 'towards' and 'sideways' flow applies to more than just a small patch (see Figure 6 and Figure 10). In the next section, we consider the control of heading using the detection of 'towards' flow and any deviation from this (i.e. the presence of 'sideways+vergence' flow). In this example, the proposed control strategy would be most effective if it used peripheral flow. 3.2. Some potential uses of polar flow The system of dividing up optic flow that we have described is not a general one; it is peculiarly biological. The definitions of 'sideways', 'towards', 'vergence' and 'cyclovergence' flow all rely on the assumption that the observer is fixating. But this assumption is an entirely valid one for humans and, indeed, most other animals, both in the case of static binocular viewing and for a moving observer. Hence, the division of optic flow and disparity that we suggest is relevant. Three examples are included below to illustrate why it might be useful: (i) controlling heading with respect to a fixated point; (ii) recovering the surface structure of a fixated surface and (iii) building up a representation of the visual A coordinate framework for optic flow and disparity Glennerster and Read direction of objects plus some information about their distance even when the head and eyes move. For the control of heading, consider an observer moving towards a fixated object while deviating as little as possible to the left or right. It is fairly clear that for this task, a division of flow into 'towards' flow and 'sideways + vergence' flow will be helpful because 'towards' flow corresponds to achieving the task (approaching the fixated object) while any other flow ('sideways + vergence' flow) indicates that the observer has deviated off the desired path. Glennerster et al (2001) discuss this example in more detail and consider a hierarchical method for, at the simplest level, correcting heading errors and, at the most complex level, recovering the actual direction of heading. It is hierarchical because gradually adding information can gradually increase the precision with which the direction of heading is defined. This is very similar to the notion of a hierarchical recovery of surface structure that others have proposed (Tittle et al, 1995; Glennerster et al, 1996). In relation to the recovery of surface structure, Weinshall (1990) and Gårding et al (1995) suggested that depth information about the surface could be recovered in a hierarchical manner, i.e. simple information could be recovered using a simple algorithm and then Euclidean or metric structure recovered by the addition of extra information. To do this, both papers pointed out that it would be useful to measure 'polar angle disparities' which were one component of flow (a change in the meridional angle, see row 3 of Table 1). 'Towards' flow does not change polar angle disparities (because this component is along radial lines, keeping the polar angle constant) so, by focusing on the information available from polar angle disparities they were essentially recovering information that was only available from 'sideways' flow. In particular, they showed that polar angle disparities provided useful A coordinate framework for optic flow and disparity Glennerster and Read information about the depth structure of the fixated surface (or 'relative nearness' as Gårding et al 1995 described it) -- essentially, the bas relief structure or ratio of depths of points on a surface. This is useful information but falls short of full metric depth. Because polar angle disparities are invariant to the changes in 'towards' flow, they are determined only by 'sideways + vergence' flow (although they are not exactly the same: for example, polar angle disparities are zero in the plane of regard which is not true of 'sideways + vergence' flow). Finally, in relation to building a representation that could survive head and eye movments, Glennerster et al (2001) raise a quite different reason for considering 'towards' and 'sideways + vergence' flow. The question they considered is how a moving observer might build up a representation of visual direction and approximate distance of objects surrounding them. Of course, a representation of visual direction does not involve flow, but if it is to contain information about the relative distance of objects then information from optic flow (or disparity) is required. First, Glennerster et al (2001) focus on the saccades that would take the eye between points across the entire optic array (i.e. pure rotations of the eye). Together, these saccades triangulate the whole sphere and so define the relative visual direction of objects surrounding the observer. Then, the authors consider the changes to this set of relative visual directions (hence optic flow) either when the observer moves or for the difference between the two eyes' views in binocular vision. Changes in the two coordinates of polar flow (r and q in row 3 of Table 1) provide information about the change in length of the saccades and changes in the angle between different saccades respectively. If saccade lengths and the angles between saccades are useful primitives in a primal sketch of visual direction then changes in r and q (i.e. optic flow decomposed into these two directions) are useful in building up information about the distance of objects. From this persepective, r and A coordinate framework for optic flow and disparity Glennerster and Read q are a particularly relevant pair of basis vectors for optic flow. For a fuller discussion of this proposal, see Glennerster et al 2001, Glennerster et al 2009 and Glennerster, 2016). 4. Conclusion We have set out a general framework to describe flow that includes both binocular disparity and optic flow generated by observer (or camera) movement. We have included a description that applies to any translation and rotation of the eye (or camera), given that the translation of the eye is small compared to the distance to the scene points but we have then added constraints that are generally true for human vision such that the rotation of the eye as it translates is constrained (the eye fixates a single point and its torsion around the line of sight is such that the horizon tends to project to the same meridian on the retina). The consequence is that the types of image change (or flow) that can occur during small translations of the optic centre (including the translation from left to right eye in binocular vision) are very restricted compared to what they might be if eye rotation was not yoked to eye translation as it is in human vision. Given this restriction, we have discussed some of the advantages of separating radial flow with respect to the fovea from flow in an orthogonal direction. Overall, this treatment is more general than, and quite different from, the traditional separation of disparity into 'horizontal' and 'vertical' components. 5. Methods A coordinate framework for optic flow and disparity Glennerster and Read In this section, we set out the notation for describing the relationship between world and eye coordinates and for flow as the eye translates and rotates. There is some overlap with the notation developed by Read et al (2009) for binocular vision. However, in making this treatment more general, there are also some important differences particularly in relation to the orientation of the cyclopean eye (or, for optic flow, the eye at a mid-point between time 1 and time 2). 5.1. Eye and world frames Figure 7 showed the relationship between the eye-centered coordinate system (Xe, Ye, Ze) and the world-centered coordinate system (Xw, Yw, Zw). We will always include superscripts when we are mixing frames. When we omit superscripts, this indicates that any frame can be chosen so long as it applies to all vectors. For example, the dot product between two position vectors is independent of which frame is chosen to express the vectors: a.b = aw.bw= ae.be. When we refer to an eye's "location", we mean the location of its center of projection. Eye 1 and Eye 2 are located at N1w and N2w respectively in world-centered coordinates. We define the epipolar vector which points from Eye 1 to Eye 2: e = (N2-N1)/2. The cyclopean eye is defined to be in between the two eyes: Nc = (N2+N1)/2 In general, we therefore have N1 = Nc - e ; N2 = Nc + e. In cyclopean coordinates, the cyclopean eye is at the origin: Ncc=0. In cyclopean coordinates, therefore, the locations of the two eyes are simply ±ec . Each eye has its own rotation matrix Rew specifying its orientation with respect to world- A coordinate framework for optic flow and disparity Glennerster and Read centered coordinates. Given a direction in eye-centered coordinates, ve, to express this direction in world-centered coordinates we rotate by Rew: vw = Rew ve. Conversely, to express a world-centered direction in eye-centered coordinates, we rotate by Rwe , the inverse of Rew (i.e. its transpose, since this is a rotation matrix): ve = Rwe vw. In particular, the rotation matrix determines the direction of the eye's visual axis in world- centered coordinates, Zew. We defined the eye-centered coordinate system such that the visual axis is the Z-axis, so in world-centered coordinates Zew = RewZ. The visual axis specifies the elevation and azimuth of the eye's gaze; the torsion about the visual axis is then needed to specify the rotation matrix. 5.2. The pseudofixation point Figure A1 shows the three eyes along with their visual axes, which are parallel to Z1,2,c respectively. In general, as in this example, the visual axes do not intersect. F is the pseudofixation point, defined to be the midpoint of the line joining the points where the visual axes approach most closely. This line is perpendicular to both visual axes, i.e. is parallel to Z1´Z2. We define the vergence vector d to be the vector Z1´Z2. Note that this is not a unit vector; its magnitude is equal to the sine of the angle between the visual axes, i.e. the vergence angle HD. Let 2dd be the vector through F linking the visual axes; if the eyes are fixating, d=0. A coordinate framework for optic flow and disparity Glennerster and Read dd F dd F2Z2 O2 dd e Oc f2 f1 FcZc F1Z1 O1 Figure A1. Definition of a 'pseudofixation point'. The locations of the three eyes, and the pseudofixation point F. The distance 2dd is the minimum separation of the visual axes from Eye 1 and Eye 2; if the eyes are fixating, d=0. The shaded plane is the generalised plane of regard, formed by vectors parallel to the visual axes of the three eyes, Z1,Z2 and Zc. In this example, the visual axes starting from Eye 1 and Eye 2 (heavy red, blue lines labelled Z1,Z2) do not intersect, but when offset by the vector dZ1´Z2 (light lines of the same colour) they lie in the generalised plane of regard along with the vector Zc starting from the cyclopean eye. For a general eye position, the epipolar vector e (pink) does not lie in the generalised plane of regard. If the eyes are fixating, then d=0 and e does also lie in this plane. F1,2,c are the distances from each eye to the point on its visual axis which is closest to the pseudofixation point. We have not indicated the frame, since the same relationships hold whether the vectors are expressed in world-centered coordinates, Z1w,Z2w, Zcw and ew, or cyclopean-eye centered coordinates Z1c,Z2c, Z and ec. A coordinate framework for optic flow and disparity Glennerster and Read We define F1,2,c to be the distance from each eye along its visual axis to the point of closest approach to the pseudofixation point F. Examining Figure 14, we see that Fc Zc = e + F2Z2 -- d d = -- e + F1Z1 + d d Eq A1 We can solve Eq A1 to obtain Zc. Taking the inner product of both sides with Z1´Z2, and using the fact that Z1.(Z1´Z2)= Z2.(Z1´Z2)=0, we obtain Fc Zc.d = e .d -- d = -- e.d + d . Eq A2 This equation states that Fc Zc. d is equal to its own negative, so must be zero. Since the fixated object is not inside the eye, Fc>0 and so and thus Zc. d = Zc. (Z1´Z2) = 0. That is, with our definition of the cyclopean eye, the visual axes of all three eyes lie in a plane, the generalised plane of regard, shown shaded in Figure 14. This definition is fully general and does not require the eyes to be fixating on a single point in space. If the eyes are fixating on a single point, then the distance d in Figure 14 is zero. It follows from Eq A2 that e.d = 0 for a fixating eye posture i.e. for fixation, the axis of vergence must lie at 90 degrees from the epipole. In the body of the paper, we have referred to the plane of regard as plane O1O2F. This remains true and is one special case of the generalised plane of regard defined here, i.e. with d = 0. A coordinate framework for optic flow and disparity Glennerster and Read 5.3. Deriving the cyclopean eye's rotation matrix We have chosen to make the cyclopean eye look directly at the pseudofixation point. This fixes the azimuth and elevation of the cyclopean eye. As we stated in the main text, we define its torsion to be exactly in between that of Eye 1 and Eye 2. We now go through how this works. 5.3.1. Deriving the cyclopean eye's rotation matrix given the locations and orientations of Eye 1 and Eye 2 in world-centered coordinates To begin with, we assume we are given N1w and N2w and the rotation matrices R1w and R2w, which specify how the two physical eyes are positioned with respect to world-centered coordinate axes. The location of the cyclopean eye in world-centered coordinates is easy: Ncw = (N2w+N1w)/2. We now explain how we define its rotation matrix, Rcw. The rotation matrix R1w gives us the coordinate axes of Eye 1 expressed in world-centered coordinates. For example, the visual axis of Eye 1, expressed in world-centered coordinates, is: Z1w = R1wZ Conversely, the inverse rotation, Rw1 , rotates the visual axis of Eye 1, expressed in world- centered coordinates, back onto the world-centered Z axis: Rw1Z1w = Z For Eye 2, similarly Z2w = R2wZ and so Z2w = R2wRw1Z1w. A coordinate framework for optic flow and disparity Glennerster and Read The rotation matrix R2wRw1 rotates the coordinate frame of Eye 1 onto the coordinate frame of Eye 2, when both frames are expressed in world-centered coordinates (Fig A2A). If there were no cyclovergence between the eyes, R2wRw1 would be a pure vergence rotation about dw : R2wRw1 = Rot[Z1w´Z2w] = Rot[dw] (for zero cyclovergence). where we have introduced the notation that Rot[v] is a rotation matrix defining a rotation through angle arcsin(v) about vector v. Z1´Z2 is parallel to the vergence axis d, and its magnitude is equal to the sine of the vergence angle between the visual axes of Eye 1 and Eye 2, HD (cyan rotation in Fig A2B). However, in general, the eyes will also be cycloverged relative to one another and therefore we also need to add a rotation around the visual axis through the cyclovergence angle TD (gold rotation in Fig A2B). It doesn't matter whether we first cycloverge through TD about visual axis 1 and then do the vergence from Eye 1 to Eye 2, as shown in Fig A2B: R2wRw1 = Rot[dw]Rot[Z1wsinTD] or whether we first verge from Eye 1 to Eye 2 and then cycloverge about visual axis 2: R2wRw1 = Rot[Z2wsinTD]Rot[dw] or whether we verge from Eye 1 to the cyclopean eye, cycloverge about the cyclopean visual axis, and then continue the vergence onto Eye 2: R2wRw1 = Rot[Zcw´Z2w] Rot[Zcwsin(TD)] Rot[Z1w´Zcw] A rotation about axis 1 followed by a rotation about axis 2 is identical to a rotation about axis Eq A3 1, followed by a rotation about the new axis 2: Rot[v2]Rot[v1] = Rot[ Rot[v2] v1 ] Rot[v2] A coordinate framework for optic flow and disparity Glennerster and Read Using this identity, it is straightforward to confirm that the three alternative expressions for R2wRw1 are the same. We now need to solve for the cyclovergence angle. Taking Eq A3, for example, we have Rot[Z2wsinTD] = R2wRw1 Rot[-dw] All the quantities in the right-hand side are known, so it is easy to solve for the cyclovergence TD. Using the fact that the trace of a rotation matrix is 1 plus twice the cosine of the rotation angle, we have 2 cosTD = Tr( R2w Rw1 Rot[-dw] ) - 1 Eq A4 We are now finally in a position to define the rotation matrix of the cyclopean eye. the rotation matrix RcwRw1 rotates the coordinate frame of Eye 1 onto the coordinate of the cyclopean eye, while the rotation matrix R2wRwc rotates the coordinate frame of the cyclopean eye onto the coordinate frame of Eye 2 (all expressed in world-centered coordinates). As noted, we wish to define the cyclopean coordinates such that they are exactly midway between the coordinate frame of the two eyes. Thus, to get to cyclopean coordinates from Eye 1, we first do a vergence rotation through f1, and then rotate through half the cyclovergence angle, which we call Td in accordance with the notation of Read et al. (2009). RcwRw1 = Rot[Zcwsin(Td)] Rot[Z1w´Zcw] Then, to get to Eye 2 from cyclopean coordinates, we rotate through the other half of the cyclovergence angle and then do a vergence rotation through f2: R2wRwc = Rot[Zcw´Z2w] Rot[Zcwsin(Td)] The cyclopean rotation matrix is then easily solved for: Rcw = Rot[Zcwsin(Td)] Rot[Z1w´Zcw] R1w or equivalently A coordinate framework for optic flow and disparity Glennerster and Read Rcw = Rot[-Zcwsin(Td)] Rot[Z2w´Zcw] R2w Eq A5 Thus, we have now produced a complete definition of the cyclopean eye for any eye postures R1w and R2w. Fig A2. How the rotation from Eye 1 to Eye 2 can be decomposed into a vergence and a torsion component. In each panel, the shaded plane indicates the plane of regard. AB: In world-centered coordinates; CD: in cyclopean coordinates. A: Rotation matrix R2wRw1 maps the coordinate axes of Eye 1 into the coordinate axes of Eye 2. In general, this rotation is about an arbitrary axis, shown by the heavy purple line. B: This can be decomposed into a vergence rotation through angle HD = f 1+ f 2 about the vergence axis dw, plus a A coordinate framework for optic flow and disparity Glennerster and Read cyclovergence rotation through angle TD about the visual axis. C: Matrix R1c maps the cyclopean eye to Eye 1, and Rc2 maps the cyclopean eye to Eye 2. In general, these rotations are about different axes, shown with the red,blue heavy lines. D: Shows how R2c can be decomposed into a cyclovergence component T, which is a rotation through Td about the cyclopean visual axis, followed by a vergence rotation Vc2 through f2 about the vergence axis dc. Similarly, R1c can be decomposed into a cyclovergence rotation TT through -Td about the cyclopean visual axis, followed by a vergence rotation Vc1 through - f 1 about the vergence axis (Eq A6). 5.3.2. Deriving cyclopean rotation matrices for Eye 1 and Eye 2, given cyclopean visual axes and cyclovergence For many applications, we do not care about where the eyes are located in a world-centered coordinate system, but simply want the rotation matrices relating Eye 1 and Eye 2 to the cyclopean eye. We can get these simply from Eq A5 by defining the world-centered frame to be equal to the cyclopean frame. Rcw is then the identity matrix, Zcw is simply the Z axis. We then have Rc1 = Rot[ZsinTd] Rot[Z1c´Z] ; R1c = Rot[Z´Z1c] Rot[-ZsinTd] Rc2 = Rot[-ZsinTd]Rot[Z2c´Z] ; R2c = Rot[Z´Z2c] Rot[ZsinTd] We define the cyclovergence rotation matrix T = Rot[ZsinTd] Eq A6 Eq A7 and two different vergence rotation matrices, both about the vergence axis dc but through A coordinate framework for optic flow and disparity Glennerster and Read different angles: Vc1 = Rot[Z1c´Z]; Vc2 = Rot[Z2c´Z] Eq A8 As an example, suppose we are told (i) the eyes are fixating; (ii) the distance of the fixation point Fc from the cyclopean eye, (iii) the epipolar vector in cyclopean coordinates, ec, and (iv) the cyclovergence TD. From this information, we can deduce the direction of vergence vector dc, since it is perpendicular to both Z and (for a fixating eye posture) to the epipolar vector ec. From Eq A1 with d=0 for fixation, we have Fc Z = ec + F2Z2c = -- ec + F1Z1c This gives us the direction of both visual axes: F2Z2c =Fc Z -- ec ; F1Z1c =Fc Z + ec Since Z1,2,c are unit vectors, this also gives us the distance from each eye to the fixation point: F1 = Fc Z + ec ; F2 = Fc Z -- ec and thus the direction of each visual axis: Z1c = (Fc Z + ec )/F1; Z2c = (Fc Z -- ec )/F2. We can then solve Eq A6 to obtain the rotation matrices of the two eyes in this cyclopean frame. For some applications, one may wish to define the horizontal meridian of the cyclopean eye to be the plane of regard. In this case, simply follow the scheme above but with ec parallel to X. A coordinate framework for optic flow and disparity Glennerster and Read 5.4. Relationship to Helmholtz coordinates for binocular vision In binocular vision, the "world-centered" coordinate system will usually be head-centered. Here, we consider the head-centered coordinate system of Read et al (2009). Then Eye 1 is the left eye and Eye 2 the right. 𝝐#=−&'(100+ ; 𝑵𝟏#=−𝝐#=&'(100+; 𝑵𝟐#=𝝐#=−&'(100+; 𝑵𝒄#=(000+ The head-centered rotation matrices are R w e = 1 é ê 0 ê 0 ê ë 0 V cos e V sin e 0 V sin - e V cos e ù ú ú ú û é ê ê ê ë e H cos 0 sin H - 0 1 0 e e H sin 0 H cos e ù ú ú ú û T cos é e ê T sin ê e 0 ê ë T sin - e T cos e 0 0 0 1 ù ú ú ú û where Ve, He, Te are respectively the Helmholtz elevation, azimuth and torsion for Eye e (see below for the definitions of these for the cyclopean eye). The visual axis for Eye e points in the direction sin𝐻1 −sin𝑉1cos𝐻1 cos𝑉1cos𝐻1+ 𝒁1#=( 𝜹#=𝒁:#×𝒁𝟐#=sin𝐻<( 0cos𝑉sin𝑉 + the vergence axis is For fixating eye postures, the Helmholtz elevation must be the same for all eyes: Ve=V. Then That is, for fixating eye postures, the vergence rotation is a rotation through the vergence angle HD about an axis tilted forward from the Y axis by the Helmholtz elevation angle V. The cyclovergence, defined in Eq A4, is TD=(T2-T1), i.e. the difference between the Helmholtz torsions for right and left eye. The cyclotorsion of the cyclopean eye is Tc=(T2+T1)/2, the mean A coordinate framework for optic flow and disparity Glennerster and Read of the Helmholtz torsions, as defined in Read et al (2009). The vergence angle is HD=(H2-H1), again as in Read et al (2009). However, note that we here define the azimuth of the cyclopean eye differently. In Read et al (2009), we defined the cyclopean azimuth to be the mean of the Helmholtz azimuths for right and left eye. However, this means that the visual axis of the cyclopean eye does not point at the fixation point. In order to achieve that, we here define the cyclopean azimuth slightly differently: Hc = H1+ f 1 = H2- f 2 where the angles f are defined in Figure A2 and Figure A3. sin𝜃:= sin𝜃'= cos𝐻:sin𝐻< cos𝐻'sin𝐻< ?sin'(𝐻:+𝐻') +4cos'𝐻:cos'𝐻' ?sin'(𝐻:+𝐻') +4cos'𝐻:cos'𝐻' This means that Hc = (H1+ H2 + f 1 -f2) / 2 differing by (f 1- f 2)/2 from the gaze angle of the cyclopean eye defined in the earlier paper. A coordinate framework for optic flow and disparity Glennerster and Read Figure A3. Definition of vergence angle and gaze azimuth. The vergence angle is HD = H2- H1 = f 2+ f 1. The cyclopean gaze azimuth is Hc = H1+ f 1 = H2- f 2. 5.5. Useful relationships Here we collect together for reference the terms defined in this paper and some useful relationships between them. 5.5.1. Vectors All vectors can be given superscripts to indicate which coordinate system they are defined in. N1, N2 : location (i.e. center of projection) of Eye 1, Eye 2 respectively. Nc = (N2+N1)/2 : location of the cyclopean eye e = (N2-N1)/2 : the epipolar vector. This points in the direction of the epipole, and we A coordinate framework for optic flow and disparity Glennerster and Read define it as half the vector linking Eye 1 and Eye 2. Z1,Z2, Zc: unit vectors along the visual axis of the respective eye. d = Z1´Z2 : the vergence vector, pointing along the axis for vergence rotations. Note that d .Zc=0. For fixation, e.Zc=0. P: an example scene point. 5.5.2. Rotation matrices Rcw: The rotation matrix expressing the coordinate frame of the cyclopean eye in world- centered coordinates (and similarly for R1w,R2w) Zcw = RcwZ : The visual axis of the cyclopean eye in world-centered coordinates (and similarly Z1w = R1wZ; Z2w = R2wZ). Rwc = (Rcw)T: The rotation matrix expressing the world-centered coordinate frame in the coordinates of the cyclopean eye. R1c : The rotation matrix expressing the coordinate frame of Eye 1 in cyclopean coordinates (and similarly for R2c). Z1c = R1cZ : The visual axis of Eye 1 in cyclopean coordinates. Z2c = R2cZ : The visual axis of Eye 2 in cyclopean coordinates. Z1w = R1wZ : The visual axis of Eye 1 in world-centered coordinates. Z2w = R2wZ : The visual axis of Eye 2 in world-centered coordinates. R1w = (Rw1)T R2w = (Rw2)T R1c = (Rc1)T = Rwc R1w = Rot[Z´Z1c] Rot[-ZsinTd] = Rot[-d^c sinf1] Rot[-ZsinTd] R2c = (Rc2)T = Rwc R2w = Rot[Z´Z2c] Rot[ZsinTd] = Rot[+d^c sinf2] Rot[+ZsinTd] Eq_RotationMatrices A coordinate framework for optic flow and disparity Glennerster and Read 5.5.3. Mapping between visual axes Z1w = R1wZ = R1wRc1Z1c = R1wRw2 Z2w = R1wRwc Zcw Z2w = R2wZ = R2wRc2Z2c = R2wRw1 Z1w = R2wRwc Zcw Zcw = RcwZ =RcwRw1Z1w = RcwRw2Z2w Z1c = R1cZ = R1cRc2Z2c Z2c = R2cZ = R2cRc1Z1c 5.5.4. Angles HD is the vergence angle, i.e. the angle between the visual axes Z1 and Z2. This is consistent with the definition in Read et al (2009). In our definition, the magnitude of the vergence vector is the sine of the vergence angle: d = sinHD We define f1,2 to be the angle between the cyclopean visual axis and the visual axis of Eye 1, 2. The sign is defined by 𝒁:×𝒁D=𝜹Esin𝜙: 𝒁D×𝒁'=𝜹Esin𝜙' where 𝜹E is the unit vector parallel to d. In terms of these angles: HD = f1+f2 In terms of the gaze azimuths in Helmholtz coordinates for binocular vision, as defined in Read et al. (2009) HD = H2-H1. The Helmholtz gaze azimuth of the cyclopean eye is Hc = (H1+ H2 + f1 - f2) / 2 A coordinate framework for optic flow and disparity Glennerster and Read (slightly different from the definition (H1+H2)/2 used in Read et al (2009). TD is the cyclovergence between Eye 1 and Eye 2. In terms of the torsion in Helmholtz coordinates for binocular vision, as defined in Read et al. (2009) TD=(T2-T1). 5.5.5. Converting between coordinate systems We only use the epipolar and vergence vectors as directions, so to convert between world- centered and cyclopean frames, we simply rotate: e c = Rwcew ; d c = Rwcdw ew = Rcwec ; dw = Rcwdc When projecting scene points into the eyes, we also need to take into account the translation offset between frames: P1 = Rw1 (Pw -- N1w) = Rc1 (Pc + ec) P2 = Rw2 (Pw -- N2w) = Rc2 (Pc - ec) Pc = Rwc (Pw -- Ncw) where Pw gives the world-centered coordinates of a scene point P (i.e. the world-centered vector from the world-centered origin to the scene point), and Pe gives the coordinates relative to Eye e (i.e. the eye-centered vector from the nodal point of that eye to the scene point). Acknowledgements A coordinate framework for optic flow and disparity Glennerster and Read Supported by EPSRC EP/K011766/1 and EP/N019423/1 to AG and Leverhulme Trust Research Leadership Award RL-2012-019 to JCAR. References Angelaki DE, Zhou HH, Wei M. Foveal versus full-field visual stabilization strategies for translational and rotational head movements. Journal of Neuroscience. 2003 Feb 15;23(4):1104-8. Backus BT, Banks MS, van Ee R, Crowell JA. Horizontal and vertical disparity, eye position, and stereoscopic slant perception. Vision Research. 1999 Mar 31;39(6):1143-70. Banks MS, Sprague WW, Schmoll J, Parnell JA, Love GD. Why do animal eyes have pupils of different shapes?. Science Advances. 2015 Aug 1;1(7):e1500391. Cormack, L. K., Czuba, T. B., Knöll, J., & Huk, A. C. (2017). Binocular Mechanisms of 3D Motion Processing. Annual review of vision science, 3, 297-318. Gårding J, Porrill J, Mayhew JE, Frisby JP. Stereopsis, vertical disparity and relief transformations. Vision Research. 1995 Mar 31;35(5):703-22. Gillam B, Lawergren B. The induced effect, vertical disparity, and stereoscopic theory. Perception & psychophysics. 1983 Mar 1;34(2):121-30. A coordinate framework for optic flow and disparity Glennerster and Read Glennerster A. A moving observer in a three-dimensional world. Phil. Trans. R. Soc. B. 2016 Jun 19;371(1697):20150265. Glennerster A, Hansard ME, Fitzgibbon AW. Fixation could simplify, not complicate, the interpretation of retinal flow. Vision Research. 2001 Mar 1;41(6):815-34. Glennerster A, Hansard ME, Fitzgibbon AW. View-based approaches to spatial representation in human vision. In Statistical and geometrical approaches to visual motion analysis 2009 (pp. 193-208). Springer, Berlin, Heidelberg Glennerster A, Rogers BJ, Bradshaw MF. Stereoscopic depth constancy depends on the subject's task. Vision Research. 1996 Nov 30;36(21):3441-56. Hartley R, Zisserman A. Multiple view geometry in computer vision. Cambridge University Press; 2003. Howard IP, Rogers BJ. Binocular vision and stereopsis. Oxford University Press, USA; 1995. Kaneko H, Howard IP. Relative size disparities and the perception of surface slant. Vision Research. 1996 Jul 31;36(13):1919-30. Koenderink JJ. Optic flow. Vision Research. 1986 Dec 31;26(1):161-79. Koenderink JJ, van Doorn AJ. Local structure of movement parallax of the plane. JOSA. 1976 Jul 1;66(7):717-23. A coordinate framework for optic flow and disparity Glennerster and Read Koenderink JJ, Van Doorn AJ. Affine structure from motion. JOSA A. 1991 Feb 1;8(2):377- 85. Land MF. Motion and vision: why animals move their eyes. Journal of Comparative Physiology A: Neuroethology, Sensory, Neural, and Behavioral Physiology. 1999 Oct 8;185(4):341-52. Land MF, Nilsson DE. Animal eyes. Oxford University Press; 2012. Mayhew JE, Longuet-Higgins HC. A computational model of binocular depth perception. Nature. 1982 Jun 3;297(5865):376-8. Read JC, Phillipson GP, Glennerster A. Latitude and longitude vertical disparities. Journal of Vision. 2009 Dec 1;9(13):11-. Regan D, Beverley KI. How do we avoid confounding the direction we are looking and the direction we are moving? Science. 1982 Jan 8;215(4529):194-6. Rogers BJ, Bradshaw MF. Vertical disparities, differential perspective and binocular stereopsis. Nature. 1993 Jan 21;361(6409):253-5. Saxe, J. G. The Poems of John Godfrey Saxe (Highgate Edition) Houghton, Mifflin and Company; 1881 A coordinate framework for optic flow and disparity Glennerster and Read Schreiber K, Crawford JD, Fetter M, Tweed D. The motor side of depth vision. Nature. 2001 Apr 12;410(6830):819-22. Shapiro LS, Zisserman A, Brady M. 3D motion recovery via affine epipolar geometry. International Journal of Computer Vision. 1995 Oct 1;16(2):147-82. Tittle JS, Todd JT, Perotti VJ, Norman JF. Systematic distortion of perceived three- dimensional structure from motion and binocular stereopsis. Journal of Experimental Psychology: Human Perception and Performance. 1995 Jun;21(3):663. Tweed D. Visual-motor optimization in binocular control. Vision Research. 1997 Jul 31;37(14):1939-51. Weinshall D. Qualitative depth from stereo, with applications. Computer Vision, Graphics, and Image Processing. 1990 Feb 1;49(2):222-41.
1809.02511
1
1809
2018-09-07T14:39:34
A next generation neural field model: The evolution of synchrony within patterns and waves
[ "q-bio.NC", "math.DS", "nlin.PS" ]
Neural field models are commonly used to describe wave propagation and bump attractors at a tissue level in the brain. Although motivated by biology, these models are phenomenological in nature. They are built on the assumption that the neural tissue operates in a near synchronous regime, and hence, cannot account for changes in the underlying synchrony of patterns. It is customary to use spiking neural network models when examining within population synchronisation. Unfortunately, these high dimensional models are notoriously hard to obtain insight from. In this paper, we consider a network of $\theta$-neurons, which has recently been shown to admit an exact mean-field description in the absence of a spatial component. We show that the inclusion of space and a realistic synapse model leads to a reduced model that has many of the features of a standard neural field model coupled to a further dynamical equation that describes the evolution of network synchrony. Both Turing instability analysis and numerical continuation software are used to explore the existence and stability of spatio-temporal patterns in the system. In particular, we show that this new model can support states above and beyond those seen in a standard neural field model. These states are typified by structures within bumps and waves showing the dynamic evolution of population synchrony.
q-bio.NC
q-bio
A next generation neural field model: The evolution of synchrony within patterns and waves ´ Aine Byrne∗ Center for Neural Science, New York University, New York, NY 10003 USA. and Centre for Mathematical Medicine and Biology, School of Mathematical Sciences, University of Nottingham, University Park, Nottingham, NG7 2RD, UK. Daniele Avitabile arXiv:1809.02511v1 [q-bio.NC] 7 Sep 2018 Centre for Mathematical Medicine and Biology, School of Mathematical Sciences, University of Nottingham, University Park, Nottingham, NG7 2RD, UK and Inria Sophia Antipolis Mditerran´e Research Centre, MathNeuro Team, e 2004 route des LuciolesBote Postale 93 06902 Sophia Antipolis, Cedex, France. Stephen Coombes . Centre for Mathematical Medicine and Biology, School of Mathematical Sciences, University of Nottingham, University Park, Nottingham, NG7 2RD, UK. (Dated: September 10, 2018) Neural field models are commonly used to describe wave propagation and bump attractors at a tissue level in the brain. Although motivated by biology, these models are phenomenological in nature. They are built on the assumption that the neural tissue operates in a near synchronous regime, and hence, cannot account for changes in the underlying synchrony of patterns. It is customary to use spiking neural network models when examining within population synchronisation. Unfortunately, these high dimensional models are notoriously hard to obtain insight from. In this paper, we consider a network of θ-neurons, which has recently been shown to admit an exact meanfield description in the absence of a spatial component. We show that the inclusion of space and a realistic synapse model leads to a reduced model that has many of the features of a standard neural field model coupled to a further dynamical equation that describes the evolution of network synchrony. Both Turing instability analysis and numerical continuation software are used to explore the existence and stability of spatio-temporal patterns in the system. In particular, we show that this new model can support states above and beyond those seen in a standard neural field model. These states are typified by structures within bumps and waves showing the dynamic evolution of population synchrony. I. INTRODUCTION The act of passing information between brain regions produces waves of neural activity. These waves are readily observed using non-invasive techniques such as electroencephalography (EEG) and magnetoencephalography (MEG) [1], as well as in brain slices [29]. Both experimental and theoretical work has shown that EEG/MEG recordings and evoked potentials can exhibit travelling and standing waves [38]. In particular travelling waves are seen in EEG sleep recordings propagating across the cortex at a speed of about 1.2-7.0 m/s [35]. Standing waves are often associated with idle brain states. For example, standing waves at α frequency (8-13 Hz) are observed in the vicinity of the visual cortex when the subject has their eyes closed [8]. Another commonly observed spatial pattern is the so called bump attractor. This spatially localised increase in population firing is produced in working memory tasks and the location of the bump can be linked to memory location [44]. Traditionally, neural field models are used to describe ∗ [email protected] wave and bump states in the brain. Although inspired by biology, these models are entirely phenomenological in nature. Even so, they have been particularly successful in describing neurophysiological phenomena, such as EEG/MEG rhythms [45], working memory [33], binocular rivalry [12] and orientation tuning in the visual cortex [7]. They are typically cast as a system of non-local differential equations which describe the spatiotemporal evolution of coarse grained population variables, such as the firing rate of a neuronal population, the average synaptic current, or the mean membrane potential [21]. The first attempt at a neural field mode is attributed to Beurle [9], who built a model to describe the propagation of activation in a given volume of neural tissue. This model was purely excitatory, but even so allowed him to examine the propagation of large scale brain activity. In the 1970s Wilson and Cowan [42, 43] extended this model to include a second inhibitory layer. Unlike Beurle, they were interested in spatially localised bump solutions. In his seminal paper, Amari [2, 3] created what is now known as the standard neural field equation. By introducing a Mexican hat type coupling function (local excitation and long range inhibition), he reduced the model to a single equation with a mixture of excitatory and inhibitory connections. This allowed him to con- 2 struct explicit solutions for spatially localised patterns, and assess their stability (at least for a Heaviside firing rate function). For a review of the Amari model and bumps in one spatial dimension see [17] and for a discussion of bubbles in two spatial dimensions see [11]. One of the main assumptions in many types of neural field model, and especially those for describing EEG, is that point-wise they describe a density of neurons operating in a near synchronous regime [37]. This wholly reasonable assumption can be traced back to the observation that an EEG scalp electrode, which typically experiences the activity of roughly 109 cortical pyramidal cells, can only detect an electric field if all the individual cell dipoles add coherently [22]. However, a near synchrony assumption in any neural mass model precludes its use in describing the increase and decrease of power commonly seen in given EEG/MEG frequency bands. These temporal variations are believed to be the result of changes in synchrony within the neural tissue. The former phenomenon is called event-related synchronisation (ERS), and the latter event-related desynchronisation (ERD) [40]. Consequently, there is a pressing need to develop the next generation of neural field models, which include this notion of a dynamic within-population synchrony (not fixed to be near synchronous), to more accurately describe the evolution of large scale spatiotemporal brain rhythms. When looking at within population synchronisation one typically uses a spiking neural network model. However, these high dimensional models are almost impossible to gain insight from. In an ideal world there would be a mathematical procedure for linking microscopic dynamics to macroscopic dynamics. This link has proved elusive for the majority of spiking models. However, Luke et al. [34] showed that the θ-neuron model is amenable to such a reduction for pulsatile coupling. Montbri´ et al. o [36] have used a similar approach to reduce a network of quadratic integrate-and-fire neurons. Laing [32] has also shown that the same approach can be applied to a network of spatially extended θ-neurons, in the presence of gap junction coupling. In previous papers [14, 19], we have shown that the approach of Luke et al. can be extended to incorporate a biologically realistic form of synaptic coupling. Here we build on this work to construct a neural field model that incorporates within population synchrony and a realistic form of synaptic coupling. We shall refer to this model as a next generation neural field model. In this paper we show, using a mixture of analysis and simulation, that this new neural field model can support exotic patterned states more reminiscent of high dimensional spiking networks, with spatiotemporal patterns showing the evolution of synchrony. We begin with an overview of the model formulation in §II and outline the necessary steps for reduction to a neural field. A Turing instability analysis of the model is covered in §III. Here, we show that the system can be unstable to both static and dynamic Turing patterns for a wide window of parameter space. More interestingly, we show that when the Turing bifurcation collides with a Hopf bifurcation patterned states emerge in which there exists an oscillating structure within a spatially localised bump. The Turing analysis is complimented with a numerical bifurcation analysis for the full nonlinear model in §IV. Here we further examine the emergent patterns away from bifurcation, as well as consider localised travelling waves. Finally, in §V we discuss our main results as well as natural extensions of the work presented. II. THE MODEL We first consider a network of N coupled quadratic integrate-and-fire (QIF) neurons, uniformly distributed along a line of length Λ such that the jth neuron is at position xj = −Λ/2 + (j − 1)∆x, where j = 1, . . . , N and ∆x = Λ/(N −!) is the spacing between neurons. The coupling between neuron i and neuron j depends m only upon the distance between the two neurons, wij = wm (xi − xj ), where m is a label used to keep track of neural subpopulations. We shall focus on the choice Λ 1 (so that we deal with large spatial scale systems). We write the network dynamics for the voltage, vi ≡ v(xi , t), in the form 2 vi = vi + η i + I i , i = 1, . . . , N, (1) subject to reset, vi → vr , whenever a firing threshold vth is reached by vi . The time at which the ith cell reaches threshold from below for the sth time will be denoted by Tis , s ∈ N. Here the background drives ηi will be assumed to be heterogeneous and chosen from a Lorentzian distribution, L(η) = 1 ∆ , π (η − η0 )2 + ∆2 (2) where η0 is the centre of the distribution and ∆ is the half width. We assume a synaptic input current of the form Ii (t) = m i m gm (t)(vsyn − vi (t)), (3) i for a global conductance gm , synaptic reversal potential m vsyn and local voltage vi . In what follow we will assume m = {1, 2} or m = 1, but one should note that this framework can be extended to include multiple types of synapses. The interplay of excitation and inhibition, on different spatial scales, is known to play an important role in the generation of global spatially patterned states [13, 18, 26]. Hence, we separate our synaptic conductance into two parts, such that we have one excitatory synaptic current (vsyn > 0) and one inhibitory (vsyn < 0), each with different spatial ranges. Each of the synaptic conductances will be taken to mimic that of a synapse with a finite rise and fall time 3 The mean field representation of the synaptic inputs gm are written as follows, that evolves according to κm = N i Qm gm N s∈N j=1 m wij δ(t − Tjs ), (4) Qm gm (x, t) = κm π ∞ l∈Z Qm = 2 . (5) Here τm is the synaptic time scale, and Qm has a re−2 sponse (Green's function) s(t) = τm te−t/τm for t ≥ 0 (and is zero otherwise), which is a popular choice for many synapse models in computational neuroscience [28]. It is well known that the QIF model is formally equivalent to the θ-neuron model [27] under the transformation vi = tan(θi /2), for θi ∈ [0, 2π) (when the threshold vth and reset vr are set to +∞ and −∞, respectively). This relationship allows us to construct a θ-neuron network dynamics as θi = 1 − cos θi + (1 + cos θi )ηi + m i Qm gm = 2κm N i m gm [(1 + cos θi )vsyn − sin θi ], (6) N j=1 m wij δ(θj − π). (7) To obtain (7) we have made use of the fact that δ(t − Tjs ) = δ(θj − π)θj (Tjm ). As such, we say that neuron j "spikes" whenever θj increases through π. The network formulation in terms of dynamics on a circle is particularly useful since we no longer have to worry about handling the discontinuous reset process as we would have to do for a QIF network. A. 0 × wm (x − y)eil(θ−π) , We take the large N limit, N → ∞, which allows us to describe the system in terms of a continuous probability distribution function ρ(x, η, θ, t), with x, η ∈ R, θ ∈ [0, 2π), and t ∈ R+ , which satisfies the continuity equation: (8) where vθ is the following realisation of (6), + m (10) where we have used the result 2πδ(θ − π) = l∈Z eil(θ−π) to write the right hand side in terms of exponentials. The formula for vθ (9) may be conveniently written in terms of e±iθ as vθ = φ(eiθ + e−iθ ) + χ, (11) where φ = (η + I − 1)/2 and χ = η + I + 1. Note that a similar approach has previously been considered by Laing [32]. However, his focus was on smooth (non-pulsatile) interactions with a first order model of the synapse, and he did not consider reversal potentials. To reduce the system we make use of the OttAntonsen (OA) ansatz [39]. This decomposes ρ in the form ρ(x, η, θ, t) = L(η)F (x, η, θ, t)/(2π), where F is 2π-periodic in θ with a Fourier series representation inθ F (x, η, θ, t) = . The OA ansatz ren Fn (x, η, t)e stricts the choice of Fn such that Fn (x, η, t) = α(x, η, t)n , with α(x, η, t) < 1 to ensure convergence. Hence, ρ can be written as L(η) ρ(x, η, θ, t) = 2π 1+ ∞ α(x, η, t)n einθ + cc , n=1 (12) where cc denotes the complex conjugate. Substituting (11) into the continuity equation (8) and balancing terms in eiθ gives an evolution equation for α: ∂ α − iα2 φ − iαχ − iφ = 0. ∂t 2π vθ = 1 − cos θ + (1 + cos θ)η dηρ(y, η, θ, t) −∞ (13) We define the Kuramoto order parameter as follows, Mean field limit ∂ ∂ ρ+ (ρvθ ) = 0, ∂t ∂θ dθ −∞ for some coupling strength κm , where Q is the linear 2nd order differential operator d 1 + τm dt ∞ 2π dy m gm [(1 + cos θ)vsyn − sin θ]. (9) z(x, t) = dθ ∞ dηρ(x, η, θ, t)eiθ , (14) −∞ 0 where z ≤ 1. The Kuramoto order parameter is a complex number z = ReiΨ , whose magnitude R represents the degree of within population synchrony and angle Ψ represents the average phase of the population. Substituting (12) into (14) we find that z (x, t) = ¯ ∞ dηL(η)α(x, η, t), (15) −∞ where z denotes the complex conjugate of z. As the ¯ Lorentzian has two simple poles η± = η0 ± i∆, the above integral may be performed by choosing a large semi-circular contour in the lower half η-plane and using 4 <latexit <latexit sha1_base64="UWW1l0c9YBbgdn9m09FAFdrzXfc=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q3KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcZeMlw==</latexit> 1 10 As in [19], we interpret (17) as describing the intrinsic population dynamics and (18) as the dynamics generated by synaptic coupling. In summary, we have the following evolution equations 2 <latexit sha1_base64="sptBfbGV2GFJTEj2vb/wWYtsBPk=">AAAB63icbVBNSwMxEJ31s9avqkcvwSp4Ktki6LHgxWMF+wHtWrJptg1NskuSFcrSv+DFgyJe/UPe/Ddm2z1o64OBx3szzMwLE8GNxfjbW1vf2NzaLu2Ud/f2Dw4rR8dtE6eashaNRay7ITFMcMVallvBuolmRIaCdcLJbe53npg2PFYPdpqwQJKR4hGnxOaSjx/rg0oV1/AcaJX4BalCgeag8tUfxjSVTFkqiDE9Hyc2yIi2nAo2K/dTwxJCJ2TEeo4qIpkJsvmtM3ThlCGKYu1KWTRXf09kRBozlaHrlMSOzbKXi/95vdRGN0HGVZJapuhiUZQKZGOUP46GXDNqxdQRQjV3tyI6JppQ6+IpuxD85ZdXSbte83HNv7+qNs6LOEpwCmdwCT5cQwPuoAktoDCGZ3iFN096L96797FoXfOKmRP4A+/zBwS4jXU=</latexit> R sin ∂t z = F(z) + <latexit sha1_base64="m19otoNAgPMj7waI7Dqsu1co1Ok=">AAAB8nicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cq9gOSUDbbTbt0swm7E6GU/gwvHhTx6q/x5r9x2+agrQ8GHu/NMDMvyqQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZprxlsslanuRtRwKRRvoUDJu5nmNIkk70Sj25nfeeLaiFQ94jjjYUIHSsSCUbSS/0ACIxQJmkb0qjW37s5BVolXkBoUaPaqX0E/ZXnCFTJJjfE9N8NwQjUKJvm0EuSGZ5SN6ID7liqacBNO5idPyblV+iROtS2FZK7+npjQxJhxEtnOhOLQLHsz8T/PzzG+CSdCZTlyxRaL4lwSTMnsf9IXmjOUY0so08LeStiQasrQplSxIXjLL6+S9mXdc+ve/VWtcVbEUYYTOIUL8OAaGnAHTWgBgxSe4RXeHHRenHfnY9FacoqZY/gD5/MHWZqQig==</latexit> <latexit sha1_base64="1xJWpPWoRNA2QObEozJYgFCqSF4=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipGQ7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AweuMzA==</latexit> 5 <latexit sha1_base64="LgCNgkqX27pODHNeA0cSMuNsMYA=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LFbBiyURRY8FLx4r2A9oY9lsJ+3SzSbsboQS+iO8eFDEq7/Hm//GbZuDtj4YeLw3w8y8IBFcG9f9dgorq2vrG8XN0tb2zu5eef+gqeNUMWywWMSqHVCNgktsGG4EthOFNAoEtoLR7dRvPaHSPJYPZpygH9GB5CFn1Fip5bmP2fnVpFeuuFV3BrJMvJxUIEe9V/7q9mOWRigNE1Trjucmxs+oMpwJnJS6qcaEshEdYMdSSSPUfjY7d0JOrdInYaxsSUNm6u+JjEZaj6PAdkbUDPWiNxX/8zqpCW/8jMskNSjZfFGYCmJiMv2d9LlCZsTYEsoUt7cSNqSKMmMTKtkQvMWXl0nzouq5Ve/+slI7yeMowhEcwxl4cA01uIM6NIDBCJ7hFd6cxHlx3p2PeWvByWcO4Q+czx84d467</latexit> <latexit sha1_base64="8zBnKX5YZCEjL5qhjzIjEp+gFBw=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw6XXL1fcqjsHWSVeTiqQo9Evf/UGMUsjlIYJqnXXcxPjZ1QZzgROS71UY0LZmA6xa6mkEWo/m186JedWGZAwVrakIXP190RGI60nUWA7I2pGetmbif953dSEN37GZZIalGyxKEwFMTGZvU0GXCEzYmIJZYrbWwkbUUWZseGUbAje8surpFWrem7Vu7+q1M/yOIpwAqdwAR5cQx3uoAFNYBDCM7zCmzN2Xpx352PRWnDymWP4A+fzB9rcjM4=</latexit> 1 <latexit sha1_base64="8zBnKX5YZCEjL5qhjzIjEp+gFBw=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw6XXL1fcqjsHWSVeTiqQo9Evf/UGMUsjlIYJqnXXcxPjZ1QZzgROS71UY0LZmA6xa6mkEWo/m186JedWGZAwVrakIXP190RGI60nUWA7I2pGetmbif953dSEN37GZZIalGyxKEwFMTGZvU0GXCEzYmIJZYrbWwkbUUWZseGUbAje8surpFWrem7Vu7+q1M/yOIpwAqdwAR5cQx3uoAFNYBDCM7zCmzN2Xpx352PRWnDymWP4A+fzB9rcjM4=</latexit> 1 <latexit sha1_base64="qHsPOeQzl88Jqob1MWcZmyG8L/w=">AAAB8nicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cq9gOSUDbbTbt0kw27E6GU/gwvHhTx6q/x5r9x2+agrQ8GHu/NMDMvyqQw6LrfTmltfWNzq7xd2dnd2z+oHh61jco14y2mpNLdiBouRcpbKFDybqY5TSLJO9HoduZ3nrg2QqWPOM54mNBBKmLBKFrJfyABU4YETSN61Zpbd+cgq8QrSA0KNHvVr6CvWJ7wFJmkxviem2E4oRoFk3xaCXLDM8pGdMB9S1OacBNO5idPyblV+iRW2laKZK7+npjQxJhxEtnOhOLQLHsz8T/PzzG+CScizXLkKVssinNJUJHZ/6QvNGcox5ZQpoW9lbAh1ZShTaliQ/CWX14l7cu659a9+6ta46yIowwncAoX4ME1NOAOmtACBgqe4RXeHHRenHfnY9FacoqZY/gD5/MHUdOQhQ==</latexit> R cos <latexit <latexit sha1_base64="UWW1l0c9YBbgdn9m09FAFdrzXfc=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q3KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcZeMlw==</latexit> 1 FIG. 1. Firing rate dynamics: Density plot showing the firing rate f as a function of the complex Kuramoto order parameter z = ReiΨ . Firing is highest near z = eiπ . This corresponds to highly synchronous behaviour where all of the phases of the neurons go through π simultaneously. the residue theorem, to yield z (x, t) = α(x, η− , t). ¯ We use (14) to write (10) as Qm gm (x, t) = κm ∞ −∞ dywm (x − y)f (z(y, t)), (16) ≡ κm wm ⊗ f (z), where ⊗ represents a spatial convolution and f is the population firing rate 1 f (z) = π 1+ ∞ Gm (z, gm ), (1 + τm ∂t )2 gm = κm wm ⊗ f (z), f 10 m l=1 (20) m where Gm (z, gm ) ≡ G(z, gm ; vsyn ) and we have omitted the dependence on control parameters. The dynamical system (19) -- (20) has therefore one complex (z) and m real (gm ) state variables. In what follows, we choose normalised exponentially-decaying synaptic kernels of the form wm (x) = exp(−βm x)/(2βm ), with Fourier transform wm (k) = 1/(1+(k/βm )2 ). As the Fourier transform is a rational function, the evolution equations (19) -- (20) are equivalent to the PDEs [20] ∂t z = F(z) + m Gm (z, gm ), 2 (1 − ∂xx /βm )(1 + τm ∂t )2 gm = κm f (z). (21) (22) In what follows, we will show that the model can support Turing patterns, travelling waves, and other complex global spatio-temporal patterns, for m = {1, 2} and travelling fronts for m = 1. In the former case, this is achieved by choosing β1 > β2 such that w1 has a shorter 1 2 spatial scale than w2 . For vsyn > 0 and vsyn < 0 this has the overall effect of short range excitation and long 1 2 range inhibition, and for vsyn < 0 and vsyn > 0 inhibition dominates at short distances and excitation at longer distances. III. (−1)l z l + cc (19) INSTABILITY OF THE HOMOGENEOUS STEADY STATE 2 1 1 − z , = π 1 + z + z + z2 z < 1. Note that (16) takes the form of a generalised neural field equation, where the firing rate function f is a derived quantity that depends on the within population synchrony. Also noteworthy is the fact that this firing rate function is not a sigmoid. It is a highly non-linear function that depends on the intrinsic population dynamics. The firing rate can be plotted as a function of the Kuramoto order parameter (Fig. 1). As expected, the firing rate is highest at z = eiπ (where a singe neuron fires as θ increases through π). The dynamics of z are obtained by evaluating (13) at η− = η0 − i∆, and taking the complex conjugate, which m gives ∂z/∂t = F(z; η0 , ∆) + m G(z, gm ; vsyn ), where (z − 1)2 (z + 1)2 F(z; η0 , ∆) = −i + [iη0 − ∆] , 2 2 (z + 1)2 (z 2 − 1) G(z, g; vsyn ) = g i vsyn − . 2 2 (17) (18) We require excitation and inhibition to generate Turing patterns, hence we let m = {1, 2} in this section. We will also assume β1 = 1 (without loss of generality) and define β = β2 < 1 as the parameter which measures the difference in the spatial scales of w1 and w2 . We first recast the system (21) -- (22) as a first-order evolution equation in the state variables u = (a, b, K1 , g1 , K2 , g2 ), where a = Re(z), b = Im(z) and Km = (1 + τm ∂/∂t)gm , and seek stationary and spatially homogeneous states u(x, t) = u∗ for all x and t. We apply a small perturbation of the form u(x, t) = u∗ + u(x, t), where u(x, t) = Aeλt eikx , λ ∈ C, k ∈ R and A ∈ C6 . Using the following identity wm (x) ⊗ eikx = wm (k)eikx , we obtain, to leading order, ∂t u(x, t) = J (k)u(x, t), where J is the following (k-dependent) Jacobian, evaluated at u∗ , J (k) = J11 J12 , J21 (k) J22 5 where we highlight the k-dependence  −1  −1 τ1 κ1 w1 (k)∂a f τ1 κ1 w1 (k)∂b f     0 0  . J21 (k) =  −1  −1 τ2 κ2 w2 (k)∂a f τ2 κ2 w2 (k)∂b f  0 0 and show the other elements in Appendix A. The complex eigenvalues λ = ν + iω satisfies the characteristic equation E(λ, k) = det J (k) − λI4 = 0, (23) where I4 is the 4 × 4 identity matrix. A homogeneous steady state u∗ is linearly stable to perturbations eikx if ν(k) < 0 for all k. By the implicit function theorem, a branch of solutions λ(k) to (23) touches the imaginary axis when ∂k M∂ω N − ∂ω M∂k N = 0, (24) where M = Re(E) and N = Im(E). We observe a Hopf bifurcation of the spatially uniform state u∗ if (23) holds for ν(0) = 0 and k = 0, i.e. if there is a non-zero solution of: ω 6 + p4 ω 4 + p2 ω 2 + p0 = 0, 4 2 p5 ω + p3 ω + p1 = 0. (25) (26) Here, pi are scalars which depend on u∗ and the control parameters of the problem (see Appendix B). Solving (25) -- (26), for fixed width of the Lorentzian ∆, synaptic coupling strengths κ1 , κ2 , synaptic time constants τ1 , τ2 , and relative width of the synaptic kernel β, gives the locus of Hopf bifurcations as a function of the synaptic reversal potential vsyn and background drive η0 (Fig 2(a) green curve). We fix the synaptic reversal potentials to 2 1 be equal and opposite, and define vsyn = vsyn = −vsyn . For simplicity we set κ1 = κ2 and τ1 = τ2 . For this choice of parameters, excitation dominates for short range interactions for vsyn > 0 and inhibition dominates for short range interactions for vsyn < 0. Note, also, that the Hopf bifurcation occurs for the same value of η0 for all vsyn . This is a consequence of the choice of equal coupling strengths and time constants. If this balance is disrupted the Hopf bifurcation will depend upon vsyn and for some parameter choices we see several Hopf curves in the (vsyn , η0 )-plane. The homogeneous steady state u∗ undergoes a static Turing bifurcation if there exists a non-zero critical wavenumber kc such that (23) and (24) hold for ν(kc ) = 0, ω(kc ) = 0. This leads to the conditions 1 2 p0 + q0 w1 + q0 w1 = 0, 1 q0 dw1 2 dw2 + q0 dkc dkc 1 2 p1 + q1 w1 + q1 w2 = 0, (27) (28) where wi and its derivative depend on kc and β, and the j scalars pi , qi depend on u∗ and the control parameters of the problem (see Appendix B). As with the Hopf bifurcation, the locus of Turing bifurcations can be plotted as a function of the synaptic reversal potential vsyn and background drive η0 (Fig. 2 red curve). The curve exhibits a turning point, therefore the system has two Turing bifurcations for sufficiently large values of vsyn . If we fix vsyn at a value in this region, and ascend the bifurcation diagram by increasing η0 , the spectrum λ(k) touches the imaginary axis of the (ν, ω)plane with wave number kc (lower Turing bifurcation) (Fig. 2(c)) and continues to move to the right, implying that u∗ is unstable to a range of perturbations with wavenumbers k ∈ (k1 , k2 ). As η0 is further increased, the spectrum returns to the left-hand plane (upper Turing bifurcation), which restores the stability of u∗ . Note the value of kc is not necessarily the same on the upper and lower branches. The scenario described above is robust to perturbations in other parameters: we found that changes in τ1 , τ1 and ∆ do not significantly affect the location of the Turing bifurcations. Increasing κ2 results in an increase of the gap between the two critical values of η0 , and hence a larger window of instability, while increasing κ1 decreases this gap and at a certain point the shape is inverted, such that the unstable region lies predominantly in the η0 < 0 region of the plot. As κ1 corresponds to the synaptic strength for the excitatory current for vsyn > 0, when it becomes significantly larger than κ2 excitation dominants even for long range interactions. As patterns arise from the interplay of excitation and inhibition, an inhibitory external drive (η0 < 0) is needed to observe Turing patterns in this regime. A Turing-Hopf instability of the homogeneous steady state occurs if there exists a non-zero wave number kc such that (23) and (24) hold with ν(kc ) = 0 and ω(kc ) = ±ωc = 0. This instability, which we will also refer to as dynamic Turing bifurcation, elicits wavetrains with wavenumber kc and phase velocity ωc /kc (near bifurcation). From (23) and (24) we obtain a system of the form ω 6 + p4 ω 4 + P2 ω 2 + P0 = 0, 4 2 p5 ω + P3 ω + P1 = 0, (29) (30) dw1 dw2 + Q2 (−6ω 5 + 4p4 ω 3 − 2P2 ω) 2 dk dk dw1 dw2 − Q1 + Q2 (5p5 ω 4 − 3P3 ω 2 + P1 ) = 0, 3 3 dk dk (31) Q1 2 j j 1 2 where Pi = pi + qi w1 (k) + qi w1 (k), Qj = qi ω i + qi−2 ω i−2 i j and pi , qi are scalars which depend on the control parameters (see Appendix B). In passing, we note that the characteristic equation now has an imaginary part, resulting in the two conditions (29), (30). Solving (29) -- (31) allows us to plot the locus of the Turing-Hopf bifurcations in the (vsyn , η0 )-plane (Fig. 2 blue curve), together with the Hopf and Turing bifurcation curves. Note that the Turing-Hopf curve intersects with the Hopf curve at vsyn ≈ 8. The value of kc deceases along the Turing-Hopf 6 (a) (b) <latexit sha1_base64="m8TaxYBegNcoWnev7K9hjlGCE6E=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFe7SaBmwMV0U8wHJEfY2c8mSvb1jd08IR/6BjYUitv4jO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7AST27nfeUKleSwfzTRBP6IjyUPOqLHSQ6MxKFfcqrsAWSdeTiqQozkof/WHMUsjlIYJqnXPcxPjZ1QZzgTOSv1UY0LZhI6wZ6mkEWo/W1w6I5dWGZIwVrakIQv190RGI62nUWA7I2rGetWbi/95vdSEN37GZZIalGy5KEwFMTGZv02GXCEzYmoJZYrbWwkbU0WZseGUbAje6svrpF2rem7Vu69V6hd5HEU4g3O4Ag+uoQ530IQWMAjhGV7hzZk4L86787FsLTj5zCn8gfP5Ayk3jQA=</latexit> 15 (c) II <latexit sha1_base64="+7RqAHpOFdrqnrA8oCJg4Cxbo2U=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgFe7SaBmwMV1E8wHJEfY2c8mSvb1jd08IR36CjYUitv4iO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7AST27nfeUKleSwfzTRBP6IjyUPOqLHSQ6PRGJQrbtVdgKwTLycVyNEclL/6w5ilEUrDBNW657mJ8TOqDGcCZ6V+qjGhbEJH2LNU0gi1ny1OnZFLqwxJGCtb0pCF+nsio5HW0yiwnRE1Y73qzcX/vF5qwhs/4zJJDUq2XBSmgpiYzP8mQ66QGTG1hDLF7a2EjamizNh0SjYEb/XlddKuVT236t3XKvWLPI4inME5XIEH11CHO2hCCxiM4Ble4c0Rzovz7nwsWwtOPnMKf+B8/gC9U41T</latexit> 15 (d) III <latexit sha1_base64="dTkrA75j8FBJr/7r2qvUwf7yX8M=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFe7SxDJgo10U8wHJEfY2e8mSvb1jd04IR/6BjYUitv4jO/+Nm+QKTXww8Hhvhpl5QSKFQdf9dgobm1vbO8Xd0t7+weFR+fikbeJUM95isYx1N6CGS6F4CwVK3k00p1EgeSeY3Mz9zhPXRsTqEacJ9yM6UiIUjKKVHu7ag3LFrboLkHXi5aQCOZqD8ld/GLM04gqZpMb0PDdBP6MaBZN8VuqnhieUTeiI9yxVNOLGzxaXzsilVYYkjLUthWSh/p7IaGTMNApsZ0RxbFa9ufif10sxvPYzoZIUuWLLRWEqCcZk/jYZCs0ZyqkllGlhbyVsTDVlaMMp2RC81ZfXSbtW9dyqd1+rNC7yOIpwBudwBR7UoQG30IQWMAjhGV7hzZk4L86787FsLTj5zCn8gfP5AzzrjQ0=</latexit> 15 <latexit sha1_base64="YeR0z88bmpVpcF7lqgeTsszmBoI=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoseCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6xEnC/YgOlQgFo2ilB++qX664VXcOskq8nFQgR6Nf/uoNYpZGXCGT1Jiu5yboZ1SjYJJPS73U8ISyMR3yrqWKRtz42fzSKTm3yoCEsbalkMzV3xMZjYyZRIHtjCiOzLI3E//zuimGN34mVJIiV2yxKEwlwZjM3iYDoTlDObGEMi3srYSNqKYMbTglG4K3/PIqaV1WPbfq3dcq9bM8jiKcwClcgAfXUIc7aEATGITwDK/w5oydF+fd+Vi0Fpx85hj+wPn8AecAjNY=</latexit> (e) IV <latexit <latexit sha1_base64="YbJPrs2Cp7Q0HJlgy4CzfIkbwco=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzU6gzLFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VatUrjKo+jCBdwCTfgQR0acA9NaAMDhGd4hTfn0Xlx3p2PVWvByWfO4Q+czx+pC4y6</latexit> 15 <latexit sha1_base64="YeR0z88bmpVpcF7lqgeTsszmBoI=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoseCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6xEnC/YgOlQgFo2ilB++qX664VXcOskq8nFQgR6Nf/uoNYpZGXCGT1Jiu5yboZ1SjYJJPS73U8ISyMR3yrqWKRtz42fzSKTm3yoCEsbalkMzV3xMZjYyZRIHtjCiOzLI3E//zuimGN34mVJIiV2yxKEwlwZjM3iYDoTlDObGEMi3srYSNqKYMbTglG4K3/PIqaV1WPbfq3dcq9bM8jiKcwClcgAfXUIc7aEATGITwDK/w5oydF+fd+Vi0Fpx85hj+wPn8AecAjNY=</latexit> V <latexit sha1_base64="YeR0z88bmpVpcF7lqgeTsszmBoI=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoseCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6xEnC/YgOlQgFo2ilB++qX664VXcOskq8nFQgR6Nf/uoNYpZGXCGT1Jiu5yboZ1SjYJJPS73U8ISyMR3yrqWKRtz42fzSKTm3yoCEsbalkMzV3xMZjYyZRIHtjCiOzLI3E//zuimGN34mVJIiV2yxKEwlwZjM3iYDoTlDObGEMi3srYSNqKYMbTglG4K3/PIqaV1WPbfq3dcq9bM8jiKcwClcgAfXUIc7aEATGITwDK/w5oydF+fd+Vi0Fpx85hj+wPn8AecAjNY=</latexit> Hop Tur ng Hop Tur ng 30 t 20 0 t ⌘0 20 x 20 0 85 <latexit sha1_base64="QemqZr6fsruN4EG6fJtLmwgab9A=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXTje4suHFZxV6gHUomzbShSWZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wwEdxYjL+9ldW19Y3N3FZ+e2d3b79wcFg3caopq9FYxLoZEsMEV6xmuRWsmWhGZChYIxxcT/LGA9OGx+reDhMWSNJTPOKUWGfdlXGnUMQlPBVaBn8OxavP8fgdAKqdwle7G9NUMmWpIMa0fJzYICPacirYKN9ODUsIHZAeazlURDITZNNJR+jMOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44eXcEf3HlZaiXSz4u+be4WDmFmXJwDCdwDj5cQAVuoAo1oBDBIzzDizfwnrxX721WuuLNe47gj7yPH46fj4g=</latexit> sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiB sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiBU7gADy6hBrdQhwZQCOEJXuDVGTnPzpsznZcWnEXPMfyR8/ED69yRTQ==</latexit> t <latexit <latexit sha1_base64="coXTjA2n3M9S92rki2twuEpV/uI=">AAAB6HicbVBNS8NAEJ34WetX1aOXYBU8lUQEPRa8eGzBfkAbymY7adduNmF3IpTSX+DFgyJe/Une/Ddu2xy09cHA470ZZuaFqRSGPO/bWVvf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmGDBElspxpZHEpshaO7md96Qm1Eoh5onGIQs4ESkeCMrFSnXqnsVbw53FXi56QMOWq90le3n/AsRkVcMmM6vpdSMGGaBJc4LXYzgynjIzbAjqWKxWiCyfzQqXthlb4bJdqWIneu/p6YsNiYcRzazpjR0Cx7M/E/r5NRdBtMhEozQsUXi6JMupS4s6/dvtDISY4tYVwLe6vLh0wzTjabog3BX355lTSvKr5X8evX5ep5HkcBTuEMLsGHG6jCPdSgARwQnuEV3pxH58V5dz4WrWtOPnMCf+B8/gDXI4za</latexit> R 0 0 95 <latexit <latexit sha1_base64="sYO7+g84YMjnL0FPL2Qemc7g+Xw=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcBOMlg==</latexit> t <latexit <latexit sha1_base64="coXTjA2n3M9S92rki2twuEpV/uI=">AAAB6HicbVBNS8NAEJ34WetX1aOXYBU8lUQEPRa8eGzBfkAbymY7adduNmF3IpTSX+DFgyJe/Une/Ddu2xy09cHA470ZZuaFqRSGPO/bWVvf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmGDBElspxpZHEpshaO7md96Qm1Eoh5onGIQs4ESkeCMrFSnXqnsVbw53FXi56QMOWq90le3n/AsRkVcMmM6vpdSMGGaBJc4LXYzgynjIzbAjqWKxWiCyfzQqXthlb4bJdqWIneu/p6YsNiYcRzazpjR0Cx7M/E/r5NRdBtMhEozQsUXi6JMupS4s6/dvtDISY4tYVwLe6vLh0wzTjabog3BX355lTSvKr5X8evX5ep5HkcBTuEMLsGHG6jCPdSgARwQnuEV3pxH58V5dz4WrWtOPnMCf+B8/gDXI4za</latexit> <latexit sha1_base64="0Yq5TIqZy8gmjXAagVPJ0JaLsro=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw2XN7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZJzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwhs/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdKqVT236t1fVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFKdI0J</latexit> 20 0 60 x 20 R <latexit <latexit sha1_base64="zlFPBJwQd38NXQhvuYcrr8P1bxk=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cW7Ae0oWy2k3btZhN2N2IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52Sc6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasIbP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSuqx6btVrXFVqZ3kcRTiBU7gAD66hBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB3TOM3g==</latexit> 0 95 <latexit sha1_base64="mFoNdpVbpfRu9/IYpphVUlYZwtU=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lKQY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzV3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17uuVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AODxjNI=</latexit> <latexit <latexit sha1_base64="sYO7+g84YMjnL0FPL2Qemc7g+Xw=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcBOMlg==</latexit> 0 <latexit sha1_base64="0Yq5TIqZy8gmjXAagVPJ0JaLsro=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw2XN7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZJzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwhs/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdKqVT236t1fVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFKdI0J</latexit> <latexit <latexit sha1_base64="coXTjA2n3M9S92rki2twuEpV/uI=">AAAB6HicbVBNS8NAEJ34WetX1aOXYBU8lUQEPRa8eGzBfkAbymY7adduNmF3IpTSX+DFgyJe/Une/Ddu2xy09cHA470ZZuaFqRSGPO/bWVvf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmGDBElspxpZHEpshaO7md96Qm1Eoh5onGIQs4ESkeCMrFSnXqnsVbw53FXi56QMOWq90le3n/AsRkVcMmM6vpdSMGGaBJc4LXYzgynjIzbAjqWKxWiCyfzQqXthlb4bJdqWIneu/p6YsNiYcRzazpjR0Cx7M/E/r5NRdBtMhEozQsUXi6JMupS4s6/dvtDISY4tYVwLe6vLh0wzTjabog3BX355lTSvKr5X8evX5ep5HkcBTuEMLsGHG6jCPdSgARwQnuEV3pxH58V5dz4WrWtOPnMCf+B8/gDXI4za</latexit> 20 0 91 <latexit <latexit sha1_base64="zlFPBJwQd38NXQhvuYcrr8P1bxk=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cW7Ae0oWy2k3btZhN2N2IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52Sc6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasIbP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSuqx6btVrXFVqZ3kcRTiBU7gAD66hBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB3TOM3g==</latexit> x 20 R 0 99 <latexit sha1_base64="mFoNdpVbpfRu9/IYpphVUlYZwtU=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lKQY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzV3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17uuVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AODxjNI=</latexit> <latexit <latexit sha1_base64="sYO7+g84YMjnL0FPL2Qemc7g+Xw=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcBOMlg==</latexit> 0 <latexit sha1_base64="0Yq5TIqZy8gmjXAagVPJ0JaLsro=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw2XN7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZJzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwhs/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdKqVT236t1fVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFKdI0J</latexit> 20 0 55 <latexit <latexit sha1_base64="zlFPBJwQd38NXQhvuYcrr8P1bxk=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cW7Ae0oWy2k3btZhN2N2IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52Sc6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasIbP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSuqx6btVrXFVqZ3kcRTiBU7gAD66hBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB3TOM3g==</latexit> x 20 R 0 75 <latexit sha1_base64="mFoNdpVbpfRu9/IYpphVUlYZwtU=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lKQY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzV3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17uuVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AODxjNI=</latexit> 10 <latexit sha1_base64="3t+4UbZpWdk262G+bJ4/hq+8w+E=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZN7qz4MZlFXuBdiiZNNOGJpkhyQhl6Bu4caFYV4IrX8edb2N6WWjrD4GP/z+HnHPCRHBjMf72ciura+sb+c3C1vbO7l5x/6Bu4lRTVqOxiHUzJIYJrljNcitYM9GMyFCwRji4nuSNB6YNj9W9HSYskKSneMQpsc6683GnWMJlPBVaBn8OpavP8fgdAKqd4le7G9NUMmWpIMa0fJzYICPacirYqNBODUsIHZAeazlURDITZNNJR+jUOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44BXcEf3HlZaifl31c9m9xqXICM+XhCI7hDHy4gArcQBVqQCGCR3iGF2/gPXmv3tusNOfNew7hj7yPH40aj4c=</latexit> sha1_base64="TBz0KWW1Y5FVHwrZZOy0b06gZmg=">AAAB6XicbVA9TwJBEJ3DL8Qv1NJmI5pYkT0bLUlsLNEIksCF7C17sGFv77I7Z0Iu/AMbC42x9R/Z+W9c4AoFXzLJy3szmZkXpkpapPTbK62tb2xulbcrO7t7+wfVw6O2TTLDRYsnKjGdkFmhpBYtlKhEJzWCxaESj+H4ZuY/PgljZaIfcJKKIGZDLSPJGTrp3qf9ao3W6RxklfgFqUGBZr/61RskPIuFRq6YtV2fphjkzKDkSkwrvcyKlPExG4quo5rFwgb5/NIpOXfKgESJcaWRzNXfEzmLrZ3EoeuMGY7ssjcT//O6GUbXQS51mqHQfLEoyhTBhMzeJgNpBEc1cYRxI92thI+YYRxdOBUXgr/88ippX9Z9WvfvaK1xVsRRhhM4hQvw4QoacAtNaAGHCJ7hFd68sffivXsfi9aSV8wcwx94nz/eLIzN</latexit> sha1_base64="XaQViiIp3HEbSwArRCfYKY95vuc=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXjRncW3LisYi/QDiWTZtrQJDMkGaEMfQM3LhTt1pWv4863MdN2oa0/BD7+/xxyzgkTwY3F+NtbWV1b39gsbBW3d3b39ksHhw0Tp5qyOo1FrFshMUxwxeqWW8FaiWZEhoI1w+FNnjcfmTY8Vg92lLBAkr7iEafEOuvex91SGVfwVGgZ/DmUrz/fc01q3dJXpxfTVDJlqSDGtH2c2CAj2nIq2LjYSQ1LCB2SPms7VEQyE2TTScfozDk9FMXaPWXR1P3dkRFpzEiGrlISOzCLWW7+l7VTG10FGVdJapmis4+iVCAbo3xt1OOaUStGDgjV3M2K6IBoQq07TtEdwV9ceRkaFxUfV/w7XK6ewkwFOIYTOAcfLqEKt1CDOlCI4Ale4NUbes/emzeZla54854j+CPv4wfqV5FM</latexit> II <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> 0 IV V ! III ! ! ! I 20 0 v 20 syn 40 ⌫ ⌫ ⌫ ⌫ FIG 2 Instab t es o the homogeneous steady state (a) Two-parameter b urcat on d agram o Hop and Tur ng nstab t es o the homogeneous steady state n the (v yn η0 )-p ane where v yn = v 1yn = −v 2yn The Hop (green) Tur ng (red) and Tur ng-Hop (b ue) curves part t on the p ane nto five sectors The Hop and Tur ng curves cross at a cod mens on-2 po nt where both nstab t es occur s mu taneous y The homogeneous steady state u∗ s stab e n I (b) u∗ s unstab e to g oba per od c osc at ons n II and d rect numer ca s mu at ons c ose to the nstab ty show a stat onary Tur ng pattern We p ot R(x t) = z(x t) obta ned v a d rect numer ca s mu at on o (21) -- (22) or m = {1 2} w th exponent a y-decay ng kerne s (top pane ) The e genva ues o the spat a y c amped system at the b urcat on po nt are shown n the bottom pane (c) u∗ s Tur ng unstab e n III hence we observe stat onary per od c patterns The spectrum o the near sed operator around u∗ at b urcat on s a so shown (bottom pane ) (d) u∗ s Tur ng-Hop unstab e n IV and the nstab ty (bottom pane ) g ves r se to a per od c wavetra n as expected (e) spat o-tempora pattern obta ned n V where u∗ s Tur ng and Tur ng-Hop unstab e The spectrum n the bottom pane s at the cod mens on-2 po nt marked n (a) w th a red c rc e Parameters ∆ = 0 5 κ1 = κ2 = 5 τ1 = τ2 = 0 2 curve as vsyn s ncreased such that at the po nt where the two curves co de kc = 0 on the Tur ng-Hopf curve The curves n F g 2(a) were computed us ng XPPAUT 25 by cont nu ng a su tab e a gebra c prob em n one of the parameters of the system (vsyn ) The equat ons defin ng an equ br um are so ved s mu taneous y w th (25) -- (26) (27) -- (28) or (29) -- (31) and cont nued n parameter space The curves part t on the (vsyn η0 )space nto five sectors ( abe ed I -- V) In add t on to sectors where u∗ s stab e (I) u∗ s unstab e to bu k osc at ons (II) Tur ng nstab t es (III) and Tur ng-Hopf nstab t es (IV) a fifth sector (V) s generated by the cross ng of the Tur ng and Hopf curves At the ntersect on (cod mens on-2) po nt the spectrum of the near sed operator around u∗ has one zero e genva ue and two comp ex conjugate e genva ues where the cr t ca wavenumber kc s non zero for the zero e genva ue and equa to zero for the comp ex conjugate pa r (F g 2(e)) D rect numer ca s mu at ons c ose to the nstab ty confirm the pred ct ons of the near stab ty ana ys s bu k osc at ons are observed n n sect on I (F g 2(b)) stat onary Tur ng patterns are observed n sector III (F g 2(c)) and both wavetra ns and stand ng waves are seen n reg on IV (F g 2(d)) In sector V the s mu taneous Tur ng and Hopf unstab e modes compete resu t ng n a character st c comp ex spat o-tempora pattern where tempora osc at ons deve op w th n each bump of a Turng pattern (F g 2(e)) The ex stence of the ntersect on po nt of the Tur ng and Hopf curves and the observat on of the exot c spat o-tempora patterns are robust to changes n parameters In pass ng we note that c ose to onset we cou d find on y arge amp tude patterns of the type shown n F gs 2(c)-(e) nd cat ng that the Turng and Tur ng-Hopf b furcat ons are subcr t ca (as confirmed be ow by numer ca cont nuat on) Interest ng y we a so observe dynam c Tur ng patterns n reg on II mp y ng that the Hopf b furcat on does not stab ze the system nstead t creates b stab ty n th s reg on where the system supports both bu k osc at ons and dynam c g oba patterns We a so found other comp ex spat o-tempora patterns away from b furcat on onset us ng d rect numer ca s mu at on (F g 3) The system supports t me-per od c patterns conta n ng structures w th n bumps (F g 3(a)) Th s pattern s observed as we move away from the Turng b furcat on but stay c ose to the Hopf curve Struc- 7 (a) (b) 15 7 <latexit sha1_base64="YeR0z88bmpVpcF7lqgeTsszmBoI=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoseCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6xEnC/YgOlQgFo2ilB++qX664VXcOskq8nFQgR6Nf/uoNYpZGXCGT1Jiu5yboZ1SjYJJPS73U8ISyMR3yrqWKRtz42fzSKTm3yoCEsbalkMzV3xMZjYyZRIHtjCiOzLI3E//zuimGN34mVJIiV2yxKEwlwZjM3iYDoTlDObGEMi3srYSNqKYMbTglG4K3/PIqaV1WPbfq3dcq9bM8jiKcwClcgAfXUIc7aEATGITwDK/w5oydF+fd+Vi0Fpx85hj+wPn8AecAjNY=</latexit> (c) 40 t <latexit <latexit sha1_base64="coXTjA2n3M9S92rki2twuEpV/uI=">AAAB6HicbVBNS8NAEJ34WetX1aOXYBU8lUQEPRa8eGzBfkAbymY7adduNmF3IpTSX+DFgyJe/Une/Ddu2xy09cHA470ZZuaFqRSGPO/bWVvf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmGDBElspxpZHEpshaO7md96Qm1Eoh5onGIQs4ESkeCMrFSnXqnsVbw53FXi56QMOWq90le3n/AsRkVcMmM6vpdSMGGaBJc4LXYzgynjIzbAjqWKxWiCyfzQqXthlb4bJdqWIneu/p6YsNiYcRzazpjR0Cx7M/E/r5NRdBtMhEozQsUXi6JMupS4s6/dvtDISY4tYVwLe6vLh0wzTjabog3BX355lTSvKr5X8evX5ep5HkcBTuEMLsGHG6jCPdSgARwQnuEV3pxH58V5dz4WrWtOPnMCf+B8/gDXI4za</latexit> <latexit sha1_base64="6orIuPHYBBLcIp8lBRkdCUHcgEA=">AAAB6nicbZC7SgNBFIbPxluMt6ilzWAUrMKujVoZsLGMaC6QLGF2cjYZMju7zMwKYckj2FgoYusT2dnZ+hQ6uRSa+MPAx/+fw5xzgkRwbVz3w8ktLa+sruXXCxubW9s7xd29uo5TxbDGYhGrZkA1Ci6xZrgR2EwU0igQ2AgGV+O8cY9K81jemWGCfkR7koecUWOt217H6xRLbtmdiCyCN4PS5ffXxScAVDvF93Y3ZmmE0jBBtW55bmL8jCrDmcBRoZ1qTCgb0B62LEoaofazyagjcmydLgljZZ80ZOL+7shopPUwCmxlRE1fz2dj87+slZrw3M+4TFKDkk0/ClNBTEzGe5MuV8iMGFqgTHE7K2F9qigz9joFewRvfuVFqJ+WPbfs3bilyhFMlYcDOIQT8OAMKnANVagBgx48wBM8O8J5dF6c12lpzpn17MMfOW8/S8SQrQ==</latexit> sha1_base64="UHdBlgNWRTh7KOJmg1d/aACfaKs=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBU0m81GPBi8eK9gPaUDbbSbp0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSAXXxnW/ndLG5tb2Tnm3srd/cHhUPT7p6CRTDNssEYnqBVSj4BLbhhuBvVQhjQOB3WByO/e7T6g0T+SjmaboxzSSPOSMGis9RENvWK25dXcBsk68gtSgQGtY/RqMEpbFKA0TVOu+56bGz6kynAmcVQaZxpSyCY2wb6mkMWo/X5w6I5dWGZEwUbakIQv190ROY62ncWA7Y2rGetWbi/95/cyEN37OZZoZlGy5KMwEMQmZ/01GXCEzYmoJZYrbWwkbU0WZselUbAje6svrpHNd99y6d+/WmhdFHGU4g3O4Ag8a0IQ7aEEbGETwDK/w5gjnxXl3PpatJaeYOYU/cD5/AOd1jW0=</latexit> sha1_base64="603WMFkXVnL7ufUw5ulHaw6PUwA=">AAAB6nicbZC7TsMwFIZPuJZSoMDIYlGQmCqHBZioxMJYBL2gNqoc10mtOk5kO0hV1EdgYQAhVp6IjYEHgJcA9zJAyy9Z+vT/58jnHD8RXBuM352FxaXlldXcWn69sLG5Vdzeqes4VZTVaCxi1fSJZoJLVjPcCNZMFCORL1jD71+M8sYdU5rH8sYMEuZFJJQ84JQYa12HHbdTLOEyHgvNgzuF0vn359nXx22h2im+tbsxTSMmDRVE65aLE+NlRBlOBRvm26lmCaF9ErKWRUkipr1sPOoQHVqni4JY2ScNGru/OzISaT2IfFsZEdPTs9nI/C9rpSY49TIuk9QwSScfBalAJkajvVGXK0aNGFggVHE7K6I9ogg19jp5ewR3duV5qB+XXVx2r3CpcgAT5WAP9uEIXDiBClxCFWpAIYR7eIQnRzgPzrPzMildcKY9u/BHzusPCE+R+Q==</latexit> g1 g1 t <latexit sha1_base64="vV/tSZsqp0PZDp4B5kGJej5F2X0=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmj0U4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUa5ZLnlryaW6ycw1w5OIUzuAQPrqACd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QN+BpET</latexit> <latexit sha1_base64="NShdlnjkNDfv4ynZNqa/uWih2wQ=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0CiXPLfk1dxi5RzmysEpnMEleHAFFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fIMmPTg==</latexit> sha1_base64="uAhRiQUyUfrbUCj+oLzLI3Qe8O4=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VabqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9x24yU</latexit> <latexit <latexit sha1_base64="sYO7+g84YMjnL0FPL2Qemc7g+Xw=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcBOMlg==</latexit> 0 20 <latexit <latexit sha1_base64="zlFPBJwQd38NXQhvuYcrr8P1bxk=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cW7Ae0oWy2k3btZhN2N2IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52Sc6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasIbP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSuqx6btVrXFVqZ3kcRTiBU7gAD66hBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB3TOM3g==</latexit> x 0 20 20 <latexit sha1_base64="mFoNdpVbpfRu9/IYpphVUlYZwtU=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lKQY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzV3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17uuVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AODxjNI=</latexit> 15 x R <latexit <latexit sha1_base64="coXTjA2n3M9S92rki2twuEpV/uI=">AAAB6HicbVBNS8NAEJ34WetX1aOXYBU8lUQEPRa8eGzBfkAbymY7adduNmF3IpTSX+DFgyJe/Une/Ddu2xy09cHA470ZZuaFqRSGPO/bWVvf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmGDBElspxpZHEpshaO7md96Qm1Eoh5onGIQs4ESkeCMrFSnXqnsVbw53FXi56QMOWq90le3n/AsRkVcMmM6vpdSMGGaBJc4LXYzgynjIzbAjqWKxWiCyfzQqXthlb4bJdqWIneu/p6YsNiYcRzazpjR0Cx7M/E/r5NRdBtMhEozQsUXi6JMupS4s6/dvtDISY4tYVwLe6vLh0wzTjabog3BX355lTSvKr5X8evX5ep5HkcBTuEMLsGHG6jCPdSgARwQnuEV3pxH58V5dz4WrWtOPnMCf+B8/gDXI4za</latexit> <latexit <latexit sha1_base64="sYO7+g84YMjnL0FPL2Qemc7g+Xw=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcBOMlg==</latexit> 0 20 0.9 <latexit sha1_base64="iKaRjd0LPTdxRu1o0jMO04B10+I=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOCzBRiYWxCHqR2qhyXKe16jiR7SBVUR+BhQEUsTCw8DpsvA3uZYCWX7L06f/Pkc85QSK4Nhh/O4WV1bX1jeJmaWt7Z3evvH/Q0HGqKKvTWMSqFRDNBJesbrgRrJUoRqJAsGYwvJnkzUemNI/lgxklzI9IX/KQU2KsdY/dq265gl08FVoGbw6V6888fweAWrf81enFNI2YNFQQrdseToyfEWU4FWxc6qSaJYQOSZ+1LUoSMe1n01HH6NQ6PRTGyj5p0NT93ZGRSOtRFNjKiJiBXswm5n9ZOzXhpZ9xmaSGSTr7KEwFMjGa7I16XDFqxMgCoYrbWREdEEWosdcp2SN4iysvQ+Pc9bDr3eFK9QRmKsIRHMMZeHABVbiFGtSBQh+e4AVeHeE8O7nzNistOPOeQ/gj5+MHBESPxw==</latexit> sha1_base64="FGw24HvrkUCpD8krQVNS0aWfITI=">AAAB6nicbVBNSwMxEJ2tX7V+VT16CVbB05L1ot4KXjxWtB/QLiWbZtvQJLskWaEs/QlePCji1V/kzX9j2u5BWx8MPN6bYWZelApuLMbfXmltfWNzq7xd2dnd2z+oHh61TJJpypo0EYnuRMQwwRVrWm4F66SaERkJ1o7GtzO//cS04Yl6tJOUhZIMFY85JdZJD9i/6Vdr2MdzoFUSFKQGBRr96ldvkNBMMmWpIMZ0A5zaMCfacirYtNLLDEsJHZMh6zqqiGQmzOenTtG5UwYoTrQrZdFc/T2RE2nMREauUxI7MsveTPzP62Y2vg5zrtLMMkUXi+JMIJug2d9owDWjVkwcIVRzdyuiI6IJtS6digshWH55lbQu/QD7wT2u1c+KOMpwAqdwAQFcQR3uoAFNoDCEZ3iFN094L96797FoLXnFzDH8gff5A1VWjQ0=</latexit> sha1_base64="3fi41L6rWpOkNKkUnCUAyJVaLj0=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOCzBRiYWxCHqR2qhyXKe16jiR7SBVUR+BhQFUWFl4HTbeBqftAC2/ZOnT/58jn3OCRHBtMP52Ciura+sbxc3S1vbO7l55/6Ch41RRVqexiFUrIJoJLlndcCNYK1GMRIFgzWB4k+fNR6Y0j+WDGSXMj0hf8pBTYqx1j92rbrmCXTwVWgZvDpXrz0mut1q3/NXpxTSNmDRUEK3bHk6MnxFlOBVsXOqkmiWEDkmftS1KEjHtZ9NRx+jUOj0Uxso+adDU/d2RkUjrURTYyoiYgV7McvO/rJ2a8NLPuExSwySdfRSmApkY5XujHleMGjGyQKjidlZEB0QRaux1SvYI3uLKy9A4dz3sene4Uj2BmYpwBMdwBh5cQBVuoQZ1oNCHJ3iBV0c4z87EeZ+VFpx5zyH8kfPxA2GBkYw=</latexit> t R <latexit <latexit sha1_base64="coXTjA2n3M9S92rki2twuEpV/uI=">AAAB6HicbVBNS8NAEJ34WetX1aOXYBU8lUQEPRa8eGzBfkAbymY7adduNmF3IpTSX+DFgyJe/Une/Ddu2xy09cHA470ZZuaFqRSGPO/bWVvf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmGDBElspxpZHEpshaO7md96Qm1Eoh5onGIQs4ESkeCMrFSnXqnsVbw53FXi56QMOWq90le3n/AsRkVcMmM6vpdSMGGaBJc4LXYzgynjIzbAjqWKxWiCyfzQqXthlb4bJdqWIneu/p6YsNiYcRzazpjR0Cx7M/E/r5NRdBtMhEozQsUXi6JMupS4s6/dvtDISY4tYVwLe6vLh0wzTjabog3BX355lTSvKr5X8evX5ep5HkcBTuEMLsGHG6jCPdSgARwQnuEV3pxH58V5dz4WrWtOPnMCf+B8/gDXI4za</latexit> <latexit sha1_base64="Bm6L7INbJx8XEYQzKxIxOiRrWCU=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCzByEcCF7K3zMHK3t5ld8+EEH6BjYXGYGnn37Hz37gHFIq+ySZP3ncmOzNBIrg2rvvl5FZW19Y38puFre2d3b3i/kFDx6liWGexiFUroBoFl1g33AhsJQppFAhsBsPrLG8+oNI8lndmlKAf0b7kIWfUWKt22y2W3LI7E/kL3gJKVx/TTG/VbvGz04tZGqE0TFCt256bGH9MleFM4KTQSTUmlA1pH9sWJY1Q++PZoBNyap0eCWNlnzRk5v7sGNNI61EU2MqImoFezjLzv6ydmvDSH3OZpAYlm38UpoKYmGRbkx5XyIwYWaBMcTsrYQOqKDP2NgV7BG955b/QOC97btmruaXKCcyVhyM4hjPw4AIqcANVqAMDhEd4hhfn3nlyXp3pvDTnLHoO4Zec92+uhpEz</latexit> <latexit sha1_base64="qoVWhpdOghMvJ6ZyOV73cMAR63I=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMRdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmj0h4GP/z+HOecEieDauO6Xk1tZXVvfyG8WtrZ3dveK+wcNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQyvp3nzAZXmsbwzowT9iPYlDzmjxlq1226x5Jbdmchf8BZQuvqYTN4AoNotfnZ6MUsjlIYJqnXbcxPjZ1QZzgSOC51UY0LZkPaxbVHSCLWfzQYdk1Pr9EgYK/ukITP3Z0dGI61HUWArI2oGejmbmv9l7dSEl37GZZIalGz+UZgKYmIy3Zr0uEJmxMgCZYrbWQkbUEWZsbcp2CN4yyv/hcZ52XPLXs0tVU5grjwcwTGcgQcXUIEbqEIdGCA8wjO8OPfOk/PqTOalOWfRcwi/5Lx/A1FJj24=</latexit> sha1_base64="Bm6L7INbJx8XEYQzKxIxOiRrWCU=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCzByEcCF7K3zMHK3t5ld8+EEH6BjYXGYGnn37Hz37gHFIq+ySZP3ncmOzNBIrg2rvvl5FZW19Y38puFre2d3b3i/kFDx6liWGexiFUroBoFl1g33AhsJQppFAhsBsPrLG8+oNI8lndmlKAf0b7kIWfUWKt22y2W3LI7E/kL3gJKVx/TTG/VbvGz04tZGqE0TFCt256bGH9Mle <latexit sha1_base64="Bm6L7INbJx8XEYQzKxIxOiRrWCU=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCzByEcCF7K3zMHK3t5ld8+EEH6BjYXGYGnn37Hz37gHFIq+ySZP3ncmOzNBIrg2rvvl5FZW19Y38puFre2d3b3i/kFDx6liWGexiFUroBoFl1g33AhsJQppFAhsBsPrLG8+oNI8lndmlKAf0b7kIWfUWKt22y2W3LI7E/kL3gJKVx/TTG/VbvGz04tZGqE0TFCt256bGH9MleFM4KTQSTUmlA1pH9sWJY1Q++PZoBNyap0eCWNlnzRk5v7sGNNI61EU2MqImoFezjLzv6ydmvDSH3OZpAYlm38UpoKYmGRbkx5XyIwYWaBMcTsrYQOqKDP2NgV7BG955b/QOC97btmruaXKCcyVhyM4hjPw4AIqcANVqAMDhEd4hhfn3nlyXp3pvDTnLHoO4Zec92+uhpEz</latexit> <latexit sha1_base64="qoVWhpdOghMvJ6ZyOV73cMAR63I=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMRdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmj0h4GP/z+HOecEieDauO6Xk1tZXVvfyG8WtrZ3dveK+wcNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQyvp3nzAZXmsbwzowT9iPYlDzmjxlq1226x5Jbdmchf8BZQuvqYTN4AoNotfnZ6MUsjlIYJqnXbcxPjZ1QZzgSOC51UY0LZkPaxbVHSCLWfzQYdk1Pr9EgYK/ukITP3Z0dGI61HUWArI2oGejmbmv9l7dSEl37GZZIalGz+UZgKYmIy3Zr0uEJmxMgCZYrbWQkbUEWZsbcp2CN4yyv/hcZ52XPLXs0tVU5grjwcwTGcgQcXUIEbqEIdGCA8wjO8OPfOk/PqTOalOWfRcwi/5Lx/A1FJj24=</latexit> sha1_base64="af6fmYZ3mFiUWU/zWR5QDen5nZM=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLElsLMHIRwIXsrfMwcre3mV3z4Rc+AU2Fhpj60+y89+4wBUKvmSSl/dmMjMvSATXxnW/ncLG5tb2TnG3tLd/cHhUPj5p6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWByO/c7T6g0j+WDmSboR3QkecgZNVZq3g/KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGNn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o13Ur9Io+jCGdwDlfgQQ3qcAcNaAEDhGd4hTfn0Xlx3p2PZWvByWdO4Q+czx+iW4y0</latexit> 04 <latexit sha1_base64="0Yq5TIqZy8gmjXAagVPJ0JaLsro=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw2XN7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZJzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwhs/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdKqVT236t1fVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFKdI0J</latexit> 0 20 <latexit sha1_base64="HfC0yXjkT7KaNK62NtBBsNYHfxA=">AAAB6XicbZDLSgMxFIZPvNZ6q7p0E6yCq5IRQXcW3LisYi/QDiWTZtrQTGZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wgkcJYQr7R0vLK6tp6biO/ubW9s1vY26+ZONWMV1ksY90IqOFSKF61wkreSDSnUSB5Pehfj/P6A9dGxOreDhLuR7SrRCgYtc66OyftQpGUyER4EbwZFK8+R6N3AKi0C1+tTszSiCvLJDWm6ZHE+hnVVjDJh/lWanhCWZ92edOhohE3fjaZdIhPnNPBYazdUxZP3N8dGY2MGUSBq4yo7Zn5bGz+lzVTG176mVBJarli04/CVGIb4/HauCM0Z1YOHFCmhZsVsx7VlFl3nLw7gje/8iLUzkoeKXm3pFg+hqlycAhHcAoeXEAZbqACVWAQwiM8wwvqoyf0it6mpUto1nMAf4Q+fgCRqY+K</latexit> sha1_base64="Z3rR9Dc+KjBBKspqjleHMueVzmw=">AAAB6XicbVDLSgNBEOyNrxhfUY9eBqPgKcyKoMeAF49RzAOSJcxOZpMhs7PLTK8QQv7AiwdFvPpH3vwbJ8keNLGgoajqprsrTJW0SOm3V1hb39jcKm6Xdnb39g/Kh0dNm2SGiwZPVGLaIbNCSS0aKFGJdmoEi0MlWuHodua3noSxMtGPOE5FELOBlpHkDJ30cEV75Qqt0jnIKvFzUoEc9V75q9tPeBYLjVwxazs+TTGYMIOSKzEtdTMrUsZHbCA6jmoWCxtM5pdOyblT+iRKjCuNZK7+npiw2NpxHLrOmOHQLnsz8T+vk2F0E0ykTjMUmi8WRZkimJDZ26QvjeCoxo4wbqS7lfAhM4yjC6fkQvCXX14lzcuqT6v+Pa3UzvI4inACp3ABPlxDDe6gDg3gEMEzvMKbN/JevHfvY9Fa8PKZY/gD7/MH4ruM0A==</latexit> sha1_base64="kjBXwcmh55kaQv7/BX/Af1Thz10=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZEXRnwY3LKvYC7VAyaaYNzWSGJCOUoW/gxoWi3bryddz5NmamXWjrD4GP/z+HnHP8WHBtMP52Ciura+sbxc3S1vbO7l55/6Cpo0RR1qCRiFTbJ5oJLlnDcCNYO1aMhL5gLX90k+WtR6Y0j+SDGcfMC8lA8oBTYqx1f4F75Qqu4lxoGdw5VK4/3zNN673yV7cf0SRk0lBBtO64ODZeSpThVLBJqZtoFhM6IgPWsShJyLSX5pNO0Kl1+iiIlH3SoNz93ZGSUOtx6NvKkJihXswy87+sk5jgyku5jBPDJJ19FCQCmQhla6M+V4waMbZAqOJ2VkSHRBFq7HFK9gju4srL0Dyvurjq3uFK7QRmKsIRHMMZuHAJNbiFOjSAQgBP8AKvzsh5dt6c6ay04Mx7DuGPnI8f7uaRTw==</latexit> t <latexit <latexit sha1_base64="zlFPBJwQd38NXQhvuYcrr8P1bxk=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cW7Ae0oWy2k3btZhN2N2IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52Sc6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasIbP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSuqx6btVrXFVqZ3kcRTiBU7gAD66hBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB3TOM3g==</latexit> x 20 g1 t 2 20 0 9 40 <latexit sha1_base64="YeR0z88bmpVpcF7lqgeTsszmBoI=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoseCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6xEnC/YgOlQgFo2ilB++qX664VXcOskq8nFQgR6Nf/uoNYpZGXCGT1Jiu5yboZ1SjYJJPS73U8ISyMR3yrqWKRtz42fzSKTm3yoCEsbalkMzV3xMZjYyZRIHtjCiOzLI3E//zuimGN34mVJIiV2yxKEwlwZjM3iYDoTlDObGEMi3srYSNqKYMbTglG4K3/PIqaV1WPbfq3dcq9bM8jiKcwClcgAfXUIc7aEATGITwDK/w5oydF+fd+Vi0Fpx85hj+wPn8AecAjNY=</latexit> 8 2 2 <latexit sha1_base64="0Yq5TIqZy8gmjXAagVPJ0JaLsro=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw2XN7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZJzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwhs/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdKqVT236t1fVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFKdI0J</latexit> 40 7 <latexit sha1_base64="aIMXJcem4EvcfdCCABkUCBoBcYg=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmzwU4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUaVyXPLXk1t1g5h7lycApncAkelKECd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QOFmpEY</latexit> <latexit sha1_base64="LN8NJGYOzXCGSISdV1ByBPyZiB0=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdi1iZ0BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0LgqeW7Jq7nFyjnMlYNTOINL8KAMFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fKF2PUw==</latexit> sha1_base64="wQdLoBJ/A11xQHtdeM5u5orD4es=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbKElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NWHdz7hMUoOSrRaFqSAmJouvyYgrZEbMLKFMcXsrYROqKDM2m5INwVt/eZN0bqueW/VabqVxlcdRhAu4hBvwoAYNuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w95b4yZ</latexit> <latexit sha1_base64="yQfmJRBoTgLxijYwuTWw/rLs+7k=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgFe5sYmfAxjKi+YDkCHubuWTJ3t6xuyeEIz/BxkIJNhY2/h07/42bj0ITHww83pthZl6QCK6N6347ubX1jc2t/HZhZ3dv/6B4eNTQcaoY1lksYtUKqEbBJdYNNwJbiUIaBQKbwfBm6jcfUWkeywczStCPaF/ykDNqrHTvlr1useSW3RnIKvEWpHT9OZm8A0CtW/zq9GKWRigNE1Trtucmxs+oMpwJHBc6qcaEsiHtY9tSSSPUfjY7dUzOrdIjYaxsSUNm6u+JjEZaj6LAdkbUDPSyNxX/89qpCa/8jMskNSjZfFGYCmJiMv2b9LhCZsTIEsoUt7cSNqCKMmPTKdgQvOWXV0njsuzZxO7cUvUM5sjDCZzCBXhQgSrcQg3qwKAPT/ACr45wnp2J8zZvzTmLmWP4A+fjB/gVj78=</latexit> sha1_base64="MMH0J0T+HrKP6TiQR5Y2+qcfYYo=">AAAB6nicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLUlsLDEKksCF7C1zsGFv97K7Z0Iu/AQbC42x9RfZ+W9c4AoFXzLJy3szmZkXpYIb6/vfXmltfWNzq7xd2dnd2z+oHh61jco0wxZTQulORA0KLrFluRXYSTXSJBL4GI1vZv7jE2rDlXywkxTDhA4ljzmj1kn3fj3oV2t+3Z+DrJKgIDUo0OxXv3oDxbIEpWWCGtMN/NSGOdWWM4HTSi8zmFI2pkPsOippgibM56dOyblTBiRW2pW0ZK7+nshpYswkiVxnQu3ILHsz8T+vm9n4Osy5TDOLki0WxZkgVpHZ32TANTIrJo5Qprm7lbAR1ZRZl07FhRAsv7xK2pf1wCV259caZ0UcZTiBU7iAAK6gAbfQhBYwGMIzvMKbJ7wX7937WLSWvGLmGP7A+/wBSTaNBQ==</latexit> sha1_base64="e+R001ceKoRS5MEsQ4XGScHHhr4=">AAAB6nicbVA9SwNBEJ3zM8avqKXNYhSswp2NdgZsLCOaD0iOsLfZS5bs7R27c0I48hNsLJRoa+PfsfPfuJek0MQHA4/3ZpiZFyRSGHTdb2dldW19Y7OwVdze2d3bLx0cNkycasbrLJaxbgXUcCkUr6NAyVuJ5jQKJG8Gw5vcbz5ybUSsHnCUcD+ifSVCwSha6d6teN1S2a24U5Bl4s1J+fpzkuOt1i19dXoxSyOukElqTNtzE/QzqlEwycfFTmp4QtmQ9nnbUkUjbvxseuqYnFmlR8JY21JIpurviYxGxoyiwHZGFAdm0cvF/7x2iuGVnwmVpMgVmy0KU0kwJvnfpCc0ZyhHllCmhb2VsAHVlKFNp2hD8BZfXiaNi4pnE7tzy9VTmKEAx3AC5+DBJVThFmpQBwZ9eIIXeHWk8+xMnPdZ64oznzmCP3A+fgBVYZGE</latexit> <latexit <latexit sha1_base64="sYO7+g84YMjnL0FPL2Qemc7g+Xw=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1Fip6Q7KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGtn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o1a5X6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AcBOMlg==</latexit> <latexit sha1_base64="mFoNdpVbpfRu9/IYpphVUlYZwtU=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lKQY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzV3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17uuVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AODxjNI=</latexit> <latexit sha1_base64="0Yq5TIqZy8gmjXAagVPJ0JaLsro=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpgh4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw2XN7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZJzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwhs/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdKqVT236t1fVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFKdI0J</latexit> 0 20 <latexit <latexit sha1_base64="zlFPBJwQd38NXQhvuYcrr8P1bxk=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cW7Ae0oWy2k3btZhN2N2IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52Sc6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasIbP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSuqx6btVrXFVqZ3kcRTiBU7gAD66hBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB3TOM3g==</latexit> x x 20 40 09 t R 01 0.1 0 20 20 <latexit sha1_base64="mFoNdpVbpfRu9/IYpphVUlYZwtU=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lKQY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKDzV3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17uuVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AODxjNI=</latexit> x 20 FIG 3 Further spat o-tempora patterns supported by the neura fie d mode (a) Pattern ound n sector V o F g 2(a) or η0 = 60 v 1yn = −v 2yn = 10 The pattern d sp ays a t me-per od c structure w th n each bump o a Tur ng pattern (b) Wander ng bump observed n reg on IV o F g 2 or η0 = 2 5 v 1yn = −v 2yn = −20 S m ar patterns are a so ound n sector II (not shown) (c) Weak y unstab e stand ng wave obta ned or η0 = 15 v yn = −10 A ter undergo ng a breath ng nstab ty or a ong trans ent the stand ng wave evo ves towards a stab e wavetra n (not shown) Other parameter va ues as n F g 2 tures of th s type n wh ch the bumps of a Tur ng pattern are per od ca y modu ated n t me were observed n t a y n sector IV of F g 2(a) Spat o-tempora patterns of th s form w th modu at on at the core were found on y n parameter sets where the Tur ng and Hopf cod mens on-2 po nt s present as expected Structures w th n bump so ut ons are not seen n standard neura fie d mode s They are however common y observed n sp k ng neuron mode s 16 30 wh ch emphas ses that th s next generat on neura mass mode reta ns nformat on about the under y ng sp k ng mode In reg on IV we a so observe an number of nterestng patterns such as the wandering bump (F g 3(b)) and a form of stand ng wave where both the w dth and the he ght of the bumps s per od ca y modu ated (F g 3(c)) Numer ca s mu at ons nd cate that the wanderng bump s stab e for a arge area of parameter space and can coex st w th a per od c wavetra n The stand ng wave however pers sts for a ong t me but u t mate y s unstab e and trans t ons to a per od c trave ng wave (not shown) IV NUMERICAL BIFURCATION ANALYSIS In order to study patterns away from b furcat ons we emp oy numer ca b furcat on techn ques wh ch a ows us to compute coherent structures determ ne the r stab ty and track the r dependence on contro parameters Here we emp oy the numer ca too k t deve oped by Av tab e 5 and emp oyed n the context of standard neura fie ds n 6 41 Th s too k t can be used to compute waves and patterns and the r stab ty We refer the reader to recent rev ews on numer ca b furcat on ana ys s for coherent structures 15 31 F rst we resca ed space such that x ∈ −0 5 0 5 to h gh ght the dependence on the sca e of the doma n s ze Λ The equ va ent PDE formu at on (21) -- (22) becomes ∂t z = F(z) + G(z I) 2 2 (Λ − ∂xx )(1 + τ1 ∂t )2 g1 = Λ2 κ1 f (z) 2 2 (32) 2 (Λ − ∂xx /β )(1 + τ2 ∂t ) g2 = Λ κ2 f (z) We perform numer ca b furcat on ana ys s for heterogeneous spat a y patterned steady states and per od c trave ng waves of the system above We construct the 8 5 where (Z, G1 , G2 ) is a reference template solution, such as one of the solutions obtained via direct simulation. We discretised the differential operators (34) using standard differentiation matrices, which are also used to compute linear stability of the coherent structures [5]. <latexit sha1_base64="++U/7ibNG5dsvlS4pWSL3QxJKUc=">AAAB6HicbZA9SwNBEIbn4leMX1FLm8UoWIU7QbQzYGOZgPmA5Ah7m7lkzd7esbsnhJBfYGOhSCzt/Dt2/hv3khQafWHh4X1n2JkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DgpKRFg==</latexit> <latexit sha1_base64="jWIA7LPy9Pp3r6oVJFUvj5dZ/0Y=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuINoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSVVj1E=</latexit> sha1_base64="uueKIUk3ohN/s04dlh2z82vluXQ=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mEoseCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipWRuUK27VXYCsEy8nFcjRGJS/+sOYpRFKwwTVuue5ifEzqgxnAmelfqoxoWxCR9izVNIItZ8tDp2RS6sMSRgrW9KQhfp7IqOR1tMosJ0RNWO96s3F/7xeasJbP+MySQ1KtlwUpoKYmMy/JkOukBkxtYQyxe2thI2poszYbEo2BG/15XXSvq56btVrupX6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AdmeMlw==</latexit> Homogeneous k = 0.738 k = 0.969 Turing Hopf Sym. break <latexit sha1_base64="LBvt9H+9GHywVwK9vmREOuBtsis=">AAAB8nicbVA9SwNBEN3zM8avqKXNYhCswl0aLQM2KSOYD0iOsLeZS5bs7R67c0I48jNsLBSx9dfY+W/cJFdo4oOBx3szzMyLUiks+v63t7W9s7u3XzooHx4dn5xWzs47VmeGQ5trqU0vYhakUNBGgRJ6qQGWRBK60fR+4XefwFih1SPOUggTNlYiFpyhk/pNnegxKNCZHVaqfs1fgm6SoCBVUqA1rHwNRppnCSjkklnbD/wUw5wZFFzCvDzILKSMT9kY+o4qloAN8+XJc3rtlBGNtXGlkC7V3xM5S6ydJZHrTBhO7Lq3EP/z+hnGd2EuVJohKL5aFGeSoqaL/+lIGOAoZ44wboS7lfIJM4yjS6nsQgjWX94knXot8GvBQ73aoEUcJXJJrsgNCcgtaZAmaZE24USTZ/JK3jz0Xrx372PVuuUVMxfkD7zPH4yikU0=</latexit> <latexit sha1_base64="TatFJ50VQZC/E/hWsbl1ikTYf4M=">AAAB7nicbZC7SgNBFIbPxluMt6ilzWAQrJZZLZJGDNhYRjAXSJYwO5lNhp2dXWZmhbjkIWwsFLGx8HnsLH0TJ5dCoz8MfPz/Ocw5J0gF1wbjT6ewsrq2vlHcLG1t7+zulfcPWjrJFGVNmohEdQKimeCSNQ03gnVSxUgcCNYOoqtp3r5jSvNE3ppxyvyYDCUPOSXGWu3oArvV81q/XMEungn9BW8Blcu3+686ADT65Y/eIKFZzKShgmjd9XBq/Jwow6lgk1Iv0ywlNCJD1rUoScy0n8/GnaAT6wxQmCj7pEEz92dHTmKtx3FgK2NiRno5m5r/Zd3MhDU/5zLNDJN0/lGYCWQSNN0dDbhi1IixBUIVt7MiOiKKUGMvVLJH8JZX/gutM9fDrneDK3UEcxXhCI7hFDyoQh2uoQFNoBDBAzzBs5M6j86L8zovLTiLnkP4Jef9G+j6kMA=</latexit> sha1_base64="Dv+N+wkEIegEZBk0NwjU/LerIDI=">AAAB7nicbVBNS8NAEJ3Ur1o/WvXoZbEInsJGD+1FKHjxWMF+QBvKZrtpl2w2YXcjlNAf4cWDIl79Pd78N27bHLT1wcDjvRlm5gWp4Npg/O2UtrZ3dvfK+5WDw6Pjau3ktKuTTFHWoYlIVD8gmgkuWcdwI1g/VYzEgWC9ILpb+L0npjRP5KOZpcyPyUTykFNirNSLbrHbuGmOanXs4iXQJvEKUocC7VHtazhOaBYzaaggWg88nBo/J8pwKti8Msw0SwmNyIQNLJUkZtrPl+fO0aVVxihMlC1p0FL9PZGTWOtZHNjOmJipXvcW4n/eIDNh08+5TDPDJF0tCjOBTIIWv6MxV4waMbOEUMXtrYhOiSLU2IQqNgRv/eVN0r12Pex6D7jeQkUcZTiHC7gCDxrQgntoQwcoRPAMr/DmpM6L8+58rFpLTjFzBn/gfP4AisCOQg==</latexit> sha1_base64="m8ohI5H75Lden12QNBVfOtJR2zc=">AAAB7nicbZC7SgNBFIbPxluMt3jpbAaDYLXMapE0YsBCywjmAskSZiezybCzF2ZmhbjkIWwsFLGxsPJh7Cx9EyeXQhN/GPj4/3OYc46XCK40xl9Wbml5ZXUtv17Y2Nza3inu7jVUnErK6jQWsWx5RDHBI1bXXAvWSiQjoSdY0wsux3nzjknF4+hWDxPmhqQfcZ9Too3VDM6xXT6rdIslbOOJ0CI4MyhdvN1/X30cZLVu8bPTi2kaskhTQZRqOzjRbkak5lSwUaGTKpYQGpA+axuMSMiUm03GHaFj4/SQH0vzIo0m7u+OjIRKDUPPVIZED9R8Njb/y9qp9ituxqMk1Syi04/8VCAdo/HuqMclo1oMDRAquZkV0QGRhGpzoYI5gjO/8iI0Tm0H284NLlURTJWHQziCE3CgDFW4hhrUgUIAD/AEz1ZiPVov1uu0NGfNevbhj6z3H5oxkgQ=</latexit> <latexit sha1_base64="TEM37w6PcFQLCBxv+NzlDSPPh4U=">AAAB6HicbZA9SwNBEIbn4leMX1FLm8UoWIU7EbQzYGOZgPmA5Ah7m7lkzd7esbsnhJBfYGOhSCzt/Dt2/hv3khQafWHh4X1n2JkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DgQ6RFQ==</latexit> <latexit sha1_base64="zDS/d1EMLhdhjGncUr18mp+9fB4=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuCNoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSPRj1A=</latexit> sha1_base64="G9IPPnJCtUtisXfD2y/+tp8gcfA=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipWRuUK27VXYCsEy8nFcjRGJS/+sOYpRFKwwTVuue5ifEzqgxnAmelfqoxoWxCR9izVNIItZ8tDp2RS6sMSRgrW9KQhfp7IqOR1tMosJ0RNWO96s3F/7xeasJbP+MySQ1KtlwUpoKYmMy/JkOukBkxtYQyxe2thI2poszYbEo2BG/15XXSvq56btVrupX6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AdOOMlg==</latexit> 4 <latexit sha1_base64="NWK7le7LCxKByUlK3CYfxrO8yFQ=">AAAB7nicbZC7SgNBFIbPeo3xFrW0GQyC1TJroaYQAzaWEcwFkiXMTmaTYWdnl5lZIS55CBsLRWwsfB47S9/EyaXQxB8GPv7/HOacE6SCa4Pxl7O0vLK6tl7YKG5ube/slvb2GzrJFGV1mohEtQKimeCS1Q03grVSxUgcCNYMoutx3rxnSvNE3plhyvyY9CUPOSXGWs3oEruVs0q3VMYunggtgjeD8tX7w3cVAGrd0menl9AsZtJQQbRuezg1fk6U4VSwUbGTaZYSGpE+a1uUJGbazyfjjtCxdXooTJR90qCJ+7sjJ7HWwziwlTExAz2fjc3/snZmwgs/5zLNDJN0+lGYCWQSNN4d9bhi1IihBUIVt7MiOiCKUGMvVLRH8OZXXoTGqeth17vF5SqCqQpwCEdwAh6cQxVuoAZ1oBDBIzzDi5M6T86r8zYtXXJmPQfwR87HD/IZkMY=</latexit> sha1_base64="3oChPboWBvplm9+XVPCIeE9dyZo=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU9h4sPYgFLx4rGA/oA1ls920SzabsLsRSuiP8OJBEa/+Hm/+G7dtDtr6YODx3gwz84JUcG0w/nZKG5tb2zvl3cre/sHhUfX4pKOTTFHWpolIVC8gmgkuWdtwI1gvVYzEgWDdILqb+90npjRP5KOZpsyPyVjykFNirNSNbrHbuG4MqzXs4gXQOvEKUoMCrWH1azBKaBYzaaggWvc9nBo/J8pwKtisMsg0SwmNyJj1LZUkZtrPF+fO0IVVRihMlC1p0EL9PZGTWOtpHNjOmJiJXvXm4n9ePzPhjZ9zmWaGSbpcFGYCmQTNf0cjrhg1YmoJoYrbWxGdEEWosQlVbAje6svrpHPletj1HnCtiYo4ynAG53AJHtShCffQgjZQiOAZXuHNSZ0X5935WLaWnGLmFP7A+fwBk9+OSA==</latexit> sha1_base64="nSQFn0zCANkIUdnHQyAJFcq8TdE=">AAAB7nicbZC7SgNBFIbPxluMt3jpbAaDYLXMWqgpxICFlhHMBZIlzE5mk2FnL8zMCnHJQ9hYKGJjYeXD2Fn6Jk4uhSb+MPDx/+cw5xwvEVxpjL+s3MLi0vJKfrWwtr6xuVXc3qmrOJWU1WgsYtn0iGKCR6ymuRasmUhGQk+whhdcjvLGHZOKx9GtHiTMDUkv4j6nRBurEZxju3xS7hRL2MZjoXlwplC6eLv/vvrYy6qd4me7G9M0ZJGmgijVcnCi3YxIzalgw0I7VSwhNCA91jIYkZApNxuPO0SHxukiP5bmRRqN3d8dGQmVGoSeqQyJ7qvZbGT+l7VS7Z+5GY+SVLOITj7yU4F0jEa7oy6XjGoxMECo5GZWRPtEEqrNhQrmCM7syvNQP7YdbDs3uFRBMFEe9uEAjsCBU6jANVShBhQCeIAneLYS69F6sV4npTlr2rMLf2S9/wCjUJIK</latexit> <latexit sha1_base64="/UB/vzmNQFDYws8N3H8eJhkay4k=">AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SaBmwsYyQL0iOsLfZJGv2do/dOSEc+Q82ForY+n/s/Ddukis08cHA470ZZuZFiRQWff/bK2xt7+zuFfdLB4dHxyfl07O21alhvMW01KYbUculULyFAiXvJobTOJK8E03vFn7niRsrtGriLOFhTMdKjASj6KR2MzVCjQflil/1lyCbJMhJBXI0BuWv/lCzNOYKmaTW9gI/wTCjBgWTfF7qp5YnlE3pmPccVTTmNsyW187JlVOGZKSNK4Vkqf6eyGhs7SyOXGdMcWLXvYX4n9dLcXQbZkIlKXLFVotGqSSoyeJ1MhSGM5QzRygzwt1K2IQaytAFVHIhBOsvb5J2rRr41eChVqmTPI4iXMAlXEMAN1CHe2hACxg8wjO8wpunvRfv3ftYtRa8fOYc/sD7/AGjtY8L</latexit> <latexit sha1_base64="FNo+pe5sTlLBpYZLTW80sYEDLr8=">AAAB63icbVDLSgNBEOyNrxhfUY9eBoPgKezmYo4BLzlGMA9IljA7mU2GzGOZmRXCkl/w4kERr/6QN//G2WQPmljQUFR1090VJZwZ6/vfXmlnd2//oHxYOTo+OT2rnl/0jEo1oV2iuNKDCBvKmaRdyyyng0RTLCJO+9H8Pvf7T1QbpuSjXSQ0FHgqWcwItrnUVkk8rtb8ur8C2iZBQWpQoDOufo0miqSCSks4NmYY+IkNM6wtI5wuK6PU0ASTOZ7SoaMSC2rCbHXrEt04ZYJipV1Ji1bq74kMC2MWInKdAtuZ2fRy8T9vmNq4GWZMJqmlkqwXxSlHVqH8cTRhmhLLF45gopm7FZEZ1phYF0/FhRBsvrxNeo164NeDh0athYo4ynAF13ALAdxBC9rQgS4QmMEzvMKbJ7wX7937WLeWvGLmEv7A+/wB8CmOCw==</latexit> <latexit sha1_base64="QwqXlsUJ+2ImkUZ9N/vpak8Rr0I=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuROC+0ksbGERD4SuJC9ZQ5W9vYuu3smhPALbCw0Bks7/46d/8Y9oFD0TTZ58r4z2ZkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8Df4qRFA==</latexit> <latexit sha1_base64="2gar9dxBtUHQpybzRUsQDoHhgGc=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuFtoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSJNj08=</latexit> sha1_base64="fhRWf68oYtxnqsuxSJW7mk5YLhM=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFqNgFe5MoWXAxjIB8wHJEfY2c8mavb1jd08IR36BjYUitv4kO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7ASTu7nfeUKleSwfzDRBP6IjyUPOqLFSszYoV9yquwBZJ15OKpCjMSh/9YcxSyOUhgmqdc9zE+NnVBnOBM5K/VRjQtmEjrBnqaQRaj9bHDojl1YZkjBWtqQhC/X3REYjradRYDsjasZ61ZuL/3m91IS3fsZlkhqUbLkoTAUxMZl/TYZcITNiagllittbCRtTRZmx2ZRsCN7qy+ukfV313KrXdCv1izyOIpzBOVyBBzdQh3toQAsYIDzDK7w5j86L8+58LFsLTj5zCn/gfP4Ac1+MlQ==</latexit> 3 <latexit sha1_base64="Wn12gZIJOOepDb/UZbaNJFgs+Jo=">AAAB8XicbVA9T8MwFHwpX6V8FRhZLCokpijpAmMlFsYiSFvRRpXjOq1V24lsBymK+i9YGECIlX/Dxr/BbTNAy0mWTnfv5PcuSjnTxvO+ncrG5tb2TnW3trd/cHhUPz7p6CRThAYk4YnqRVhTziQNDDOc9lJFsYg47UbTm7nffaJKs0Q+mDylocBjyWJGsLHS430uXBTZwHRYb3iutwBaJ35JGlCiPax/DUYJyQSVhnCsdd/3UhMWWBlGOJ3VBpmmKSZTPKZ9SyUWVIfFYuMZurDKCMWJsk8atFB/JwostM5FZCcFNhO96s3F/7x+ZuLrsGAyzQyVZPlRnHFkEjQ/H42YosTw3BJMFLO7IjLBChNjS6rZEvzVk9dJp+n6nuvfNRstVNZRhTM4h0vw4QpacAttCICAhGd4hTdHOy/Ou/OxHK04ZeYU/sD5/AHhwJBG</latexit> kg1 k1 2 <latexit sha1_base64="z47r0npwszHz2/UHFt3Oq9Oi4sE=">AAAB/3icbZA9SwNBEIbnNGqMX1HBxmYxClbhzkbLgI1lBPMByXHubfaSJXt7x+6ccMQU/hUbC0Vs/Rt2/hs3H4UmvrDw8M4MM/uGqRQGXffbWVktrK1vFDdLW9s7u3vl/YOmSTLNeIMlMtHtkBouheINFCh5O9WcxqHkrXB4Pam3Hrg2IlF3mKfcj2lfiUgwitYKykfdJtdI+oFHphR0hYowD8oVt+pORZbBm0OlVoiiewCoB+Wvbi9hWcwVMkmN6Xhuiv6IahRM8nGpmxmeUjakfd6xqGjMjT+a3j8mZ9bpkSjR9ikkU/f3xIjGxuRxaDtjigOzWJuY/9U6GUZX/kioNEOu2GxRlEmCCZmEQXpCc4Yyt0CZFvZWwgZUU4Y2spINwVv88jI0L6qeW/VubRqnMFMRjuEEzsGDS6jBDdShAQwe4Rle4c15cl6cd+dj1rrizGcO4Y+czx8qnZbr</latexit> sha1_base64="Wpyl60zfUfdviD9FUEBemg951cw=">AAAB/3icbZDNSsNAFIUn9a/Wv6jgxs1gFVyVxI0uC25cVrCt0IQwmU7aoZNJmLkRQuzCV3HjQhG3voY738ZpmoW2Hhj4OPde7p0TpoJrcJxvq7ayura+Ud9sbG3v7O7Z+wc9nWSKsi5NRKLuQ6KZ4JJ1gYNg96liJA4F64eT61m9/8CU5om8gzxlfkxGkkecEjBWYB95PaYAjwIXlxR4XEaQB3bTaTml8DK4FTRRpU5gf3nDhGYxk0AF0XrgOin4BVHAqWDThpdplhI6ISM2MChJzLRflPdP8ZlxhjhKlHkScOn+nihIrHUeh6YzJjDWi7WZ+V9tkEF05RdcphkwSeeLokxgSPAsDDzkilEQuQFCFTe3YjomilAwkTVMCO7il5ehd9FynZZ76zTbp1UcdXSMTtA5ctElaqMb1EFdRNEjekav6M16sl6sd+tj3lqzqplD9EfW5w85wZV7</latexit> sha1_base64="GA6ely70JSOYBIcvitrigLFlMo8=">AAAB/3icbZDNSsNAFIUntWqtf1HBjZvBKrgqiRtdFty4rGB/oAlhMp20QyeTMHMjhNiFr+LGhSJufQ13vo3TtAttPTDwce693DsnTAXX4DjfVmWtur6xWduqb+/s7u3bB4ddnWSKsg5NRKL6IdFMcMk6wEGwfqoYiUPBeuHkZlbvPTCleSLvIU+ZH5OR5BGnBIwV2MdelynAo8DFJQUelxHkgd1wmk4pvAruAhqtalSqHdhf3jChWcwkUEG0HrhOCn5BFHAq2LTuZZqlhE7IiA0MShIz7Rfl/VN8bpwhjhJlngRcur8nChJrnceh6YwJjPVybWb+VxtkEF37BZdpBkzS+aIoExgSPAsDD7liFERugFDFza2YjokiFExkdROCu/zlVeheNl2n6d6ZNM7QXDV0gk7RBXLRFWqhW9RGHUTRI3pGr+jNerJerHfrY95asRYzR+iPrM8fy4iYIw==</latexit> A. <latexit sha1_base64="vV/tSZsqp0PZDp4B5kGJej5F2X0=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmj0U4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUa5ZLnlryaW6ycw1w5OIUzuAQPrqACd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QN+BpET</latexit> <latexit sha1_base64="NShdlnjkNDfv4ynZNqa/uWih2wQ=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0CiXPLfk1dxi5RzmysEpnMEleHAFFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fIMmPTg==</latexit> sha1_base64="uAhRiQUyUfrbUCj+oLzLI3Qe8O4=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VabqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9x24yU</latexit> 1 <latexit sha1_base64="hguXnqry5RuPdMY+wTR8icGo8UA=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCwhkY8ELmRvmYOVvb3L7p4JIfwCGwuNwdLOv2Pnv3EPKBR8k02evO9MdmaCRHBtXPfbya2tb2xu5bcLO7t7+wfFw6OGjlPFsM5iEatWQDUKLrFuuBHYShTSKBDYDIa3Wd58RKV5LO/NKEE/on3JQ86osVbN6xZLbtmdiayCt4DSzec003u1W/zq9GKWRigNE1Trtucmxh9TZTgTOCl0Uo0JZUPax7ZFSSPU/ng26IScW6dHwljZJw2Zub87xjTSehQFtjKiZqCXs8z8L2unJrz2x1wmqUHJ5h+FqSAmJtnWpMcVMiNGFihT3M5K2IAqyoy9TcEewVteeRUal2XPLXs1t1Q5g7nycAKncAEeXEEF7qAKdWCA8AQv8Oo8OM/OmzOdl+acRc8x/JHz8QN8gpES</latexit> <latexit sha1_base64="9tJtMG+ucSzlgnQupHZj3PyqRTA=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMBdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmjiDwMf/38Oc84JEsG1cd1vJ7e2vrG5ld8u7Ozu7R8UD48aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeTvPmIyrNY3lvRgn6Ee1LHnJGjbVqXrdYcsvuTGQVvAWUbj4nk3cAqHaLX51ezNIIpWGCat323MT4GVWGM4HjQifVmFA2pH1sW5Q0Qu1ns0HH5Nw6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G0K9gje8sqr0Lgse27Zq7mlyhnMlYcTOIUL8OAKKnAHVagDA4QneIFX58F5dt6cybw05yx6juGPnI8fH0WPTQ==</latexit> sha1_base64="0q2hcy/Ngoze85Zh4Z8FouvdSlI=">AAAB6HicbVBNT8JAEJ3iF+IX6tHLRjTxRFoveiTx4hESCyTQkO0yhZXtttndmpCGX+DFg8Z49Sd589+4QA8KvmSSl/dmMjMvTAXXxnW/ndLG5tb2Tnm3srd/cHhUPT5p6yRTDH2WiER1Q6pRcIm+4UZgN1VI41BgJ5zczf3OEyrNE/lgpikGMR1JHnFGjZVa3qBac+vuAmSdeAWpQYHmoPrVHyYsi1EaJqjWPc9NTZBTZTgTOKv0M40pZRM6wp6lksaog3xx6IxcWmVIokTZkoYs1N8TOY21nsah7YypGetVby7+5/UyE90GOZdpZlCy5aIoE8QkZP41GXKFzIipJZQpbm8lbEwVZcZmU7EheKsvr5P2dd1z617LrTUuijjKcAbncAUe3EAD7qEJPjBAeIZXeHMenRfn3flYtpacYuYU/sD5/AFwV4yT</latexit> <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> <latexit sha1_base64="iEuqQD02qntqL9OyWq6wx5NE8ZI=">AAAB6nicbZDLSgMxFIZP6q3WW9Wlm2AV3FgybnQjFnThsqK9QDuUTJppQzOZIckIZegjuHGhiFufyJ3v4EOYXhba+kPg4//PIeecIJHCWEK+UG5peWV1Lb9e2Njc2t4p7u7VTZxqxmsslrFuBtRwKRSvWWElbyaa0yiQvBEMrsd545FrI2L1YIcJ9yPaUyIUjFpn3Z96pFMskTKZCC+CN4PS1ffNZQUAqp3iZ7sbszTiyjJJjWl5JLF+RrUVTPJRoZ0anlA2oD3ecqhoxI2fTUYd4WPndHEYa/eUxRP3d0dGI2OGUeAqI2r7Zj4bm/9lrdSGF34mVJJartj0ozCV2MZ4vDfuCs2ZlUMHlGnhZsWsTzVl1l2n4I7gza+8CPWzskfK3h0pVY5gqjwcwCGcgAfnUIFbqEINGPTgCV7gFUn0jN7Q+7Q0h2Y9+/BH6OMH4XKO8A==</latexit> sha1_base64="TAV6G+83+fR0ZQt+gj6L/MD1f/w=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgY9iz0TJgYxnRfEByhL3NXLJkb+/Y3RNCyE+wsVDE1l9k579xk1yhiQ8GHu/NMDMvTKUwltJvr7C2vrG5Vdwu7ezu7R+UD4+aJsk0xwZPZKLbITMohcKGFVZiO9XI4lBiKxzdzvzWE2ojEvVoxykGMRsoEQnOrJMeLn3aK1dolc5BVomfkwrkqPfKX91+wrMYleWSGdPxaWqDCdNWcInTUjczmDI+YgPsOKpYjCaYzE+dknOn9EmUaFfKkrn6e2LCYmPGceg6Y2aHZtmbif95ncxGN8FEqDSzqPhiUZRJYhMy+5v0hUZu5dgRxrVwtxI+ZJpx69IpuRD85ZdXSfOq6tOqf08rtbM8jiKcwClcgA/XUIM7qEMDOAzgGV7hzZPei/fufSxaC14+cwx/4H3+AEevjQQ=</latexit> sha1_base64="88ZukDOpZjqFoMNUke8ECvdzrD8=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquLFkutGNWFDEZUV7gXYomTTThmYyQ5IRytBHcONCEbc+gY/iznfwAVyaXhba+kPg4//PIeccPxZcG4w/nczC4tLySnY1t7a+sbmV396p6ShRlFVpJCLV8IlmgktWNdwI1ogVI6EvWN3vX4zy+j1Tmkfyzgxi5oWkK3nAKTHWuj12cTtfwEU8FpoHdwqF86/Ls6v3/nelnf9odSKahEwaKojWTRfHxkuJMpwKNsy1Es1iQvuky5oWJQmZ9tLxqEN0aJ0OCiJlnzRo7P7uSEmo9SD0bWVITE/PZiPzv6yZmODUS7mME8MknXwUJAKZCI32Rh2uGDViYIFQxe2siPaIItTY6+TsEdzZleehViq6uOje4EL5ACbKwh7swxG4cAJluIYKVIFCFx7gCZ4d4Tw6L87rpDTjTHt24Y+ctx+xd5EL</latexit> 0 10 <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> 0 <latexit sha1_base64="ze+1l/6iGkHT3xP4qpdARxj2xMI=">AAAB7XicbZDLSgMxFIbP1Futt6pLN8EquCoZN+rKghuXFewF2qFk0kwbm0mGJCOUoe/gxoUibn0fd+7c+hSaXhba+kPg4//PIeecMBHcWIw/vNzS8srqWn69sLG5tb1T3N2rG5VqympUCaWbITFMcMlqllvBmolmJA4Fa4SDq3HeuGfacCVv7TBhQUx6kkecEuuseptZ0sGdYgmX8URoEfwZlC6/vy4+AaDaKb63u4qmMZOWCmJMy8eJDTKiLaeCjQrt1LCE0AHpsZZDSWJmgmwy7QgdO6eLIqXdkxZN3N8dGYmNGcahq4yJ7Zv5bGz+l7VSG50HGZdJapmk04+iVCCr0Hh11OWaUSuGDgjV3M2KaJ9oQq07UMEdwZ9feRHqp2Ufl/0bXKocwVR5OIBDOAEfzqAC11CFGlC4gwd4gmdPeY/ei/c6Lc15s559+CPv7QeJCJH5</latexit> sha1_base64="dS2N130qxsYrry7G8r/PP74zoGU=">AAAB7XicbVA9SwNBEJ3zM8avqKXNYRSswp6NlgEbywjmA5Ij7G32kjV7u8funBCO/AcbC0Vs/T92/hs3yRWa+GDg8d4MM/OiVAqLhHx7a+sbm1vbpZ3y7t7+wWHl6LhldWYYbzIttelE1HIpFG+iQMk7qeE0iSRvR+Pbmd9+4sYKrR5wkvIwoUMlYsEoOqnV40j7pF+pkhqZw18lQUGqUKDRr3z1BpplCVfIJLW2G5AUw5waFEzyabmXWZ5SNqZD3nVU0YTbMJ9fO/UvnDLwY21cKfTn6u+JnCbWTpLIdSYUR3bZm4n/ed0M45swFyrNkCu2WBRn0kftz173B8JwhnLiCGVGuFt9NqKGMnQBlV0IwfLLq6R1VQtILbgn1fp5EUcJTuEMLiGAa6jDHTSgCQwe4Rle4c3T3ov37n0sWte8YuYE/sD7/AEkyI65</latexit> sha1_base64="syNpdZ1daUwggsxpUJgHnz1nKE4=">AAAB7XicbZDLSgMxFIYzXmutWnXpJlgFVyXjRl1ZcOOygr1IW0omPdPGZpIhyQhl6Du4caGIW9/HnQsfQF9C08tCW38IfPz/OeScE8SCG0vIu7ewuLS8sppZy67nNja38ts7VaMSzaDClFC6HlADgkuoWG4F1GMNNAoE1IL+xSiv3YE2XMlrO4ihFdGu5CFn1Dqr2gRL26SdL5AiGQvPgz+Fwvn359nXx02u3M6/NTuKJRFIywQ1puGT2LZSqi1nAobZZmIgpqxPu9BwKGkEppWOpx3iQ+d0cKi0e9Lisfu7I6WRMYMocJURtT0zm43M/7JGYsPTVsplnFiQbPJRmAhsFR6tjjtcA7Ni4IAyzd2smPWopsy6A2XdEfzZleehelz0SdG/IoXSAZoog/bQPjpCPjpBJXSJyqiCGLpF9+gRPXnKe/CevZdJ6YI37dlFf+S9/gBFk5NF</latexit> ⌘0 10 <latexit sha1_base64="3t+4UbZpWdk262G+bJ4/hq+8w+E=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZN7qz4MZlFXuBdiiZNNOGJpkhyQhl6Bu4caFYV4IrX8edb2N6WWjrD4GP/z+HnHPCRHBjMf72ciura+sb+c3C1vbO7l5x/6Bu4lRTVqOxiHUzJIYJrljNcitYM9GMyFCwRji4nuSNB6YNj9W9HSYskKSneMQpsc6683GnWMJlPBVaBn8OpavP8fgdAKqd4le7G9NUMmWpIMa0fJzYICPacirYqNBODUsIHZAeazlURDITZNNJR+jUOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44BXcEf3HlZaifl31c9m9xqXICM+XhCI7hDHy4gArcQBVqQCGCR3iGF2/gPXmv3tusNOfNew7hj7yPH40aj4c=</latexit> sha1_base64="TBz0KWW1Y5FVHwrZZOy0b06gZmg=">AAAB6XicbVA9TwJBEJ3DL8Qv1NJmI5pYkT0bLUlsLNEIksCF7C17sGFv77I7Z0Iu/AMbC42x9R/Z+W9c4AoFXzLJy3szmZkXpkpapPTbK62tb2xulbcrO7t7+wfVw6O2TTLDRYsnKjGdkFmhpBYtlKhEJzWCxaESj+H4ZuY/PgljZaIfcJKKIGZDLSPJGTrp3qf9ao3W6RxklfgFqUGBZr/61RskPIuFRq6YtV2fphjkzKDkSkwrvcyKlPExG4quo5rFwgb5/NIpOXfKgESJcaWRzNXfEzmLrZ3EoeuMGY7ssjcT//O6GUbXQS51mqHQfLEoyhTBhMzeJgNpBEc1cYRxI92thI+YYRxdOBUXgr/88ippX9Z9WvfvaK1xVsRRhhM4hQvw4QoacAtNaAGHCJ7hFd68sffivXsfi9aSV8wcwx94nz/eLIzN</latexit> sha1_base64="XaQViiIp3HEbSwArRCfYKY95vuc=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXjRncW3LisYi/QDiWTZtrQJDMkGaEMfQM3LhTt1pWv4863MdN2oa0/BD7+/xxyzgkTwY3F+NtbWV1b39gsbBW3d3b39ksHhw0Tp5qyOo1FrFshMUxwxeqWW8FaiWZEhoI1w+FNnjcfmTY8Vg92lLBAkr7iEafEOuvex91SGVfwVGgZ/DmUrz/fc01q3dJXpxfTVDJlqSDGtH2c2CAj2nIq2LjYSQ1LCB2SPms7VEQyE2TTScfozDk9FMXaPWXR1P3dkRFpzEiGrlISOzCLWW7+l7VTG10FGVdJapmis4+iVCAbo3xt1OOaUStGDgjV3M2K6IBoQq07TtEdwV9ceRkaFxUfV/w7XK6ewkwFOIYTOAcfLqEKt1CDOlCI4Ale4NUbes/emzeZla54854j+CPv4wfqV5FM</latexit> 20 <latexit sha1_base64="QemqZr6fsruN4EG6fJtLmwgab9A=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXTje4suHFZxV6gHUomzbShSWZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wwEdxYjL+9ldW19Y3N3FZ+e2d3b79wcFg3caopq9FYxLoZEsMEV6xmuRWsmWhGZChYIxxcT/LGA9OGx+reDhMWSNJTPOKUWGfdlXGnUMQlPBVaBn8OxavP8fgdAKqdwle7G9NUMmWpIMa0fJzYICPacirYKN9ODUsIHZAeazlURDITZNNJR+jMOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44eXcEf3HlZaiXSz4u+be4WDmFmXJwDCdwDj5cQAVuoAo1oBDBIzzDizfwnrxX721WuuLNe47gj7yPH46fj4g=</latexit> sha1_base64="LOhMDFdaD7Ni8csKulxBbULGOyw=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxiPiA5wt5mL1myt3fszgkh5B/YWChi6z+y89+4Sa7QxAcDj/dmmJkXpkpapPTbK2xsbm3vFHdLe/sHh0fl45OWTTLDRZMnKjGdkFmhpBZNlKhEJzWCxaES7XB8O/fbT8JYmehHnKQiiNlQy0hyhk56qNF+uUKrdAGyTvycVCBHo1/+6g0SnsVCI1fM2q5PUwymzKDkSsxKvcyKlPExG4quo5rFwgbTxaUzcumUAYkS40ojWai/J6YstnYSh64zZjiyq95c/M/rZhjdBFOp0wyF5stFUaYIJmT+NhlIIziqiSOMG+luJXzEDOPowim5EPzVl9dJq1b1adW/p5X6RR5HEc7gHK7Ah2uowx00oAkcIniGV3jzxt6L9+59LFsLXj5zCn/gff4A37GMzg==</latexit> sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiBU7gADy6hBrdQhwZQCOEJXuDVGTnPzpsznZcWnEXPMfyR8/ED69yRTQ==</latexit> FIG. 4. Continuation of spatially periodic patterns. A bifurcation diagram in η0 of two stationary patterned states, with different wavenumbers. Each branch of patterned states are connected to the branch of homogeneous states via a subcritical Turing bifurcations and reconnect to the steady state within the region of instability. Only the rightmost and left most Turing points correspond to a change in stability of the system. The homogeneous steady state undergoes a Hopf bifurcation at η0 = 3.298 as expected from the instability analysis in §III. More interestingly, we also see Hopf bifurcations along the branch of periodic solutions. Solid (dashed) lines represent stable (unstable) solutions. Parameter values: vsyn = 15, Λ = 12π, other parameters as in Fig. 2(a). stationary patterns by solving for (z, g1 , g2 ) the boundary value problem F(z) + G1 (z, g1 ) + +G2 (z, g2 ) = 0, 2 (Λ2 − ∂xx )g1 − Λ2 κ1 f (z) = 0, 2 (33) 2 (Λ − ∂xx /β )g2 − Λ κ2 f (z) = 0, with Neumann boundary conditions. Wavetrains are solutions to (32) with z(x, t) = Z(x − ct), g1 (x, t) = G1 (x−ct) and g2 (x, t) = G2 (x−ct) for c ∈ R, where Z(ξ), G1 (ξ) and G2 (ξ) are Λ-periodic. To compute wavetrains, we solve for (Z, G1 , G2 , c) the boundary value problem c∂ξ z + Λ F(Z) + G1 (Z, G1 ) + G2 (Z, G2 ) = 0, (Λ2 − ∂ξξ )(1 − cτ ∂ξ )2 G1 − Λ2 κf (Z) = 0, (Λ2 − ∂ξξ /β 2 )(1 − cτ ∂ξ )2 G2 − Λ2 κf (Z) = 0, ψ(Z, G1 , G2 ) = 0, (34) posed on ξ ∈ [−1/2, 1/2] with periodic boundary conditions. The last equation in (33) and (34) is a standard phase condition [23] 1/2 dξ d Z(ξ) Z(ξ) − Z(ξ) dξ dξ d G1 (ξ) G(ξ) − G1 (ξ) dξ dξ ψ(Z, G1 , G2 ) = d G2 (ξ) G(ξ) − G2 (ξ) , dξ −1/2 1/2 + −1/2 1/2 + −1/2 (35) Turing patterns We first analyse the stationary patterns seen in §III, using numerical continuation to verify the analytical results, determine the criticality of the Turing bifurcation, and examine the behaviour of these solutions away from the onset of the instability. We continued solutions to the boundary value problem (33) in the parameter η0 with Λ = 12π, vsyn = 15 and all other parameters as in Fig. 2(a). This corresponds to making a vertical excursion through the 2-parameter bifurcation diagram Fig. 2(a). Solving (27) -- (28) we found two Turing bifurcations at η0 = −0.648, kc = 0.738 and η0 = 12.67, kc = 0.969. Hence, we numerically continued patterns states with k = 0.738 (blue) and k = 0.969 (red) (Fig. 4). As expected, the homogeneous steady state bifurcates to patterns at η0 = −0.648 (blue) and η0 = 12.67 (red), corresponding to the Turing bifurcations found in §III. The stability of the patterned states, as well as the homogeneous state, was numerically calculated along the continuation branch. Along with the Turing bifurcations, the homogeneous steady state undergoes a Hopf bifurcation at η0 = 3.298, which matches the value found analytically. The patterned Turing solutions were found to go unstable to a globally oscillating periodic pattern, through a Hopf bifurcation at η0 = 2.8380 (blue) and η0 = 5.8546 (red). The patterned solutions are also unstable (through a symmetry breaking bifurcation) to broader patterns at low values of η0 .As anticipated, the bifurcations from the homogeneous state steady state are subcritical, hence we have bistability between a spatial pattern and the homogeneous state and the bistability region occurs in a wide region of parameter space. Further numerical continuation results (not shown) indicate that the region of bi-stability increases as the reversal potential vsyn is increased. As vsyn decreases towards 0, the static Turing bifurcation points collide, and the patterned states cease to exist, as predicted from the Turing analysis in §III. We found the scenario presented above to be robust to changes in the other parameters. B. Wavetrains We now shift our focus to wavetrain solutions originating at a Turing-Hopf bifurcation of the homogeneous steady state. We recall that we find these states as solutions to (34) with periodic boundary conditions, hence Λ corresponds to the spatial period of the wavetrain profile. In addition, the phase velocity c of a wavetrain is accessible from the boundary-value problem solution. 9 (a) 10 (b) <latexit sha1_base64="ka3t2/48Wh23hsAl88gQOgU1RLk=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqMQm3CXRsuAjWVE8wHJEfY2e8mSvb1jd04IR36CjYUitv4iO/+Nm+QKTXww8Hhvhpl5QSKFQdf9dgobm1vbO8Xd0t7+weFR+fikbeJUM95isYx1N6CGS6F4CwVK3k00p1EgeSeY3M79zhPXRsTqEacJ9yM6UiIUjKKVHqr0alCuuDV3AbJOvJxUIEdzUP7qD2OWRlwhk9SYnucm6GdUo2CSz0r91PCEsgkd8Z6likbc+Nni1Bm5tMqQhLG2pZAs1N8TGY2MmUaB7Ywojs2qNxf/83ophjd+JlSSIldsuShMJcGYzP8mQ6E5Qzm1hDIt7K2EjammDG06JRuCt/ryOmnXa55b8+7rlcZFHkcRzuAcquDBNTTgDprQAgYjeIZXeHOk8+K8Ox/L1oKTz5zCHzifP38FjSo=</latexit> 20 <latexit sha1_base64="QemqZr6fsruN4EG6fJtLmwgab9A=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXTje4suHFZxV6gHUomzbShSWZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wwEdxYjL+9ldW19Y3N3FZ+e2d3b79wcFg3caopq9FYxLoZEsMEV6xmuRWsmWhGZChYIxxcT/LGA9OGx+reDhMWSNJTPOKUWGfdlXGnUMQlPBVaBn8OxavP8fgdAKqdwle7G9NUMmWpIMa0fJzYICPacirYKN9ODUsIHZAeazlURDITZNNJR+jMOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44eXcEf3HlZaiXSz4u+be4WDmFmXJwDCdwDj5cQAVuoAo1oBDBIzzDizfwnrxX721WuuLNe47gj7yPH46fj4g=</latexit> sha1_base64="LOhMDFdaD7Ni8csKulxBbULGOyw=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxiPiA5wt5mL1myt3fszgkh5B/YWChi6z+y89+4Sa7QxAcDj/dmmJkXpkpapPTbK2xsbm3vFHdLe/sHh0fl45OWTTLDRZMnKjGdkFmhpBZNlKhEJzWCxaES7XB8O/fbT8JYmehHnKQiiNlQy0hyhk56qNF+uUKrdAGyTvycVCBHo1/+6g0SnsVCI1fM2q5PUwymzKDkSsxKvcyKlPExG4quo5rFwgbTxaUzcumUAYkS40ojWai/J6YstnYSh64zZjiyq95c/M/rZhjdBFOp0wyF5stFUaYIJmT+NhlIIziqiSOMG+luJXzEDOPowim5EPzVl9dJq1b1adW/p5X6RR5HEc7gHK7Ah2uowx00oAkcIniGV3jzxt6L9+59LFsLXj5zCn/gff4A37GMzg==</latexit> sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiBU7gADy6hBrdQhwZQCOEJXuDVGTnPzpsznZcWnEXPMfyR8/ED69yRTQ==</latexit> 8 <latexit sha1_base64="RX0r+KSeKWVCDBmt2yVKtPMLU2A=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmzkU4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJih3Uy7jxKBk84+CRBATkWxr0ucKmRFjC5QpbmclbEgVZcbeJm+P4C2vvAqNq5LnlryaW6ycw1w5OIUzuAQPrqECd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QOHHpEZ</latexit> <latexit sha1_base64="TWl08uTjLS78G4SobYD0XsFYmto=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdi1MZ0BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzVh2c+4TFKDks0/ClNBTEymW5MeV8iMGFmgTHE7K2EDqigz9jZ5ewRveeVVaFyVPLfk1dxi5RzmysEpnMEleHANFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fKeGPVA==</latexit> sha1_base64="5jynEd0nF7EC3wxivqcZY9TrYJs=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbKElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg/LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NWHdz7hMUoOSrRaFqSAmJouvyYgrZEbMLKFMcXsrYROqKDM2m5INwVt/eZN0bqueW/VabqVxlcdRhAu4hBvwoAYNuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9684ya</latexit> (c) <latexit sha1_base64="8MdXaT6mU7jtdNwNdCpBw6ai3X4=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqMQm3CXRsuAjWVE8wHJEfY2e8mSvb1jd04IR36CjYUitv4iO/+Nm+QKTXww8Hhvhpl5QSKFQdf9dgobm1vbO8Xd0t7+weFR+fikbeJUM95isYx1N6CGS6F4CwVK3k00p1EgeSeY3M79zhPXRsTqEacJ9yM6UiIUjKKVHqrB1aBccWvuAmSdeDmpQI7moPzVH8YsjbhCJqkxPc9N0M+oRsEkn5X6qeEJZRM64j1LFY248bPFqTNyaZUhCWNtSyFZqL8nMhoZM40C2xlRHJtVby7+5/VSDG/8TKgkRa7YclGYSoIxmf9NhkJzhnJqCWVa2FsJG1NNGdp0SjYEb/XlddKu1zy35t3XK42LPI4inME5VMGDa2jAHTShBQxG8Ayv8OZI58V5dz6WrQUnnzmFP3A+fwCAio0r</latexit> <latexit sha1_base64="3t+4UbZpWdk262G+bJ4/hq+8w+E=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZN7qz4MZlFXuBdiiZNNOGJpkhyQhl6Bu4caFYV4IrX8edb2N6WWjrD4GP/z+HnHPCRHBjMf72ciura+sb+c3C1vbO7l5x/6Bu4lRTVqOxiHUzJIYJrljNcitYM9GMyFCwRji4nuSNB6YNj9W9HSYskKSneMQpsc6683GnWMJlPBVaBn8OpavP8fgdAKqd4le7G9NUMmWpIMa0fJzYICPacirYqNBODUsIHZAeazlURDITZNNJR+jUOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44BXcEf3HlZaifl31c9m9xqXICM+XhCI7hDHy4gArcQBVqQCGCR3iGF2/gPXmv3tusNOfNew7hj7yPH40aj4c=</latexit> sha1_base64="TBz0KWW1Y5FVHwrZZOy0b06gZmg=">AAAB6XicbVA9TwJBEJ3DL8Qv1NJmI5pYkT0bLUlsLNEIksCF7C17sGFv77I7Z0Iu/AMbC42x9R/Z+W9c4AoFXzLJy3szmZkXpkpapPTbK62tb2xulbcrO7t7+wfVw6O2TTLDRYsnKjGdkFmhpBYtlKhEJzWCxaESj+H4ZuY/PgljZaIfcJKKIGZDLSPJGTrp3qf9ao3W6RxklfgFqUGBZr/61RskPIuFRq6YtV2fphjkzKDkSkwrvcyKlPExG4quo5rFwgb5/NIpOXfKgESJcaWRzNXfEzmLrZ3EoeuMGY7ssjcT//O6GUbXQS51mqHQfLEoyhTBhMzeJgNpBEc1cYRxI92thI+YYRxdOBUXgr/88ippX9Z9WvfvaK1xVsRRhhM4hQvw4QoacAtNaAGHCJ7hFd68sffivXsfi9aSV8wcwx94nz/eLIzN</latexit> sha1_base64="XaQViiIp3HEbSwArRCfYKY95vuc=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXjRncW3LisYi/QDiWTZtrQJDMkGaEMfQM3LhTt1pWv4863MdN2oa0/BD7+/xxyzgkTwY3F+NtbWV1b39gsbBW3d3b39ksHhw0Tp5qyOo1FrFshMUxwxeqWW8FaiWZEhoI1w+FNnjcfmTY8Vg92lLBAkr7iEafEOuvex91SGVfwVGgZ/DmUrz/fc01q3dJXpxfTVDJlqSDGtH2c2CAj2nIq2LjYSQ1LCB2SPms7VEQyE2TTScfozDk9FMXaPWXR1P3dkRFpzEiGrlISOzCLWW7+l7VTG10FGVdJapmis4+iVCAbo3xt1OOaUStGDgjV3M2K6IBoQq07TtEdwV9ceRkaFxUfV/w7XK6ewkwFOIYTOAcfLqEKt1CDOlCI4Ale4NUbes/emzeZla54854j+CPv4wfqV5FM</latexit> 16 <latexit sha1_base64="0LyhbvMZEh6YU73egR9mzvArekA=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqMQm3CXRsuAjWVE8wHJEfY2e8mSvb1jd04IR36CjYUitv4iO/+Nm+QKTXww8Hhvhpl5QSKFQdf9dgobm1vbO8Xd0t7+weFR+fikbeJUM95isYx1N6CGS6F4CwVK3k00p1EgeSeY3M79zhPXRsTqEacJ9yM6UiIUjKKVHqrsalCuuDV3AbJOvJxUIEdzUP7qD2OWRlwhk9SYnucm6GdUo2CSz0r91PCEsgkd8Z6likbc+Nni1Bm5tMqQhLG2pZAs1N8TGY2MmUaB7Ywojs2qNxf/83ophjd+JlSSIldsuShMJcGYzP8mQ6E5Qzm1hDIt7K2EjammDG06JRuCt/ryOmnXa55b8+7rlcZFHkcRzuAcquDBNTTgDprQAgYjeIZXeHOk8+K8Ox/L1oKTz5zCHzifP4IPjSw=</latexit> ⇤ = ⇡/k ⇤ = 2⇡/k ⇤ = 4⇡/k Hop Tu ng Hop Hop 16 <latexit sha1_base64="VOoKRtLZHEXx8Q5B69L4Znh8o9U=">AAAB6XicbZC7SgNBFIbPeo3xFrW0GYyCVdi1UDsDNpZRzAWSJcxOZpMhs7PLzFkhLHkDGwvFWAlWvo6db+PkUmjiDwMf/38Oc84JEikMuu63s7S8srq2ntvIb25t7+wW9vZrJk4141UWy1g3Amq4FIpXUaDkjURzGgWS14P+9TivP3BtRKzucZBwP6JdJULBKFrrzjtvF4puyZ2ILII3g+LV52j0DgCVduGr1YlZGnGFTFJjmp6boJ9RjYJJPsy3UsMTyvq0y5sWFY248bPJpENyYp0OCWNtn0IycX93ZDQyZhAFtjKi2DPz2dj8L2umGF76mVBJilyx6UdhKgnGZLw26QjNGcqBBcq0sLMS1qOaMrTHydsjePMrL0LtrOS5Je/WLZaPYaocHMIRnIIHF1CGG6hAFRiE8AjP8OL0nSfn1Xmbli45s54D+CPn4weWMo+N</latexit> sha1_base64="vJGKWfYypACzq1YKgzLOrYsRbkA=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0k8qMeCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jNs1BWx8MPN6bYWZekEhh0HW/ndLa+sbmVnm7srO7t39QPTxqmzjVjLdYLGPdDajhUijeQoGSdxPNaRRI3gkmt3O/88S1EbF6xGnC/YiOlAgFo2ilB+9qUK25dTcHWSVeQWpQoDmofvWHMUsjrpBJakzPcxP0M6pRMMlnlX5qeELZhI54z1JFI278LL90Rs6tMiRhrG0pJLn6eyKjkTHTKLCdEcWxWfbm4n9eL8Xwxs+ESlLkii0WhakkGJP522QoNGcop5ZQpoW9lbAx1ZShDadiQ/CWX14l7cu659a9e7fWOCviKMMJnMIFeHANDbiDJrSAQQjP8ApvzsR5cd6dj0VrySlmjuEPnM8f50SM0w==</latexit> sha1_base64="QjWfb9jF5DpSDol7iRWsNBO7stA=">AAAB6XicbZC7SgNBFIbPxluMt6ilzWAUrMKuhdoZsLGMYi6QLGF2cjYZMju7zMwKYckb2FgomtbK17HzbZxcCk38YeDj/89hzjlBIrg2rvvt5FZW19Y38puFre2d3b3i/kFdx6liWGOxiFUzoBoFl1gz3AhsJgppFAhsBIObSd54RKV5LB/MMEE/oj3JQ86osda9d9EpltyyOxVZBm8OpevP94nG1U7xq92NWRqhNExQrVuemxg/o8pwJnBUaKcaE8oGtIcti5JGqP1sOumInFqnS8JY2ScNmbq/OzIaaT2MAlsZUdPXi9nE/C9rpSa88jMuk9SgZLOPwlQQE5PJ2qTLFTIjhhYoU9zOSlifKsqMPU7BHsFbXHkZ6udlzy17d26pcgIz5eEIjuEMPLiECtxCFWrAIIQneIFXZ+A8O2/OeFaac+Y9h/BHzscP82+RUg==</latexit> 12 6 kg1 k1 <latexit sha1_base64="m24jJNLOQa0iEnae+s0oEb657D4=">AAAB6XicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8so5gLJEmYns8mQ2dll5qwQlryBjYVirAQrX8fOt3FyKTTxh4GP/z+HOecEiRQGXffbWVldW9/YzG3lt3d29/YLB4d1E6ea8RqLZaybATVcCsVrKFDyZqI5jQLJG8HgepI3Hrg2Ilb3OEy4H9GeEqFgFK1155U7haJbcqciy+DNoXj1OR6/A0C1U/hqd2OWRlwhk9SYlucm6GdUo2CSj/Lt1PCEsgHt8ZZFRSNu/Gw66YicWadLwljbp5BM3d8dGY2MGUaBrYwo9s1iNjH/y1ophpd+JlSSIlds9lGYSoIxmaxNukJzhnJogTIt7KyE9ammDO1x8vYI3uLKy1Avlzy35N26xcopzJSDYziBc/DgAipwA1WoAYMQHuEZXpyB8+S8Om+z0hVn3nMEf+R8/ACQIo+J</latexit> sha1_base64="kyPywSf8BeFi+Z8wSsY7iGTmWbg=">AAAB6XicbZA9TwJBEIbn8AvxC7W02YgmVuSORjtJbCzRyEcCF7K37MGGvb3L7pwJIfwDGwuN0lr5d+z8N+4BhYJvssmT953JzkyQSGHQdb+d3Nr6xuZWfruws7u3f1A8PGqYONWM11ksY90KqOFSKF5HgZK3Es1pFEjeDIY3Wd585NqIWD3gKOF+RPtKhIJRtNa9V+kWS27ZnYmsgreA0vXne6ZprVv86vRilkZcIZPUmLbnJuiPqUbBJJ8UOqnhCWVD2udti4pG3Pjj2aQTcm6dHgljbZ9CMnN/d4xpZMwoCmxlRHFglrPM/C9rpxhe+WOhkhS5YvOPwlQSjEm2NukJzRnKkQXKtLCzEjagmjK0xynYI3jLK69Co1L23LJ355aqZzBXHk7gFC7Ag0uowi3UoA4MQniCF3h1hs6z8+ZM56U5Z9 <latexit sha1_base64="p01vSfJLyQ7pt6DtMz/RCNM7Mb4=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTOQu0ksbGERD4SuJC9ZQ5W9vYuu3smhPALbCw0Bks7/46d/8Y9oFD0TTZ58r4z2ZkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DhBaRFw==</latexit> <latexit sha1_base64="Qe7te+9YtxHKH1q/XgFz9DrYbZE=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuhdoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSbZj1I=</latexit> sha1_base64="qqlJvqD69kDjRh93ah+hTYFl6dI=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFqNgFe4somXAxjIB8wHJEfY2c8mavb1jd08IR36BjYUitv4kO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7ASTu7nfeUKleSwfzDRBP6IjyUPOqLFSszYoV9yquwBZJ15OKpCjMSh/9YcxSyOUhgmqdc9zE+NnVBnOBM5K/VRjQtmEjrBnqaQRaj9bHDojl1YZkjBWtqQhC/X3REYjradRYDsjasZ61ZuL/3m91IS3fsZlkhqUbLkoTAUxMZl/TYZcITNiagllittbCRtTRZmx2ZRsCN7qy+ukfV313KrXdCv1izyOIpzBOVyBBzdQh3toQAsYIDzDK7w5j86L8+58LFsLTj5zCn/gfP4Ad+uMmA==</latexit> 12 c <latexit sha1_base64="m24jJNLOQa0iEnae+s0oEb657D4=">AAAB6XicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8so5gLJEmYns8mQ2dll5qwQlryBjYVirAQrX8fOt3FyKTTxh4GP/z+HOecEiRQGXffbWVldW9/YzG3lt3d29/YLB4d1E6ea8RqLZaybATVcCsVrKFDyZqI5jQLJG8HgepI3Hrg2Ilb3OEy4H9GeEqFgFK1155U7haJbcqciy+DNoXj1OR6/A0C1U/hqd2OWRlwhk9SYlucm6GdUo2CSj/Lt1PCEsgHt8ZZFRSNu/Gw66YicWadLwljbp5BM3d8dGY2MGUaBrYwo9s1iNjH/y1ophpd+JlSSIlds9lGYSoIxmaxNukJzhnJogTIt7KyE9ammDO1x8vYI3uLKy1Avlzy35N26xcopzJSDYziBc/DgAipwA1WoAYMQHuEZXpyB8+S8Om+z0hVn3nMEf+R8/ACQIo+J</latexit> sha1_base64="b3xfLwgBQyq/rB87I5MS71An4JA=">AAAB6XicbVBNT8JAEJ3iF+IX6tHLRjTxRFoueiTx4hGNBRJoyHbZwobtttmdmpCGf+DFg8Z49R9589+4QA8KvmSSl/dmMjMvTKUw6LrfTmljc2t7p7xb2ds/ODyqHp+0TZJpxn2WyER3Q2q4FIr7KFDybqo5jUPJO+Hkdu53nrg2IlGPOE15ENOREpFgFK304DUG1Zpbdxcg68QrSA0KtAbVr/4wYVnMFTJJjel5bopBTjUKJvms0s8MTymb0BHvWapozE2QLy6dkUurDEmUaFsKyUL9PZHT2JhpHNrOmOLYrHpz8T+vl2F0E+RCpRlyxZaLokwSTMj8bTIUmjOUU0so08LeStiYasrQhlOxIXirL6+TdqPuuXXv3q01L4o4ynAG53AFHlxDE+6gBT4wiOAZXuHNmTgvzrvzsWwtOcXMKfyB8/kD4TSMzw==</latexit> sha1_base64="kyPywSf8BeFi+Z8wSsY7iGTmWbg=">AAAB6XicbZA9TwJBEIbn8AvxC7W02YgmVuSORjtJbCzRyEcCF7K37MGGvb3L7pwJIfwDGwuN0lr5d+z8N+4BhYJvssmT953JzkyQSGHQdb+d3Nr6xuZWfruws7u3f1A8PGqYONWM11ksY90KqOFSKF5HgZK3Es1pFEjeDIY3Wd585NqIWD3gKOF+RPtKhIJRtNa9V+kWS27ZnYmsgreA0vXne6ZprVv86vRilkZcIZPUmLbnJuiPqUbBJJ8UOqnhCWVD2udti4pG3Pjj2aQTcm6dHgljbZ9CMnN/d4xpZMwoCmxlRHFglrPM/C9rpxhe+WOhkhS5YvOPwlQSjEm2NukJzRnKkQXKtLCzEjagmjK0xynYI3jLK69Co1L23LJ355aqZzBXHk7gFC7Ag0uowi3UoA4MQniCF3h1hs6z8+ZM56U5Z9FzDH/kfPwA7V+RTg==</latexit> <latexit sha1_base64="z47r0npwszHz2/UHFt3Oq9Oi4sE=">AAAB/3icbZA9SwNBEIbnNGqMX1HBxmYxClbhzkbLgI1lBPMByXHubfaSJXt7x+6ccMQU/hUbC0Vs/Rt2/hs3H4UmvrDw8M4MM/uGqRQGXffbWVktrK1vFDdLW9s7u3vl/YOmSTLNeIMlMtHtkBouheINFCh5O9WcxqHkrXB4Pam3Hrg2IlF3mKfcj2lfiUgwitYKykfdJtdI+oFHphR0hYowD8oVt+pORZbBm0OlVoiiewCoB+Wvbi9hWcwVMkmN6Xhuiv6IahRM8nGpmxmeUjakfd6xqGjMjT+a3j8mZ9bpkSjR9ikkU/f3xIjGxuRxaDtjigOzWJuY/9U6GUZX/kioNEOu2GxRlEmCCZmEQXpCc4Yyt0CZFvZWwgZUU4Y2spINwVv88jI0L6qeW/VubRqnMFMRjuEEzsGDS6jBDdShAQwe4Rle4c15cl6cd+dj1rrizGcO4Y+czx8qnZbr</latexit> sha1_base64="Wpyl60zfUfdviD9FUEBemg951cw=">AAAB/3icbZDNSsNAFIUn9a/Wv6jgxs1gFVyVxI0uC25cVrCt0IQwmU7aoZNJmLkRQuzCV3HjQhG3voY738ZpmoW2Hhj4OPde7p0TpoJrcJxvq7ayura+Ud9sbG3v7O7Z+wc9nWSKsi5NRKLuQ6KZ4JJ1gYNg96liJA4F64eT61m9/8CU5om8gzxlfkxGkkecEjBWYB95PaYAjwIXlxR4XEaQB3bTaTml8DK4FTRRpU5gf3nDhGYxk0AF0XrgOin4BVHAqWDThpdplhI6ISM2MChJzLRflPdP8ZlxhjhKlHkScOn+nihIrHUeh6YzJjDWi7WZ+V9tkEF05RdcphkwSeeLokxgSPAsDDzkilEQuQFCFTe3YjomilAwkTVMCO7il5ehd9FynZZ76zTbp1UcdXSMTtA5ctElaqMb1EFdRNEjekav6M16sl6sd+tj3lqzqplD9EfW5w85wZV7</latexit> sha1_base64="GA6ely70JSOYBIcvitrigLFlMo8=">AAAB/3icbZDNSsNAFIUntWqtf1HBjZvBKrgqiRtdFty4rGB/oAlhMp20QyeTMHMjhNiFr+LGhSJufQ13vo3TtAttPTDwce693DsnTAXX4DjfVmWtur6xWduqb+/s7u3bB4ddnWSKsg5NRKL6IdFMcMk6wEGwfqoYiUPBeuHkZlbvPTCleSLvIU+ZH5OR5BGnBIwV2MdelynAo8DFJQUelxHkgd1wmk4pvAruAhqtalSqHdhf3jChWcwkUEG0HrhOCn5BFHAq2LTuZZqlhE7IiA0MShIz7Rfl/VN8bpwhjhJlngRcur8nChJrnceh6YwJjPVybWb+VxtkEF37BZdpBkzS+aIoExgSPAsDD7liFERugFDFza2YjokiFExkdROCu/zlVeheNl2n6d6ZNM7QXDV0gk7RBXLRFWqhW9RGHUTRI3pGr+jNerJerHfrY95asRYzR+iPrM8fy4iYIw==</latexit> 4 <latexit sha1_base64="TEM37w6PcFQLCBxv+NzlDSPPh4U=">AAAB6HicbZA9SwNBEIbn4leMX1FLm8UoWIU7EbQzYGOZgPmA5Ah7m7lkzd7esbsnhJBfYGOhSCzt/Dt2/hv3khQafWHh4X1n2JkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DgQ6RFQ==</latexit> <latexit sha1_base64="zDS/d1EMLhdhjGncUr18mp+9fB4=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuCNoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSPRj1A=</latexit> sha1_base64="G9IPPnJCtUtisXfD2y/+tp8gcfA=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipWRuUK27VXYCsEy8nFcjRGJS/+sOYpRFKwwTVuue5ifEzqgxnAmelfqoxoWxCR9izVNIItZ8tDp2RS6sMSRgrW9KQhfp7IqOR1tMosJ0RNWO96s3F/7xeasJbP+MySQ1KtlwUpoKYmMy/JkOukBkxtYQyxe2thI2poszYbEo2BG/15XXSvq56btVrupX6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AdOOMlg==</latexit> <latexit sha1_base64="RX0r+KSeKWVCDBmt2yVKtPMLU2A=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmzkU4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJih3Uy7jxKBk84+CRBATkWxr0ucKmRFjC5QpbmclbEgVZcbeJm+P4C2vvAqNq5LnlryaW6ycw1w5OIUzuAQPrqECd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QOHHpEZ</latexit> <latexit sha1_base64="TWl08uTjLS78G4SobYD0XsFYmto=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdi1MZ0BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzVh2c+4TFKDks0/ClNBTEymW5MeV8iMGFmgTHE7K2EDqigz9jZ5ewRveeVVaFyVPLfk1dxi5RzmysEpnMEleHANFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fKeGPVA==</latexit> sha1_base64="5jynEd0nF7EC3wxivqcZY9TrYJs=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbKElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg/LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NWHdz7hMUoOSrRaFqSAmJouvyYgrZEbMLKFMcXsrYROqKDM2m5INwVt/eZN0bqueW/VabqVxlcdRhAu4hBvwoAYNuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9684ya</latexit> <latexit sha1_base64="RX0r+KSeKWVCDBmt2yVKtPMLU2A=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmzkU4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJih3Uy7jxKBk84+CRBATkWxr0ucKmRFjC5QpbmclbEgVZcbeJm+P4C2vvAqNq5LnlryaW6ycw1w5OIUzuAQPrqECd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QOHHpEZ</latexit> <latexit sha1_base64="TWl08uTjLS78G4SobYD0XsFYmto=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdi1MZ0BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzVh2c+4TFKDks0/ClNBTEymW5MeV8iMGFmgTHE7K2EDqigz9jZ5ewRveeVVaFyVPLfk1dxi5RzmysEpnMEleHANFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fKeGPVA==</latexit> sha1_base64="5jynEd0nF7EC3wxivqcZY9TrYJs=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbKElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg/LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NWHdz7hMUoOSrRaFqSAmJouvyYgrZEbMLKFMcXsrYROqKDM2m5INwVt/eZN0bqueW/VabqVxlcdRhAu4hBvwoAYNuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9684ya</latexit> ⇤ = ⇡/k ⇤ = 2⇡/k ⇤ = 4⇡/k Hop Tu ng Hop 2 <latexit sha1_base64="vV/tSZsqp0PZDp4B5kGJej5F2X0=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmj0U4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUa5ZLnlryaW6ycw1w5OIUzuAQPrqACd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QN+BpET</latexit> <latexit sha1_base64="NShdlnjkNDfv4ynZNqa/uWih2wQ=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0CiXPLfk1dxi5RzmysEpnMEleHAFFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fIMmPTg==</latexit> sha1_base64="uAhRiQUyUfrbUCj+oLzLI3Qe8O4=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VabqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9x24yU</latexit> <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> 0 c 8 <latexit sha1_base64="rCVsgeNE0e8YDyRFxTSTN12NQJo=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCwhkY8ELmRvmYOVvb3L7p4JIfwCGwuNwdLOv2Pnv3EPKBR8k02evO9MdmaCRHBtXPfbya2tb2xu5bcLO7t7+wfFw6OGjlPFsM5iEatWQDUKLrFuuBHYShTSKBDYDIa3Wd58RKV5LO/NKEE/on3JQ86osVaNdYslt+zORFbBW0Dp5nOa6b3aLX51ejFLI5SGCap123MT44+pMpwJnBQ6qcaEsiHtY9uipBFqfzwbdELOrdMjYazsk4bM3N8dYxppPYoCWxlRM9DLWWb+l7VTE177Yy6T1KBk84/CVBATk2xr0uMKmREjC5QpbmclbEAVZcbepmCP4C2vvAqNy7Lnlr2aW6qcwVx5OIFTuAAPrqACd1CFOjBAeIIXeHUenGfnzZnOS3POoucY/sj5+AHISpFE</latexit> <latexit sha1_base64="/LXIjfRVWR37uD2eTarUm8xNMK0=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMBdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmjiDwMf/38Oc84JEsG1cd1vJ7e2vrG5ld8u7Ozu7R8UD48aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeTvPmIyrNY3lvRgn6Ee1LHnJGjbVqrFssuWV3JrIK3gJKN5+TyTsAVLvFr04vZmmE0jBBtW57bmL8jCrDmcBxoZNqTCgb0j62LUoaofaz2aBjcm6dHgljZZ80ZOb+7shopPUoCmxlRM1AL2dT87+snZrw2s+4TFKDks0/ClNBTEymW5MeV8iMGFmgTHE7K2EDqigz9jYFewRveeVVaFyWPbfs1dxS5QzmysMJnMIFeHAFFbiDKtSBAcITvMCr8+A8O2/OZF6acxY9x/BHzscPaw2Pfw==</latexit> sha1_base64="7Ii0uNmP2RfOgM1BBfSq2ARZFww=">AAAB6HicbVBNT8JAEJ3iF+IX6tHLRjTxRFoveiTx4hESCyTQkO0yhZXtttndmpCGX+DFg8Z49Sd589+4QA8KvmSSl/dmMjMvTAXXxnW/ndLG5tb2Tnm3srd/cHhUPT5p6yRTDH2WiER1Q6pRcIm+4UZgN1VI41BgJ5zczf3OEyrNE/lgpikGMR1JHnFGjZVabFCtuXV3AbJOvILUoEBzUP3qDxOWxSgNE1TrnuemJsipMpwJnFX6mcaUsgkdYc9SSWPUQb44dEYurTIkUaJsSUMW6u+JnMZaT+PQdsbUjPWqNxf/83qZiW6DnMs0MyjZclGUCWISMv+aDLlCZsTUEsoUt7cSNqaKMmOzqdgQvNWX10n7uu65da/l1hoXRRxlOINzuAIPbqAB99AEHxggPMMrvDmPzovz7nwsW0tOMXMKf+B8/gC8H4zF</latexit> 10 0 10 20 ⌘0 30 8 <latexit sha1_base64="TEM37w6PcFQLCBxv+NzlDSPPh4U=">AAAB6HicbZA9SwNBEIbn4leMX1FLm8UoWIU7EbQzYGOZgPmA5Ah7m7lkzd7esbsnhJBfYGOhSCzt/Dt2/hv3khQafWHh4X1n2JkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DgQ6RFQ==</latexit> <latexit sha1_base64="zDS/d1EMLhdhjGncUr18mp+9fB4=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuCNoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSPRj1A=</latexit> sha1_base64="G9IPPnJCtUtisXfD2y/+tp8gcfA=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipWRuUK27VXYCsEy8nFcjRGJS/+sOYpRFKwwTVuue5ifEzqgxnAmelfqoxoWxCR9izVNIItZ8tDp2RS6sMSRgrW9KQhfp7IqOR1tMosJ0RNWO96s3F/7xeasJbP+MySQ1KtlwUpoKYmMy/JkOukBkxtYQyxe2thI2poszYbEo2BG/15XXSvq56btVrupX6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AdOOMlg==</latexit> <latexit sha1_base64="TEM37w6PcFQLCBxv+NzlDSPPh4U=">AAAB6HicbZA9SwNBEIbn4leMX1FLm8UoWIU7EbQzYGOZgPmA5Ah7m7lkzd7esbsnhJBfYGOhSCzt/Dt2/hv3khQafWHh4X1n2JkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DgQ6RFQ==</latexit> <latexit sha1_base64="zDS/d1EMLhdhjGncUr18mp+9fB4=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuCNoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSPRj1A=</latexit> sha1_base64="G9IPPnJCtUtisXfD2y/+tp8gcfA=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipWRuUK27VXYCsEy8nFcjRGJS/+sOYpRFKwwTVuue5ifEzqgxnAmelfqoxoWxCR9izVNIItZ8tDp2RS6sMSRgrW9KQhfp7IqOR1tMosJ0RNWO96s3F/7xeasJbP+MySQ1KtlwUpoKYmMy/JkOukBkxtYQyxe2thI2poszYbEo2BG/15XXSvq56btVrupX6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AdOOMlg==</latexit> 40 4 10 0 10 20 ⌘0 30 40 <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> 4 0 <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> 0 10 <latexit sha1_base64="3t+4UbZpWdk262G+bJ4/hq+8w+E=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZN7qz4MZlFXuBdiiZNNOGJpkhyQhl6Bu4caFYV4IrX8edb2N6WWjrD4GP/z+HnHPCRHBjMf72ciura+sb+c3C1vbO7l5x/6Bu4lRTVqOxiHUzJIYJrljNcitYM9GMyFCwRji4nuSNB6YNj9W9HSYskKSneMQpsc6683GnWMJlPBVaBn8OpavP8fgdAKqd4le7G9NUMmWpIMa0fJzYICPacirYqNBODUsIHZAeazlURDITZNNJR+jUOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44BXcEf3HlZaifl31c9m9xqXICM+XhCI7hDHy4gArcQBVqQCGCR3iGF2/gPXmv3tusNOfNew7hj7yPH40aj4c=</latexit> sha1_base64="TBz0KWW1Y5FVHwrZZOy0b06gZmg=">AAAB6XicbVA9TwJBEJ3DL8Qv1NJmI5pYkT0bLUlsLNEIksCF7C17sGFv77I7Z0Iu/AMbC42x9R/Z+W9c4AoFXzLJy3szmZkXpkpapPTbK62tb2xulbcrO7t7+wfVw6O2TTLDRYsnKjGdkFmhpBYtlKhEJzWCxaESj+H4ZuY/PgljZaIfcJKKIGZDLSPJGTrp3qf9ao3W6RxklfgFqUGBZr/61RskPIuFRq6YtV2fphjkzKDkSkwrvcyKlPExG4quo5rFwgb5/NIpOXfKgESJcaWRzNXfEzmLrZ3EoeuMGY7ssjcT//O6GUbXQS51mqHQfLEoyhTBhMzeJgNpBEc1cYRxI92thI+YYRxdOBUXgr/88ippX9Z9WvfvaK1xVsRRhhM4hQvw4QoacAtNaAGHCJ7hFd68sffivXsfi9aSV8wcwx94nz/eLIzN</latexit> sha1_base64="XaQViiIp3HEbSwArRCfYKY95vuc=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXjRncW3LisYi/QDiWTZtrQJDMkGaEMfQM3LhTt1pWv4863MdN2oa0/BD7+/xxyzgkTwY3F+NtbWV1b39gsbBW3d3b39ksHhw0Tp5qyOo1FrFshMUxwxeqWW8FaiWZEhoI1w+FNnjcfmTY8Vg92lLBAkr7iEafEOuvex91SGVfwVGgZ/DmUrz/fc01q3dJXpxfTVDJlqSDGtH2c2CAj2nIq2LjYSQ1LCB2SPms7VEQyE2TTScfozDk9FMXaPWXR1P3dkRFpzEiGrlISOzCLWW7+l7VTG10FGVdJapmis4+iVCAbo3xt1OOaUStGDgjV3M2K6IBoQq07TtEdwV9ceRkaFxUfV/w7XK6ewkwFOIYTOAcfLqEKt1CDOlCI4Ale4NUbes/emzeZla54854j+CPv4wfqV5FM</latexit> 20 <latexit sha1_base64="QemqZr6fsruN4EG6fJtLmwgab9A=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXTje4suHFZxV6gHUomzbShSWZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wwEdxYjL+9ldW19Y3N3FZ+e2d3b79wcFg3caopq9FYxLoZEsMEV6xmuRWsmWhGZChYIxxcT/LGA9OGx+reDhMWSNJTPOKUWGfdlXGnUMQlPBVaBn8OxavP8fgdAKqdwle7G9NUMmWpIMa0fJzYICPacirYKN9ODUsIHZAeazlURDITZNNJR+jMOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44eXcEf3HlZaiXSz4u+be4WDmFmXJwDCdwDj5cQAVuoAo1oBDBIzzDizfwnrxX721WuuLNe47gj7yPH46fj4g=</latexit> sha1_base64="LOhMDFdaD7Ni8csKulxBbULGOyw=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxiPiA5wt5mL1myt3fszgkh5B/YWChi6z+y89+4Sa7QxAcDj/dmmJkXpkpapPTbK2xsbm3vFHdLe/sHh0fl45OWTTLDRZMnKjGdkFmhpBZNlKhEJzWCxaES7XB8O/fbT8JYmehHnKQiiNlQy0hyhk56qNF+uUKrdAGyTvycVCBHo1/+6g0SnsVCI1fM2q5PUwymzKDkSsxKvcyKlPExG4quo5rFwgbTxaUzcumUAYkS40ojWai/J6YstnYSh64zZjiyq95c/M/rZhjdBFOp0wyF5stFUaYIJmT+NhlIIziqiSOMG+luJXzEDOPowim5EPzVl9dJq1b1adW/p5X6RR5HEc7gHK7Ah2uowx00oAkcIniGV3jzxt6L9+59LFsLXj5zCn/gff4A37GMzg==</latexit> sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiBU7gADy6hBrdQhwZQCOEJXuDVGTnPzpsznZcWnEXPMfyR8/ED69yRTQ==</latexit> <latexit sha1_base64="kj037EtbvNSvtMYW6MIcVdK9BxM=">AAAB7nicbVC7SgNBFL3rM8ZX1NJmMApWYddGOwM2FhYRzAOSJczOziZDZmeXmbtCCPkIG8GI2Orv2Pk3ziYpNPHAhcM553IfQSqFQdf9dlZW19Y3Ngtbxe2d3b390sFhwySZZrzOEpnoVkANl0LxOgqUvJVqTuNA8mYwuMn95iPXRiTqAYcp92PaUyISjKKVmp07Gw1pt1R2K+4UZJl4c1K+/nzJMal1S1+dMGFZzBUySY1pe26K/ohqFEzycbGTGZ5SNqA93rZU0ZgbfzRdd0zOrBKSKNG2FJKp+rtjRGNjhnFgkzHFvln0cvE/r51hdOWPhEoz5IrNBkWZJJiQ/HYSCs0ZyqEllGlhdyWsTzVlaD9UtE/wFk9eJo2LiudWvHu3XD2FGQpwDCdwDh5cQhVuoQZ1YDCAJ5jAq5M6z86b8z6LrjjzniP4A+fjBw2Kk7o=</latexit> <latexit sha1_base64="pceq4LgYQObjNbA2MmlZS3Lp9Io=">AAAB7nicbVC7SgNBFL0bXzG+opY2g1GwCrs22hmwsbCIYB6QLGF29iYZMju7zMwKYclH2AhRREv9HTv/xsmj0MQDFw7nnMt9BIng2rjut5NbWV1b38hvFra2d3b3ivsHdR2nimGNxSJWzYBqFFxizXAjsJkopFEgsBEMrid+4wGV5rG8N8ME/Yj2JO9yRo2VGu1bGw1pp1hyy+4UZJl4c1K6+hyP3wGg2il+tcOYpRFKwwTVuuW5ifEzqgxnAkeFdqoxoWxAe9iyVNIItZ9N1x2RU6uEpBsrW9KQqfq7I6OR1sMosMmImr5e9Cbif14rNd1LP+MySQ1KNhvUTQUxMZncTkKukBkxtIQyxe2uhPWposzYDxXsE7zFk5dJ/bzsuWXvzi1VTmCGPBzBMZyBBxdQgRuoQg0YDOARnuHFSZwn59V5m0VzzrznEP7A+fgBsD6R9Q==</latexit> sha1_base64="4wl+DUzlXPnXRZK/m8ZBYmSFeGA=">AAAB7nicbVC7SgNBFL0bXzG+opY2g1GwCrs2sQzYWFhEMA9IljA7ezcZMju7zMwKYclH2FgoYuv32Pk3TpItNPHAwOGcc5l7T5AKro3rfjuljc2t7Z3ybmVv/+DwqHp80tFJphi2WSIS1QuoRsEltg03AnupQhoHArvB5Hbud59QaZ7IRzNN0Y/pSPKIM2qs1B3c22hIh9WaW3cXIOvEK0gNCrSG1a9BmLAsRmmYoFr3PTc1fk6V4UzgrDLINKaUTegI+5ZKGqP288W6M3JplZBEibJPGrJQf0/kNNZ6Ggc2GVMz1qveXPzP62cmuvFzLtPMoGTLj6JMEJOQ+e0k5AqZEVNLKFPc7krYmCrKjG2oYkvwVk9eJ53ruufWvQe31rwo6ijDGZzDFXjQgCbcQQvawGACz/AKb07qvDjvzscyWnKKmVP4A+fzBwFfjzs=</latexit> ⇤ <latexit sha1_base64="iUlx6QDkke9z6q0+mzE0gtXqTEo=">AAAB6XicbZDLSgMxFIZPvNZ6q7p0E6yCq5LRhe4suHFZxV6gHUomzbShmcyQZIQy9A3cuFCsK8GVr+POtzG9LLT1h8DH/59DzjlBIoWxhHyjpeWV1bX13EZ+c2t7Z7ewt18zcaoZr7JYxroRUMOlULxqhZW8kWhOo0DyetC/Huf1B66NiNW9HSTcj2hXiVAwap11d07ahSIpkYnwIngzKF59jkbvAFBpF75anZilEVeWSWpM0yOJ9TOqrWCSD/Ot1PCEsj7t8qZDRSNu/Gwy6RCfOKeDw1i7pyyeuL87MhoZM4gCVxlR2zPz2dj8L2umNrz0M6GS1HLFph+FqcQ2xuO1cUdozqwcOKBMCzcrZj2qKbPuOHl3BG9+5UWonZU8UvJuSbF8DFPl4BCO4BQ8uIAy3EAFqsAghEd4hhfUR0/oFb1NS5fQrOcA/gh9/ACQJI+J</latexit> sha1_base64="NQpDjORV9LqRKjp0vLHsTT3vYQ0=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFfa00DJgYxnFfEByhL3NXrJkb+/YnRNCyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvTJW0SOm3V1hb39jcKm6Xdnb39g/Kh0dNm2SGiwZPVGLaIbNCSS0aKFGJdmoEi0MlWuHodua3noSxMtGPOE5FELOBlpHkDJ30cEV75Qqt0jnIKvFzUoEc9V75q9tPeBYLjVwxazs+TTGYMIOSKzEtdTMrUsZHbCA6jmoWCxtM5pdOyblT+iRKjCuNZK7+npiw2NpxHLrOmOHQLnsz8T+vk2F0E0ykTjMUmi8WRZkimJDZ26QvjeCoxo4wbqS7lfAhM4yjC6fkQvCXX14lzcuqT6v+Pa3UzvI4inACp3ABPlxDDe6gDg3gEMEzvMKbN/JevHfvY9Fa8PKZY/gD7/MH4TaMzw==</latexit> sha1_base64="zMxGvBYD59aqPpvpyRUdHOQUixk=">AAAB6XicbZC7TsMwFIZPyq2UW4GRxaIgMVUODLBRiYWxIHqR2qhyXKe16jiR7SBVUd+AhQEEXZl4HTbeBiftAC2/ZOnT/58jn3P8WHBtMP52Ciura+sbxc3S1vbO7l55/6Cpo0RR1qCRiFTbJ5oJLlnDcCNYO1aMhL5gLX90k+WtR6Y0j+SDGcfMC8lA8oBTYqx1f4F75Qqu4lxoGdw5VK4/3zNN673yV7cf0SRk0lBBtO64ODZeSpThVLBJqZtoFhM6IgPWsShJyLSX5pNO0Kl1+iiIlH3SoNz93ZGSUOtx6NvKkJihXswy87+sk5jgyku5jBPDJJ19FCQCmQhla6M+V4waMbZAqOJ2VkSHRBFq7HFK9gju4srL0Dyvurjq3uFK7QRmKsIRHMMZuHAJNbiFOjSAQgBP8AKvzsh5dt6c6ay04Mx7DuGPnI8f7WGRTg==</latexit> 30 FIG 5 Numer ca b urcat on ana ys s o wavetra ns (a) B urcat on d agram o wavetra ns n the parameter η0 or var ous va ues o the doma n s ze where kc = 0 739 Here we show on y the max mum o the synapt c conductance g1 as our so ut on measure The green so ut on b urcates off the steady state so ut on (grey) at a Tur ng-Hop b urcat on and the other so ut ons go unstab e at s ght y arger va ues o η0 (b) the branches n (a) are p otted us ng the phase ve oc ty c as so ut on measure The d agrams n (a) and (b) confirm that the Tur ng-Hop b urcat on s subcr t ca At the cr t ca po nts the wavetra n emerges w th a non-zero phase ve oc ty (ωc /kc ) (c) Branches o wavetra ns n the cont nuat on parameter Λ or η0 = 0 Th s curve s effect ve y a d spers on re at on show ng the wave speed as unct on o the spat a per od In th s reg on o parameter space the branch s an so a there ore the wavetra n ex sts on y or a fin te range o per ods Parameter va ues v yn = −30 other parameters as n F g 2 As n §IV A the spat a frequency kc at b furcat on was found by so v ng (29) -- (31) and we cont nued patterns w th th s wave number by sett ng Λ = 2π/kc where kc = 0 739 (F g 5(a) -- (b) green curve) We a so cont nued spat a patterns w th k = 2kc (red curve) and k = kc /2 (b ue curve) th s was ach eved by pos ng the system on doma ns Λ = π/kc and Λ = 4π/kc respect ve y The wavetra ns b furcate from the homogeneous steady state w th a non-zero phase ve oc ty at d fferent va ues of η0 (F g 5(b)) The va ue for wh ch per od c waves emerge for Λ = 2π/kc corresponds to the Tur ngHopf b furcat on (green dot) found n §III and the speed s equa to ωc /kc The wavetra ns are unstab e for ow va ues of η0 For the sma er doma n s ze they are unstab e to the stab e steady state whereas for the arger doma n s ze they trans t on to finer patterns w th a sma er spat a wave ength k To fu y exp ore the re at onsh p between the wave speed and the spat a per od we fix the va ue of η0 = 0 and use Λ as a cont nuat on parameter Th s opens up the poss b ty to trace branches of so ut ons n the (c Λ)p ane and hence approx mate the d spers on curve of the waves (F g 5(c)) We found that wavetra ns occur n so as and therefore on y ex st for a fin te range of spat a per ods For 5 2 < Λ < 13 1 a fast (stab e) wavetra n coex st w th s ower (unstab e) one Whereas when 3 2 < Λ < 5 2 and 13 1 < Λ < 28 0 the two waves are unstab e For a other va ues of Λ we do not see wavetra n so ut ons If η0 s ncreased the d spers on curve s no onger an so a but rather a monoton ca y ncreas ng funct on Thus n th s reg me the system supports per od c trave ng waves for a va ues of the spat a per od above a thresho d va ue C Fronts Standard neura fie d mode s are known to support trave ng fronts (for exponent a y decay ng kerne s) wh ch are trave ng waves whose profi e connects a un form h gh-act v ty state to another ow-act v ty state 10 24 It s therefore natura to search for these coherent structures n our new neura fie d mode (19) -- (20) As n any other non oca neura fie d the ex stence of the h gh- and ow-act v ty states depends on the cho ce of the synapt c kerne As nh b t on s not necessary for the generat on of oca sed patterns such as trave ng fronts n th s sect on we cons der on y a s ng e synapt c conductance m = 1 and suppress the m abe ead ng to the equ va ent PDE formu at on (1 − 2 ∂x )(1 ∂t z = F(z) + G(z g) + τ ∂t )2 g = κf (z) (36) wh ch we pose on ξ ∈ R We now set z(x t) = Z(x − ct) g(x t) = G(x − ct) and we seek trave ng fronts as bounded so ut ons U (ξ) = (U1 (ξ) U6 (ξ)) to the boundary-va ue prob em ∂ξ U = N (U ) ξ∈R m U (ξ) = U ξ→±∞ (37) 10 where N : R6 → R6 is the real-valued nonlinear function   − Re[F(U1 + iU2 ) + G(U1 + iU2 , U3 )]/c − Im[F(U1 + iU2 ) + G(U1 + iU2 , U3 )]/c   (U3 − U4 )/(τ c)   N (U ) =  . U5     U6 U5 + U6 + κf (U1 + iU2 ) − U4 /(τ c) In this spatial-dynamical system formulation of the problem, the first three components of U have a direct interpretation in terms of the state variables (z, g) of (36), U1 = A ≡ Re Z, U2 = B ≡ Im Z, U3 = G, whereas U4 , U5 and U6 are auxiliary variables, necessary to cast the problem as a system of first-order differential equations in ξ. The equilibria U ± = ± ± ± ± (U1 , U2 , U3 , U3 , 0, 0) of the spatial-dynamical system (37) correspond to high- and low-activity homogeneous steady states of (36) and are completely determined by solving for (U1 , U2 , U3 ) the algebraic problem Re[F(U1 + iU2 ) + G(U1 + iU2 , U3 )] = 0, Im[F(U1 + iU2 ) + G(U1 + iU2 , U3 )] = 0, κf (U1 + iU2 ) − U3 = 0. (38) We define U + as the high activity state, which displays high synaptic conductance and U − as the low activity state, displaying lower synaptic conductance. Note that we also seek the symmetric counterpart solution where limξ→±∞ U (ξ) = U ± . We have continued solutions to (38) in η0 and vsyn using XPPAUT [25] (Fig. 6(a)). The system has three fixed points in the region enclosed by the saddle-node curves, two of which are stable (U ± ). Therefore, we look for travelling waves in this shaded region of parameter space. We note that, as τ is decreased, a Hopf bifurcation of the U − state arises, opening up the possibility of creating heteroclinic connections to periodic orbits, which we will not consider further here. The state at ξ → −∞ (ξ → ∞) displays a highconductance (low-conductance) G, hence a high (low) firing, and it is therefore referred to as the high (low) activity state (Fig. 6(b)). We computed travelling waves using the routines from [5], which allows us to study the spectral stability of the waves. Computations are performed on a large truncated domain Λ = 60, and plotted on [−10, 10] for convenience. Interestingly, we see ripples in the wake of the front, which indicates that the solution connects a node to a focus. The solution in the phase space (A, B, G) illustrates that the high-activity state (U − ) is a focus, and the low activity one (U + ) a node. We continued the travelling wave shown in Fig. 6(b) (blue), and its symmetric counterpart (red), in the mean background drive η0 , using c as a solution measure (Fig. 6(d)). Fronts live on an isolated branch and destabilise at saddle-node bifurcations. The bifurcation diagram is symmetric with respect to the axis c = 0, as solutions on the branch with c > 0 and c < 0 are related via the transformation U (ξ) → U (−ξ). Up to 6 coexisting waves exist in a wide region of η0 parameter space, albeit only two of them are stable. In addition, we found stable waves in which the high activity state U + is moving across the tissue, invading the low activity state U − , and vice versa. The inversion of velocity occurs where the blue and red stable branches overlap. As the isola of travelling fronts is traced, these solutions gain or lose oscillations in the wake of the wave. To investigate further this aspect, we monitor the spectrum of U ± (as equilibria of the spatial-dynamical system (37)) as we move along the branch. Taking a vertical excursion in Fig. 6(d), we find three independent solutions, one of which is stable and two of which unstable. We examine the stability of the fixed points U ± in the travelling wave frame for each of these solutions. As expected, oscillations in the wake of the wave are absent where the unstable spectrum of U − is purely real (Fig. 7(a)), and they develop when a complex-conjugate pair of eigenvalues crosses the imaginary axis (Fig. 7(b)). We also find that the amplitude of the oscillations is small when the real part of the complex eigenpair is small (Fig. 7(c)). V. DISCUSSION We have presented the derivation of an atypical neural field model from a network of spatially distributed θneurons. In the reduced model, which we dub a next generation neural field model, within population synchrony drives the population firing rate. The new model supports a range of patterns, such as bumps, waves and breathers. Noteworthy is the state characterised by structures within bumps, as these states are not seen in standard neural mass models. These structures instead typify patterns seen in networks of spiking neurons, which signifies that by maintaining the notion of within population synchrony this neural field model can retain information about the underlying spiking network. Exotic states, with within bump oscillations, were found in the region where the Hopf and Turing bifurcations collided. A Turing instability analysis provided us with an understanding of how the system behaved close to bifurcation points, allowing us to determine when the system transitioned from the homogeneous steady state. However, unlike the Amari model we cannot use the Heaviside approximation to make further analytical progress since the firing rate is now a fixed real valued function of the Kuramoto order parameter. As such, we have moved to numerical continuation techniques to analyse the behaviour of the system away from these bifurcation points. Numerical techniques were also used to examine the existence and stability of travelling fronts. Previous work [14] illustrated that the point version of this model with an external time-dependent drive (without spatial extent) could support β-rebound, an event- 11 (a) 0 (b) 6 U 4 G 2 0 Bistability 0.8 R 0.6 0.4 10 14 10 + (c) c <latexit sha1_base64="hQMIBOfHiUFR97X8IUpOplkAurY=">AAAB6nicbZA9SwNBEIbn4leMX1FLm8UgCEK4s9HOgI1lRC8JJGfY2+wlS/b2jt05IYb8BBsLRWzzi+ws/SduPgpNfGHh4X1n2JkJUykMuu6Xk1tZXVvfyG8WtrZ3dveK+wc1k2SacZ8lMtGNkBouheI+CpS8kWpO41Dyeti/nuT1R66NSNQ9DlIexLSrRCQYRWvd+Q9n7WLJLbtTkWXw5lC6Gj99VwCg2i5+tjoJy2KukElqTNNzUwyGVKNgko8KrczwlLI+7fKmRUVjboLhdNQRObFOh0SJtk8hmbq/O4Y0NmYQh7Yyptgzi9nE/C9rZhhdBkOh0gy5YrOPokwSTMhkb9IRmjOUAwuUaWFnJaxHNWVor1OwR/AWV16G2nnZc8verVuqEJgpD0dwDKfgwQVU4Aaq4AODLjzDK7w50nlx3p2PWWnOmfccwh854x8egY/O</latexit> sha1_base64="u5KeCHKWxRt8fJA1ZXTs81m2SUg=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBZBEEriRY8FLx4rmrbQxrLZbtqlm03YnQgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5YSqFQdf9dkpr6xubW+Xtys7u3v5B9fCoZZJMM+6zRCa6E1LDpVDcR4GSd1LNaRxK3g7HNzO//cS1EYl6wEnKg5gOlYgEo2ile//xol+tuXV3DrJKvILUoECzX/3qDRKWxVwhk9SYruemGORUo2CSTyu9zPCUsjEd8q6lisbcBPn81Ck5s8qARIm2pZDM1d8TOY2NmcSh7YwpjsyyNxP/87oZRtdBLlSaIVdssSjKJMGEzP4mA6E5QzmxhDIt7K2EjaimDG06FRuCt/zyKmld1j237t25tQYp4ijDCZzCOXhwBQ24hSb4wGAIz/AKb450Xpx352PRWnKKmWP4A+fzB8A4jVA=</latexit> sha1_base64="n9CRbNUT769UliLxqXWDfwJpoKU=">AAAB6nicbZDLSgMxFIbP1Futt3rZuQkWQRBKxo3uLLjQZUWnLbRjyaSZNjSTGZKMUIc+ghsXirjt2odx59I3MW1daOsPgY//P4ecc4JEcG0w/nRyC4tLyyv51cLa+sbmVnF7p6bjVFHm0VjEqhEQzQSXzDPcCNZIFCNRIFg96F+M8/o9U5rH8tYMEuZHpCt5yCkx1rrx7o7bxRIu44nQPLg/UDofPXxdvu9l1Xbxo9WJaRoxaaggWjddnBg/I8pwKtiw0Eo1Swjtky5rWpQkYtrPJqMO0aF1OiiMlX3SoIn7uyMjkdaDKLCVETE9PZuNzf+yZmrCMz/jMkkNk3T6UZgKZGI03ht1uGLUiIEFQhW3syLaI4pQY69TsEdwZ1eeh9pJ2cVl9xqXKgimysM+HMARuHAKFbiCKnhAoQuP8AwvjnCenFfnbVqac356duGPnNE3z6mREg==</latexit> <latexit sha1_base64="O4DjIWLmL6adnSkN0S35wp1C5+0=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJLLSERD4SuJC9ZQ5W9vYuu3smhPALbCw0Bks7/46d/8Y9oFD0TTZ58r4z2ZkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxheZ3nzAZXmsbwzowT9iPYlDzmjxlq1m26x5Jbdmchf8BZQuvqYZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1ISX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBBxdQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DndqRKA==</latexit> <latexit sha1_base64="xBvVSkDKiTj8V/S1f2KEkuvucXw=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGLLRMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8nubNB1Sax/LOjBL0I9qXPOSMGmvVbrrFklt2ZyJ/wVtA6epjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvoZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQAVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DUCdj2M=</latexit> sha1_base64="7qQyQ5h77susj5tKug7kTRGCfIY=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFqNgFe5sYhmw0DIB8wHJEfY2c8mavb1jd08IR36BjYUitv4kO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7AST27nfeUKleSwfzDRBP6IjyUPOqLFS825QrrhVdwGyTrycVCBHY1D+6g9jlkYoDRNU657nJsbPqDKcCZyV+qnGhLIJHWHPUkkj1H62OHRGLq0yJGGsbElDFurviYxGWk+jwHZG1Iz1qjcX//N6qQlv/IzLJDUo2XJRmApiYjL/mgy5QmbE1BLKFLe3EjamijJjsynZELzVl9dJ+7rquVWv6VbqF3kcRTiDc7gCD2pQh3toQAsYIDzDK7w5j86L8+58LFsLTj5zCn/gfP4Aka+MqQ==</latexit> <latexit sha1_base64="X4aDnUVPoJrzRBXnqG9AIZwZhKo=">AAAB6nicbZDLSgMxFIbPeK31VnXpJlgFN5ZMN7oRC7pwWdFeoB1KJs20oZlkSDJCGfoIblwo4tYncuc7+BCml4W2/hD4+P9zyDknTAQ3FuMvb2l5ZXVtPbeR39za3tkt7O3XjUo1ZTWqhNLNkBgmuGQ1y61gzUQzEoeCNcLB9ThvPDJtuJIPdpiwICY9ySNOiXXW/VkZdwpFXMIToUXwZ1C8+r65rABAtVP4bHcVTWMmLRXEmJaPExtkRFtOBRvl26lhCaED0mMth5LEzATZZNQROnFOF0VKuyctmri/OzISGzOMQ1cZE9s389nY/C9rpTa6CDIuk9QySacfRalAVqHx3qjLNaNWDB0QqrmbFdE+0YRad528O4I/v/Ii1MslH5f8O1ysHMNUOTiEIzgFH86hArdQhRpQ6METvMCrJ7xn7817n5YuebOeA/gj7+MH4veO8Q==</latexit> sha1_base64="2SEM3eGqeoRW4lYWolzVAeeQtpI=">AAAB6nicbVA9SwNBEJ3zM8avqKXNYhRsDHtptAzYWEY0H5AcYW8zlyzZ2zt294QQ8hNsLBSx9RfZ+W/cJFdo4oOBx3szzMwLUymMpfTbW1vf2NzaLuwUd/f2Dw5LR8dNk2SaY4MnMtHtkBmUQmHDCiuxnWpkcSixFY5uZ37rCbURiXq04xSDmA2UiARn1kkPV1XaK5Vphc5BVomfkzLkqPdKX91+wrMYleWSGdPxaWqDCdNWcInTYjczmDI+YgPsOKpYjCaYzE+dkgun9EmUaFfKkrn6e2LCYmPGceg6Y2aHZtmbif95ncxGN8FEqDSzqPhiUZRJYhMy+5v0hUZu5dgRxrVwtxI+ZJpx69IpuhD85ZdXSbNa8WnFv6fl2nkeRwFO4QwuwYdrqMEd1KEBHAbwDK/w5knvxXv3Phata14+cwJ/4H3+AEk0jQU=</latexit> sha1_base64="hbtbN+ratGNP6Oc+jDdNo3ukzAw=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquLFkutGNWFDEZUV7gXYomTTThmYyQ5IRytBHcONCEbc+gY/iznfwAVyaXhba+kPg4//PIeccPxZcG4w/nczC4tLySnY1t7a+sbmV396p6ShRlFVpJCLV8IlmgktWNdwI1ogVI6EvWN3vX4zy+j1Tmkfyzgxi5oWkK3nAKTHWuj0u4Xa+gIt4LDQP7hQK51+XZ1fv/e9KO//R6kQ0CZk0VBCtmy6OjZcSZTgVbJhrJZrFhPZJlzUtShIy7aXjUYfo0DodFETKPmnQ2P3dkZJQ60Ho28qQmJ6ezUbmf1kzMcGpl3IZJ4ZJOvkoSAQyERrtjTpcMWrEwAKhittZEe0RRaix18nZI7izK89DrVR0cdG9wYXyAUyUhT3YhyNw4QTKcA0VqAKFLjzAEzw7wnl0XpzXSWnGmfbswh85bz+y/JEM</latexit> 10 ⌘0 20 30 U+ 1 <latexit sha1_base64="vV/tSZsqp0PZDp4B5kGJej5F2X0=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmj0U4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUa5ZLnlryaW6ycw1w5OIUzuAQPrqACd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QN+BpET</latexit> <latexit sha1_base64="NShdlnjkNDfv4ynZNqa/uWih2wQ=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0CiXPLfk1dxi5RzmysEpnMEleHAFFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fIMmPTg==</latexit> sha1_base64="uAhRiQUyUfrbUCj+oLzLI3Qe8O4=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VabqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9x24yU</latexit> G <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> U <latexit sha1_base64="Xjxj1j2+WvC32wZ3DhYVHVhLrkc=">AAAB6nicbZA9TwJBEIbn8AvxC7W02UhMbCR3NtpJYmOJ0QMSOMnesgcb9vYuu3MmSPgJNhYaY8svsrP0n7h8FAq+ySZP3ncmOzNhKoVB1/1yciura+sb+c3C1vbO7l5x/6Bmkkwz7rNEJroRUsOlUNxHgZI3Us1pHEpeD/vXk7z+yLURibrHQcqDmHaViASjaK07/+GsXSy5ZXcqsgzeHEpX46fvCgBU28XPVidhWcwVMkmNaXpuisGQahRM8lGhlRmeUtanXd60qGjMTTCcjjoiJ9bpkCjR9ikkU/d3x5DGxgzi0FbGFHtmMZuY/2XNDKPLYChUmiFXbPZRlEmCCZnsTTpCc4ZyYIEyLeyshPWopgztdQr2CN7iystQOy97btm7dUsVAjPl4QiO4RQ8uIAK3EAVfGDQhWd4hTdHOi/Ou/MxK805855D+CNn/AMhiY/Q</latexit> sha1_base64="be7E9tnS0JRK5EnNBq6r9S2JMTI=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBiyXxoseCF48VTVtoY9lsN+3SzSbsToQS+hO8eFDEq7/Im//GbZuDtj4YeLw3w8y8MJXCoOt+O6W19Y3NrfJ2ZWd3b/+genjUMkmmGfdZIhPdCanhUijuo0DJO6nmNA4lb4fjm5nffuLaiEQ94CTlQUyHSkSCUbTSvf940a/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+6pScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZn+TgdCcoZxYQpkW9lbCRlRThjadig3BW355lbQu655b9+7cWoMUcZThBE7hHDy4ggbcQhN8YDCEZ3iFN0c6L86787FoLTnFzDH8gfP5A8NAjVI=</latexit> sha1_base64="5Pib3wNvET9FX8+4YrCKT2Dosbo=">AAAB6nicbZC7TgJBFIbP4g3xhpfOZiIxsZHM2mgniYWWGF0ggZXMDrMwYXZ2MzNrghsewcZCY2ypfRg7S9/EASwU/JNJvvz/OZlzTpAIrg3Gn05uYXFpeSW/Wlhb39jcKm7v1HScKso8GotYNQKimeCSeYYbwRqJYiQKBKsH/YtxXr9nSvNY3ppBwvyIdCUPOSXGWjfe3XG7WMJlPBGaB/cHSuejh6/L972s2i5+tDoxTSMmDRVE66aLE+NnRBlOBRsWWqlmCaF90mVNi5JETPvZZNQhOrROB4Wxsk8aNHF/d2Qk0noQBbYyIqanZ7Ox+V/WTE145mdcJqlhkk4/ClOBTIzGe6MOV4waMbBAqOJ2VkR7RBFq7HUK9gju7MrzUDspu7jsXuNSBcFUediHAzgCF06hAldQBQ8odOERnuHFEc6T8+q8TUtzzk/PLvyRM/oG0rGRFA==</latexit> <latexit sha1_base64="PZxHYbmKRx3q9oyWN7/XOFJc40s=">AAAB8nicbVBNS8NAEJ34WetX1aOXxSJ4Kkkveix68VjBfkAayma7aZduNmF3IoTQn+HFgyJe/TXe/Ddu2xy09cHA470ZZuaFqRQGXffb2djc2t7ZrexV9w8Oj45rJ6ddk2Sa8Q5LZKL7ITVcCsU7KFDyfqo5jUPJe+H0bu73nrg2IlGPmKc8iOlYiUgwilbyb+0GGgopMB/W6m7DXYCsE68kdSjRHta+BqOEZTFXyCQ1xvfcFIOCahRM8ll1kBmeUjalY+5bqmjMTVAsTp6RS6uMSJRoWwrJQv09UdDYmDwObWdMcWJWvbn4n+dnGN0EhVBphlyx5aIokwQTMv+fjITmDGVuCWVa2FsJm1BNGdqUqjYEb/XlddJtNjy34T006y1SxlGBc7iAK/DgGlpwD23oAIMEnuEV3hx0Xpx352PZuuGUM2fwB87nD36gkUQ=</latexit> c <latexit sha1_base64="ICeM7tEyNJx4+UKss+ChiFUGcGU=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOC92oxMJYBL1IbVQ5rtNadZzIdpCqqI/AwgCKWBhYeB023gb3MkDLL1n69P/nyOecIBFcG4y/ncLa+sbmVnG7tLO7t39QPjxq6jhVlDVoLGLVDohmgkvWMNwI1k4UI1EgWCsY3Uzz1iNTmsfywYwT5kdkIHnIKTHWusdutVeuYBfPhFbBW0Dl+jPP3wGg3it/dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNQJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Op+V/WSU1Y9TMuk9QwSecfhalAJkbTvVGfK0aNGFsgVHE7K6JDogg19jolewRveeVVaF66Hna9O1ypncFcRTiBU7gAD66gBrdQhwZQGMATvMCrI5xnJ3fe5qUFZ9FzDH/kfPwAAsCPxg==</latexit> sha1_base64="+sSWKmzLMcZXlXqR/BBBx6myqzA=">AAAB6nicbVBNSwMxEJ2tX7V+VT16CVbB05L1Yo8FLx4r2g9ol5JNs21okl2SrFCW/gQvHhTx6i/y5r8xbfegrQ8GHu/NMDMvSgU3FuNvr7SxubW9U96t7O0fHB5Vj0/aJsk0ZS2aiER3I2KY4Iq1LLeCdVPNiIwE60ST27nfeWLa8EQ92mnKQklGisecEuukB+zXB9Ua9vECaJ0EBalBgeag+tUfJjSTTFkqiDG9AKc2zIm2nAo2q/Qzw1JCJ2TEeo4qIpkJ88WpM3TplCGKE+1KWbRQf0/kRBozlZHrlMSOzao3F//zepmN62HOVZpZpuhyUZwJZBM0/xsNuWbUiqkjhGrubkV0TDSh1qVTcSEEqy+vk/a1H2A/uMe1xkURRxnO4ByuIIAbaMAdNKEFFEbwDK/w5gnvxXv3PpatJa+YOYU/8D5/AFPSjQw=</latexit> sha1_base64="LI+k/Cwz6I63jyYJKGDnd8CGGCc=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOC92oxMJYBL1IbVQ5rtNadZzIdpCqqI/AwgAqrCy8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/6B8eNTUcaooa9BYxKodEM0El6xhuBGsnShGokCwVjC6yfPWI1Oax/LBjBPmR2QgecgpMda6x261V65gF8+EVsFbQOX6c5rrrd4rf3X7MU0jJg0VROuOhxPjZ0QZTgWblLqpZgmhIzJgHYuSREz72WzUCTq3Th+FsbJPGjRzf3dkJNJ6HAW2MiJmqJez3Pwv66QmrPoZl0lqmKTzj8JUIBOjfG/U54pRI8YWCFXczorokChCjb1OyR7BW155FZqXrodd7w5XamcwVxFO4BQuwIMrqMEt1KEBFAbwBC/w6gjn2Zk67/PSgrPoOYY/cj5+AF/9kYs=</latexit> <latexit sha1_base64="Bm6L7INbJx8XEYQzKxIxOiRrWCU=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCzByEcCF7K3zMHK3t5ld8+EEH6BjYXGYGnn37Hz37gHFIq+ySZP3ncmOzNBIrg2rvvl5FZW19Y38puFre2d3b3i/kFDx6liWGexiFUroBoFl1g33AhsJQppFAhsBsPrLG8+oNI8lndmlKAf0b7kIWfUWKt22y2W3LI7E/kL3gJKVx/TTG/VbvGz04tZGqE0TFCt256bGH9MleFM4KTQSTUmlA1pH9sWJY1Q++PZoBNyap0eCWNlnzRk5v7sGNNI61EU2MqImoFezjLzv6ydmvDSH3OZpAYlm38UpoKYmGRbkx5XyIwYWaBMcTsrYQOqKDP2NgV7BG955b/QOC97btmruaXKCcyVhyM4hjPw4AIqcANVqAMDhEd4hhfn3nlyXp3pvDTnLHoO4Zec92+uhpEz</latexit> <latexit sha1_base64="qoVWhpdOghMvJ6ZyOV73cMAR63I=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMRdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmj0h4GP/z+HOecEieDauO6Xk1tZXVvfyG8WtrZ3dveK+wcNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQyvp3nzAZXmsbwzowT9iPYlDzmjxlq1226x5Jbdmchf8BZQuvqYTN4AoNotfnZ6MUsjlIYJqnXbcxPjZ1QZzgSOC51UY0LZkPaxbVHSCLWfzQYdk1Pr9EgYK/ukITP3Z0dGI61HUWArI2oGejmbmv9l7dSEl37GZZIalGz+UZgKYmIy3Zr0uEJmxMgCZYrbWQkbUEWZsbcp2CN4yyv/hcZ52XPLXs0tVU5grjwcwTGcgQcXUIEbqEIdGCA8wjO8OPfOk/PqTOalOWfRcwi/5Lx/A1FJj24=</latexit> sha1_base64="af6fmYZ3mFiUWU/zWR5QDen5nZM=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLElsLMHIRwIXsrfMwcre3mV3z4Rc+AU2Fhpj60+y89+4wBUKvmSSl/dmMjMvSATXxnW/ncLG5tb2TnG3tLd/cHhUPj5p6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWByO/c7T6g0j+WDmSboR3QkecgZNVZq3g/KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGNn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o13Ur9Io+jCGdwDlfgQQ3qcAcNaAEDhGd4hTfn0Xlx3p2PZWvByWdO4Q+czx+iW4y0</latexit> 0c 0 05 <latexit sha1_base64="rCVsgeNE0e8YDyRFxTSTN12NQJo=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCwhkY8ELmRvmYOVvb3L7p4JIfwCGwuNwdLOv2Pnv3EPKBR8k02evO9MdmaCRHBtXPfbya2tb2xu5bcLO7t7+wfFw6OGjlPFsM5iEatWQDUKLrFuuBHYShTSKBDYDIa3Wd58RKV5LO/NKEE/on3JQ86osVaNdYslt+zORFbBW0Dp5nOa6b3aLX51ejFLI5SGCap123MT44+pMpwJnBQ6qcaEsiHtY9uipBFqfzwbdELOrdMjYazsk4bM3N8dYxppPYoCWxlRM9DLWWb+l7VTE177Yy6T1KBk84/CVBATk2xr0uMKmREjC5QpbmclbEAVZcbepmCP4C2vvAqNy7Lnlr2aW6qcwVx5OIFTuAAPrqACd1CFOjBAeIIXeHUenGfnzZnOS3POoucY/sj5+AHISpFE</latexit> <latexit sha1_base64="/LXIjfRVWR37uD2eTarUm8xNMK0=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMBdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmjiDwMf/38Oc84JEsG1cd1vJ7e2vrG5ld8u7Ozu7R8UD48aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeTvPmIyrNY3lvRgn6Ee1LHnJGjbVqrFssuWV3JrIK3gJKN5+TyTsAVLvFr04vZmmE0jBBtW57bmL8jCrDmcBxoZNqTCgb0j62LUoaofaz2aBjcm6dHgljZZ80ZOb+7shopPUoCmxlRM1AL2dT87+snZrw2s+4TFKDks0/ClNBTEymW5MeV8iMGFmgTHE7K2EDqigz9jYFewRveeVVaFyWPbfs1dxS5QzmysMJnMIFeHAFFbiDKtSBAcITvMCr8+A8O2/OZF6acxY9x/BHzscPaw2Pfw==</latexit> sha1_base64="7Ii0uNmP2RfOgM1BBfSq2ARZFww=">AAAB6HicbVBNT8JAEJ3iF+IX6tHLRjTxRFoveiTx4hESCyTQkO0yhZXtttndmpCGX+DFg8Z49Sd589+4QA8KvmSSl/dmMjMvTAXXxnW/ndLG5tb2Tnm3srd/cHhUPT5p6yRTDH2WiER1Q6pRcIm+4UZgN1VI41BgJ5zczf3OEyrNE/lgpikGMR1JHnFGjZVabFCtuXV3AbJOvILUoEBzUP3qDxOWxSgNE1TrnuemJsipMpwJnFX6mcaUsgkdYc9SSWPUQb44dEYurTIkUaJsSUMW6u+JnMZaT+PQdsbUjPWqNxf/83qZiW6DnMs0MyjZclGUCWISMv+aDLlCZsTUEsoUt7cSNqaKMmOzqdgQvNWX10n7uu65da/l1hoXRRxlOINzuAIPbqAB99AEHxggPMMrvDmPzovz7nwsW0tOMXMKf+B8/gC8H4zF</latexit> <latexit sha1_base64="e3n4Hdl+h1VLOIkzSLBAapQ/qJk=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOA7BRiYWxCHqR2qhyXKe16jiR7SBVUR+BhQEUsTCw8DpsvA3uZYCWX7L06f/Pkc85QSK4Nhh/O4WV1bX1jeJmaWt7Z3evvH/Q0HGqKKvTWMSqFRDNBJesbrgRrJUoRqJAsGYwvJnkzUemNI/lgxklzI9IX/KQU2KsdY/di265gl08FVoGbw6V6888fweAWrf81enFNI2YNFQQrdseToyfEWU4FWxc6qSaJYQOSZ+1LUoSMe1n01HH6NQ6PRTGyj5p0NT93ZGRSOtRFNjKiJiBXswm5n9ZOzXhlZ9xmaSGSTr7KEwFMjGa7I16XDFqxMgCoYrbWREdEEWosdcp2SN4iysvQ+Pc9bDr3eFK9QRmKsIRHMMZeHAJVbiFGtSBQh+e4AVeHeE8O7nzNistOPOeQ/gj5+MH/6mPxA==</latexit> sha1_base64="Mjqpvk/HCTohdRb7NMpW18Eqhj0=">AAAB6nicbVBNSwMxEJ2tX7V+VT16CVbB05L1oB4LXjxWtB/QLiWbZtvQJLskWaEs/QlePCji1V/kzX9j2u5BWx8MPN6bYWZelApuLMbfXmltfWNzq7xd2dnd2z+oHh61TJJpypo0EYnuRMQwwRVrWm4F66SaERkJ1o7GtzO//cS04Yl6tJOUhZIMFY85JdZJD9i/6ldr2MdzoFUSFKQGBRr96ldvkNBMMmWpIMZ0A5zaMCfacirYtNLLDEsJHZMh6zqqiGQmzOenTtG5UwYoTrQrZdFc/T2RE2nMREauUxI7MsveTPzP62Y2vglzrtLMMkUXi+JMIJug2d9owDWjVkwcIVRzdyuiI6IJtS6digshWH55lbQu/QD7wT2u1c+KOMpwAqdwAQFcQx3uoAFNoDCEZ3iFN094L96797FoLXnFzDH8gff5A1DKjQo=</latexit> sha1_base64="4QKjlfxVwMICjxmqKE8H8GBSuXU=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOA7BRiYWxCHqR2qhyXKe16jiR7SBVUR+BhQFUWFl4HTbeBqftAC2/ZOnT/58jn3OCRHBtMP52Ciura+sbxc3S1vbO7l55/6Ch41RRVqexiFUrIJoJLlndcCNYK1GMRIFgzWB4k+fNR6Y0j+WDGSXMj0hf8pBTYqx1j92LbrmCXTwVWgZvDpXrz0mut1q3/NXpxTSNmDRUEK3bHk6MnxFlOBVsXOqkmiWEDkmftS1KEjHtZ9NRx+jUOj0Uxso+adDU/d2RkUjrURTYyoiYgV7McvO/rJ2a8MrPuExSwySdfRSmApkY5XujHleMGjGyQKjidlZEB0QRaux1SvYI3uLKy9A4dz3sene4Uj2BmYpwBMdwBh5cQhVuoQZ1oNCHJ3iBV0c4z87EeZ+VFpx5zyH8kfPxA1z1kYk=</latexit> <latexit sha1_base64="Au8SkmTnn5P4zSn4ESbNz8+h6LY=">AAAB6nicbZDLSgMxFIbPeK31VnXpJlgFN5aMLnQjFnThsqK9QDuUTJppQ5PMkGSEMvQR3LhQxK1P5M538CFMLwtt/SHw8f/nkHNOmAhuLMZf3sLi0vLKam4tv76xubVd2NmtmTjVlFVpLGLdCIlhgitWtdwK1kg0IzIUrB72r0d5/ZFpw2P1YAcJCyTpKh5xSqyz7k/OcLtQxCU8FpoHfwrFq++byzIAVNqFz1YnpqlkylJBjGn6OLFBRrTlVLBhvpUalhDaJ13WdKiIZCbIxqMO0ZFzOiiKtXvKorH7uyMj0piBDF2lJLZnZrOR+V/WTG10EWRcJallik4+ilKBbIxGe6MO14xaMXBAqOZuVkR7RBNq3XXy7gj+7MrzUDst+bjk3+Fi+RAmysE+HMAx+HAOZbiFClSBQhee4AVePeE9e2/e+6R0wZv27MEfeR8/5HyO8g==</latexit> sha1_base64="d/TDdsrptIgjwHYEIVa6RaW7xwQ=">AAAB6nicbVA9TwJBEJ3DL8Qv1NJmI5rYSPa00JLExhKjCAlcyN4yBxv29i67eyaE8BNsLDTG1l9k579xgSsUfMkkL+/NZGZemEphLKXfXmFldW19o7hZ2tre2d0r7x88miTTHBs8kYluhcygFAobVliJrVQji0OJzXB4M/WbT6iNSNSDHaUYxKyvRCQ4s066P7+k3XKFVukMZJn4OalAjnq3/NXpJTyLUVkumTFtn6Y2GDNtBZc4KXUygynjQ9bHtqOKxWiC8ezUCTl1So9EiXalLJmpvyfGLDZmFIeuM2Z2YBa9qfif185sdB2MhUozi4rPF0WZJDYh079JT2jkVo4cYVwLdyvhA6YZty6dkgvBX3x5mTxeVH1a9e9opXaSx1GEIziGM/DhCmpwC3VoAIc+PMMrvHnSe/HevY95a8HLZw7hD7zPH0q5jQY=</latexit> sha1_base64="2L9p5nh5Zd0n0A9q33bilGFiCX8=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquLFkdKEbsaCIy4r2Au1QMmmmDc1khiQjlKGP4MaFIm59Ah/Fne/gA7g0vSy09YfAx/+fQ845fiy4Nhh/Opm5+YXFpexybmV1bX0jv7lV1VGiKKvQSESq7hPNBJesYrgRrB4rRkJfsJrfuxjmtXumNI/knenHzAtJR/KAU2KsdXt4jFv5Ai7ikdAsuBMonH9dnl29977LrfxHsx3RJGTSUEG0brg4Nl5KlOFUsEGumWgWE9ojHdawKEnItJeORh2gfeu0URAp+6RBI/d3R0pCrfuhbytDYrp6Ohua/2WNxASnXsplnBgm6fijIBHIRGi4N2pzxagRfQuEKm5nRbRLFKHGXidnj+BOrzwL1aOii4vuDS6U9mCsLOzALhyACydQgmsoQwUodOABnuDZEc6j8+K8jkszzqRnG/7IefsBtIGRDQ==</latexit> (d) <latexit sha1_base64="rCVsgeNE0e8YDyRFxTSTN12NQJo=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCwhkY8ELmRvmYOVvb3L7p4JIfwCGwuNwdLOv2Pnv3EPKBR8k02evO9MdmaCRHBtXPfbya2tb2xu5bcLO7t7+wfFw6OGjlPFsM5iEatWQDUKLrFuuBHYShTSKBDYDIa3Wd58RKV5LO/NKEE/on3JQ86osVaNdYslt+zORFbBW0Dp5nOa6b3aLX51ejFLI5SGCap123MT44+pMpwJnBQ6qcaEsiHtY9uipBFqfzwbdELOrdMjYazsk4bM3N8dYxppPYoCWxlRM9DLWWb+l7VTE177Yy6T1KBk84/CVBATk2xr0uMKmREjC5QpbmclbEAVZcbepmCP4C2vvAqNy7Lnlr2aW6qcwVx5OIFTuAAPrqACd1CFOjBAeIIXeHUenGfnzZnOS3POoucY/sj5+AHISpFE</latexit> <latexit sha1_base64="/LXIjfRVWR37uD2eTarUm8xNMK0=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMBdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmjiDwMf/38Oc84JEsG1cd1vJ7e2vrG5ld8u7Ozu7R8UD48aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeTvPmIyrNY3lvRgn6Ee1LHnJGjbVqrFssuWV3JrIK3gJKN5+TyTsAVLvFr04vZmmE0jBBtW57bmL8jCrDmcBxoZNqTCgb0j62LUoaofaz2aBjcm6dHgljZZ80ZOb+7shopPUoCmxlRM1AL2dT87+snZrw2s+4TFKDks0/ClNBTEymW5MeV8iMGFmgTHE7K2EDqigz9jYFewRveeVVaFyWPbfs1dxS5QzmysMJnMIFeHAFFbiDKtSBAcITvMCr8+A8O2/OZF6acxY9x/BHzscPaw2Pfw==</latexit> sha1_base64="7Ii0uNmP2RfOgM1BBfSq2ARZFww=">AAAB6HicbVBNT8JAEJ3iF+IX6tHLRjTxRFoveiTx4hESCyTQkO0yhZXtttndmpCGX+DFg8Z49Sd589+4QA8KvmSSl/dmMjMvTAXXxnW/ndLG5tb2Tnm3srd/cHhUPT5p6yRTDH2WiER1Q6pRcIm+4UZgN1VI41BgJ5zczf3OEyrNE/lgpikGMR1JHnFGjZVabFCtuXV3AbJOvILUoEBzUP3qDxOWxSgNE1TrnuemJsipMpwJnFX6mcaUsgkdYc9SSWPUQb44dEYurTIkUaJsSUMW6u+JnMZaT+PQdsbUjPWqNxf/83qZiW6DnMs0MyjZclGUCWISMv+aDLlCZsTUEsoUt7cSNqaKMmOzqdgQvNWX10n7uu65da/l1hoXRRxlOINzuAIPbqAB99AEHxggPMMrvDmPzovz7nwsW0tOMXMKf+B8/gC8H4zF</latexit> <latexit sha1_base64="TEM37w6PcFQLCBxv+NzlDSPPh4U=">AAAB6HicbZA9SwNBEIbn4leMX1FLm8UoWIU7EbQzYGOZgPmA5Ah7m7lkzd7esbsnhJBfYGOhSCzt/Dt2/hv3khQafWHh4X1n2JkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DgQ6RFQ==</latexit> <latexit sha1_base64="zDS/d1EMLhdhjGncUr18mp+9fB4=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuCNoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSPRj1A=</latexit> sha1_base64="G9IPPnJCtUtisXfD2y/+tp8gcfA=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LFbBU0mkoMeCF48t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgobm1vbO8Xd0t7+weFR+fikreNUMWyxWMSqG1CNgktsGW4EdhOFNAoEdoLJ3dzvPKHSPJYPZpqgH9GR5CFn1FipWRuUK27VXYCsEy8nFcjRGJS/+sOYpRFKwwTVuue5ifEzqgxnAmelfqoxoWxCR9izVNIItZ8tDp2RS6sMSRgrW9KQhfp7IqOR1tMosJ0RNWO96s3F/7xeasJbP+MySQ1KtlwUpoKYmMy/JkOukBkxtYQyxe2thI2poszYbEo2BG/15XXSvq56btVrupX6RR5HEc7gHK7Agxuowz00oAUMEJ7hFd6cR+fFeXc+lq0FJ585hT9wPn8AdOOMlg==</latexit> <latexit sha1_base64="iEuqQD02qntqL9OyWq6wx5NE8ZI=">AAAB6nicbZDLSgMxFIZP6q3WW9Wlm2AV3FgybnQjFnThsqK9QDuUTJppQzOZIckIZegjuHGhiFufyJ3v4EOYXhba+kPg4//PIeecIJHCWEK+UG5peWV1Lb9e2Njc2t4p7u7VTZxqxmsslrFuBtRwKRSvWWElbyaa0yiQvBEMrsd545FrI2L1YIcJ9yPaUyIUjFpn3Z96pFMskTKZCC+CN4PS1ffNZQUAqp3iZ7sbszTiyjJJjWl5JLF+RrUVTPJRoZ0anlA2oD3ecqhoxI2fTUYd4WPndHEYa/eUxRP3d0dGI2OGUeAqI2r7Zj4bm/9lrdSGF34mVJJartj0ozCV2MZ4vDfuCs2ZlUMHlGnhZsWsTzVl1l2n4I7gza+8CPWzskfK3h0pVY5gqjwcwCGcgAfnUIFbqEINGPTgCV7gFUn0jN7Q+7Q0h2Y9+/BH6OMH4XKO8A==</latexit> sha1_base64="88ZukDOpZjqFoMNUke8ECvdzrD8=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquLFkutGNWFDEZUV7gXYomTTThmYyQ5IRytBHcONCEbc+gY/iznfwAVyaXhba+kPg4//PIeccPxZcG4w/nczC4tLySnY1t7a+sbmV396p6ShRlFVpJCLV8IlmgktWNdwI1ogVI6EvWN3vX4zy+j1Tmkfyzgxi5oWkK3nAKTHWuj12cTtfwEU8FpoHdwqF86/Ls6v3/nelnf9odSKahEw 6 <latexit sha1_base64="p01vSfJLyQ7pt6DtMz/RCNM7Mb4=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTOQu0ksbGERD4SuJC9ZQ5W9vYuu3smhPALbCw0Bks7/46d/8Y9oFD0TTZ58r4z2ZkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DhBaRFw==</latexit> <latexit sha1_base64="Qe7te+9YtxHKH1q/XgFz9DrYbZE=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuhdoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSbZj1I=</latexit> sha1_base64="qqlJvqD69kDjRh93ah+hTYFl6dI=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFqNgFe4somXAxjIB8wHJEfY2c8mavb1jd08IR36BjYUitv4kO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7ASTu7nfeUKleSwfzDRBP6IjyUPOqLFSszYoV9yquwBZJ15OKpCjMSh/9YcxSyOUhgmqdc9zE+NnVBnOBM5K/VRjQtmEjrBnqaQRaj9bHDojl1YZkjBWtqQhC/X3REYjradRYDsjasZ61ZuL/3m91IS3fsZlkhqUbLkoTAUxMZl/TYZcITNiagllittbCRtTRZmx2ZRsCN7qy+ukfV313KrXdCv1izyOIpzBOVyBBzdQh3toQAsYIDzDK7w5j86L8+58LFsLTj5zCn/gfP4Ad+uMmA==</latexit> U B <latexit sha1_base64="VCjjAZWhE+5inR/AeWjuT606hjU=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquBoyIujOghuXFe0F2qFk0kwbmskMSUYoQx/BjQtlcOPCja/jzrcxvSy09YfAx/+fQ845QSK4Nhh/O4WV1bX1jeJmaWt7Z3evvH/Q0HGqKKvTWMSqFRDNBJesbrgRrJUoRqJAsGYwvJnkzUemNI/lgxklzI9IX/KQU2KsdY/di265gl08FVoGbw6V6888fweAWrf81enFNI2YNFQQrdseToyfEWU4FWxc6qSaJYQOSZ+1LUoSMe1n01HH6NQ6PRTGyj5p0NT93ZGRSOtRFNjKiJiBXswm5n9ZOzXhlZ9xmaSGSTr7KEwFMjGa7I16XDFqxMgCoYrbWREdEEWosdcp2SN4iysvQ+Pc9bDr3eFK9QRmKsIRHMMZeHAJVbiFGtSBQh+e4AVeHeE8O7nzNistOPOeQ/gj5+MH/KGPwg==</latexit> sha1_base64="sCtn71eGklQ5AsfjVDmEuZPAGvw=">AAAB6nicbVBNSwMxEJ2tX7V+VT16CVbB05IVQY8FLx4r2g9ol5JNs21okl2SrFCW/gQvHhTx6i/y5r8xbfegrQ8GHu/NMDMvSgU3FuNvr7S2vrG5Vd6u7Ozu7R9UD49aJsk0ZU2aiER3ImKY4Io1LbeCdVLNiIwEa0fj25nffmLa8EQ92knKQkmGisecEuukB+xf9as17OM50CoJClKDAo1+9as3SGgmmbJUEGO6AU5tmBNtORVsWullhqWEjsmQdR1VRDIT5vNTp+jcKQMUJ9qVsmiu/p7IiTRmIiPXKYkdmWVvJv7ndTMb34Q5V2lmmaKLRXEmkE3Q7G804JpRKyaOEKq5uxXREdGEWpdOxYUQLL+8SlqXfoD94B7X6mdFHGU4gVO4gACuoQ530IAmUBjCM7zCmye8F+/d+1i0lrxi5hj+wPv8AU3CjQg=</latexit> sha1_base64="3K7piSf/jwgTwlHwmGZD/VlF/8s=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquBoyIujOghuXFe0F2qFk0kwbmskMSUYoQx/BjQulunXj67jzbcy0XWjrD4GP/z+HnHOCRHBtMP52Ciura+sbxc3S1vbO7l55/6Ch41RRVqexiFUrIJoJLlndcCNYK1GMRIFgzWB4k+fNR6Y0j+WDGSXMj0hf8pBTYqx1j92LbrmCXTwVWgZvDpXrz0mut1q3/NXpxTSNmDRUEK3bHk6MnxFlOBVsXOqkmiWEDkmftS1KEjHtZ9NRx+jUOj0Uxso+adDU/d2RkUjrURTYyoiYgV7McvO/rJ2a8MrPuExSwySdfRSmApkY5XujHleMGjGyQKjidlZEB0QRaux1SvYI3uLKy9A4dz3sene4Uj2BmYpwBMdwBh5cQhVuoQZ1oNCHJ3iBV0c4z87EeZ+VFpx5zyH8kfPxA1ntkYc=</latexit> 2 <latexit sha1_base64="vV/tSZsqp0PZDp4B5kGJej5F2X0=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmj0U4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUa5ZLnlryaW6ycw1w5OIUzuAQPrqACd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QN+BpET</latexit> <latexit sha1_base64="NShdlnjkNDfv4ynZNqa/uWih2wQ=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0CiXPLfk1dxi5RzmysEpnMEleHAFFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fIMmPTg==</latexit> sha1_base64="uAhRiQUyUfrbUCj+oLzLI3Qe8O4=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VabqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9x24yU</latexit> <latexit sha1_base64="p01vSfJLyQ7pt6DtMz/RCNM7Mb4=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTOQu0ksbGERD4SuJC9ZQ5W9vYuu3smhPALbCw0Bks7/46d/8Y9oFD0TTZ58r4z2ZkJEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeZHnzAZXmsbwzowT9iPYlDzmjxlq1i26x5Jbdmchf8BZQuv6YZnqrdoufnV7M0gilYYJq3fbcxPhjqgxnAieFTqoxoWxI+9i2KGmE2h/PBp2QU+v0SBgr+6QhM/dnx5hGWo+iwFZG1Az0cpaZ/2Xt1IRX/pjLJDUo2fyjMBXExCTbmvS4QmbEyAJlittZCRtQRZmxtynYI3jLK/+FxnnZc8tezS1VTmCuPBzBMZyBB5dQgVuoQh0YIDzCM7w4986T8+pM56U5Z9FzCL/kvH8DhBaRFw==</latexit> <latexit sha1_base64="Qe7te+9YtxHKH1q/XgFz9DrYbZE=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKuhdoZsLFMwFwgWcLs5GwyZnZ2mZkVwpInsLFQJJ3Y+Tp2vo2TS6HRHwY+/v8c5pwTJIJr47pfTm5ldW19I79Z2Nre2d0r7h80dJwqhnUWi1i1AqpRcIl1w43AVqKQRoHAZjC8mebNB1Sax/LOjBL0I9qXPOSMGmvVLrrFklt2ZyJ/wVtA6fpjMnkDgGq3+NnpxSyNUBomqNZtz02Mn1FlOBM4LnRSjQllQ9rHtkVJI9R+Nht0TE6t0yNhrOyThszcnx0ZjbQeRYGtjKgZ6OVsav6XtVMTXvkZl0lqULL5R2EqiInJdGvS4wqZESMLlCluZyVsQBVlxt6mYI/gLa/8FxrnZc8tezW3VDmBufJwBMdwBh5cQgVuoQp1YIDwCM/w4tw7T86rM5mX5pxFzyH8kvP+DSbZj1I=</latexit> sha1_base64="qqlJvqD69kDjRh93ah+hTYFl6dI=">AAAB6HicbVA9SwNBEJ2LXzF+RS1tFqNgFe4somXAxjIB8wHJEfY2c8mavb1jd08IR36BjYUitv4kO/+Nm+QKTXww8Hhvhpl5QSK4Nq777RQ2Nre2d4q7pb39g8Oj8vFJW8epYthisYhVN6AaBZfYMtwI7CYKaRQI7ASTu7nfeUKleSwfzDRBP6IjyUPOqLFSszYoV9yquwBZJ15OKpCjMSh/9YcxSyOUhgmqdc9zE+NnVBnOBM5K/VRjQtmEjrBnqaQRaj9bHDojl1YZkjBWtqQhC/X3REYjradRYDsjasZ61ZuL/3m91IS3fsZlkhqUbLkoTAUxMZl/TYZcITNiagllittbCRtTRZmx2ZRsCN7qy+ukfV313KrXdCv1izyOIpzBOVyBBzdQh3toQAsYIDzDK7w5j86L8+58LFsLTj5zCn/gfP4Ad+uMmA==</latexit> 6 vsyn <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> <latexit sha1_base64="3t+4UbZpWdk262G+bJ4/hq+8w+E=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZN7qz4MZlFXuBdiiZNNOGJpkhyQhl6Bu4caFYV4IrX8edb2N6WWjrD4GP/z+HnHPCRHBjMf72ciura+sb+c3C1vbO7l5x/6Bu4lRTVqOxiHUzJIYJrljNcitYM9GMyFCwRji4nuSNB6YNj9W9HSYskKSneMQpsc6683GnWMJlPBVaBn8OpavP8fgdAKqd4le7G9NUMmWpIMa0fJzYICPacirYqNBODUsIHZAeazlURDITZNNJR+jUOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44BXcEf3HlZaifl31c9m9xqXICM+XhCI7hDHy4gArcQBVqQCGCR3iGF2/gPXmv3tusNOfNew7hj7yPH40aj4c=</latexit> sha1_base64="TBz0KWW1Y5FVHwrZZOy0b06gZmg=">AAAB6XicbVA9TwJBEJ3DL8Qv1NJmI5pYkT0bLUlsLNEIksCF7C17sGFv77I7Z0Iu/AMbC42x9R/Z+W9c4AoFXzLJy3szmZkXpkpapPTbK62tb2xulbcrO7t7+wfVw6O2TTLDRYsnKjGdkFmhpBYtlKhEJzWCxaESj+H4ZuY/PgljZaIfcJKKIGZDLSPJGTrp3qf9ao3W6RxklfgFqUGBZr/61RskPIuFRq6YtV2fphjkzKDkSkwrvcyKlPExG4quo5rFwgb5/NIpOXfKgESJcaWRzNXfEzmLrZ3EoeuMGY7ssjcT//O6GUbXQS51mqHQfLEoyhTBhMzeJgNpBEc1cYRxI92thI+YYRxdOBUXgr/88ippX9Z9WvfvaK1xVsRRhhM4hQvw4QoacAtNaAGHCJ7hFd68sffivXsfi9aSV8wcwx94nz/eLIzN</latexit> sha1_base64="XaQViiIp3HEbSwArRCfYKY95vuc=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXjRncW3LisYi/QDiWTZtrQJDMkGaEMfQM3LhTt1pWv4863MdN2oa0/BD7+/xxyzgkTwY3F+NtbWV1b39gsbBW3d3b39ksHhw0Tp5qyOo1FrFshMUxwxeqWW8FaiWZEhoI1w+FNnjcfmTY8Vg92lLBAkr7iEafEOuvex91SGVfwVGgZ/DmUrz/fc01q3dJXpxfTVDJlqSDGtH2c2CAj2nIq2LjYSQ1LCB2SPms7VEQyE2TTScfozDk9FMXaPWXR1P3dkRFpzEiGrlISOzCLWW7+l7VTG10FGVdJapmis4+iVCAbo3xt1OOaUStGDgjV3M2K6IBoQq07TtEdwV9ceRkaFxUfV/w7XK6ewkwFOIYTOAcfLqEKt1CDOlCI4Ale4NUbes/emzeZla54854j+CPv4wfqV5FM</latexit> <latexit sha1_base64="8JFEz9ga3t3ijN9xsE8ihgO3Gvk=">AAAB6XicbZC7SgNBFIbPeo3xFrW0GYyCVdgVQTsDNpZRzAWSJcxOZpMhs7PLzFkhLHkDGwvFWAlWvo6db+PkUmjiDwMf/38Oc84JEikMuu63s7S8srq2ntvIb25t7+wW9vZrJk4141UWy1g3Amq4FIpXUaDkjURzGgWS14P+9TivP3BtRKzucZBwP6JdJULBKFrrzjtvF4puyZ2ILII3g+LV52j0DgCVduGr1YlZGnGFTFJjmp6boJ9RjYJJPsy3UsMTyvq0y5sWFY248bPJpENyYp0OCWNtn0IycX93ZDQyZhAFtjKi2DPz2dj8L2umGF76mVBJilyx6UdhKgnGZLw26QjNGcqBBcq0sLMS1qOaMrTHydsjePMrL0LtrOS5Je/WLZaPYaocHMIRnIIHF1CGG6hAFRiE8AjP8OL0nSfn1Xmbli45s54D+CPn4weTKo+L</latexit> sha1_base64="CvyJ0JfrVfxBFWefWisicFWF2Eg=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0lE0GPBi8cq9gPaUDbbTbt0swm7E6GE/gMvHhTx6j/y5r9xm+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlsslrHuBtRwKRRvoUDJu4nmNAok7wST27nfeeLaiFg94jThfkRHSoSCUbTSg3c1qNbcupuDrBKvIDUo0BxUv/rDmKURV8gkNabnuQn6GdUomOSzSj81PKFsQke8Z6miETd+ll86I+dWGZIw1rYUklz9PZHRyJhpFNjOiOLYLHtz8T+vl2J442dCJSlyxRaLwlQSjMn8bTIUmjOUU0so08LeStiYasrQhlOxIXjLL6+S9mXdc+vevVtrnBVxlOEETuECPLiGBtxBE1rAIIRneIU3Z+K8OO/Ox6K15BQzx/AHzucP5DyM0Q==</latexit> sha1_base64="W8ftobPAPJmyZ3+M7ZPsIDZ8VGw=">AAAB6XicbZC7SgNBFIbPxluMt6ilzWAUrMKuCNoZsLGMYi6QLGF2cjYZMju7zMwKYckb2FgomtbK17HzbZxcCk38YeDj/89hzjlBIrg2rvvt5FZW19Y38puFre2d3b3i/kFdx6liWGOxiFUzoBoFl1gz3AhsJgppFAhsBIObSd54RKV5LB/MMEE/oj3JQ86osda9d9EpltyyOxVZBm8OpevP94nG1U7xq92NWRqhNExQrVuemxg/o8pwJnBUaKcaE8oGtIcti5JGqP1sOumInFqnS8JY2ScNmbq/OzIaaT2MAlsZUdPXi9nE/C9rpSa88jMuk9SgZLOPwlQQE5PJ2qTLFTIjhhYoU9zOSlifKsqMPU7BHsFbXHkZ6udlzy17d26pcgIz5eEIjuEMPLiECtxCFWrAIIQneIFXZ+A8O2/OeFaac+Y9h/BHzscP8GeRUA==</latexit> <latexit sha1_base64="iEuqQD02qntqL9OyWq6wx5NE8ZI=">AAAB6nicbZDLSgMxFIZP6q3WW9Wlm2AV3FgybnQjFnThsqK9QDuUTJppQzOZIckIZegjuHGhiFufyJ3v4EOYXhba+kPg4//PIeecIJHCWEK+UG5peWV1Lb9e2Njc2t4p7u7VTZxqxmsslrFuBtRwKRSvWWElbyaa0yiQvBEMrsd545FrI2L1YIcJ9yPaUyIUjFpn3Z96pFMskTKZCC+CN4PS1ffNZQUAqp3iZ7sbszTiyjJJjWl5JLF+RrUVTPJRoZ0anlA2oD3ecqhoxI2fTUYd4WPndHEYa/eUxRP3d0dGI2OGUeAqI2r7Zj4bm/9lrdSGF34mVJJartj0ozCV2MZ4vDfuCs2ZlUMHlGnhZsWsTzVl1l2n4I7gza+8CPWzskfK3h0pVY5gqjwcwCGcgAfnUIFbqEINGPTgCV7gFUn0jN7Q+7Q0h2Y9+/BH6OMH4XKO8A==</latexit> sha1_base64="TAV6G+83+fR0ZQt+gj6L/MD1f/w=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgY9iz0TJgYxnRfEByhL3NXLJkb+/Y3RNCyE+wsVDE1l9k579xk1yhiQ8GHu/NMDMvTKUwltJvr7C2vrG5Vdwu7ezu7R+UD4+aJsk0xwZPZKLbITMohcKGFVZiO9XI4lBiKxzdzvzWE2ojEvVoxykGMRsoEQnOrJMeLn3aK1dolc5BVomfkwrkqPfKX91+wrMYleWSGdPxaWqDCdNWcInTUjczmDI+YgPsOKpYjCaYzE+dknOn9EmUaFfKkrn6e2LCYmPGceg6Y2aHZtmbif95ncxGN8FEqDSzqPhiUZRJYhMy+5v0hUZu5dgRxrVwtxI+ZJpx69IpuRD85ZdXSfOq6tOqf08rtbM8jiKcwClcgA/XUIM7qEMDOAzgGV7hzZPei/fufSxaC14+cwx/4H3+AEevjQQ=</latexit> sha1_base64="88ZukDOpZjqFoMNUke8ECvdzrD8=">AAAB6nicbZDLSgMxFIbP1Futt6pLN8EquLFkutGNWFDEZUV7gXYomTTThmYyQ5IRytBHcONCEbc+gY/iznfwAVyaXhba+kPg4//PIeccPxZcG4w/nczC4tLySnY1t7a+sbmV396p6ShRlFVpJCLV8IlmgktWNdwI1ogVI6EvWN3vX4zy+j1Tmkfyzgxi5oWkK3nAKTHWuj12cTtfwEU8FpoHdwqF86/Ls6v3/nelnf9odSKahEwaKojWTRfHxkuJMpwKNsy1Es1iQvuky5oWJQmZ9tLxqEN0aJ0OCiJlnzRo7P7uSEmo9SD0bWVITE/PZiPzv6yZmODUS7mME8MknXwUJAKZCI32Rh2uGDViYIFQxe2siPaIItTY6+TsEdzZleehViq6uOje4EL5ACbKwh7swxG4cAJluIYKVIFCFx7gCZ4d4Tw6L87rpDTjTHt24Y+ctx+xd5EL</latexit> <latexit sha1_base64="8gXrJDo2rDm2cnFDpnm24pOAyEs=">AAAB9XicbVC7TsNAEFyHV0h4BChpLAISVWTTQBmJhjJI5CElJjpfzskp57N1tw5YVr4DGgoQouUX+AY6/oMP4PIoIGGklUYzu9rd8WPBNTrOl5VbWV1b38hvFopb2zu7pb39ho4SRVmdRiJSLZ9oJrhkdeQoWCtWjIS+YE1/eDnxmyOmNI/kDaYx80LSlzzglKCRbkfdDrJ7VGGmUznulspOxZnCXibunJSrxe+HDwCodUufnV5Ek5BJpIJo3XadGL2MKORUsHGhk2gWEzokfdY2VJKQaS+bXj22T4zSs4NImZJoT9XfExkJtU5D33SGBAd60ZuI/3ntBIMLL+MyTpBJOlsUJMLGyJ5EYPe4YhRFagihiptbbTogilA0QRVMCO7iy8ukcVZxnYp7bdI4hhnycAhHcAounEMVrqAGdaCg4BGe4cW6s56sV+tt1pqz5jMH8AfW+w/+OpWh</latexit> sha1_base64="Wfpr9RfiCD93JW9anc3+Z7CQTxI=">AAAB9XicbVA9TwJBEJ3DL8Qv1NLmIppYkTsbLUlsLDGRjwSQ7C17sGF377I7h14u/A8bC42x9b/Y+W9c4AoFXzLJy3szmZkXxIIb9Lxvp7C2vrG5Vdwu7ezu7R+UD4+aJko0ZQ0aiUi3A2KY4Io1kKNg7VgzIgPBWsH4Zua3JkwbHql7TGPWk2SoeMgpQSs9TPpdZE+oZWZSNe2XK17Vm8NdJX5OKpCj3i9/dQcRTSRTSAUxpuN7MfYyopFTwaalbmJYTOiYDFnHUkUkM71sfvXUPbfKwA0jbUuhO1d/T2REGpPKwHZKgiOz7M3E/7xOguF1L+MqTpApulgUJsLFyJ1F4A64ZhRFagmhmttbXToimlC0QZVsCP7yy6ukeVn1vap/51VqZ3kcRTiBU7gAH66gBrdQhwZQ0PAMr/DmPDovzrvzsWgtOPnMMfyB8/kDgN2TCw==</latexit> sha1_base64="G+ykR14Eo10eA6wSvBQ0j1CgUis=">AAAB9XicbVDLSgNBEJz1GRMfUY9eBqPgKex60WPEiwhCBPOAZA2zk0kyZGZ2memNLku+Qy8eFPHq0f/w5n/4AU4eB00saCiquunuCiLBDbjul7OwuLS8sppZy+bWNza38ts7VRPGmrIKDUWo6wExTHDFKsBBsHqkGZGBYLWgfz7yawOmDQ/VDSQR8yXpKt7hlICVbgetJrB70DI1iRq28gW36I6B54k3JYVS7vvh4+zyqtzKfzbbIY0lU0AFMabhuRH4KdHAqWDDbDM2LCK0T7qsYakikhk/HV89xIdWaeNOqG0pwGP190RKpDGJDGynJNAzs95I/M9rxNA59VOuohiYopNFnVhgCPEoAtzmmlEQiSWEam5vxbRHNKFgg8raELzZl+dJ9bjouUXv2qZxgCbIoD20j46Qh05QCV2gMqogijR6RM/oxblznpxX523SuuBMZ3bRHzjvPxLElnA=</latexit> 0 ⇠ 10 1 1 <latexit sha1_base64="3t+4UbZpWdk262G+bJ4/hq+8w+E=">AAAB6XicbZDLSgMxFIbP1Futt6pLN8EquCoZN7qz4MZlFXuBdiiZNNOGJpkhyQhl6Bu4caFYV4IrX8edb2N6WWjrD4GP/z+HnHPCRHBjMf72ciura+sb+c3C1vbO7l5x/6Bu4lRTVqOxiHUzJIYJrljNcitYM9GMyFCwRji4nuSNB6YNj9W9HSYskKSneMQpsc6683GnWMJlPBVaBn8OpavP8fgdAKqd4le7G9NUMmWpIMa0fJzYICPacirYqNBODUsIHZAeazlURDITZNNJR+jUOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44BXcEf3HlZaifl31c9m9xqXICM+XhCI7hDHy4gArcQBVqQCGCR3iGF2/gPXmv3tusNOfNew7hj7yPH40aj4c=</latexit> sha1_base64="TBz0KWW1Y5FVHwrZZOy0b06gZmg=">AAAB6XicbVA9TwJBEJ3DL8Qv1NJmI5pYkT0bLUlsLNEIksCF7C17sGFv77I7Z0Iu/AMbC42x9R/Z+W9c4AoFXzLJy3szmZkXpkpapPTbK62tb2xulbcrO7t7+wfVw6O2TTLDRYsnKjGdkFmhpBYtlKhEJzWCxaESj+H4ZuY/PgljZaIfcJKKIGZDLSPJGTrp3qf9ao3W6RxklfgFqUGBZr/61RskPIuFRq6YtV2fphjkzKDkSkwrvcyKlPExG4quo5rFwgb5/NIpOXfKgESJcaWRzNXfEzmLrZ3EoeuMGY7ssjcT//O6GUbXQS51mqHQfLEoyhTBhMzeJgNpBEc1cYRxI92thI+YYRxdOBUXgr/88ippX9Z9WvfvaK1xVsRRhhM4hQvw4QoacAtNaAGHCJ7hFd68sffivXsfi9aSV8wcwx94nz/eLIzN</latexit> sha1_base64="XaQViiIp3HEbSwArRCfYKY95vuc=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXjRncW3LisYi/QDiWTZtrQJDMkGaEMfQM3LhTt1pWv4863MdN2oa0/BD7+/xxyzgkTwY3F+NtbWV1b39gsbBW3d3b39ksHhw0Tp5qyOo1FrFshMUxwxeqWW8FaiWZEhoI1w+FNnjcfmTY8Vg92lLBAkr7iEafEOuvex91SGVfwVGgZ/DmUrz/fc01q3dJXpxfTVDJlqSDGtH2c2CAj2nIq2LjYSQ1LCB2SPms7VEQyE2TTScfozDk9FMXaPWXR1P3dkRFpzEiGrlISOzCLWW7+l7VTG10FGVdJapmis4+iVCAbo3xt1OOaUStGDgjV3M2K6IBoQq07TtEdwV9ceRkaFxUfV/w7XK6ewkwFOIYTOAcfLqEKt1CDOlCI4Ale4NUbes/emzeZla54854j+CPv4wfqV5FM</latexit> A 05 1 6 2 4 ⌘0 FIG 6 Numer ca cont nuat on resu ts or trave ng ronts (a) two-parameter b urcat on d agram o stat onary homogeneous states o (36) (equ br a o (37)) n the (η0 v yn )-p ane We p ot the ocus o sadd e-node b urcat ons o the steady state co d ng at a cusp b urcat on To the e t o the cusp we dent y a ow-act v ty state U − and a h gh-act v ty one U + wh ch coex st and are stab e n the shaded area Parameter va ues ∆ = 0 5 κ = 5 τ = 1 (b) Trave ng ront profi e computed by so v ng (37) on a truncated doma n −30 30 or η0 = −3 v yn = 4 We show the profi es n the synapt c conductance var ab e G(ξ) and the synchrony var ab e R(ξ) = Z(ξ) (c) The trave ng ront n (b) s a heteroc n c orb t connect ng the equ br a U + and U − o the spat a -dynam ca system (37) We show a pro ect on o the (approx mate) heteroc n c orb t n the (A B G)-space (d) Numer ca cont nuat on o the trave ng ront U (ξ) ound n (b) (b ue) and ts symmetr c counter U (−ξ) (red) us ng η0 as cont nuat on parameter and c as so ut on measure The ronts ve on an so ated branch and destab se at sadd e-node b urcat ons The b urcat on curve s symmetr c w th respect to the ax s c = 0 So ut ons w th c > 0 and c < 0 are re ated v a the trans ormat on U (ξ) → U (−ξ) 5 08 ! 0 06 5 (a) 04 2 (b) 20 10 ! 0 10 20 1 ⌫ 0 1 05 ! 0 ⇠ ! 04 1 ⌫ 2 3 -10 0 ⇠ 2 0 2 4 ⌫ 10 ⌫ 1 2 8 05 ! 0 05 02 0 6 1 04 1 2 1 06 5 0 0 08 ! 0 ⌫ 1 06 0 2 4 10 08 1 0 05 -10 5 (c) Spectrum of U + Profi e R(⇠) Spectrum of U -10 0 ⇠ 10 1 2 0 ⌫ 2 4 FIG 7 Spectrum o the h gh- and ow-act v ty homogeneous states connected by se ected trave ng wave profi es n F g 6 (a) c = −0 8769 (b) c = 0 3594 (c) c = 0 9470 Parameter va ues η0 = −3 v yn = 4 ∆ = 0 5 κ = 5 τ = 1 re ated modu at on of the beta rhythm as seen n MEG An nterest ng extens on of th s work wou d be to nc ude an add t ona dr ve n the neura fie d mode to exam ne how the nc us on of space affects β band moduat ons and perhaps exp a n why β-rebound s seen n both the contra atera and the ps atera hem spheres dur ng movement More genera y the mode parameters can be a tered so that the popu at on osc ates at other frequenc es and hence used to exp a n other eventre ated desynchron sat on/synchron sat on phenomena n 12 the brain. In §IV, we pointed to the existence of a front which connects periodic orbits to nodes/focuses, for the model with an exponential coupling kernel. We observed such fronts by decreasing the synaptic time constant τ (Fig. 8). The numerical machinery used here doesn't allow for the continuation of these solutions, known as defects. The analysis of defects is still an open problem. Close examination of Fig. 8 reveals that there are two fronts in the connection between the node and the limit cycle, which appear to be moving at different speeds. Even more interesting, would be the analysis of fronts which connect two periodic orbits of different amplitudes. The spreading of such a wave across the cortex could be viewed as the spreading of an epileptic seizure. Another numerical challenge would be to continue the exotic patterned states seen in §III when the Turing and Hopf bifurcations collide. These patterns have both a spatial and a temporal period, which would require extending the numerical machineries [5] to continue both a spatial and a temporal pattern. This has been achieved in [4] for the Brusselator model. A natural extension to the work presented in both §III and §IV would be to include a second spatial dimension. It is more natural to view the cortex as a two dimensional sheet and examine the propagation of waves across it. We would expect that the 2D system supports the two dimensional versions of the patterned states presented here, but also potentially some more exotic states. Extending both the Turing analysis and the numerical machinery to include a second spatial extension is worthy of further exploration. VI. ∂g1 ∂t ∂K1 ∂t ∂g2 ∂t ∂K2 ∂t 1 (−g1 + K1 ), τ1 1 = (−K1 + κ1 wm ⊗ f (a + ib)), τ1 1 = (−g2 + K2 ), τ2 1 = (−K2 + κ2 wm ⊗ f (a + ib)). τ2 = The Jacobian of the system can be written as follows: J (k) = J11 J12 , J21 (k) J22 where, J11 J12  ∂ ∂a ∂ ∂a   ∂a ∂t ∂b ∂t  , =   ∂ ∂b ∂ ∂b ∂a ∂t ∂b ∂t   ∂ ∂a ∂ ∂a 0 0  ∂g1 ∂t ∂g2 ∂t      = , 0 ∂ ∂b 0 ∂ ∂b    ∂g1 ∂t ∂g2 ∂t ACKNOWLEDGEMENTS DA was supported by the Engineering and Physical Sciences Research Council under grant EP/P510993/1. SC was supported by the European Commission through the FP7 Marie Curie Initial Training Network 289146, NETT: Neural Engineering Transformative Technologies. Appendix A: Jacobian  ∂f −1 τ1 κ1 w1 ∂a    0  J21 (k) =   −1 ∂f τ2 κ2 w2  ∂a  0 To calculate the Jacobian we first write the system (19) -- (20) in its full six dimensional form, ∂a 1 2 = b(a − 1) − b(a + 1)(η0 + vsyn g1 + vsyn g2 ) ∂t 1 − (a + 1)∆ − (a2 − b2 − 1)(∆ + g1 + g2 ), 2 ∂b 1 2 = − ((a − 1) − b2 ) − b∆ − ab(∆ + g1 + g2 ) ∂t 2 1 1 2 + ((a + 1)2 − b2 )(η0 + vsyn g1 + vsyn g2 ), 2 J22  −1 −τ1  −1  τ1 =   0 0  ∂f ∂b     0  , ∂f  −1 τ2 κ2 w2  ∂b   0 −1 τ1 κ1 w1 0 0 −1 −τ1 0 0 0 −1 −τ2 −1 τ2 0   0  .  0  −1 −τ2 The variables, a, b, g1 and g2 are evaluated at the steady state, and hence, depend upon the control parameters, wi depend on k and β and f is given by (17). 13 13 13 (a) (a) (a) (b) (b) (b) <latexit sha1_base64="ka3t2/48Wh23hsAl88gQOgU1RLk=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqMQm3CXRsuAjWVE8wHJEfY2e8mSvb1jd04IR36CjYUitv4iO/+Nm+QKTXww8Hhvhpl5QSKFQdf9dgobm1vbO8Xd0t7+weFR+fikbeJUM95isYx1N6CGS6F4CwVK3k00p1EgeSeY3M79zhPXRsTqEacJ9yM6UiIUjKKVHqr0alCuuDV3AbJOvJxUIEdzUP7qD2OWRlwhk9SYnucm6GdUo2CSz0r91PCEsgkd8Z6likbc+Nni1Bm5tMqQhLG2pZAs1N8TGY2MmUaB7Ywojs2qNxf/83ophjd+JlSSIldsuShMJcGYzP8mQ6E5Qzm1hDIt7K2EjammDG06JRuCt/ryOmnXa55b8+7rlcZFHkcRzuAcquDBNTTgDprQAgYjeIZXeHOk8+K8Ox/L1oKTz5zCHzifP38FjSo=</latexit> 20 16 <latexit sha1_base64="QemqZr6fsruN4EG6fJtLmwgab9A=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXTje4suHFZxV6gHUomzbShSWZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wwEdxYjL+9ldW19Y3N3FZ+e2d3b79wcFg3caopq9FYxLoZEsMEV6xmuRWsmWhGZChYIxxcT/LGA9OGx+reDhMWSNJTPOKUWGfdlXGnUMQlPBVaBn8OxavP8fgdAKqdwle7G9NUMmWpIMa0fJzYICPacirYKN9ODUsIHZAeazlURDITZNNJR+jMOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44eXcEf3HlZaiXSz4u+be4WDmFmXJwDCdwDj5cQAVuoAo1oBDBIzzDizfwnrxX721WuuLNe47gj7yPH46fj4g=</latexit> sha1_base64="LOhMDFdaD7Ni8csKulxBbULGOyw=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxiPiA5wt5mL1myt3fszgkh5B/YWChi6z+y89+4Sa7QxAcDj/dmmJkXpkpapPTbK2xsbm3vFHdLe/sHh0fl45OWTTLDRZMnKjGdkFmhpBZNlKhEJzWCxaES7XB8O/fbT8JYmehHnKQiiNlQy0hyhk56qNF+uUKrdAGyTvycVCBHo1/+6g0SnsVCI1fM2q5PUwymzKDkSsxKvcyKlPExG4quo5rFwgbTxaUzcumUAYkS40ojWai/J6YstnYSh64zZjiyq95c/M/rZhjdBFOp0wyF5stFUaYIJmT+NhlIIziqiSOMG+luJXzEDOPowim5EPzVl9dJq1b1adW/p5X6RR5HEc7gHK7Ah2uowx00oAkcIniGV3jzxt6L9+59LFsLXj5zCn/gff4A37GMzg==</latexit> sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiBU7gADy6hBrdQhwZQCOEJXuDVGTnPzpsznZcWnEXPMfyR8/ED69yRTQ==</latexit> 20 <latexit sha1_base64="QemqZr6fsruN4EG6fJtLmwgab9A=">AAAB6XicbZDLSgMxFIbPeK31VnXpJlgFVyXTje4suHFZxV6gHUomzbShSWZIMkIZ+gZuXCjWleDK13Hn25heFtr6Q+Dj/88h55wwEdxYjL+9ldW19Y3N3FZ+e2d3b79wcFg3caopq9FYxLoZEsMEV6xmuRWsmWhGZChYIxxcT/LGA9OGx+reDhMWSNJTPOKUWGfdlXGnUMQlPBVaBn8OxavP8fgdAKqdwle7G9NUMmWpIMa0fJzYICPacirYKN9ODUsIHZAeazlURDITZNNJR+jMOV0Uxdo9ZdHU/d2REWnMUIauUhLbN4vZxPwva6U2ugwyrpLUMkVnH0WpQDZGk7VRl2tGrRg6IFRzNyuifaIJte44eXcEf3HlZaiXSz4u+be4WDmFmXJwDCdwDj5cQAVuoAo1oBDBIzzDizfwnrxX721WuuLNe47gj7yPH46fj4g=</latexit> sha1_base64="LOhMDFdaD7Ni8csKulxBbULGOyw=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxiPiA5wt5mL1myt3fszgkh5B/YWChi6z+y89+4Sa7QxAcDj/dmmJkXpkpapPTbK2xsbm3vFHdLe/sHh0fl45OWTTLDRZMnKjGdkFmhpBZNlKhEJzWCxaES7XB8O/fbT8JYmehHnKQiiNlQy0hyhk56qNF+uUKrdAGyTvycVCBHo1/+6g0SnsVCI1fM2q5PUwymzKDkSsxKvcyKlPExG4quo5rFwgbTxaUzcumUAYkS40ojWai/J6YstnYSh64zZjiyq95c/M/rZhjdBFOp0wyF5stFUaYIJmT+NhlIIziqiSOMG+luJXzEDOPowim5EPzVl9dJq1b1adW/p5X6RR5HEc7gHK7Ah2uowx00oAkcIniGV3jzxt6L9+59LFsLXj5zCn/gff4A37GMzg==</latexit> sha1_base64="Ep4JJqG/YQT2v9XWicYV5I5p4bs=">AAAB6XicbZC7TsMwFIZPyq2UW4CRxaIgMVVOF9ioxMJYEL1IbVQ5rtNadZzIdpCqqG/AwgCCrky8Dhtvg9N2gJZfsvTp/8+RzzlBIrg2GH87hbX1jc2t4nZpZ3dv/8A9PGrqOFWUNWgsYtUOiGaCS9Yw3AjWThQjUSBYKxjd5HnrkSnNY/lgxgnzIzKQPOSUGGvdV3HPLeMKngmtgreA8vXne65pved+dfsxTSMmDRVE646HE+NnRBlOBZuUuqlmCaEjMmAdi5JETPvZbNIJOrdOH4Wxsk8aNHN/d2Qk0nocBbYyImaol7Pc/C/rpCa88jMuk9QwSecfhalAJkb52qjPFaNGjC0QqridFdEhUYQae5ySPYK3vPIqNKsVD1e8O1yuncFcRTiBU7gADy6hBrdQhwZQCOEJXuDVGTnPzpsznZcWnEXPMfyR8/ED69yRTQ==</latexit> <latexit sha1_base64="VOoKRtLZHEXx8Q5B69L4Znh8o9U=">AAAB6XicbZC7SgNBFIbPeo3xFrW0GYyCVdi1UDsDNpZRzAWSJcxOZpMhs7PLzFkhLHkDGwvFWAlWvo6db+PkUmjiDwMf/38Oc84JEikMuu63s7S8srq2ntvIb25t7+wW9vZrJk4141UWy1g3Amq4FIpXUaDkjURzGgWS14P+9TivP3BtRKzucZBwP6JdJULBKFrrzjtvF4puyZ2ILII3g+LV52j0DgCVduGr1YlZGnGFTFJjmp6boJ9RjYJJPsy3UsMTyvq0y5sWFY248bPJpENyYp0OCWNtn0IycX93ZDQyZhAFtjKi2DPz2dj8L2umGF76mVBJilyx6UdhKgnGZLw26QjNGcqBBcq0sLMS1qOaMrTHydsjePMrL0LtrOS5Je/WLZaPYaocHMIRnIIHF1CGG6hAFRiE8AjP8OL0nSfn1Xmbli45s54D+CPn4weWMo+N</latexit> sha1_base64="vJGKWfYypACzq1YKgzLOrYsRbkA=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0k8qMeCF49V7Ae0oWy2m3bpZhN2J0IJ/QdePCji1X/kzX/jNs1BWx8MPN6bYWZekEhh0HW/ndLa+sbmVnm7srO7t39QPTxqmzjVjLdYLGPdDajhUijeQoGSdxPNaRRI3gkmt3O/88S1EbF6xGnC/YiOlAgFo2ilB+9qUK25dTcHWSVeQWpQoDmofvWHMUsjrpBJakzPcxP0M6pRMMlnlX5qeELZhI54z1JFI278LL90Rs6tMiRhrG0pJLn6eyKjkTHTKLCdEcWxWfbm4n9eL8Xwxs+ESlLkii0WhakkGJP522QoNGcop5ZQpoW9lbAx1ZShDadiQ/CWX14l7cu659a9e7fWOCviKMMJnMIFeHANDbiDJrSAQQjP8ApvzsR5cd6dj0VrySlmjuEPnM8f50SM0w==</latexit> sha1_base64="QjWfb9jF5DpSDol7iRWsNBO7stA=">AAAB6XicbZC7SgNBFIbPxluMt6ilzWAUrMKuhdoZsLGMYi6QLGF2cjYZMju7zMwKYckb2FgomtbK17HzbZxcCk38YeDj/89hzjlBIrg2rvvt5FZW19Y38puFre2d3b3i/kFdx6liWGOxiFUzoBoFl1gz3AhsJgppFAhsBIObSd54RKV5LB/MMEE/oj3JQ86osda9d9EpltyyOxVZBm8OpevP94nG1U7xq92NWRqhNExQrVuemxg/o8pwJnBUaKcaE8oGtIcti5JGqP1sOumInFqnS8JY2ScNmbq/OzIaaT2MAlsZUdPXi9nE/C9rpSa88jMuk9SgZLOPwlQQE5PJ2qTLFTIjhhYoU9zOSlifKsqMPU7BHsFbXHkZ6udlzy17d26pcgIz5eEIjuEMPLiECtxCFWrAIIQneIFXZ+A8O2/OeFaac+Y9h/BHzscP82+RUg==</latexit> <latexit sha1_base64="SrO2bRa0WdAu7JXEQrWADLenrn0=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCwhkY8ELmRvmYOVvb3L7p4JIfwCGwuNwdLOv2Pnv3EPKBR8k02evO9MdmaCRHBtXPfbya2tb2xu5bcLO7t7+wfFw6OGjlPFsM5iEatWQDUKLrFuuBHYShTSKBDYDIa3Wd58RKV5LO/NKEE/on3JQ86osVbNdIslt+zORFbBW0Dp5nOa6b3aLX51ejFLI5SGCap123MT44+pMpwJnBQ6qcaEsiHtY9uipBFqfzwbdELOrdMjYazsk4bM3N8dYxppPYoCWxlRM9DLWWb+l7VTE177Yy6T1KBk84/CVBATk2xr0uMKmREjC5QpbmclbEAVZcbepmCP4C2vvAqNy7Lnlr2aW6qcwVx5OIFTuAAPrqACd1CFOjBAeIIXeHUenGfnzZnOS3POoucY/sj5+AHiDpFV</latexit> <latexit sha1_base64="zJ7fb7NcMlJMd2W8pIpS4sZe4uI=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMBdIljA7mU3GzM4uM2eFsOQJbCwUSSd2vo6db+PkUmjiDwMf/38Oc84JEikMuu63k1tb39jcym8Xdnb39g+Kh0cNE6ea8TqLZaxbATVcCsXrKFDyVqI5jQLJm8Hwdpo3H7k2Ilb3OEq4H9G+EqFgFK1Vw26x5JbdmcgqeAso3XxOJu8AUO0Wvzq9mKURV8gkNabtuQn6GdUomOTjQic1PKFsSPu8bVHRiBs/mw06JufW6ZEw1vYpJDP3d0dGI2NGUWArI4oDs5xNzf+ydorhtZ8JlaTIFZt/FKaSYEymW5Oe0JyhHFmgTAs7K2EDqilDe5uCPYK3vPIqNC7Lnlv2am6pcgZz5eEETuECPLiCCtxBFerAgMMTvMCr8+A8O2/OZF6acxY9x/BHzscPhNGPkA==</latexit> sha1_base64="I2Iy4gyD75g/wQG3iiULld6zRIU=">AAAB6HicbVDLTgJBEOzFF+IL9ehlIpp4Irte9EjixSMk8khgQ2aHXhiZfWSm14QQvsCLB43x6id5828cYA8KVtJJpao73V1BqqQh1/12ChubW9s7xd3S3v7B4VH5+KRlkkwLbIpEJboTcINKxtgkSQo7qUYeBQrbwfhu7refUBuZxA80SdGP+DCWoRScrNSgfrniVt0F2DrxclKBHPV++as3SEQWYUxCcWO6npuSP+WapFA4K/UygykXYz7ErqUxj9D408WhM3ZplQELE20rJrZQf09MeWTMJApsZ8RpZFa9ufif180ovPWnMk4zwlgsF4WZYpSw+ddsIDUKUhNLuNDS3srEiGsuyGZTsiF4qy+vk9Z11XOrXsOt1C7yOIpwBudwBR7cQA3uoQ5NEIDwDK/w5jw6L86787FsLTj5zCn8gfP5A9XjjNY=</latexit> <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> t R <latexit sha1_base64="iKaRjd0LPTdxRu1o0jMO04B10+I=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOCzBRiYWxCHqR2qhyXKe16jiR7SBVUR+BhQEUsTCw8DpsvA3uZYCWX7L06f/Pkc85QSK4Nhh/O4WV1bX1jeJmaWt7Z3evvH/Q0HGqKKvTWMSqFRDNBJesbrgRrJUoRqJAsGYwvJnkzUemNI/lgxklzI9IX/KQU2KsdY/dq265gl08FVoGbw6V6888fweAWrf81enFNI2YNFQQrdseToyfEWU4FWxc6qSaJYQOSZ+1LUoSMe1n01HH6NQ6PRTGyj5p0NT93ZGRSOtRFNjKiJiBXswm5n9ZOzXhpZ9xmaSGSTr7KEwFMjGa7I16XDFqxMgCoYrbWREdEEWosdcp2SN4iysvQ+Pc9bDr3eFK9QRmKsIRHMMZeHABVbiFGtSBQh+e4AVeHeE8O7nzNistOPOeQ/gj5+MHBESPxw==</latexit> sha1_base64="3fi41L6rWpOk sha1_base64="3fi41L6rWpOkNKkUnCUAyJVaLj0=">AAAB6nicbZC7TsMwFIZPyq2UW4GRxaIgMUUOCzBRiYWxCHqR2qhyXKe16jiR7SBVUR+BhQFUWFl4HTbeBqftAC2/ZOnT/58jn3OCRHBtMP52Ciura+sbxc3S1vbO7l55/6Ch41RRVqexiFUrIJoJLlndcCNYK1GMRIFgzWB4k+fNR6Y0j+WDGSXMj0hf8pBTYqx1j92rbrmCXTwVWgZvDpXrz0mut1q3/NXpxTSNmDRUEK3bHk6MnxFlOBVsXOqkmiWEDkmftS1KEjHtZ9NRx+jUOj0Uxso+adDU/d2RkUjrURTYyoiYgV7McvO/rJ2a8NLPuExSwySdfRSmApkY5XujHleMGjGyQKjidlZEB0QRaux1SvYI3uLKy9A4dz3sene4Uj2BmYpwBMdwBh5cQBVuoQZ1oNCHJ3iBV0c4z87EeZ+VFpx5zyH8kfPxA2GBkYw=</latexit> <latexit sha1_base64="SrO2bRa0WdAu7JXEQrWADLenrn0=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCwhkY8ELmRvmYOVvb3L7p4JIfwCGwuNwdLOv2Pnv3EPKBR8k02evO9MdmaCRHBtXPfbya2tb2xu5bcLO7t7+wfFw6OGjlPFsM5iEatWQDUKLrFuuBHYShTSKBDYDIa3Wd58RKV5LO/NKEE/on3JQ86osVbNdIslt+zORFbBW0Dp5nOa6b3aLX51ejFLI5SGCap123MT44+pMpwJnBQ6qcaEsiHtY9uipBFqfzwbdELOrdMjYazsk4bM3N8dYxppPYoCWxlRM9DLWWb+l7VTE177Yy6T1KBk84/CVBATk2xr0uMKmREjC5QpbmclbEAVZcbepmCP4C2vvAqNy7Lnlr2aW6qcwVx5OIFTuAAPrqACd1CFOjBAeIIXeHUenGfnzZnOS3POoucY/sj5+AHiDpFV</latexit> <latexit sha1_base64="zJ7fb7NcMlJMd2W8pIpS4sZe4uI=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMBdIljA7mU3GzM4uM2eFsOQJbCwUSSd2vo6db+PkUmjiDwMf/38Oc84JEikMuu63k1tb39jcym8Xdnb39g+Kh0cNE6ea8TqLZaxbATVcCsXrKFDyVqI5jQLJm8Hwdpo3H7k2Ilb3OEq4H9G+EqFgFK1Vw26x5JbdmcgqeAso3XxOJu8AUO0Wvzq9mKURV8gkNabtuQn6GdUomOTjQic1PKFsSPu8bVHRiBs/mw06JufW6ZEw1vYpJDP3d0dGI2NGUWArI4oDs5xNzf+ydorhtZ8JlaTIFZt/FKaSYEymW5Oe0JyhHFmgTAs7K2EDqilDe5uCPYK3vPIqNC7Lnlv2am6pcgZz5eEETuECPLiCCtxBFerAgMMTvMCr8+A8O2/OZF6acxY9x/BHzscPhNGPkA==</latexit> sha1_base64="I2Iy4gyD75g/wQG3iiULld6zRIU=">AAAB6HicbVDLTgJBEOzFF+IL9ehlIpp4Irte9EjixSMk8khgQ2aHXhiZfWSm14QQvsCLB43x6id5828cYA8KVtJJpao73V1BqqQh1/12ChubW9s7xd3S3v7B4VH5+KRlkkwLbIpEJboTcINKxtgkSQo7qUYeBQrbwfhu7refUBuZxA80SdGP+DCWoRScrNSgfrniVt0F2DrxclKBHPV++as3SEQWYUxCcWO6npuSP+WapFA4K/UygykXYz7ErqUxj9D408WhM3ZplQELE20rJrZQf09MeWTMJApsZ8RpZFa9ufif180ovPWnMk4zwlgsF4WZYpSw+ddsIDUKUhNLuNDS3srEiGsuyGZTsiF4qy+vk9Z11XOrXsOt1C7yOIpwBudwBR7cQA3uoQ5NEIDwDK/w5jw6L86787FsLTj5zCn8gfP5A9XjjNY=</latexit> <latexit sha1_base64="Bm6L7INbJx8XEYQzKxIxOiRrWCU=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCzByEcCF7K3zMHK3t5ld8+EEH6BjYXGYGnn37Hz37gHFIq+ySZP3ncmOzNBIrg2rvvl5FZW19Y38puFre2d3b3i/kFDx6liWGexiFUroBoFl1g33AhsJQppFAhsBsPrLG8+oNI8lndmlKAf0b7kIWfUWKt22y2W3LI7E/kL3gJKVx/TTG/VbvGz04tZGqE0TFCt256bGH9MleFM4KTQSTUmlA1pH9sWJY1Q++PZoBNyap0eCWNlnzRk5v7sGNNI61EU2MqImoFezjLzv6ydmvDSH3OZpAYlm38UpoKYmGRbkx5XyIwYWaBMcTsrYQOqKDP2NgV7BG955b/QOC97btmruaXKCcyVhyM4hjPw4AIqcANVqAMDhEd4hhfn3nlyXp3pvDTnLHoO4Zec92+uhpEz</latexit> <latexit sha1_base64="qoVWhpdOghMvJ6ZyOV73cMAR63I=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMRdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmj0h4GP/z+HOecEieDauO6Xk1tZXVvfyG8WtrZ3dveK+wcNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQyvp3nzAZXmsbwzowT9iPYlDzmjxlq1226x5Jbdmchf8BZQuvqYTN4AoNotfnZ6MUsjlIYJqnXbcxPjZ1QZzgSOC51UY0LZkPaxbVHSCLWfzQYdk1Pr9EgYK/ukITP3Z0dGI61HUWArI2oGejmbmv9l7dSEl37GZZIalGz+UZgKYmIy3Zr0uEJmxMgCZYrbWQkbUEWZsbcp2CN4yyv/hcZ52XPLXs0tVU5grjwcwTGcgQcXUIEbqEIdGCA8wjO8OPfOk/PqTOalOWfRcwi/5Lx/A1FJj24=</latexit> sha1_base64="af6fmYZ3mFiUWU/zWR5QDen5nZM=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLElsLMHIRwIXsrfMwcre3mV3z4Rc+AU2Fhpj60+y89+4wBUKvmSSl/dmMjMvSATXxnW/ncLG5tb2TnG3tLd/cHhUPj5p6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWByO/c7T6g0j+WDmSboR3QkecgZNVZq3g/KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGNn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o13Ur9Io+jCGdwDlfgQQ3qcAcNaAEDhGd4hTfn0Xlx3p2PZWvByWdO4Q+czx+iW4y0</latexit> <latexit sha1_base64="vV/tSZsqp0PZDp4B5kGJej5F2X0=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmj0U4SG0tI5CMBQvaWOVjZ27vs7pmQC7/AxkJjsLTz79j5b9wDCgXfZJMn7zuTnRk/Flwb1/121tY3Nre2czv53b39g8PC0XFDR4liWGeRiFTLpxoFl1g33AhsxQpp6Ats+qPbLG8+otI8kvdmHGM3pAPJA86osVat3CsU3ZI7E1kFbwHFm89ppvdqr/DV6UcsCVEaJqjWbc+NTTelynAmcJLvJBpjykZ0gG2Lkoaou+ls0Am5sE6fBJGyTxoyc393pDTUehz6tjKkZqiXs8z8L2snJrjuplzGiUHJ5h8FiSAmItnWpM8VMiPGFihT3M5K2JAqyoy9Td4ewVteeRUa5ZLnlryaW6ycw1w5OIUzuAQPrqACd1CFOjBAeIIXeHUenGfnzZnOS9ecRc8J/JHz8QN+BpET</latexit> <latexit sha1_base64="NShdlnjkNDfv4ynZNqa/uWih2wQ=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdhNo50BG8sEzAWSJcxOziZjZmeXmVkhLHkCGwtF0omdr2Pn2zi5FJr4w8DH/5/DnHOCRHBtXPfbWVvf2Nzazu3kd/f2Dw4LR8cNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQxvp3nzEZXmsbw3owT9iPYlDzmjxlq1crdQdEvuTGQVvAUUbz4nk3cAqHYLX51ezNIIpWGCat323MT4GVWGM4HjfCfVmFA2pH1sW5Q0Qu1ns0HH5MI6PRLGyj5pyMz93ZHRSOtRFNjKiJqBXs6m5n9ZOzXhtZ9xmaQGJZt/FKaCmJhMtyY9rpAZMbJAmeJ2VsIGVFFm7G3y9gje8sqr0CiXPLfk1dxi5RzmysEpnMEleHAFFbiDKtSBAcITvMCr8+A8O2/OZF665ix6TuCPnI8fIMmPTg==</latexit> sha1_base64="uAhRiQUyUfrbUCj+oLzLI3Qe8O4=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUqg3LFbfqLkE2iZeTCuRoDstfg1HM0gilYYJq3ffcxPgZVYYzgfPSINWYUDalY+xbKmmE2s+Wh87JtVVGJIyVLWnIUv09kdFI61kU2M6Imole9xbif14/NeGtn3GZpAYlWy0KU0FMTBZfkxFXyIyYWUKZ4vZWwiZUUWZsNiUbgrf+8ibp1KqeW/VabqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9x24yU</latexit> <latexit sha1_base64="1Ha30VsQtYTp3nW5WCh6LZpzBGw=">AAAB5HicbZA9T8MwEIYv5auEr8LKYlGQmKqEBTYqsTAWiX5IbVQ57qU1dZzIdpCqqL+AhQGEYGLl77Dxb3BaBmh5JUuP3vdOvrswFVwbz/tySiura+sb5U13a3tnd6/i7rd0kimGTZaIRHVCqlFwiU3DjcBOqpDGocB2OL4q8vY9Ks0TeWsmKQYxHUoecUaNtW68fqXq1byZyDL4P1C9/Hgt9NboVz57g4RlMUrDBNW663upCXKqDGcCp24v05hSNqZD7FqUNEYd5LNBp+TEOgMSJco+acjM/d2R01jrSRzaypiakV7MCvO/rJuZ6CLIuUwzg5LNP4oyQUxCiq3JgCtkRkwsUKa4nZWwEVWUGXsb1x7BX1x5GVpnNd+r+dX6McxVhkM4glPw4RzqcA0NaAIDhAd4gmfnznl0XuaFJeen4wD+yHn/BgA4j+s=</latexit> <latexit sha1_base64="DhnS4SVaUrVH9wMXxPE2LEZZOpU=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKsjXYGbCwTMBdIljA7OZuMmb0wMyuEJU9gY6FIOrHzdex8GyeXQhN/GPj4/3OYc46fSKENpd9Obm19Y3Mrv13Y2d3bPygeHjV0nCqOdR7LWLV8plGKCOtGGImtRCELfYlNf3g7zZuPqLSIo3szStALWT8SgeDMWKtGu8USLdOZyCq4CyjdfE4m7wBQ7Ra/Or2YpyFGhkumddulifEypozgEseFTqoxYXzI+ti2GLEQtZfNBh2Tc+v0SBAr+yJDZu7vjoyFWo9C31aGzAz0cjY1/8vaqQmuvUxESWow4vOPglQSE5Pp1qQnFHIjRxYYV8LOSviAKcaNvU3BHsFdXnkVGpdll5bdGi1VzmCuPJzAKVyAC1dQgTuoQh04IDzBC7w6D86z8+ZM5qU5Z9FzDH/kfPwAHcGPTA==</latexit> sha1_base64="PuXa0rEGfX7FfmF6HqDLb+ymO3E=">AAAB5HicbVA9SwNBEJ3zM55f0dZmMQpWYc9Gy4CNZQTzAckR9jZzyZq9vWN3TwhHfoGNhWLrb7Lz37hJrtDEBwOP92aYmRdlUhhL6be3sbm1vbNb2fP3Dw6Pjqv+SdukuebY4qlMdTdiBqVQ2LLCSuxmGlkSSexEk7u533lGbUSqHu00wzBhIyViwZl10gMdVGu0Thcg6yQoSQ1KNAfVr/4w5XmCynLJjOkFNLNhwbQVXOLM7+cGM8YnbIQ9RxVL0ITF4tAZuXTKkMSpdqUsWai/JwqWGDNNIteZMDs2q95c/M/r5Ta+DQuhstyi4stFcS6JTcn8azIUGrmVU0cY18LdSviYacaty8Z3IQSrL6+T9nU9oPWg1rgow6jAGZzDFQRwAw24hya0gAPCC7zBu/fkvXofy8YNr5w4hT/wPn8ABgmLbA==</latexit> <latexit sha1_base64="1cX1pw6RAXuT7gr52Ruqbpbqy0Y=">AAAB63icbVBNSwMxEJ2tX7V+VT16CVbBiyXbix4LXjxWsB/QLiWbZtvQJLskWaEs/QtePCji1T/kzX9jtt2Dtj4YeLw3w8y8MBHcWIy/vdLG5tb2Tnm3srd/cHhUPT7pmDjVlLVpLGLdC4lhgivWttwK1ks0IzIUrBtO73K/+8S04bF6tLOEBZKMFY84JTaXrn2Mh9UaruMF0DrxC1KDAq1h9WswimkqmbJUEGP6Pk5skBFtORVsXhmkhiWETsmY9R1VRDITZItb5+jSKSMUxdqVsmih/p7IiDRmJkPXKYmdmFUvF//z+qmNboOMqyS1TNHloigVyMYofxyNuGbUipkjhGrubkV0QjSh1sVTcSH4qy+vk06j7uO6/9CoNS+KOMpwBudwBT7cQBPuoQVtoDCBZ3iFN096L96797FsLXnFzCn8gff5A7aFjUA=</latexit> -100 -50 <latexit sha1_base64="vQWLT/ViCnpqQmH+dGECu1vyhWY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpiB4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw+WV2y9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5ufOiXnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhDd+xmWSGpRssShMBTExmf1NBlwhM2JiCWWK21sJG1FFmbHplGwI3vLLq6RVq3pu1buvVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFOY40K</latexit> 0 x <latexit <latexit sha1_base64="kRDnxcx7hyKVH6+joAsSPWBDiHM=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUcoflilt1lyCbxMtJBXI0h+WvwShmaYTSMEG17ntuYvyMKsOZwHlpkGpMKJvSMfYtlTRC7WfLQ+fk2iojEsbKljRkqf6eyGik9SwKbGdEzUSvewvxP6+fmvDWz7hMUoOSrRaFqSAmJouvyYgrZEbMLKFMcXsrYROqKDM2m5INwVt/eZN0alXPrXqtWqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9vc4yU</latexit> <latexit sha1_base64="79wqBERO1FRfA1Dy5mKGWRmSyn4=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmz0U4SG0tI5CMBQvaWOVjZ27vs7hnJhV9gY6ExWNr5d+z8N+4BhYJvssmT953JzowfC66N6347K6tr6xubua389s7u3n7h4LCuo0QxrLFIRKrpU42CS6wZbgQ2Y4U09AU2/OFNljceUGkeyTszirET0r7kAWfUWKv62C0U3ZI7FVkGbw7F689JpvdKt/DV7kUsCVEaJqjWLc+NTSelynAmcJxvJxpjyoa0jy2LkoaoO+l00DE5s06PBJGyTxoydX93pDTUehT6tjKkZqAXs8z8L2slJrjqpFzGiUHJZh8FiSAmItnWpMcVMiNGFihT3M5K2IAqyoy9Td4ewVtceRnqFyXPLXlVt1g+hZlycAwncA4eXEIZbqECNWCA8AQv8OrcO8/OmzOZla44854j+CPn4wfoHpFZ</latexit> <latexit sha1_base64="gBTNO7IH+P1mduEskbFWJ6WrbPE=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdi10c6AjWUC5gLJEmYnZ5Mxs7PLzKwYljyBjYUi6cTO17HzbZxcCk38YeDj/89hzjlBIrg2rvvtrKyurW9s5rby2zu7e/uFg8O6jlPFsMZiEatmQDUKLrFmuBHYTBTSKBDYCAY3k7zxgErzWN6ZYYJ+RHuSh5xRY63qY6dQdEvuVGQZvDkUrz/H43cAqHQKX+1uzNIIpWGCat3y3MT4GVWGM4GjfDvVmFA2oD1sWZQ0Qu1n00FH5Mw6XRLGyj5pyNT93ZHRSOthFNjKiJq+Xswm5n9ZKzXhlZ9xmaQGJZt9FKaCmJhMtiZdrpAZMbRAmeJ2VsL6VFFm7G3y9gje4srLUL8oeW7Jq7rF8inMlINjOIFz8OASynALFagBA4QneIFX5955dt6c8ax0xZn3HMEfOR8/iuGPlA==</latexit> sha1_base64="lEcjT01rh+kON+avtnUjSMl7nNU=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLElsLCGRjwQuZG+Zg5W9vcvunpFc+AU2Fhpj60+y89+4wBUKvmSSl/dmMjMvSATXxnW/ncLG5tb2TnG3tLd/cHhUPj5p6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWByO/c7j6g0j+W9mSboR3QkecgZNVZqPg3KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGNn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o13Ur9Io+jCGdwDlfgQQ3qcAcNaAEDhGd4hTfnwXlx3p2PZWvByWdO4Q+czx/b84za</latexit> 50 <latexit sha1_base64="BWuP9VBNfr/tVrmk+3RJQHf+H1w=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0kKRY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKD3V3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17muVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AOTgjNM=</latexit> 100 <latexit sha1_base64="XiFhIYx2rFbNJAJ/xJ6g+YVEH2k=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxoPiA5wt5mkizZ2zt294Rw5CfYWChi6y+y89+4Sa7QxAcDj/dmmJkXJlIYS+m3V9jY3NreKe6W9vYPDo/KxyctE6eaY5PHMtadkBmUQmHTCiuxk2hkUSixHU5u5377CbURsXq00wSDiI2UGArOrJMefEr75Qqt0gXIOvFzUoEcjX75qzeIeRqhslwyY7o+TWyQMW0Flzgr9VKDCeMTNsKuo4pFaIJsceqMXDplQIaxdqUsWai/JzIWGTONQtcZMTs2q95c/M/rpnZ4E2RCJalFxZeLhqkkNibzv8lAaORWTh1hXAt3K+Fjphm3Lp2SC8FffXmdtGpVn1b9+1qlfpHHUYQzOIcr8OEa6nAHDWgChxE8wyu8edJ78d69j2VrwctnTuEPvM8fTOKNCQ==</latexit> R <latexit sha1_base64="Bm6L7INbJx8XEYQzKxIxOiRrWCU=">AAAB6HicbZA9TwJBEIbn8AvxC7W02YgmVuTORjtJbCzByEcCF7K3zMHK3t5ld8+EEH6BjYXGYGnn37Hz37gHFIq+ySZP3ncmOzNBIrg2rvvl5FZW19Y38puFre2d3b3i/kFDx6liWGexiFUroBoFl1g33AhsJQppFAhsBsPrLG8+oNI8lndmlKAf0b7kIWfUWKt22y2W3LI7E/kL3gJKVx/TTG/VbvGz04tZGqE0TFCt256bGH9MleFM4KTQSTUmlA1pH9sWJY1Q++PZoBNyap0eCWNlnzRk5v7sGNNI61EU2MqImoFezjLzv6ydmvDSH3OZpAYlm38UpoKYmGRbkx5XyIwYWaBMcTsrYQOqKDP2NgV7BG955b/QOC97btmruaXKCcyVhyM4hjPw4AIqcANVqAMDhEd4hhfn3nlyXp3pvDTnLHoO4Zec92+uhpEz</latexit> <latexit sha1_base64="qoVWhpdOghMvJ6ZyOV73cMAR63I=">AAAB6HicbZC7SgNBFIbPxluMt6ilzWAUrMKujXYGbCwTMRdIljA7OZuMmZ1dZmaFsOQJbCwUSSd2vo6db+PkUmj0h4GP/z+HOecEieDauO6Xk1tZXVvfyG8WtrZ3dveK+wcNHaeKYZ3FIlatgGoUXGLdcCOwlSikUSCwGQyvp3nzAZXmsbwzowT9iPYlDzmjxlq1226x5Jbdmchf8BZQuvqYTN4AoNotfnZ6MUsjlIYJqnXbcxPjZ1QZzgSOC51UY0LZkPaxbVHSCLWfzQYdk1Pr9EgYK/ukITP3Z0dGI61HUWArI2oGejmbmv9l7dSEl37GZZIalGz+UZgKYmIy3Zr0uEJmxMgCZYrbWQkbUEWZsbcp2CN4yyv/hcZ52XPLXs0tVU5grjwcwTGcgQcXUIEbqEIdGCA8wjO8OPfOk/PqTOalOWfRcwi/5Lx/A1FJj24=</latexit> sha1_base64="af6fmYZ3mFiUWU/zWR5QDen5nZM=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLElsLMHIRwIXsrfMwcre3mV3z4Rc+AU2Fhpj60+y89+4wBUKvmSSl/dmMjMvSATXxnW/ncLG5tb2TnG3tLd/cHhUPj5p6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWByO/c7T6g0j+WDmSboR3QkecgZNVZq3g/KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGNn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o13Ur9Io+jCGdwDlfgQQ3qcAcNaAEDhGd4hTfn0Xlx3p2PZWvByWdO4Q+czx+iW4y0</latexit> 2 0 t 0.9 0 <latexit sha1_base64="yQfmJRBoTgLxijYwuTWw/rLs+7k=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgFe5sYmfAxjKi+YDkCHubuWTJ3t6xuyeEIz/BxkIJNhY2/h07/42bj0ITHww83pthZl6QCK6N6347ubX1jc2t/HZhZ3dv/6B4eNTQcaoY1lksYtUKqEbBJdYNNwJbiUIaBQKbwfBm6jcfUWkeywczStCPaF/ykDNqrHTvlr1useSW3RnIKvEWpHT9OZm8A0CtW/zq9GKWRigNE1Trtucmxs+oMpwJHBc6qcaEsiHtY9tSSSPUfjY7dUzOrdIjYaxsSUNm6u+JjEZaj6LAdkbUDPSyNxX/89qpCa/8jMskNSjZfFGYCmJiMv2b9LhCZsTIEsoUt7cSNqCKMmPTKdgQvOWXV0njsuzZxO7cUvUM5sjDCZzCBXhQgSrcQg3qwKAPT/ACr45wnp2J8zZvzTmLmWP4A+fjB/gVj78=</latexit> sha1_base64="MMH0J0T+HrKP6TiQR5Y2+qcfYYo=">AAAB6nicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLUlsLDEKksCF7C1zsGFv97K7Z0Iu/AQbC42x9RfZ+W9c4AoFXzLJy3szmZkXpYIb6/vfXmltfWNzq7xd2dnd2z+oHh61jco0wxZTQulORA0KLrFluRXYSTXSJBL4GI1vZv7jE2rDlXywkxTDhA4ljzmj1kn3fj3oV2t+3Z+DrJKgIDUo0OxXv3oDxbIEpWWCGtMN/NSGOdWWM4HTSi8zmFI2pkPsOippgibM56dOyblTBiRW2pW0ZK7+nshpYswkiVxnQu3ILHsz8T+vm9n4Osy5TDOLki0WxZkgVpHZ32TANTIrJo5Qprm7lbAR1ZRZl07FhRAsv7xK2pf1wCV259caZ0UcZTiBU7iAAK6gAbfQhBYwGMIzvMKbJ7wX7937WLSWvGLmGP7A+/wBSTaNBQ==</latexit> sha1_base64="e+R001ceKoRS5MEsQ4XGScHHhr4=">AAAB6nicbVA9SwNBEJ3zM8avqKXNYhSswp2NdgZsLCOaD0iOsLfZS5bs7R27c0I48hNsLJRoa+PfsfPfuJek0MQHA4/3ZpiZFyRSGHTdb2dldW19Y7OwVdze2d3bLx0cNkycasbrLJaxbgXUcCkUr6NAyVuJ5jQKJG8Gw5vcbz5ybUSsHnCUcD+ifSVCwSha6d6teN1S2a24U5Bl4s1J+fpzkuOt1i19dXoxSyOukElqTNtzE/QzqlEwycfFTmp4QtmQ9nnbUkUjbvxseuqYnFmlR8JY21JIpurviYxGxoyiwHZGFAdm0cvF/7x2iuGVnwmVpMgVmy0KU0kwJvnfpCc0ZyhHllCmhb2VsAHVlKFNp2hD8BZfXiaNi4pnE7tzy9VTmKEAx3AC5+DBJVThFmpQBwZ9eIIXeHWk8+xMnPdZ64oznzmCP3A+fgBVYZGE</latexit> <latexit sha1_base64="1cX1pw6RAXuT7gr52Ruqbpbqy0Y=">AAAB63icbVBNSwMxEJ2tX7V+VT16CVbBiyXbix4LXjxWsB/QLiWbZtvQJLskWaEs/QtePCji1T/kzX9jtt2Dtj4YeLw3w8y8MBHcWIy/vdLG5tb2Tnm3srd/cHhUPT7pmDjVlLVpLGLdC4lhgivWttwK1ks0IzIUrBtO73K/+8S04bF6tLOEBZKMFY84JTaXrn2Mh9UaruMF0DrxC1KDAq1h9WswimkqmbJUEGP6Pk5skBFtORVsXhmkhiWETsmY9R1VRDITZItb5+jSKSMUxdqVsmih/p7IiDRmJkPXKYmdmFUvF//z+qmNboOMqyS1TNHloigVyMYofxyNuGbUipkjhGrubkV0QjSh1sVTcSH4qy+vk06j7uO6/9CoNS+KOMpwBudwBT7cQBPuoQVtoDCBZ3iFN096L96797FsLXnFzCn8gff5A7aFjUA=</latexit> -100 -50 <latexit sha1_base64="vQWLT/ViCnpqQmH+dGECu1vyhWY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LFbBiyUpiB4LXjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfj25nffkKleSwfzSRBP6JDyUPOqLHSw+WV2y9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5ufOiXnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhDd+xmWSGpRssShMBTExmf1NBlwhM2JiCWWK21sJG1FFmbHplGwI3vLLq6RVq3pu1buvVepneRxFOIFTuAAPrqEOd9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AFOY40K</latexit> 0 x <latexit <latexit sha1_base64="kRDnxcx7hyKVH6+joAsSPWBDiHM=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsaLElsLCGRjwQuZG+Zg5W9vcvungm58AtsLDTG1p9k579xgSsUfMkkL+/NZGZekAiujet+O4Wt7Z3dveJ+6eDw6PikfHrW0XGqGLZZLGLVC6hGwSW2DTcCe4lCGgUCu8H0buF3n1BpHssHM0vQj+hY8pAzaqzUcoflilt1lyCbxMtJBXI0h+WvwShmaYTSMEG17ntuYvyMKsOZwHlpkGpMKJvSMfYtlTRC7WfLQ+fk2iojEsbKljRkqf6eyGik9SwKbGdEzUSvewvxP6+fmvDWz7hMUoOSrRaFqSAmJouvyYgrZEbMLKFMcXsrYROqKDM2m5INwVt/eZN0alXPrXqtWqVxlcdRhAu4hBvwoA4NuIcmtIEBwjO8wpvz6Lw4787HqrXg5DPn8AfO5w9vc4yU</latexit> <latexit sha1_base64="79wqBERO1FRfA1Dy5mKGWRmSyn4=">AAAB6HicbZA9TwJBEIbn/ET8Qi1tNqKJFbmz0U4SG0tI5CMBQvaWOVjZ27vs7hnJhV9gY6ExWNr5d+z8N+4BhYJvssmT953JzowfC66N6347K6tr6xubua389s7u3n7h4LCuo0QxrLFIRKrpU42CS6wZbgQ2Y4U09AU2/OFNljceUGkeyTszirET0r7kAWfUWKv62C0U3ZI7FVkGbw7F689JpvdKt/DV7kUsCVEaJqjWLc+NTSelynAmcJxvJxpjyoa0jy2LkoaoO+l00DE5s06PBJGyTxoydX93pDTUehT6tjKkZqAXs8z8L2slJrjqpFzGiUHJZh8FiSAmItnWpMcVMiNGFihT3M5K2IAqyoy9Td4ewVtceRnqFyXPLXlVt1g+hZlycAwncA4eXEIZbqECNWCA8AQv8OrcO8/OmzOZla44854j+CPn4wfoHpFZ</latexit> <latexit sha1_base64="gBTNO7IH+P1mduEskbFWJ6WrbPE=">AAAB6HicbZC7SgNBFIbPeo3xFrW0GYyCVdi10c6AjWUC5gLJEmYnZ5Mxs7PLzKwYljyBjYUi6cTO17HzbZxcCk38YeDj/89hzjlBIrg2rvvtrKyurW9s5rby2zu7e/uFg8O6jlPFsMZiEatmQDUKLrFmuBHYTBTSKBDYCAY3k7zxgErzWN6ZYYJ+RHuSh5xRY63qY6dQdEvuVGQZvDkUrz/H43cAqHQKX+1uzNIIpWGCat3y3MT4GVWGM4GjfDvVmFA2oD1sWZQ0Qu1n00FH5Mw6XRLGyj5pyNT93ZHRSOthFNjKiJq+Xswm5n9ZKzXhlZ9xmaQGJZt9FKaCmJhMtiZdrpAZMbRAmeJ2VsL6VFFm7G3y9gje4srLUL8oeW7Jq7rF8inMlINjOIFz8OASynALFagBA4QneIFX5955dt6c8ax0xZn3HMEfOR8/iuGPlA==</latexit> sha1_base64="lEcjT01rh+kON+avtnUjSMl7nNU=">AAAB6HicbVA9TwJBEJ3DL8Qv1NJmI5pYkTsbLElsLCGRjwQuZG+Zg5W9vcvunpFc+AU2Fhpj60+y89+4wBUKvmSSl/dmMjMvSATXxnW/ncLG5tb2TnG3tLd/cHhUPj5p6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWByO/c7j6g0j+W9mSboR3QkecgZNVZqPg3KFbfqLkDWiZeTCuRoDMpf/WHM0gilYYJq3fPcxPgZVYYzgbNSP9WYUDahI+xZKmmE2s8Wh87IpVWGJIyVLWnIQv09kdFI62kU2M6ImrFe9ebif14vNeGNn3GZpAYlWy4KU0FMTOZfkyFXyIyYWkKZ4vZWwsZUUWZsNiUbgrf68jppX1c9t+o13Ur9Io+jCGdwDlfgQQ3qcAcNaAEDhGd4hTfnwXlx3p2PZWvByWdO4Q+czx/b84za</latexit> 50 <latexit sha1_base64="BWuP9VBNfr/tVrmk+3RJQHf+H1w=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0kKRY8FLx6r2A9oQ9lsN+3SzSbsToQS+g+8eFDEq//Im//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1iNOE+xEdKREKRtFKD3V3UK64VXcBsk68nFQgR3NQ/uoPY5ZGXCGT1Jie5yboZ1SjYJLPSv3U8ISyCR3xnqWKRtz42eLSGbm0ypCEsbalkCzU3xMZjYyZRoHtjCiOzao3F//zeimGN34mVJIiV2y5KEwlwZjM3yZDoTlDObWEMi3srYSNqaYMbTglG4K3+vI6adeqnlv17muVxkUeRxHO4ByuwINraMAdNKEFDEJ4hld4cybOi/PufCxbC04+cwp/4Hz+AOTgjNM=</latexit> 0.1 100 <latexit sha1_base64="XiFhIYx2rFbNJAJ/xJ6g+YVEH2k=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFqNgFfbSaBmwsYxoPiA5wt5mkizZ2zt294Rw5CfYWChi6y+y89+4Sa7QxAcDj/dmmJkXJlIYS+m3V9jY3NreKe6W9vYPDo/KxyctE6eaY5PHMtadkBmUQmHTCiuxk2hkUSixHU5u5377CbURsXq00wSDiI2UGArOrJMefEr75Qqt0gXIOvFzUoEcjX75qzeIeRqhslwyY7o+TWyQMW0Flzgr9VKDCeMTNsKuo4pFaIJsceqMXDplQIaxdqUsWai/JzIWGTONQtcZMTs2q95c/M/rpnZ4E2RCJalFxZeLhqkkNibzv8lAaORWTh1hXAt3K+Fjphm3Lp2SC8FffXmdtGpVn1b9+1qlfpHHUYQzOIcr8OEa6nAHDWgChxE8wyu8edJ78d69j2VrwctnTuEPvM8fTOKNCQ==</latexit> FIG 8 Trave ng wave connect ng a per od c orb t to a node Sur ace p ots show ng the evo ut on o (a) the synapt c conductance g and wave connecting or a ront wh ch connects an osc plots state to the evolution of S mu synaptic FIG. 8. Travelling wave connecting a periodic orbit to a node: Surface atory showing the evolution of (a) the synaptic FIG. 8. Travelling (b) the synchronyaRperiodic orbit to a node: Surface plots showing a fixed po nt state(a) the at ons or the system defined(b) the synchrony=R for vayn = 10 ∆ =connects an τoscillatory state to a fixed point state. Simulations for by (36) synchrony R for a front which connects 5 oscillatory state to a fixed point state. Simulations for w th η0 −5 conductance g and front which 0 5 κ = an = 0 2 conductance g and (b) the the system defined by (36), with ⌘0 = 5, vsyn = 10, = 0.5,  = 5, ⌧ = 0.2 system defined by (36), with ⌘0 = 5, vsyn = 10, = 0.5,  = 5, ⌧ = 0.2 the The der vat ves are computed as fo ows chinery to include a second spatial extension is worthy of chinery to include a second spatial extension is worthy of ∂ ∂a ∂ further exploration. ∂b further exploration. = ∂a ∂t ∂b ∂t 1 2 = b − (a + 1)∆ − b(η0 + vsyn g1 + vsyn g2 ) − a(g1 + g2 ) VI. VI. ACKNOWLEDGEMENTS ACKNOWLEDGEMENTS 2 2 p2 = (τ1 + τ2 + 4τ1 τ2 ) A2 + B 2 + 1 The Jacobian of the system can be written )2 follows: The Jacobian of the system 1canτbe written2as follows: −4(τ + 2 )A /(τ1 τ as ✓ ◆ ✓ 2 2 p3 = 2 (τ1 τ2 + τ1 τ2 ) A2 + 112 + 12◆ + τ2 − 4A) J B J (τ J11 J12 1 , J (k) = 2 J (k) = J (k) J , J 21 τ 2 J22 −(τ121 (k)2 )A22/(τ1 τ2 )2 + −2 2 2 p4 = −2 ∂ ∂b ∂ ∂a where, where, τ1 + τ2 + A + B + 4 (1 − (τ1 + τ2 )A) /(τ1 τ2 ) =− −1 −1 ∂b ∂t supported by the Engineering and Physical ∂a ∂t p5 = 2 τ1 + τ2 − 0 A 0 @ @a @ @a1 1 DA was DA was supported by the Engineering and Physical @ @a @ @a 1 2 = (a − Research Council under grant EP/P510993/1. Sciences 1) − (a + 1)(η0 + vsyn g1 +grant 2EP/P510993/1.g2 ) q0 = κ C /(τ τj )2 Research Council under vsyn g ) + b(∆ + g1 + Sciences B @a @t @b @t C B @a @t @b @t C B C SC was supported by the European Commission through supported by the European Commission through C SC was∂a J11 = B ∂ 2 2 @ A q1 = κ f (a b)J112τj κ C /(τ τj )2 A ,, + =@ = b(a + 1)vsyn − Training 1) the FP7 Marie Curie Initial Training Network 289146, @ @b @ @b the FP7 Marie Curie Initial (a − b −Network 289146, @ @b @ @b ∂g ∂t 2 q2 = 2κ f (a b) + τj κ C@a @t τ@b @t NETT: Neural Engineering Transformative Technologies. NETT: Neural Engineering Transformative Technologies. @t @b @a/(τ j ) @t ∂ ∂b 1 1 2 = ((a + 1)2 − b2 )vsyn − ab q3 = κ f (a b)/τ ∂g ∂t 2 1 0 1 0 @ @a @ @a Appendix Appendix A: Jacobian 2 (a + 1)2 − b2 A: Jacobian ∂f 0 @ @a 0 @ @a C 0 B =− B0 @g1 @t @g2 @t C @g2 @t C B C B @g1 @t ∂a π ((a + 1)2 − b2 )2 B C C, J12 = B =B To calculate the Jacobian we first write the system calculate the Jacobian we first write the system where i j ∈ {1 2} B and @ @b J 12 B To C, @ @b C 4 b(a + 1) ∂f C @ @b 0 @ @b C B0 (19) -- (20) in its full six dimensional form, =− @ A (19) -- (20) in its full2six dimensional form, @ 0 @g1 @t 0 @g2 @t A ∂b π ((a + 1) − b2 )2 @g1 @t @g2 @t @a @a = b(a 1) b(a + 1)(⌘ + v1 g + v2 g ) 1 2 0 syn 1 syn 2 ∂ ∂a ∂ ∂b @t = b(a 1) b(a + 1)(⌘0 + vsyn g1 + vsyn g2 ) @t A≡ Append x B 1 Tur ng coeffic ents 0 1 0 = ∂b ∂t ∂a ∂t 1 1 (a2 b2 1)( + g + g ), 2 2 @f @f 1 1 (a + 1) + 1) 1 1 w1 @f 1 1 w1 @f 1 + g2 ), 2 (a b 1)( + g1 b ⌧1 b b @a b @b C 2(a B⌧ ∂ B⌧1 1 w1∂@a ∂b⌧1 1 w1 @b C ∂a 1 2 B C The coeffic ents of the character st c equat on (23) n B≡ B =− C @b 1 B C @b = 1 ((a 1)2 b2 ) b 2 2 C ∂b B ∂t ∂a ∂t ab( + g1 + g2 ) + g1 + g2 ) B C §III are ca cu ated 1) fo bows b as ) ab( B C 0 0 0 0 B C @t = 2 ((a 2 B C, @t J21 (k) = B B 2b C , C 21 (k) = B J 12 B C = −Af (a b) 1 vsyn @f 1 ((a + 2 2/(τ1b22)(⌘ + v1 g + v2 g ), 2 2 )2 1 2 B ⌧ 1+ π @f − (1  w 2@f C @f C 1 p0 = A + B1) τ )(⌘ 0 + v syng 1 + v syng 2), + ((a + 1) B⌧ 2  2w 2 + b2 C a) b2 2 w2 b ⌧2 1 2 w 2 + C b + 2 b B 2 b @a ⌧2 2 @b C 0 syn 1 syn 2 B C B 2 @a @b A 2 2 2 @ A @ @g1 p = 2 (τ + τ ) A + B − A /(τ1 τ2 ) 1 @g1 =1 1 ( g 1 K 2), 1+ 1 0 0 = ⌧ ( g1 + K1 ), 0 0 @t 1 @t ⌧1 @K1 1 @K1 = 1 ( K +  w ⌦ f (a + ib)), 1 1 m 0 1 1 0 ⌧ 1 @t = ⌧1 ( K1 + 1 wm ⌦ f (a + ib)), ⌧1 0 0 0 1 @t 1 ⌧1 0 0 0 @g2 1 B C @g2 = 1 ( g + K ), B ⌧ 1 B 11 C ⌧1 1 0 0 C 2 2 ⌧1 1 0 Cke 1 @t = ⌧ 2 ( g2 + K2 ), D A 2 S Amar = B ⌧1 Homogeneous nets 0o neuron-C . e ements B @t M ⌧2 exander C Trengove and C van Leeuwen DonJ22 = B J 22 Cybernet cs 17 211 -- 2201 1975 C . B C ders B o og ca B 0 0 ⌧2 0 C @ @K2 s1dead Cort ca trave ng waves and the m ts o 0 ⌧2 1 0 A @ 0 cs o pattern ormat A n atera @K2 = 1 chronometrywn cogn (ave neurosc ence Cogn t ve ( menta ( K2 + 2 m ⌦ f t + ib)). 3 S Amar Dynam on 1 @t = ⌧2 ng K2 + 2 wm ⌦ f (a + ib)). ⌧ 0 0 ⌧2 o ⌧ 1 @t Process2 16(4) 365 -- 375 2015 nh b t on type neura fie ds ⌧B 1 og ca2 1 0 0 ⌧ 2 Cybernet cs 27 2 14 (2):77 -- 87, 1977. [4] D. Avitabile. Computation of planar patterns and their stability. PhD thesis, University of Surrey, 2008. [5] D. Avitabile. Numerical computation of coherent structures in spatially-extended neural networks. Tutorial at Second International Conference on Mathematical Neuroscience, Antibes Juan-les-Pins, 2016. URL https: //www.maths.nottingham.ac.uk/personal/pmzda/. [6] D. Avitabile and H. Schmidt. Snakes and ladders in an inhomogeneous neural field model. Physica D, 294:24 -- 36, 2015. [7] R. Ben-Yishai, R. L. Bar-Or, and H. Sompolinsky. Theory of orientation tuning in visual cortex. Proceedings of the National Academy of Sciences of the United States of America, 92(9):3844 -- 8, 1995. u [8] H. Berger. Uber das Elektroenzenkephalogram des Menschen. Archiv fur Psychiatrie und Nervenkrankheiten, 87: u 527 -- 70, 1929. [9] R. L. Beurle. 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, 1956. [10] P. C. Bressloff. Spatiotemporal dynamics of continuum neural fields. Journal of Physics A: Mathematical and Theoretical, 45(3):033001, 2012. [11] P. C. Bressloff and S. Coombes. Neural 'bubble' dynamics revisited. Cognitive Computation, 5:281 -- 294, 2013. [12] P. C. Bressloff and M. A. Webber. Neural field model of binocular rivalry waves. Journal of Computational Neuroscience, 32(2):233 -- 52, 2012. [13] P. C. Bressloff, J. D. Cowan, M. Golubitsky, P. J. Thomas, and M. C. Wiener. Geometric visual hallucinations, Euclidean symmetry and the functional architecture of striate cortex. Philosophical Transactions of the Royal Society of London. Series B,: Biological Sciences, 356(1407):299 -- 330, 2001. [14] A. Byrne, M. J. Brookes, and S. Coombes. A mean field model for movement induced changes in the beta rhythm. Journal of Computational Neuroscience, 43:143 -- 158, 2017. [15] A. R. Champneys and B. Sandstede. Numerical computation of coherent structures. In B. Krauskopf, H. M. Osinga, and J. Gal´n-Vioque, editors, Numerical Contina uation Methods for Dynamical Systems: Path following and boundary value problems, pages 331 -- 358. Springer, 2007. [16] C. C. Chow and S. Coombes. Existence and wandering of bumps in a spiking neural network model. SIAM Journal on Applied Dynamical Systems, 5(4):552 -- 574, 2006. [17] S. Coombes. Waves, bumps, and patterns in neural field theories. Biological Cybernetics, 93:91 -- 108, 2005. [18] S. Coombes. Large-scale neural dynamics: Simple and complex. NeuroImage, 52(3):731 -- 9, 2010. [19] S. Coombes and A. Byrne. Next generation neural mass models. In A. Torcini and F. Corinto, editors, Lecture Notes Nonlinear Dynamics in Computational Neuroscience: from Physics and Biology to ICT, PoliTO, pages 1 -- 16. Springer, 2019. [20] S. Coombes, G. Lord, and M. Owen. Waves and bumps in neuronal networks with axo-dendritic synaptic interactions. Physica D, 178(3):219 -- 241, 2003. [21] S. Coombes, P. P. Beim Graben, R. Potthast, and J. Wright. Neural fields : Theory and Applications. Springer, 2014. [22] F. L. da Silva and A. V. Rotterdam. Biophysical aspects of EEG and magnetoencephalogram generation. In E. Niedermeyer and F. L. da Silva, editors, Electroencephalography: Basic Principles, Clinical Applications and Related Fields, pages 107 -- 126. Lippincott Williams & Wilkins, 2005. [23] E. J. Doedel. Auto: A program for the automatic bifurcation analysis of autonomous systems. Congr. Numer, 30:265 -- 284, 1981. [24] B. Ermentrout. Neural networks as spatio-temporal pattern-forming systems. Reports on Progress in Physics, 61:353 -- 430, 1998. [25] G. B. Ermentrout. Simulating, Analyzing, and Animating Dynamical Systems: A Guide to XPPAUT for Researchers and Students. SIAM Books, 2002. [26] G. B. Ermentrout and J. D. Cowan. A mathematical theory of visual hallucination patterns. Biological Cybernetics, 34:137 -- 150, 1979. [27] G. B. Ermentrout and N. Kopell. Parabolic bursting in an excitable system coupled with a slow oscillation. SIAM Journal on Applied Mathematics, 46:233 -- 253, 1986. [28] F. Gabbiani and S. J. S. J. Cox. Mathematics for neuroscientists. Elsevier Academic Press, 2010. [29] D. Golomb and Y. Amitai. Propagating neuronal discharges in neocortical slices: Computational and experimental study. Journal of Neurophysiology, 78:1199 -- 1211, 1997. [30] C. Laing and C. C. Chow. Stationary bumps in networks of spiking neurons. Neural Computation, 13:1473 -- 1494, 2001. [31] C. R. Laing. Numerical bifurcation theory for highdimensional neural models. The Journal of Mathematical Neuroscience, 4(1):13, 2014. ISSN 2190-8567. [32] C. R. Laing. Exact neural fields incorporating gap junctions. SIAM Journal on Applied Dynamical Systems, 14: 1899 -- 1929, 2015. [33] C. R. Laing and C. C. Chow. A spiking neuron model for binocular rivalry. Journal of Computational Neuroscience, 12:39 -- 53, 2002. [34] T. B. Luke, E. Barreto, and P. So. Complete classification of the macroscopic behaviour of a heterogeneous network of theta neurons. Neural Computation, 25:3207 -- 3234, 2013. [35] M. Massimini. The sleep slow oscillation as a traveling wave. Journal of Neuroscience, 24(31):6862 -- 6870, 2004. [36] E. Montbri´, D. Paz´, and A. Roxin. Macroscopic deo o scription for networks of spiking neurons. Physical Review X, 5:021028, 2015. [37] P. L. Nunez and R. Srinivasan. Electric Fields of the Brain: The Neurophysics of EEG. Oxford University Press, 2nd edition, 2005. [38] P. L. Nunez and R. Srinivasan. A theoretical basis for standing and traveling brain waves measured with human EEG with implications for an integrated consciousness. Clinical Neurophysiology, 117(11):2424 -- 35, 2006. [39] E. Ott and T. M. Antonsen. Low dimensional behavior of large systems of globally coupled oscillators. Chaos, 18:037113, 2008. [40] G. Pfurtscheller and F. H. L. da Silva. Event-related EEG/MEG synchronization and desynchronization: Basic principles. Clinical Neurophysiology, 110:1842 -- 1857, 1999. 15 [41] J. Rankin, D. Avitabile, J. Baladron, G. Faye, and D. J. B. Lloyd. Continuation of localized coherent structures in nonlocal neural field equations. SIAM Journal on Scientific Computing, 36(1):B70 -- B93, Jan. 2014. [42] H. R. Wilson and J. D. Cowan. Excitatory and inhibitory interactions in localized populations of model neurons. Biophysical Journal, 12:1 -- 24, 1972. [43] H. R. Wilson and J. D. Cowan. A mathematical theory of the functional dynamics of cortical and thalamic nervous tissue. Biological Cybernetics, 13(2):55 -- 80, 1973. [44] K. Wimmer, D. Q. Nykamp, C. Constantinidis, and A. Compte. Bump attractor dynamics in prefrontal cortex explains behavioral precision in spatial working memory. Nature Neuroscience, 17(3):431 -- 439, 2014. [45] K. Zhang. Representation of spatial orientation by the intrinsic dynamics of the head-direction cell ensemble: A theory. Journal of Neuroscience, 16(6), 1996.
1609.08320
2
1609
2017-03-05T20:49:36
Synchronization of oscillators through time-shifted common inputs
[ "q-bio.NC", "nlin.AO" ]
Shared upstream dynamical processes are frequently the source of common inputs in various physical and biological systems. However, due to finite signal transmission speeds and differences in the distance to the source, time shifts between otherwise common inputs are unavoidable. Since common inputs can be a source of correlation between the elements of multi-unit dynamical systems, regardless of whether these elements are directly connected with one another or not, it is of importance to understand their impact on synchronization. As a canonical model that is representative for a variety of different dynamical systems, we study limit-cycle oscillators that are driven by stochastic time-shifted common inputs. We show that if the oscillators are coupled, time shifts in stochastic common inputs do not simply shift the distribution of the phase differences, but rather the distribution actually changes as a result. The best synchronization is therefore achieved at a precise intermediate value of the time shift, which is due to a resonance-like effect with the most probable phase difference that is determined by the deterministic dynamics.
q-bio.NC
q-bio
Synchronization of oscillators through time-shifted common inputs Ehsan Bolhasani,1, 2 Yousef Azizi,1 Alireza Valizadeh,1, 2 and Matjaz Perc3, 4 1Department of physics, Institute for Advanced Studies in Basic Sciences, Zanjan, Iran 2School of Cognitive Science, Institute for Research in Fundamental Sciences (IPM), P. O .Box 1954851167, Tehran, Iran 3Faculty of Natural Sciences and Mathematics, University of Maribor, Koroska cesta 160, SI-2000 Maribor, Slovenia 4CAMTP -- Center for Applied Mathematics and Theoretical Physics, University of Maribor, Krekova 2, SI-2000 Maribor, Slovenia Shared upstream dynamical processes are frequently the source of common inputs in various physical and biological systems. However, due to finite signal transmission speeds and differences in the distance to the source, time shifts between otherwise common inputs are unavoidable. Since common inputs can be a source of correlation between the elements of multi-unit dynamical sys- tems, regardless of whether these elements are directly connected with one another or not, it is of importance to understand their impact on synchronization. As a canonical model that is represen- tative for a variety of different dynamical systems, we study limit-cycle oscillators that are driven by stochastic time-shifted common inputs. We show that if the oscillators are coupled, time shifts in stochastic common inputs do not simply shift the distribution of the phase differences, but rather the distribution actually changes as a result. The best synchronization is therefore achieved at a precise intermediate value of the time shift, which is due to a resonance-like effect with the most probable phase difference that is determined by the deterministic dynamics. PACS numbers: 05.45.Xt, 87.19.lm, 89.75.Fb, 89.75.Kd INTRODUCTION Synchronization has been observed in a wide variety of physical and biological systems, and it is indeed a universal phenomenon in nonlinear sciences [1, 2]. The functionality of systems often depends on the degree of synchronization, with examples including power grid sys- tems [3 -- 5], Josephson junction arrays [6, 7], cardiac pace- maker cells [8, 9], and gamma rhythms in nervous sys- tems [10 -- 12], to name but a few examples. The sub- ject is an evergreen field of research in statistical physics, attracting recurrent attention in relation to chaotic sys- tems, neuronal dynamics, and network science [13 -- 23]. The coordination of activity of nonlinear oscillators is conventionally assigned to the presence of attractive in- teractions between them [24]. However, synchronization among coupled oscillators in the presence of delayed in- teractions is more challenging to understand [25 -- 29]. De- layed interactions between coupled oscillators are in gen- eral attributed to finite information transmission speeds. As has been shown before, such interactions can change attractive coupling to a repulsive coupling and vice versa, and thus crucially affect the emergence of synchroniza- tion in networks of coupled oscillators [30 -- 34]. But regardless of presence of direct interactions be- tween the oscillators, synchronization is still attainable through a common input. The presence of such in- puts evokes statistical correlations between the oscilla- tors, which can ultimately lead to synchronization [35 -- 41]. In studies of the synchronization through com- mon inputs, signal transmission delays can be ignored if the common source affects the target oscillators through equidistant routes, but of course not if the distances to the downstream oscillators differ. In the latter case, the input cross-correlation function for every pair of oscilla- tors peaks at a different time that depends on the specific time shift. As yet, this realistic setup constitutes an im- portant unsolved problem in the realm of synchronization among nonlinear oscillators. In this paper, we therefore study the emergence of phase-synchronization among limit-cycle oscillators that are driven by time-shifted common inputs. In particular, we determine the effects on the output cross-correlation function of two oscillators when they receive time-shifted common inputs and the time shift is smaller than the pe- riod of oscillation of the oscillators. If the oscillators are uncoupled, we show that the effect of a time-shifted com- mon input is simply a shift in the output cross-correlation function. For coupled oscillators, however, time shifts in the common inputs non-trivially affect the output cor- relation so that in addition to the shift, the correla- tion function and the distribution function for the phase difference of the oscillators are dependent on the input time-shift. We show analytically and confirm numerically that the degree of synchrony depends on the time-shift of common inputs, such that optimal synchronization is obtained only when the common inputs are differentially shifted by a non-zero time lag. MATHEMATICAL FRAMEWORK We consider two bidirectionally coupled oscillators with state vector Xi, (i = 1, 2) which evolves according 2 FIG. 1: (a) Phase difference distribution function of two identical uncoupled oscillators, receiving a fully correlated Gaussian white noise with different time shifts shown above the plots. (b) Same results for non-identical oscillators with the frequency mismatch ∆ω = 0.5. Here D = 0.5, and τs is −2,−1, 0, 1, 2 from left to right. to X1(t) = F X1(t), I1 (cid:16) (cid:16) (cid:16) (cid:16) (cid:17) (cid:17) (cid:17) (cid:17) + εg12G12 X1(t), X2(t) X2(t) = F X2(t), I2 + εg21G21 X2(t), X1(t) √ √ + + Dεξ(t), Dεξ(t − τs), (1) where vector function F (Xi, Ii) describes the inherent dynamic of the isolated oscillators (gij = 0). We as- sume that the isolated oscillators have a stable limit cycle XLC(t) = XLC(t + T ) with the period T which is con- trolled by the scalar parameter Ii. Gij determines the in- teraction function whose strength is defined by g12. The last terms in the equations describe common stochastic input to the two oscillators where ξ(t) is Gaussian white noise with zero mean and unit variance and D determines the noise intensity. τs is the key parameter of the present study which is assumed to be smaller than the period of the oscillation of the two oscillators (τs < T ) and deter- mines the time shift between the two stochastic inputs, i.e., the second neuron receives exactly the same stochas- tic input to the first neuron, but with a time shift equal to τs. Both the interaction term and stochastic inputs are scaled with the small parameter ε (cid:28) 1. Furthermore the oscillators are assumed to be in general nonidenti- cal due to the small difference in their control parameter I1 − I2 = ε∆I. Applying the standard phase reduction method in the regime of weak perturbation [24, 36] leads to the follow- ing Ito stochastic differential equations, for the time evo- lution of the phase variable θ(X) in the vicinity of the unperturbed limit cycle XLC √ √ θ1(t) = ω1 + εg12Z (θ1) · G (θ1(t), θ2(t)) + DεZ (θ1) · ξ (t) θ2(t) = ω2 + εg21Z (θ2) · G (θ2(t), θ1(t)) DεZ (θ2) · ξ (t − τs) + (2) where Z(θ) = ∇X θXLC(θ) is the phase sensitivity [42]. Since we assumed small inhomogeneity in the bifurcation parameter ∆I ∼ O(ε), difference in natural frequencies will be also small ω1 − ω2 = ε∆ω. We take ω1 = 1 and ω2 = 1 − ε∆ω and define time- shifted phase difference as θ(t) = θ1(t)− θ2(t + τs). From Eqs. 2 we obtain the evolution of θ(t) to the order of ε (cid:16) here f θ1(t), θ2(t(cid:48)) = D Z (cid:17) √ (cid:16) (3) (cid:16) (cid:17)(cid:105) (cid:17) · G (θi(t), θj(t)). We as- θ2(t(cid:48)) θ1(t) and θi(t) Hij(θi(t), θj(t)) = gijZ sume θi(t) = t + ϕi(t) where the first term t captures the intrinsic dynamics of isolated oscillators and the second term is slow varying deviation from natural oscillations. Averaging the equation over one period [43], we have dϕ(t) dt = ε∆ω + ε √ + (cid:16) ¯H12(ϕ) − ¯H21(−ϕ) (cid:16) (cid:17) (cid:17) ε ¯f ϕ − τs η(t) (4) (cid:16) (cid:104) θ(t) = ε∆ω + ε H12(θ1(t), θ2(t)) −H21(θ2(t + τs), θ1(t + τs)) √ (cid:16) + εf θ1(t), θ2(t + τs) (cid:17) (cid:17) · ξ(t) (cid:16) (cid:17) − Z 3 FIG. 2: (a) Phase difference distribution function of two coupled oscillators in the locked mode, receiving a fully correlated Gaussian white noise with different time shifts shown above each plot. (b) Same results for oscillators in the running mode. The results shown by gray bar plots are calculated by numeric integration of Eq. 3 and the solid lines show the analytical result given by Eq. 6. The vertical dashed lines in (a) show the fixed point of Eq. 3. Other parameters are D = 0.5 and ∆ω = 0.2 for the locked case, and ∆ω = −0.1 for the running case. (cid:105)2 (cid:82) 2π (cid:114) 1 (cid:82) 2π 2π (cid:104) where ϕ = ϕ1 − ϕ2 and the averaged func- tions ¯Hij and ¯f are defined as ¯Hij(θi(t), θj(t)) = 1 = 2π 0 dtHij(θi(t), θj(t)) ¯f (ϕ − τs) and 0 dt f (θ1(t), θ2(t + τs)) (see Appendix). Note that in the above equation the lag in the common inputs τs acts like delay in the connections with a change of variable ϕ → ϕ + τs. We finally derive the Fokker-Planck equation for the distribution of the phase differences of the two oscillators, described by Eq. 4 (ϕ, t) = − ε ∂ρ ∂t (cid:17) (5) + ε [Γ (ϕ) ρ (ϕ, t)] ∂ ∂ϕ ∂2 ∂ϕ2 (cid:16) ¯f (ϕ − τs)ρ (φ, t) (cid:16) ¯H12(ϕ) − ¯H21(−ϕ) (cid:17) where Γ (ϕ) = ∆ω + and ρ(ϕ, t) depicts the distribution of ϕ. The stationary distribu- tion of the phase differences can be calculated by letting ∂ρ/∂t = 0 (cid:34) (cid:82) 2π e−M (2π) − 1 0 e−M (x)dx (cid:82) ϕ 0 e−M (x)dx + 1 (cid:35) (6) eM (ϕ) N ¯f (ϕ − τs) ρ0(ϕ) = where M (ϕ) = (cid:82) ϕ 0 Γ(x) ¯f (x−τs) dx and N is a normalization factor. In the following the analytical results are ob- tained by calculation of the stationary distribution func- tion from the above equation (see Appendix for the de- tails of derivations). RESULTS To check the validity of the Eq. 4 and the correspond- ing solution of Fokker-Planck equation Eq. 6, we take Z(θ) = 1 − cos(θ) which is the canonical form of the phase sensitivity for the type-I oscillators near SNIC bi- furcation [44]. Furthermore, we assume the oscillators are pulse-coupled, i.e., Gij = Σnδ(cid:0)t − tn (cid:1) in which δ is Dirac's delta function and tn j is the instant of θj = 2πn. Pulse coupling approximation for interaction between os- cillators is valid in the systems where the interaction term activates over a time which is small compared to period of the oscillation [9, 45 -- 47]. To assess the degree of syn- chrony regardless of the value of the phase lag we use the synchronization index γ2 = (cid:104)cos ϕ(cid:105)2 +(cid:104)sin ϕ(cid:105)2, where the brackets denote the averaging over time [48]. j First, we considered two uncoupled oscillators receiv- ing a common noise with a time shift τs. For the identical oscillators ∆ω = 0, it has been shown that the oscillators synchronize when receiving common noises [36]. Our re- sults show that time shift in the inputs results in the same time shift in the synchronized output without changing the synchrony index (Fig. 1a and Fig. 3). When the oscil- lators are not identical ∆ω (cid:54)= 0, phase difference distri- bution function spreads and synchrony index decreases (Fig. 1b and Fig. 3). In the presence of heterogeneity (mismatch of the firing rates) the most probable phase lag φ∗ is not zero when the inputs have zero time-shift [49]. Interestingly this effect of heterogeneity can be com- pensated by a non-zero time shift in the inputs, i.e., the most probable phase difference of the oscillators could be around zero despite the heterogeneity by a suitable choice 4 with the maximum synchrony index is attained in a cer- tain value of time lag of the inputs (τs (cid:39) 1, see Fig. 2b and Fig. 3). Analytical results from Eq. 6 are again val- idated by the direct numeric solution of Eq. 3. SUMMARY Summarizing, we have studied the emergence of syn- chronization between two limit-cycle oscillators subject to a common but time shifted stochastic input. In addi- tion to showing that for uncoupled oscillators the effect of time shifts is a trivial corresponding shift in the distribu- tion of the phase differences between the two oscillators, we have derived fundamental conditions for optimal syn- chronization when the oscillators are coupled. Namely, we have shown analytically that the time shift between the inputs changes the distribution of relative phases, and with it also the degree of synchronization in the system. Specifically, with time shifts that are in accord with the most probable phase difference between the oscillators due to their deterministic dynamics, a resonance-like ef- fect can be observed that leads to optimally phase locked oscillators. Due to the generality of the considered setup, we expect our results to significantly improve our under- standing of phase synchronization in networks of nonlin- ear elements, especially within the realm of multilayer networks, where common inputs across different layers might be particularly likely [50, 51]. Appendix We define θ(t) = θ1(t) − θ2(t + τs) and subtract the equations Eq. 2 after shifting the time in second equation by τs to obtain the evolution equation for θ(t) (Eq. 3). Then we define θi(t) = t + ϕi(t) (A1) where the first term in R.H.S. captures the intrinsic dy- namics of isolated oscillators with ωi (cid:39) 1, and the second term is slow varying deviation from natural oscillations. Using the method of averaging [43], we substitute Eq. A1 into Eq. 3 and average Eq. 3 over one period. For the second term in the R.H.S. of Eq. 3 we define (cid:90) 2π (cid:90) 2π (cid:90) 2π 0 ¯H(ϕ) = = = 1 2π 1 2π 1 2π H(θ1(t), θ2(t))dt H(t + ϕ1, t + ϕ2)dt H(¯t, ¯t − ϕ)d¯t (A2) 0 0 and for the third term in R.H.S. of Eq. 3 we should take into account the integral of the correlation term as follows FIG. 3: (a) Synchronization index versus input time-shift for the case of two identical uncoupled phase oscillators (black dashed line), non-identical uncoupled phase oscillators (solid black line), coupled phase oscillators in phase-locked mode (solid gray line), and non-identical coupled phase oscillators in the running mode (gray dashed line). (b) Most probable phase lag which shows the location of the peak of the phase lag distributions in Figs. 1 and 2 is plotted versus input time shift. of the time lag of inputs (see Fig. 1b with τs (cid:39) −1 and Fig. 3b). This means that the maximum zero-lag correla- tion of the two non-identical oscillators is achieved when the input to the high-frequency oscillator is lagged. Note that in this case changing the lag in the inputs does not change the functional form of the distribution function and so the synchrony index. Unlike the case of uncoupled oscillators the effect of time lag in the inputs to the coupled oscillators is not restricted to the shift of the distribution of the phase lags. For a system of two coupled oscillators, two cases can be recognized. In the first case (locked mode) the deterministic version of Eq. 4 (with no stochastic input) has a stable fixed point, and in the second case there is no fixed point for Eq. 4 and the system is in the running mode. The effect of common inputs in these two cases are shown in Fig. 2a and b, respectively. In the locked mode the synchrony index is no longer independent of the time lag of the inputs and peaks when the time lag matches the phase lag of deterministic case (Fig. 3). Accordingly, while the location of the peak of the distribution func- tion shifts with changing input time lag, its maximum value and the width of the distribution around the most probable phase lag also changes (Fig. 2a). Analytical results obtained by the solving the stationary Fokker- Planck equation (Eq. 6) confirm the results of the simu- lation. In the running mode the results are qualitatively sim- ilar to the locked mode except for the overall decrease in the synchrony index and wider distribution of the phase differences (Fig. 2b and Fig. 3). Changing the time-lag of the inputs the distribution is shifted while its width and maximum are changed. Again, the best entrainment 5 f (θ1(t), θ2(t) + τs) · ξ(t) f (θ1(t(cid:48)), θ2(t(cid:48)) + τs) · ξ(t(cid:48)) fi(θ1(t), θ2(t) + τs)ξi(t) fj(θ1(t(cid:48)), θ2(t(cid:48)) + τs)ξj(t(cid:48)) fi(θ1(t), θ2(t) + τs) fj(θ1(t(cid:48)), θ2(t(cid:48)) + τs) ξi(t)ξj(t(cid:48)) fi(θ1(t), θ2(t) + τs) fi(θ1(t(cid:48)), θ2(t(cid:48)) + τs) δ(t − t(cid:48)) (cid:105)(cid:69) (cid:105)(cid:69) (cid:69) (cid:105)(cid:68) (cid:105) (cid:105) ×(cid:104) (cid:105)(cid:104) (cid:105)(cid:104) (cid:105)(cid:104) 1 2π 0 1 2π (cid:90) 2π (cid:88) (cid:88) (cid:88) 1 2π ij ij = = = 0 (cid:90) dt dt(cid:48)(cid:68)(cid:104) (cid:90) (cid:90) 2π dt dt(cid:48)(cid:68)(cid:104) (cid:90) 2π (cid:90) dt dt(cid:48)(cid:104) (cid:90) (cid:90) 2π dt dt(cid:48)(cid:104) (cid:105)2 (cid:105)2 0 0 1 2π dt f (θ1(t), θ2(t) + τs) i (cid:104) 0 (cid:90) 2π (cid:90) 2π (cid:104) (cid:90) 2π dt 0 1 2π 0 = 1 2π = 1 2π = (cid:104) ¯f (ϕ − τs) (cid:105)2 = 1 2π According to Eqs. A3 and A4 we can define f (t + ϕ1, t + τs + ϕ2) (cid:105)2 f (¯t, ¯t − ϕ + τs) . (A4) d¯t (cid:104) (cid:90) 2π (cid:104) (cid:105)2 dt f (θ1(t), θ2(t) + τs) 0 (A5) and using Eq. 3 and Eq. A2 to A5 we obtain the evolution equation for φ = φ1 − φ2 dϕ(t) dt = ε∆ω + ε √ + (cid:17) (cid:16) ¯H(ϕ) − ¯H(−ϕ) (cid:16) (cid:17) ϕ − τs η(t) ε ¯f where η(t) is another Gaussian white noise with zero mean and unit variance. In a more compact form we would have = εΓ(ϕ) + η(t) (A7) dϕ(t) dt where √ ε ¯f (cid:16) ϕ − τs (cid:17) (cid:16) ¯H(ϕ) − ¯H(−ϕ) (cid:17) Γ(ϕ) = ∆ω + . (A8) The Fokker-Planck equation for the phase difference distribution of Eq. A7 is (ϕ, t) = − ε ∂ρ ∂t + ε [Γ (ϕ) ρ (ϕ, t)] (cid:16) ¯f (ϕ − τs)ρ (φ, t) ∂ ∂ϕ ∂2 ∂ϕ2 (cid:17) (A9) (A3) and the steady state solution of this equation can be achieved by solving ∂ρ(ϕ, t)/∂t = 0. Defining R(ϕ) = ¯f (ϕ − τs)ρ (φ, t) (A10) we can rewrite Eq. A9 for ¯f (ϕ − τs) (cid:54)= 0 as (ϕ, t) = − ε ∂R ∂t + ε (cid:21) R (ϕ, t) (cid:20) Γ (ϕ) ¯f (ϕ − τs) ∂ ∂ϕ ∂2 ∂ϕ2 R (φ, t) . (A11) (A12) (cid:105) General solution of Eq. A11 is R(ϕ) = 1 N e−M (x)dx) + 1 where N and A are constant which can be derived by normalization and periodicity conditions, and (cid:90) ϕ (A eM (ϕ)(cid:104) ϕ(cid:90) (A6) M (ϕ) = dx Γ (x) ¯f (x − τs) , (A13) and then we derive steady state phase difference distri- bution ρ0(ϕ) as Eq. 6 . To check the validity of the results we take Z(θ) = 1 − cos(θ) (A14) which is the canonical form of the phase sensitivity for the type-I oscillators near SNIC bifurcation [44] and assumed that oscillators are pulse-coupled (cid:1) Gij = Σnδ(cid:0)t − tn (cid:16) g12 − g21 j (cid:17) substituting these equations in equations A1 to A8 results Γ(ϕ) = 2π 1 − cos(ϕ) (A16) (A15) and (cid:114)(cid:16) (cid:17) 1 − cos(ϕ − τs) ¯f (ϕ − τs) = . (A17) We have then solved the integrals of Eq. 6 numerically to obtain the analytical result for steady state phase dif- ference distribution which is presented in Fig. 2 of the manuscript. Acknowledgments This research was supported by the Iranian National Elites Foundation, by the Cognitive Sciences and Tech- nologies Council, and by the Slovenian Research Agency (Grant Nos. J1-7009 and P5-0027). [1] A. Pikovsky, M. Rosenblum, and J. Kurths, Synchroniza- tion: A Universal Concept In Nonlinear Sciences (Cam- bridge University Press, Cambridge, 2003). [2] S. H. Strogatz, Sync: The Emerging Science of Sponta- neous Order (Hyperion, New York, 2003). [3] F. Blaabjerg, R. Teodorescu, M. Liserre, and A. V. Tim- bus, IEEE Transactions on Industrial Electronics 53, 1398 (2006). [4] M. Rohden, A. Sorge, M. Timme, and D. Witthaut, Phys. Rev. Lett. 109, 064101 (2012). [5] A. E. Motter, S. A. Myers, M. Anghel, and T. Nishikawa, Nat. Phys. 9, 191 (2013). 6 [21] J. G´omez-Gardenes, Y. Moreno, and A. Arenas, Phys. Rev. Lett. 98, 034101 (2007). [22] J. G´omez-Gardenes, S. G´omez, A. Arenas, Y. Moreno, Phys. Rev. Lett. 106, 128701 (2011). and [23] J. Aguirre, R. Sevilla-Escoboza, R. Guti´errez, D. Papo, and J. Buld´u, Phys. Rev. Lett. 112, 248701 (2014). [24] Y. Kuramoto, Chemical oscillations, waves, and turbu- lence (Springer, New York, 1984). [25] K. Pyragas, Phys. Rev. E 58, 3067 (1998). [26] M. S. Yeung and S. H. Strogatz, Phys. Rev. Lett. 82, 648 (1999). [27] W. S. Lee, E. Ott, and T. M. Antonsen, Phys. Rev. Lett. 103, 044101 (2009). [28] Q. Wang, M. Perc, Z. Duan, and G. Chen, Phys. Rev. E 80, 026206 (2009). [29] S. Sadeghi and A. Valizadeh, Journal of computational neuroscience 36, 55 (2014). [30] M. Dhamala, V. K. Jirsa, and M. Ding, Phys. Rev. Lett. 92, 074104 (2004). [31] Z. G. Esfahani, L. L. Gollo, and A. Valizadeh, Sci. Rep. 6, 23471 (2016). [32] A. Pototsky and N. B. Janson, Physical Review E 80, 066203 (2009). [33] M. K. Sen, B. C. Bag, K. G. Petrosyan, and C.-K. Hu, Journal of Statistical Mechanics: Theory and Experi- ment 2010, P08018 (2010). [34] O. D'Huys, T. Jungling, and W. Kinzel, Physical Review E 90, 032918 (2014). [35] D. S. Goldobin and A. Pikovsky, Physica A 351, 126 (2005). [36] J.-n. Teramae and D. Tanaka, Phys. Rev. Lett. 93, 204103 (2004). [37] A. S. Pikovsky, Physics Letters A 165, 33 (1992). [38] L. Yu, E. Ott, and Q. Chen, Physical review letters 65, [6] P. Barbara, A. Cawthorne, S. Shitov, and C. Lobb, Phys. 2935 (1990). Rev. Lett. 82, 1963 (1999). [39] B. Hauschildt, N. Janson, A. Balanov, and E. Scholl, [7] K. K. Likharev, Dynamics of Josephson Junctions and Physical Review E 74, 051906 (2006). Circuits (Gordon & Breach, Philadelphia, 1986). [40] Y. Kawamura and H. Nakao, Physical Review E 94, [8] J. Jalife, J. Physiol. 356, 221 (1984). [9] C. S. Peskin, Mathematical Aspects of Heart Physiology (Courant Institute, New York, 1975). [10] C. M. Gray and W. Singer, Proc. Natl. Acad. Sci. USA 86, 1698 (1989). 032201 (2016). [41] E. Bolhasani, Y. Azizi, and A. Valizadeh, Correlated neu- ronal activity and its relationship to coding, dynamics and network architecture p. 192 (2014). [42] A. T. Winfree, The geometry of biological time (Springer, [11] M. I. Rabinovich, R. Huerta, and P. Varona, Phys. Rev. New York, 1980). Lett. 96, 014101 (2006). [43] G. B. Ermentrout and N. Kopell, J. Math. Biol. 29, 195 [12] M. Besserve, S. C. Lowe, N. K. Logothetis, B. Scholkopf, (1991). and S. Panzeri, PLoS Biol. 13, e1002257 (2015). [13] L. M. Pecora and T. L. Carroll, Phys. Rev. Lett. 64, 821 (1990). [14] N. F. Rulkov, M. M. Sushchik, L. S. Tsimring, and H. D. Abarbanel, Phys. Rev. E 51, 980 (1995). [15] M. G. Rosenblum, A. S. Pikovsky, and J. Kurths, Phys. Rev. Lett. 76, 1804 (1996). [16] A. Neiman, L. Schimansky-Geier, A. Cornell-Bell, and F. Moss, Phys. Rev. Lett. 83, 4896 (1999). [44] E. M. Izhikevich, Dynamical Systems in Neuroscience (MIT press, Cambridge, Massachusetts, 2007). [45] W. Gerstner, Phys. Rev. Lett. 76, 1755 (1996). [46] D. Hansel and G. Mato, Phys. Rev. Lett. 86, 4175 (2001). [47] S. H. Strogatz, Nature 410, 268 (2001). [48] P. Tass, M. Rosenblum, J. Weule, J. Kurths, A. Pikovsky, J. Volkmann, A. Schnitzler, and H.-J. Freund, Phys. Rev. Lett. 81, 3291 (1998). [49] S. D. Burton, G. B. Ermentrout, and N. N. Urban, J. [17] S. Boccaletti, J. Kurths, G. Osipov, D. Valladares, and Neurophysiol. 108, 2115 (2012). C. Zhou, Phys. Rep. 366, 1 (2002). [18] A. Arenas, A. D´ıaz-Guilera, and C. J. P´erez-Vicente, Phys. Rev. Lett. 96, 114102 (2006). [19] A. Arenas, A. D´ıaz-Guilera, J. Kurths, Y. Moreno, and C. Zhou, Phys. Rep. 469, 93 (2008). [20] I. Fischer, R. Vicente, J. M. Buld´u, M. Peil, C. R. Mi- rasso, M. Torrent, and J. Garc´ıa-Ojalvo, Phys. Rev. Lett. 97, 123902 (2006). [50] S. Boccaletti, G. Bianconi, R. Criado, C. I. Del Ge- nio, J. G´omez-Gardenes, M. Romance, I. Sendina-Nadal, Z. Wang, and M. Zanin, Phys. Rep. 544, 1 (2014). [51] M. Kivela, A. Arenas, M. Barthelemy, J. P. Gleeson, Y. Moreno, and M. A. Porter, J. Complex Netw. 2, 203 (2014).
1703.05406
1
1703
2017-03-15T22:25:29
A Pulse-Gated, Predictive Neural Circuit
[ "q-bio.NC" ]
Recent evidence suggests that neural information is encoded in packets and may be flexibly routed from region to region. We have hypothesized that neural circuits are split into sub-circuits where one sub-circuit controls information propagation via pulse gating and a second sub-circuit processes graded information under the control of the first sub-circuit. Using an explicit pulse-gating mechanism, we have been able to show how information may be processed by such pulse-controlled circuits and also how, by allowing the information processing circuit to interact with the gating circuit, decisions can be made. Here, we demonstrate how Hebbian plasticity may be used to supplement our pulse-gated information processing framework by implementing a machine learning algorithm. The resulting neural circuit has a number of structures that are similar to biological neural systems, including a layered structure and information propagation driven by oscillatory gating with a complex frequency spectrum.
q-bio.NC
q-bio
A Pulse-Gated, Predictive Neural Circuit Yuxiu Shao∗, Andrew T. Sornborger†, Louis Tao‡ ∗Center for Bioinformatics, National Laboratory of Protein Engineering and Plant Genetic Engineering, College of Life Sciences, Peking University, Beijing, China Email: [email protected] † Department of Mathematics, University of California, Davis, USA Email: [email protected] ‡Center for Bioinformatics, National Laboratory of Protein Engineering and Plant Genetic Engineering, College of Life Sciences, and Center for Quantitative Biology, Peking University, Beijing, China Email: [email protected] ATS and LT are corresponding authors. Abstract -- Recent evidence suggests that neural information is encoded in packets and may be flexibly routed from region to region. We have hypothesized that neural circuits are split into sub-circuits where one sub-circuit controls information propagation via pulse gating and a second sub-circuit processes graded information under the control of the first sub-circuit. Using an explicit pulse-gating mechanism, we have been able to show how information may be processed by such pulse- controlled circuits and also how, by allowing the information processing circuit to interact with the gating circuit, decisions can be made. Here, we demonstrate how Hebbian plasticity may be used to supplement our pulse-gated information processing framework by implementing a machine learning algorithm. The resulting neural circuit has a number of structures that are similar to biological neural systems, including a layered structure and information propagation driven by oscillatory gating with a complex frequency spectrum. Index Terms -- neural circuit, neuromorphic circuit, informa- tion gating, pulse gating, autoregressive prediction. I. INTRODUCTION A considerable experimental literature indicates that 1) oscillations of various frequencies are important for the mod- ulation of interactions in neural systems [1], [2], [3], 2) each individual pulse that makes up an oscillation may be implicated in the parallel transfer of a packet of information [4], [5], [6], [7], and 3) neural systems exist for the control of information propagation [8]. We have begun to build a neural information processing framework based on the pulse-gated propagation of graded (amplitude dependent) information [9], [10]. A key question that arises in understanding information processing in the brain is: How may neural plasticity be used to form computational circuits? In this paper, we explore this question by creating a predictive neural circuit based on the pulse-gated control of firing rate information. We outline a set of sub-circuits re- sponsible for sub-computations needed to estimate the process coefficients of an autoregressive (AR) learning circuit. We then combine the sub-circuits to demonstrate that a neural system can use Hebbian learning in concert with pulse-gating to predict an AR process. A. Autoregressive Processes II. METHODS AR and moving-average (MA) processes are the two princi- ple linear models that are used to make statistical predictions [11], [12]. AR models are filters of length n on a time series and we will consider these here. The object of this work is to implement an algorithm for predicting an AR process in a substrate of neurons. Below, we denote the filter by a, the individual process covariances by σi, and the process variance- covariance matrix by Σ. An AR(n) process for a zero-mean random variable, xt, is defined xt = aixt−i + t . To find the filter, we need to solve the Yule-Walker equa- tions, which arise from the covariance structure of the process as follows: n(cid:88) i=1 i=1 n(cid:88) n(cid:88) n(cid:88) i=1 i=1 (cid:104)xtxt(cid:105) = (cid:104)xtxt−1(cid:105) = ... (cid:104)xtxt−n(cid:105) = ai(cid:104)xt−ixt(cid:105) + (cid:104)txt(cid:105) ai(cid:104)xt−ixt−1(cid:105) ai(cid:104)xt−ixt−n(cid:105) where (cid:104)(cid:105) denotes an expectation value over t. Defining σi = (cid:104)xt−ixt(cid:105) = (cid:104)xtxt−i(cid:105), from the second through n'th equations above, we have where σ = (σ1, . . . , σn), σ0  Σ = σ = Σa , σ1 σ0 ... . . . σn−1 ... ... σ0  , σ1 ... σn−1 (1) (2) and a = (a1, . . . , an). B. Push-Me Pull-You Neuron Pairs In our AR process, xt, since we are considering an im- plementation in a neural circuit where firing rates encode information, the input to our system will be positive semi- definite. Both negative and positive values of xt − m0, where m0 is the mean, must be able to be represented by the circuit. We therefore define neuron pairs where one neuron of the pair represents positive values (i.e. the amplitude above the mean) and the other neuron represents negative values (i.e. the absolute value of the amplitude below the mean). At any given moment, only one neuron will encode non-zero amplitude in such a pair. In order to read in and encode positive and negative values of the input in a pair of neurons, we let m0 = (cid:104)xt(cid:105), and τ dm1 dt τ dm2 dt = −m1 + [x − m0 + p(t) − g0]+ = −m2 + [m0 − x + p(t) − g0]+ , where p(t) is a square pulse of duration T . Here, g0 is a threshold value and we assume that x − m0 is less than the amplitude of the pulse, p(t), so neuron 1 (firing rate m1) will only fire when the input is simultaneously pulsed and above the mean and, similarly, neuron 2 will only fire when the input is simultaneously pulsed and below the mean. We call such a pair of neurons a push-me pull-you (PMPY) pair. Additionally, PMPY amplitudes may be exactly propagated along a chain via pulse-gating, and we will also designate pairs along the chain as PMPY pairs. C. Recursive, Gradient Descent Solution to σ = Σa To make predictions, we need to estimate a for the process. To solve σ = Σa, we find the (unique) zero of σ − Σa using gradient descent: (cid:0)(cid:107)σ − Σa(cid:107)2(cid:1) τ da dt = − 1 2 = Σ (σ − Σa) . Discretizing to first order in time gives δ δa x(t) x+ x+ n+1 n+1 m0 x− x− n+1 n+1 c1+ c1+ n+1 n+1 c1− c1− n+1 n+1 c2+ c2+ n+1 n+1 c2− c2− n+1 n+1 x+ x+ n n x− x− n n c1+ c1+ n n c1− c1− n n c2+ c2+ n n c2− c2− n n . . . . . . . . . . . . . . . . . . . . . . . . Hebb . . . . . . . . . . . . x+ x+ 2 2 x− x− 2 2 c1+ c1+ 2 2 c1− c1− 2 2 c2+ c2+ 2 2 c2− c2− 2 2 x+ x+ 1 1 x− x− 1 1 c1+ c1+ 1 1 c1− c1− 1 1 c2+ c2+ 1 1 c2− c2− 1 1 , σ+− Fig. 1. The neural circuit that computes covariances, σ++ , etc., between lagged values, x(t − T ), of the input time series. The mean, m0, is first removed from the series. Then, using PMPY pairs, positive and negative values of the series are propagated for n+1 lags. These same values are copied into two sets of populations among which are feedforward, all-to-all connections. Hebbian plasticity (Hebb) acts on the synapses between the two sets of populations, generating a synaptic connectivity that is an estimate of the lagged covariance. i i i ≡ (cid:104)x− σ−− Note also that if we define t−ix− t (cid:105). Note that σi = σ++ i − σ+− i − σ−+ i + σ−− i .  ,  (4) Ξ = Γ = 0 0 0 0  1 −1  σ++ 0 σ−+ 0 σ++ 1 σ−+ ... ... (σ++ 2 , . . .), 1 0 1 1 0 0 0 1 −1 0 0 σ+− σ−− σ+− σ−− ... ... , σ+− 0 0 0 0 1 −1 σ+− σ−− σ+− σ−− ... ... , . . .), σ++ 1 σ−+ 1 σ++ 0 σ−+ ... ... , σ+− , σ++ 1 0 0 0 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... ... . . . . . . . . . . . . ... ... and an+1 = an + Σ (σ − Σan) , ∆t τ (3) and, since the eigenvalues of the symmetric matrix Σ are positive and as long as ∆t/τ is sufficiently small, iteration will give a∞ → Σ−1σ as τ → ∞. In order to see how this solution may be computed using positive-only elements such as when performing the calcula- tion with a set of PMPY pairs, we consider a pair of positive, semi-definite time series derived from xt, (x+ t ), where t − x− x+ t = xt. That is, the part of xt that is greater than the mean in a PMPY pair is contained in x+ t , and the absolute value of the part of xt that is less than the mean is contained in x− t . The covariances associated with this time series are i ≡ (cid:104)x+ t (cid:105), and σ++ i ≡ (cid:104)x− i ≡ (cid:104)x+ t (cid:105), σ−+ t (cid:105), σ+− t−ix− t , x− t−ix+ t−ix+ = 1 then Ξγ = σ, and ΞΓΞT = Σ. p 2 2 1 2 , a− = γ 1 , a− (a+ 1 , a+ Thus, if we first compute the solution to = Γ (γ − Γp) τ dp dt then, Ξp = a. It may be seen directly that this gradient descent projects to the one described above since = ΞΓΞT(cid:0)Ξγ − ΞΓΞT Ξp(cid:1) τ da dt = τ Ξ dp dt = Σ (σ − Σa) . (5) The reason this works is that multiplication of a vector, v, by Ξ returns a new vector containing the differences of each pair of elements of v, and the matrix product ΞXΞT returns pn c1 Q1Γ c2 m1 m2 m3 c1 Q1Γ c2 ∆t τ I pn+1 u c1 Γ c2 Q2 y − + pn Fig. 2. The gradient descent algorithm: the coefficient estimate, pn, a vector of PMPY pairs, here represented as one node in the graph, is copied via pulse gating into c1, the vector of PMPY pairs in the middle two rows of Fig. 1, here represented as one node. This vector is then transformed via the synaptic connectivities, Γ, learned with Hebbian plasticity, using appropriate pulsing, Q1 acts on the output, with the result that Q1Γpn is written into the PMPY vector c2, in the lower two rows of Fig. 1. This result is then propagated through a set of memory populations, m1,...,3. After a sufficient delay such that c1 has had time to decay to a firing rate of approximately 0, u is read into c1, then transformed such that Γu is now in c2. The fixed connectivity Q2 acts on this result giving Q2Γu in y. The streams are then merged with inhibitory connectivity acting to subtract m3 from y, such that the value Q2Γu − Q1Γpn is in z. Finally, via the last two operations, pn+1 is set to pn + ∆t τ Q1Γ (γ − Q1Γpn). In this way, one step of gradient descent is achieved. a new matrix containing the difference of the sum of diagonal elements of each 2× 2 block submatrix and the sum of its off diagonal elements. Finally, it will be useful in the discussion of the neural circuit used to implement gradient descent to extend the above considerations by using a matrix Γ(cid:48), which is defined as in (4), but based on an (n + 1) × (n + 1) matrix Σ. Thus, Γ(cid:48) will contain covariances up to a lag of n + 1. In this case, we can explicitly write the gradient descent algorithm as Q1Γ(cid:48) (γ(cid:48) − Q1Γ(cid:48)p(cid:48) p(cid:48) n+1 = p(cid:48) n) , n + (6) ∆t τ where γ(cid:48) ≡ Q2Γ(cid:48)u, u ≡ (1, 0, 0, 0, . . .), Q1 = diag([1, 1, . . . , 1, 1, 0, 0]) Q2 = diag([1, 1, . . . , 1, 1]n, 2) (i.e. Q2 is a matrix of zeros, except for a 2nd superdiagonal of 1's), and p(cid:48) = (a+ 1 , a− 1 , . . . , a+ n , a− n , 0, 0) . We have done this seemingly nonsensical extension because now γ(cid:48) is generated by a vector input to the covariance matrix Γ(cid:48), as opposed to being a fixed vector. Thus, the algorithm will work for an arbitrary AR(n) time series covariance. D. The Neural Circuit Throughout our circuits, we use the pulse-gated propagation of firing rates described in [9]. i i , σ+− 1) Computing Lagged Covariances: The first major struc- ture in the circuit is set up to use Hebbian plasticity to calculate covariances, σ++ , etc., between successively delayed random variables in the input. To do this, the input, x(t), is discretized and bound via a pulse into the circuit as current i ≡ [x(ti) − m0]+ and or firing rate amplitude packets, x+ i ≡ [m0 − x(ti)]+, in PMPY pairs. These packets are x− propagated in a chain of n + 1 pairs. Additionally, two copies of each pair are made using a pulse gated copying procedure. Hebbian plasticity is meant to mimic synaptic plasticity. It is a phenomenon in which coincident spiking activity between two neurons serves to increase the synaptic weight between the neurons. Since the probability of two spikes being coincident is proportional to the product of the firing rates, over time Hebb Hebb 1 1 Hebb . . . Hebb Hebb 1 1 a+ 1 a− 1 . . . a+ n a− n a+ 1 a− 1 x+ n−1 x− n−1 x+ n−1a+ 1 x− n−1a− 1 + . . . xn+1 a+ n a− n x+ 1 a+ n x− 1 a− n . . . x+ 1 x− 1 Fig. 3. Prediction: On the left, populations initialized with the value 1 are connected one-to-one with populations containing estimated coefficients. Hebbian plasticity causes the values of the estimated coefficients to be encoded in the synapses. On the right, once the synapses are sufficiently stable, lagged time series values are written into the populations previously initialized with 1 (1-populations). With a pulse, the positive components are computed, then summed, resulting in a prediction of x(t + 1)+. Initializing the 1-populations with swapped + and − components (not shown) of x results in x(t + 1)−. Estimates of future values of x may subsequently be made with the same circuit. the synaptic strength becomes an estimate of the covariance between the firing rates of the two neurons (or in our case, neuronal populations). In our AR circuit, the populations of copied PMPY pairs interact with all-to-all connectivity (i.e. both populations in each pair in the first set of copies are connected with both populations in each pair in the second set of copies, see Fig. 1). Synaptic weights, {sj}, are set according to τs dsj dt = −(sj − x(t)y(t)) , functions, z(t) =(cid:80) where x and y are firing rates from two neuronal populations. We assume that the input is in the form of a train of delta i ziδ(t − ti), that are read in by a pulse- gated neural population. After gating into a neural population, we have x(t) = z(t)∗ G(t), where G(t) is the pulse envelope Fig. 4. The Basic Pulse-Gated Algorithm: Connectivity matrix (top left), pulse sequence (top center) and neuronal population firing rates (top right for a single update and bottom for the complete evolution). Covariance, gradient descent, and prediction circuits are activated such that they do not interfere with each other, see text for complete description. from the gating operation, (cid:26) t τ e−t/τ , τ e−t/τ , T < t < ∞ . 0 < t < T T G(t) = In the limit that the Hebbian timescale is much greater than the synaptic timescale, τs (cid:29) τ, and the gated pulses overlap only negligibly, we have (cid:80) sj(t) ≈ σj , where σj = 4τ /τs between x and y over τs/4τ samples. i zizi−j is an estimate of the covariance Once the synaptic strengths in the copy populations have reached an equilibrium, and using pulse gating such that the synapses are feedforward between PMPY copy 1 and PMPY copy 2, we have c2 = Γc1. Here, for simplicity, we have discarded the (cid:48) and just write Γ. 2) Gradient Descent: To implement the gradient descent part of the algorithm, we implement (6) using pulse gating. A method to do this is shown in Fig. 2. Here, a vector of PMPY pairs containing pn is gated into the first copy of lagged values from the time series. The synaptic connectivity, learned using Hebbian plasticity, that contains an estimate of the covariance matrix is used (along with appropriate pulses) to perform the operation c2 := Q1Γpn. This result is then stored in a short- term memory. During this calculation, but after sufficient delay that c1 may be reused, u is gated into c1 and, via the matrix Γ and another connectivity matrix that encodes Q2, Q2Γu is computed and stored in y. The two computation streams are then combined with appropriate inhibition to subtract m3 from y giving z = γ − Q1Γpn. Two subsequent operations result in pn+1 being set to pn + ∆t τ Q1Γ (γ − Q1Γpn) 3) Prediction: To predict future values of x(t), we need to form estimates of x(t) using the AR(n) process coefficients. The computation to do this requires us to compute n(cid:88) n(cid:88) i=1 and (cid:104)x+ t (cid:105) = (cid:104)x− t (cid:105) = ((cid:104)a+ t−i(cid:105) + (cid:104)a− i x− t−i(cid:105)) i x+ ((cid:104)a+ i x− t−i(cid:105) + (cid:104)a− i x+ t−i(cid:105)) i=1 1 , a− using the coefficients, p = (a+ 1 , . . .), that we have esti- mated with the gradient descent algorithm. One way that this may be done is depicted in Fig. 3. Here, by initializing a set of populations to 1 (1- populations), connecting them one-to-one with the elements of pn, Hebbian plasticity causes the synapses to encode the values of the elements of p. This is a method for creating a long-term memory. Subsequently, x is copied to the 1- populations, then gated to compute {(cid:104)a+ t−i(cid:105), . . .}. These values are then summed to compute (cid:104)x+ t (cid:105). By swapping the + and − elements of x, then copying to the 1-populations, the same circuit results in (cid:104)x− t (cid:105). The output may be stored in memory or used in other sub-circuits in the usual way. t−i(cid:105),(cid:104)a− i x− i x+ III. RESULTS In Fig. 4, we depict results for the AR(2) process, xn+1 = a1xn + a2xn−1 + , where  is a zero mean Gaussian noise process, a1 = 3/4 and a2 = −1/2. At the beginning of the sequence, AR input to short term memory is silenced to allow operation of the gradient descent circuit. At the same time, Using only pulse-gating, Hebbian learning and standard neuronal synaptic properties, we implemented a short-term memory, a gradient descent algorithm, a long-term memory and a method for computing an inner product to make a prediction. Additionally, the structure of the algorithm defines the brain rhythms that arise during information processing. In Fig. 6, we show the spectra of the gating pulses in the pulse sequence. For this figure, we used pulses of 10ms duration, resulting in strong gamma and theta rhythms. To our knowledge, this is the first time a neuronal model has been used to relate algorithmic structure with neuronal oscillation structure. ACKNOWLEDGMENT L.T. thanks the UC Davis Mathematics Department for its hospitality. A.T.S. would like to thank Liping Wei and the Center for Bioinformatics at the College of Life Sciences at Peking University for their hospitality. This work was supported by the Natural Science Foundation of China grant 91232715 (YXS and LT), by the Open Research Fund of the State Key Laboratory of Cognitive Neuroscience and Learning grant CNLZD1404 (YXS and LT), and by the Beijing Mu- nicipal Science and Technology Commission under contract Z151100000915070 (YXS and LT). REFERENCES [1] R. Azouz and C. Gray, "Dynamic spike threshold reveals a mechanism for synaptic coincidence detection in cortical neurons in vivo," Proc. Natl. Acad. Sci. USA, vol. 97, pp. 8110 -- 8115, 2000. [2] P. Fries, T. Womelsdorf, R. Oostenveld, and R. Desimone, "The effects of visual stimulation and selective visual attention on rhythmic neuronal synchronization in macaque area V4," J. Neurosci., vol. 28, pp. 4823 -- 4835, 2008. [3] A. Markowska, D. Olton, and B. Givens, "Cholinergic manipulations in the medial septal area: Age-related effects on working memory and hippocampal electrophysiology," J. Neurosci., vol. 15, pp. 2063 -- 2073, 1995. [4] P. Fries, "A mechanism for cognitive dynamics: Neuronal communica- tion through neuronal coherence," Trends Cogn. Sci., vol. 9, pp. 474 -- 480, 2005. [5] E. Salinas and T. Sejnowski, "Impact of correlated synaptic input on output firing rate and variability in simple neuronal models," J. Neurosci., vol. 20, pp. 6193 -- 6209, 2000. [6] A. Luczak, P. Bartho, S. L. Marguet, G. Buzsaki, and K. D. Harris, "Sequential structure of neocortical spontaneous activity in vivo," Proc. Natl. Acad. Sci. U.S.A., vol. 104, pp. 347 -- 352, 2007. [7] A. Luczak, B. L. McNaughton, and K. D. Harris, "Packet-based commu- nication in the cortex," Nat. Rev. Neurosci., vol. 16, no. 12, pp. 745 -- 755, Dec 2015. [8] T. Gisiger and M. Boukadoum, "Mechanisms Gating the Flow of Information in the Cortex: What They Might Look Like and What Their Uses may be," Front Comput Neurosci, vol. 5, p. 1, 2011. [9] A. Sornborger, Z. Wang, and L. Tao, "A mechanism for graded, dynamically routable current propagation in pulse-gated synfire chains and implications for information coding," J. Comput. Neurosci., vol. 39, pp. 181 -- 195, August 2015. [10] Z. Wang, A. T. Sornborger, and L. Tao, "Graded, Dynamically Routable Information Processing with Synfire-Gated Synfire Chains," PLoS Com- put. Biol., vol. 12, no. 6, p. e1004979, Jun 2016. [11] S. Pandit and S.-M. Wu, Eds., Time Series and System Analysis with Applications. New York: John Wiley & Sons, 1983. [12] D. Percival and A. Walden, Eds., Spectral Analysis for Physical Appli- cations. Cambridge, UK: Cambridge University Press, 1993. Fig. 5. Covariance and AR Coefficient Estimates: The evolution of estimates converges after approximately 30, 000 updates (30 pulses per update). Fig. 6. Spectrum of AR Pulse Sequence: For gating pulses of length 10ms, the pulse sequence spectrum has peaks in the theta and gamma bands. the process coefficients, encoded in firing rate amplitudes are propagated to the coefficient learning and prediction circuit to be encoded in synaptic weights. Once the gradient descent circuit and learning and prediction circuits have updated the process coefficient estimates, a1 and a2, information needed for memory is propagated internally to each circuit and AR input is allowed to enter the covariance learning circuit. This update sequence is then repeated to obtain progressively better estimates over a long time (100, 000 repeats for this simulation). In Fig. 5, we show the convergence of the algorithm to steady state estimates for both the covariance matrix elements and AR process coefficients. Final AR process coefficients are within 6% of theoretical values. Better estimates may be obtained by increasing the learning time constant, τ. IV. CONCLUSION Using a pulse-gating paradigm for propagating information in a neural circuit in concert with Hebbian learning, we have described how to implement an algorithm for predicting future values of an arbitrary AR(n) process. We have implemented the algorithm for an AR(2) process. The algorithm consists of three sub-circuits responsible for 1) learning the covariance matrix, 2) computing the process coefficients and 3) making a prediction by learning the process coefficients, then perform- ing the prediction.
1902.03815
1
1902
2019-02-11T11:06:50
Interspike interval correlations in networks of inhibitory integrate-and-fire neurons
[ "q-bio.NC", "math.DS", "physics.bio-ph", "physics.data-an" ]
We study temporal correlations of interspike intervals (ISIs), quantified by the network-averaged serial correlation coefficient (SCC), in networks of both current- and conductance-based purely inhibitory integrate-and-fire neurons. Numerical simulations reveal transitions to negative SCCs at intermediate values of bias current drive and network size. As bias drive and network size are increased past these values, the SCC returns to zero. The SCC is maximally negative at an intermediate value of the network oscillation strength. The dependence of the SCC on two canonical schemes for synaptic connectivity is studied, and it is shown that the results occur robustly in both schemes. For conductance-based synapses, the SCC becomes negative at the onset of both a fast and slow coherent network oscillation. Finally, we devise a noise-reduced diffusion approximation for current-based networks that accounts for the observed temporal correlation transitions.
q-bio.NC
q-bio
Interspike interval correlations in networks of inhibitory integrate-and-fire neurons Wilhelm Braun1, 2, ∗ and André Longtin2 1Institut für Genetik, Universität Bonn, Kirschallee 1, 53115 Bonn, Germany 2Department of Physics and Centre for Neural Dynamics, University of Ottawa, 598 King Edward, Ottawa K1N 6N5, Canada (Dated: February 12, 2019) 9 1 0 2 b e F 1 1 ] . C N o i b - q [ 1 v 5 1 8 3 0 . 2 0 9 1 : v i X r a We study temporal correlations of interspike intervals (ISIs), quantified by the network-averaged serial correlation coefficient (SCC), in networks of both current- and conductance-based purely inhibitory integrate-and-fire neurons. Numerical simulations reveal transitions to negative SCCs at intermediate values of bias current drive and network size. As bias drive and network size are increased past these values, the SCC returns to zero. The SCC is maximally negative at an intermediate value of the network oscillation strength. The dependence of the SCC on two canonical schemes for synaptic connectivity is studied, and it is shown that the results occur robustly in both schemes. For conductance-based synapses, the SCC becomes negative at the onset of both a fast and slow coherent network oscillation. Finally, we devise a noise-reduced diffusion approximation for current-based networks that accounts for the observed temporal correlation transitions. I. INTRODUCTION Quantifying the statistics of spiking in single neuron models is an important goal in computational neuro- science. A large body of work has dealt with first-order statistics of interspike intervals in canonical single neu- ron models, often computed in the context of memory- free renewal first-passage problems [1]. In recent years, progress has been made in tackling the much harder prob- lem of second-order spiking statistics in non-renewal neu- ron models, among them the serial correlation coefficient (SCC) at lag k, which is defined as the Pearson correla- tion between ordered ISI sequences shifted by an amount k, e.g., for k = 1, the SCC at lag 1 is the correlation coef- ficient between adjacent ISIs. Its computation has uncov- ered enhanced information transmission in single neurons via the effect of noise shaping [2] and increased detectabil- ity of weak signals [3]. In particular, it is now possible to compute, in some cases even analytically, the SCC for single-neuron models in the presence of spike-frequency adaptation [4], colored noise [5] and time-dependent de- terministic or stochastic firing thresholds [6]. Recently, it was also shown that, at short observation times, negative ISI correlations can enhance the resolution of a nonlinear dynamical sensor whose design was inspired by a sim- ple non-renewal neuron model [7]. There also have been recent efforts to characterize the patterning of ISI se- quences using ordinal analysis, revealing parameter sets that maximize the probability of certain patterns [8]. In contrast, very few studies deal with the computation of temporal correlations in networks of neurons. In [9], a theory to compute the SCC in networks of adapting neurons based on hazard functions is put forward; the main focus of that study is the computation of first-order statistics, such as the mean ISI or the average activity of neurons in the network. When it comes to the compu- ∗ [email protected] tation of SCCs in networks, a notable exception are the recent studies [10], were the main focus is on the self- consistent computation of the spectra of neuronal spike trains in asynchronous networks of excitatory and in- hibitory neurons. Iterative numerical methods are de- scribed to compute the spike train power spectrum self- consistently simulating one single representative neuron only. This adds to previous analyses of network syn- chrony in terms of a single "effective" neuron [11]. Be- cause the approach directly simulates spike trains, the SCC is also computed and shows weak positive and nega- tive values. However, the validity of the proposed scheme is limited to the asynchronous state without a global os- cillation of the network activity. Moreover, there are a number of studies dealing with properties of networks in which each constituent neuron is endowed with an adap- tation mechanism as an intrinsic correlation-generating, highlighting the benefits of adaptation for information transmission, reliable neural coding and noise shaping [12, 13]. Here we focus solely on the ISI correlations of single neurons in networks that transition from asynchronous to synchronous behavior. The motivation to study this setup is twofold. First of all, it is desirable to under- stand how the statistics of a neuron embedded in a net- work changes compared to the well-studied isolated case, because neuronal coding often relies on populations of neurons. In particular, in networks of neurons without correlation-generating intrinsic mechanisms, under what conditions do correlations arise and how are they maxi- mized? Secondly, precise spiking patterns generated by neurons embedded in a network can potentially be se- lected by post-synaptic plasticity mechanisms in target cells downstream from the network, e.g. spike-timing de- pendent plasticity [14 -- 16] or short-term facilitation and depression [17, 18], leading to enhanced weak signal de- tectability [19]. It is thus desirable to understand how networks of spiking neurons generate temporal correla- tions in the constituting neurons' output spike trains. Our study explores the array of questions above by fo- cusing solely on very short range ISI correlations (in fact, only correlations between neighboring ISIs), for the sake of conciseness. We stress from the outset that in strongly oscillatory network states, correlations extend beyond more than one lag, as e.g. observed in [20, 21]. Also, we do not consider correlations between neurons in this study, which is set in the context of sparse networks. A large body of work exists that describes the pairwise cor- relation coefficient between the spike trains or membrane voltages of two different neurons in networks with differ- ent topologies [22 -- 25]. The paper is organized as follows. In Sec. II, we present the network model, synaptic connectivity rules and the central quantities of interest for this study, among them the mean network SCC. We describe the parameter de- pendence of the mean network SCC at the onset of a coherent global network oscillation in Sec. III. Whereas the focus in this section is on networks with current-based synapses, we briefly extend our findings to networks with conductance-based synapses that are capable of generat- ing both slow and fast gamma oscillations. Finally, in Sec. IV, we explore how the mean network SCC can be computed by an effective single offline neuron receiving time-dependent input recorded online during a full net- work simulation. We conclude with a brief discussion of our results in Sec. V. II. NETWORK OF INHIBITORY NEURONS A. Neuron and synaptic dynamics 1. Networks with current-based synapses We consider N leaky integrate-and-fire (LIF) neurons with membrane voltages Xi, i ∈ [1, 2, ..., N ], with Xi ∈ (−∞, xth]. They evolve according to [11] X. i t. = γt [I0 − Xi + I syn i (t)] + √γtσ0ξi(t) . (1) i For conciseness, a factor carrying units of resistance in front of I0 and I syn has been set to one and omitted. When Xi(t) reaches the threshold xth from below, a spike is recorded and Xi is reset to xr immediately; when we consider a refractory period of length τr, the neuron is kept at this membrane voltage for a time τr. The mem- brane time constant is τt = 1 . ξi is a standard zero- γt mean Gaussian white noise. If not mentioned otherwise, we choose xr = 10 mV, xth = 20 mV, τt = 20 ms. The current-based inhibitory synaptic current is given by I syn i (t) = τtJ δ(t − tk j − D) . (2) C(cid:88)k=1(cid:88)j J < 0 is the inhibitory synaptic strength. We choose J = −0.1 mV and fix the synaptic delay D = 2 ms. 2 Hence, the outer sum in Eq. 2 is over C neurons in the presynaptic neighborhood of size C, while the inner sum j . For more is over the spikes j of the neuron k at times tk details on implementation, we refer to Appendix A 1. Finally, I0 is the external bias current, and σ0 is the strength of the external noise. If not mentioned other- wise, we choose σ0 = 1 mV. The single neuron fires periodically in the absence of private noise and synaptic input (i.e. is in the suprathreshold regime) if I0 > 20 mV. In the diffusion approximation, we may express the synaptic current as follows [11]: (t) = µ(t) + σ(t)√τtξi(t) , I syn i (3) with µ(t) = µloc , (4) where the local part of the average synaptic current is related to the firing rate ν of the network at time t − D: µloc = CJν(t − D)τt. The strength of the fluctuations is similarly given by σ(t) = σloc , (5) with σloc = J(cid:112)Cν(t − D)τt. 2. Networks with conductance-based synapses For networks with conductance-based synapses, membrane voltage evolves according to the X. i(t) t. CI = gI inh(t)(EI l (cid:0)EI rest − Xi(t)(cid:1)+gI inh−Xi(t))+I0 . (6) Every neuron receives inputs from other inhibitory neu- rons in the network via synapses with time-dependent conductance gI inh(t), whose time course is given by a bi- exponential function: gI inh(t) = gI inh,peaksI inh(cid:34)exp(cid:32)− inh,d(cid:33) − exp(cid:32)− t − τl τ I inh,r(cid:33)(cid:35) , t − τl τ I (7) inh is a constant for t ≥ τl, gI inh,peak (which that ensures gI is defined by computing the maximum of the term in brackets in Eq. 7): inh(t) = 0 otherwise. Here, sI inh reaches its maximum gI 1 s τd−τr τd−τr τd(cid:19) τr τd(cid:19) τd =(cid:18) τr −(cid:18) τr inh. With inh,d and s ≡ sI inh,r, τd ≡ τ I where we set τr ≡ τ I these definitions, increasing τ I inh,r has the effect of pro- longing the time gI inh(t) needs to reach its maximum, and , therefore extends the effect of one pre-synaptic inhibitory pulse by increasing the area under the time course of inh(t). For all remaining parameter values, we refer to gI Appendix C. zero. The main quantity of interest in this paper, which we call the mean network SCC, is the neuron-averaged SCC in a network of N neurons: 3 B. Synaptic connection rules There are different strategies to choose the number of presynaptic neurons C (see Eq. 2) per neuron. When we only fix the connection probability p, not all neurons have the same number of inhibitory synapses. Excluding autapses, the distribution of the number of synapses onto a neuron (i.e. the in-degree) follows a Bernoulli distribu- tion B(N − 1, p) with mean C = p(N − 1) and standard deviation (cid:112)(N − 1)p(1 − p). We will call this scenario the P -fixed case. For N = 1000 and p = 0.2, the stan- dard deviation is ≈ 13, whereas the mean is C ≈ 200. For more details on the implementation, we refer to Ap- pendix A 2. This scenario is in contrast to the case where every neu- ron has exactly C inhibitory synapses. This is the C-fixed case. We always choose C = C for comparisons of the P -fixed vs C-fixed connectivity scenarios. We will see below that these two different prescriptions for synaptic connectivity have a profound influence on the second- order statistics of the network, as e.g. reflected in the mean and standard deviation of the distribution of the SCC across neurons, which we now define. ρ(k = 1) = 1 N N(cid:88)i=1 ρi(k = 1) . (9) The standard deviation of the SCC across neurons is de- fined as std (ρ) =(cid:118)(cid:117)(cid:117)(cid:116) 1 N N(cid:88)i=1 (ρi(k = 1) − ρ(k = 1))2 . (10) It is used in the following to quantify the spread of SCC values at lag 1 across the population. The second quantity of interest is the mean network ISI, which is defined as the mean of the average ISI across neurons: (cid:104)T(cid:105) = 1 N N(cid:88)i=1 E(T i) , (11) where T i is the set of stationary ISIs for neuron i. Again, the standard deviation of the mean ISI, std((cid:104)T(cid:105)), is de- fined analogously. C. The mean network SCC and ISI D. Population activity and power spectra For a single spike train of neuron i, the SCC ρ is defined as ρi(n, k) = E(cid:0)T i n+k(cid:1) − Qi nT i Qi 2(n, k) 1(n, k) , (8) where and n)E(T i n+k) , 1(n, k) = E(T i Qi 2(n, k) =(cid:113)Var(T i Qi n)Var(T i n+k) . Here, T i n denotes the nth ISI in the spike train of neu- ron i. Furthermore, E(...) denotes the expectation across time for one single neuron. In general, the SCC depends on both the lag k and the position n of the ISI in the spike train [5, 26]. We here consider only stationary spike trains, for which the SCC depends only on the lag k be- tween different ISIs. We will denote this quantity by ρ(k) and focus on k = 1, i.e. on successive ISI correlations. Significant short-term correlations between ISIs will be typically reflected in deviations of this coefficient from i=1(cid:80)j δ(t− ti We denote the population activity(cid:80)N j) of a network of purely inhibitory neurons by ν(t). The power spectral density (PSD) of the population activity is com- puted from the population activity ν(t). Similarly, the averaged single-neuron spectrum (cid:104)Sxx(f )(cid:105) is computed It is the average from all spikes of the neuron group. PSD of spike trains xi =(cid:80)j δ(t− ti j), scaled such that as f → ∞, (cid:104)Sxx(f )(cid:105) approaches the inverse mean ISI given by Eq. 11. For more details on the implementation, we refer to Appendix A 3. Formal definitions of the popula- tion PSD and the averaged single neuron power spectrum can be found e.g. in [27]. III. BEHAVIOR OF THE MEAN NETWORK SCC WITH VARYING NETWORK SIZE AND BIAS CURRENT In the presence of a noisy network oscillation, one expects the SCC of a single neuron in the network to be negative. A simple toy problem illustrates this point. Assume that firing times ti are generated according to the following prescription: ti = i(cid:104)t(cid:105) + σξ(i), where (cid:104)t(cid:105) is the period of the oscillation and σ > 0 is a parameter that determines how strong the independent standard normal Gaussian random variables ξ(i) influence the dynamics. We show in Appendix B that the SCC between adjacent ISIs in 2. To explain this this case is negative: ρ(k = 1) = − 1 result, note that when an ISI larger than the mean (cid:104)t(cid:105) is generated, we must have a large value of ξ. Because the random variables ξ(i) are independent, it is unlikely that the next firing time ti+1 will be larger than the mean again; instead, it will more likely lie around the mean, therefore introducing negative correlations between ad- jacent ISIs in the spike train. In real networks of spiking neurons, the periodic oscillation will furthermore modu- late the membrane potential, and thus spiking will occur preferentially near the phases corresponding to the max- imal values of the network oscillation. A different mechanism occurs in sparse networks of LIF neurons. In these networks, a global oscillation of the network activity A(t) develops as the bias current I0 is increased [11]. The frequency of this oscillation is ap- proximately independent of I0 and depends mainly on the synaptic delay D [11]. In Fig. 1, we show that the mean network SCC turns negative at the onset of a coherent network oscillation with frequency f ≈ 200 Hz (vertical red dashed line in plots for the PSD in Fig. 1 A-C); parameters can be ad- justed to produce much slower rhythms as well. The av- erage single-neuron spectrum (cid:104)Sxx(f )(cid:105) has a larger peak (green dashed line) at the inverse of the mean network ISI (approximately 61 Hz), and another smaller peak at the network frequency f. This entails that an individual neuron in the network does not participate in every up- state of the network- instead, it skips cycles of the global oscillation. As the network oscillates more synchronously with increasing I0, the first peak in the average single- neuron power spectrum decreases. A state of high synchrony, in which each neuron has a higher probability of firing per cycle, i.e. skipping is nearly absent, is seen for larger I0 (Fig. 1 C) and the mean network SCC moves back towards zero. Hence we see a non-monotonic change in the mean network SCC as a function of I0. At the onset of negative mean net- work SCCs (Fig. 1 C), we have verified that the neuron- averaged SCC at lag 2 (Eq. 8 for k = 2) is slightly positive (≈ 0.07) and at lag 3, it is slightly negative (≈ −0.017), which is consistent with the underlying network oscilla- tion causing the negative mean network SCC. A. SCC in networks with current-based synapses We restrict our numerical simulations to the suprathresh- old regime and intermediate values of I0 and N. Increas- ing network size or bias current past these values can re- sult, respectively, in unrealistically synchronous network states (for which the SCC is either zero or positive) or strong inhibition that drives neurons far below their reset potential xr, which is biologically not plausible. 4 FIG. 1. Activity in networks with current-based synapses and SCC distributions at the onset of negative mean network SCCs. Network activity (spike rastergram, network activity A(t), power spectral density of network activity and aver- age single-neuron spectrum (cid:104)Sxx(f )(cid:105)), SCC distribution and mean ISI distribution for N = 500 for three different values of I0: panels A, B, C: I0 = 40, 50, 60 mV respectively, which corresponds to the location of the three white stars in Fig. 2. The mean network SCC is maximally negative at I0 = 50 mV, i.e. when power at the network frequency (hor- izontal red dashed lines in plots for PSD are at a fixed value of 102 for comparison) increases. Remaining parameter val- ues are as in Fig. 2. The green dashed lines show the mean network ISI (or its inverse in the average single-neuron power spectrum). No refractory period. 636063806400t[ms]0500neuron#A636063806400t[ms]0100A[Hz]200400600f[Hz]100PSD[1/s]200400600f[Hz]101hSxx(f)i[1/s]−0.15−0.10−0.050.00ρi(k=1)050neuronnumber16.216.316.4E(cid:0)Ti(cid:1)[ms]050neuronnumber5710572057305740t[ms]0500neuron#B5710572057305740t[ms]0100A[Hz]200400600f[Hz]100PSD[1/s]200400600f[Hz]101103hSxx(f)i[1/s]−0.30−0.25−0.20ρi(k=1)050neuronnumber11.5011.5511.60E(cid:0)Ti(cid:1)[ms]050neuronnumber10520105301054010550t[ms]0500neuron#C10520105301054010550t[ms]0200A[Hz]2004006008001000f[Hz]101PSD[1/s]2004006008001000f[Hz]101104hSxx(f)i[1/s]−0.10−0.050.000.050.10ρi(k=1)050neuronnumber8.888.908.928.94E(cid:0)Ti(cid:1)[ms]050neuronnumber 1. Onset of negative mean network SCCs In Fig. 2, we show the dependence of the mean network SCC and ISI on network size N and bias current drive I0. As N increases for an intermediate value of I0, the mean network SCC goes from around zero to negative values and then back to zero (Fig. 2 A). In contrast, the mean network ISI (Fig. 2 B) increases (decreases) mono- tonically with N (I0). Standard deviations of the two quantities across neurons (Eq. 10) remain small (Fig. 2A, B right). Lineouts of Fig. 2 along the white and black dashed lines, together with control simulations with a dif- ferent detailed connectivity, are shown in Fig. 3, where it is also shown that the mean network SCC behaves non- monotonically as a function of I0, too (cf. Fig. 1). FIG. 2. Temporal correlation-decorrelation transitions as a function of network size N and external bias I0 for the C-fixed scenario in networks with current-based synapses. Mean net- work SCC (Eq. 9) (A, left) and ISI (Eq. 11) (B, left) together with standard deviations of SCC (A, right) and of the mean ISI distribution across neurons (B, right) in the C-fixed con- nectivity scenario. The number of synapses onto each neuron is fixed at C = p(N − 1) with p = 0.2. d = 200 s simulation time. No refractory period. In summary, we observed a transition to negative mean network SCCs as the bias drive I0 is increased. In Fig. 4 A, we plot the value of the PSD at its maximum (which is attained at the frequency f of the network os- cillation) as a function of I0, together with the frequency f of the network oscillation and the mean network SCC. This plot corresponds to the white vertical line in Fig. 2 A. The maximal value of the PSD serves as a measure of coherence for the network oscillation. We see that as the mean network SCC decreases towards negative values, 5 FIG. 3. Two types of temporal correlation-decorrelation tran- sitions are present for the C-fixed case in networks with current-based synapses. Lineout of Fig. 2 A along the white dashed vertical line (A, orange squares) and along the black dashed horizontal line (B, orange squares). Small green cir- cles in A are for a control simulation with a different random seed and hence for a different network connectivity. Also, the duration of the control simulation was increased to d = 300 s in contrast to Fig. 2, where d = 200 s. The green dashed lines are the minimum and the maximum of the SCC distribution for the control simulation. The blue squares are for smaller values of N = 10 (panel A) or I0 = 25 mV (panel B) before the transition to negative mean network SCCs. the coherence of the network oscillation increases. The frequency f of the global network oscillation increases only slightly with I0. In Fig. 4 B (which corresponds to the black horizontal line in Fig. 2 A), we show analogous results for the PSD maximum as a function of N for fixed I0. For larger network sizes, the network frequency f de- creases monotonically with N, whereas the peak value of the PSD increases. Again, a minimum of the mean network SCC (observed at N = 500) is accompanied by an intermediate value of the maximum of the PSD of the network activity. To determine the microscopic properties of the spike trains at the transition to negative SCCs, we plot con- secutive ISIs in Fig. 5 for three different values of I0, corresponding to the white stars in Fig. 2 A. The slope of a linear regression function in these plots then is the SCC at lag 1 [2], which is close to zero for Fig. 5 A and C, and negative in panel B. We see in Fig. 5 A that all ISIs scatter around the mean value and the SCC stays close to zero. In Fig. 5 B, short ISIs tend to be followed by long ones, and vice versa, so that the SCC becomes negative. We checked that the slope is not negative due to the three large outlier ISIs. In Fig. 5 C, most ISIs cluster again around the mean value, but there are also smaller clusters reflecting the strongly driven state of the system in which more complex firing patterns, reflecting higher-order mode locking, can occur. As such, the mechanism for the onset of negative mean network SCCs is similar to the onset of negative ISI cor- relations in single perfect IF neuron models with another form of negative activity-dependent feedback, namely a spike-triggered adaptation current with a single exponen- 30405060I0[mV]−0.2−0.10.0ρ(k=1)AN=10N=50010050010001500N−0.3−0.2−0.10.0ρ(k=1)BI0=25mVI0=50mV 6 FIG. 4. Maximum of PSD of network activity, mean network SCC and network frequency f as a function of I0 for N = 500 (panel A) and as a function of N for I0 = 50 mV (panel B). This corresponds to the vertical white line and to the horizontal black line in Fig. 2. Thus in both cases (A) and (B), the transition correlates with the onset of increase in power of the network oscillation. d = 200 s simulation time. tial time scale [28, 29]. In these models, the single-neuron SCC becomes maximally negative at an intermediate value of the ratio between the timescale of the adaptation current providing negative feedback and the determinis- tic mean ISI of the neuron. In our network setup, the bias drive I0 sets the time scale for single-neuron firing, whereas the global rhythm is mainly determined by net- work properties, in particular the synaptic transmission delay D. The analogy, however, is not perfect as firing of the single neuron determines the network rhythm, which in turn determines the negative feedback. We now show that the transition to negative mean net- work SCCs also occurs in networks with a more bi- FIG. 5. Plot of successive ISIs (n + 1st ISI plotted as function of nth ISI) for three different values of I0, corresponding to the three white stars in Fig. 2. A: I0 = 40 mV, A: I0 = 50 mV, C: I0 = 60 mV. The red solid vertical and horizontal lines denote the mean ISI. The dashed red line is a linear regression, whose slope is the SCC at lag 1, which is given by -0.06, -0.23, -0.03 for panels A, B, C, respectively. These values are close to the mean network SCC. Other parameter values, except duration of simulation (d = 50 s), are as in Fig. 2. ologically realistic fluctuating number of synapses (P - fixed case, Fig. 6 A). Here, in contrast with the C-fixed scenario, the standard deviation of the SCC computed 304050600.00.5max[1/s]×107A30405060−0.20.0ρ(k=1)30405060I0[mV]150200250f[Hz]50010001500200001000020000max[1/s]B500100015002000−0.2−0.10.0ρ(k=1)500100015002000N150200250f[Hz]12.515.017.520.022.525.0T(n)[ms]12.515.017.520.022.525.0T(n+1)[ms]A10121416T(n)[ms]10121416T(n+1)[ms]B681012T(n)[ms]681012T(n+1)[ms]C across neurons increases at the onset of negative mean network SCCs (Fig. 6 B), which is also shown in the lineouts in Fig. 7 C and D. Overall, the temporal cor- relations are slightly less negative than in the C-fixed scenario shown in Fig. 2. The mean network ISI is gen- erally larger than in the C-fixed case, i.e. the average neuron is subject to more inhibition than in the C-fixed case, which at the same time slightly reduces the mean network SCC for fixed values of N and I0. 7 FIG. 7. Correlation-decorrelation transitions for current- based networks in the P -fixed scenario. The mean network SCC (panels A, B) deviates from zero as the standard devia- tion of the SCC across neurons (panels C, D) increases. Line- out of Fig. 6 A along the white dashed vertical line (A, orange line) and along the black dashed horizontal line (B, orange line). The green symbols are a control simulation started with a different random seed compared to the simulation shown in orange, and with different network realizations for each value of I0. The green dashed lines are the minimum and the max- imum of the SCC distributions for this simulation. The ma- genta diamonds are another control simulation with exactly the same network connectivity for each value of I0. Hence, the rugged dependence of the mean network SCC for N = 500 is not an artifact of finite simulation time or the setup of the network connectivity. C,D: Ensemble standard deviations of the SCC distribution (lineouts of Fig. 6 A right panel along the white and black dashed lines). The blue squares are for a smaller value of N (panels A and C) or I0 (panels C and D) before the transition to negative mean network SCCs. B. SCC in networks with conductance-based synapses The mechanism for the generation of negative temporal correlations relies on the onset of a coherent network os- cillation in a simplified model of neural dynamics. In par- ticular, the current-based synapses are instantaneous and do not follow biologically realistic dynamics. We there- fore investigated whether transitions to negative mean network SCCs could also be observed in more biologically plausible neuronal networks capable of fast oscillations. To that end, we considered networks of inhibitory LIF neurons with single-neuron parameters similar to those for fast-spiking basket cells in region CA1 of hippocam- FIG. 6. A fluctuating number of synapses across neurons in networks with current-based synapses decreases the mean net- work SCC below zero for a larger range of parameters. The transition (panel A left) to negative mean network SCCs in this case is accompanied by an increase in the standard de- viation of the SCC (panels A right) across neurons. Mean network SCC (Eq. 9) (A, left) and ISI (Eq. 11) (B, left) to- gether with standard deviations of SCC (A, right) and of the mean ISI distribution across neurons (B, right) for the P -fixed connectivity scenario. The average number of synapses onto a neuron is given by C = p(N − 1) with p = 0.2. No refractory period. 2. Effect of refractory period In Fig. 8, we show that for the C-fixed scenario in the presence of an absolute refractory period, the mean net- work SCC transitions to small negative values for smaller N as I0 is increased, and to intermediate positive values for larger N. The critical value for I0 for the generation of negative mean network SCCs remains similar to the case without a refractory period, however, the transition is observed at smaller values of N. 30405060I0[mV]−0.6−0.4−0.20.0ρ(k=1)AN=10N=50010050010001500N−0.20−0.15−0.10−0.050.00ρ(k=1)BI0=25mVI0=50mV30405060I0[mV]0.000.050.100.15std(ρ)CN=10N=50010050010001500N0.000.020.040.060.080.10std(ρ)DI0=25mVI0=50mV 8 FIG. 8. Mean network SCC (A) and its standard deviation across neurons (B) in the presence of a refractory period in the C-fixed scenario. The parameter values are like in Fig. 2, but with a refractory period τr = 2 ms. pus [30]. The model is described in detail in Appendix C. The refractory period is always set to 1 ms in the following plots. Instead of a constant bias current I0, we drive every neuron with excitatory Poisson processes, whose strength (measured by their constant rate νext) de- termines the coherence of a nascent network oscillation. The frequency f of the network oscillation is largely inde- pendent of the Poisson drive and depends mainly on the synaptic parameters [31]. For a network oscillation in the fast gamma range, the mean network SCC transitions to negative values at a value of νext ≈ 60 Hz (Fig. 9), well in the suprathreshold regimes, which requires νext ≈ 13 Hz ,as determined by numerical simulation. Can transitions to negative mean network SCCs still be observed in the presence of a slower oscillation in the gamma range? We increased the synaptic rise time from inh,r = 0.45 ms in Fig. 9 to three higher values in Fig. 10, τ I which reduces the network frequency [31]. We observe transitions to negative mean network SCCs in Fig. 10, which are strongest for for τ I inh,r on the order of the re- fractory period. For τ I inh,r ≥ 3.5 ms, the network activity does not oscillate coherently anymore and therefore, the mean network SCC does not decrease as much as for smaller τ I Finally, we show that as in current-based networks, the mean network SCC also transitions to negative values as the network size is increased (Fig. 11), although there is seemingly no flat part near a mean network SCC close to zero before the transition to negative values unless it happens at low N , as in e.g. Fig. 6 B- the transition to negative mean network SCCs as a function of N occurs faster than in current-based networks. Taken together, we have shown that at the onset of a coherent global network oscillation, neural networks with both current- and conductance-based synapses generate negative ISI correlations in the output spike trains of their constituent neurons, with a parameter dependence that is generally not monotonic. inh,r. FIG. 9. Transition to negative mean network SCCs in a net- work of conductance-based IF neurons with increasing exter- nal drive. Top: Network activity at νext = 70 Hz. The network oscillation has a high frequency f ≈ 140 Hz. Bot- tom: Mean network SCC and ISI. Thin lines with crosses show minimum and maximum of the SCC distribution. C- fixed scenario with pN = 25, corresponding e.g. to p = 0.05 for N = 500. Parameter values: nP I = 800, for remaining parameters, see Appendix C. IV. NOISE-REDUCED DIFFUSION APPROXIMATION Which features of the statistics of the input spike train (cf. Eq. 2) to a given neuron causes its output spike to show non-renewal statistics? One way to address this question is to consider the sem- inal diffusion approximation (DA). In the DA for purely inhibitory networks, the synaptic current fed into one av- erage "typical" neuron in the network is given by Eq. 3. Assuming that the coupling between neurons is sparse, pairwise correlations between neurons can be neglected. This entails that the statistics of one neuron are repre- 10001020104010600500neuron#1000102010401060t[ms]01000A[Hz]406080100−0.4−0.20.0ρ(k=1)N=500N=1000N=2000406080100νext[Hz]10203040hTi[ms] 9 FIG. 11. Transition to negative mean network SCCs in net- work of conductance-based IF neurons with increasing net- work size N. C-fixed scenario with p = 0.05. Thin lines with crosses show minimum and maximum of the SCC distribution. For N = 500, the network oscillation has the same frequency as in Fig. 9 at the corresponding values of νext. Parameter values: nP I = 800, for remaining parameters, see Appendix C. to self-consistently determine spike train power spectra in networks of EI neurons using an iterative numerical algorithm in the C-fixed scenario. This self-consistent approach does not apply, however, if the connectivity is not sparse or the network activity is synchronous. A. Sensitivity of SCC to the number of input spikes First, we determined how sensitive the SCC of a neu- ron embedded in a network is to changes in the size of its pre-synaptic neighborhood (Fig. 12). We recorded all spike trains present in the corresponding pre-synaptic neighborhood and then ran an offline simulation in which the recorded inhibitory spikes were manually fed into the simulation. As the size of the pre-synaptic neighborhood increases, the mean ISI (middle panel of Fig. 12) increases monotonically, in contrast to the SCC, which surprisingly shows non-monotonic behavior with increasing size of the presynaptic neighborhood until it reaches its online value when the size of the presynaptic neighborhood reaches its online value given by C = p(N − 1). Thus, we have verified that feeding back all pre-synaptic spike trains to a chosen neuron in an 'offline' simulation of that neuron reproduced the output statistics obtained during the full network simulation. The behavior of the FIG. 10. Transition to negative mean network SCCs in net- work of conductance-based IF neurons with increasing ex- ternal drive. Top: Network activity at νext = 70 Hz for inh,r = 2.0 ms. The network oscillation has a lower fre- τ I quency f ≈ 80 Hz. Bottom: Mean network SCC and ISI. Thin lines with crosses show minimum and maximum of the SCC distribution. N = 500. C-fixed scenario with p = 0.05. Parameter values: nP I = 800, for remaining parameters, see Appendix C. sentative of the whole neuronal network. With some ad- ditional technical assumptions, a mean-field type descrip- tion of the dynamics is possible [11], because it is possi- ble to write a one-dimensional time-dependent Fokker -- Planck equation (FPE) for the evolution of the membrane potential of one single cell. The formulation of the FPE together with its boundary conditions is self-consistent, because the solution of the FPE gives the time-dependent firing rate ν(t) of the neuron; this quantity also appears in the FPE via the time-dependent drift and diffusion coefficients Eqs. 4 and 5, respectively. If the network is sparse, this approach means that the output spike train of any given neuron in the network and its input spike train share the same statistics. This was used in [10] 10001020104010600500neuron#1000102010401060t[ms]01000A[Hz]406080100−0.4−0.20.0ρ(k=1)τIinh,r=3.5msτIinh,r=2.0msτIinh,r=1.0ms406080100νext[Hz]255075hTi[ms]500100015002000−0.20.0ρ(k=1)νext=30Hzνext=70Hzνext=80Hz500100015002000N050100hTi[ms] SCC with increasing pre-synaptic neighborhood size is non-monotonic, and nearly all input spikes (bottom panel of Fig. 12) are needed to reproduce the online value. We have checked that the behavior of the SCC does not de- pend strongly on which neuron in the network is chosen, which is a result of the low standard deviation of the SCC distribution across neurons for the C-fixed case (cf. Fig. 2 A right panel). 10 also checked that the input statistics did not depend on the chosen neuron in the network, as expected for the C- fixed case. In addition to the mean ISI and the SCC, we also computed the coefficient of variation (CV) of the in- put spike sequence, defined as the ratio between standard deviation and mean of the pooled ISI sequence. FIG. 12. The SCC (top panel) of one neuron embedded in the network is very sensitive to the number of presynaptic spikes (bottom panel), in contrast to the monotonically increasing mean ISI (middle panel). The maximal value of C is 140, and the abscissa is swept from left to right by letting the pres-synaptic neighborhood size increase from 0 to 140. The dashed horizontal lines are at an SCC value of zero (upper line) and the online SCC value for the chosen neuron (lower line). The dashed horizontal line in the plot for the mean ISI is the online value of the mean ISI for the chosen neuron. The different lines (hardly distinguishable, because they largely overlap) are for 10 random permutations of the order of the presynaptic spike trains to ensure that the non-monotonic behavior of the SCC is not a result of a particular arrangement of the order of pre-synaptic spike trains. Parameter values: N = 700, p = 0.2 (C-fixed scenario). I0 = 50 mV, d = 100 s simulation time. Other parameters as in Fig. 2. B. Statistics of the input spike trains impinging on a given neuron So far, we have not yet quantified the statistics of the input spike trains. However, for the DA to be applica- ble, the network has to be in an asynchronous regime, in which neurons discharge according to a Poisson process with time-dependent rate ν(t). We show the statistics of the pre-synaptic input to one neuron in a network in Fig. 13 as a function of the size of its presynaptic neigh- borhood C, which is obtained by the superposition of a C , 1] of actual pre-synaptic spike trains im- fraction  ∈ [ 1 pinging on the chosen neuron. These pre-synaptic spike trains were recorded in a full network simulation. We FIG. 13. Statistics of pre-synaptic input of one fixed neuron in a network of current-based neurons without refractory period. From top to bottom: input mean ISI, SCC and CV.  varies C and 1 from left to right. A: small . B: large . between 1 Parameters: N = 500, other parameter values as in Fig. 2, in particular, C = p(N − 1) = 100. 020406080100120140−0.250.00ρ0204060801001201407.510.012.5hTi[ms]020406080100120140sizeofpresynapticneighborhood01000000nspikespre05101501020hTini[ms]AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAI0=21mVI0=40mVI0=50mVI0=60mV051015−1.0−0.50.0ρin051015C012CVin607080901000.20.4hTini[ms]BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBI0=21mVI0=40mVI0=50mVI0=60mV607080901000.00.2ρin60708090100C123CVin C , i.e. when only one input spike Note that for  = 1 train is considered, we recover the results for the out- put statistics of the neuron. The mean input ISI (top panel in Fig. 13) decreased with increasing number of pre-synaptic spike trains approximately like 1 C , as ex- pected in the Poisson limit. For small suprathreshold I0 = 21 mV (purple circles in Fig. 13), the input CV (bottom panel) is close to 1 and the input SCC (mid- dle panel) is close to zero, and thus, the network is in a regime where the input spike trains follow Poisson statis- tics. Increasing I0 to larger suprathreshold values (small red circles, green crosses and blue diamonds in Fig. 13, the same values as the three white stars in Fig. 2) results in positive input SCCs and input CVs exceeding 1, so that the input statistics are then manifestly non-Poissonian. C. Noise-reduced diffusion approximation in an offline simulation In light of these findings, we now explore whether a DA can still be used to compute both the mean network SCC and ISI. We first describe an open-loop or offline simula- tion scheme to address this question; it uses an adjustable form of the DA for the simulation of one effective neu- ron in the network. We then compare the results for mean network SCC and ISI obtained during an online full network simulation and the offline effective single-neuron simulation. The algorithm to compare online and offline simulations results is as follows: 1. Simulate full network online, and obtain the network activity A(t) (smoothed with rectangular kernel of width σw = 0.01 ms). 2. Simulate one single offline neuron (cf. Eq. 1) accord- ing to the Langevin equation (12) X =γt (−X + I0 − CJν(t − D)τt) + γt(cid:113)αJ J 2Cν(t − D)τt + σ2 0√τtξ(t) , with a timestep h = 10−2 ms using the DA with ν(t) = A(t) in Eq. 3 as drive. We introduced a factor αJ that scales the contribution of the network activity ν to the noise strength in the DA. As for the full network simula- tions, the Euler-Maruyama method is used for the inte- gration of the stochastic differential equation 12. 3. Compute mean SCC at lag 1 and its standard devia- tion for the network. 4. Compute SCC at lag 1 for the chosen offline single neuron. 5. Compare the SCC result from online network simula- tion and offline single-neuron simulation. Another way for estimating the mean network SCC is to simulate an inhomogeneous Poisson process (iPP) [32, 33] with the same time-varying smoothed rate A(t) as the full network and then compute the mean ISI and the SCC for 11 the generated spike train. We simulated the iPP using the numerical scheme proposed in [34]. The idea is that this simple surrogate for the network activity, which has temporal correlations solely due to its time-varying rate (i.e. there are no intrinsic correlations), can give insight into the behavior of the SCC we observe below. We will now only consider the C-fixed scenario, since the DA should then reproduce the behavior of one typical neuron accurately. It would also be possible to compute the mean network SCC in the P -fixed scenario by aver- aging SCC results over the (binomial) distribution of the C-value, a computationally expensive task. We verified that the choice of integration timestep h of the single neuron dynamics as well as the smoothing width σw do not impact the following results (see Fig. 16 below). In Fig. 14, we show the performance of the DA and iPP approximation schemes for both the mean network SCC (top panels) and ISI (bottom panels) as I0 is increased (cf. the vertical white dashed line in Fig. 2). For small I0 before the transition to negative mean network SCCs, both approximation schemes agree well with the full net- work simulations. At the onset of negative mean network SCCs, both the DA and the iPP approximation schemes fail to reproduce the negative mean network SCC of the full online simulation, although the DA shows a slight decrease of the SCC. The mean network ISI, in contrast, is always approximated well by both schemes. To understand the origin of the mismatch between the full network simulations and the DA at the onset of non- vanishing mean network ISI correlations, we computed the CVs of the spike train generated by the DA and com- pared it to the mean network CV (the CV averaged over all neurons in the network). We found that as I0 exceeded approximately 40 mV, the CV obtained by the DA was slightly larger than the mean network CV. Therefore, we decreased the contribution of the network activity to the noise in the DA by setting αJ = 0 (see Eq. 3). In Fig. 15, we show that this simple adjustment leads to good agree- ment between DA and full network simulations for both the mean network ISI and SCC. D. Effect of smoothing width These results do not depend on the width of the smooth- ing kernel σw as long as it is chosen small enough (Fig. 16). Also, αJ must be chosen as small as possi- ble for the noise-reduced DA to perform well. It is not necessary to scale down the noise in the particu- lar way we have pursued so far with Eq. 12 to improve the accuracy of the DA. We show in Fig. 17 that a re-scaling of the noise term in the DA with a factor αg = 0.66 according to X =γt (−X + I0 − CJν(t − D)τt) + αgγt(cid:113)J 2Cν(t − D)τt + σ2 0√τtξ(t) , (13) 12 FIG. 14. Diffusion approximation for the mean network SCC (top of each panel) and ISI (bottom of each panel) as a func- tion of I0. A: N = 500. B: N = 1000. Red circles: full network simulation (FN). Green squares: single neuron offline simulation using the DA Eq. 12 with αJ = 1.0. Errorbars for the full network simulation indicate ± one standard deviation of the SCC distribution. Parameter values: C-fixed scenario with C = p(N − 1). Simulation time d = 200 s. FIG. 15. Noise-reduced diffusion approximation for the mean network SCC (top) and ISI (bottom). A: N = 500. B: N = 1000. Red circles: full network simulation (FN). Blue diamonds: iPP approximation (iPP). Green squares: single neuron offline simulation using the DA Eq. 12 with αJ = 0. Errorbars for the full network simulation indicate ± one stan- dard deviation of the SCC. The thin red lines indicate the minimum and maximum of the online SCC distribution for each value of I0. Parameter values as in Fig. 14. has a similar effect on the accuracy of the DA as reducing the time-varying factor in the noise strength in Eq. 12. However, this version of the DA performs slightly worse for the mean network ISI at small I0, where it also under- estimates the mean network CV (not shown). We found this particular value of αg by systematic numerical sim- ulations, gradually decreasing αg from 1 (equivalent to Eq. 12 with αJ = 1). This confirms that the spiking dy- namics of the full network generates input spike trains that are less stochastic than one would expect from a Poisson process with time-varying intensity. Finally, we show in Fig. 18 that the noise-reduced DA remains valid for a larger range of connection probabil- ities p, where we also observe a transition to negative mean network SCCs. Importantly, our standard value of p = 0.2 is in the regime where the DA accurately re- produces the mean network ISI. Increasing p for fixed N is, in the C-fixed scenario, equivalent to an increase of N for fixed p, for which we have already shown that the mean network SCC decreases to negative values for large enough I0 (black horizontal dashed line in Fig. 2). In summary, our explorations of three forms of the DA (two of which are effectively noise-reduced) using offline 30405060−0.3−0.2−0.10.0ρ(k=1)ADAiPPFN30405060I0[mV]2040hTi[ms]30405060−0.4−0.20.0ρ(k=1)BDAiPPFN30405060I0[mV]204060hTi[ms]30405060−0.3−0.2−0.10.0ρ(k=1)ADAiPPFN30405060I0[mV]2040hTi[ms]30405060−0.4−0.20.0ρ(k=1)BDAiPPFN30405060I0[mV]20406080hTi[ms] 13 FIG. 16. Effect of smoothing width σw and αJ on the perfor- mance of the DA (Eq. 12). Solid lines: single neuron offline simulation using the DA Eq. 12 with three values of αJ as indicated in the legend. The black dashed lines are results from the online full network simulation, which do not depend on αJ or σw. Parameter values (analoguous to Fig. 15 A for I0 = 50 mV): N = 500, C = p(N − 1) fixed, p = 0.2, I0 = 50 mV. Simulation length d = 300 s. simulations show that it is possible to obtain conditions where a DA with the proper smoothed spike input A(t) produces correct first and second order ISI statistics. This will help constrain the search for a self-consistent effective neuron model that can produce such statistics. V. DISCUSSION We have studied ISI correlations in spike trains of neu- rons embedded in purely inhibitory networks with both current- and conductance-based synapses. Our results re- veal that these simple networks in general do not generate renewal dynamics in their spike trains. This is not sur- prising given their propensity to oscillate (Fig. 1). How- ever, the manner in which serial ISI correlations appear and their non-monotonic behavior (which implies max- imal ISI correlation at intermediate parameter values) and link to SCC variance in the P -fixed case (Fig. 6) are new features revealed by our analysis. In particular, strong negative serial correlations are ob- served at an intermediate value for the bias drive and network size in networks without and with a refractory period (Figs. 2 - 8). We have also shown that the mean network SCC is maximally negative at an intermediate value of the network oscillation strength (as quantified by the peak value of the PSD of the network activity A(t), Fig. 4) as it decreases below zero with increasing bias current or system size. In conductance-based networks, negative ISI correlations FIG. 17. Noise-reduced DA (Eq. 13 with αg = 0.66). Dif- fusion approximation for the mean network SCC (top) and ISI (bottom). A: N = 500. B: N = 1000. Red circles: full network simulation (FN). Blue diamonds: iPP approximation (iPP). Green squares: single neuron offline simulation using the DA Eq. 13 with αg = 0.66. Errorbars for the full network simulation indicate ± one standard deviation of the SCC. The thin red lines indicate the minimum and maximum of the on- line SCC distribution for each value of I0. Parameter values as in Fig. 14. also arise generically with increasing bias drive for both slow and fast gamma network oscillations (Figs. 9 - 11). Importantly, our results highlight the importance of nearly all input spikes for shaping the mean network SCC (Fig. 12). Even if the input statistics to a given neuron are non-Poissonian (Fig. 13), the onset of neg- ative temporal correlations can be approximately quan- tified by an effectively noise-reduced diffusion approxi- mation (Eq. 12), showing that the network dynamically suppresses noise at increasing values of the bias current 10−210−1100101σw[ms]−0.25−0.20−0.15−0.10−0.050.00ρ(k=1)αJ=0αJ=0.1αJ=130405060−0.2−0.10.0ρ(k=1)ADAiPPFN30405060I0[mV]2040hTi[ms]30405060−0.20.0ρ(k=1)BDAiPPFN30405060I0[mV]255075hTi[ms] 14 tive ISI correlations arise from slowly drifting firing rates caused by heterogeneous innervation of more than one haircell synapse, whereas negative ISI correlations arise from synaptic depletion alone. It would be interesting to extend our results to these two scenarios, which would require the inclusion of non-stationary rate dynamics and synaptic plasticity into our model. We would also expect transitions to negative mean net- work SCCs in networks of Hodgkin-Huxley or IF neu- rons with adaptation, as long as noisy oscillatory states are present. This is because the detailed voltage dynam- ics are only important insofar as they interact with the global rhythm, i.e. the form of the subthreshold voltage time course is not crucial for the onset of negative mean network SCCs [37]. Concerning a possible full analytical understanding of SCC transitions, the existence of ISI correlations points to stochastic dynamics in dimension greater than one, and an analytic explanation of our results is beyond the scope of our study. Recent first-passage time results for a two-dimensional Wiener process [38] nevertheless give hope that the present numerical approach will eventually lead to an analytical description. It is also an interesting subject for future study to de- termine when non-vanishing ISI correlations exist also in networks with both excitatory and inhibitory subpopula- tions. Furthermore, how our results relate to transitions in networks with these two types of neurons observed by increasing coupling strength [39, 40] and whether they can be computed using the framework of [5] or the frame- work of doubly stochastic point processes [41] remain open questions. VI. ACKNOWLEDGMENTS FIG. 18. DA for the mean network ISI and SCC is valid even for large p. DA for the mean network SCC (top) and ISI (bot- tom). A: N = 500. B: N = 1000. Red circles: full network simulation (FN). Blue diamonds: iPP approximation (iPP). Green squares: single neuron offline simulation using the DA Eq. 12 with αJ = 0. Error bars for the full network simula- tion indicate ± one standard deviation of the SCC. Parameter values: C = p(N − 1) fixed, I0 = 50 mV. Simulation time d = 300 s. We would like to thank Raoul-Martin Memmesheimer, Sven Goedeke, Alexandre Payeur and Richard Naud for interesting discussions and helpful comments. We would like to thank NSERC Canada (grant number 121891-2013) for funding this work and the German Fed- eral Ministry of Education and Research (BMBF) for support via the Bernstein Network (Bernstein Award 2014, 01GQ1501 and 01GQ1710). W.B. thanks Denisa Dervishi and Karl Braun for hospitality. drive. Whereas this decrease of the noise intensity does not have a strong effect on the validity of the DA for the mean network ISI, we have shown that it has a pro- nounced effect for the mean network SCC (Figs. 14 - 18). Overall, our results constitute a first step to understand how pre-synaptic spiking shapes the statistics of post- synaptic spiking in neural networks (Figs. 12- 13), a ques- tion that has recently received increased interest in the biological literature, with positive ISI correlations ob- served in the zebrafish lateral line system [35] and nega- tive correlations observed in the auditory system afferent neurons [36]. In these systems, it is thought that posi- Appendix A: Details for network simulation 1. Simulation details We implemented all simulations in brian2 [42]. The inte- gration time step was set to h = 10−2 ms if not mentioned otherwise and a standard Euler-Maruyama scheme was used. Except in Fig. 8, the refractory period is set to 0 ms for the current-based synapse model. To ensure that we are in a stationary regime, we discard the first 1000 ISIs both for the computation of the mean network SCC 0.00.20.4−0.3−0.2−0.10.0ρ(k=1)ADAiPPFN0.00.20.4p5101520hTi[ms]0.00.20.4−0.3−0.2−0.10.0ρ(k=1)BDAiPPFN0.00.20.4p102030hTi[ms] and ISI. The statistics of the SCC and ISI are computed across long realizations (at least 100 s) of the network ac- tivity. Increasing the duration did not change the results presented here. Note that the network activity of interest will sometimes consist in a rhythm, in which case the ISI distribution, and the SCC, vary across the duration of the period of the rhythm. In this case, the averages are taken over many cycles of the rhythm to obtain a mean and standard deviation of the ISI and the SCC during the rhythm. 2. Connectivity details For the P -fixed case, we use the standard syntax of brian2 in the Synapses.connect() object. We wire up C-fixed networks by first making a masterlist of all the neurons, each labeled by its number between 1 and N. To choose which neurons are connected to e.g. the neu- ron with label k, this neuron k is deleted from the list. Then the remaining list is randomly permuted. Finally, the first C neurons from that list are chosen as the pre- synaptic neurons to neuron k. The procedure is repeated for every neuron in the masterlist. 3. Spectral measures A smoothed version A(t) of the population activity is computed using the function smooth_rate() of the PopulationRateMonitor class in brian2. This is the in- stantaneous firing rate of the network. If not mentioned otherwise, A(t) is obtained from the pooled spike trains of all neurons by convolution with a Gaussian rectan- gular or Gaussian kernel of width σw. For illustration purposes, we choose σw = 1 ms; for computations in- volving the diffusion approximation, we choose a smaller σw = 0.01 ms or σw = 0.1 ms. We stress that the choice of σw does not influence our results as long as it is small enough to fully capture the dynamics of the population activity A(t), i.e. it is important not to smoothen out too much the fluctuations present in A(t). For the computation of spectra (both the PSD of the net- work activity as well as the averaged single-neuron spec- trum), we used two functions provided in [43]. For the average single-neuron spectrum, only 100 neurons were included in the average. For both types of spectra, the mean activity before taking the Fourier transform was subtracted and the first 1000 ms of the simulation period were discarded to obtain stationary time series. The code can be found in the github repository https://github. com/EPFL-LCN/neuronaldynamics-exercises. For the PSD of the network activity, the frequency res- olution was set to 2 Hz and the time series for ν was split into 6 parts that were Fourier-transformed inde- pendently and then averaged to obtain the PSD. Sim- 15 ilarly, the averaged single-neuron spectrum (cid:104)Sxx(f )(cid:105) was computed with a sampling rate of 6000 Hz. It was checked that the functions scipy.signal.periodogram and scipy.signal.welch gave the same results for both types of spectra. Appendix B: SCC for a noisy periodic oscillation Consider firing times ti, i ≥ 1 generated according to ti = ni(cid:104)t(cid:105) + ξi (cid:104)ξi(cid:105) = 0,(cid:104)ξiξj(cid:105) = σ2δij , (B1) where ni = 1, 2, 3, ... and (cid:104)t(cid:105) is the period of firing around which the spike times are perturbed. The ith ISI is given by Ti = ti − ti−1 . (B2) It follows that (cid:104)Ti(cid:105) = (cid:104)t(cid:105). Similarly, we can show that E (Ti+1Ti) = (cid:104)t(cid:105)2 − σ2. Moreover, we have Var(Ti) = (cid:104)T 2 i (cid:105) − (cid:104)Ti(cid:105)2 = (cid:104)t(cid:105)2 + 2σ2 − (cid:104)t(cid:105)2 . (B3) Hence, the SCC is given by ρ(k = 1) = (cid:104)t(cid:105)2 − σ2 − (cid:104)t(cid:105)2 (cid:104)t(cid:105)2 + 2σ2 − (cid:104)t(cid:105)2 = − 1 2 . (B4) This result can be generalized to the case when ni itself is a random number drawn from a certain (e.g. binomial) distribution. Appendix C: Inhibitory integrate-and-fire neuron model with conductance-based synapses We consider a model with parameters similar to [30]. The parameters are given by CI = 100 pF, gI l = 10 nS (so that rest = −65 mV, the membrane time constant is 10 ms), EI thresh = −52 mV, inh = −75 mV. When Xi(t) exceeds X I EI reset = −67 mV, followed by a refractory it is reset to X I ref = 1 ms. In the absence of synap- period of duration τ I tic inputs, the rheobase is I0 = 0.13 nA. In our simula- tions, we choose I0 = 0. The only excitatory input to the population is delivered via excitatory Poisson processes. Each neuron receives input from nP I independent Poisson processes, each with frequency f P I . Each of these random inputs induces instantaneous jumps of the membrane po- tential Xi by an amplitude κP The latency between a presynaptic spike and the start of the increase of gI inh is τl = 1 ms. The remaining param- eters are given by gI inh,d = 1.2 ms and inh,r = 0.45 ms. τ I inh,peak = 5.0 nS, τ I I = 0.1 mV. 16 Biol. 10, 1 (2014). 052127 (2017). Rev. E 66 (2002). (2015). [29] T. Schwalger, The statistics of non-renewal neuron models, Ph.D. thesis, Humboldt- Universität zu Berlin (2013). interspike-interval [30] J. R. Donoso, D. Schmitz, N. Maier, and R. Kempter, J. Neurosci. 38, 3124 (2018). [31] N. Brunel and X.-J. Wang, J. Neurophysiol. 90, 415 (2003). [32] D. R. Cox and P. A. W. Lewis, The Statistical Analysis of Series of Events, Methuen's Monographs on Applied Probability and Statistics (John Wiley, London, 1966). [33] P. A. W. Lewis and G. S. Shedler, Naval Research Logis- tics Quarterly 26, 403 (1979). [34] P. J. Laub, T. Taimre, and P. K. Pollett, arXiV (2015). [35] S. Song, J. A. Lee, I. Kiselev, V. Iyengar, J. G. Trapani, and N. Tania, Sci. Rep. 8 (2018). [36] A. J. Peterson, D. R. Irvine, and P. Heil, J. Neurosci. 34, 15097 (2014). 94, 4344 (2005). [37] C. Geisler, N. Brunel, and X.-J. Wang, J. Neurophysiol. [38] L. Sacerdote, M. Tamborrino, and C. Zucca, J. Comput. Appl. Math 296, 275 (2016). [39] S. Ostojic, Nat. Neurosci. 17, 594 (2014). [40] S. Wieland, D. Bernardi, T. Schwalger, and B. Lindner, Phys. Rev. E 92, 040901 (2015). [41] S. B. Lowen and M. C. Teich, J. Acoust. Soc. Am. 92, 803 (1992). [42] M. Stimberg, D. F. Goodman, V. Benichoux, and R. Brette, BMC Neuroscience 14, P38 (2013). [43] W. Gerstner, W. M. Kistler, R. Naud, and L. Panin- ski, Neuronal Dynamics, 1st ed. (Cambridge University Press, 2014). [1] G. L. Gerstein and B. Mandelbrot, Biophys. J. 4, 41 [25] M. Helias, T. Tetzlaff, and M. Diesmann, PLOS Comp. [26] W. Braun, R. Thul, and A. Longtin, Phys. Rev. E 95, [27] B. Lindner, L. Schimansky-Geier, and A. Longtin, Phys. [28] T. Schwalger and B. Lindner, Phys. Rev. E 92, 062703 [2] M. J. Chacron, B. Lindner, and A. Longtin, Phys. Rev. Lett. 92 (2004); O. Avila-Akerberg and M. J. Chacron, Exp. Brain Research 210, 353 (2011); W. H. Nesse, L. Maler, and A. Longtin, Proc. Nat. Acad. Sci. USA 107, 21973 (2010); M. J. Chacron, A. Longtin, and L. Maler, J. Neurosci. 21, 5328 (2001). [3] R. Ratnam and M. E. Nelson, J. Neurosci. 20, 6672 (1964). (2000). [4] E. Urdapilleta, Phys. Rev. E 84, 041904 (2011); T. Schwalger and B. Lindner, Front. Comput. Neurosci. 7 (2013); E. Urdapilleta, EPL 115 (2016). [5] T. Schwalger, F. Droste, and B. Lindner, Journal of Computational Neuroscience 39, 29 (2015); R. Mankin and A. Rekker, Phys. Rev. E 94, 062103 (2016). [6] J. Benda, L. Maler, and A. Longtin, J. Neurophysiol. 104, 2806 (2010); W. Braun, P. C. Matthews, and R. Thul, Physical Review E 91 (2015); E. Urdapilleta, Phys. Rev. E 83 (2011); J. Phys. A: Math. Theor. 45 (2012). [7] A. P. Nikitin, N. G. Stocks, and A. R. Bulsara, Phys. Rev. Lett. 109, 238103 (2012). [8] J. A. Reinoso, M. C. Torrent, and C. Masoller, Phys. Rev. E 94, 032218 (2016); M. Masoliver and C. Masoller, Sci. Rep. 8, 8276 (2018). [9] E. Muller, L. Buesing, J. Schemmel, and K. Meier, Neu- ral Comput. 19, 2958 (2007). [10] B. Dummer, S. Wieland, and B. Lindner, Front. Com- put. Neurosci 8, 104 (2014); R. F. O. Pena, S. Vellmer, D. Bernardi, A. C. Roque, and B. Lindner, Front. Com- put. Neurosci. 12, 9 (2018). [11] N. Brunel and V. Hakim, Neural. Comp. 11, 1621 (1999). [12] F. Farkhooi, E. Muller, and M. P. Nawrot, Phys. Rev. E 83, 050905 (2011); F. Farkhooi, A. Froese, E. Muller, R. Menzel, and M. P. Nawrot, PLOS Comp. Biol. 9, 1 (2013). [13] D. J. Mar, C. C. Chow, W. Gerstner, R. W. Adams, and J. J. Collins, Proc. Natl. Acad. Sci. U.S.A. 96, 10450 (1999). [14] G. Bi and M. Poo, J. Neurosci. 18, 10464 (1998). [15] P. J. Sjöström, G. G. Turrigiano, and S. B. Nelson, Neu- [16] R. Kempter, W. Gerstner, and J. L. van Hemmen, Phys. ron 32, 1149 (2001). Rev. E 59, 4498 (1999). [17] L. Abbott and W. G. Regehr, Nature 431, 706. [18] M. V. Tsodyks and H. Markram, Proc. Nat. Acad. Sci. USA 94, 719 (1997). [19] N. Lüdtke and M. E. Nelson, Neural Comput. 18, 2879 [20] A. B. Neiman and D. F. Russell, Phys. Rev. E 71, 061915 [21] C. Bauermeister, T. Schwalger, D. F. Russel, A. B. Neiman, and B. Linder, PLOS Comput. Biol. 9 (2013). [22] I. Ginzburg and H. Sompolinsky, Phys. Rev. E 50, 3171 [23] R. Moreno-Bote and N. Parga, Phys. Rev. Lett. 96, 028101 (2006). [24] T. Tetzlaff, S. Rotter, E. Stark, M. Abeles, A. Aert- sen, and M. Diesmann, Neural Comput. 20, 2133 (2008), pMID: 18439140. (2006). (2005). (1994).
1308.3363
1
1308
2013-08-15T11:30:56
How Chaotic is the Balanced State?
[ "q-bio.NC", "cond-mat.dis-nn", "physics.bio-ph" ]
Large sparse circuits of spiking neurons exhibit a balanced state of highly irregular activity under a wide range of conditions. It occurs likewise in sparsely connected random networks that receive excitatory external inputs and recurrent inhibition as well as in networks with mixed recurrent inhibition and excitation. Here we analytically investigate this irregular dynamics in finite networks keeping track of all individual spike times and the identities of individual neurons. For delayed, purely inhibitory interactions we show that the irregular dynamics is not chaotic but in fact stable. Moreover, we demonstrate that after long transients the dynamics converges towards periodic orbits and that every generic periodic orbit of these dynamical systems is stable. We investigate the collective irregular dynamics upon increasing the time scale of synaptic responses and upon iteratively replacing inhibitory by excitatory interactions. Whereas for small and moderate time scales as well as for few excitatory interactions, the dynamics stays stable, there is a smooth transition to chaos if the synaptic response becomes sufficiently slow (even in purely inhibitory networks) or the number of excitatory interactions becomes too large. These results indicate that chaotic and stable dynamics are equally capable of generating the irregular neuronal activity. More generally, chaos apparently is not essential for generating high irregularity of balanced activity, and we suggest that a mechanism different from chaos and stochasticity significantly contributes to irregular activity in cortical circuits.
q-bio.NC
q-bio
How Chaotic is the Balanced State? Sven Jahnke1,2, Raoul-Martin Memmesheimer1−3 and Marc Timme1,2 1Network Dynamics Group, Max-Planck-Institute for Dynamics & Self-Organization (MPIDS), 2Bernstein Center for Computational Neuroscience (BCCN), 37073 Göttingen, Germany, and 3Center for Brain Science, Faculty of Arts and Sciences Harvard University, Cambridge, MA02138, USA 3 1 0 2 g u A 5 1 ] . C N o i b - q [ 1 v 3 6 3 3 . 8 0 3 1 : v i X r a Large sparse circuits of spiking neurons exhibit a balanced state of highly irregular activity under a wide range of conditions. It occurs likewise in sparsely connected random networks that receive excitatory external inputs and recurrent inhibition as well as in networks with mixed recurrent in- hibition and excitation. Here we analytically investigate this irregular dynamics in finite networks keeping track of all individual spike times and the identities of individual neurons. For delayed, purely inhibitory interactions we show that the irregular dynamics is not chaotic but in fact stable. Moreover, we demonstrate that after long transients the dynamics converges towards periodic or- bits and that every generic periodic orbit of these dynamical systems is stable. We investigate the collective irregular dynamics upon increasing the time scale of synaptic responses and upon itera- tively replacing inhibitory by excitatory interactions. Whereas for small and moderate time scales as well as for few excitatory interactions, the dynamics stays stable, there is a smooth transition to chaos if the synaptic response becomes sufficiently slow (even in purely inhibitory networks) or the number of excitatory interactions becomes too large. These results indicate that chaotic and stable dynamics are equally capable of generating the irregular neuronal activity. More generally, chaos apparently is not essential for generating high irregularity of balanced activity, and we suggest that a mechanism different from chaos and stochasticity significantly contributes to irregular activity in cortical circuits. I. INTRODUCTION Most neurons in the brain communicate by emitting and receiving electrical pulses, called action potentials or spikes, via chemically operating synaptic connections. Local cortical circuits often exhibit spiking dynamics that is highly irregular and appears as if it were random. Such irregular activity at low neuronal firing rate is thus considered a basic "ground state". It is characterized by individual neurons that display largely fluctuating membrane potentials and highly variable inter-spike-intervals (ISIs) as well as by low correlations between the neurons (Brunel, 2000, Kumar et al., 2007, v. Vreeswijk and Sompolinsky, 1996, 1998, Vogels and Abbott, 2005). Originally, this dynamical state seemed to be in contradiction to cortical anatomy, where each neuron receives a huge number of synapses, typically 103 − 104 (Braitenberg and Schüz, 1998): One might expect that a large number of uncorrelated, or weakly correlated synaptic inputs to one neuron, given the central limit theorem, sums up to a regular total input signal with only small relative fluctuations, therefore excluding the emergence of irregular dynamics. So the finding of highly irregular activity might be surprising. This issue was resolved by the idea of a "balanced state" (v. Vreeswijk and Sompolinsky, 1996), in which excitatory (positive) and inhibitory (negative) input balances such that the average membrane potential is sub-threshold and strong fluctuations once in a while are sufficiently depolarizing to initiate a spike. The original work "Chaos in neuronal networks with balanced excitatory and inhibitory activity" (v. Vreeswijk and Sompolinsky, 1996) was an analysis of a self-consistent, highly irregular "balanced" activity for sparse random networks of binary model neurons. It was shown that the balanced state naturally and robustly occurs in large networks if the inhibitory and excitatory coupling strengths and their respective numbers of synapses appropriately scale with each other. Moreover, flipping the state of only one of the binary neurons in a large network (i.e. applying the smallest possible non-zero perturbation in such a system) leads to a supra-exponential divergence between the perturbed and the unperturbed realizations of the network dynamics, exemplifying the extremely chaotic nature of the balanced activity (v. Vreeswijk and Sompolinsky, 1996, 1998). Later analysis (Amit and Brunel, 1997, Brunel, 2000) of networks of integrate-and-fire neurons demonstrated that the mean field description of the balanced activity for these continuous-state neuron networks is very similar to that of the original binary neuron networks. These and related results (Amit and Brunel, 1997, Brunel, 2000, Timme and Wolf, 2008, Timme et al., 2002) indicate that statistically the same balanced activity persists both in networks with external excitatory inputs and recurrent inhibition only as well as in networks with equal total amounts of recurrent inhibition and recurrent excitation. These findings about the robustness of the balanced state, the original work (v. Vreeswijk and Sompolinsky, 1996, 1998) together with common intuition may suggest that highly irregular activity originates from chaotic network dynamics. This hypothesis, however, has not been systematically investigated so far. Recent research even points towards the contrary: it shows that in globally coupled networks without delay the dynamics tends to converge to stable periodic orbits if inhibition dominates (Jin, 2002). Numerical investigations of −2 0 40 80 120 160 time t 2 50 ν p B 60 V C p C 0 0.15 ν i 0 0.5 0.35 50 ν p E 60 V C p CVi 1.4 F A n o r u e n ) t ( V 1 40 20 2 40 20 2 D n o r u e n ) t ( V 1 −2 0 40 80 120 160 time t 0 0.15 ν i 0 0.5 0.35 CVi 1.4 Figure 1: Highly irregular spiking activity equally emerges from chaotic and from stable circuit dynamics. Irregular dynamics in purely inhibitorily (A-C) and inhibitorily and excitatorily (D-F) coupled random networks of identical leaky integrate-and-fire neurons (N = 400, γi ≡ 1, VΘ,i ≡ 1, τij ≡ 0.1T free , Pre(i) ≡ 80). (A,D) Spiking dynamics (A)Pj εi,j ≡ −16, Ii ≡ 4, NE = 0; (D)Pj εi,j ≡ −11, Ii ≡2.7, NE = 1000, where the NE excitatory couplings are distributed such that each neuron has the same number of excitatory inputs. The upper panel displays the spiking times (blue lines) of the first 40 neurons. The lower panel displays the membrane potential trajectory of neuron i = 1 (spikes of height ∆V = 1 added at firing times). (B,E) Histogram of mean firing rates νi. (C,F) Histogram of the coefficients of variation CVi := σi/µi; µi = (cid:10)ts i,k(cid:11)k ; i := D(cid:0)ts σ2 i,k − µi(cid:1)2Ek averaged over time. i,k+1 − ts i,k+1 − ts i weakly diluted inhibitorily coupled networks without delay even show that although the dynamics may be irregular, its Lyapunov exponent is negative (Zillmer et al., 2006). These numerical simulations also demonstrate by example that the dynamics converges to a periodic orbit after long quasi-stationary transients. Interestingly, a related article (Zillmer et al., 2009) also gives numerical evidence that there can be chaos and long transients even in networks with only inhibitory connections. In this article we show analytically in the limit of fast synaptic response, that in inhibitory networks with inhomo- geneous delay distribution and arbitrary, strongly connected topology (A network is strongly connected if there is a directed path of connections between any ordered pair of neurons.) any generic trajectory is asymptotically stable. After a (typically long) stable transient characterized by irregular activity the dynamics converges to a periodic orbit that is also stable, in agreement with the results presented in (Memmesheimer and Timme, 2006a, Timme and Wolf, 2008). In particular the transients are not chaotic in contrast to the ones occurring in purely excitatorily coupled net- works (Zumdieck et al., 2004). We show that this collective dynamics is robust upon increasing the synaptic response time from zero and upon replacing some inhibitory by excitatory interactions. Nevertheless, if the synaptic response becomes too slow or the number of excitatory interaction too large, the collective dynamics becomes chaotic via a transition where the Lyapunov exponent changes smoothly and the spiking activity stays highly irregular. Thus the irregularity equally prevails in networks with stable as well as in networks with chaotic dynamics, leaving no evidence that chaos generates the irregularity. Some analytically accessible aspects of the stable irregular dynamics have been briefly reported before (Jahnke et al., 2008a). Parts of this work have been presented in (Memmesheimer, 2007) and at a Bernstein Symposium (Jahnke et al., 2008b). II. NETWORK MODEL We consider networks of N neurons with directed couplings that interact by sending and receiving spikes. If such a directed connection exists from neuron j to neuron i, we call i postsynaptic to neuron j. We denote the set of all postsynaptic neurons of neuron j by Post(j). Neuron j is then presynaptic to neuron i, the set of all presynaptic neurons is denoted by Pre(i). The membrane potential Vi(t) of some neuron i evolves according to d dt Vi = fi(Vi) + NXj=1Xk∈Z εij δ(cid:0)t − ts jk − τij(cid:1) , (1) 3 A VΘ,1 1 V 0 C φ Θ,1 1 φ s t1k time 0 s t1k time B Θ,2V 2 V 0 D φ Θ,2 2 φ 0 ε21 s t1k s +τ t1k 21 time φ( ) 21 2 φ( ) − 21 2 s +τ t1k s +τ t1k s t1k s +τ t1k 21 time Figure 2: Relation between membrane potential and phase dynamics. (A) Membrane potential of neuron 1. At t = ts 1k the membrane potential V1(t) crosses the threshold potential VΘ,1 which leads to a reset of the potential to V1(ts 1k) = 0, and spikes are emitted. (B) Membrane potential of neuron 2 which is postsynaptic to neuron 1, i.e. 2 ∈ Post(1). The spike is received at ts 1k + τ21 and induces a jump of size ε21 in the potential. (C) Phase dynamics of neuron 1. The phase increases linearly until the phase threshold φΘ,1 is reached, then it is set to zero. (D) Phase dynamics of neuron 2. When the spike is received at t = ts 1k + τ21 it induces a phase jump: φ2(t1k + τ21) = H (2) 2 (U2(φ2((t1k + τ21)−)) + ε21). ε21 (φ2(t1k + τ21)) = U −1 where a smooth function fi specifies the internal dynamics, εij are the coupling strengths from presynaptic neurons j to i, δ(.) is the Dirac delta-distribution, τij > 0 are the delay times of the connection and ts jk denotes the time when neuron j sends the kth spikes. If not stated otherwise, in the following we consider inhibitory networks, i.e. εij ≤ 0. Spikes sent to postsynaptic neurons with different delay times from neuron j are considered as separate spikes, so if the delays are all different, e.g. if they are chosen randomly, neuron j sends Post(j) spikes at time ts jk. When neuron j reaches the threshold VΘ,j of the potential, i.e. Vj(t−) = VΘ,j, it generates spikes at t =: ts jk for some k and is reset, Vj(ts jk) = 0. The neuronal dynamics is therefore smooth except at times when events, namely sendings or receivings of spikes happen. Simultaneous sendings of spikes by one neuron are treated as one event as well as simultaneous ′(Vj) < 0 for all j and receptions of spikes sent by the same neuron. We require that the fj satisfy fj(Vj) > 0 and fj Vj ≤ VΘ,j, such that in isolation each neuron j exhibits oscillatory dynamics with a period T free . j The network dynamics can equivalently be described by a phase-like variable φj (t) ∈ (−∞, φΘ,j] satisfying at all non-event times (Mirollo and Strogatz, 1990). When the phase threshold is reached, φj (ts jk is reset, φj (ts time τij at neuron i and induces a phase change according to −) = φΘ,j, the phase jk) := 0 and a spike is generated. This spike travels to the postsynaptic neurons, arrives after a delay dφj/dt = 1 (2) with the transfer function φi(ts jk + τij) = H (i) εij(cid:16)φi(cid:16)(cid:0)ts jk + τij(cid:1)−(cid:17)(cid:17) . H (i) ε (φ) := U −1 i [Ui(φ) + ε] , (3) (4) where each Ui(t) is the free (all εij = 0) solution of (1) through the initial condition Ui(0) = 0, yielding U ′ i < 0 and φΘ,j = U −1 U ′′ dynamics and membrane potential. i > 0 and (VΘ,j), cf. (Memmesheimer and Timme, 2006a). Fig. 2 illustrates the relation between phase j The analysis below is valid for general Ui(φ); in the numerical simulations we employ leaky integrate-and-fire i Ii > VΘ,i, the membrane potential i I i(1 − exp(−γiφ)) on φ and the oscillation period of a free neuron is given neurons, fi(V ) := Ii − γiV with time scale γ−1 has the functional dependence Ui(φ) = γ−1 i > 0 and equilibrium potential γ−1 := γ−1 i i by T free ln(Ii/(Ii − γiVΘ,i)). We consider arbitrary generic spike sequences in which all neurons are active, i.e. there is a finite constant T > 0, such that in every time interval [t, t + T ), t ∈ R, every neuron fires at least once. Further, we assume that the dynamics is sufficiently irregular such that two events occur at the same time with zero probability. 4 Due to the delay, the state space is formally infinite dimensional. However, it becomes finite dimensional after some finite time t′ (cf. (Ashwin and Timme, 2005)). At a given time t > t′ the network dynamics is completely determined by the phases φi(t) and by the spikes which have been sent but not yet received by the postsynaptic neurons at t. Their number is bounded by some constant N D′. Due to the inhibitory character of the network couplings, each neuron i needs at least the free oscillation period φΘ,i = T free to generate a spike after the last reset. Consequently, at most i D′ = N(cid:24) max i,j∈{1,...,N }(cid:18) τij φΘ,i(cid:19)(cid:25) spikes per neuron are in transit and the state space stays finite, with dimensionality smaller than or equal to N ·(1+D′) (cf. also (Ashwin and Timme, 2005)). Here⌈x⌉ denotes the ceiling function, the smallest integer larger or equal to x. We now introduce variables to describe spikes, which are already sent at time t by neuron j to the postsynaptic neuron i and not yet received. A single spike in transit is characterized by the state variable σijk(t) ∈ [0, τij]. The index k = 1, 2, 3 . . . ≤ D′/N numbers the different spikes traveling from neuron j to i at time t in the order of arrival at the postsynaptic neuron i, starting with k = 1 for the next spike to arrive. When spikes are emitted at time ts jn for some n, σijk(t) is set to σijk(ts jn) = 0. The spike index k equals the number of spikes already in transit plus one. It thus depends on the actual network state at time ts jn. Between two events σijk(t) increases linearly with slope one, when σijk(ts jn + τij) = τij the spike is received by the postsynaptic neuron i, where it induces a phase jump according to Eq. (3). After spike reception we cancel the spike arrived (which has index k = 1) and renumber the indices k > 1 as k → k − 1 such that σij1(ts jn + τij) specifies the spike sent by neuron j which arrives next at the postsynaptic neuron i (cf. Fig. 3(A,B) for illustration). III. RESULTS A. Lyapunov stability of arbitrary generic spike sequences In this section we study the stability properties of the spike sequences. We compare the microscopic dynamics of two sequences, that slightly differ in the timing of the spikes, but have the same ordering. We show that the distance between these trajectories is bounded by the initial distance. Assuming that one sequence is generated by a perturbation of the other, this implies Lyapunov stability for the considered spike pattern. Distances and perturbation sizes are measured using the maximum norm. Since the distance between trajectories only changes at event times, we can choose an event-based perspective. The time of the nth event in the entire network is denoted by tn in the first sequence, and by tn in the second one. Analogously, we denote the phases of a neuron i at a given time t by φi(t) and φi(t) and the spikes in transit at time t by σijk(t) and σijk(t) in the different sequences. Let (5) (6) (7) ∆(n) i :=(cid:16)φi (tn) − φi(cid:0)tn(cid:1)(cid:17) −(cid:0)tn − tn(cid:1) = δ(n) i − δt(n) denote the difference between the phase difference, δ(n) after the nth and before the (n + 1)th event (cf. Fig. 3(C)). Similarly i := φi (tn) − φi(cid:0)tn(cid:1), and the temporal offset, δt(n) := tn − tn, ∆σ(n) ijk :=(cid:0)σijk (tn) − σijk(cid:0)tn(cid:1)(cid:1) −(cid:0)tn − tn(cid:1) = δσ(n) ijk − δt(n) labels the shift of the kth spike sent by neuron j and not yet arrived at neuron i after the nth and before the (n + 1)th event. Between two consecutive events, both φi(t), φi(t) and σijk(t), σijk(t) increase linearly and only at event times the phases and spike variables are updated nonlinearly as described above. Therefore, to study the stability of the system it is sufficient to consider the phase shifts after events. In the following we investigate the evolution of shifts at the discrete event times. There are two different kinds of events: (i) sending and (ii) receiving of spikes. In the first case, the shifts ∆(n) ijk stay unchanged, but new spikes with new spike variables are generated. These variables inherit the perturbation of the sending neuron. In the second case, the phase shift of the neuron which receives the spike changes and the spikes in transit are reordered. The resulting phase shift of the neuron receiving a spike turns out to be a weighted sum of previous shifts. This can be shown by studying both cases in detail. and ∆σ(n) i 5 A j ) t ( φ φ Θ, j 0 B ) t ( j k σi τ ij C i ) t ( φ ~ , ) t ( φ i time +τ tn ij φ i(t) ∆ (n−1) i t−δ (n) (n)∆ i δ (n) i k=1 k=2 tn+1 tn k=1 time 0 i(t)~ φ tn ~ tn time Figure 3: (A) Phase dynamics of neuron j and definition of state variables. At t = tn and t = tn+1 the phase φj(t) reaches the threshold and is reset to 0. (B) The spikes emitted travel to the postsynaptic neurons i ∈ Post(j). They are described by σijk(t). In this example we show two spikes traveling from neuron j to one specific postsynaptic neuron i, described by σij1(t) (black) and σij2(t) (red). At t = tn, σij1(tn) is set to 0; here k = 1 because there is no spike in transit at t = t− n . When neuron j spikes again at t = tn+1, σij2(tn+1) is set to 0. At t = tn + τij the spike emitted at t = tn arrives at the postsynaptic neuron i and induces a phase jump in φi(t) (not shown, cf. Fig. 2). After spike reception, we renumber k → k − 1 such that σij2 → σij1. (C) Definition of the phase shifts. The phase curve φi(t) of neuron i (blue) before and after the reception of a spike at t = tn i = φi (tn) − φi (cid:0)tn(cid:1) is the difference of neuron i's phases in the unperturbed and the is shown together with φi(t) (black). δ(n) perturbed dynamics taken at corresponding event times tn and tn. δt(n) = tn − tn denotes the difference of event times tn and tn, i.e. the temporal offset between both sequences. Finally, ∆(n) i − δt(n) is some phase shift of neuron i with the temporal offset taken into account. i = δ(n) 1. Transfer of perturbations without change of size. If as (n + 1)th event the phase of some neuron j∗ reaches its threshold and a spike is emitted, the shifts of all neurons' phases stay unchanged, Similarly, the shifts of the spikes in transit stay unchanged j ∆(n+1) = φi (tn) + (tn+1 − tn) −(cid:16) φi(cid:0)tn(cid:1) +(cid:0)tn+1 − tn(cid:1)(cid:17) − δt(n+1) = ∆(n) ijk = σijk (tn) + (tn+1 − tn) −(cid:0)σijk(cid:0)tn(cid:1) +(cid:0)tn+1 − tn(cid:1)(cid:1) − δt(n+1) = ∆σ(n) . j ijk . ∆σ(n+1) Additionally new spikes are generated σij∗ k∗ (tn+1) = 0 and σij∗ k∗ (tn+1) = 0 where k∗ = k∗(i, j∗, n + 1) is the appropriate spike number, cf. Fig. 3. The shifts of the new spike variables depend on the phase shift of the sending neuron j∗according to (8) (9) (10) ∆σ(n+1) ij∗ k∗ = −δt(n+1) = −h(tn + φΘ,j∗ − φj∗ (tn)) −(cid:16)tn + φΘ,j∗ − φj∗(cid:0)tn(cid:1)(cid:17)i = ∆(n) j∗ . 2. Averaging of prior perturbations. induces a phase jump in the postsynaptic neuron i∗. According to Eq. (3), the phase shift ∆(n+1) computed as If as (n + 1)th event some spike arrives, say σi∗ j∗1 (tn+1) = τi∗j∗ , it can be i∗ where φi∗(cid:0)t− ∆(n+1) i∗ = H (i∗) εi∗ j∗(cid:0)φi∗(cid:0)t− n+1(cid:1) = φi∗ (tn) + (tn+1 − tn) and φi∗(cid:0)t− −(cid:1) = φi∗(cid:0)t− φi∗(cid:0)tn+1 n+1(cid:1)(cid:17) − δt(n+1), εi∗ j∗(cid:16) φi∗(cid:0)t− n+1(cid:1)(cid:1) − H (i∗) n+1(cid:1) = φi∗ (etn) +(cid:0)etn+1 −etn(cid:1) are the phases "just before" n+1(cid:1) + ∆(n) i∗ + δt(n+1). (11) (12) Applying the mean value theorem in Eq. (11) and the relation spike reception. Using the definitions (6) we find the identity δt(n+1) =tn+1 − tn+1 =tn + τi∗j∗ − σi∗ j∗1(tn) − tn − τi∗j∗ + σi∗ j∗1(tn) = − ∆σ(n) i∗j∗1 (13) A 1 ) φ ( ) i ε ( H 0 −1 −1 B 3 6 φ ) ( ' i U 1.5 0 −1 0 phase φ 1 0 phase φ 1 Figure 4: (A) The transfer function H (i) 0 (φ) = φ is the identity (black), for ε < 0 (blue: ε = −0.5, green: ε = −1) the phase φ after receiving an input is smaller than before. For all inhibitory inputs we find H (i) ε (φ)/∂φ ≤ 1 for ε ≤ 0 (cf. Eq. (22)). ε (φ) < φ. (B) The derivative of Ui(φ) is monotonic decreasing, therefore ∂H (i) ε (φ) for a leaky integrate-and-fire neuron. For ε = 0 (black), H (i) yields ∆(n+1) i∗ = c(n+1) i∗j∗ where c(n+1) i∗ is given by the derivative c(n+1) i∗j∗ = ∂H (i∗) εi∗ j∗ (φ) ∂φ · ∆(n) i∗ j∗1, i∗ j∗ (cid:17) · ∆σ(n) i∗ +(cid:16)1 − c(n+1) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)φ∈[φi∗(t− n+1(cid:1), φ takes values φ ∈h φi∗(cid:0)t− n+1), φi∗(t− n+1)] for φi∗(cid:0)t− n+1(cid:1) ≤ φi∗(cid:0)t− n+1(cid:1); for φi∗(cid:0)t− n+1(cid:1) > φi∗(cid:0)t− i∗ is not connected to neuron j∗, εi∗j∗ = 0, the function H (i∗) that the phase shift stays unchanged, ∆(n+1) For εi∗j∗ < 0, c(n+1) is bounded by = ∆(n) i∗ , indeed c(n+1) εi∗ j∗ (φ) = H (i∗) i∗j∗ = dH (i∗) i∗ 0 0 i∗ n+1(cid:1) , φi∗(cid:0)t− n+1(cid:1)i. If neuron (φ) = φ becomes the identity, such (φ)/dφ = 1 independent of φ. (14) (15) (17) (18) (19) (20) where εmax := max {εij } and εmin := min i,j:εij 6=0 i,j:εij 6=0 increasing with ε  ∂ε(cid:18)hU ′ ∂ ∂φ cmin := inf εmin (φ)(cid:17) ∂(cid:16)H (k) φ,k ∂φ ) = U ′ ∂ε( ∂H (k) (cid:16)U ′ = − k(φ) (φ) ∂ ε k(H (k) ε U ′ k(φ) k(H (k) ε (φ))(cid:17)2 · U ′′ ≤ c(n+1) i∗ ≤ sup ∂(cid:16)H (k) εmax (φ)(cid:17)  {εij}. We have used that ∂(cid:16)H (k) ∂φ ε φ,k (φ))i−1(cid:19) k (H (k) ε (φ)) · 1 k(H (k) U ′ ε (φ)) > 0. =: cmax, (16) (φ)(cid:17) /∂φ is monotonic The shifts of traveling spikes stay unchanged on spike reception, cf. Eq. (9), ∆σ(n+1) ijk = ∆σ(n) ijk for all spikes with i 6= i∗ ∨ j 6= j∗. For i = i∗, j = j∗ the spike variables are renumbered and ∆σ(n+1) i∗ j∗k−1 = ∆σ(n) i∗j∗ k holds, except for k = 1, because σi∗ j∗1(t) is the variable describing the spike received and therefore canceled. We will now ascertain that the coefficients c(n+1) in Eq. (14) lie in a compact interval within (0, 1), such that a true averaging takes places when interactions happen. Formally, the phases of neurons can achieve values φi ∈ (−∞, φΘ,i]. i∗j∗ Each neuron fires at least once within a time interval of length T , therefore the phases are certainly bounded to the compact interval φi ∈ [−T + φΘ,i, φΘ,i]. Further, in inhibitory networks the phase after an interaction is smaller than before, 7 because together with Ui also U −1 i is strictly monotonic increasing and ε < 0. The strict concavity of Ui (φ) implies H (i) ε (φ) = U −1 i (Ui(φ) + ε) < U −1 i (Ui(φ)) = φ, (21) 0 < ∂H (i) ε ∂φ (φ) = U ′ i (φ) U ′ i(cid:16)H (i) ε (φ)(cid:17) < 1 (22) for any finite φ (cf. Fig. 4 for illustration). The derivative ∂H(i) ε (φ) ∂φ is continuous in φ, therefore the image of [−T + φΘ,i, φΘ,i] under the map (φ) ∂H(i) εij ∂φ is compact. Together with Eq. (22) it follows that Taken together, Eq. (8, 9, 10, 14, 23) imply that a true averaging between shifts already present in the system takes place when a spike is received. For other events the shifts stay unchanged. As a consequence, the maximal and minimal shift after the nth event, 0 < cmin, cmax < 1. (23) ∆(n) max := max i,j,kn∆(n) i , ∆σ(n) ijko and ∆(n) min := min i,j,kn∆(n) i are bounded by the initial shifts for all future events, ∆(n) max ≤ ∆(0) max and ∆(n) min ≥ ∆(0) min, , ∆σ(n) ijko , (24) (25) as long as the order of events in both sequences is the same. Here the minima and maxima are taken over i, j ∈ {1, . . . , N } and k numbers the spikes traveling from neuron j to i at time tn. An initial perturbation cannot grow, thus the trajectory is Lyapunov stable. We note that we did not make any assumptions about the network connectivity, the results hold for any network structure and the described class of trajectories. B. Asymptotic stability In this section we prove that for strongly connected networks even asymptotic stability holds under the condition that the perturbed and the unperturbed sequences have the same order of events, i.e. the order of events is unchanged by small perturbations. The central idea is as follows: We study the dynamics and convergence of two neighboring trajectories. We will track the propagation of the perturbation of one specific neuron l0 through the entire network. Since there is a directed connection between every pair of neurons in the network and any spike reception leads to an averaging of shifts, there is an averaging over all perturbations in the network. For large times all perturbations converge towards the same value, such that both sequences become equivalent, only shifted by a constant temporal offset. Further details and the derivation of the following Eqs. (26),(27) and (28) is provided in the appendix. We find the upper bound for the maximal perturbation after K := 2N M events and analogously the lower bound ∆(K) max ≤ c∗ · ∆(0) l0 + [1 − c∗] · ∆(0) max ∆(K) min ≥ c∗ · ∆(0) l0 + [1 − c∗] · ∆(0) min (26) (27) for the minimal perturbation. The averaging factor c∗ is determined by the network parameters and bounded to A bound for the difference of the maximal and minimal perturbation after K events is therefore given by 3/4 ≤ 1 − c∗ < 1. ∆(K) max − ∆(K) min ≤ (1 − c∗)(cid:16)∆(0) max − ∆(0) min(cid:17) . (28) (29) 8 A 0.01 ) n ( κ , ) n ( δ i x a m 10−10 10−14 0 B 0 ) n ( t δ −5 −10 # events n 3500 0 # events n 900 Figure 5: Stable dynamics in the network of Fig. 1(A-C). (A) Exponential decay of the maximal perturbation maxi (cid:12)(cid:12)(cid:12) (blue dots) and minimal margin κ(n) (gray line) for one microscopic dynamics. The initial perturbation max(cid:12)(cid:12)(cid:12) decays exponentially fast. (B) Exponential convergence of the temporal offset δt(n). For large n the original and the perturbed sequences are equivalent, just shifted by a constant offset δt = limn→∞ nδt(n)o . (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) ≈ 10−5 δ(n) δ(0) i i The spread of any perturbation through the network has a contracting effect on the total perturbation, it leads to a decay of the difference between the extremal perturbations at least by a factor (1 − c∗). Inequality (29) implies together with the bound (28) that for the considered trajectories in the long-time limit the maximal and minimal perturbation are the same, Thus, for t → ∞ the events are just shifted by a constant temporal offset ∆(n) max δt = lim n→∞ δt(n) = − lim n→∞ lim n→∞ ∆(n) max = lim n→∞ ∆(n) min. (30) (31) (cf. Fig. 5 for illustration), and both sequences become equivalent limn→∞ δ(n) i = 0. Thus all sequences considered are asymptotically stable for all strongly connected networks and perturbations decay exponentially fast with at least ∆(n) max − ∆(n) min ≤ (1 − c∗)⌊ n max − ∆(0) (32) K ⌋(cid:16)∆(0) min(cid:17) , where ⌊.⌋ is the floor function. The actual and numerically measured exponential decay is much faster than the estimation given by Eq. (32), because for deriving Eqs. (26) and (27) in the appendix we had to assume a worst-case scenario. The main reasons for the faster decay are that (i) the mean path length is more meaningful for estimating the number of events until all neurons have received an input from the starting one and (ii) it is impossible that the neurons receive the worst-case perturbation at each reception. C. Margins The stability results in the previous sections hold for the class of patterns, where a small perturbation does not change the order of spikes. In this section we show that typical spike patterns, generated by networks with a complex connectivity, belong to this class. In heterogeneous networks with purely inhibitory interactions the occurrence of events at identical times has a zero probability. There is no mechanism causing simultaneous spiking, like supra-threshold inputs in excitatorily coupled networks. As long as two events do not occur at the same event time, there is a non-zero perturbation size keeping the order unchanged in any finite time interval. However, the requirement of an unchanged event order yields more and more conditions over time such that the allowed size of a perturbation could decay more quickly with time than the actual perturbation. This is excluded if a temporal margin µ(n) (cf. (Jin, 2002)) stays larger than the dynamical perturbation for infinite time. Formally, after time tn denote the kth potential future event time (of the original trajectory) that would arise if there were no future interactions by θn,k, k ∈ N, and the temporal margin by µ(n) := θn,2 − θn,1. A sufficiently small perturbation, satisfying ∆(n) min < µ(n), cannot change the order of the (n + 1)th event. max − ∆(n) Stability of generic periodic orbits. This directly implies that almost all periodic orbits (all those with non- degenerate event times tn) consisting of a finite number of P events are stable because there is a minimal margin κ(P ) := min n∈{1,...,P } µ(n) (33) for every non-degenerate periodic pattern (c.f. also (Memmesheimer and Timme, 2009)). Stability of irregular non-periodic spiking activity. We study the stability properties of irregular non- periodic spike sequences by considering the minimal margin κ(n) over the first n events. For simplicity, we consider delay distributions where τij is independent of the spike receiving neuron i. The irregular spiking dynamics of the entire network is well modeled by a Poisson point process with rate ν s, where ν s specifies the mean firing rate of the network (Brunel, 2000, Brunel and Hakim, 1999, Burkitt, 2006, Tuckwell, 1988). We assume that, along with the irregular dynamics, the temporal margins are also generated by a Poisson point process with the network event rate ν = 2ν s, because any spike sending time generates one receiving event due to the independence of τij from i and the definition of events. The distribution function of margins is given by 9 Therefore, the probability that the minimal margin κ(n) after n events is smaller or equal to µ is determined by the probabilities that not all individual margins µ(n) are larger than µ such that P(cid:16)µ(n) ≤ µ(cid:17) = 1 − e−νµ. P(cid:16)κ(n) ≤ µ(cid:17) = 1 − nYm=1 P(cid:16)µ(m) > µ(cid:17) = 1 − e−nνµ (34) (35) with density ρn(µ) := dP(cid:0)κ(n) ≤ µ(cid:1) /dµ = nν exp(−nνµ). This implies an algebraic decay with the number n of events for the expected minimal margin κ(n) =Z ∞ 0 µρn(µ)dµ = (νn)−1 (36) that depends only on the event rate and is independent of the specific network parameters. Numerical simulations show excellent agreement with this algebraic decay (36); a typical example is shown in Fig. 6(C). This already strongly suggests that a sufficiently small perturbation stays smaller than the minimal margin for all times. However, in each step, the exponential distribution of margins has finite density for arbitrary small values of µ, i.e. in each step the margin can fall below the level of perturbation with finite probability. We will show that P(cid:16)∃n ∈ N : µ(n) ≤ ∆(n) perturbation size, goes to zero if the size of the initial perturbation goes to zero. Thus, also the probability that there is a change in the order of events goes to zero. Of course, we cannot expect to reach zero for nonzero perturbation. min(cid:17), the probability that there is at least one step in which the margin falls below the max − ∆(n) We derive a lower bound for the probability that the margin stays larger than ∆(n) min for infinite time. We show that it converges to one when the size of the perturbation goes to zero and thus prove that sufficiently small perturbations have arbitrarily high probability to stay smaller than the minimal margin for all times. We start from max − ∆(n) K − 1 and Eq. (28) leads to K(cid:5) ≥ n the upper bound for the evolution of the perturbation, Eq. (32). Using(cid:4) n min(cid:17) , n(cid:19) ∆(0) K −1(cid:16)∆(0) = exp(cid:18) log(1 − c∗) min ≤ (1 − c∗) max − ∆(n) max − ∆(0) 1 − c∗ ∆(n) n max − ∆(0) min K = C exp(−αn), where we introduced C := ∆(0) max − ∆(0) min 1 − c∗ log(1 − c∗) K , > 0. α := − In particular, C → 0 if the initial perturbation goes to zero, i.e. ∆(0) initial perturbation. The probability, that all margins are larger than all perturbations is given by max − ∆(0) min → 0, while α is independent of the P(cid:16)∀n : µ(n) > ∆(n) max − ∆(n) min(cid:17) = ∞Yn=1 P(cid:16)µ(n) > ∆(n) max − ∆(n) min(cid:17) , (40) , (37) (38) (39) A 1 ) µ < ) 1 ( κ ( P 0.4 B 1 ) µ < ) 5 0 1 ( κ ( P 0.4 0.2 0 0 0.2 0 0 0.1 0.2 µ 0.1 10µ / −6 0.2 C 1 ) n ( κ , ) n ( κ 10−4 10−8 1 3 10 # events n 106 10 Figure 6: Margins in the network given in Fig. 1(A-C). (A,B) Probability distribution P (κ(n) ≤ µ) after (A) n = 1 and (B) n = 105 events. The blue curve is measured distribution over 2500 samples, the red dotted line is the analytical prediction (no free fit parameter, rate ν is measured; cf. Eq. (35)). (C) Algebraic decay of the average minimal margin, κ(n) (green dashed line, averaged over 250 random initial conditions) and its analytical prediction (no free fit parameter; black solid line). Additionally we show the minimal margin κ(n) for three exemplary initial conditions (gray lines), including that of Fig. 5(A). since the margins are independent. Using Eq. (37) and P(cid:0)µ(n) > µ(cid:1) = exp(−νµ) yields P(cid:16)µ(n) > C exp(−αn)(cid:17) P(cid:16)µ(n) > ∆(n) min(cid:17) ≥ max − ∆(n) ∞Yn=1 exp(−νC exp(−αn)) = ∞Yn=1 ∞Yn=1 = exp −νC = exp(cid:18)−νC exp(−αn)! ∞Xn=1 exp(α) − 1(cid:19) , 1 (41) which goes to one if the initial perturbation (and thus C) goes to zero. We note that the assumption of a constant lower bound of the minimal margin is not necessary in contrast to (Jin, 2002). Indeed this assumption would be highly problematic, because in an irregular dynamics arbitrarily small margins naturally occur with positive probability in every step. So, assuming some lower bound would exclude generic irregular dynamics. In contrast, the novel approach introduced above enabled us to show that the generic irregular dynamics is stable. For infinitely large networks in a mean field approach, where one takes the limit of infinitely many neurons in the beginning (e.g. Brunel and Hakim (1999), v. Vreeswijk and Sompolinsky (1996, 1998)), our method is not applicable in a straightforward way. In this limit the average inter-spike interval goes to zero and a positive margin cannot be presumed. In our analysis we employ the fact that the minimal margin stays finite, i.e. a sufficiently small perturbation does not change the order of events. This assumption and therefore our results hold for arbitrarily large finite systems. D. Convergence to periodic orbits Interestingly, the asymptotic stability together with the finite phase space imply that generic spike sequences converge to a periodic orbit. To show convergence analytically, we extend and explicate the ideas presented in (Jahnke et al., 2008a, Jin, 2002). In the following we focus on a finite subsequence s∗ of a spike sequence generated by a given network. The number E∗ of events in s∗ is called the length E∗ of s∗. As discussed towards the end of section II, the considered system is finite dimensional with dimension bounded from above by N + N D′. Thus, if s∗ is sufficiently long it contains at least two disjoint subsequences s1 and s2 of length E, where the ordering of events is identical. The maximal E for which the existence of s1 and s2 is guaranteed, is given by the largest integer E satisfying When increasing the observation length E∗ also the possible length E of the subsequences increases. Both the phases and the variables encoding the spikes in transit are bounded to a finite interval, φi(t) ∈ [−T + φΘ,i, φΘ,i] and E∗ ≥ (N + N D′)E + E. (42) A 40 1 1 11 B 0.1 ) n ( κ 10−5 κ ) n ( ) , n ( µ 1 10−3 10−6 10−9 1 10 4 7 10 10 # events n C 40 n o r u e n 5 280 40 D n o r u e n time 294 40 10−9 1 A B 410 # events n 108 5 1830270 time 1830284 Figure 7: Convergence towards a periodic orbit in a random network (N = 40, γi ≡ 1, Ii ≡ 3.0, VΘ,i ≡ 1.0, τij ≡ T free /10, Pre(i) ≡ 8, Pj ǫij ≡ −3.3). (A) Coupling matrix, each realized connection is indicated by a black square. (B) Average minimal margin κ(n) (averaged over 250 random initial conditions, cf. Fig. 6(C)) decays as a power-law (region A) and saturates after about 107 events (region B) when the periodic orbit is reached. Inset: Margin µ(n) (black) and minimal margin κ(n) (gray) for a trajectory started from one specific initial condition. The margin µ(n) fluctuates strongly on the transient and is comparatively large after the sequence becomes periodic; thus the minimal margin κ(n) does not decrease further for future events n. (C), (D): Snapshots of irregular spike sequences (C) after n ≈ 15.000 events on the transient and (D) after n ≈ 108 events on the periodic orbit. i σijk(t) ∈ [0, maxij {τij }] at any given time t. Therefore the maximal event-based distance between two trajectories is also bounded to a finite size, Φmax := max i,j {T, τij } . (43) Thus comparing sequences s1 and s2 the initial event-based distance between their underlying trajectories at the beginning of s1 and at the beginning of s2 is bounded by Φmax, i.e. ∆(0) min ≤ Φmax. By definition, the order of events in both sequences s1 and s2 is the same; therefore the distance between them shrinks according to Eq. (32). After E events the distance is bounded by (cf. Eq. (37)) max − ∆(0) ∆(E) max − ∆(E) min ≤ Φ∗ max := Φmax 1 − c∗ exp (−αE) , (44) where we used the definition (39). We note that, if we increase our observation length E∗ and therefore the subsequence length E, the maximal possible distance Φ∗ max between the trajectories underlying the sequences s1 and s2 decreases. If Φ∗ max is sufficiently small, i.e. the distance between s1 and s2 after E events is smaller than the average margin, there is a high probability that also the order of events in the sequence following s1 and s2 are the same. Analogous to Eq. (41) the probability is given by PE ∗ = exp(cid:18)−ν 1 (1 − c∗) (exp(α) − 1) Φ∗ max(cid:19) , (45) which goes to one when the observation length E∗ tends to infinity. This implies a periodic dynamics: Assume that s1 occurs first in s∗ at the ath event and after a certain event number a + L ≤ E∗ − E the sequence s2 begins. We have seen that (with a certain probability) the ordering of events for (infinite) sequences starting at the ath and the (a + L)th event is the same for all future times. Therefore also the sequence starting at the (a + 2L)th event is identical to the ones mentioned before, so the ordering of events is periodic. Together with the exponential convergence of equally ordered sequences this implies that also the spike timings converge towards a periodic orbit. For arbitrarily long observed sequences this happens with an arbitrary large probability that tends to one as E∗ → ∞. In Fig. 7 we show a typical example: The mean margin, κ(n), decays algebraically on the transient and saturates after the periodic orbit is reached. Interestingly, the periodic attractor (shown in Fig. 7(D)) of the sparse random network resembles the "splay state" known from globally coupled networks (Strogatz and Mirollo, 1993). In a splay state the firing pattern is characterized by equally spaced inter-spike-intervals. It has been shown that it is possible to design networks, which exhibit more complex periodic spike patterns (Memmesheimer and Timme, 2006a,b). Indeed, in different parameter regimes we observe such complex periodic orbits, with a large periodicity, cf. the heterogeneous 50 n o r u e n 1 0 12 2000 time t 4000 Figure 8: Long periodic orbit in a random network (N = 50, γi ≡ 1, Ii ≡ 1.2, VΘ,i ≡ 1.0, τi,j ≡ 0.1T free are randomly independently drawn from the uniform distribution [−1, 0] and normalized afterwards (PN final spike pattern (shown after a transient of 9 · 106 events) repeats every 11012 events and is highly irregular. ). The coupling strength j=1 εij = −6.5). The i globally coupled network in Fig. 8, where the periodic orbit is reached after a small number of events compared to sparse networks. As shown previously (Timme and Wolf, 2008, Timme et al., 2002), highly irregular spiking activity may coexist with even the simplest (fully synchronous) periodic orbits that exhibits regular, maximally ordered activity. increasing the network size N , we leave the sum I := Ii +PN internal inhibition,PN Although the attractor is reached after a finite number of events, the transient becomes very important in sys- tems with strong inhibition or large number of neurons. As formerly found in networks with excitatory coupling (Zumdieck et al., 2004), and also in weakly diluted networks with purely inhibitory interactions (Zillmer et al., 2006), the transient length grows rapidly with network size such that the dynamics is governed by the transient for large time scales. We study inhibitory random networks with an arbitrary network structure, typically far away from the weakly diluted topology. To perform numerical measurements of the transient length in dependence on the network size N , we define the length of the transient, tr by the number of events after which the order of events stays periodic. When j=1 εij of the external excitatiory current, Ii, and the j=1 εij, constant. Thus, on average each neuron receives a constant effective input independent of N and the mean firing rate of a single neuron hνii is approximately conserved. Fig. 9 shows the increase of tran- sient lengths with network size for different sizes of internal inhibition and different scalings of the single connection strengths. We observe an exponential increase of the mean transient length with network size N independent of the scaling of the coupling strengths, εij, and the number of incoming connections, Ki := Pre(i). This is qualitatively similar to the scaling of the transient lengths in weakly diluted networks (Zillmer et al., 2006). Assuming that the rapid increase continues for much larger networks, the transient will dominate the dynamics essentially forever in net- works of biological relevant sizes. In this sense the transient becomes quasi-stationary (cf. also Zillmer et al. (2009, 2006)). If a larger network is in the balanced state (cf. Fig. 1), the stable transients typically dominate the network dynamics. E. Robustness of stability and smooth transition to chaos In the following we will check the robustness of our results. The considerations above hold for networks with inhibitory interactions without temporal extent. We investigate the influence of excitatory interaction and pulses with a finite duration. For small deviation from the networks considered above the stability properties are similar, for large fractions of excitatory neurons and large temporal extent we observe a transition to a chaotic regime. For temporally extended couplings we assume that after a neuron is reset all previous input is lost. Therefore the state of a neuron is specified by its last spike time and all spikes it has received afterwards. The phase representation is thus not meaningful anymore and we track two trajectories by comparing the differences in the last spiking times of the neurons and in the spike arrival times since these last spikings. To keep the section consistent, we adopt this view when studying the Lyapunov exponents of the excitatory dynamics. 6 8 10 . > r t < . 4 8 10 10 K 30 A 6 3 10 . > r t B < . 5 3 10 −0.34 −0.46 e0 13 D C 106 > r t < 105 −1 N ε Σ j=1 ij −2 310 r t h 210 110 ε ij 0 (N) ε ij −1/2 (N) ε ij −1 (N) 60 100 140 60 100 network size N 140 60 100 140 0 0 0.4 0.8 1.2 t r 106 810 > r t < 610 410 Figure 9: Scaling of the transient lengths in sparse random networks. (A) The sum of the external excitation and the internal inhibition is fixed, Ii := PN j=1 εij + Ii ≡ 1.0, such that the mean firing rate, hνiii, of the single neurons stays approximately constant. The number of incoming connections to each neurons, Ki := Pre(i), is fixed to Ki ≡ 15 (green +), Ki ≡ 20 (black ×), Ki ≡ 25 (red ∗). The single connections are set to εij = e0/√Ki with e0 = −0.38. The solid lines show the best exponential fit. We observe an exponential increase of the trial averaged transient length htri with network size N . The inset shows the dependence of the transient length on the in-degree for fixed network size (N = 100). (B) Networks with fixed j=1 εij = e0√Ki, such that fraction of connections, where Ki ≡ ⌊pN⌋ and p = 1/4. We scale the coupling strengths as PN εij ∝ 1/√Ki, and the external current Ii is adjusted appropriately to fix the mean firing rate hνiii. The mean transient lengths for different e0 are shown together with the best exponential fit (e0 = −0.34 (green +), e0 = −0.38 (black ×), e0 = −0.42 (red ∗)). Again, we observe an exponential increase of the transient lengths with N . The inset shows the increase of the (C) Networks with fixed fraction of transient length with inhibitory coupling strength for fixed network size (N = 100). connections, where Ki ≡ ⌊pN⌋ and p = 1/4. The external currents Ii ≡ 3.0 are fixed and the internal coupling is normalized to PN j=1 εij ≡ −1.1 (green +), −1.5 (blue ×), −1.75 (black ∗), −1.9 (red (cid:3)), such that the mean firing rate, hνiii, of the single neurons is approximately the same for each curve. The number of incoming connections per neuron, Ki, increases linearly with network size, so the coupling strengths are scaled as εij ∝ 1/Ki. Here we also observe an exponential increase of htri with network size. The inset shows the fast increase of the transient length with inhibitory coupling strength for fixed network size and external current (N = 100, Ii ≡ 3.0). (D) The transient length is broadly distributed as shown in the histogram for 2500 trials started from random initial conditions, where the initial phases were randomly independently drawn from the uniform distribution on [0, 1] (N = 100, Ii ≡ 3.0, PN j=1 εij ≡ −1.5) Excitatory interactions We have shown that in networks with purely inhibitory interactions the dynamics is typically stable. If the con- nection from neuron j to neuron i is excitatory, the phase shift ∆(n+1) of the postsynaptic neuron i after receiving before and the shift ∆σ(n) a spike from neuron j as the (n + 1)th event may exceed its shift ∆(n) ij1 of the received spike. Fig. 10(A,B) gives an illustration: A spike is simultaneously received in the perturbed and in the unperturbed dynamics (i.e. ∆σ(n) ε (φ) is shown for (A) an inhibitory input and (B) an excitatory input. For inhibitory input the phase shift ∆(n) leads to the stable dynamics as described in the sections above. For excitatory input, the phase shift ∆(n) when the spike is received. ij1 = 0). The phase shift before and after the application of the transfer function H (i) is reduced, this increases i i i i Indeed, since the inverse of Ui(φ) is monotonically increasing with φ, we find for a given ε > 0 In contrast to Eq. (22) the derivative of the transfer function H (i) ε (φ) is bounded from below by H (i) ε (φ) = U −1 i (U (φ) + ε) > U −1 i (Ui(φ)) = φ. (46) (47) dH (i) ε dφ (φ) = U ′ U ′ i (φ) i(cid:16)H (i) ε (φ)(cid:17) > 1. According to Eq. (14), this can lead to an increase of (in particular extremal) perturbations and to a destabilization of the trajectory. The upper bound of c(n) ij , cmax < 1 (cf. Eq. (23)) does not hold anymore. However, in a network with a small fraction of excitatory connections, the trajectory is still stable. At an interaction the perturbation may increase, but the stabilizing effect of inhibitory inputs dominates the dynamics. We study the transition from the 14 stable regime to chaotic dynamics (a discussion of the chaotic dynamics in networks with purely excitatory interactions can be found in (Zumdieck et al., 2004)). When increasing the number of excitatory couplings, we increase the mean effective input current to the neurons. Thus we additionally decrease the external input Ii to keep the network rate ν constant. Indeed, in good approximation, the current has to be decreased linearly with NE, the number of excitatory connections, Ii ≡ I − kNE where I is the original input current. To quantify the transition we estimate the largest Lyapunov exponent of the system: At the nth event, we denote n − W (n) the earliest event which still influences the future dynamics of the system explicitly. We apply an initial perturbation of size k∆0k1 to the event times t0, t−1, . . . , t−W (0), where ∆n = (∆tn, ∆tn−1, . . . , ∆tn−W (n)) is the perturbation vector at the nth event time and ∆ti is the perturbation of ti. We evolve the system and rescale the perturbation vector ∆n by an after each event, such that the rescaled perturbation vector is of the same size as the initial perturbation, The largest Lyapunov exponent, λmax, is then given by k∆′ nk1 = k∆n · ank1 = k∆0k1 . λmax = lim n→∞ λ(n) with λ(n) := 1 n nXi=1 ln(a−1 i ). (48) (49) We observe a transition from a stable to a chaotic regime, characterized by a positive Lyapunov exponent. For small fraction of excitatory neurons the dynamics is typically stable, the effect of the inhibitory pulses dominates the dynamics and, on average, a perturbation do not grow over time. With increasing NE the Lyapunov exponent increases until the dynamic becomes chaotic. Of course, in our simulations we can only study finite time Lyapunov exponents with very large n and estimate the value to which they converge. The chaotic dynamics may thus be transient. However, it dominates the dynamics at least over very long times. Estimating the maximal Lyapunov exponent in networks including excitatory interactions can be difficult (Brette et al., 2007, Cessac and Vieville, 2008, Hansel et al., 1998, Kirst and Timme, 2009, Zumdieck et al., 2004). Suprathreshold excitation, together with the infinitely fast response of neurons receiving a spike and the sharp thresh- old may induce synchronous events. Thus even an infinitesimal small perturbation may change the order of spikes. Nonetheless generically the perturbation will stay infinitesimal small, in particular for a small fraction of excitatory connections, such that we estimate the largest Lyapunov exponent in the following way: We evaluate at each time step the resulting temporal perturbation on the actual event as a result of earlier perturbation under the assumption that the order of spikes stays the same. This gives us the new perturbation vector ∆n, which is rescaled according to (48). For long times λ(n) then will give an estimate of the largest Lyapunov exponent and describe the generic behavior of the trajectory under the influence of sufficiently small perturbations. However, we cannot exclude the occurrence of macroscopic perturbations in general. Fig. 10(C,D) shows some numerical results: In Fig. 10(C) the largest Lyapunov exponent is measured for an increasing fraction of excitatory connections starting with the network of Fig. 1(A-C) and ending with the network of Fig. 1(D-F). The number of excitatory connections is increased by successively choosing one incoming inhibitory connection per neuron to be excitatory. The external current, Ii, is reduced linearly to keep the network rate unchanged according to Ii ≡ I − kNE, where I = 4.0 is the initial external current, k ≈ 0.052 and NE is the number of excitatory connections. For a large fraction of excitatory couplings we observe a transition to an unstable, chaotic regime. The inset demonstrate the convergence of 1 ) → λmax (exemplary shown for NE = 1000). In panel (D) Average firing rate hνiii. For a constant external current the rate increases with increasing fraction of excitatory connections (black crosses, for Ii ∈ {2.25, 2.5, 2.75, 3.0, 3.25, 3.5, 3.75, 4.0, 4.25, 4.5}) . The neurons' firing rate stays almost constant, if we reduce the external current linearly with the number of excitatory neurons (blue crosses). We determined the value of Ii at the intersection point of the hνiii vs. NE curves with the desired frequency by linear interpolation. The values (I i, NE) that give rise to the desired frequency lie in good approximation on a straight line with slope −k ≈ −0.052 (cf. Inset to (D)). nPn i=1 ln(a−1 i The result is particularly remarkable since in mean-field descriptions of balanced networks, as long as the mean input to each cell is the same, the regime where NE = 0 is comparable to the regime where NE > 0 with appropriately reduced external excitatory current Ii. Up to now, we considered δ-coupling, where the response to an action potential is instantaneous. However, in In the following, we investigate biological neuronal systems the postsynaptic current has finite temporal extent. Temporally extended interactions A ) φ ( U B ) φ ( U C ∆i (n) (n+1) ∆i phaseφ (n+1) ∆ i (n)∆ i phaseφ 15 4 iI 2 0 NE 12000 7.5 3 − 0 1 / ) n ( λ 3.5 1 EN = 10000 n / 107 5 3 − 0 1 / x a m λ 4 2 0 −2 D 0.28 i i > ν < 0.24 0.20 0 12000 number of excitatory connections N 4000 8000 E 0 4000 8000 number of excitatory connections N 12000 E Figure 10: Destabilizing effect of excitation. (A) Simultaneous receiving of an inhibitory input decreases the phase shift. (B) Simultaneous receiving of an excitatory input increases the phase shift. (C) The largest Lyapunov exponent is measured for an increasing fraction of excitatory neurons starting with the network of Fig. 1(A-C) and ending with the network of Fig. 1(D-F). Here we keep the firing rate constant. Inset shows exemplarily the convergence of a finite time Lyapunov exponent with n. (D) Average firing rate hνiii versus number of excitatory connections in the network for different external currents Ii (black crosses, values belonging to the same Ii are connected by a dashed line). The neurons' firing rate stays almost constant, if we reduce the external current linearly with the number of excitatory neurons (blue crosses). The inset displays the current strength employed to maintain firing rate of hνiii ≈ 0.23. (Further details see text.) the influence of such temporally extended interactions. The analysis gets more complicated, because neurons are permanently influenced by incoming signals. As mentioned above, in our model we assume that the neuron looses the information about previously received spikes when it reaches the threshold and is reset. We modify Eq. (1) by introducing a temporally extended interaction kernel g(t), such that the evolution of the membrane potential is given by d dt Vi = fi(Vi) + NXj=1Xk∈Z εi,jg(cid:0)t − ts jk − τij(cid:1) . (50) In the following analysis we consider single exponential couplings, g(t) = Θ(t) · βe−βt with time scales γ−1 > β−1 > 0, the time constant of the postsynaptic current is shorter than the membrane time constant. As an exemplary neuron model we study the leaky integrate-and-fire neuron, fi(Vi) = −γVi(t) + Ii, but the analysis can easily be extended to more complex neuron models and interaction kernels. Numerical simulations show that the stability of the dynamics is robust against introduction of synaptic currents with small temporal extent, but on increase of temporal extension a transition to chaos occurs. In Fig. 11(A), the largest Lyapunov exponent, λmax, in a random network is estimated in dependence of the decay time constant β−1 of the synaptic current. For small time constant β−1, the dynamics behaves similar to the dynamics with δ-pulse interactions, in particular it is stable, the largest Lyapunov exponent is negative. For increasing β−1 the temporal extension becomes more and more influential and there is a transition to an unstable, chaotic regime with positive largest Lyapunov exponent. We now study the linear stability properties analytically. We denote the last spiking time of neuron i before tn by t0(n, i) = max k∈Z ( ts ik ts ik ≤ tn) ; (51) at t = t0(n, i) the potential of neuron i was reset to zero. The solution of Eq. (50) together with the initial condition Vi(t0(n, i)) = 0 is then given between the nth and (n + 1)th network event by Vi,n(t) = Ii γ (cid:16)1 − e−γ(t−t0(n,i))(cid:17) + εijΘ(cid:0)tr NXj=1Xk∈Z β − γ β ijk − t0(n, i)(cid:1) Θ(cid:0)t − tr ijk(cid:1)(cid:16)e−γ(t−tr ijk)(cid:17) , ijk) − e−β(t−tr (52) ijk = ts jk + τij is the reception time of the spike sent at ts where tr jk by neuron j at neuron i. The sum in Eq. (52) takes into account all spikes which are received by neuron i between t0(n, i) ≤ t ≤ tn and therefore influence the potential Vi,n(t). In the limit of very short temporal extension of the postsynaptic current, β → ∞, Eq. (52) becomes a solution of Eq. (1). After the nth event neuron i would reach the threshold at some time t′ under the assumption that there are no further inputs after tn. According to Eq. (52), t′ is implicitly given by 16 where A is the vector of the original event times tn, . . . , tn−W , VΘ,i − Vi,n(A, t = t′) = 0, A :=(cid:0) tn , . . . , tn−W (cid:1) , where we introduced W = maxn {W (n)}. We now estimate the effect of a small perturbation ∆tn, . . . , ∆tn−W of the event times tn, . . . , tn−W on the hypothetical event time t′. By Eq. (53), the Jacobian of t′, Dt′, with respect to former spike times, tn, . . . , tn−W , is given as The linearized estimation of the displacement ∆t′ of t′ is then given by (A) , . . . , ∂t′ ∂tn−W (A)(cid:19) = −(cid:18) ∂Vi,n ∂t (A, t′)(cid:19)−1 · DVi,n (A, t′) . The special structure of Vi,n(t) (cf. Eq. (52)), more precisely the fact that Vi,n(t) depends on t via t − tk for k ∈ {n − W, . . . , n}, yields the identity (A, t′)(cid:19)−1 · nXk=n−W(cid:18)− ∂Vi,n ∂tk (A, t′)(cid:19) · ∆tk. ∆tn ... ∆tn−W ∂t  =(cid:18) ∂Vi,n nXk=n−W ∂Vi,n ∂tk − (A, t′) = ∂Vi,n ∂t (A, t′) . ∂tn Dt′ (A) =(cid:18) ∂t′ = Dt′ (A) · . ∆t′ (53) (54) (55) (56) (57) (58) (59) (60) Under the condition, ∂Vi,n ∂tk (A, t′) ≤ 0 for all k = n − W, . . . , n, we can combine Eq. (56) and Eq. (57) and find bounds for the the displacement min k={n−W,...n} ∆tk ≤ ∆t′ ≤ max k={n−W,...n} ∆tk. Condition (58) implies that if neuron i sends or receives a spike earlier, also the threshold is crossed earlier. This always holds for δ-couplings, for interactions with temporal extend it restricts the class of patterns as we show below. Eq. (59) is an analog to Eq. (25), sufficiently small perturbations stay bounded by the initial ones for finite times. This directly implies Lyapunov stability for periodic orbits. For general irregular dynamics and to prove asymptotic stability, the propagation of pulses through the network has to be studied as for the nonlinear stability analysis in the main part. We now want to specify a class of periodic patterns which are stable in a network with temporally extended synaptic currents. The influence of various events on Vi,n(A, t′) is as follows: For an influential spike receiving tk, Eq. (52) yields ∂Vi,n ∂tk (A, t′) = β β − γ ε∗(cid:16)γe−γ(t′−tk) − βe−β(t′−tk)(cid:17) , where ε∗ < 0 is the coupling strength from the sending neuron. For the last spike sending of neuron i, tk = t0(n, i), For any other event tk a displacement of tk has no influence on Vi,n(t), here ∂Vi,n ∂t0(n, i) (A, t′) = −I ie−γ(t−t0(n,i)) < 0. ∂Vi,n ∂tk (A, t′) = 0. (61) (62) A λ 12 max −3 6 10 0 −6 0 −1β 0.3 time 8 C n o r u e n 5 1 0 17 time 8 n o r u e n 5 1 B 0 t ∆ 2 1 0 T +d τ 0 time t 300 t ∆ 1 0.5 0 0 d τ (T + )/2 time t 150 (A) Largest Figure 11: Robustness and transition to chaos with increasing temporal extent of the postsynaptic current. Lyapunov exponent λmaxin a random network (N = 50, Pre(i) ≡ 10, γ = 1, Ii ≡ 4.0, Pj εij ≡ −3.3,τij ≡ 0.1T free ) versus time constant β−1 of the synaptic current. (B) Stable periodic pattern in a network with temporally extended interactions. (C) Time compressed, thus unstable pattern in a network with temporally extended interactions. (For details see text.) i Therefore condition (58) reduces to a condition on the left and right hand side of Eq. (60) and can be reformulated as t′ − tk > 1 β − γ ln(cid:18) β γ(cid:19) := Td, (63) where tk are the spikes arrival times at neuron i since the last reset t0(n, i). This means that the class of patterns where each neuron i does not cross the threshold for a time period Td after receiving a spike are stable. For β → ∞ the system tends to the δ - pulse coupled system and indeed Td vanishes, limβ→∞ Td = 0, such that any non-degenerated orbit is stable. However, for temporally extended interactions unstable periodic orbits exists and also chaotic dynamics is possible (cf. Fig. 11(A)). To illustrate our analytical findings, in Fig 11(B,C), we used a generalization of a recently introduced method (Memmesheimer and Timme, 2006a,b) to design two networks realizing predefined spike patterns in a network of five neurons (VΘ,i ≡ 1.0, Ii ≡ 2.4, β = 8, τ := τij ≡ 0.125) with temporally extended couplings. Both patterns are the same, but with different inter-spike-intervals. In (B) all spikes are separated by ∆T = T d + τ , which ensures that a neuron never spikes within a time period Td after receiving a spike; in (C) we choose the inter-spike-intervals smaller ∆T = (Td + τ ) /2. The lower panels illustrates the stability properties: The spike times of the different neurons are plotted relative to the spike time of neuron 1 in vertical direction. The horizontal direction is simulated time, different colors indicate spike times of the five different neurons. At certain points in time (blue arrows) the network dynamics is perturbed. The dynamics in (B) is stable: After perturbations of size ≈ 0.2 (maximum norm), the dynamics converge towards the periodic orbit. (C) The dynamics is unstable: a perturbation of size ≈ 10−12 leads to a divergence from the unstable periodic orbit. IV. DISCUSSION Irregular spiking activity that robustly arises in balanced state models, constitutes a generic feature of cortical dynamics. Here, for a class of models including e.g. the common leaky integrate-and-fire neuron, we have shown that generic trajectories which give rise to the irregular balanced state can be exponentially stable. In particular, they are stable in purely inhibitory strongly connected networks of neurons with delayed couplings and with infinitesimal synaptic time course. We numerically illustrated and refined our analytical results. For small sparse networks we showed that the dynamics even converges to a periodic orbit. However, the length of the irregular yet stable transient grows rapidly with network size such that for larger networks, in particular for biological relevant sizes, transients dominate the dynamics on all relevant time scales. Furthermore, we found that the phenomenon of stable yet irregular dynamics is robust against introducing some excitatory interactions or against increasing the synaptic time scales from zero. If the synaptic responses become too slow or excitatory interactions too many, we revealed a smooth transition from stable irregular dynamics to chaotic, equally irregular dynamics. We emphasize that we kept the network rate during this transition (and thus keep the balance) and that the mean field descriptions (Brunel, 2000, v. Vreeswijk and Sompolinsky, 1998) of networks in both regimes are identical when the parameters are suitably chosen. Thus, highly irregular spiking dynamics occurs independent of the stability properties of the network. 18 Earlier studies on balanced neural activity considered a priori the limit of infinitely many neurons in sparse networks (Amit and Brunel, 1997, Brunel, 2000, Brunel and Hakim, 1999, v. Vreeswijk and Sompolinsky, 1996, 1998). In this mean field limit the collective dynamics is well understood. In particular in infinitely large networks of binary neurons with balanced excitatory and inhibitory interactions the dynamics are chaotic (v. Vreeswijk and Sompolinsky, 1996, 1998). Further studies of finite networks found stable dynamics in weakly diluted networks of inhibitory coupled neurons (Zillmer et al., 2006), as well as in globally coupled networks with dominating inhibition (Jin, 2002). Recent analytical evidence confirmed the existence of stable dynamics in inhibitory coupled networks of integrate-and-fire neurons with a more complex structure (Jahnke et al., 2008a). As the inter-event times that underly our analysis shrink inversely proportional to the network size (at a given individual neuron-spiking-rate), the methods applied here, however, are not applicable in a straightforward way in associated mean field models. Thus, one cannot make strict statements about stability in the limit of infinitely many neurons. Nevertheless, as shown above, generic transients and periodic trajectories in arbitrarily large inhibitory coupled networks are stable. Taken together, the results show that the microscopic dynamics in purely inhibitory coupled networks differs substantially from the dynamics of networks that explicitly include excitatory couplings. Whereas the latter is chaotic, the former is completely different and generically stable - despite both showing the same irregularity features. In particular, chaotic as well as stable dynamics are equally well capable of generating highly irregular spiking activity. The smoothness of the transition to chaos, without essential change of the irregularity (e.g. of the large coefficients of variation) further suggests that chaos is not the main dynamical origin of the high irregularity. We thus suggest that a mechanism different from chaos contributes to the irregularity of cortical firing patterns in a substantial way. Moreover, the location of the transition from chaos to stability as well as the dynamical mechanism underlying this transition, remain unknown. Nevertheless, chaos as well as stochastic network properties such as unreliable synapses, may support the robust occurrence of irregular activity in cortical networks and also modify its computational features. It is thus an important future task to investigate which anatomical and dynamical features of cortical networks are indeed of crucial relevance for their spiking activity and their functions. Acknowledgments We thank Fred Wolf for helpful comments, David Hansel for pointing out the article (Zillmer et al., 2009) and the Federal Ministry of Education and Research (BMBF) Germany for partial support under Grant No. 01GQ430. RMM thanks the Sloan Swartz Foundation for partial support. 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. Appendix In the following we will derive the bounds (26), (27) and (28) stated in section III B. Therefore, we will track the propagation of the perturbation of one specific neuron l0 through the entire network. All neurons spike at least once in a sufficiently large but finite time interval T . Moreover, after τmax = maxi,j (τij ) all spikes in transit have certainly arrived at the postsynaptic neurons. We label the maximal number of events possible in the time interval [t, t + max {T, τmax}] by M . For purely inhibitory networks, M < ∞ due to the bounded neural spike rate. We denote the set of postsynaptic neurons of l0 by Posti(l0) := Post ◦ Post ◦ . . . ◦ Post(l0) , (64) i times {z } thus a neuron li ∈ Posti(l0) is connected to l0 by a directed path of length i (cf. Fig. 12). Further, we define Post0(l0) := {l0}. We estimate the bounds of the perturbation following one specific path from a neuron l0 to a neuron l. In a strongly connected network, l ∈ Postj(l0) for some j ≤ N − 1, so there is a directed path between l0 and l = lj via neurons l1, . . . , lj−1 in the network. As the consideration holds for an arbitrary path, the result is an universal bound of the perturbation. Initially, at n = 0, the neurons are perturbed by ∆(0) . The first spiking of neuron l0 after n = 0 is i 19 1 2 3 events s0 r1 s1 r2 s2 r3 l0 l1 l2 l3 Figure 12: Tracking the propagation of a pulse. Here neuron l0 = 1 (black) is fixed as the initial neuron. The sets of postsynaptic neurons are: Post(1) = {2} (blue), Post2(1) = {1, 3} (green), Post3(1) = {1, 2} (red). Following one specific path through the network, we label the first spike event of neuron li ∈ Posti(l0) after receiving an input from li−1 ∈ Posti−1(l0) si. The event when the generated spike is received by li+1 is labeled ri+1. In a strongly connected network of size N the union SN −1 i=0 (cid:0)Posti(l0)(cid:1) = {1, . . . , N} contains all the neurons, because any two of them are connected by a directed path of maximal length N − 1. labeled by s0 ≤ M . After a delay time τl1l0 this spike is received by the postsynaptic neuron l1 ∈ Post(l0), we call the event r1 ≤ 2M . After at most M further events, at s1 ≤ 3M , the neuron l1 emits a spike. In general, we recursively define si as the first spiking event of neuron li ∈ Posti(l0) after ri and ri as the event when the spike generated by li−1 ∈ Posti−1(l0) at si−1 is received (cf. Fig. 12). Due to the definition of M , the relations si ≤ (2i + 1)M and ri ≤ 2iM hold. First we prove by induction that the perturbation of the neuron li before sending of a spike at si is bounded from above by ∆(si−1) li ≤h(1 − cmax)i · csi−1 min i ∆(0) l0 +h1 − (1 − cmax)i · csi−1 min i ∆(0) max. (65) 1. Initially neuron l0 is perturbed by ∆(0) l0 . Before l0 generates a spike it receives at most s0 − 1 inputs. According to Eq. (14), if the neuron l0 indeed receives an input the perturbation of neuron l0 may increase. To find an upper bound, we assume that at every event 0 < n < s0 neuron l0 receives an input with the maximal initial perturbation, ∆(0) max, and a minimal averaging constant cmin, which moves the average into the direction of the maximal possible perturbation. Repeated application of (14) yields ∆(1) l0 ∆(2) l0 ≤ cmin∆(0) l0 ≤ cmin∆(1) l0 + (1 − cmin) ∆(0) max, + (1 − cmin) ∆(0) max ≤ c2 min∆(0) l0 ∆(s0−1) l0 . . . min ∆(0) l0 ≤ cs0−1 +(cid:0)1 − cs0−1 min (cid:1) ∆(0) max. which is the inductive statement (65) for i = 0. +(cid:0)1 − c2 min(cid:1) ∆(0) max (66) 2. We assume that the statement (65) holds for ∆(si−1) , which is neuron li's perturbation as inherited by the spike sent at si (cf. Eq. (10)). After at most M events the spike is received by the postsynaptic neuron li+1 at event ri+1. In our worst- (or worse than worst-) case estimation, we assume that neuron li+1 is maximally perturbed before it receives the spike, ∆(ri+1−1) is maximal, cmax, such that again the average is moved into the direction of the maximal perturbation. Therefore the perturbation after the interaction is bounded by max, and that the interaction factor c(ri+1) = ∆(0) li+1 li+1 li ∆ (ri+1) li+1 ≤ cmax · ∆(0) max + (1 − cmax) ∆(si−1) li ≤ h(1 − cmax)i+1 · csi−1 min i · ∆(0) l0 +h1 − (1 − cmax)i+1 · csi−1 min i · ∆(0) max. (67) Before si+1 > ri+1 > si, neuron li+1 receives at most (si+1 − 1 − ri+1) inputs. Analogously to Eq. (66), we assume that with each event li+1 receives a spike which is maximally perturbed (with ∆(0) max), and the averaging constant is minimal, cmin. This yields ∆(si+1−1) li+1 ≤h(1 − cmax)i+1 · csi−1+si+1−1−ri+1 min i ∆(0) l0 +h1 − (1 − cmax)i+1 · csi−1+si+1−1−ri+1 min max. i ∆(0) (68) We replace csi−1+si+1−1−ri+1 0. This directly yields the induction statement for ∆(si+1−1) by csi+1−1 min min . li+1 in Eq. (68), thereby increasing the right-hand side, because si−1−ri+1 < 20 Based on Eq. (65) we now derive an upper bound of the perturbation of all neurons after event sN −1. After this event every neuron has sent at least one spike which is influenced by the initial perturbation of neuron l0, because Posti(l0) contains all neurons of the network (cf. Fig. 12). After the in a strongly connected network the union sith event, neuron li can still receive spikes. Before the sN −1th event, taken as reference, it receives in the worst case scenario (sN −1 − si) inputs with maximal initial perturbation ∆(0) max and minimal averaging factor cmin. Using Eq. (65) we repeatedly apply Eq. (14) (sN −1 − si) times which leads to N −1Si=0 ∆(sN −1) li ≤ h(1 − cmax)i · csi−1+sN −1−si min i ∆(0) l0 +h1 − (1 − cmax)i · csi−1+sN −1−si min i ∆(0) max. The right-hand side increases with i, therefore the perturbation of an arbitrary neuron j ∈ {1, . . . , N } after sN −1 events is bounded from above by (69) (70) (71) ∆(sN −1) j ≤h(1 − cmax)N −1 · c(sN −1−1) min i · ∆(0) l0 +h1 − (1 − cmax)N −1 · c(sN −1−1) min i · ∆(0) max. At the sN −1th event there can be D′ spikes per neuron in transit which are, in the worst-case scenario, assumed to have the maximal perturbation. Due to their arrival after the sN −1th event, the perturbations of neurons can still increase. However, after sN −1 + M events all spikes generated before the sN −1th event have arrived at the corresponding postsynaptic neurons. Taking into account the arrival of these spikes using Eqs. (14,70) , we find an upper bound for the perturbation after sN −1 + M events, ∆(sN −1+M) j ≤h(1 − cmax)N −1 · c(sN −1−1+M) min i · ∆(0) l0 +h1 − (1 − cmax)N −1 · c(sN −1−1+M) min i · ∆(0) max. Due to the fact that generated spikes inherit a perturbation present in the phases at spike sending time, the bound (71) holds also for the perturbation of spikes generated after the sN −1th and before the (sN −1 + M )th event, because the bound (71) limits the maximal perturbation for all neurons between the sN −1th and the (sN −1 + M )th event. We conclude that the perturbations of the neurons and the spikes in transit after K := 2N M ≥ sN −1 + M events are bounded by ∆(K) j ≤h(1 − cmax)N −1 · c2N M−1 min i · ∆(0) l0 +h1 − (1 − cmax)N −1 · c2N M−1 min Therefore we find an upper bound for the maximal perturbation ∆(K) max after K events, ∆(K) max ≤ c∗ · ∆(0) l0 + [1 − c∗] · ∆(0) max. with 0 < c∗ := (1 − cmax)N −1 · c2N M−1 min ≤ (1 − cmax) · cmax ≤ 1/4, Similarly, we find a lower bound for the minimal perturbation after K events ∆(K) min ≥ c∗ · ∆(0) l0 + [1 − c∗] · ∆(0) min, 3/4 ≤ (1 − c∗) < 1. i · ∆(0) max. (72) (73) (74) (75) (76) by an estimation analogous to the one above, where only ∆(0) be replaced by "≥". We note, that we did not have to specify the perturbation ∆(0) l0 max has to be replaced ∆(0) min and the relation "≤" has to to derive this result. Amit, D., Brunel, N. (1997). Model of global spontaneous activity and local structured activity during delay periods in the cerebral cortex. Cereb. Cortex 7, 237–252. 21 Ashwin, P., Timme, M. (2005). Unstable attractors: Existence and robustness in networks of oscillators with delayed pulse coupling. Nonlinearity 18, 2035–2060. Braitenberg, V., Schüz, A. (1998). Cortex: Statistics and Geometry of Neuronal Connectivity. Springer, Berlin. Brette, R., Rudolph, M., Carnevale, T., Hines, M., Beeman, D., Bower, J. M., Diesmann, M., Morrison, A., Goodman, P. H., Harris, F. C., Jr. Zirpe, M., Natschlager, T., Pecevski, D., Ermentrout, B., Djurfeldt, M., Lansner, A., Rochel, O., Vieville, T., Muller, E., Davison, A. P., El Boustani, S., and Destexhe, A. (2007). Simulation of networks of spiking neurons: a review of tools and strategies. J. Comput. Neurosci. 23, 349–98. Brunel, N. (2000). Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. J. Comput. Neurosci. 8, 183–208. Brunel, N., Hakim, V. (1999). Fast global oscillations in networks of integrate-and-fire neurons with low firing rate. Neural Comput. 11, 1621–1671. Burkitt, A. (2006). A review of the integrate-and-fire neuron model: I. Homogeneous synaptic input. Biol. Cybern. 95, 1–19. Cessac, B., Vieville, T. (2008). On Dynamics of Integrate-and-Fire Neural Networks with Conductance Based Synapses. Fron- tiers in Comput. Neurosci. 2, 2–20 Hansel, D., Mato, G., Meunier, C., Neltner, L. (1998). On numerical simulations of integrate-and-fire neural networks. Neural Comput. 10, 467–483 Jahnke, S., Memmesheimer, R. M., Timme, M. (2008a). Stable irregular dynamics in complex neural networks. Phys. Rev. Lett. 100, 048102. Jahnke, S., Memmesheimer, R. M., Timme, M. (2008b). Chaos is unlikely to underly the irregularity of cortical dynamics. Frontiers in Comput. Neurosci.Doi: 10.3389/conf.neuro.10.2008.01.006. Jin, D. (2002). Fast convergence of spike sequences to periodic patterns in recurrent networks. Phys. Rev. Lett. 89, 208102. Kirst, C., Timme, M. (2009). How Precise is the Timing of Action Potentials? Frontiers in Comput. Neurosci. 3, 2–3 Kumar, A., Schrader, S., Aertsen, A., Rotter, S. (2007). The high-conductance state of cortical networks. Neural Comp. 20, 1–34. Memmesheimer, R. (2007). Precise spike timing in complex neural networks. Doctoral thesis, Georg-August-Universität zu Göttingen. Memmesheimer, R., Timme, M. (2006a). Designing complex networks. Physica D 224, 182–201. Memmesheimer, R., Timme, M. (2006b). Designing the dynamics of spiking neural networks. Phys. Rev. Lett. 97, 188101. Memmesheimer, R., Timme, M. (2009). Stable and unstable periodic orbits in complex networks of spiking neurons with delays (submitted). Mirollo, R., Strogatz, S. (1990). Synchronization of pulse-coupled biological oscillators. SIAM J. Appl. Math. 50, 1645–1662. Strogatz, S., Mirollo, R. (1993). Splay states in globally coupled Josephson arrays: Analytical prediction of Floquet multipliers. Phys. Rev. E 47, 220–227. Timme, M., Wolf, F. (2008). The simplest problem in the collective dynamics of neural networks: is synchrony stable? Nonlin- earity 21, 1579–1599. Timme, M., Wolf, F., Geisel, T. (2002). Coexistence of regular and irregular dynamics in complex networks of pulse-coupled oscillators. Phys. Rev. Lett. 89, 258701. Tuckwell, H. (1988). Introduction to theoretical neurobiology: Volume 2. Nonlinear and stochastic theories. Cambridge Univ. Press, Cambridge. v. Vreeswijk, C., Sompolinsky, H. (1996). Chaos in neuronal networks with balanced excitatory and inhibitory activity. Science 274, 1724–1726. v. Vreeswijk, C., Sompolinsky, H. (1998). Chaotic balanced state in a model of cortical circuits. Neural Comput. 10, 1321–1371. Vogels, T., Abbott, L. (2005). Signal propagation and logic gating in networks of integrate-and-fire neurons. J. Neurosci. 25, 10786–10795. Zillmer, R., Brunel, N., Hansel, D. (2009). Very long transients, irregular firing and chaotic dynamics in networks of randomly connected inhibitory integrate-and-fire neurons. Phys. Rev. E 79, 031909 Zillmer, R., Livi, R., Politi, A., Torcini, A. (2006). Desynchronization in diluted neural networks. Phys. Rev. E 74, 036203. Zumdieck, A., Timme, M., Geisel, T., Wolf, F. (2004). Long chaotic transients in complex networks. Phys. Rev. Lett. 93, 244103.
1706.05656
6
1706
2018-01-10T13:35:26
Lexical representation explains cortical entrainment during speech comprehension
[ "q-bio.NC", "cs.CL" ]
Results from a recent neuroimaging study on spoken sentence comprehension have been interpreted as evidence for cortical entrainment to hierarchical syntactic structure. We present a simple computational model that predicts the power spectra from this study, even though the model's linguistic knowledge is restricted to the lexical level, and word-level representations are not combined into higher-level units (phrases or sentences). Hence, the cortical entrainment results can also be explained from the lexical properties of the stimuli, without recourse to hierarchical syntax.
q-bio.NC
q-bio
Lexical representation explains cortical entrainment during speech comprehension Stefan L. Frank Centre for Language Studies, Radboud University Jinbiao Yang Institute of Brain and Cognitive Science, NYU Shanghai Abstract Results from a recent neuroimaging study on spoken sentence comprehension have been interpreted as evidence for cortical entrainment to hierarchical syntactic structure. We present a simple computational model that predicts the power spectra from this study, even though the model's linguistic knowledge is restricted to the lexical level, and word-level representations are not combined into higher-level units (phrases or sentences). Hence, the cortical entrainment results can also be explained from the lexical properties of the stimuli, without recourse to hierarchical syntax. 1 Introduction There is considerable debate on the precise role of hierarchical syntactic structure during sentence compre- hension, with some arguing that full hierarchical analysis is part and parcel of the comprehension process (e.g., [1, 2], among many others) and others claiming that more shallow [3] or even non-hierarchical [4, 5] processing is common. Ding, Melloni, Zhang, Tian, and Poeppel [6] recently presented evidence that cortical entrainment during speech perception reflects the neural tracking of hierarchical structure of simple sentences, which would support the view that hierarchical processing is inevitable. They had participants listen to sequences of linguistic material consisting of English monosyllabic words or Chinese syllables, presented at a fixed rate of one syllable every 250 ms (or at a slightly slower rate for English). Depending on the experimental condition, each four-unit subsequence contained linguistic units at one or two hierarchically higher levels, or lacked meaningful structure beyond the syllable. For example, part of the stimulus in the English four-word sentence condition could be "... dry fur rubs skin fat rat sensed fear ..." where a group of four consecutive words forms a sentence with the following hierarchical structure: 1 (1) S NP VP Adj N V N dry fur rubs skin Here, the labels Adj, N, and V stand for the syntactic categories adjective, noun, and verb, respectively; NP and VP represent noun phrase and verb phrase; and S denotes the complete sentence. Magnetoencephalography (MEG) signals were recorded from participants listening to such a word se- quence. Power spectra computed from these signals displayed peaks at precisely the presentation rates of words, phrases, and sentences. Ding et al. interpreted this finding as evidence for neural tracking of language at these three, hierarchically related levels. A similar result was obtained from Chinese sentences consisting of a two-syllable noun followed by a two-syllable verb or verb phrase. In this condition, the power spectrum showed peaks at the occurrence frequency of the syllable, the word/phrase, and the sentence; that is, at 4 Hz, 2 Hz, and 1 Hz, respectively. When the sentence-level structure was removed by presenting only the Chinese verbs and verb phrases, only the 4 Hz and 2 Hz peak remained; and in a shuffled-syllable condition without any consistent lexical or sentential structure, the power spectrum was reduced to a 4 Hz peak only. Finally, when sentences consisted of a monosyllablic verb followed by a three-syllable noun or noun phrase, there were significant peaks in the power spectrum at 4 Hz and 1 Hz but not at 2 Hz, as there was no longer any linguistic unit occurring at a 2 Hz rate. To summarize the Ding et al. results: Across conditions, spectral power peaks occur exactly at the presentation frequencies of the linguistic units at the various hierarchical levels present in the stimulus. However, this does not yet imply a causal relation between cortical entrainment and the processing of hierarchical syntactic structure. As we will demonstrate by means of a simple computational model that does not incorporate syntactic structures or any other linguistic knowledge beyond the word level, the same entrainment results can follow from lexical representation alone. Our model represents the stimuli from the Ding et al. experiments as sequences of high-dimensional nu- merical vectors. These are assigned by a distributional semantics model (trained on large amounts of English or Chinese text) such that two words that tend to occur in similar local contexts receive similar vectors. Consequently, similarities between vectors reflect similarities between the semantic/syntactic properties of the represented words. For example, the inanimate nouns "fur" and "skin" will have similar representa- tions, which will differ somewhat different from the vector for the animate noun "rat", which in turn will be quite distinct from the vector for the action verb "rubs". Psycholinguists have argued that information about linguistic distributions partly underlies knowledge of word meaning [7]. Distributional semantics is also increasingly influential in semantic theory [8, 9] and widely applied in computational linguistics where it yields state-of-the-art results in applications for natural language processing [10]. Vector representations for words account for experimental findings from psycholinguistics [11, 12] and are not unrelated to cortical representations: They have been shown to allow for the decoding of neural activity during single word com- 2 prehension [13] or narrative reading [14,15], and distances between vectors are predictive of neural activation during written and spoken language comprehension [16, 17]. We use these vectors differently here: Rather than comparing vector (distances) to neural activation, we compute power spectra directly over sequences of vectors, as Ding et al. do for recorded MEG signals. Vector sequences that represent the Ding et al. stimuli in the different conditions result in power spectra that are very similar to those from human participants. Hence, the cortical entrainment results need not be indicative of the detection or construction of hierarchical structures. 2 Methods 2.1 Materials All materials were taken directly from the Ding et al. experiments. For English, these were 60 four-word sentences with structure as in (1) except that in seven sentences, the first word was not an adjective but a numeral or possessive pronoun. All English words were monosyllabic. For Chinese, the written items provided by Ding et al. were converted into pinyin (a phonological representation of Mandarin Chinese) by Pypinyin (version 0.12.1; we replaced the package's word-pinyin dictionary with a larger one from ZDIC) after which the result was manually checked and corrected when needed. Converting the written characters to pinyin is essential because the original experiments used auditory stimuli presentation. The same pinyin syllable can correspond to many different written characters, making word co-occurrence patterns differ strongly between spoken and written Chinese. Consequently, a distributional semantics model can only adequately capture the lexical information in spoken stimuli if it is applied to the pinyin form. There were two sets of 50 four-syllable Chinese sentences in the Ding et al. experiments. Sentences in the first set consisted of a two-syllable noun and a two-syllable verb or verb phrase, for example "laoni´u g¯engd`ı"("Old cattle ploughs the field"). The second set of Chinese sentences consisted of a monosyllablic verb followed by a three-syllable noun or noun phrase, for instance, "zh¯eng gu`ant¯angb¯ao"("Braising the soup dumplings"). Following Ding et al., syllable strings containing only two-syllable verbs or verb phrases were constructed by taking the verb (phrase) parts from the [N V(P)] sentences. Furthermore, four-syllable sequences without any consistent word, phrase, or sentence structure were obtained by randomly reassigning syllables to the [N V(P)] sentences while retaining each syllable's position in the sentence. To summarize, there are five experimental conditions using the same stimuli as Ding et al.: one in English ([NP VP] sentences) and four in Chinese ([N V(P)] sentences, [V N(P)] sentences, verb (phrases) only, and shuffled syllable sequences). Depending on condition, each stimuli sequence is composed of either four-syllable sentences, two-syllable verbs or verb phrases, or individual syllables without further linguistic structure. 3 2.2 Representing lexical knowledge 2.2.1 Distributional semantics Vector representations of English and Chinese words were generated by the Skipgram distributional semantics model [20], which is a feedforward neural network with N hidden units and input/output units that each represent a word type. The network is exposed to a large amount of English or Chinese text and, for each word token, learns simultaneously to predict the five following words and to retrodict the five preceding words. Words that are paradigmatically related to one another tend to occur in similar contexts, resulting in similar weight updates in the network. Consequently, connection weights come to represent words such that they capture paradigmatic relations among the represented words. More precisely, after training, a word is represented by the N -dimensional weight vector of connections emanating from that word's input unit, and words with similar syntactic or semantic properties will have similar vectors. Words that occurred less than five times in the training corpus were excluded to reduce processing time and memory requirements, and because the distributional information for infrequent words is less reliable. For each language, we obtained twelve different sets of vectors (i.e., simulated twelve participants) by running the Skipgram model twelve times with hidden-layer size N randomly drawn from a normal distribu- tion with mean 300 and standard deviation 25, and then rounded to the nearest integer. Other parameters of distributional semantics model training were identical to those in [16]. 2.2.2 Training corpora To get representations of English words, the model was trained on the first slice of the ENCOW14 web corpus [21], comprising 28.9 million sentences with 644.5 million word tokens of 2.8 million types (token and type counts include punctuation, numbers, etc.). This is the same corpus that was used in earlier work to obtain word-vector distances that predict neural activation during sentence reading [16]. For Chinese, the model was trained on the Chinese Wikipedia full-text corpus (downloaded 1 November 2016). We used Wikipedia Extractor (version 2.66) to extract cleaned text from the downloaded Wikipedia XML. Traditional Chinese text was converted into simplified Chinese by OpenCC (version 0.42). Chinese is standardly written without explicit word boundaries but these are required by the distributional semantics model. Therefore, the Chinese corpus was segmented into words, using Jieba (version 0.38). Following this, the corpus was converted to pinyin as described under 'Materials' above but without manual checking. The resulting corpus comprised almost 898,000 articles with a total of 210.8 million pinyin word tokens of 2 million types. 2.2.3 Representing incomplete words The Chinese materials include multisyllabic words, for which cortical entrainment was crucially established at the syllabic rate. Capturing this in the model requires a vector representation at each syllable position. How- ever, distributional semantics models generate only vectors that represent complete words. The construction of syllable-level representations is loosely based on the Cohort model of spoken word recognition [22]: The syllable sequence s1, . . . , sn (possibly containing only one syllable) activates all words that begin with that 4 sequence; this set of words is called the cohort. A word has to occur at least 5 times in the Wikipedia corpus to be considered part of the cohort. The vector at syllable position n, representing the sequence s1, . . . , sn, equals the average vector of words in the cohort, weighted by the words' corpus frequencies. The cohort becomes empty when s1, . . . , sn does not form the beginning of any word, in which case syllable sn starts a new cohort. Note that this method can be applied to the sentence sequences as well as the shuffled syllable sequence because it does not depend on (knowledge of) word boundaries. Qualitatively identical results were obtained with a slightly alternative scheme in which cohorts are not included at word-final position, that is, the representation at each word-final syllable equals the single vector for that word. 2.3 Lexical information over time In Ding et al.'s Chinese experiments, syllables come in at a fixed rate of 4 Hz, or every 250 ms. The English experiments used a slightly slower presentation rate, but for simplicity we model English and Chinese experiments using the same 4 Hz rate. Let v = (v1, . . . , vN ) be the N -dimensional column vector that represents the English word or Chinese syllable sequence currently being presented. We assume that the lexical information does not immediately appear at word onset (t = 0 ms) but some time later, at τ ≥ 0. The value of τ is randomly sampled at each word/syllable presentation, from a uniform distribution with mean µ = 40 and width β = 50 (how this choice of parameter values came about is discussed below). The available lexical information at t milliseconds after word onset (for 0 ≤ t ≤ 250) is represented by a column vector w(t) = (w1(t), . . . , wN (t)) with wi(t) =   εi(t) if t < τ vi + εi(t) if t ≥ τ where εi(t) denotes Gaussian noise with mean 0 and standard deviation σ = 0.5. Hence, the lexical infor- mation vector w(t) starts with representing only noise but at t = τ the information in v becomes available (if τ < 0, it becomes becomes available immediately, at t = 0). In practice, we discretize continuous time t into 5 ms bins, corresponding to the 200 Hz low-pass filter frequency applied by Ding et al., so that there are 50 time steps between two syllable onsets. All vectors w for the stimuli sequence of an experimental condition are concatenated into a single matrix W that captures the entire session's time sequence, with a different random order of trials for each of the twelve simulated participants. This matrix has N rows and 50 columns per syllable of the stimulus sequence. The values of the model's three free parameters (µ, β, and σ) were chosen to obtain results on the English sentences that were visually similar to those of Ding et al., using a different set of word vectors (see Appendix D). 2.4 Analysis Following Ding et al., we applied a Discrete Fourier Transform to obtain a power spectrum for each experi- mental condition and each simulated participant. The individual rows of matrix W, each representing the 5 time course in a single dimension of word vector space, were transformed to the frequency domain. Next, these per-dimension power spectra were averaged over the N dimensions to obtain the power at each fre- quency bin. Frequency bin width was 1/9 Hz for Chinese and 1/11 Hz for English, as in Ding et al. Again following Ding et al., we tested whether the power at each frequency bin significantly exceeded the average of the previous and next two bins using one-tailed t-tests with false discovery rate correction [23]. 3 Results Figure 1 displays the model predictions for all five conditions, side by side with the original results from Ding et al. Ignoring differences in scale, the two sets of power spectra are strikingly similar. For English four-word [NP VP] sentences, the model predicts peaks in the power spectrum at the presentation rate of words (4 Hz), phrases (2 Hz), and sentences (1 Hz). Results for Chinese four-syllable [N V(P)] sentences look very similar to those for English and, crucially, to those of Ding et al.: Peaks occur at the presentation rates of syllables, words/phrases, and sentences. The minor peaks at 3 Hz in both conditions are also visible in the Ding et al. results, although it only reaches significance for the English sentences. Although the 3 Hz peak in the English condition is significant in our reanalysis of Ding et al.'s data, it is not in their original analysis. Possibly, this is because because we applied a less conservative false-discovery rate correction method. The 3 Hz peaks most likely occur as the second harmonic of the 1 Hz signal [18] so they do not reflect any interesting property of the input. The 2 Hz and 4 Hz peaks do not merely arise as harmonics because they remain when only the verbs and verb phrases of the Chinese [N V(P)] sentences are presented, even though the power spectrum lacks the 1 Hz peak in this condition. This is the case both in the model predictions and the neural data. When the syllables are randomly reassigned to sentences, breaking up the word and sentence structure, only the 4 Hz peak should remain. In the model, the 2 Hz and 3 Hz peaks are indeed no longer present and the 1 Hz peak has almost completely vanished: The peak size (defined as peak power minus the average power of the previous and next two frequency bins) is significantly reduced compared to the [N V(P)] sentences (paired t-test: t(11) = 26.0; p < .00001). Finally, when the stimuli sequence consists of sentences with a one-syllable verb followed by a three- syllable noun or noun phrase, no linguistic unit occurs at a 2 Hz rate. The model results in Figure 1 show that the 2 Hz peak is indeed strongly reduced compared to the [N V(P)] condition (paired t-test: t(11) = 30.8; p < .00001). The corresponding results from Ding et al. also show a small 2 Hz peak in this condition, although it fails to reach significance. However, note that individual differences between simulated participants are much smaller than between human participants; Hence, very small model effects can reach significance whereas they are washed out by noise in the corresponding human data. Most likely, the small 2 Hz peaks are merely the first harmonic of the 1 Hz signal. 6 r e w o P r e w o P r e w o P r e w o P r e w o P * 1 * Human * 10 dB 2 * 3 Frequency (Hz) * 10 dB 1 2 3 Frequency (Hz) * 10 dB 1 2 3 Frequency (Hz) 10 dB 2 3 Frequency (Hz) * 3 10 dB 2 1 * 1 * 4 * 4 * 4 * 4 * 4 r e w o P r e w o P r e w o P r e w o P r e w o P * 1 * 1 1 * 1 * 1 Model English [NP VP] * 3 * 4 * 2 5 dB Frequency (Hz) Chinese [N V(P)] * 2 5 dB * 2 5 dB * 3 3 Frequency (Hz) Chinese V(P) * 4 * 4 * 4 * 4 Frequency (Hz) Chinese shuffled 5 dB 2 3 Frequency (Hz) Chinese [V N(P)] 5 dB * 2 * 3 Frequency (Hz) Frequency (Hz) Figure 1: Power spectra from human MEG signal (left) and corresponding model predictions (right) in all five conditions. Shaded areas in the human results represent standard errors from the mean over eight subjects. Grey lines in the model results depict individual model runs (simulated subjects). Blue lines are the averages over (simulated) subjects. Statistically significant peaks (P < .025; one-tailed) are indicated by asterisks. In the top left panel (human results on English stimuli) the frequency scale has been adjusted to match the simulated presentation rates. 7 4 Discussion For all five stimuli types, our model predicts power spectra that are qualitatively almost identical to those from Ding et al.'s MEG study. Only the very small but statistically significant 1 Hz peak in the Chinese shuffled syllable sequence condition was not present in the human results. Apparently, some aspects of the syllables still recur at a 1 Hz rate, albeit very weakly. This is in fact not surprising considering that the syllables kept their original position in the sentence. A possible explanation for the difference between model prediction and human data is that the model treats the shuffled syllable sequence and grammatical sentences exactly the same, in that at each point all possible word candidates are selected (see under 'Representing incomplete words' in the Methods section). In contrast, human participants are likely to forgo pro-active word activation when listening to non-word sequences, even though their task in this condition was to detect the occasional correct sentence. Crucially, however, the predicted 1 Hz peak was much reduced compared to the [N V(P)] sentence condition, which shows that shuffling the syllables affects the model predictions in the same direction as the MEG power spectra. 4.1 Lexical versus structural accounts of the Ding et al. results The model's only linguistic representations are formed by the distributional semantics vectors. Although these are learned from word strings (see Methods), they do not explicitly encode information about word sequences (such as transitional probabilities) and can therefore not be used to predict, for example, that a noun is likely to be followed by a verb [16]. More importantly, a word's vector does not depend on its position or neighbors in the stimulus sentence, and the model lacks higher-level representations of phrases and sentences. This, of course, raises a crucial question: What is the origin of the predicted power peaks at the presentation rates of phrases and sentences? The word vectors represent lexical properties by virtue of the fact that similarities between vectors mirror paradigmatic relations between words: Words that share more syntactic/semantic properties are encoded by more similar vectors. Consequently, if certain lexical properties occur at a fixed rate in the stimulus sequence, this will be reflected as a recurring approximate numerical pattern in the model's time-series of vectors. For example, in Ding et al.'s English four-word sentences condition, every other word is a noun, most often referring to some entity, and every forth word is a transitive verb, usually referring to an action. Because semantically and syntactically similar words are represented by similar vectors, the vector sequence corresponding to this experimental condition shows spectral power peaks at exactly the occurrence rates of two-word phrases and four-word sentences. Crucially, this does not rely on any hierarchical structure or process: Vectors represent only lexical information and the spectral power analysis is applied over a sequence of vectors that is not processed, integrated, or interpreted. Note, however, that whether a word is a noun or verb (or, in terms of semantics: refers to an entity or action) depends on its role within the sentence. In the stimulus sentence "fat rat sensed fear", for example, the individual word "fat" could be a noun instead of an adjective and "fear", on its own, could be a verb instead of a noun. The vectors do not distinguish between the different senses of these ambiguous words: There is only one representation of "fat". Nevertheless, there is apparently enough repetition in the stimulus word's properties to account for the MEG results. In fact, we 8 Full word salad Partial word salad r e w o P 5 dB 1 2 3 Frequency (Hz) * 4 r e w o P * 1 * 2 5 dB * 3 * 4 Frequency (Hz) Figure 2: Model predictions on word salad conditions. Stimuli sequences were constructed by randomly drawing (with replacement) words from the English [NP VP] sentences. Left: all words randomly drawn. Right: adjectives and verbs keep their original position in the [NP VP] stimuli. obtained qualitatively similar results when each word was represented by a vector that merely identifies its most frequent syntactic category, that is, independently of the word's role in the sentence (see Appendix A). The model shows how Ding et al.'s cortical entrainment results can be explained without recourse to hierarchical structures or integrative processes. However, this does not rule out that cortical entrainment to hierarchical structure exists or even that the Ding et al. results in fact do reflect processing of hierarchical syntax. Indeed, it has recently been demonstrated that a hierarchical sentence processing model predicts similar power spectra, at least on English materials [19]. Whether lexical or phrasal/sentential properties of the speech signal are in fact responsible for cortical entrainment can possibly be established by testing on a 'word salad' version of the English stimuli sequence, created by keeping all verbs and adjectives in place but replacing all nouns by random words from different syntactic categories. This breaks any (consistent) syntactic structure so no 1 Hz peak should be visible if its occurrence depends on the repeated presence of 4-word sentences. In contrast, our lexical model does predict 1 Hz power in this word-salad condition because both verbs and adjectives still occur at a 1 Hz rate (see Figure 2). Returning to the stimuli sequences tested by Ding et al., our model demonstrates that hierarchical syntax is not necessary to explain these entrainment results: Representing only lexical properties of the stimuli suffices. Compared to structural accounts (such as [19]), our lexical explanation has the advantage of parsimony. This is because constructing a sentence's hierarchical structure requires information about the words' possible syntactic categories (i.e., whether a word can be a noun, verb, or adjective); or in semantic terms: Understanding a sentence requires at least some knowledge of word meaning. Thus, lexical properties are necessary in a structural explanation of the entrainment results. Conversely, however, the vector representations in our model do not depend on the sentence in which the words happen to appear. Hence, the lexical model forms a more parsimonious account. 4.2 Word vectors as cortical representations The vector representations are not intended to be neurally realistic, that is, we do not claim they are isomorphic to cortical representations. Rather, the relation between model and brain is a higher-order one: The rhythmic recurrence of patterns in a vector sequence corresponds to the patterns in the MEG signal 9 caused by perceiving the stimuli represented by the vectors. Ding et al. report two additional experiments that probe the cortical representations more directly. First, they show that overall MEG activity strongly decreases around phrase and sentence boundaries. Second, they find a spatial dissociation between the cortical areas that show a phrasal- or sentential-rate response. Although there was no reason to expect our vectors to account for these findings too, we did find that vector lengths show a sharp drop right after phrase and sentence boundaries (Appendix B) and that (apparent) effects of phrases and sentences can be localized in different vector dimensions (Appendix C). Hence, the correspondence between the distributional semantic vectors and cortical representations may be stronger than anticipated. Acknowledgments We are very grateful to Nai Ding for sharing his data and to Jona Sassenhagen, Tal Linzen, and two anonymous reviewers for their comments on earlier versions of this paper. The research presented here was funded by the Netherlands Organisation for Scientific Research (NWO) Gravitation Grant 024.001.006 to the Language in Interaction Consortium. References [1] Berwick RC, Friederici AD, Chomsky N, Bolhuis JJ. Evolution, brain, and the nature of language. Trends in cognitive sciences. 2013;17:89–98. [2] Hale JT. What a rational parser would do. Cognitive Science. 2011;35:399–443. [3] Sanford AJ, Sturt P. Depth of processing in language comprehension: Not noticing the evidence. Trends in Cognitive Sciences. 2002;6:382–386. [4] Frank SL, Bod R, Christiansen MH. How hierarchical is language use? Proceedings of the Royal Society B: Biological Sciences. 2012;279:4522–4531. [5] Jackendoff R, Wittenberg E. Linear grammar as a possible stepping-stone in the evolution of language. Psychonomic Bulletin & Review. 2017;24:219–224. [6] Ding N, Melloni L, Zhang H, Tian X, Poeppel D. Cortical tracking of hierarchical linguistic structures in connected speech. Nature Neuroscience. 2016;19:158–164. [7] Andrews M, Vigliocco G, Vinson DP. Integrating experiential and distributional data to learn semantic representations. Psychological Review. 2009;116:463–498. [8] Baroni M, Bernardi R, Zamparelli R. Frege in space: A program for compositional distributional semantics. Linguistic Issues in Language Technologies. 2014;9:5–110. [9] Coecke B, Sadrzadeh M, Clark S. Mathematical foundations for distributed compositional model of meaning. Linguistic Analysis. 2010;36:345–384. 10 [10] Bellegarda JR, Monz C. State of the art in statistical methods for language and speech processing. Computer Speech & Language. 2016;35:163–184. [11] Mandera P, Keuleers E, Brysbaert M. Explaining human performance in psycholinguistic tasks with models of semantic similarity based on prediction and counting: A review and empirical validation. Journal of Memory and Language. 2017;92:57–78. [12] Rotaru AS, Vigliocco G, Frank SL. Modelling the structure and dynamics of semantic processing. Cognitive Science. in press;. [13] Mitchell T, Shinkareva SV, Carlson A, Chang KM, Malave VL, Mason RA, et al. Predicting human brain activity associated with the meanings of nouns. Science. 2008;320:1191–1195. [14] Wehbe L, Murphy B, Talukdar P, Fyshe A, Ramdas A, Mitchell T. Simultaneously uncovering the patterns of brain regions involved in different story reading subprocesses. PLoS ONE. 2014;9:e112575. [15] Wehbe L, Vaswani A, Knight K, Mitchell T. Aligning context-based statistical models of language with brain activity during reading. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. Doha, Qatar: Association for Computational Linguistics; 2014. p. 233–243. [16] Frank SL, Willems RM. Word predictability and semantic similarity show distinct patterns of brain activity during language comprehension. Language, Cognition and Neuroscience. 2017;32:1192–1203. [17] Li J, Brennan J, Mahar A, Hale J. Temporal Lobes as Combinatory Engines for both Form and Meaning. In: Proceedings of the Workshop on Computational Linguistics for Linguistic Complexity (CL4LC). Osaka, Japan: The COLING 2016 Organizing Committee; 2016. p. 186–191. [18] Zhou H, Melloni L, Poeppel D, Ding N. Interpretations of frequency domain analyses of neural en- trainment: periodicity, fundamental frequency, and harmonics. Frontiers in Human Neuroscience. 2016;10:274. [19] Martin AE, Doumas LAA. A mechanism for the cortical computation of hierarchical linguistic structure. PLoS Biology. 2017;15:e2000663. [20] Mikolov T, Chen K, Corrado G, Dean J. Efficient estimation of word representations in vector space. In: Proceedings of the ICLR Workshop; 2013. [21] Schafer R. Processing and querying large web corpora with the COW14 architecture. In: Ba´nski P, Biber H, Breiteneder E, Kupietz M, Lungen H, Witt A, editors. Proceedings of the 3rd Workshop on the Challenges in the Management of Large Corpora; 2015. p. 28–34. [22] Marslen-Wilson WD. Functional parallelism in spoken word-recognition. Cognition. 1987;25:71–102. [23] Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society B. 1995;57:289–300. 11 r e w o P * 1 * 2 5 dB * 3 * 4 Frequency (Hz) Figure 3: Power spectra resulting from processing English [NP VP] sentences with each word replaced by its most frequent syntactic category. A Representing words as syntactic categories Under the model's explanation of the Ding et al.'s results, similar outcomes should be obtained if vectors represent only the syntactic categories of the stimulus words. The question remains if this can be done without interpreting the stimuli beyond the individual words, because for many words the syntactic category (also known as part-of-speech; POS) is ambiguous until the word is understood in its sentence context. As a first test of the feasibility of a 'POS only'-account, we replaced each word of the English [NP VP] stimuli by its most frequent POS in the ENCOW corpus (i.e., without considering the word's role in the sentence), yielding 13 different POS tags. Each was assigned a vector such that all 13 vectors are orthogonal, that is, they identify POS without encoding any notion of similarity between syntactic categories. As much as possible without sacrificing orthogonality, vector values were randomly drawn from the original vectors (i.e., representing words), again for 12 simulated participants. The resulting frequency spectra (Figure 3) again shows the power peaks at 1, 2, and 4 Hz. This suggests that it is possible that POS representations underlie the MEG findings. However, note that this result should not be taken as evidence for the cognitive or neural representation of syntactic categories (let alone for the particular set of POS tags in the ENCOW corpus) because similar outcomes were obtained using the original vector representations in which words only approximately cluster by POS. B Vector lengths over the course of a sentence Ding et al. presented participants with 90 Chinese sentences consisting of a 3- or 4-syllable NP followed by a 4- or 5-syllable VP. Overall MEG activity decreased strongly near the NP and sentence boundaries, which was taken as further evidence for neural tracking of phrasal and sentential structure. We obtained vector representations for pinyin versions of the same stimuli and took each vector's euclidean length as a measure for overall 'activation'. Figure 4 plots the average length as a function of syllable position in the sentence. There are indeed sharp drops after NP and sentence (VP) boundaries, corresponding to the human results. This is most likely caused by the fact that syllable strings that cross a phrase boundary rarely form a possible word beginning. Consequently, the first syllable of a phrase often starts a new cohort, and averaging over its many words yields a relatively short vector. 12 4.5 4 3.5 3 2.5 h t g n e l r o t c e v e g a r e v A N 1 N 2 N 4 V −4 N V −2 3 Syllable position V −3 V −1 V 0 Figure 4: Average euclidean length of vector representations at each point of 7- and 8-syllable Chinese [NP VP] sentences. Error bars indicate 95% confidence intervals. Plots are left aligned on the 3- or 4-syllable NP (labeled N1 to N4) as well as right aligned on the 4- or 5-syllable VP (labeled V−4 to V0). Sentence with a 3-syllable NP have no N4 and 4-syllable NP sentences have no V−4. C Localization of sentential- and phrasal-rate responses In an electrocorticography (ECoG) study using the English four-word [NP VP] sentences, Ding et al. identi- fied the electrodes that selectively respond at the phrasal or sentential rate but not at the word-presentation rate. Across these electrodes, power at the phrasal and sentential frequencies were negatively correlated, which "demonstrates spatially dissociable neural tracking of the sentential and phrasal structures" (p. 162). For each individual dimension of vector representations for the same stimuli sequences, we computed the peak size at each frequency as the power at that frequency minus the average over all frequencies within 0.5 Hz on either side. These peak size were then expressed as z-scores relative to the range from 0.5 to 4.5 Hz. Next, we identified vector dimensions that show a sentential- or phrasal-rate pattern but no word- rate pattern by selecting those with z > 2 at 1 or 2 Hz and z < 1 at 4 Hz. Across the 12 simulated subjects, 10.0% of vector dimensions met these criteria. On these dimensions, peak sizes at 1 and 2 Hz were negatively correlated (r = −.43; p < .00001) which shows that the lexical properties represented by the vectors are dissociated across vector dimensions in a way that corresponds to the the spatial dissociation between sentential- and phrasal-rate responses in Ding et al.'s ECoG experiment. D Setting model parameters The model has three free parameters: • µ: The average arrival time of lexical information, in simulated milliseconds from word onset • β: The width of the uniform distribution of arrival time of lexical information • σ: The standard deviation of Gaussian noise Appropriate values for these parameters were found by getting a new set of word vectors (with N = 300) for the English [NP VP] sentences. The objective was to find µ, β, and σ such that the frequency spectrum 13 resembles the corresponding results from Ding et al. in this condition, that is, there are peaks at 1 Hz, 2 Hz, and 4 Hz, possibly a minor peak at 3 Hz, no other peaks, and a slight increase in power for lower frequencies. Figure 5 shows the power spectra for each combination of µ ∈ {25, 75, 125, 175, 225}, β ∈ {0, 25, 50, 75, 100}, and σ ∈ {0, .25, .5, .75, 1}. The three major peaks at 1 Hz, 2 Hz, and 4 Hz, and the minor peak at 3 Hz, are clearly visible for most combinations of parameter values. Only for very high levels of noise σ or when both µ and β are high (which means that much lexical information never arrives) are the power spectra mostly flat. These results motivated us to fix β at 50 and perform a more fine-grained search through parameter space for low-to-medium levels of µ and σ. Based on visual inspection of the resulting power spectra in Figure 6, we settled for µ = 40 and σ = 0.5. The parameters are not over-fitted, as is clear from the fact that the same combination of values resulted in very similar power spectra using another set of vectors as well as on the Chinese sentences (Figure 1). 14 β = 0 β = 25 β = 50 β = 75 β = 100 5 2 = µ 5 7 = µ 5 2 1 = µ 5 7 1 = µ 5 2 2 = µ 2 1 4 Frequency (Hz) 3 3 2 1 4 Frequency (Hz) σ = 0 σ = 0.25 3 2 1 4 Frequency (Hz) σ = 0.5 3 2 4 1 Frequency (Hz) σ = 0.75 σ = 1 2 4 1 Frequency (Hz) 3 Figure 5: Power spectra resulting from processing English [NP VP] sentences, for different combinations of parameter values. 15 σ = 0 σ = 0.1 σ = 0.2 σ = 0.3 σ = 0.4 σ = 0.5 0 4 = µ 0 6 = µ 0 8 = µ 0 0 1 = µ 0 2 1 = µ 2 1 Frequency (Hz) 3 4 2 1 Frequency (Hz) 3 4 2 1 Frequency (Hz) 3 4 2 1 Frequency (Hz) 4 3 2 1 Frequency (Hz) 4 3 2 1 Frequency (Hz) 4 3 Figure 6: Power spectra resulting from processing English [NP VP] sentences, for β = 50 and different values of µ and σ. 16
1709.00339
1
1709
2017-08-31T11:14:59
How images determine our visual search strategy
[ "q-bio.NC", "cond-mat.other" ]
When searching for a target within an image our brain can adopt different strategies, but which one does it choose? This question can be answered by tracking the motion of the eye while it executes the task. Following many individuals performing various search tasks we distinguish between two competing strategies. Motivated by these findings, we introduce a model that captures the interplay of the search strategies and allows us to create artificial eye-tracking trajectories, which could be compared to the experimental ones. Identifying the model parameters allows us to quantify the strategy employed in terms of ensemble averages, characterizing each experimental cohort. In this way we can discern with high sensitivity the relation between the visual landscape and the average strategy, disclosing how small variations in the image induce changes in the strategy.
q-bio.NC
q-bio
How images determine our visual search strategy Tatiana A. Amor,1, 2 Mirko Lukovi´c,1 Hans J. Herrmann,1, 2 and Jos´e S. Andrade, Jr.2, ∗ 1Computational Physics IfB, ETH Zurich, Stefano-Franscini-Platz 3, CH-8093, Zurich, Switzerland 2Departamento de F´ısica, Universidade Federal do Cear´a, 60451-970, Fortaleza, Cear´a, Brazil Abstract When searching for a target within an image our brain can adopt different strategies, but which one does it choose? This question can be answered by tracking the motion of the eye while it executes the task. Following many individuals performing various search tasks we distinguish between two competing strategies. Motivated by these findings, we introduce a model that captures the interplay of the search strategies and allows us to create artificial eye-tracking trajectories, which could be compared to the experimental ones. Identifying the model parameters allows us to quantify the strategy employed in terms of ensemble averages, characterizing each experimental cohort. In this way we can discern with high sensitivity the relation between the visual landscape and the average strategy, disclosing how small variations in the image induce changes in the strategy. Keywords: Visual Search Eye Movement Search Strategies 7 1 0 2 g u A 1 3 ] . C N o i b - q [ 1 v 9 3 3 0 0 . 9 0 7 1 : v i X r a ∗ Correspondence: [email protected] 1 INTRODUCTION What visual strategy do we employ when searching for a familiar face in a crowd and how would it change if we had to find a friend in a well organized choir? For instance, do we explore each face sequentially or do we pick them randomly until reaching our target. Along these lines, it is not at all obvious whether there exists a characteristic strategy that is related to the scene content. To shed light into this problem it is necessary to find a method that allows us to identify and quantify particular features associated to the strategy adopted while searching for a hidden target. Different models have been developed with the goal of understanding what guides eye movement during visual search (see Ref. [1] for a review). One family of these models is based on the construction of saliency maps, which define regions of interest derived from properties of scene objects (such as luminescence, color, orientation) [2 -- 7]. By definition these salient regions stand out from other parts of the scene and are therefore more susceptible to frequent eye fixations. In the context of visual search, these models prove to be more suitable for tasks involving a small number of equally relevant distractors, which are essentially all items in the scene that are not targets of the search [8, 9]. On the other hand, in more complex visual search tasks, the salient regions might not necessarily be relevant. In order to address this issue, other implementations of the saliency model consider the relative information of an object with respect to the global information of the scene [10 -- 12]. In the absence of salient elements, it is not possible to use these models as there are no a priori privileged regions within the scene. As a consequence, another family of visual searching models have been proposed, namely, the saccadic targeting models [13 -- 17]. In these studies, the main hypothesis is that saccadic eye movements are directed to locations within the scene that contain elements similar to the target. This similarity can be due to the image content as well as to a neurobiological filter. Within this framework, Najemnik & Geisler propose a model where each point in space has a certain probability of being explored and the saccadic movement is directed to the most probable regions [18, 19]. These probabilities are then updated over time so that regions that were already explored are less likely to be revisited, introducing in a natural way the notion of persistence while searching. Moreover, other implementations of this model have taken into account the proximity between consecutive saccadic movements by adding a cost function that punishes 2 longer saccades [20]. Scanpaths produced by eye movements in general reveal different forms of persistence, at the level of saccades [21] and also within the fixations [22, 23]. During most forms of random search, persistence of the jittering movements (i.e. saccades in visual search) plays a crucial role since it unveils the strategy involved [24, 25]. In fact, while looking for a hidden target in a field of distractors, a variety of patterns have been documented [21, 26 -- 28], ranging from systematic or completely persistent to random. However, while some of the models described above do make use of some form of persistence, they do not relate it to the overall strategy of the search. Beside some exceptions [29, 30], the saliency models do not focus on understanding the sequence in which the fixations are performed but instead determine the regions where they are more likely to occur. On the other hand, the saccadic targeting models, although dealing with the saccadic sequence, do not account for different strategies. Here we propose a visual search model (VSM) that quantifies the global persistent be- havior and the overall strategy employed while looking for a hidden target. The parameters of the VSM define the saccadic orientation distribution, which has experimentally proven to provide information regarding the strategy [21] and the scene content [27]. By studying this distribution obtained from experimental data, we were able to identify different search strategies that emerge from exploring the same scene and quantify them through the VSM parameters. We analyzed the search strategies adopted in three different visual search scenes that differ in the features and arrangement of the scene items. We found that the average strategy changes with these scenes suggesting that scene content and structure influence the way subjects execute their search. In what follows we first describe the VSM and how to extract the model parameters from the experimental scanpaths in the "cloud number (CN)" experiment. In this visual search task the participants are requested to find a unique number "5" embedded in a field of numbers ranging from "1" to "9" serving as distractors. By means of an efficiency measure about the eye paths, we validate our model with respect to the experimental data. Then, we identify the location of the experimental trajectories for three different visual search tasks in the model parameter space and compare the average strategies applied for each task. Given that the average position within the parameter space differs for the three cohorts, the VSM can serve as a tool to predict the average strategy of similar experiments. Finally, we discuss the implications of our results in the frame of visual search and the possibilities of exploiting 3 the model to be used in other fields. FIG. 1. Different types of visual trajectories are found in the "Cloud number (CN)" experiment. During a visual searching task, where participants have to find a unique number "5" embedded in a cloud of numbers between "1" and "9", participants perform very different trajectories. These trajectories range from a systematic search (Left) to a trajectory resembling more a random search with a strong persistence imprint (Right). VISUAL SEARCH MODEL (VSM) While searching for a hidden target during a visual search episode, participants perform different search strategies that vary from a very systematic to a seemingly random, as depicted in Fig. 1. These strategies are identified by studying the relative orientation between saccadic movements performed while searching. The saccadic relative angle distribution provides information on the type of strategy employed [21] and may as well provide insight regarding the scene an observer is looking at [27]. Therefore, the aim of the VSM is to emulate visual search trajectories via the study of the distribution of the inter-saccadic angles as main ingredient. We sample the relative angle, θ, between two consecutive saccadic movements from two distributions, as exemplified in Figs. 2.A,B. The first distribution corresponds to a uniform distribution of all possible angles between 0 and 2π (see Fig. 2.A), and the second one is a uniform distribution with angles restricted to the interval [−δθ, δθ] (see Fig. 2.B). Angles 4 FIG. 2. Schematic definition of the Visual Search Model (VSM). Two parameters, λ and δθ govern the way to select the relative angles between two consecutive saccadic movements. We consider two uniform distributions: one over all possible angles (A) and one in the range [−δθ, δθ] (B). With probability 1 − λ angles are chosen from the first distribution and with probability λ from the second one. Different combinations of these parameters result in a variety of distributions (D) responsible for different visual search strategies, ranging from a pure random walk (C) to a systematic search (E). are chosen from the second distribution with probability λ, and from the first distribution with probability 1 − λ. Therefore, only two parameters, λ and δθ, govern the statistical properties of the relative angle distribution. The combination of these parameters gives rise to a wide range of different trajectories. In the case λ = 0, angles can only be sampled from the first distribution, corresponding to a purely random strategy, (see Fig. 2.C). On the other hand, if λ = 1, the angle is only sampled from the step distribution, so that only angles in the range [−δθ, δθ] are allowed. If in the latter case we further assume that δθ 5 P(θ)0πP:1-λP(θ) -δθ δθ0πP:λ+ABCDEP(θ) -δθ δθ0P(θ)0π -δθ δθπP(θ)0πλ=1and δθ≠πλ=0 for any δθ0<λ<1 and δθ≠π λ=0λ=1, δθ= 0.004λ=0.6, δθ= 0.3 is very small, so that each saccadic movement persists in the direction of the previous one, then the emergent trajectory will correspond to a systematic type of search (see Fig. 2.E). Between the two extreme cases, we find a repertoire of trajectories that result from the interplay between the parameters λ and δθ (see Fig. 2.D). The parameter λ is related to the strategy employed during the search and δθ to the persistent behavior of the search. For instance, an intermediate value of λ combined with a small value of δθ results in a trajectory that can be identified as a random search with a strong persistence imprint. Besides the relative angle distribution, another important ingredient in our model is the length of the saccadic movements, i.e the length of the jumps. We consider them to be sampled from a log-normal distribution, as reported in Ref. [26]. Estimation of λ and δθ from the experimental data. From the VSM we propose the inter-saccadic relative angles, θ, to be sampled from a combination of a uniform distribution over all angles and one in the interval [−δθ, δθ] (see Fig. 2). A conventional method to acquire the parameters from experimental data would be to study the histogram of θ. However, one issue of this approach is that the parameter δθ is bin-size dependent and thus, is not a robust method for the parameter estimation. Moreover, the average number of θ values obtained in each trajectory is around 250, thus making it difficult to determine with precision the distribution from each of the scanpaths. One way around this issue is to extract the model parameters directly from the experimental data. When the angles are ranked in increasing order from −π to π, they present a characteristic shape of three linear segments, as shown in Fig. 3.A with green dots. This shape is similar for all the trajectories (see Fig.S1). In the extreme case that θ is sampled from the uniform distribution between −π and π, the ranked angles form a straight line with a slope equal to 2π. On the other hand, if θ is sampled from the uniform distribution between −δθ and δθ, the ranked angles form a straight line where θmax− θmin = 2δθ. Considering the evident threefold behavior of the ranked angles obtained from the experiments and its origin as a combination of these three extreme cases, we fit the data using three consecutive linear segments (see Fig. 3.A, solid orange lines). The middle region depends on δθ, and the first and third regions are related to the uniform distribution between −π and π. The estimation of δθ is straightforwardly performed as the difference between the angle values defining the middle section equals 2δθ. The parameter λ is the ratio between the number of angles originated by the oriented distribution, Nλ, and the total 6 number of angles, N . Within this framework, Nλ corresponds to the difference between the total number of angles and nstep, defined as the angle index value for which the extension of the first linear segment is equal to π. This allows us to estimate the parameters of the VSM using the experimental data without the need of any intermediate manipulation. RESULTS In order to compare the VSM with the experimental data we study the efficiency, ε, defined in terms of the space-filling attributes of the trajectories investigated (similar to the method applied in Ref. [31]) The underlying hypothesis is that if the trajectory is able to fill the image space then the chances of finding the target increases, as more locations are examined. The efficiency is measured in the following way: given a trajectory of a finite amount of steps, N , we divide the figure being explored with a grid of m× m cells. Next, we determine the percentage of cells that were explored by the trajectory, as shown in Fig. 3.B. When a fixation point falls into a cell, this cell is considered to be explored, since it is during fixational events that most visual information is gathered [32]. On the other hand, if a saccadic movement passes over a cell but there are no fixations inside it, the cell is considered unexplored. For a fixed length N of the trajectory, we have different mean ε values within the param- eter space. An efficient trajectory is the one that visits many locations, i.e has a large value for ε. The efficiency values obtained in the parameter space are depicted in Fig. 3.C. As expected, the values corresponding to λ = 0 and any δθ are the same as the ones obtained for δθ = π and any λ value, since these trajectories emerge from the same distribution. The optimal values are those having a large λ value and a small δθ. These are typical of a systematic type of search, in which very few locations are inspected again. Therefore, the final trajectory explores more locations as compared to the random strategy. From the study of the relative inter-saccadic angle distribution of the CN cohort, we estimate the corresponding parameters λ and δθ. When placed in the parameter space, we see that the experimental data presents a dispersion in the parameter λ ranging from 0.1 to 0.8 (see Fig. 3.C). Interestingly, the estimated experimental values for δθ are confined within a range related to small angles, therefore, associated to a persistent movement. We find that the center of mass of these data points, i.e the mean values, is close to 0.4 for λ, 7 FIG. 3. Comparison of the VSM with the experimental eye paths. (A) Estimation of λ and δθ from the experimental trajectories. We rank θ (circles) for each experimental trajectory and perform a fit of three linear segments (solid lines). The limits of the middle region (yellow shaded region) define δθ. The difference between the extended line of the first region and the end of the third region (green shaded region) defines the amount of angles, Nλ, originated by the distribution restricted by δθ. The parameter λ is defined as the ratio between Nλ and the total number of angles N . (B) Definition of space filling efficiency, ε. ε corresponds to the percentage of visited locations (gray cells) after parceling the image into m × m cells. In this case, m = 10. (C) Mean efficiency values obtained from the VSM parameters space for 100 realizations of the model. The experimental data from the CN cohort (white dots) are placed on their corresponding λ and δθ values. (D) Comparison between ε values obtained with the VSM and the experimental ones. For each trajectory (i.e black points) we run the model with the same number of jumps and the estimated λ and δθ values. The shaded area corresponds to the standard deviation of 100 realizations of the model. The data presented in (A) and the data point highlighted in (C) correspond to the trajectory depicted in (B) with length ∼ 140. and to π/8 for δθ. 8 ε=0.64B3.142.541.941.340.740.14δθ0.00.20.40.60.81.0λCRandomSystematicRandomPersistent0.600.620.640.660.680.700.720.74ε05101520253035Trials0.00.20.40.60.81.0εDVSMCloud number020406080100120140Angle Index3210123θ2δθNλλ=Nλ/NAExp. DataSegment Fit For the estimated values of λ and δθ, we run the VSM with the corresponding values of N for each trajectory and compare the obtained ε values. The VSM matches 78% of the experimental data, as shown in Fig. 3.D. For some points, however the model fails to predict the expected ε. For instance, one participant initiates with a systematic search and later breaks away from that strategy to go for a systematic search truncated in space, i.e without reaching the right end of the image (see Fig. S2). Our model is not able to reproduce this behavior and, therefore, for the same set of parameters, λ and δθ, it generates instead a more efficient search. Interestingly, for the CN cohort we do not find an average behavior that is purely sys- tematic, namely a λ value close to unity. As a consequence, one could ask whether the average searching behavior changes for different searching tasks. We analyzed two other experimental cohorts, namely, the "5-2" and "Where's Wally? (WW)" sets, as reported in Ref. [26] (See Mat. and Methods). The 5-2 experiment consists in finding a target, namely, a single number "5", in a regular array of distractors, namely, numbers "2", that can be in red or green. The WW experiment corresponds to scanning images from the famous children's book where the goal is to find "Wally", a character wearing a striped shirt in a crowd of people [33]. A visual search task as the 5-2 experiment, where the target and distractors are placed on a regular lattice, might evoke predominantly systematic search strategies as the regularity of the lattice over which the items are placed serves as a guide to the eye. If indeed this is the case, this cohort should be characterized in the parameter space with large λ values and small δθ. On the other hand, an experiment as WW, which is composed by a very complex scene with no apparent regularities, might evoke a random strategy, as there is no preferential searching direction. In the parameter space this would be achieved with a small λ value and, in principle, any value of δθ. Within this hypothesis the CN experiment can be viewed as an intermediate configuration between the aforementioned experiments, as the items are positioned on a lattice but with random displacements. We compute the parameters from the 5-2 and WW cohorts (see Fig.S3) and validate the VSM to the experimental data (see Fig.S4). The location of all the trajectories composing the experimental cohorts in parameter space are depicted in Fig. 4.A. The large variability across the λ parameter appears within subjects as well as within trials performed by the same subject in a searching episode (see Fig. S5). However, even though there is variability present in each of these visual tasks, it is possible to identify a mean strategy employed 9 FIG. 4. Different tasks are placed in distinctive positions in the parameter space. (A) Three visual search experimental cohorts going from numbers placed in a regular lattice to a random position exhibit distinctive λ values. The shaded area shows where the experimental data falls and the circle represents the center of mass of each group. (B) The 5-2 experiment presents a median λ value of 0.6 (green), the CN experiment, 0.4 (red), and the WW experiment, 0.2 (blue). (C) For the median value of δθ there is no clear distinction and all experiments are confined to the range [0,π/4]. by each experimental cohort. The 5-2 experiment presents a median λ value of about 0.6, whereas for the WW experiment it becomes about 0.2, as shown in Fig. 4.B. The values for δθ are always confined within [0,π/4], as shown in Fig. 4.C. By combining the efficiency results from the VSM with the location of the experimental data in parameter space we realize that the 5-2 cohort is the most efficient one. The other trajectories are relatively less efficient. This raises the question whether the efficiency ana- lyzed here, as well as the criteria reviewed in the literature, such as the overall time to find a target, are adequate when it comes to compare different visual search tasks. Through the study of the inter-saccadic relative angle distribution we were able to quan- tify the average strategies employed in three visual searching experiments and catalog them according to their λ and δθ values. The average strategy changes along the different cohorts but always evoking persistent searches (small δθ values), Fig. 4.C. One could argue that the 10 0.00.51.01.52.02.53.0δθ0.00.20.40.60.81.0λWhere'sWally?(WW)5-2CloudNumber(CN)A5-2CNWW0.00.20.40.60.81.0λB5-2CNWW0.00.51.01.5δθC discrepancy within the average strategy employed appears as a consequence of the amount of distractors present in the image. However, the 5-2 experiment was performed with differ- ent amounts of distractors and significant differences between the average λ values have not been observed (See Fig. S6). This suggests that the difference between the average strategies appears as a consequence of the scene content and structure. DISCUSSION Our work focuses on the study of ocular patterns that emerge during a visual search task. Through a simple model containing two parameters we show that the knowledge of the saccadic relative orientation distribution gives the relevant information on the visual trajectories found in experiments. The validation of the model with experimental scanpaths through an indirect efficiency measurement confirms this fact. Moreover, the distribution of the experimental data points over the parameter space shows that indeed there is a large variability between the different visual strategies employed while searching, even for the same subject performing the same experiment. Nevertheless, we are able to discriminate changes in the average strategy across different scenes. This is shown through the analysis of the model parameters λ and δθ, which suggests that the visual strategy employed is linked to the underlying structure of the scene, even when the cognitive task remains the same, i.e visual search. The experimental cohorts presented in this work correspond to participants performing visual searches in different scenarios. Even though these scenarios are easily distinguishable from one another, they have the following common properties: a) the scenes are static, no moving items were analyzed in these experiments; b) the scenes have no salient regions, therefore there are no a priori privileged regions to explore in the image; c) the task always involves finding a unique target hidden in a set of distractors; e) the distractors have the same size as the target and, f) all distractors are equal, in principle, there are no distractors less relevant than others. The layout of the items (from a regular lattice to a crowded space) and the features associated to the distractors are the main sources of differences between the experiments. In the 5-2 experiment, the items can either be a number "5" or a number "2". On the other hand, for the WW experiment, the items presented in the scene, although they always involve people, present a variety of shapes and colors that make them richer in 11 visual content, thus, making the scene more complex. The difficulty of the task in which the participant is engaged while performing visual search has a strong effect on the number of items that can be processed during a fixation [34, 35]. Following this idea, it has been suggested that what determines a visual task to be easy or difficult is linked to the discriminability of the target [36, 37]. Accordingly, we can arrange our set of experiments in the following order of increasing difficulty: 5-2, CN and WW. In this way, our results indicate that a difficult task is more likely to be executed in a random fashion with a relative degree of persistence, whereas an easy task evokes a more systematic search. An intermediate difficulty task, such as the CN experiment, leads to a mixed type of strategy within the aforementioned extremes. Our method is simple enough to be used as a tool to further understand or diagnose mental disorders. For instance, a recent study investigated how the visual patterns, while performing different free viewing tasks, changes for healthy control versus schizophrenic patients [38]. The authors presented a set of scenes, with different levels of image complexity, for the participants to freely explore and recorded the position of theirs eyes while doing so. They found that both groups reduced the area of exploration as the images become less complex, while schizophrenic participants tend to maintain the same type of scanpath for all the scenes. Therefore, it would be interesting to measure the parameters λ and δθ over these trajectories and study how they change from one group to the other. We believe, based on the reported trajectories in Ref. [38], that the average λ should change for the healthy control group, whereas for the schizophrenic group it should remain constant along the different levels of image complexity. Our findings and methods can be also applied to other types of search such as way-finding, where similar patterns appear for people looking for a target in a confined or open space [39]. Furthermore, within the framework of foraging and random walks in general, we propose a very simple model that exhibits persistent behavior similar to the effect of "wind", namely, a directional bias in a random walk search [40]. Additionally, our model comes with the advantage of being able to quantify the type of search by means of two parameters which can be easily determined from experimental data. In summary, we were able to use the parameter space of the model to correlate a particular scene with a distinct average strategy. With this new relationship between visual strategies and scene content we can predict the average strategy applied in a new experiment by 12 comparing the new scene structure to the ones discussed here. As a perspective work, it would be also interesting to extend our findings to search tasks that involve moving objects, or perhaps to general visual tasks where the sequence of fixations is relevant. Furthermore, given that the scenes are shown on a 2D display with well defined boundaries, it would be worth investigating whether the systematic search strategies remain once the boundaries are removed. MATERIALS AND METHODS Both experimental cohorts were conducted in the Universidade Federal do Cear´a with partici- pants who had normal or corrected to normal vision. The experimental data corresponding to the CN cohort is published in Ref. [21] and the one corresponding the 5-2 and WW cohort, in Ref. [26]. The study has been approved by the Ethics Research Committee of the Universidade Federal do Cear´a (COMEPE) under the protocol number 056/11. All methods used in this study were carried out in accordance with the approved guidelines and all experimental protocols were approved by COMEPE. Informed consent was obtained from all subjects. "Cloud number (CN)" cohort. Experimental data was recorded with an EyeLink 1000 system (SR Research Ltd., Mississauga, Canada), with an acquisition frequency of 1kHz on a monocular recording over ten subjects (Mean age: 24). Each subject carried out a sequence of four trials, each one with a maximum duration of five minutes. In each trial we presented the participant an image with numbers randomly distributed in a 1024 × 1280px image, where the goal was to find a unique number "5" within 1499 distractors. Between each trial, the subject had the possibility of relaxing and before starting the recording we performed a new calibration. At the beginning of each trial, the participant was asked to fixate his/her eyes on the center of the screen, in case a drift correction needed to be performed. The classification of the fixation and saccades was made using the EyeLink online filter [41, Section 4.3]. Fixations in the EyeLink system are identified using a saccade-pick algorithm. The system analyzes the moment-to-moment velocity and acceleration of the eye using fixed thresholds for both eye velocity and acceleration. If the eye goes above either the velocity or acceleration threshold, the start of a saccade is marked. Analogously, when both the velocity and the accelera- tion drop back below their thresholds, the algorithm identifies the saccade end. By default, every 13 movement which does not lie within this definition is considered as being part of a fixation. The saccade velocity threshold was set to be 30◦/s, the saccade acceleration threshold, 8000◦/s2, and the saccade motion threshold, 0.15◦. "5-2" and "Where's Wally? (WW)" cohorts. Eye movements were recorded with a Tobii T120 eye-tracking system (Tobii Technology), over 11 subjects (Mean age: 23). The stimuli were presented on a 17 TFT-LCD monitor with resolution 1024 × 1280px and acquisition frequency of 60Hz. The 5-2 experiment consists on finding a number "5" within an array of numbers "2" serving as distractors. All numbers (target and distractors) are positioned on a square lattice and are randomly colored red or green, hindering the visual detection of the target through the identification of patterns on the peripheral vision. The number of distractors present in each task is related to a degree of difficulty: DF0 (207 distractors), DF1 (857 distractors) and DF2 (1399 distractors). The maximum time given to search the target was 1, 1.5 and 2 minutes respectively. The WW experiment consists in finding "Wally", the famous character from the series of books with the same name [33], who is hidden within a very complex background of crowded characters, with a maximum searching time of 2 minutes. As explained in Ref. [26] the identification of fixations and saccades was carried out with a modified version of the fixation filter developed by Olsson [42]. Data availability. The datasets analyzed in the current study are available from the corresponding author upon reasonable request. Acknowledgments. The authors would like to thank H.F.Credidio for making available the experimental data corresponding to the 5-2 and WW cohorts. This work was supported by the Brazilian agencies CNPq, CAPES, FUNCAP, and National Institute of Science and Technology for Complex Systems in Brazil. We acknowledge financial support from the European Research Council (ERC) Advanced Grant 319968-FlowCCS. Author contribution. T.A.A., M.L., H.J.H. and J.S.A. designed research; T.A.A. performed the experiments; T.A.A. and M.L. analyzed data; T.A.A., M.L., H.J.H. and J.S.A. performed research; T.A.A., M.L., H.J.H. and J.S.A. wrote the paper. 14 Competing financial interests. The authors declare no conflict of interest. [1] M. P. Eckstein, J. Vis. 11, 14 (2011). [2] L. Itti, C. Koch, and E. Niebur, IEEE Trans. Pattern Anal Mach. Intell. , 1254 (1998). [3] L. Itti and C. Koch, Vision Res. 40, 1489 (2000). [4] E. Over, I. Hooge, B. Vlaskamp, and C. Erkelens, Vision Res. 47, 2272 (2007). [5] T. Foulsham and G. Underwood, J. Vis. 8, 6 (2008). [6] K. A. Ehinger, B. Hidalgo-Sotelo, A. Torralba, and A. Oliva, Vis. cogn. 17, 945 (2009). [7] K. Nakayama and P. Martini, Vision Res. 51, 1526 (2011). [8] Z. Li, Trends Cogn. Sci. 6, 9 (2002). [9] R. M. Shiffrin and G. T. Gardner, J Exp Psychol 93, 72 (1972). [10] L. Zhang, M. H. Tong, T. K. Marks, H. Shan, and G. W. Cottrell, J Vis 8, 32 (2008). [11] N. D. Bruce and J. K. Tsotsos, J. Vis. 9, 5 (2009). [12] D. Gao, S. Han, and N. Vasconcelos, IEEE Trans. Pattern Anal Mach. Intell. 31, 989 (2009). [13] R. P. Rao, G. J. Zelinsky, M. M. Hayhoe, and D. H. Ballard, Vision Res. 42, 1447 (2002). [14] B. R. Beutter, M. P. Eckstein, and L. S. Stone, JOSA A 20, 1341 (2003). [15] M. P. Eckstein, B. A. Drescher, and S. S. Shimozaki, Psychol. Sci. 17, 973 (2006). [16] M. Pomplun, Vision Res. 46, 1886 (2006). [17] G. J. Zelinsky, Psychol Rev. 115, 787 (2008). [18] J. Najemnik and W. S. Geisler, Nature 434, 387 (2005). [19] J. Najemnik and W. S. Geisler, J. Vis. 8, 4 (2008). [20] C. Araujo, E. Kowler, and M. Pavel, Vision Res. 41, 3613 (2001). [21] T. A. Amor, S. D. Reis, D. Campos, H. J. Herrmann, and J. S. Andrade, Sci. Rep. 6, 20815 (2016). [22] R. Engbert, K. Mergenthaler, P. Sinn, and A. Pikovsky, Proc. Natl. Acad. Sci. USA 108, E765 (2011). [23] C. A. Marlow, I. V. Viskontas, A. Matlin, C. Boydston, A. Boxer, and R. P. Taylor, PLoS One 10, e0139379 (2015). [24] E. A. Codling, M. J. Plank, and S. Benhamou, J. R. Soc. Interface 5, 813 (2008). [25] V. Tejedor, R. Voituriez, and O. B´enichou, Phys. Rev. Lett. 108, 088103 (2012). 15 [26] H. F. Credidio, E. N. Teixeira, S. D. Reis, A. A. Moreira, and J. S. Andrade, Sci. Rep. 2 (2012), 10.1038/srep00920. [27] O. Le Meur and A. Coutrot, Vision Res. 121, 72 (2016). [28] B. C. Motter and E. J. Belky, Vision Res. 38, 1805 (1998). [29] D. Brockmann and T. Geisel, Neurocomputing 32, 643 (2000). [30] G. Boccignone and M. Ferraro, Physica A: Statistical Mechanics and its Applications 331, 207 (2004). [31] R. Engbert and K. Mergenthaler, Proc. Natl. Acad. Sci. USA 103, 7192 (2006). [32] D. H. Hubel, J. Wensveen, and B. Wick, Eye, brain, and vision (Scientific American Library New York, 1995). [33] M. Handford, Wheres Wally? The Wonder Book (Walker Books, London, 1997). [34] B. C. Motter and D. A. Simoni, Vision Res. 48, 2382 (2008). [35] W. S. Geisler and K.-L. Chou, Psychol. Rev. 102, 356 (1995). [36] A. H. Young and J. Hulleman, J. Exp. Psychol. Hum. Percept. Perform 39, 168 (2013). [37] J. Hulleman and C. N. Olivers, Psychon. Bull. Rev. 21, 652 (2014). [38] J. I. Egana, C. Devia, R. Mayol, J. Parrini, G. Orellana, A. Ruiz, and P. E. Maldonado, Front. Psychiatry 4 (2013), 10.3389/fpsyt.2013.00037. [39] T. J. Pingel and V. R. Schinazi, Cartographic Perspectives , 33 (2014). [40] V. V. Palyulin, A. V. Chechkin, and R. Metzler, Proc. Natl Acad. Sci. USA 111, 2931 (2014). [41] S. EyeLink, EyeLink 1000 User Manual (2010). [42] P. Ollson, KTH Electrical Engineering (2007). 16
1102.5428
1
1102
2011-02-26T16:00:08
Reservoirs of Stability: Flux Tubes in the Dynamics of Cortical Circuits
[ "q-bio.NC", "cond-mat.dis-nn", "nlin.CD" ]
Triggering a single additional spike in a cerebral cortical neuron was recently demonstrated to cause a cascade of extra spikes in the network that is likely to rapidly decorrelate the network's microstate. The mechanisms involved in this extreme sensitivity of cortical networks are currently not well understood. Here, we show in a minimal model of cortical circuit dynamics that exponential state separation after single spike and even single synapse perturbations coexists with dynamical stability to infinitesimal state perturbations. We propose a unifying picture of exponentially separating flux tubes enclosing unique stable trajectories composing the networks' state spaces.
q-bio.NC
q-bio
Reservoirs of Stability: Flux Tubes in the Dynamics of Cortical Circuits Michael Monteforte∗ and Fred Wolf Max Planck Institute for Dynamics and Self-Organization 37073 Göttingen, Germany BCCN, BFNT and Faculty of Physics, University Göttingen, 37073 Göttingen, Germany Triggering a single additional spike in a cerebral cortical neuron was recently demonstrated to cause a cascade of extra spikes in the network that is likely to rapidly decorrelate the network's microstate. The mechanisms involved in this extreme sensitivity of cortical networks are currently not well understood. Here, we show in a minimal model of cortical circuit dynamics that exponential state separation after single spike and even single synapse perturbations coexists with dynamical stability to infinitesimal state perturbations. We propose a unifying picture of exponentially separating flux tubes enclosing unique stable trajectories composing the networks' state spaces. PACS numbers: 87.19.lj, 87.10.-e, 05.45.-a Understanding the dynamical characteristics of cerebral cortex networks is fundamental for the understanding of sen- sory information processing in the brain. Bottom-up inves- tigations of different generic neuronal network models have led to a variety of results ranging from stable [1 -- 3] to chaotic dynamics [4 -- 6]. In top-down attempts to construct classifica- tion and discrimination systems with such networks, the 'edge of chaos' was proposed to be computationally optimal [7]. Near this transition between ordered and chaotic dynamics, a network can combine the fading memory and the separa- tion property, both of which are important for the efficacy of computing applications [8]. While fading memory (informa- tion about perturbations of the microstate die out over time) is achieved by a stable dynamics, the separation property (dis- tinguishable inputs lead to significantly different macrostates) is best supported by a chaotic dynamics. Widely used in reservoir computing [8] and one of the most simple models of cortical circuits are networks of randomly coupled inhibitory leaky integrate and fire (LIF) neurons [9]. These networks exhibit stable chaos, characterized by stable dynamics with respect to infinitesimal perturbations despite an irregular network activity [1, 2]. They thus exhibit fading memory. Whether and how such networks realize the separa- tion property is however unclear. Motivated by the recent observation that real cortical net- works are highly sensitive to single spike perturbations [6], we examine in this letter how single spike and single synapse per- turbations evolve in a formally stable model of generic corti- cal circuits. We show that random networks of inhibitory LIF neurons exhibit negative definite Lyapunov spectra, confirm- ing the existence of stable chaos. The Lyapunov spectra are invariant to the network size, indicating that stable dynamics is representative for large networks, extensive and preserved in the thermodynamic limit. Remarkably, in the limit of large connectivity, perturbations decay as fast as in uncoupled neu- rons. Single spike perturbations induce only minute firing rate responses but surprisingly lead to exponential state separation causing complete decoherence of the networks' microstates within milliseconds. By examining the transition from un- stable dynamics to stable dynamics for arbitrary perturbation size, we derive a picture of tangled flux tubes composing the networks' phase space. These flux tubes form reservoirs of stability enclosing unique stable trajectories, whereas adja- cent trajectories separate exponentially fast. In the thermody- namic limit the flux tubes become vanishingly small, implying that even in the limit of infinitesimal weak perturbations the dynamics would be unstable. This contradicts the prediction from the Lyapunov spectrum analysis and reveals that charac- terizing the dynamics of such networks qualitatively depends on the order in which the weak perturbation limit and the ther- modynamic limit are taken. We studied large sparse networks of N LIF neurons ar- ranged on directed Erdös-Rényi random graphs of mean in- degree K. The neurons' membrane potentials Vi ∈ (−∞, VT) with i = 1 . . . N satisfy τm Vi = −Vi + Ii(t) (1) between spike events. When Vi reaches the threshold VT ≡ 1, neuron i emits a spike and Vi is reset to VR ≡ 0. The membrane time constant is denoted τm. The synaptic input currents are Ii(t) = √KI0 − J0√K τm Xj∈pre(i)Xs δ(t − t(s) j ), (2) composed of constant excitatory external currents √KI0 and inhibitory nondelayed δ pulses of strength −J0/√K, re- ceived at the spike times t(s) of the presynaptic neurons j ∈ pre(i). The external currents I0 were chosen to obtain a target network-averaged firing rate ¯ν. j Equivalent to the voltage representation, Eq. (1), is a phase representation in which each neuron is described by a phase φi ∈ [−∞, 1], obtained through φi = − ln((Vi − √KI0)/(VR − √KI0))/T free i = − ln((VT − √KI0)/(VR− √KI0)) is the interspike interval of an isolated neuron), with a constant phase velocity and the phase transi- tion curve U (φi) = − ln(exp(−φiT free ) + J0/(KI0))/T free describing the phase updates at spike reception. The neurons' phases thus evolve, from time ts−1 after the last spike in the network through the next spike time ts, at which neuron j ∗ i (where T free i i 30 (a) n o r u e n 1 2 0 V 0 (b) 0 0.2 t (s) 0.8 (c) t s d i 0 (d) t s d i ν (Hz) 40 1 1 0 (e) 20 cv 1 ) z H ( ν 0 0 ν bal (Hz) 20 1 χ 0 Figure 1: The balanced state in inhibitory LIF networks: (a) Asyn- chronous irregular spike pattern of 30 randomly chosen neurons, (b) fluctuating voltage trace of one neuron (voltage increased to V = 2 at spikes), (c),(d) broad distributions of individual neu- rons' firing rates ν and coefficients of variation cv, (e) network- averaged firing rate ¯ν and synchrony measure χ versus predicted rate ¯νbal = I0/(J0τm) (dotted line: guide to the eye for ¯ν = ¯νbal, χ = STD([φi]) [STD(φi)] where [·] denotes population average), (parameters: N = 10 000, K = 1000, ¯ν = 10 Hz, J0 = 1, τm = 10 ms). fires, according to the map i i φi(ts) =(φi(ts−1) + (ts − ts−1)/T free U (φi(ts−1) + (ts − ts−1)/T free (3) The neurons postsynaptic to the spiking neuron j ∗ are i∗ ∈ post (j ∗). We used the exact phase map (3) for numerically exact event-based simulations and to analytically calculate the single spike Jacobian D(ts) = ∂ ~φ(ts) ∂ ~φ(ts−1) for i 6= i∗ for i = i∗. ) : di∗ (ts) 1 − di∗ (ts) δij for i = j = i∗ for i = i∗, j = j ∗ otherwise. (4) Dij(ts) =  This matrix depends on the spiking neuron j ∗ and on the phases of the spike receiving neurons i∗ through the derivative s )) evalu- of the phase transition curve di∗ (ts) = ∂φU (φi∗ (t− ated at time t− s just before spike reception [10]. Describing the evolution of infinitesimal phase perturbations, the single spike Jacobians (4) were used for numerically exact calcula- tions of all Lyapunov exponents λ1 ≥ ··· ≥ λN in a standard reorthogonalization procedure [11]. As expected from the construction of the LIF networks, the dynamics converged to a balanced state. Figure 1 shows a representative spike pattern and voltage trace illustrating the irregular and asynchronous firing and strong membrane poten- tial fluctuations. A second characteristic feature of balanced networks is a substantial heterogeneity in the spike statistics across neurons, indicated by broad distributions of coefficients of variation (cv) and firing rates (ν). Independent of model de- tails, the network-averaged firing rate ¯ν in the balanced state can be predicted as ¯νbal ≈ I0/(J0τm) [10]. The good agree- ment of this prediction with the numerically obtained firing 0 (a) ) 1 - s ( λ i -50 -1 τ m N= 2000 N=10000 N=50000 K= 100 K= 200 K= 500 K=1000 K=2000 2 λ max λ mean RMT -50 (b) ) 1 - s ( λ -1 τ m -50 ) 1 - s ( λ -1 τ m 0 (c) 10000 K 0 ν (Hz) 20 0 0.2 i / N 0.8 1 Figure 2: Stable dynamics with respect to infinitesimal perturba- tions: (a) Spectrum of Lyapunov exponents {λi} of networks of N = 10 000 LIF neurons for different connectivities K, inset: close- up of spectra for K = 100 and different network sizes N, (b),(c) largest Lyapunov exponent λ2 = λmax and mean Lyapunov exponent λmean = 1 i=1 λi versus connectivity K and average firing rate ¯ν (dashed lines: random matrix theory for λmean [10]), (parameters: N = 100 000, K = 1000, ¯ν = 10 Hz, J0 = 1, τm = 10 ms; averages of 10 initial conditions). N PN rate confirms the dynamical balance of excitation and inhibi- tion in the studied networks. Although the voltage trajectory of each neuron and the net- work state were very irregular, the collective dynamics of the networks was apparently completely stable (Fig. 2). For all firing rates, coupling strengths and connection probabilities, the whole spectrum of Lyapunov exponents (disregarding the zero exponent for perturbations tangential to the trajectory) was negative, confirming the occurrence of so-called stable chaos in LIF networks [1, 2]. The invariance of the Lya- punov spectra to the network size N , to our knowledge, for the first time demonstrates that this type of dynamics is ex- tensive. With increasing connectivity K all Lyapunov ex- ponents approached a constant λi ≈ −1/τm. This is de- duced from the mean Lyapunov exponent given by λmean ≈ −1/τm + (VT − hV i)/(√KI0) + O(1/K) in random matrix approximation and the numerical observation that the largest exponent approached λmean in the large K-limit ([10] and Fig. 2(b)). These results suggest that in the thermodynamic limit arbitrary weak perturbations decay exponentially on the single neuron membrane time constant. As will become clear in the following, however, this issue is quite delicate. Experimentally realizable and well-controlled state pertur- bations to the dynamics of cortical networks are the addition or suppression of individual spikes [6, 12]. Such minimalis- tic neurostimulation can elicit complex behavioral responses [12] and can trigger a measurable rate response in intact cor- tical networks [6]. We therefore examined how such single spike perturbations affected the collective dynamics of our networks. Here, the simplest single spike perturbation is the suppression of a single spike. Figure 3 illustrates the firing rate response if one spike is skipped at t = 0. The miss- ing inhibition immediately triggered additional spikes in the K postsynaptic neurons such that the network-averaged firing rate increased abruptly by δ¯ν ∼ K ¯ν/N . Since the induced ex- tra spikes inhibited further neurons in the network, the over- shoot in the firing rate quickly settled back to the stationary 20 (a) n o r u e n 10.8 (b) ) z H ( ν 10 skipped 0 t (ms) 3 t (ms) 3 0 (c) 1 1 -40 0 t (ms) 40 reference K=1000 N=10000 K= 100 N=10000 K= 100 N= 1000 a r t x e S 0 0 t · Kν 3 PSfrag replacements Figure 3: Weak firing rate response after single spike failure: (a) Sample spike pattern of 20 randomly chosen neurons (gray: refer- ence trajectory, black: single spike skipped at t = 0), (b) network- averaged firing rate of reference trajectory ¯ν and in response to skipped spike ν versus time for different connectivities K and net- work sizes N, (c) number of extra spikes Sextra = N R (ν − ¯ν)dt in the entire network versus time (rescaled with average input rate K ¯ν), (parameters: ¯ν = 10 Hz, J0 = 1, τm = 10 ms; averages of 100 initial conditions with 10 000 calculations each). εft ∼ eK ¯ν t ε 1 t τm 10-3 e− 1 √N K ¯ν 100 (a) D 10-4 K=1000 K= 500 K= 200 t (ms) 6 0 0 (b) 20 ) 1 - s m ( p λ 0 0 10-1 (c) 10-2 λ p~ 0.9Kν / 2 1 K D / 10 Hz 5 Hz 2 Hz 6 t (ms) 10-4 10-5 N=103 N=104 N=105 K 2000 0 ν (Hz) 20 0 t · Kν 15 Figure 4: Sensitivity to single spike failures: (a) Distance D be- tween trajectory after spike failure and reference trajectory versus time in log-lin plots for different connectivities K and average fir- ing rates ¯ν, (b) pseudo Lyapunov exponent λp from exponential fits D ∼ exp(λpt) before reaching saturation versus connectivity K and average firing rate ¯ν, (c) distance-evolution of all parameter sets (rescaled with approximate perturbation strength KJ0/√K) versus time (rescaled with average input rate K ¯ν) collapse to characteristic exponential state separation with rate λp ∼ 0.9K ¯ν (inset: differ- ent network sizes N for K = 100), (parameters: N = 100 000, K = 1000, ¯ν = 10 Hz, J0 = 1, τm = 10 ms; averages of 10 initial conditions with 100 calculations each). state within a time of order δt ∼ 1/(K ¯ν). The overall number of additional spikes in the networks therefore was N δ¯νδt ≈ 1 and the one skipped spike was immediately compensated by a single extra spike [10]. N Pi φi(t) − φi(t) between the perturbed trajec- Even though the failure of one individual spike resulted in very weak and brief firing rate responses, it nevertheless in- duced rapid state decoherence. Figure 4 displays the distance D(t) = 1 tory (spike failure at t = 0) and the reference trajectory. After the spike failure, all trajectories separated exponentially fast at a surprisingly high rate. Because this exponential separation of nearby trajectories is reminiscent of deterministic chaos, we call its separation rate the pseudo Lyapunov exponent λp. The pseudo Lyapunov exponent was network size invariant, 100 D 10-8 (a) ε < ε ft 0 t(ms) 20 0 1 (b) (d) ε > ε ft 0 20 t(ms) ε ~~ ε t(ms) ft 20 e s p a n y s 1 10-4 ε s P 0.5 0 (c) 10-2 t f 1 − exp(−ε/εft) i e k p s ε ft 10-2 1 PSfrag replacements 1 − exp(−ε/εft) 100 εft ∼ √N K ¯ν 3 1 t τm e− eK ¯νt 1 103 N 105 102 104 K 100 ν(Hz) 102 Figure 5: (Color online) Sensitivity to finite-size perturbations: (a) Distance D between perturbed and reference trajectory measured at spike times of reference trajectory (projecting out possible time- shifts) for perturbations of strengths ε = 0.00002, 0.002, 0.2 in log- lin plots (gray lines: 20 examples for initial perturbations of same size pointing in different random directions perpendicular to trajec- tory, color lines: averages of exponentially separating/converging cases), (b) probability Ps of exponential state separation versus per- turbation strength ε in lin-log plot (dashed line: fit to Ps(ε) = 1− exp(−ε/εft), dotted line: characteristic perturbation size εft sep- arating stable from unstable dynamics, shaded areas: strengths cor- responding to single synapse and single spike failures), (c) character- istic perturbation size εft versus network size N, connectivity K and average firing rate ¯ν in log-log plots, (d) symbolic picture of stable flux tubes with radius εft (stable dynamics inside flux tube but expo- nential separation of adjacent flux tubes), (parameters: N = 10 000, K = 1 000, ¯ν = 10 Hz, J0 = 1, τm = 10 ms; averages of 10 initial conditions with 100 calculations and 100 random directions each). but showed a completely different behavior compared to the classical Lyapunov exponents. With increasing connectivity, it appears to diverge linearly λp ∼ K ¯ν. It is thus expected to grow to infinity in the high connectivity limit, reminiscent of binary neuron networks exhibiting an infinite Lyapunov expo- nent in the thermodynamic limit [4]. In the same balanced LIF networks, we thus find stable dy- namics in response to infinitesimal perturbations and unstable dynamics in response to single spike failures. To further ana- lyze the transition between these completely opposite behav- iors, we applied finite perturbations of variable size perpendic- ular to the state trajectory (Fig. 5). Depending on the pertur- bation strength ε and direction −→δφ (withPi δφ2 i = 1), the per- turbed trajectory either converged back to the reference trajec- tory or diverged exponentially fast. The probability Ps(ε) that a perturbation of strength ε induced exponential state separa- tion was very well-fitted by Ps(ε) = 1− exp(−ε/εft). Hence, εft is a characteristic phase space distance separating stable from unstable dynamics. Intriguingly, this distance decreased as εft ∼ 1/(√KN ¯ν). For large K and N the dynamics in the thermodynamic limit (N → ∞) would be unstable even to infinitesimal perturbations (ε → 0). Contrary, the analysis of the Lyapunov spectra has shown that taking the limit ε → 0 first and then N → ∞ yields stable dynamics. Thus, the order of the limits appears crucial in defining the dynamical nature of balanced LIF networks. The evolution of finite perturbations suggests a picture of stable flux tubes around unique trajectories (Fig. 5(d)). Perturbations within these flux tubes decayed exponentially, whereas perturbations greater than the typical flux tube ra- Single synaptic failures correspond to small perturbations of dius εft ∼ 1/(√KN ¯ν) induced exponential state separation. size εsyn ≈ J0/√KN and therefore had a N and K inde- pendent probability of inducing exponential state separation. This probability increased linearly with the average firing rate ¯ν [10]. Summarizing, our analysis revealed the co-occurence of dynamical stability to infinitesimal state perturbations and sensitive dependence on single spike and even single synapse perturbations in the dynamics of networks of inhibitory LIF neurons in the balanced state. They exhibit a negative defi- nite extensive Lyapunov spectrum that at first sight suggests a well-defined thermodynamic limit of the network dynamics characterized by stable chaos as previously proposed [1, 2]. In this dynamics, single spike failures induce extremely weak firing rate responses that become basically negligible for large networks. Nevertheless, such single spike perturbations typi- cally put the network state on a very different dynamical path that diverges exponentially from the original one. The rate of exponential state separation was quantified with the so called pseudo Lyapunov exponent λp. The scaling of λp ∼ K ¯ν im- plies extremely rapid, practically instantaneous, decorrelation of network microstates. Our results suggest that the seemingly paradoxical coexistence of local stability and exponential state separation reflects the partitioning of the networks' phase space into a tangle of flux tubes. States within a flux tube are attracted to a unique, dynamically stable trajectory. Different flux tubes, however, separate exponentially fast. The decreas- ing flux tube radius in the large system limit suggests that an unstable dynamics dominates the thermodynamic limit. The resulting sensitivity to initial conditions is described by the rate of flux tube separation, the pseudo Lyapunov exponent, that showed no sign of saturation. These findings suggest that the previously reported infinite Lyapunov exponent on the one hand [4] and local stability on the other hand [1, 2] resulted from the order in which the weak perturbation limit and the thermodynamic limit were taken. For finite networks, the phase space structure revealed here may provide a basis for insensitivity to small perturbations (e.g. noise or variations in external inputs) and strong sensitiv- 4 ity to larger perturbations. In the context of reservoir comput- ing, the flux tube radius defines a border between the fading property (variations of initial conditions smaller εft die out exponentially) and the separation property (input variations larger εft cause exponentially separating trajectories). Appli- cations of LIF neuron networks in reservoir computing may thus strongly benefit if the flux tube structure of the network phase space is taken into account. Our results of a very high pseudo Lyapunov exponent also reveal that the notion of an 'edge of chaos' is not applicable in these networks. We thank E. Bodenschatz, T. Geisel, S. Jahnke, J. Jost, P. E. Latham, R. M. Memmesheimer, H. Sompolinsky, M. Timme and C. van Vreeswijk for fruitful discussions. This work was supported by BMBF (01GQ07113, 01GQ0811), GIF (906-17.1/2006), DFG (SFB 889) and the Max Planck Society. ∗ Electronic address: [email protected] [1] R. Zillmer et al. Phys. Rev. E 74, 036203 (2006); R. Zillmer, N. Brunel and D. Hansel, ibid. 79, 031909 (2009). [2] S. Jahnke, R. M. Memmesheimer and M. Timme, Phys. Rev. Lett. 100, 048102 (2008); Front. Comput. Neurosci. 3, 13 (2009). [3] D. Z. Jin, Phys. Rev. Lett. 89, 208102 (2002). [4] C. van Vreeswijk and H. Sompolinsky, Science 274, 1724 (1996); Neural Comput. 10, 1321 (1998). [5] H. Sompolinsky, A. Crisanti, and H. J. Sommers, Phys. Rev. Lett. 61, 259 (1988); D. Hansel and H. Sompolinsky, J. Com- put. Neurosci. 3, 7 (1996); A. Zumdieck et al., Phys. Rev. Lett. 93, 244103 (2004); A. Banerjee, P. Seriés, and A. Pouget, Neu- ral Comput. 20, 974 (2008); D. Zhou et al., Phys. Rev. E 80, 031918 (2009); S. El Boustani and A. Destexhe, Int. J. Bifur- cat. Chaos 20, 1687 (2010); M. Monteforte and F. Wolf, Phys. Rev. Lett. 105, 268104 (2010). [6] M. London et al., Nature 466, 123 (2010). [7] N. Bertschinger and T. Natschlaeger, Neural Comput. 16, 1413 (2004). [8] W. Maass, T. Natschlaeger and H. Markram, Neural Comput. 14, 2531 (2002); H. Jaeger, GMD Report 148 (2001); D. Sus- sillo and L. F. Abbott, Neuron 63, 544 (2009). [9] A. N. Burkitt, Biol. Cybern. 95, 1 (2006); 95, 97 (2006). [10] See EPAPS Document No. xxx.yyy for more details. [11] G. Benettin et al., Meccanica 15, 21 (1980). [12] M. Brecht et al., Nature 427, 704 (2004); D. Huber et al., Nature 451, 61 (2008); A. R. Houweling and M. Brecht, Nature 451, 65 (2008).
1311.4206
2
1311
2015-03-03T13:52:23
Fluctuations and information filtering in coupled populations of spiking neurons with adaptation
[ "q-bio.NC" ]
Finite-sized populations of spiking elements are fundamental to brain function, but also used in many areas of physics. Here we present a theory of the dynamics of finite-sized populations of spiking units, based on a quasi-renewal description of neurons with adaptation. We derive an integral equation with colored noise that governs the stochastic dynamics of the population activity in response to time-dependent stimulation and calculate the spectral density in the asynchronous state. We show that systems of coupled populations with adaptation can generate a frequency band in which sensory information is preferentially encoded. The theory is applicable to fully as well as randomly connected networks, and to leaky integrate-and-fire as well as to generalized spiking neurons with adaptation on multiple time scales.
q-bio.NC
q-bio
Fluctuations and information filtering in coupled populations of spiking neurons with adaptation∗ Moritz Deger,1 Tilo Schwalger,1 Richard Naud,2 and Wulfram Gerstner1 1School of Computer and Communication Sciences and School of Life Sciences, Brain Mind Institute, École polytechnique fédérale de Lausanne, Station 15, 1015 Lausanne EPFL, Switzerland† 2Department of Physics, University of Ottawa, 150 Louis Pasteur, K1N-6N5 Ottawa, Ontario, Canada (Dated: June 26, 2018) 5 1 0 2 r a M 3 ] . C N o i b - q [ 2 v 6 0 2 4 . 1 1 3 1 : v i X r a Finite-sized populations of spiking elements are fundamental to brain function, but also used in many areas of physics. Here we present a theory of the dynamics of finite-sized populations of spiking units, based on a quasi-renewal description of neurons with adaptation. We derive an integral equation with colored noise that governs the stochastic dynamics of the population activity in response to time-dependent stimulation and calculate the spectral density in the asynchronous state. We show that systems of coupled populations with adaptation can generate a frequency band in which sensory information is preferentially encoded. The theory is applicable to fully as well as randomly connected networks, and to leaky integrate-and-fire as well as to generalized spiking neurons with adaptation on multiple time scales. I. INTRODUCTION Multi-scale modeling of complex systems has led to important advances in fields as diverse as complex fluid dynamics, chemical biology, soft matter physics, mete- orology, computer science, and neuroscience [1–6]. In these approaches, mathematical methods such as mean- field theories and coarse-graining provide the basis to link properties of microscopic elements to macroscopic vari- ables. In many cases, macroscopic variables fluctuate due to a finite number of microscopic elements. For instance, in the brain, neurons can be grouped into populations of 50 to 1000 neurons [7] with similar properties [8, 9]. Fluc- tuations of the global activity of such populations are not captured in classical mean-field theories [10, 11], which assume infinite system size. Here we put forward a the- ory for the fluctuating macroscopic activity in networks of pulse-coupled elements occurring in neuronal networks [12], queuing theory [13] and synchronizing fireflies [14]. Finite-size effects in networks of spiking elements have been approached by different methods, including exten- sions of the Fokker-Planck equation for neuronal mem- brane potentials [10, 15], stochastic field theory [16], mo- ment expansions in networks of generalized linear models (GLM) [17], modified Hawkes processes [18–20], simpli- fied Markov neuron models [21–27] and the use of a linear response formalism for spike trains perturbed by finite- size fluctuations [28–30]. These studies lack, however, slow cellular feedback mechanisms mediating adaptation. Adaptation characterized by a reduced response of a ∗ This article was published as: M. Deger, T. Schwalger, R. Naud, W. Gerstner, Fluctuations and information filtering in coupled populations of spiking neurons with adaptation, Phys. Rev. E 90, 062704 (2014), DOI: 10.1103/PhysRevE.90.062704 . Typesetting errors were corrected in the paragraph of Eq. (5) and in Eqs. (8) and (13). † Corresponding authors: M. D., [email protected] ; T. S., [email protected] . neuron to slow compared to fast inputs is a wide-spread phenomenon in the brain and has important implications for signal processing [31–34] and the spontaneous activity of single neurons [35–37]. On the population level, adap- tation has been recently analyzed using a quasi-renewal (QR) theory [38]. The QR framework uses tools of re- newal point process theory [39] to treat neurons with ar- bitrary refractoriness. In particular, the dynamics of the population activity is determined by an integral equa- tion [11, 40, 41]. These studies, however, have assumed an infinitely large population. Here we present a theory for the interaction of finite- sized populations of adapting neurons. The theory is valid for the broad class of neuron models that can be approximated by a QR point process. This includes integrate-and-fire (IF) as well as GLM neurons, for which parameters can be reliably extracted from experimental data [9, 12, 42]. Based on this theory we analyze information filtering (noise shaping) in neuronal populations. We show that in a single population no noise shaping occurs, but that in coupled populations, band-pass-like noise shaping is possible due to adaptation and connectivity. This article is structured as follows: First, we present the general dynamics of the population activity and its fluctuations. We then describe how randomly connected, adapting neurons can be treated in this framework. For fluctuations about a stationary state, we linearize the dynamics and compute the spectral density of the pop- ulation activity. We then determine its coherence with external input signals and quantify information transmis- sion. 2 still large enough to include many spikes of the popula- tion. For large N, the number of neurons that fire in the time bin at t and had their last spike in bin t is a Gaussian random number with mean and variance n0(t)PH(t, t)∆t, where n0(t) = N A(s) ds is the past spike count at t < t. Summing over t and treating ∆t as macroscopi- cally infinitesimal, we find the conditional mean activity (see Appendix B) ´ t+∆t t t a(t) = −∞ PH(t, t)A(t) dt, (1) Figure 1. (Color) Schematic of the spike response model (a GLM for spike generation) with exponential escape noise and the derived quasi-renewal population model. II. RESULTS A. Dynamics of globally coupled renewal models. ity A(t) = N−1(cid:80)N i=1 si(t), where si(t) = (cid:80) Our main quantity of interest is the population activ- k δ(t − tk i ) i and is the spike train of neuron i with spike times tk N denotes the number of neurons. In experiments or simulations, the measured activity ¯A(t) would be deter- ´ ∞ mined by temporal filtering of the population activity, −∞ A(t − s)f (s)ds with a normalized filter i.e. ¯A(t) = function f (s) with finite support. Below we will use a rectangular filter f (s) = θ(s)θ(∆t − s)/∆t, where θ(s) is the Heaviside step function. To determine the fluctuation statistics of A(t), we gen- eralize the integral equation of an infinite population [11] to large but finite N. Let us first consider a homoge- neous population of all-to-all connected renewal neurons. In this case, the spikes of each neuron occur with an in- stantaneous rate or hazard function ρH(t, t), which only depends on its last spike time t ≤ t and the synaptic input determined by the history H(t) = {A(t(cid:48))}t(cid:48)<t of the population activity. Note that for uncoupled station- ary networks, the hazard reduces to ρH(t, t) = ρ(t − t) as it should be for a renewal model. The probabil- ity density of the next spike time t given t is given by PH(t, t) = ρH(t, t)SH(t, t), with the survivor function de- fined as SH(t, t) = exp(− ´ t ρH(t(cid:48), t)dt(cid:48)). Our approach is to use the Gaussian approximation for large N, i.e. we calculate the first- and second-order statistics of A(t) as a functional of its past activity (see Appendix B). However, the dynamics of A depends on its own history H(t) and on the occupation density of refrac- tory states t−t across the population [43]. Thus, we have to average over the possible refractory states consistent with a given history H. To perform the Gaussian approx- imation, the dynamics of the full system is coarse-grained by discretizing time with a small time step ∆t which is t which is equal to the population integral [11] for the in- finite system. For finite N, A(t) will be of the form (2) A(t) = a(t) + δA(t), standard deviation (cid:112)a(t)/(N ∆t) because the variance where the deviation δA(t) has zero mean and a diverging of the spike count in [t, t + ∆t] is given by N ∆t a(t). Importantly, δA(t) cannot be described by a white noise process but future values δA(t + τ ), τ > 0, are correlated with δA(t), because they share a common history H(t). In fact, a neuron that fired its last spike at t < t cannot have its next spike at both times t and t+τ, which induces a negative correlation for the deviations at t and t + τ. We find (see Appendix B) for τ ≥ 0 the conditional correlation function (cid:104)δA(t + τ )δA(t)(cid:105)H(t) = N−1a(t)δ(τ )− N−1 t −∞ PH(t + τ, t)PH(t, t)A(t) dt, (3) where (cid:104)·(cid:105)H(t) denotes the average conditioned on the his- tory of A before t. Thus, the correlation function is in general explicitly time-dependent. B. Adaptation and random connectivity. In the presence of adaptation, the instantaneous rate of a neuron depends on all its previous spikes so that it can no longer be described by renewal theory. Here we describe how adapting neurons in networks may still be approximated by a quasi-renewal process. Specifically, we consider a homogeneous population of neurons mod- eled by the spike-response model with escape-noise [11], also known as GLM [9, 12, 42], with hazard function ρi(t) = c exp [(hi(t) − ϑi(t))/δu] , (4) (cid:104) κ ∗(cid:16)(cid:80)N (Fig. 1). That is, neuron i produces a spike in a [t, ∆t) with probability ρi(t)∆t. small time interval This probability depends on the input potential (t), which is driven hi(t) = by presynaptic spike trains sj(t) (with synaptic weight wij) and external input I(t). The membrane filter kernel j=1 wijsj(t) + I (cid:17)(cid:105) membranefilternonlinearityspikingmechanismrefractoriness & adaptation filternetwork (other neurons)+externalinputsingle neuron modelpopulation model-+-+other populationsinterval densitymean activity &fluctuationsneuronal population is given by κ(t) = θ(t − τs) exp(−(t − τs)/τm), where τs and τm are the synaptic delay and the membrane ´ ∞ time constant, respectively. The operation ∗ denotes the convolution (f ∗ g)(t) = −∞ f (t − s)g(s)ds and The variable θ(t) is the Heaviside step function. ϑi defined as ϑi(t) = (si ∗ η)(t) can be interpreted as a dynamic firing threshold that is triggered by the neuron's own output spike train si(t) [9]. Here, η(t) is a feedback kernel that consists of two parts, η = ηa + ηr: a short-range refractory kernel ηr(t) = θ(t) [Jrθ(t − τabs) exp(−(t − τabs)/τm) + θ(τabs − t)D] mainly affected by the last spike, and a long-range adap- tation kernel ηa(t) = Jaθ(t) exp(−t/τa) that accumulates the spike history on a longer time scale τa. An absolute refractory period is included in ηr(t) by setting it to D = 1012 for 0 < t < τabs. Our choice of the kernels corresponds to a leaky IF model with dynamic threshold [44] and a reset by a constant amount −Jr after each spike. The parameter c in Eq. (4) sets a baseline firing rate and δu sets the strength of intrinsic noise ("softness" of threshold). Fits of this model to pyramidal neuron recordings yielded δu ≈ 4mV [45]. Our standard param- eter set given below corresponds to an amplitude of single post-synaptic potentials of 0.25mV (excitatory, exc.) and −1.1mV (inhibitory, inh.). In the following, we measure voltage in units of δu, so that δu = 1 in dimensionless units. For the synaptic weights wij, we use a homoge- neous random network as specified in Appendix A, below. The dependence of the term exp(−ϑi) in Eq. (4) which describes the feedback of the neuron's own spiking history si can be approximated by the explicit contribution of the last spike of the neuron at t and the average effect of previous spikes up to t [38]: ´ e−ϑi(t) ≈ e−η(t−t)(cid:104)e − t−∞ si(t(cid:48))η(t−t(cid:48)) dt(cid:48)(cid:105)tk (cid:16) (cid:104)´ t Here, the average is taken over all previous spike times i < t. As shown in [38], this average can be approxi- tk mated by exp . Replac- ing further the firing rate (cid:104)si(t(cid:48))(cid:105) by the population ac- tivity A(t(cid:48)), the threshold ϑi(t) becomes (cid:17)(cid:104)si(t(cid:48))(cid:105) dt(cid:48)(cid:105) e−η(t−t(cid:48)) − 1 i <t. −∞ ϑ(t, t) = η(t − t) + (γt−t ∗ A)(t), (5) for all neurons with last spike at t. The kernel γτ (s) = θ(s − τ )(1 − e−η(s)) represents the effect of adaptation in the quasi-renewal (QR) approximation. Furthermore, for homogeneous random networks and large N, the lo- cal field hi caused by synaptic input to neuron i is de- i,j wij termined by A and an effective weight ¯w = N−2(cid:80) [10], hence h(t) = [κ ∗ (JsA + I)](t), (6) where Js = N ¯w. The above steps enable us to treat neuronal adaptation and network coupling in a quasi- renewal framework with hazard function ρH(t, t) = c exp(cid:2)h(t) − ϑ(t, t)(cid:3) . (7) 3 Figure 2. (Color) Quasi-renewal approximation of the pop- ulation dynamics. Coarse-grained population activity ¯A(t) ( ¯A(t) = n0(t)/(N ∆t), where n0(t) is the spike count in [t, t + ∆t), with ∆t = 2ms, black) resulting from 500 ran- domly connected adapting neurons (4) receiving a common input current (I(t), blue). The theoretical expectation a(t) (green, Eq. 1) based on QR approximation (7), and expected fluctuations (red, one std., a(t) ± (N ∆t)− 1 2 ). At time t, a(t) depends on the actual history ¯A(t(cid:48)) (black) for t(cid:48) < t. Standard parameters (see Appendix A), except for I(t) as shown and c = 5s−1. √ Inset shows mean (green) and std. (red) of deviations δA(t) = ¯A(t) − a(t) as a function of a, averaged over 25 repetitions of the displayed I(t) (dots) vs. theory (lines). 2 a(t) 1 Note that ρH(t, t) is identical for all neurons which have fired their last spike at t. In Fig. 2 the population activity of a spiking neural network simulation is compared to the theoretical pre- diction (2). To evaluate (1) numerically, we iteratively compute SH(t + ∆t, t) = SH(t, t)(1 − ρH(t, t)∆t), with SH(t, t) = 1, and use PH = ρHSH. The QR population integral describes the response of the population activity and its fluctuations for stationary, as well as for slowly or rapidly varying inputs. C. Linearized population dynamics. The amount of information transmitted and processed in sensory areas of the brain is limited by the fluctuations of the population activities [46–48]. Likewise, in decision networks, finite-size induced fluctuations determine the reliability of decisions [49, 50]. The spontaneous activ- ity of cortical networks is typically asynchronous, and is believed to underlie cortical information processing [51]. In order to analytically determine the power spectrum of the spontaneous population activity, we linearize the dynamics around the large N limit. To this end, we as- sume that in the limit N → ∞ and for constant external input I(t) = I0 the network dynamics has an equilibrium point with activity A0 corresponding to an asynchronous firing state. With this equilibrium activity we can asso- ciate a renewal neuron model that is obtained from the original model, Eq. (4), by replacing hi(t) and ϑi(t, t) by h0 = κ∗ (JA0 + I0) and ϑ0(t− t) = η(t− t) + (γt−t ∗ A0), respectively. In the following, we will use the subscript 0510152025A(t)[1/s]1.01.21.41.61.82.02.2time [s]1800180I(t)[1/s]012345pa[1/ps]024A[1/s] "0" to refer to quantities of the associated renewal model. For finite system size, N < ∞, the activity will deviate from A0. Through (5) and (6) the fluctuations ∆A(t) = A(t) − A0 also lead to fluctuations ∆h(t) = h(t) − h0(t) and ∆ϑ(t, t) = ϑ(t, t)− ϑ0(t− t) , which in turn influence A(t). Our goal is to determine the spectral properties of ∆A(t). To simplify the derivations, we approximate the QR kernel γτ by its average over the inter-spike-interval den- sity P0(τ ) 1, ∞ P0(τ )γτ (s)dτ = (1 − e−η(s))(1 − S0(s)) .(8) γ(s) = 0 d to first order Expanding (1)-(3) in ∆A, ∆h and ∆ϑ, yields the linearized stochastic dynamics (see Ap- pendix C) A(t) = A0 + (Q ∗ ∆A)(t) +(cid:112)A0/N ξ(t). (9a) dt (L ∗ [κJs − γ])(t) determines Here, Q(t) = P0(t) + A0 the linear response of the expected activity a to a per- ´ ∞ turbation ∆A. For our model (4), the kernel L in this expression is given by L(t) = θ(t) 0 ρ0(s)S0(s+t)ds but L can be derived for most common neuron models [11], or, alternatively, may be estimated from neural record- ings. The noise term ξ(t) is stationary Gaussian noise with correlation function (cid:104)ξ(t)ξ(t + τ )(cid:105) = δ(τ ) − P0(s + τ )P0(s)ds ∞ (9b) 0 for all τ; cf. (3). Eq. (9) shows that the population activity in the stationary state is a Gaussian process with memory, where finite-size fluctuations are described by the colored noise ξ(t). D. Fluctuations in coupled populations. Let us now turn to K populations consisting of (cid:126)N = (N1, . . . , NK) neurons. Parameters of neurons and cou- pling are homogeneous within each population but may differ between one group and the next. To incorporate network coupling, the network input (6) becomes hk(t) = (κk ∗ (J (cid:126)A)k)(t) for k = 1, . . . , K, where J is the coupling matrix and (cid:126)A(t) is a vector of population activities. For each population the dynamics are given by (9) but Q(t) ´ ∞ is now a K × K matrix of coupling kernels. Using the −∞ f (t)e−iωtdt, this matrix Fourier transform f (ω) = can be written as Q(ω) = P0 + iωA0 L( KJ − G), where the matrices P0, A0, L, K, G are defined as the diagonal matrices of the vectors (cid:126)P0, (cid:126)A0, (cid:126)L, (cid:126)κ, (cid:126)γ, respectively. The power spectrum, defined as the Fourier transform of the 1 An alternative would be to average γτ (s) over the backward re- currence time A0S0(τ ). (cid:104) † 0)(1− P 1 − R0(J − K−1 G) (cid:105)−1 4 correlation function CA(τ ) = (cid:104)∆ (cid:126)A(t)∆ (cid:126)AT (t + τ )(cid:105), can be obtained from the transformed Eq. (9) as CA(ω) = † 0)(1 − Q†)−1. Here, † de- (1 − Q)−1N−1A0(1 − P0 P notes the adjoint matrix (conjugate transpose). It is instructive to rewrite this expression in terms of the power spectrum of the associated renewal model [52] † C0(ω) = A0(1− P0)−1(1− P0 P 0)−1 as follows: CA(ω) = BN−1 C0 B†, B = (10) Here, R0 = iω(1 − P0)−1 A0 L K is the diagonal matrix containing the linear response functions of the associated renewal models with respect to current perturbations (cid:126)I(t) [11]. Eq. (10) shows that finite-size fluctuations are char- acterized by the renewal spectrum C0(ω), shaped by re- current input (via J and K) and adaptation (via G), and reduced by the factor N−1. There are several known limit cases: first, for vanishing adaptation, G = 0, we re- cover the linear response result of [28–30] for networks of white-noise driven IF neurons. Our formula shows that adaptation appears as an additional diagonal term in the effective coupling matrix J − K−1 G, and hence can be interpreted as an inhibitory self-coupling. Second, if both adaptation and recurrent connections vanish, G = 0 and J = 0, we arrive at CA(ω) = N−1 C0(ω) because the su- perposition of independent spike trains does not change the shape of the power spectrum. Third, our result also includes the frequently employed Hawkes process [18–20], which is recovered for a constant single neuron spectrum C0 = A0 and vanishing adaptation, G = 0. For a com- parison of our result to simulations, see Fig. 3, which will be discussed below. In section II F we make use of Eq. (10) to quantify information filtering in neural pop- ulations (Fig. 4). A comparison to the special cases of Eq. (10) described above is shown in Fig. 5. In Fig. 3(a)-(b) the spectral density is shown compared to simulations, where the frequency equals ω/(2π). The spectra are well described by the novel theory, which cap- tures refractoriness, recurrent feedback and the reduction of power at low frequencies due to adaptation. The lat- ter arises from negative correlations between ISIs typical for adapting neurons [37]. Interestingly, this purely non- renewal effect is well accounted for by our quasi-renewal theory. Since adaptation effects are most prominent at low frequency, we examined the dependence of the power on the model parameters for ω → 0 (Fig. 3(c)-(d)). Our theory describes the simulations well across the studied parameter range. E. Influence of correlated external signals. Neuronal networks in the brain are subject to exter- nal influences, either due to sensory input or ongoing activity in other brain areas. How do neuronal pop- ulations respond to small, time-dependent input cur- 5 Figure 3. (Color) Spectral density of the population activity in random networks. Simulation (light colored lines) vs. theory (10) (solid lines). (a): single populations (K = 1), (b-d): coupled exc.-inh. network (K = 2), (in (b) green lines show the amplitude of the cross-spectrum). (c-d): The limit ω → 0 in dependence on parameters for the K = 2 case shown in (b); symbols mark simulation results for different connection probabilities ((cid:53),(cid:52),◦: p = 0.2, 0.5, 1) while keeping Js = N ¯w fixed, solid lines show theory (10). Symbols mostly fall on top of each other indicating independence of results with respect to p. In (c) colors are as in (b), in (d1-d3) only exc. population is shown, colors denote number of neurons (blue, green, red, cyan: N = 50, 250, 500, 1000). Standard parameters (marked by arrows in (c-d), cf. Appendix A) were used except as indicated. rents (cid:126)I(t) = (I1, . . . , IL) with spectral density CI (ω)? To answer this question, we proceed as before and lin- earize Eqs. (1)-(3) with respect to the small fluctuations ∆hk(t) = (κk ∗ (J∆ (cid:126)A + M(cid:126)I)k)(t) of the local field. Here, we restrict our analysis to independent inputs (cid:126)I, such that CI is diagonal with entries CI,i, but also included a K×L mixing matrix M, which allows us to model shared input. The resulting spectral density is given by the sum CA(ω) = B[N−1 C0 + R0M CIM† R † 0] B†. (11) Thus, additional fluctuations due to the stimulus (cid:126)I are shaped by both the single neuron filter R0M and the net- work and adaptation filter B (10), combining the effects of recurrent connectivity and adaptation. ´ ∞ 0 can be regarded as a frequency resolved measure of infor- mation transmission. Information theory [53–55] states that the mutual information rate is bounded from below by − log2[1 − Γij(ω)] dω 2π . Since adaptation attenuates the response to slowly changing signals, one might expect that it also atten- uates low frequency information content. For a single population, however, this is not the case, but instead the coherence is low-pass, i.e. it monotonically decreases for increasing frequency [29]. Here we show that in coupled populations of adapting neurons, coherences can be non- monotonic allowing the neural circuit to preferentially encode information in certain frequency bands. Put dif- ferently, a multi-population setup can realize an informa- tion filter. Using Eq. (12), we find the general form of the coher- ence matrix: F. Information transmission. Γij(ω) = Our theory allows us to quantify the transmission of information from external input signals (cid:126)I(t) through a system of coupled neural populations. The coherence between the signal j and the activity of population i Γij(ω) = (cid:104) I∗ j (ω) Ai(ω)(cid:105)2 CI,j(ω)( CA(ω))ii (12) k=1 Bik2 C0,k +(cid:80)L (cid:80)K ( B R0M)ij2 CI,j l=1 ( B R0M)il2 CI,l . (13) In this expression, the numerator represents the contri- bution of the signal Ij(t) to the power spectrum of pop- ulation i. This effective signal power is divided by the total power spectrum of population i, which consists of direct (k = i) and indirect (k (cid:54)= i) sources of variabil- ity. Both sources contain internally generated noise due (a)singlepopulationnetworks(b)excitation-inhibitionnetwork10-1100101102frequency [Hz]0.51.0spectral density [0.01/s]Js=5Js=0Js=-510-1100101102frequency [Hz]01234spectral density [0.01/s]exc.crossinh.(c)(d1)(d2)(d3)0250500750N110CA(0)[0.01/s]0.51.01.5u0240246810Js024543210Ja024 6 power and hence exhibits a very similar reduction of low- frequency power. Consequently, the coherence (being the ratio of these two spectra) is rather flat at low fre- quencies and shows a decay at higher frequencies (Fig. 4, red lines). This low-pass characteristics changes at later stages in the chain (green & blue): The signal term is in- creasingly more shaped by adaptation and coupling prop- erties, whereas the noise spectrum changes less. As a result, the coherence shows a maximum at a finite fre- quency (Fig. 4, blue lines). This band-pass structure be- comes more pronounced from layer to layer, representing a form of information filtering. Coherence functions with band-pass characteristics have been observed in neurons postsynaptic to electro-receptor afferents in electric fish [58, 59]. III. CONCLUSIONS Figure 4. (Color) Signal processing and coherence shaping in a feed-forward chain of recurrently connected neural popula- tions, simulation vs. theory. Left: Schematic of the network. Right: spectral coherences Γi1 (13) of I1 and Aexc. , theory (13) (lines) vs. simulation result (light colored lines). Col- ors indicate coherences of signals I1 and population activity Ai. The mutual information rate is closely related to the spectral coherence, see text. Standard parameters, except for increased coupling Js = 10. The input current I1(t) is a Gaussian white noise with spectral density CI,1(ω) = 9s−1. i to finite size Nk as well as signal power. However, the diagonal elements (k = i) of the shaping matrices can be much stronger than the off-diagonal elements (k (cid:54)= i) depending on the coupling matrix J. Therefore, we ex- pect that the direct source of variability dominates in the denominator. If there is only one population and signal (K = 1, L = 1), the term B11(ω)2 occurs in both numerator and denominator and cancels. Thus, coupling and adap- tation do not shape the coherence in a single popula- tion. Furthermore, the signal term R0,1(ω)2M11 CI,1(ω) is matched in both numerator and denominator, which leads to a flat coherence at frequencies where the sig- nal dominates the finite N noise. At high frequen- cies, the neural response amplitude R02 decays due to the leaky membrane, but the spontaneous spectrum C0 has a constant high-frequency limit equal to A0. One therefore typically observes a low-pass like information transfer characteristics of single neurons or populations [29, 56, 57]. For several populations (K > 1), however, we can dis- tinguish two cases: If the signal is read out at a receiving population, Mij (cid:54)= 0, the signal power in the numerator is matched by the dominating direct signal power in the denominator, and hence the shaping of the signal power cancels. In contrast, if read out at a different popula- tion, Mij = 0, the signal j contributes only indirectly to the power spectrum of population i via synaptic connec- tions. Thus, we expect that the shape of signal power and power spectrum (i.e. numerator and denominator, re- spectively) is generally different if the transmission path involves multiple populations. As an example of this mechanism, we show a feed- forward chain of excitatory and inhibitory populations (Fig. 4, K = 6, L = 1). In the first layer, the effec- tive signal power is reduced at low frequencies because of adaptation and inhibitory feedback. However, the power spectrum of Aexc,1(t) is dominated by the same signal We have shown that fluctuations in finite-sized net- works of spiking neurons are captured by a colored noise term added to the population integral equation of the in- finite system. Our approach yields spectral densities of the population activity in randomly or fully connected multi-population networks which are in excellent agree- ment with simulation results. Our quasi-renewal theory includes refractory effects and adaptation on multiple time scales. In contrast to earlier treatments of neuronal refractory effects in population dynamics [17, 19, 20] or linear response formulas for adaptive neurons [60], the QR population integral, derived directly from the neu- ron model definition, captures the time-dependent, non- linear dynamics and adaptation of neural population ac- tivity. We applied our theory to information filtering by cou- pled populations of spiking neurons with adaptation. We showed that, although impossible in single popula- tions due to a cancellation of signal and noise terms of the coherence, coupled populations can filter informa- tion through adaptation mechanisms and neuronal inter- actions. This mechanism might be exploited in the lay- ered structure of cortical circuits, or in sensory systems of insects where signals traverse a sequence of nuclei. In this paper we treated populations of point neurons with static synapses and applied linear response theory. How to generalize our theory to incorporate effects of nonlinear dendritic integration, spike-synchrony detec- tion and short-term synaptic plasticity, which all con- tribute to information filtering [61–63], is an important question that merits further investigation. Nonetheless, due to the versatility of GLM models, our theory already provides a useful tool for interpreting neural data at the population level. For example, our theory suggests that in in-vitro experiments with optogenetically evoked in- put currents and simultaneous measurements of neural activity [64], system parameters may be identified based on the relation of the spectra, Eq. (11). Moreover, large- scale neural systems can now be analyzed as coupled 10-1100101102frequency [Hz]10-310-210-1100spectral coherence [1] populations of model neurons with single-cell parame- ters extracted from experiments, and simulated using a muti-scale approach. n0(t) # of neurons which spike in [t, t + ∆t] nk(t) # of neurons which spiked in [t − k∆t, t − (k − 1)∆t] ; nk(t) = n0(t − k∆t) 7 ACKNOWLEDGEMENTS Research was supported by the European Research Council (no. 268 689, T. Schwalger and W. Gerst- ner) and by the Swiss National Science Foundation (no. 200020_147200, M. Deger). We thank Laureline Logiaco for helpful discussions. M. D. and T. S. contributed equally to this work. Appendix A: Network simulations We compare our theoretical results to simulations of networks of excitatory and inhibitory neurons defined by (4). In the case K = 1 (Fig. 3(a)), we use (cid:126)N = N and J = Js, in the case K = 2 (Fig. 3(b)-(d)), (cid:126)N = (4/5N, 1/5N ), and J = ((Js,−1.1 Js), (Js,−1.1 Js)). In the case K = 6 (Fig. 4), each exc. (inh.) population consists of 4/5N (1/5N) neurons and J with entries as in Fig. 4A, where exc. (inh.) couplings are Js (−1.1 Js). Generally, neurons of population i receive synapses from a random subset of pNj neurons of population j, each with synaptic weight wij = Jij/(pNj) and delay τs. Un- less the connection probability p is 1, self-connections are excluded. Networks were simulated for 2 · 104s us- ing NEST [65] (neuron model pp_psc_delta, temporal resolution 2ms). Standard parameters, unless indicated otherwise: N = 500, c = 10s−1, τm = 0.01s, τa = 0.3s, ∆t = τs = τabs = 2ms, Jr = 3, Ja = 1, Js = 5, p = 0.2, I0 = 0, CI = 0. In the exc.-inh. network, for the inh. neurons which typically show little adaptation [9] we de- activated adaptation by setting Ja = 0 and c = 5s−1. While it is possible to theoretically approximate the sta- tionary interval distribution P0 by searching for a self- consistent rate A0 as described in [38], here we use P0 ´ ∞ from simulated inter-spike-intervals of each population. ´ ∞ From the measured P0(t) we derive A0 = 1/ tP0(t)dt, t P0(t(cid:48))dt(cid:48) and ρ0(t) = P0(t)/S0(t). S0(t) = 0 Appendix B: Detailed derivation of Eq. (3) The aim is to find a dynamical equation for the popu- lation activity N(cid:88) i=1 1 N si(t) = lim ∆t→0 n0(t) N ∆t , (B1) A(t) = mk(t) # of neurons with last spike in [t − k∆t, t − (k − 1)∆t] δmk(t) # of neurons with last spike in [t − k∆t, t − (k − 1)∆t] and next spike in [t, t + ∆t] ρ(t, t) hazard function: rate at t given last spike at t S(t, t) P (t, t) inter-spike-interval density: probability density survivor function: probability of no spike in [t, t] of next spike at t given last spike at t; P = ρ · S Table I. Definitions of symbols used in Appendix B. ∞(cid:88) furthermore the total number of neurons that spiked in the time bin [t − k∆t, t − (k − 1)∆t], k = 1, 2, . . . , but had no further spike until time t. Let us denote this number by mk(t) (Fig. 6 and Table I). The number of neurons that spike in [t, t + ∆t] and had their last spike in the bin [t − k∆t, t − (k − 1)∆t] shall be denoted by δmk(t). These neurons decrease the number mk(t) of neurons from group k that had survived until time t + ∆t in the next time step, i.e. δmk(t) = mk(t)− mk+1(t + ∆t), The total number of spikes at time t, n0(t), is the sum over all possible last spike times, hence k = 1, 2, . . . . (B2) n0(t) = δmk(t). (B3) k=1 = (cid:104)·(cid:105){nk(t)}k=1,2,... We will now express the activity n0(t) in terms of the past activity Ht = {nk(t)}k=1,2,..., using the Gaus- sian approximation. This requires to compute the mean and correlation function of n0(t) given the past values nk(t), k = 1, 2, . . . . In the following, the averaging bracket (cid:104)·(cid:105) has to be understood as the conditional aver- age (cid:104)·(cid:105)Ht , i.e. we will omit the condi- tioning subscript for simplicity. Although nk(t), the total number of spikes in bin t−k∆t, is fixed, the number mk(t) of neurons that had their last spike in bin t− k∆t is vari- able. It is this variability that we will average over (This corresponds to a statistical ensemble of populations that all have an identical history of population activity nk(t), k = 1, 2, . . . .). Suppose we know the value mk(t) of the group of neu- rons with their last spike in [t− k∆t, t− (k− 1)∆t]. Then the expected number of spikes from that group in the next interval is (cid:104)δmk(t)(cid:105)mk(t) = ρ(t, t − k∆t) · ∆t · mk(t) , (B4) where n0(t) is the total number of spikes in the interval [t, t + ∆t]. More generally, we define nk(t), k ∈ Z, as the total number of spikes in [t− k∆t, t− (k − 1)∆t], i.e. the activity k time bins in the past. It is useful to consider where ρ(t, t) is the hazard function of the neurons (in- stantaneous rate at time t given last spike at t), and (cid:104)x(cid:105)y denotes the expectation of x conditioned on y (in ad- dition to the overall condition of a fixed history Ht = 8 Figure 5. (Color) Spectral density of the population activity in random networks (as Fig. 3(a)-(b)) with comparison to earlier theories (special cases). Simulation (light colored lines) vs. theory (10) (solid lines). For comparison: uncoupled renewal processes ( B = 1), dotted; coupled renewal processes ( B−1 = 1 − R0J), dash-dotted; Hawkes process ( B−1 = 1 − R0J, C0 = A0), dashed lines. (a): single populations (K = 1), (b): coupled exc.-inh. network (K = 2). Blue dash-dotted and solid lines coincide because inhibitory neurons here have no adaptation (Ja = 0). Standard parameters were used except as indicated. The average number of neurons that fired their last spike in [t − k∆t, t − (k − 1)∆t] and survived up to t can be expressed using the survival probability S(t, t) = exp(− ´ t ρ(t(cid:48), t)dt(cid:48)) as follows: t (cid:104)mk(t)(cid:105) = S(t, t − k∆t)nk(t) = S(t, t − k∆t)n0(t − k∆t) . (B7) We can now take the limit ∆t → 0 in (B6) and find (cid:104)n0(t)(cid:105) ∆t → N t −∞ P (t, t)A(t) dt(cid:48), ∆t → 0, (B8) Figure 6. Illustration of negative correlations between δmk(t) and δmk+1(t + ∆t). (a): Expected number of spikes from group k in bins [t, t + ∆t] and [t + ∆t, t + 2∆t]. (b): A large fluctuation of δmk(t) leads to reduction of the number of avail- able spikes mk+1(t + ∆t). As a consequence, δmk+1(t + 1) tends to be small. {nk(t)}k=1,2,...). But since we do not know the exact value of mk(t) we need to average (cid:104)δmk(t)(cid:105) = (cid:104)(cid:104)δmk(t)(cid:105)mk(t)(cid:105) = ρ(t, t − k∆t) · ∆t · (cid:104)mk(t)(cid:105) (B5) where we have used (B4). We now use this result to calculate the expected number of spikes in the interval [t, t + ∆t]. Averaging over (B3) yields ∞(cid:88) where P (t, t) = ρ(t, t)S(t, t) is the inter-spike-interval density. Eq. (B8) is equivalent to Eq. (1). q ∈ Z To obtain the correlation function we can write for (cid:104)n0(t)n0(t + q∆t)(cid:105) = (cid:104)δmk(t)δml(t + q∆t)(cid:105). (B9) Here, the spike numbers δmk(t) and δml(t + q∆t) that refer to different groups k and l − q are uncorrelated. Correlations only arise for δmk(t) and δmk+q(t + q∆t), i.e. spikes that refer to the same group in the past. Thus, (cid:104)n0(t)n0(t + q∆t)(cid:105) = (cid:104)δmk(t)(cid:105)(cid:104)δml(t + q∆t)(cid:105) ∞(cid:88) k,l=1 ∞(cid:88) k,l=1 (cid:104)n0(t)(cid:105) = ρ(t, t − k∆t) · ∆t · (cid:104)mk(t)(cid:105) . (B6) (cid:104)δmk(t)(cid:105)(cid:104)δmk+q(t + q∆t)(cid:105), (B10) k=1 k=1 + (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) ∞(cid:88) ∞(cid:88) k=1 − (a)singlepopulationnetworks(b)excitation-inhibitionnetwork10-1100101102frequency [Hz]0.51.0spectral density [0.01/s]Js=5Js=0Js=-510-1100101102frequency [Hz]01234spectral density [0.01/s]exc.crossinh.(a)(b) ∞(cid:88) so that the covariance is (cid:104)∆n0(t)∆n0(t + q∆t)(cid:105) = ∞(cid:88) k=1 (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) − (cid:104)δmk(t)(cid:105)(cid:104)δmk+q(t + q∆t)(cid:105), (B11) k=1 where ∆n0(t) = n0(t) − (cid:104)n0(t)(cid:105). Therefore we need to compute (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) . (B12) To this end, let us consider the cases q = 0 and q > 0 separately. For q = 0 and large N, the number of neurons that spike in [t, t + ∆t] and had their last spike at t − k∆t is a Poisson variable with mean and variance ∆t P (t, t − k∆t) nk(t). Thus (B12) becomes (cid:10)[δmk(t)]2(cid:11) = ∆t P (t, t − k∆t) nk(t) + O(∆t3). (B13) 9 in bin t− k∆t that survived until time t. Thus mk(t) can be regarded as a binomially distributed random number with n = nk(t) trials and survival probability p = S(t, t− k∆t). This random number has mean np and variance np(1 − p). Hence, the second moment reads (cid:104)[mk(t)]2(cid:105) = (cid:104)[mk(t) − (cid:104)mk(t)(cid:105)]2(cid:105) + (cid:104)mk(t)(cid:105)2 O(∆t2) (cid:124) (cid:123)(cid:122) (cid:125) = nk(t)S(t, t − k∆t) [1 − S(t, t − k∆t)] + O(∆t2). (B17a) Likewise, (cid:104)[mk+1(t + ∆t)]2(cid:105) = nk(t)S(t + ∆t, t − k∆t) × [1 − S(t + ∆t, t − k∆t)] + O(∆t2) (B17b) because nk+1(t+∆t) = nk(t). Inserting (B17) into (B16) we find For q > 0, we employ (B2) twice and obtain (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) = (cid:104)mk(t) mk+q(t + q∆t)(cid:105) − (cid:104)mk+1(t + ∆t) mk+q(t + q∆t)(cid:105) − (cid:104)mk(t) mk+q+1(t + (q + 1)∆t)(cid:105) (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) = nk(t) ∆t2 × S(t + (q + 1)∆t, t − k∆t) − S(t + q∆t, t − k∆t) × S(t, t − k∆t) − S(t + ∆t, t − k∆t) ∆t ∆t + (cid:104)mk+1(t + ∆t) mk+q+1(t + (q + 1)∆t)(cid:105). (B14) = −P (t + q∆t, t − k∆t)P (t, t − k∆t) nk(t) ∆t2 , . (B18) Here, we have identified the derivative d/dtS(t, t) = −P (t, t). Note that this expression is of order O(∆t3), whereas (cid:104)δmk(t)(cid:105)(cid:104)δmk+q(t+q∆t)(cid:105) is of order O(∆t4). So we can neglect the second term on the right-hand side of Eq. (B11). Putting all together, we find 1 ∆t2(cid:104)∆n0(t)∆n0(t(cid:48) = t + q∆t)(cid:105) ∞(cid:88) (cid:34) ∞(cid:88) k=1 k=1 = 1 ∆t2 (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) δq,0 P (t, t − k∆t)nk(t) + O(∆t) (cid:35) P (t + q∆t, t − k∆t)P (t, t − k∆t)nk(t) 1 ∆t ∞(cid:88) = − k=1 t −∞ P (t, t(cid:48)(cid:48))A(t(cid:48)(cid:48)) dt(cid:48)(cid:48) − N P (t, t(cid:48)(cid:48))P (t(cid:48), t(cid:48)(cid:48))A(t(cid:48)(cid:48)) dt(cid:48)(cid:48). (B19) −−−−→ ∆t→0 N δ(t − t(cid:48)) t −∞ t A(t) = −∞ In order to evaluate each of these four correlators, we note that the probability that a neuron from group k "survives" until time t + q∆t given that it survived until time t is S(t + q∆t, t − k∆t)/S(t, t − k∆t) according to Bayes law. Thus, out of the mk(t) neurons that survived until time t, on average (cid:104)mk+q(t + q∆t)(cid:105) mk(t) = S(t + q∆t, t − k∆t) S(t, t − k∆t) ·mk(t), (B15) also survive until t + q∆t. Therefore, the correlator for 0 ≤ l < q can be written as (cid:68) (cid:104)mk+l(t + l∆q)mk+q(t + q∆t)(cid:105) = = mk+l(t + l∆q)(cid:104)mk+q(t + q∆t)(cid:105)mk+l(t+l∆t) (cid:69) ·(cid:10)m2 k+l(t + l∆t)(cid:11) . , S(t + q∆t, t − k∆t) S(t + l∆t, t − k∆t) = Applying this result to (B14), we obtain (cid:104)δmk(t)δmk+q(t + q∆t)(cid:105) = [S(t + (q + 1)∆t, t − k∆t) − S(t + q∆t, t − k∆t)] (cid:18) (cid:104)[mk+1(t + ∆t)]2(cid:105) S(t + ∆t, t − k∆t) × − (cid:104)[mk(t)]2(cid:105) S(t, t − k∆t) (cid:19) How can we calculate the second moment of mk(t)? Recall that mk(t) is the part of the nk(t) neurons firing (B16) Thus, using (B8), we arrive at the final result P (t, t(cid:48))A(t(cid:48)) dt(cid:48) + δA(t) (B20) 10 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)∆g=0 The perturbation is given by ∞ −∞ · ∆g(s, t) ds. (cid:12)(cid:12)(cid:12)(cid:12)∆g=0 t ρ0(t(cid:48) − t)e∆g(t(cid:48),t)dt(cid:48)(cid:17) t δS(t, t) δ∆g(s, t) (cid:16)− ´ δ∆g(s, t) The functional derivative at ∆g = 0 reads (cid:12)(cid:12)(cid:12)(cid:12)∆g=0 δS(t, t) δ∆g(s, t) δ exp = = −θ(t − s)θ(s − t)S0(t, t)ρ0(s − t). We insert this expression to compute the integral in (C2) t −∞ − ∞ ∞ −∞ −∞ = [8] = t −∞ ∆S(t, t)dt = −∞ × S0(t − t)ρ0(s − t ∞ θ(s − t)θ(t − s) ∞ (cid:124)(cid:123)(cid:122)(cid:125) θ(x)S0(t−s+x)ρ0(x)∆g(s, s−x)dxds ∞ )∆g(s, t) ds dt x(t) 0 θ(x)S0(t − s + x)ρ0(x) dx ds . ∆g(s) θ(t − s) θ(t−s) (cid:124) 0 (cid:123)(cid:122) L(t−s) (cid:125) In the last step, we have used the approximation Eq. (8), ∆g(t, t) = ∆g(t) = [(κJs − γ) ∗ ∆A](t), which has no dependence on the time of the last spike. Hence (C2) becomes a(t) = A0 + (P0 ∗ ∆A)(t) + A0 (L ∗ ∆g)(t). (C3) d dt This linearized equation for the mean activity is valid for any small ∆A and ∆g in the past, also if due to finite size fluctuations. Eqs. (C1)-(C3) generalize to additional time-dependent inputs ∆I(t) by extending the definition of ∆g to ∆g(t) = [(κJs − γ) ∗ ∆A + κ ∗ ∆I](t). Furthermore, the noise correlation function Eq. (B21) becomes to leading order for τ ≥ 0. In the latter expression we extended the limits of integration to infinity, which is possible if we assume P (t, t(cid:48)) = 0 for t < t(cid:48). Equation (B21) tells us that the noise correlation function consists of two parts: a white (δ-correlated) part and a negative correlation due to neural refractoriness. Intuitively, since δmk(t) and δmk+1(t + ∆t) share the same number of available neu- rons mk(t), a positive fluctuation of the number of spikes in the bin t of the neurons of group k reduces the number of neurons with last spike in bin k more than on average. Thus, the number of neurons of group k that can still fire in time bin t + ∆t is smaller than on average, which explains the negative correlations (Fig. 6). Appendix C: Detailed derivation of Eq. (9) t ´ We aim to linearize the population integral a(t) = −∞ P (t, t)A(t) dt, Eq. (1), around an equilibrium point A0, with small fluctuations ∆A(t) = A(t) − A0 with a mean of zero. To this end, let us first note that the haz- ard function Eq. (7) then can be written as ρ(t, t) = ceh0−ϑ0(t−t)e∆h(t)−∆ϑ(t,t) = ρ0(t − t)e∆g(t,t). (C1) (cid:16)− (cid:17) t ´ t ρ(s, t)ds Furthermore, recall that P (t, t) = −d/dtS(t, t), where . Expanding to first order S(t, t) = exp in ∆A yields S(t, t) = S0(t − t) + ∆S(t, t) and P (t, t) = P0(t− t)− d/dt∆S(t, t), where S0(t) = exp(− 0 ρ0(t(cid:48))dt(cid:48)) and P0(t) = −d/dtS0(t) define the zeroth-order terms. Thus, the linearized population integral reads ´ t a(t) = A0 + (P0 ∗ ∆A)(t) − A0 d dt where δA(t) is Gaussian with conditional correlation function where we used the normalization of P0(t) and the bound- ary condition ∆S(t, t) = 0. (cid:104)δA(t)δA(t + τ )(cid:105) = N−1δ(τ ) P (t, t(cid:48))A(t(cid:48)) dt(cid:48) ∆S(t, t) = − N−1 P (t + τ, t(cid:48))P (t, t(cid:48))A(t(cid:48)) dt(cid:48) (B21) ∞ −∞ ∞ −∞ t −∞ ∆S(t, t) dt, (C2) (cid:104)ξ(t + τ )ξ(t)(cid:105) = δ(τ )− P0(t + τ − t)P0(t− t) dt, with(cid:112)A0/N ξ(t) = δA(t). Eqs. (C3) and (C4) are equiv- (C4) alent to Eq. (9). t −∞ [1] M. L. Klein and W. Shinoda, Science 321, 798 (2008). [2] G. S. Ayton, W. G. Noid, and G. A. Voth, Curr Opin Struct Biol 17, 192 (2007). [4] R. A. Pielke, Mesoscale Meteorological Modeling (Aca- demic Press, 2002). [5] M. Benaïm and J.-Y. Le Boudec, Perform Evaluation 65, [3] C. Peter and K. Kremer, Soft Matter 5, 4357 (2009). 823 (2008). [6] W. Gerstner, H. Sprekeler, and G. Deco, Science 338, 60 (2012). [7] S. Lefort, C. Tomm, J.-C. F. Sarria, and C. C. H. Pe- tersen, Neuron 61, 301 (2009). [8] D. H. Hubel and T. N. Wiesel, J Physiol 160, 106 (1962). [9] S. Mensi, R. Naud, C. Pozzorini, M. Avermann, C. C. H. Petersen, and W. Gerstner, J Neurophysiol 107, 1756 (2012). [10] N. Brunel and V. Hakim, Neural Comput 11, 1621 (1999). [11] W. Gerstner, Neural Comput 12, 43 (2000). [12] J. W. Pillow, J. Shlens, L. Paninski, A. Sher, A. M. Litke, E. J. Chichilnisky, and E. P. Simoncelli, Nature 454, 995 (2008). [13] V. Frost and B. Melamed, IEEE C M 32, 70 (1994). [14] R. E. Mirollo and S. H. Strogatz, SIAM J Appl Math 50, [15] M. Mattia and P. Del Giudice, Phys Rev E 66, 051917 1645 (1990). (2002). e1002872 (2013). put 21, 1203 (2009). [18] A. G. Hawkes, Journal of the Royal Statistical Society. Series B (Methodological) 33, 438 (1971). Phys Rev E 85, 031916 (2012). [20] M. Helias, T. Tetzlaff, and M. Diesmann, New J Phys 15, 023002 (2013). [21] H. Soula and C. C. Chow, Neural Comput 19, 3262 [22] M. A. Buice and J. D. Cowan, Phys Rev E 75, 051919 (2007). (2007). [23] P. C. Bressloff, SIAM J Appl Math 70, 1488 (2009). [24] M. Benayoun, J. D. Cowan, W. van Drongelen, E. Wallace, PLoS Comput Biol 6, e1000846 (2010). [25] J. D. Touboul and G. B. Ermentrout, J Comput Neurosci and 11 164 (2013). e1002711 (2012). Hall, 1980). [36] T. Schwalger, The statistics of non-renewal neuron models, Ph.D. thesis, Humboldt- Universität zu Berlin (2013). interspike-interval [37] T. Schwalger and B. Lindner, Front Comput Neurosci 7, [38] R. Naud and W. Gerstner, PLoS Comput Biol 8, [39] D. Cox and V. Isham, Point processes (Chapman and [40] H. R. Wilson and J. D. Cowan, Biophys J 12, 1 (1972). [41] M. Deger, M. Helias, S. Cardanobile, F. M. Atay, and S. Rotter, Phys Rev E 82, 021129 (2010). [42] W. Truccolo, L. R. Hochberg, and J. P. Donoghue, Nat Neurosci 13, 105 (2010). [43] C. Meyer and C. van Vreeswijk, Neural Comput 14, 369 [44] Y. H. Liu and X. J. Wang, J Comput Neurosci 10, 25 (2002). (2001). [45] R. Jolivet, A. Rauch, H.-R. Lüscher, and W. Gerstner, [46] E. Zohary, M. N. Shadlen, and W. T. Newsome, Nature [47] D. J. Mar, C. C. Chow, W. Gerstner, R. W. Adams, and J. J. Collins, Proc Natl Acad Sci U S A 96, 10450 (1999). [48] H. Sompolinsky, H. Yoon, K. Kang, and M. Shamir, [49] X.-J. Wang, Neuron 36, 955 (2002). [50] G. Deco and E. T. Rolls, Eur J Neurosci 24, 901 (2006). [51] A. Renart, J. de la Rocha, P. Bartho, L. Hollender, N. Parga, A. Reyes, and K. D. Harris, Science 327, 587 (2010). [52] R. L. Stratonovich, Topics in the theory of random noise I, Mathematics and its applications; vol. 3 (New York: Gordon and Breach, 1963). [53] C. E. Shannon, Bell Syst. Tech. J. 27, 379 (1948). [54] F. Gabbiani, Network Comp Neural 7, 61 (1996). [55] A. Borst and F. E. Theunissen, Nat Neurosci 2, 947 [16] M. A. Buice and C. C. Chow, PLoS Comput Biol 9, J Comput Neurosci 21, 35 (2006). [17] T. Toyoizumi, K. R. Rad, and L. Paninski, Neural Com- 370, 140 (1994). [19] V. Pernice, B. Staude, S. Cardanobile, and S. Rotter, Phys Rev E 64, 051904 (2001). [26] G. Dumont, G. Northoff, and A. Longtin, Phys Rev E [56] M. J. Chacron, B. Lindner, and A. Longtin, Phys Rev [27] F. Lagzi and S. Rotter, Front Comput Neurosci 8, 142 [57] R. D. Vilela and B. Lindner, Phys Rev E 80, 031909 [28] B. Lindner, B. Doiron, and A. Longtin, Phys Rev E 72, [58] M. J. Chacron, L. Maler, and J. Bastian, Nat Neurosci [29] O. Ávila Åkerberg and M. J. Chacron, Phys Rev E 79, [59] R. Krahe, J. Bastian, and M. J. Chacron, J Neurophysiol Lett 92, 080601 (2004). (1999). (2009). 8, 673 (2005). 100, 852 (2008). 31, 453 (2011). 90, 012702 (2014). (2014). 061919 (2005). 011914 (2009). [30] J. Trousdale, Y. Hu, E. Shea-Brown, and K. Josić, PLoS Comput Biol 8, e1002408 (2012). [60] M. J. E. Richardson, Phys Rev E 80, 021928 (2009). [61] R. Rosenbaum, J. Rubin, and B. Doiron, PLoS Comput [31] J. Benda and A. V. M. Herz, Neural Comput 15, 2523 Biol 8, e1002557 (2012). [32] B. N. Lundstrom, M. H. Higgs, W. J. Spain, and A. L. 34, 285 (2013). Fairhall, Nat Neurosci 11, 1335 (2008). [63] F. Droste, T. Schwalger, and B. Lindner, Front Comput [33] C. Pozzorini, R. Naud, S. Mensi, and W. Gerstner, Nat Neurosci 7, 86 (2013). Neurosci 16, 942 (2013). [34] F. Farkhooi, A. Froese, E. Muller, R. Menzel, and M. P. Nawrot, PLoS Comput Biol 9, e1003251 (2013). [35] S. A. Prescott and T. J. Sejnowski, J Neurosci 28, 13649 [64] A. El Hady, G. Afshar, K. Bröking, O. M. Schlüter, T. Geisel, W. Stühmer, and F. Wolf, Front Neural Cir- cuits 7, 167 (2013). [65] M.-O. Gewaltig and M. Diesmann, Scholarpedia 2, 1430 [62] N. Sharafi, J. Benda, and B. Lindner, J Comput Neurosci (2003). (2008). (2007).
1602.03493
1
1602
2016-02-10T19:26:32
Winning versus losing during gambling and its neural correlates
[ "q-bio.NC" ]
Humans often make decisions which maximize an internal utility function. For example, humans often maximize their expected reward when gambling and this is considered as a "rational" decision. However, humans tend to change their betting strategies depending on how they "feel". If someone has experienced a losing streak, they may "feel" that they are more likely to win on the next hand even though the odds of the game have not changed. That is, their decisions are driven by their emotional state. In this paper, we investigate how the human brain responds to wins and losses during gambling. Using a combination of local field potential recordings in human subjects performing a financial decision-making task, spectral analyses, and non-parametric cluster statistics, we investigated whether neural responses in different cognitive and limbic brain areas differ between wins and losses after decisions are made. In eleven subjects, the neural activity modulated significantly between win and loss trials in one brain region: the anterior insula ($p=0.01$). In particular, gamma activity (30-70 Hz) increased in the anterior insula when subjects just realized that they won. Modulation of metabolic activity in the anterior insula has been observed previously in functional magnetic resonance imaging studies during decision making and when emotions are elicited. However, our study is able to characterize temporal dynamics of electrical activity in this brain region at the millisecond resolution while decisions are made and after outcomes are revealed.
q-bio.NC
q-bio
Winning versus losing during gambling and its neural correlates P. Sacr´e∗, M. S. D. Kerr∗, S. Subramanian†, K. Kahn∗, ∗The Johns Hopkins University, Baltimore, MD. †University of Cambridge, UK. ‡Cleveland Clinic, Cleveland, OH. J. Gonzalez-Martinez‡, M. A. Johnson‡, J. T. Gale‡, and S. V. Sarma∗ 6 1 0 2 b e F 0 1 ] . C N o i b - q [ 1 v 3 9 4 3 0 . 2 0 6 1 : v i X r a Abstract -- Humans often make decisions which maximize an internal utility function. For example, humans often maximize their expected reward when gambling and this is considered as a "rational" decision. However, humans tend to change their betting strategies depending on how they "feel". If someone has experienced a losing streak, they may "feel" that they are more likely to win on the next hand even though the odds of the game have not changed. That is, their decisions are driven by their emotional state. In this paper, we investigate how the human brain responds to wins and losses during gambling. Using a combination of local field potential recordings in human subjects performing a financial decision-making task, spectral analyses, and non-parametric cluster statistics, we investigated whether neural responses in different cognitive and limbic brain areas differ between wins and losses after decisions are made. In eleven subjects, the neural activity modulated significantly between win and loss trials in one brain region: the anterior insula (p = 0.01). In particular, gamma activity (30 -- 70 Hz) increased in the anterior insula when subjects just realized that they won. Modulation of metabolic activity in the anterior insula has been observed previously in functional magnetic resonance imaging studies during decision making and when emotions are elicited. However, our study is able to characterize temporal dynamics of electrical activity in this brain region at the millisecond resolution while decisions are made and after outcomes are revealed. I. INTRODUCTION Decision-making links cognition to behavior and is a key driver of human personality, fundamental for survival, and essential for our ability to learn and adapt. It has been well- established that humans often make emotional-based deci- sions [1]. Thus, psychiatric patients who have dysfunctional cognitive circuitry, frequently have alterations in decision- making that are poorly understood. Understanding the neural basis of decision-making is there- fore essential toward patient management. However, access to the human brain has been limited to a few case studies wherein subjects have lesions in the key decision-making structures such as the orbital frontal cortex [2] -- [4], or wherein func- tional magnetic resonance imaging (fMRI) is used to measure neural activity in several healthy subjects during decision- making [5]. Both of these approaches have limitations. Lesions don't provide actual neural data to ascertain a specific brain regions's role during behavior. Rather, the region's function is inferred by absence of behaviors from lesioned subjects when compared to healthy subjects. On the other hand, fMRI provides a correlate of neural activity (metabolic activity) but suffers from poor temporal resolution. fMRI resolution is on the order of multiple seconds, while decisions are often made on the order of tenths of a second. Here, we took advantage of a technique called stereoelec- troencephalography (SEEG) that allowed us to record high temporal resolution electrophysiological data (electrical activ- ity at the millisecond scale) directly from deep and peripheral brain regions in human subjects while they performed a gambling task. These subjects are implanted with electrodes for clinical purposes, and each contact in the brain generates a local field potential (LFP) signal. For our gambling task, eleven subjects played a game of high card where they won virtual money if their card was higher than the computer's card. On each trial, subjects had to decide to bet "high" ($20) or "low" ($5) on their card being higher than the hidden computer's card. Eighty percent of the trials lead to clear rational decisions where the expected reward for one choice is higher than another. On twenty percent of the trials, however, there is no clear rational decision since the expected reward is equal for both choices. Here, we focused on characterizing differences in brain activity between when a subject realizes that he/she has won versus when he/she has lost. Thus, the goal of this study is to assess the role of different brain regions in responding to outcomes of gambling decisions. To identify neural correlates that encode a subject's response to a win versus a loss, we computed spectrograms for each brain region (i.e., electrode channel) and each patient across all trials when subjects won and compared them to spectrograms when subjects lost. Specifically, we examined spectral content before, during, and after the computer's card was shown (and hence outcome is revealed) and implemented a non-parametric cluster statistic to test whether the spectral activity differed between wins and losses in the given brain region. The non- parametric cluster test generates a cluster defined by a set of adjacent time-frequency windows that gives rise to the smallest p-value. If this p-value is smaller than 0.05, then the brain region (defined by location of the channel) was designated as encoding response to gambling outcome. limbic and hippocampal networks, We found that upon examining several brain regions in cognitive, the anterior insula encoded information that separated wins from losses. Specifically, when gamma activity (30 -- 70 Hz), was prevalent in this region, then the player was more likely to have won after he/she sees the computer's card. The anterior insula has been implicated in the role of emotions in risk-related decision making [6] as well as being a neural basis for encoding how someone "feels" [7]. But no study has had access to electrical activity from these regions at a millisecond time-scale. Hence, these findings are the first to show evidence of temporal dynamics of rhythms in the brain at a fine resolution when decisions are made and when subjects "feel" the consequences of their decisions. A. Subjects II. METHODS At the Cleveland Clinic, patients with medically intractable epilepsy routinely undergo SEEG recordings in order to lo- calize the seizure focus. See II-B for details on the SEEG procedure. In this study, aside from the behavioral exper- iments, no alterations were made to the patient's clinical care, including the placement of the electrodes [8]. Subjects enrolled voluntarily and gave informed consent under criterion approved by the Cleveland Clinic Institutional Review Board. A total of eleven subjects volunteered to perform the task. Details on these recordings and eventual annotated seizure focus of these eleven patients are noted in Table I. Subjects were implanted with 8 to 13 depth electrodes. Implantation was performed using robot-assisted surgery along with co-registered functional MRIs and angiograms to ensure safe implantation [9]. Once inserted, SEEG electrophysiologi- cal data were acquired using a Nihon Kohden 1200 EEG diag- nostic and monitoring system (Nihon Kohden America, USA) at a sampling rate of 2 kHz. Behavioral event data were si- multaneously acquired through the MonkeyLogic MATLAB® toolbox [10]. There are standard concerns in analyzing data from epileptic patients. First, patients are often on medication, which might affect the neurophysiology of the brain. For clinical purposes, patients were kept off of their anti-seizure medication for their entire stay at Cleveland Clinic, so these effects would be minimized. Secondly, actual seizures might the neurophysiology around the seizure focus. Human epilepsy recordings are taken to localize the seizure focus, so overlap is expected between seizure focus and areas recorded. impact B. Stereoelectroencephalography The Cleveland Clinic is a world-renowned center for the evaluation and treatment of epilepsy and brain tumors, as- sessing around 9500 patients every year from all 50 states and more than 10 countries. More than 400 epilepsy surgeries are performed every year, including a growing number of stereotactically implanted depth electrodes or stereoelectroen- cephalography (SEEG) [11] -- [14]. This surgical procedure was developed in France [15] -- [27], and brought to the United States by Dr. Jorge Gonzalez-Martinez [28]. The innovative approach using SEEG methodology relies on its capability in accessing large-scale networks, providing precise human brain data, from cortical to subcortical areas, in three-dimensional fashion. In routine placement of depth electrodes, burr-holes that are each 15 mm in diameter are required for safe vi- sualization of cortical vessels, and therefore only a small TABLE I THIS TABLE PROVIDES CLINICALLY RELEVANT INFORMATION ON EACH SUBJECT: THE GENDER, THE AGE IN YEARS, AND THE DURATION OF EPILEPSY ("DUR.") IN YEARS. THIS TABLE ALSO PROVIDES THE NUMBER OF WIN, LOSS AND DRAW TRIALS. ID Gender male female female female female female female female male female 1 2 3 4 5 6 7 8 9 10 11 male Age [yr.] 26 41 55 31 53 60 36 23 32 32 28 Dur. [yr.] 3 38 52 13 23 8 36 5 11 13 11 Win Loss Draw 73 79 23 53 50 63 56 46 65 57 67 73 58 34 57 61 88 68 50 59 62 78 39 25 15 34 25 21 33 36 36 35 37 number of electrodes are placed. SEEG placement, however, uses several small drill holes (1.8 mm in diameter), allowing many electrodes to be inserted. Since direct visualization of the cortical surface is not pos- sible with small drills (Fig. 1A), the SEEG technique requires detailed pre-procedural vascular mapping using pre-operative imaging with magnetic resonance angiography (MRA) and cerebral angiography. Angiography is an X-ray examination of the blood vessels. The mapping procedure is performed under fluoroscopy using general anesthesia, and an expert neuro-anesthesiologist correctly titrates anesthesia to permit measurement of intracranial electroencephalography (EEG). The number and location of implanted electrodes are pre- operatively planned based on a pre-implantation hypothesis, which is formulated in accordance with non-invasive pre- implantation data, such as seizure semiology, ictal and inter- ictal scalp EEG, magnetic resonance imaging (MRI), positron emission tomography (PET) and ictal single-photon emission computed tomography (SPECT) scans. Thus, the implanta- tion strategy has the goal of accepting or rejecting the pre- implantation hypothesis of the location of the epileptogenic zone (EZ). SEEG provides a more complete coverage of the brain, from lateral, intermediate and/or deep structures in a three- dimensional arrangement recorded over hundreds of channels. Using strict techniques, this procedure is safe and minimally invasive: only 1/1176 implantations last year resulted in an asymptomatic intracranial hemorrhage. The rate of complica- tions in SEEG implantations is less than 1 % [13], [29]. C. Gambling task Subjects performed the gambling task in their Epilepsy Monitoring Unit room. The task was displayed via a computer screen and the subject interacted with the task using an InMo- tion2 robotic manipulandum (Interactive Motion Technologies, Fig. 1. Imaging fusion and placement of multiple electrodes using the SEEG method. Fig. A is a photograph showing 14 electrodes at the skin surface. Fig. B is a fluoroscopy image of an SEEG-implanted subject (coronal view with eye forward). Note the precise parallel placement, with tips terminating at the midline or dural surface. USA). The manipulandum is controlled by the subject's hand and allows for 2D planar motion, which translated directly to the position of a cursor on screen. The gambling task (Fig. 2A) is based on a simple game of high card where subjects would win virtual money if their card beat the computer's card. Specifically, in the beginning of each trial, the subject controlled a cursor via a planar manipulandum to a fixation target. Afterwards, the subject is shown his card (2, 4, 6, 8, or 10) that is randomly chosen with equal distribution. The computer's card is initially hidden. The screen then shows their two choices: a high bet ($20) or a low bet ($5). The subject has 6 seconds to select one with his cursor. Following selection, the computer's card, which follows the same distribution, is revealed. The final screen depicts the amount won or lost. D. Data analysis All electrophysiological and behavioral analyses were con- ducted offline using custom MATLAB® scripts. Data for 16 anatomical regions in cognitive, limbic and hip- pocampal networks were separated into win and loss trials. For each brain region, differences in the neural responses between the task conditions during the 250 ms before and 1000 ms after the computer card were examined by means of a non- parametric cluster statistic. Specifically, spectrograms were constructed for each trial time-locked to when the computer's card is shown. Then the spectrograms for win trials were compared to those for loss trials. To see if spectrograms for each group were statistically significantly different, we used a nonparametric cluster-based test [30]. Clusters are defined as a set of adjacent time-frequency windows whose activity is statistically significant between trials where the subject ends up winning versus losing. 1) Spectral analysis: We calculated the oscillatory power using multitapers from the Chronux toolbox [31]. We used three orthogonal tapers with a 300 ms window sliding at 50 ms steps. We dropped frequencies under 10 Hz because of the Nyquist criterion and analyzed upwards to 100 Hz. Afterwards, Fig. 2. Gambling task and behavioral results. (A) Timeline of the behavioral task. After fixation, subjects were shown their card. Once the bets were shown, subjects selected one of the choices and then were shown the computer's card following a delay. Feedback was provided afterwards by displaying the amount won or lost. (B) Average bet decisions across cards. Subjects predominantly bet low for 2 and 4 cards and bet high for 8 and 10 cards. There was no predominant strategy for 6 cards, which had bout 33 % chance of eliciting a high bet. (C) Reaction times across cards. Subjects reacted faster for cards whose rewards had lower variability. we normalize each frequency bin's power by first taking the natural log of the power in each frequency bin, and then performing a Gaussian normalization based on the power in each frequency bin over the entire recording session. 2) Non-parametric cluster statistical test: Significant dif- ferences between the neural response data in each anatomical region are defined by a non-parametric cluster statistic run on data aggregated from trials by all relevant subjects [30]. This test leverages the dependency between adjacent time- frequency windows in order to avoid over-penalizing with multiple comparison corrections. For each time-frequency window in the spectrogram, a null distribution was created by shuffling these wins and loss labels 1000 times between trials within each subject. Within each shuffle, the average difference between the newly labeled win and loss trial spectrograms was calculated. A p-value was assigned for each window by comparing the difference acquired from the true labels with the distribution of differences acquired from the shuffled labels. Clusters were formed by grouping windows with significant p-values (p < 0.05) that were adjacent in either time or frequency. The test statistic for each cluster was calculated by taking the sum of the log of the p-values for each window in the cluster. This prioritizes clusters that both have strong differences as well as large sizes. A null distribution of cluster statistics was created using the same process but with the 1000 spectrograms obtained from the originally shuffled labels. The observed cluster statistic was then compared against this null distribution of cluster statistics in order to obtain the final p-value of the test. AB$5$20FixationShow CardGo-CueDelayShow DeckFeedback<8s2s<6s0.35-0.6s1.3-1.55s1.3sHigh bet (%)0100246810Cardʼs valueReaction time (s)246810Cardʼs value01.5Win $5!ABC501.00.5 Fig. 3. Non-parametric cluster statistic for different brain regions. For each brain region, differences in the neural responses between the task conditions during the 250 ms before and 1000 ms after the show computer card epoch were summarized by means of a non-parametric cluster statistic (p-value). In addition, we provide the number of subjects with recording in this area (Subj.), and the number of win and loss trials. III. RESULTS This section summarizes the main findings of our analysis. First, we show the p-value associated with 16 brain regions. Then, we focus on the brain region identified in our ex- ploratory analysis and show time-frequency differences of the neural responses around the show computer card epoch. While multiple brain regions appear to show some response (data not shown), the two task conditions only significantly differ after the computer card is shown in one brain region: anterior insula (p = 0.01) (see Fig. 3). All p-values are computed from a non-parametric cluster statistic described in Section II. Differences in the neural responses around the show com- puter card epoch was examined for the anterior insula (Fig. 4). Spectrograms of the neural responses between win and loss conditions show large differences in the gamma band (30 -- 70 Hz) after the computer card is shown and the gambling outcome is revealed. We summarize the neural activity in the time-frequency domain by averaging the responses over 70 -- 100 Hz. Gamma activity increases in anterior insula quickly around 250 ms after the computer card epoch and then decreases slowly until 1000 ms after the computer card epoch (Fig. 5). The gamma activity for win trials decreases more slowly than gamma activity for loss trials. Fig. 4. Differences in the neural responses during the 250 ms before and 1000 ms after the show computer card epoch for anterior insula. (First and second rows) Spectrograms of the neural responses show differences in the time-frequency domain between win and loss conditions. (Third row) The cluster emphasizes the region of the time-frequency domain where the neural responses show significant differences between win and loss conditions (p < 0.05). Fig. 5. Gamma activity (30 -- 70 Hz) increases in anterior insula quickly around 250 ms after the computer card epoch and then decreases slowly until 1000 ms after the computer card epoch. The gamma activity for win trials decreases more slowly than gamma activity for loss trials. amygdala 2 124 140angular gyrus 8 444 495anterior cingulate 3 173 207anterior insula 2 103 118cuneus 10 582 627hippocampus 10 576 620hippocampus (posterior) 4 225 226 inferior frontal gyrus 6 362 384inferior temporal gyrus 10 576 620insula 7 423 420intraparietal sulcus 6 374 399middle frontal gyrus 5 270 308orbitofrontal cortex 4 249 254parieto-occipital sulcus 7 397 433posterior cingulate 9 552 592precuneus 9 517 568 010.05Brain region Subj. Win Loss p-valueclusterfrequency (Hz)time relative to show deck (s)-0.6-0.4-0.200.20.40.6frequency (Hz)frequency (Hz)clusterloss trialswin trialstime relative to show deck (s)gamma activity (-)losswin IV. CONCLUSION In our subject population, the neural activity modulated significantly between win and loss trials in the anterior in- sula (p = 0.01). In particular, gamma-band activity increased around 500 ms after the show computer card for win trials. The anterior insula has been implicated in the role of emotions in risk-related decision making [6] as well as being a neural basis for encoding how someone "feels" [7], but no study has had access to electrical activity from these regions at a millisecond time-scale. Hence, these findings are the first to show evidence of temporal dynamics of rhythms in the brain at a fine resolution when decisions are made and when subjects "feel" the consequences of their decisions. Still, the exact mechanisms by which this region in involved in risk-based decision making are not known and will therefore remain the focus of future work. ACKNOWLEDGMENT The authors would like to thank J. Bulacio, J. Jones, H.-J. Park and S. Thompson (Cleveland Clinic, Cleveland, OH) for facilitating experiments, collecting, de-identifying and transferring data, and providing anatomical labels. This work was supported by a National Science Foundation grant (EFRI- MC3: # 1137237) awarded to S.V.S., J.A.G., J.B. and J.T.G. REFERENCES [1] M. Toda, "Emotion and decision making," Acta Psychologica, vol. 45, no. 1 -- 3, pp. 133 -- 155, 1980. [2] A. Bechara, "The role of emotion in decision-making: Evidence from neurological patients with orbitofrontal damage," Brain and Cognition, vol. 55, no. 1, pp. 30 -- 40, 2004. Development of Orbitofrontal Function. [3] J. Kim and M. E. Ragozzino, "The involvement of the orbitofrontal cortex in learning under changing task contingencies," Neurobiology of Learning and Memory, vol. 83, no. 2, pp. 125 -- 133, 2005. [4] I. Bohn, C. Giertler, and W. Hauber, "Orbital prefrontal cortex and guidance of instrumental behaviour in rats under reversal conditions," Behavioural Brain Research, vol. 143, no. 1, pp. 49 -- 56, 2003. [5] N. K. Logothetis, "What we can do and what we cannot do with fmri," Nature, vol. 453, pp. 869 -- 878, 06 2008. [6] A. G. Sanfey, J. K. Rilling, J. A. Aronson, L. E. Nystrom, and J. D. Cohen, "The neural basis of economic decision-making in the ultimatum game," Science, vol. 300, no. 5626, pp. 1755 -- 1758, 2003. [7] A. D. B. Craig, "How do you feel -- now? the anterior insula and human awareness," Nat Rev Neurosci, vol. 10, pp. 59 -- 70, Jan. 2009. [8] M. A. Johnson, S. Thompson, J. Gonzalez-Martinez, H.-J. Park, J. Bula- cio, I. Najm, K. Kahn, M. Kerr, S. V. Sarma, and J. T. Gale, "Performing behavioral tasks in subjects with intracranial electrodes," Journal of Visualized Experiments, no. 92, p. e51947, 2014. [9] D. Ongur and J. Price, "The organization of networks within the orbital and medial prefrontal cortex of rats, monkeys and humans," Cerebral Cortex, vol. 10, no. 3, pp. 206 -- 219, 2000. [10] W. F. Asaad and E. N. Eskandar, "A flexible software tool for temporally- precise behavioral control in matlab," Journal of Neuroscience Methods, vol. 174, no. 2, pp. 245 -- 258, 2008. [11] J. Gonzalez-Martinez and D. Lachhwani, "Stereoelectroencephalography technique and results," Child's in children with cortical dysplasia: Nervous System, vol. 30, no. 11, pp. 1853 -- 1857, 2014. [12] J. Gonzalez-Martinez, J. Mullin, J. Bulacio, A. Gupta, R. Enatsu, I. Najm, W. Bingaman, E. Wyllie, and D. Lachhwani, "Stereoelec- troencephalography in children and adolescents with difficult-to-localize refractory focal epilepsy," Neurosurgery, vol. 75, no. 3, pp. 258 -- 268, 2014. [13] J. Gonzalez-Martinez, J. Mullin, S. Vadera, J. Bulacio, G. Hughes, S. Jones, R. Enatsu, and I. Najm, "Stereotactic placement of depth electrodes in medically intractable epilepsy," Journal of Neurosurgery, vol. 120, no. 3, pp. 639 -- 644, 2014. [14] J. Gonzalez-Martinez and I. M. Najm, "Indications and selection criteria for invasive monitoring in children with cortical dysplasia," Child's Nervous System, vol. 30, no. 11, pp. 1823 -- 1829, 2014. [15] J. Bancaud, J. Talairach, S. Geier, A. Bonis, S. Trottier, and M. Man- rique, "[Behavioral manifestations induced by electric stimulation of the anterior cingulate gyrus in man]," Rev Neurol, vol. 132, pp. 705 -- 724, Oct. 1976. [16] J. Bancaud, J. Talairach, M. Lamarche, A. Bonis, and S. Trottier, "[Neurophysiopathological hypothesis on startle epilepsy in man]," Rev Neurol, vol. 131, pp. 559 -- 571, Aug. 1975. [17] J. Bancaud, J. Talairach, P. Morel, and M. Bresson, "[Ammon's horn and amygdaline nucleus: clinical and electric effects of their stimulation in man]," Revue neurologique, vol. 115, pp. 329 -- 352, Sept. 1966. [18] J. Bancaud, J. Talairach, P. Morel, M. Bresson, A. Bonis, S. Geier, E. Hemon, and P. Buser, ""Generalized" epileptic seizures elicited by electrical stimulation of the frontal lobe in man," Electroencephalogr Clin Neurophysiol, vol. 37, pp. 275 -- 282, Sept. 1974. [19] J. Bancaud, J. Talairach, P. Waltregny, M. Bresson, and P. Morel, "[Stimulation of focal cortical epilepsies by megimide in topographic di- agnosis (clinical EEG and SEEG study)]," Revue neurologique, vol. 119, pp. 320 -- 325, Sept. 1968. [20] J. Bancaud, J. Talairach, P. Waltregny, M. Bresson, and P. Morel, "Activation by Megimide in the topographic diagnosis of focal cortical epilepsies (clinical EEG and SEEG study)," Electroencephalography and clinical neurophysiology, vol. 26, p. 640, June 1969. [21] J. Talairach and G. Szikla, "[Amygdalo-hippocampal partial destruction by yttrium-90 in the treatment of certain epilepsies of rhinencephalic manifestation]," Neurochirurgie, vol. 11, no. 3, pp. 233 -- 240, 1965. [22] J. Talairach and G. Szikla, "Application of stereotactic concepts to the surgery of epilepsy," in Advances in Stereotactic and Functional Neurosurgery 4 (F. Gillingham, J. Gybels, E. Hitchcock, G. Rossi, and G. Szikla, eds.), vol. 30 of Acta Neurochirurgica Supplementum, pp. 35 -- 54, Springer Vienna, 1980. [23] J. Talairach, P. Tournoux, A. Musolino, and O. Missir, "Stereotaxic exploration in frontal epilepsy," Advances in neurology, vol. 57, pp. 651 -- 688, 1992. [24] J. Bancaud, J. Talairach, A. Bonis, C. Schaub, G. Szikla, P. Morel, and M. Bordas-Ferer, La st´er´eo-´electroenc´ephalographie dans l'´epilepsie. Informations neurophysiopathologiques apport´ees par l'investigation fonctionnelle st´er´eotaxique. Masson, 1965. [25] C. Munari, J. Bancaud, A. Bonis, P. Buser, J. Talairach, and G. Szikla, "[Role of the amygdala in the occurence of oro-alimentary signs of during epileptic seizures in man]," Revue d'´electroenc´ephalographie et de neurophysiologie clinique, vol. 9, no. 3, pp. 236 -- 240, 1979. [26] C. Munari, L. Bossi, C. Stoffels, P. Brunet, J. Bancaud, J. Talairach, and P. L. Morselli, "[Cerebral concentrations of anticonvulsants in patients with epilepsy of tumoral origin]," Revue d'´electroenc´ephalographie et de neurophysiologie clinique, vol. 12, pp. 38 -- 43, Apr 1982. [27] C. Munari, A. Giallonardo, P. Brunet, D. Broglin, and J. Bancaud, "Stereotactic investigations in frontal lobe epilepsies," in Advances in Stereotactic and Functional Neurosurgery 8 (G. Broggi, J. Burzaco, E. Hitchcock, B. Meyerson, and S. T´oth, eds.), vol. 46 of Acta Neu- rochirurgica Supplementum, pp. 9 -- 12, Springer Vienna, 1989. [28] J. Gonzalez-Martinez, J. Bulacio, A. Alexopoulos, L. Jehi, W. Bingaman, and I. Najm, "Stereoelectroencephalography in the "difficult to localize" refractory focal epilepsy: Early experience from a north american epilepsy center," Epilepsia, vol. 54, no. 2, pp. 323 -- 330, 2013. [29] F. Cardinale and G. Lo Russo, "Stereo-electroencephalography safety and effectiveness: Some more reasons in favor of epilepsy surgery," Epilepsia, vol. 54, no. 8, pp. 1505 -- 1506, 2013. [30] E. Maris and R. Oostenveld, "Nonparametric statistical testing of EEG- and MEG-data," Journal of Neuroscience Methods, vol. 164, no. 1, pp. 177 -- 190, 2007. [31] H. Bokil, P. Andrews, J. E. Kulkarni, S. Mehta, and P. P. Mitra, "Chronux: A platform for analyzing neural signals," Journal of Neu- roscience Methods, vol. 192, no. 1, pp. 146 -- 151, 2010.
1709.01116
1
1709
2017-09-04T18:55:35
Musical NeuroPicks: a consumer-grade BCI for on-demand music streaming services
[ "q-bio.NC", "cs.CY", "cs.HC", "cs.MM" ]
We investigated the possibility of using a machine-learning scheme in conjunction with commercial wearable EEG-devices for translating listener's subjective experience of music into scores that can be used in popular on-demand music streaming services. Our study resulted into two variants, differing in terms of performance and execution time, and hence, subserving distinct applications in online streaming music platforms. The first method, NeuroPicks, is extremely accurate but slower. It is based on the well-established neuroscientific concepts of brainwave frequency bands, activation asymmetry index and cross frequency coupling (CFC). The second method, NeuroPicksVQ, offers prompt predictions of lower credibility and relies on a custom-built version of vector quantization procedure that facilitates a novel parameterization of the music-modulated brainwaves. Beyond the feature engineering step, both methods exploit the inherent efficiency of extreme learning machines (ELMs) so as to translate, in a personalized fashion, the derived patterns into a listener's score. NeuroPicks method may find applications as an integral part of contemporary music recommendation systems, while NeuroPicksVQ can control the selection of music tracks. Encouraging experimental results, from a pragmatic use of the systems, are presented.
q-bio.NC
q-bio
Musical NeuroPicks : a consumer-grade BCI for on-demand music streaming services Fotis P. Kalaganis1, Dimitrios A. Adamos2,3, Nikos A. Laskaris1,3 1AIIA Lab, Department of Informatics, 2School of Music Studies 3Neuroinformatics GRoup, http://neuroinformatics.gr Aristotle University of Thessaloniki 54124 Thessaloniki, Greece E-mail: [email protected]; [email protected]; [email protected] 1 Abstract We investigated the possibility of using a machine-learning scheme in conjunction with commercial wearable EEG-devices for translating listener's subjective experience of music into scores that can be used in popular on-demand music streaming services. Our study resulted into two variants, differing in terms of performance and execution time, and hence, subserving distinct applications in online streaming music platforms. The first method, NeuroPicks, is extremely accurate but slower. It is based on the well-established neuroscientific concepts of brainwave frequency bands, activation asymmetry index and cross frequency coupling (CFC). The second method, NeuroPicksVQ, offers prompt predictions of lower credibility and relies on a custom-built version of vector quantization procedure that facilitates a novel parameterization of the music-modulated brainwaves. Beyond the feature engineering step, both methods exploit the inherent efficiency of extreme learning machines (ELMs) so as to translate, in a personalized fashion, the derived patterns into a listener's score. NeuroPicks method may find applications as an integral part of contemporary music recommendation systems, while NeuroPicksVQ can control the selection of music tracks. Encouraging experimental results, from a pragmatic use of the systems, are presented. Keywords: EEG, music evaluation, recommendation-systems, human computer interaction 2 Introduction 1 Until recently, electroencephalography (EEG) was met exclusively in hospitals and clinics, where trained experts operated expensive devices. The vast majority of the EEG-related research was dedicated to the diagnosis of epilepsy, sleep disorders, Alzheimer's disease, as well as the monitoring of certain clinical procedures such as anesthesia. By the same token, the application of Brain-Computer Interfaces (BCIs) has so far been confined to neuroprosthetics and for building communication channels for the physically impaired people [1]. Recent advances in biosensors [2] offer commercial neuroimaging headsets at affordable prices. Nowadays, the procedure of recording the electrical activity of the brain via electrodes on the human scalp has been significantly simplified and does not require a clinician's level of expertise [3]. This emerging potential stimulates growth, favors innovation and anticipates novel applications of non-invasive BCIs within real-life environments [4]. Since the beginning of the twenty-first century, the digital revolution has radically affected the music industry and is continuously reforming the business model of music economy [5]. So far, well-established channels of music distribution have been replaced and new industry stakeholders have emerged. Among them, music on-demand recommendation and streaming services emerge as the ''disruptive innovators'' [6] of the new digital music ecosystem. In a previous work [7], we have introduced our vision for the integration of bio- personalized features of musical aesthetic appreciation into modern music recommendation systems to enhance user's feedback and rating processes. A feasibility study was then carried out, based on the recording of brain activity via a 14-electrode modern commercial wireless EEG device. Therein, a single-sensor discernible EEG pattern that reflected the personalized musical aesthetic experience was identified over the prefrontal cortex and a proof of concept was provided by means of regression analysis. In the current work, we aimed for a more systematic development of a machine learning scheme that could harness the information from the brain activity recorded without sacrificing the convenience of the listener. The BCI system we opted for is presented pictorially in Fig. 1. Among the desired specifications was the seamless incorporation of a wearable EEG device within an online streaming music service. 3 We have now employed an affordable 4 dry-sensor wireless EEG device that operates with higher resolution and focuses on the prefrontal cortex. The first part of our study was devoted to the identification of the optimal brain activity descriptors that, within the range of capabilities provided by the device, would reliably and consistently reflect the listeners' evaluation about the music being played. We experimented with a set of standard signal-descriptors, conventionally employed in human electroencephalography, that could readily fit in a real-time system. They were based on single-sensor measurement and satisfied the requirements of low computational cost. A particular combination of descriptors was detected that offered very high accuracy and selected as a composite biomarker of music appreciation. Since its performance was reaching satisfactory levels -only- when applied to long temporal segments of brain activity, we resorted to an alternative subject-adaptive descriptor that could facilitate a music-evaluation BCI-system operating at a faster pace. Tailored to the task of music evaluation, appears as a powerful novel biomarker consistent with the nonstationarities and nonlinearities of brain activity. The second part of this study concerned the interface of either biomarker with an extreme learning machine and the implementation/validation of the overall music-evaluation BCI system. Two distinct scenarios appeared as covered by the proposed approach. The first one relates with the need in modern music streaming services for accurate ratings by the user that will, in turn, be exploited in order to deliver suggestions that match the user's taste. The second one, relates with the necessity for an effortless interaction of the users with the application's interface, for instance, so as to change the currently delivered song whenever this is not among their preferences. For this case the BCI-system should be able to provide a prompt estimate (i.e. a prediction of the final music evaluation score) at a tolerable error. An ELM in either case takes over the translation of brain activity descriptors into a single numbered score expressing the appraisal of music within the range 1-5. The two introduced variants of BCI-system, namely NeuroPicks and NeuroPicksVQ, can be personalized with very limited amount of training and their operation induces negligible amount of delay. 4 In our experimentations, we adopted a passive listening paradigm utilizing the Spotify on- demand music streaming service (http://spotify.com). Overall, the outcomes of this work are very encouraging for conducting experiments about music perception in real-life situations and embedding brain signal analysis within the contemporary technological universe. A preliminary version of this work, describing only the NeuroPicks system, has been presented in [8]. The need for an additional, supposedly complementary biomarker that could reveal the users' intentions, was the main motivation for this extended version. The NeuroPicks system takes advantage of well-established features that describe human cognition processes, while the NeuroPicksVQ system exploits a novel descriptor of nonlinear brain response dynamics EEG. The remaining paper is structured as follows. Section 2 serves as an introduction to EEG and its role in describing and understanding the effects of music. Section 3 outlines the essential tools that were employed during data analysis. Section 4 describes the experimental setup and the adopted methodology for analyzing EEG data. Section 5 is devoted to the presentation of results, while the last section includes a short discussion about the limitations of this study and its future perspectives. Electroencephalography and Music perception studies 2 Electroencephalography is a non-invasive neuroimaging technique that monitors the electrical activity of the brain using electrodes placed on the scalp. EEG reflects mainly the summation of excitatory and inhibitory postsynaptic potentials at the dendrites of ensembles of neurons with parallel geometric orientation. While the electrical field produced by distinct neurons is too weak to be recorded with surface EEG electrode, as neural action gets to be synchronous crosswise a huge number of neurons, the electrical fields created by individual neurons aggregate, resulting to effects measurable outside the skull [9]. The EEG brain signals, also known as brainwaves, are traditionally decomposed (by means of band-pass filtering or a suitable transform) and examined within particular frequency bands, which are denoted via Greek letters and in order of increasing central frequency are defined as follows: δ (0.5-4)Hz, θ (4-8)Hz, α (8-13)Hz, β (13-30)Hz, γ(>30)Hz. EEG 5 is widely recognized as an invaluable neuroimaging technique with high temporal resolution. Considering the dynamic nature of music, EEG appears as the ideal technique to study the interaction of music, as a continuously delivered auditory stimulus, with the brainwaves. For more than two decades neuroscientists study the relationship between listening to music and brain activity from the perspective of induced emotions [10, 11, 12]. More recently, a few studies appeared which shared the goal of uncovering patterns, lurked in brainwaves, that correspond to subjective aesthetic pleasure caused by music [13, 14]. Regarding music perception, the literature has reported a wide spectrum of changes in the ongoing brain activity. This includes a significant increase of power in β-band over posterior brain regions [15]. An increase in γ band, which was confined to subjects with musical training [16], an asymmetrical activation pattern reflecting induced emotions [17] and an increase of frontal midline θ power when contrasting pleasant with unpleasant musical sounds [18]. Regarding the task of decoding the subjective evaluation of music from the recorded brain activity, the role of higher-frequency brainwaves has been identified [19], and in particular the importance of γ-band brainwaves recorded over forebrain has been reported [20]. More recently, our group has exploited the concept of nested oscillations in the brain [7] to introduce a relevant CFC biomarker for the assessment of spontaneous aesthetic brain responses during music listening. The reported experimental results indicated that β and γ EEG oscillations recorded over the left prefrontal cortex are crucial for estimating the subjective aesthetic appreciation of a piece of music and may reflect the interconnectivity of the frontal cortex with subcortical music-rewarding dopaminergic areas. 3 Methods This section describes the methodological elements employed towards the design and realization of the introduced BCI-systems. It commences by presenting, briefly, the 6 Figure 1 Flow chart of the desirable music evaluation BCI descriptors of neural activation that were examined on their ability to convey the aesthetic appreciation during music listening. These descriptors were chosen among an extended repertoire which is currently in use in neuroscientific studies. Additionally, a novel descriptor of brain dynamics is introduced, based on the reparameterization of brainwaves according to a learned dictionary of short-term activations. The potential of each individual descriptor (and their combinations) to mediate the desirable read-out was estimated using Distance Correlation, which is described next. Finally, the selected machine learning scheme, that incorporated an important class of artificial neural network (ANNs), is described. 3.1 Neural activation profiling -The composite biomarker Brainwaves are often characterized by their prominent frequency and their (signal) energy content. Here, we adopted a quasi-instantaneous parameterization of brainwaves content, by means of Hilbert transform. The signal from each sensor x(t), was first filtered within the 7 range corresponding to the frequency-band of a brain rhythm (like δ-rhythm) and the envelope of the filtered activity Arhythm(t) was considered as representing the momentary strength of the associated oscillatory activity. Apart from the amplitude of each brain rhythm, its relative contribution was also derived by normalizing with the total signal strength (summed from all brain rhythms). In neuroscience research, activation refers to the change in EEG activity in response to a stimulus and is of great interest to investigate differences in the way the two hemispheres are activated [21]. To this end, an activation asymmetry index was formed by combining measurements of activation strength from two symmetrically i.e. AI(t)=leftArhythm(t)- rightArhythm(t). The normalized version of this index was also employed as an additional alternative descriptor. located sensors, A third descriptor was based on the CFC concept, which refers to the functional interactions between distinct brain rhythms [22]. A particular estimator was employed [23] that quantified the dependence of amplitude variations of a high-frequency brain rhythm on the instantaneous phase of a lower-frequency rhythm (a phenomenon known as phase- amplitude coupling (PAC)). This estimator operated on each sensor separately and used to investigate all the possible PAC couplings among the defined brain rhythms. It is important to notice here, that the included descriptors were selected so as to cover different neural mechanisms and share a common algorithmic framework. Their implementation -and mainly their integration- within a unifying system did not induce time delays unreasonable for the purposes of our real-time application since they are based on band-pass filtering and Hilbert Transform for deriving the instantaneous amplitude and phase. 3.2 Nonlinear Dynamics Descriptor – The new biomarker In an attempt to reveal the subjective music preference, we introduce an additional descriptor, which is based on a semi-supervised data-learning scheme that adapts to the music-modulated brain dynamics of the user. The descriptor stems from the blending of nonlinear dynamics principles with pattern analysis concepts, and experimentally proved to outperform the well-established descriptors when operates on short signal segments. The 8 overall approach builds over the idea of reconstructing brain dynamics from the recorded sensor signal, representing them as trajectories in a high-D space and provide a computable relevant parametrization, by means of data learning, that incorporates the listener's appraisal scores. This representation follows a discriminative VQ scheme, which was introduced in [24] for disentangling pathological from healthy cognitive responses. The critical modification, that is incorporated here, is the codebook editing step which is based on correlation analysis of the encoded information with the provided scores. 3.2.1 VQ-encoding of single-trial traces The first step of reconstructing dynamics, is performed by means of time-delay embedding. The pair of parameters of time-delay and embedding dimension (τ, de) is set to appropriately selected values, i.e. a small τ to preserve the temporal resolution of the signal and sufficiently large de so as to track the intrinsic dynamics. Then a given 1D signal x(t), t=1,…,T is represented as a trajectory in a phase-space by forming the consecutive vectors xi=[x(ti), x(ti+τ), x(ti+2τ), …, x(ti+deτ)], i=1,…,T-deτ. The trail of this trajectory is a set of N= T- deτ, time-indexed, de- dimensional points and equivalently represented in the tabular form of a trajectory matrix: Tr{ x(t);de,τ }=[x1 x2 … xN] , where ('''' denotes the row separator). The formed trajectory matrix encapsulates the brain response dynamics in ''raw'' format. A more refined signature can be derived by means of vector-quantization (VQ). Given a set of reference points in phase- space, a partition -known as Voronoi tessellation- is induced that enables the description of the trajectory-trail in the form of a distribution pattern. Fig.2 exemplifies these steps based on 3 different single-trial signals recorded from a subject during music listening. The original signals, after band-pass filtering (Fig.2a), have been represented as 2D trajectories in Fig.2b. In the same figure, 10 reference points are also depicted together with the corresponding Voronoi diagram. The shown Voronoi regions can be thought of as 2D bins and have been used in the re-parameterization of each trajectory in the form of a histogram (see Fig.2c). Provided a codebook (i.e. a set of reference vectors, also known as codevectors) in involves the reconstructed phase-space, the algorithmic procedure of VQ-encoding, application of the 'nearest-neighbor rule' to each row-vector in the trajectory matrix so as to 9 assign it to the most similar codevector. Formally, the codebook {cj}j=1,…,k divides the phase- space into k Voronoi regions, ViÌ ℝ#$ , i=1,…,k 𝑉&= 𝐱 Î ℝ#$∶ ǁ 𝐱−𝐜𝐢ǁ. ≤ ǁ 𝐱−𝐜𝐣 ǁ. , " j,j=1,…,i−1,i+1,..,𝑘 and the single-trial encoding step transforms a reconstructed trajectory to a distribution pattern over these regions. It is prominent, that the design of codebook is a critical part towards the definition of a suitable descriptor that would emphasize the characteristics of brain dynamics that reflect the level of music appraisal. 3.2.2 Codebook Design and selection of CodeWaves The codebook design is based on a data-learning algorithmic scheme and presumes the availability of a certain amount of training data, i.e. subject-specific EEG recordings during music listening along with the subjective evaluation score. The procedure begins with the execution of Neural Gas algorithm for the initial design of codebook and proceeds with the selection of the most informative codevectors. The first stage, that is the derivation of prototypical brainwave patterns, runs in unsupervised mode and captures the underlying Figure 2 A nonlinear dynamics approach to analyzing brain signal traces: a) music listening single-trial responses recorded at FP1 sensor filtered within γlow frequency band, b) embedding of the corresponding dynamics as (approximate) trajectories in a 2D space, with the Voronoi regions defined based on a given set of 10 reference points (code-vectors), denoted with red disks. c) Describing the trajectories as histograms over the given codebook (Voronoi regions are treated as bins). 10 dynamic manifold by demarcating (a user defined number of) k Voronoi regions. The second stage, that is the selection of regions reflecting better the music appraisal, is based on correlating the ''activation'' of each Voronoi region with the subjective score. The overall scheme (see Fig.3a) is of a semi-supervised nature and enhances further the personalization of the proposed BCI-system. Given a training set of signals, s1(t), s2(t), … , sM(t) , we first form the corresponding trajectory matrices Tr{ s1(t);de,τ }, Tr{ s2(t);de,τ },…, Tr{ sM(t);de,τ }, and then create an overall data-matrix, by stacking them : Xdata=[ Tr{ s1(t);de,τ } Tr{ s2(t);de,τ }… Tr{ sM(t);de,τ }]. The [M·N × de] data matrix is then fed to Neural Gas, a self-organized network that efficiently converges to a set of k<<M·N codebook vectors using a stochastic gradient descent procedure with a ''soft-max'' adaptation rule that minimizes the average distortion error. The derived codevectors cj Î ℝ#$, j=1,2,...k constitute the initial set of reference brainwave patterns for the encoding of single trial brain dynamics. In principle, we use a relative high value for the codebook size, e.g. k=100 or 200. Such a value is needed to ensure that the reconstructed space will be partitioned at a sufficient detail. This in turn will enable the Figure 3 a) Designing and editing the codebook based on training data. b) The steps of VQ encoding of a single-trial response. 1-2: the band-pass filtered trace undergoes time-delayed embedding; 3: Phase-space is divided into regions via Voronoi tessellation and all vectors are assigned to their nearest prototype. 4: The encoded trajectory is represented in histogram format. 5: The 'projections' within the most informative regions are kept for the Biomarker. 11 accurate representation of the dynamics after the encoding and will provide sufficient flexibility in the quest for informative Voronoi regions. Next, the VQ-encoding is applied to the training set of signals. In this way, a signal si(t) is transformed to a distribution pattern hi=[hi(1), hi(2)…,hi(k)], with the rth-entry reflecting the relative activation of the rth Voronoi region during listening to the musical piece associated with the ith signal. The associated labels scorei ,i=1,2…,M expressing the subjective evaluation of the listener for each song, are utilized in a feature selection procedure that follows the principles of dynamic programming and exploits the generic character of the Distance Correlation measure (described in section 3.2). After the initial ordering of the individual codevectors based on their dependence with the level of music appreciation, i.e. based on the quantity R(hi(r), scorei), an incremental procedure is followed. It starts with the codevector r[1], the measurements hi(r[1]) of which show the highest nonlinear dependence. At each iteration, the ranked list of codevectors is traversed systematically and the current list of selected codevectors is augmented with the next in order codevector only if the overall measure of dependence increases. The procedure terminates when no further increase is Figure 4 Selecting CodeWaves: a) a 2D embedding of the initial codebook which consists of 100 code-vectors coloured by the distance correlation (with the subjective score of music appraisal) of each one separately, b) the spatial representation of the codebook editing step with the selected codewaves ranked by their distance correlation value (legend refers to the score of the hierarhically formed list) c) the seleted codewaves. 12 achieved and returns a selected list of codevectors {r[1], r[2],… r[sel]}. The selected codevectors, named hereafter as codewaves, constitute the basis for the new biomarker. Fig.4 exemplifies the step of selecting the codewaves by means of a semantic map representation [25], (a 2D scatter plot derived based on multidimensional scaling) for signals recorded at FP1 and band- pass filtering within γ band. Wrapping up section 3.2, the nonlinear dynamics biomarker (after the Codebook design and editing) is implemented with the steps depicted graphically in Fig.3b. Distance Correlation 3.3 In statistics and in probability theory, Distance Correlation is a measure of statistical dependence between two random variables or two random vectors of arbitrary, not necessarily equal, dimension. Distance Correlation, denoted by R, generalizes the idea of correlation and holds the important property that R(x,y)=0 if and only if x and y are independent. R index satisfies 0≤ R ≤1 and, contrary to Pearson's correlation coefficient, is suitable for revealing non-linear relationships [26]. In this work, it was the core mechanism for identifying neural correlates of subjective music evaluation, by detecting associations between the results of signal descriptors (or combinations of them) and the listener's scores. Extreme Learning Machines 3.4 Machine learning deals with the development and implementation of algorithms that can build models able to generalize a specific function from a set of given examples. Regression is a supervised learning task that machine learning can handle with efficiency of similar, or even higher, level than the standard statistical techniques, and -mainly- without imposing hypotheses. In this work, the decoding of subjective music evaluation was cast as a (nonlinear) regression problem. A model was then sought (i.e. learned from the experimental data) that would perform the mapping of patterns derived from the brain activity descriptors to the subjective music evaluation. ELMs appeared as a suitable choice due to their documented ability to handle efficiently difficult tasks without demanding extensive training sessions [27]. They are 13 feedforward ANNs with a single layer of hidden nodes, where the weights connecting inputs to hidden nodes are randomly assigned and never updated [28]. Implementation and Experiments 4 Our experimentations evolved in two different directions. First, a set of experiments was run so as to use the experimental data for establishing the brainwave pattern(s) that would best reflect the music appraisal of an individual and train the learning machine from music pieces of known subjective evaluation. Next, additional experiments were run that implemented the real-time scoring by the trained ELM-machine so as to justify the proposed BCIs in a more naturalistic setting. In both cases, the musical pieces were delivered through a popular music on-demand streaming service (Spotify) that facilitates the registration of the listeners' feedback ('like'-'dislike') to adapt the musical content to their taste and make suggestions about new titles. Participants 4.1 All 5 participants were healthy students of AUTH Informatics department. Their average age was 23 years and music listening was among their daily habits. They signed an informed consent after the experimental procedures had been explained to them. device (i.e. device Data Acquisition dry-sensor wireless 4.2 Having in mind the user-friendliness of the proposed scheme, we adopted a modern commercial - http://www.choosemuse.com) in our implementations. This "gadget" offers a 4-channel EEG signal, with a topological arrangement that can be seen in Fig.5. The signals are digitized at the sampling frequency of 220 Hz. Data are transmitted under open sound control (OSC) protocol, which is based on UDP/IP to facilitate communication and flexible interoperability among computers, sound synthesizers and other multimedia devices optimized for modern networking technology [29]. Interaxon's Muse 14 Figure 5 a) Topological arrangement of available Electrodes b) Interaxon's Muse headset. Experimental Procedure 4.3 Prior to placing the headset, subjects sat in a comfortable armchair and volume of speakers was set to a desirable level. They were advised to refrain from body and head movements and enjoy the music experience. Recording was divided into sessions of 30 minutes duration. The music streaming service was operated in radio mode, hence randomly selected songs (from the genre of their preference) were delivered to each participant while his/her brain activity was registered. Among the songs there were advertisements. That part of recordings was isolated from the rest. The recording procedure was integrated, in MATLAB, together with all necessary information from the streaming audio signal (i.e. song id, time stamps for the beginning and termination of each song). Participants evaluated each of the listened song, using as score one of the integers {1,2,3,4,5}, during a separate session just after the end of the recording. These scores, together with the associate patterns extracted from the recorded brain signals, comprised our training set. The overall procedure was repeated on a different day. On average, the recorded activity was corresponding to 30 songs per participant and approximately 40% were unknown to the listeners. In addition to the above data, two of the subjects participated in an extra recording session, during which the trained BCI-systems operated in real-time and their predictions were compared with the scores provided by the listeners just after the end of the recording. 15 4.4 Data Analysis The preprocessing of signal included 50Hz component removal (by a built-in notch filter in MuseIO - the software that connects to and streams data from Muse), DC offset removal, and removal of the signal segments that corresponded to the 5 first second of each song (in order to avoid artificial transients). The digital processing included band-pass filtering for deriving the brainwaves of standard brain rhythms and computation of the descriptors described in sections 3.1 and 3.2. To increase frequency resolution, we divided the β rhythm into βlow (13-20 Hz), and βhigh (20- 30 Hz) sub-bands and derived descriptors separately. Similarly, the γ rhythm was divided into γlow (30-49Hz) and γhigh (51-90Hz). As a means to increase the volume of available data, multiple segments were extracted from the bandpass-filtered brain activity corresponding to listening to a single song listening. These extracts had been selected randomly, with their length parametrized during the off-line experimentation between 10 and 100 seconds. This 'bootstrapping' step was compatible with the possibility that, in radio-listening mode, the listening may start in the middle of a song. The two types of descriptors (presented in 3.1 and 3.2 respectively), were applied to the available single-trial segments, in order to identify the optimal combination (or subset in the latter case) of features to be included in the proposed BCI-systems. In the case of NeuroPicks, the feature selection step was performed collectively for all participants, resulting in a single composite Biomarker. On the contrary, such a procedure was not compatible with the nature of NeuroPicksVQ that relies on the adaptive design of codewaves. Hence, the nonlinear dynamics biomarker was defined on a subject-by-subject basis. The derived biomarkers were utilized as ''feature-extractors'' and a training set of patterns was gathered by estimating the biomarkers from additional (independently sampled) extracts of brain activity. These patterns were employed for tailoring an ELM model to each participant. This was considered the "overall training" phase for the supervised component(s) of the BCI systems. In an ''in-situ'' testing phase (section 5.3), the subject- specific ELM model was applied to streaming data (from the biomarkers operating on line) in order to predict the listener's evaluation. 16 5 Results 5.1 NeuroPicks : the composite biomarker + ELM The biomarker in this BCI-system was designed based on a feature-engineering procedure that was meant to be universally valid, i.e. to result to a biomarker that would be applicable to any user, without the need for time-consuming adaptations. Towards this end, the ensemble of descriptors described in 3.1 was derived regarding all defined brain rhythms (7), available sensors (4), both pairs of symmetrically placed sensors (4), and legitimate pairs of cross-frequency interactions rhythmlow® rhythmhigh (21). All these measurements, from the single-trial segments with the highest length, were gathered in a tall matrix with #rows=(30songs x 10extracts x 5subjects) and #columns=140. The statistical dependence of each one dimension (column) with the associated scores of subjective music evaluation was estimated based on the Distance Correlation measure and a bootstrapping scheme during which 100 random samples (from the rows) were created and the individual estimates were averaged. Fig.6, includes the obtained correlation measurements in the form of three distinct tables, split according to the type of each information dimension. The 140 features were ranked in descending order according to R-score (i.e. regarding their music evaluation expressiveness) and a dynamic programming methodology was then applied. Starting with the feature of highest R, we traversed systematically the ranked list for the combination that would eventually maximize the Distance Correlation. This procedure led to a set of 5 features (integrated over time) that constituted the synthesized music appraisal biomarker. This biomarker included the normalized temporal asymmetry index in βlow band, relative energy of α band at temporal electrodes, γlow®γhigh PAC at FP1 sensor and relative energy of θ band at TP9. The relevance of the designed biomarker to music evaluation showed a dependence on the length of segment based on which it had been evaluated. Figure 7 indicates this trend (as averaged profile across participants). It can be observed that while the performance of the biomarker is poor when fed with short EEG segments of music listening, its effectiveness constantly increases with the duration of the listened music. This trend provided an initial indication about how fast the automated music evaluation system NeuroPicks could operate. 17 Figure 6 Distance Correlation score of all signal descriptors, derived based on 90-sec long segments of brain activity during music listening. During the offline experimentation, the available data (the biomarker patterns of each of the five participants along with the associated ratings) were randomly partitioned in training and testing set, in a 60%-40% proportion for a Monte-Carlo cross validation. An ELM was trained using the former set and its performance was quantified using the latter set. The 18 Figure 7 Distance Correlation of the composite biomarker with the score of subjective music appreciation as a function of the segment length. number of neurons in the hidden layer, the only parameter to be tuned in ELMs, was selected as the smallest number of neurons such that the training and testing error were converging at an acceptable level, lower than 0.01. This number was found ranging from 12 to 18. The overall procedure was repeated 10 times, and averaged results were reported. The normalized root mean squared error (nRMSE), as defined for regression tasks, was found to be 0.063±0.0093 (mean±std across participants). For comparison purposes, we also employed Support Vector Machines (SVMs), which performed slightly inferiorly. Although the difference was marginal, the very short training time was another factor in favor of employing ELMs. 5.2 NeuroPicksVQ : the adaptive biomarker + ELM This music evaluation BCI system stemmed from the need to provide an indication about the listener's appraisal within few seconds, i.e. without postponing estimation till the end of the song, which is the ideal scenario for the previous system. It operates on short segments of brain activity and owns its efficiency to the adaptive character of the employed descriptor. 19 The most important operational difference (with respect to NeuroPicks), is that before the realization of NeuroPicksVQ, training is required for the biomarker as well. Data learning is involved for optimizing the representation of brainwaves in a subject-specific manner. As a means to confine the search space, we examined the suitability of each sensor and brain rhythm for leading to a VQ-based biomarker that reflects well the subjective music evaluation, and parametrized this search over the length of the extracted segments (aiming for the shortest segments that could support acceptable performance). In every case, for the initial step of forming trajectories via time-delay embedding, the following strategy was followed for defining the two embedding parameters (deviating from the standard practice in the literature of nonlinear dynamics [30]). Having in mind to provide a detailed description, we always kept time delay τ=1 and leave the embedding dimension de as an additional parameter to be optimized across subjects. For the step of initial codebook design, k was set to 100 (and experimentally verified that it was not a critical parameter as long as it remained above 70). The codebook-editing step usually led to the selection of 8 (range: (6-12)) codewaves depending on the situation (subject, sensor, rhythm and initialization of Neural-Gas algorithm). The statistical dependence of the designed biomarkers with the associated scores of subjective music evaluation was first estimated for each subject independently, based on the Distance Correlation measure and a bootstrapping scheme, and the individual results were then averaged across subjects. Fig.8, represents the suitability of each sensor and brain rhythm to provide sufficient information based on 20-sec segments and when the embedding Figure 8 Distance Correlation of the VQ-based biomarker (with the subjective music appreciation) for all electrodes and brain rhythms based on 20sec-long segments of brain activity signals during music listening. The shown measurements are the averaged across subjects Distance Correlation values. 20 dimension was de =50. It can be observed that it was the activity in γlow from FP1 electrode that would lead to the most suitable biomarker. The corresponding correlation level with the subjective score was reach the value of 0.72. It should be mentioned here that the use of longer segments did lead to higher R-values, but the increase of R-values with time interval was building very slow and never exceed the value of 0.78. Considering the above parameters as the prescribed values (i.e. sensor:FP1, rhythm: γlow, segment:20sec, de=50), and working independently for each subject, we derived the VQ- biomarker for all (approximately 300=30 songs × 10 extracts ) available segments. The included codewaves had been defined in a personalized fashion and even their number was differing among subjects. The VQ-biomarker patterns were randomly partitioned in training and testing set, in a 60%-40% proportion for a Monte-Carlo cross validation. An ELM was trained using the former set and its performance was quantified using the latter set. The number of neurons in the hidden layer, was selected as described in 5.1. The overall procedure was repeated 10 times, and the normalized RMSE, averaged across repetitions, provided an indication of performance. By averaging across participants, the error for the NeuroPicksVQ system was estimated at 0.115 ±0.046. This is a satisfactory figure of performance considering the 20sec-interval needed. 5.3 Online evaluation of the BCI - systems Additional experiments were performed in order to test the performance of the two systems in a realistic setting. Two of our subjects participated in additional recording sessions during which the (already custom-made) NeuroPicks systems were providing a read-out of the subjective music evaluation. In the case of NeuroPicks /NeuroPicksVQ , 90/20 seconds long segments were used. The main purpose of this setup was to truly evaluate the performance of the proposed BCIs on new, previously unseen data. Based on three recording sessions lasting for 90 minutes in total, 18 and 21 songs were delivered for the first and second subject, through the Spotify platform. The automated evaluation scores (predictions of subjective appraisal) were registered and compared with the listeners' ratings provided at the end of each recording session. The normalized RMSE for the NeuroPicks system was estimated at 0.09 and 0.07 for the two individuals. These estimates were based on the score predicted by 21 Figure 9 Time evolving Predictions of NeuroPicksVQ BCI music evaluation system acting on recorded brain activity during listening to a song with high subjective rating. The corresponding audio signal is shown beneath, segmented into verse, pre-chorus and chorus. the system using the activity during the initial 90 seconds of each song. The same performance index was estimated at 0.19 and 0.16 for the NeuroPicksVQ system. These estimates were based on a single score (mean value) produced by integrating along time the time-course of predictions produced by the system (see Fig.9). Discussion 6 The paper reports our attempts to associate the listener's brainwaves with the subjective aesthetic pleasure induced by music. It also serves as a proof-of-concept study on the feasibility of interfacing a modern consumer EEG device with a popular on-demand music streaming service (Spotify). Our results indicate that descriptors of the signals recorded from a restricted number of sensors (located over frontal and temporal brain areas) can be combined in computable biomarkers operating on long or short temporal segments and reflecting the listener's subjective music evaluation. The derived biomarkers' patterns can be efficiently decoded by regression-ELM leading to reliable readouts directly from the listener. The main advantage of the approach is that it complies with idea of employing EEG-wearables in daily activities and is readily embedded within the contemporary on-demand music streaming services. Nevertheless, the problem of artifacts (noisy signals of biological origin) has not been addressed yet. For the presented results, the participants had been asked to limit body/head movements and facial expressions and as much as possible. Hence, before employing such a system to naturalistic recordings, methodologies for real-time artifact suppression (as in [31]) have to be incorporated. 22 Today, we live in the world of Internet of Things (IoT) where the interconnectedness among devices has already been anticipated and supportive technologies are now being realized [32]. However, to achieve a seamless integration of technology in people's lives, there is still much room for improvement. In a world beyond IoT, technology would proactively facilitate people's expectations and wearable devices would transparently interface with systems and services. To enable such scenarios would require to move from the IoT to the Internet of People (IoP) [33]. People would then participate as first-class citizens relishing the benefits of holistic human-friendly applications. As such, we present a pragmatic use of a consumer BCI that ideally fits in the anticipated forms of the digital music universe and demonstrate a first series of encouraging experimental results. Acknowledgements This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. 23 References 1. Niedermeyer, E., da Silva, F. L.: Electroencephalography: basic principles, clinical applications, and related fields. Lippincott Williams & Wilkins (2005) 2. Liao, L.D., Lin, C.T., McDowell, K., Wickenden, A.E., Gramann, K., Jung, T.P., Ko, L.W. and Chang, J.Y.: Biosensor technologies for augmented brain–computer interfaces in the next decades. Proceedings of the IEEE, 100(Special Centennial Issue), pp.1553-1566, (2012) 3. Casson,, A.J., Yates, D., Smith, S., Duncan, J.S., Rodriguez-Villegas, E.: Wearable Electroencephalography, IEEE Eng. Med. Biol. Mag. 29 44–56 (2010) 4. Lance, B. J., Kerick, S. E., Ries, A. J., Oie, K. S., & McDowell, K.: Brain–computer interface technologies in the coming decades. Proceedings of the IEEE, 100(Special Centennial Issue), 1585-1599 (2012) 5. Wikström, P., DeFillippi, R. (Eds.): Business Innovation and Disruption in the Music Industry. Edward Elgar Publishing (2016) 6. Downes, L., Nunes, P.: Big Bang Disruption. Harvard Business Review, 44-56 (2013) 7. Adamos, A.D., Dimitriadis, I.S., Laskaris, A.N.: Towards the bio-personalization of music recommendation systems: a single-sensor EEG biomarker of subjective music preference. Information Sciences, vol. 343-344, 94-108 (2016) 8. Kalaganis, F., Adamos, A.D., Laskaris, N.: A consumer BCI for Automated Music Evaluation within a popular on-demand Music Streaming Service:''Taking Listener's Brainwaves to Extremes'', 12th IFIP WG 12.5 International Conference and Workshops, AIAI 2016, (2016) 9. Cohen, M.X.: Analyzing Neural Time Series Data: Theory and Practice, MIT Press (2014) 10. Altenmüller, E.: Cortical DC-potentials as electrophysiological correlates of hemispheric dominance of higher cognitive functions. Int. J. Neurosci. 47 1-14 (1989) 11. Petsche, H., Ritcher, P., von Stein A., Etlinger, S.C., Filz, O.: EEG Coherence and Musical Thinking. Music Perception: An Interdisciplinary Journal 11 117-151 (1993) 12. Birbaumer, N., Lutzenberger, W., Rau, H., Braun, C., Mayer-Kress, G.: Perception of music and dimensional complexity of brain activity. International Journal of Bifurcation and Chaos 6, 267 (1996) 13. Hadjidimitriou, S.K., Hadjileontiadis, L.J.: Toward an EEG-Based Recognition of Music Liking Using Time-Frequency Analysis, IEEE Trans. Biomed. Eng. 59, 3498-3510 (2013) 14. Schmidt, B., Hanslmayr, S.: Resting frontal EEG alpha-asymmetry predicts the evaluation of affective musical stimuli, Neurosci. Lett. 460, 237–240 (2009) 15. Nakamura, S., Sadato, N., Oohashi, T., Nishina, E., Fuwamoto, Y., Yonekura, Y.: Analysis of music-brain interaction with simultaneous measurement of regional cerebral blood flow and electroencephalogram beta rhythm in human subjects, Neurosci Lett., 275(3), 222-226 (1999) 16. Bhattacharya, J., Petsche, H.: Musicians and the gamma band: a secret affair? Neuroreport 12(2), 371-374, (2001) 24 17. Schmidt, A.L., Trainor, L.J.: Frontal brain electrical activity (EEG) distinguishes valence and intensity of musical emotions. Cognition and Emotion, 15(4), 487-500 (2001) 18. Sammler, D., Grigutsch, M., Fritz, T., Koelsch, S.: Music and emotion: electrophysiological correlates of the processing of pleasant and unpleasant music. Psychophysiology, 44(2), 293-304 (2007) 19. Hadjidimitriou, S.K., Hadjileontiadis, L.J.: EEG-Based Classification of Music Appraisal Responses Using Time-Frequency Analysis and Familiarity Ratings, IEEE Trans. Affect. Comput. 4, 161–172 (2013) 20. Pan, Y., Guan, C., Yu, J., Ang, K.K., Chan, T.E.: Common frequency pattern for music preference identification using frontal EEG. 6th International IEEE/EMBS Conference on Neural Engineering, 505-508 (2013) 21. Coan, A.J., Allen J.B.J.: Frontal EEG asymmetry as a moderator and mediator of emotion. Biological Psychology, 67(1-2), 7-50 (2004) 22. Canolty, T.R., Knight, T.R.: The functional role of cross-frequency coupling. Trends in Cognitive Sciences, 14(11) 506-515 (2010) 23. Dimitriadis, I.S., Laskaris, A.N., Bitzidou, P.M., Tarnaras, I., Tsolaki, N.M.: A novel biomarker of amnestic MCI based on dynamic cross-frequency coupling patterns during cognitive brain responses. Front. Neurosci. (2015) 24. Laskaris, N.A., Tarnanas, I., Tsolaki, M.N., Vlaikidis, N., Karlovasitou, A.K.: Improved detection of amnestic MCI by means of discriminative vector quantization of single- trial cognitive ERP responses. Journal of Neuroscience Method, 212(2), 344-354 (2013) 25. Laskaris, N.A., Ioannides, A.A.: Semantic geodesic maps: a unifying geometrical approach for studying the structure and dynamics of single trial evoked responses. Clinical Neurophysiology, 113(8), 1209-1226 (2002) 26. Szekely, J.G., Rizzo, L.M., Bakirov, K.N.: Measuring and testing dependence by correlation of distances. The Annals of Statistics , 35(6), 2769-2794 (2007) 27. Huang, G.B., Zhu, Q.Y., Siew, C.K.: Extreme learning machine: Theory and applications. Neurocomputing, 70(1-3), 489-501 (2006) 28. Huang, G.B.: What are Extreme Learning Machines? Filling the Gap between Frank Rosenblatt's Dream and John von Neumann's Puzzle. Cogn. Comput., 7, 263-278 (2015) 29. Wright, M., Freed, A., Momeni, A.: Open Sound Control: state of the art 2003. NIME '03: Proceedings of the 3rd conference on New interfaces for Musical Expression, (2003) 30. Abarbanel, H.D.I.: Analysis of Observed Chaotic Data, Springer-Verlag New York (1996) 31. Akhtar, M.T., Jung, T.P., Makeig, S., Cauwenberghs, G.: Recursive independent component analysis for online blind source separation. IEEE Inter. Symp. on Circuits and Systems, 6, 2813 – 2816, (2012) 32. Want, R., Schilit, B. N., Jenson, S.: Enabling the internet of things. Computer, (1), 28-35 (2015). 25 33. Miranda, J., Makitalo, N., Garcia-Alonso, J., Berrocal, J., Mikkonen, T., Canal, C., Murillo, J. M.: From the Internet of Things to the Internet of People. Internet Computing, IEEE, 19(2), 40-47 (2016) 26 27
1311.6864
1
1311
2013-11-27T03:59:13
Bayesian spike inference from calcium imaging data
[ "q-bio.NC", "q-bio.QM", "stat.AP" ]
We present efficient Bayesian methods for extracting neuronal spiking information from calcium imaging data. The goal of our methods is to sample from the posterior distribution of spike trains and model parameters (baseline concentration, spike amplitude etc) given noisy calcium imaging data. We present discrete time algorithms where we sample the existence of a spike at each time bin using Gibbs methods, as well as continuous time algorithms where we sample over the number of spikes and their locations at an arbitrary resolution using Metropolis-Hastings methods for point processes. We provide Rao-Blackwellized extensions that (i) marginalize over several model parameters and (ii) provide smooth estimates of the marginal spike posterior distribution in continuous time. Our methods serve as complements to standard point estimates and allow for quantification of uncertainty in estimating the underlying spike train and model parameters.
q-bio.NC
q-bio
Bayesian spike inference from calcium imaging data Eftychios A. Pnevmatikakis, Josh Merel, Ari Pakman and Liam Paninski Department of Statistics Center for Theoretical Neuroscience Grossman Center for the Statistics of Mind Columbia University, New York, NY 3 1 0 2 v o N 7 2 ] C N . o i b - q [ 1 v 4 6 8 6 . 1 1 3 1 : v i X r a Abstract —We present ef ficient Bayesian methods for extracting neuronal spiking information from calcium imaging data. The goal of our methods is to sample from the posterior distribution of spike trains and model parameters (baseline concentration, spike amplitude etc) given noisy calcium imaging data. We present discrete time algorithms where we sample the existence of a spike at each time bin using Gibbs methods, as well as continuous time algorithms where we sample over the number of spikes and their locations at an arbitrary resolution using Metropolis-Hastings methods for point processes. We provide Rao-Blackwellized extensions that (i) marginalize over several model parameters and (ii) provide smooth estimates of the marginal spike posterior distribution in continuous time. Our methods serve as complements to standard point estimates and allow for quanti fication of uncertainty in estimating th e underlying spike train and model parameters. I . INTRODUCT ION Calcium imaging is an increasingly popular technique for large scale data acquisition in neuroscience [1]. The method detects underlying, single neuron activity indirectly through observations of fluorescent indicators for calcium concent ra- tion. A key problem in the analysis of calcium imaging data is the inference of exact spike times from the noisy calcium signal which has slower dynamics compared to neural spiking and is sampled at a relatively low acquisition rate. A variety of methods have been proposed to deal with this problem including particle filtering [2], fast nonnegative deconvo lution for approximate maximum-a-posteriori (MAP) inference [3], greedy template matching [4], and methods for estimating signals with finite rate of innovation [5]. In these methods, parameter estimation is typically performed offline or in an iterative manner using for example the expectation maximiza- tion algorithm. In this paper we propose Bayesian methods for sampling from the joint posterior distribution of the spike train and the model parameters given the fluorescence observations. W e present two efficient approaches for sampling the spikes. Th e first is a discrete time binary sampler that samples whether a spike occurred at each timebin using Gibbs sampling. By exploiting the weak interaction between spikes at distant timebins, we show that a full sample can be obtained with just O(T ) complexity, where T is the number of timebins, and that parallelization is also possible. Our second sampler operates in continuous time and samples the number of spikes and the spike times at arbitrary resolution using Metropolis-Hastings (MH) techniques. We use a proposal distribution to move the spike times around a local neighborhood that is based on the resulting signal residual. This proposal distribution enables fast mixing and tractable inference; each full sample is obtained with just O(K ) complexity where K is the total number of spikes, rendering this algorithm particularly efficient for recordings that are sparse and/or are obtained at a fine resol u- tion. Moreover, in high-SNR conditions, this method enables super-resolution spike inference (i.e. determining where each spike occurred within each timebin) and smooth estimates of the marginal spike posterior using a Rao-Blackwellized scheme. We also show that is possible to marginalize over several of the model parameters and derive collapsed Gibbs samplers that exhibit faster mixing. I I . MODE L DE SCR I P T ION AND BLOCK G IBB S SAM P L ING We assume that we observe a single neuron calcium trace for a duration of T timesteps. Let s ∈ {0, 1}T the binary spiking vector of the neuron that indicates the existence of the spike at each timebin. The calcium activity that is generated by s can be described by a simple first order autoregressive process as c(t) = γ c(t − 1) + As(t), where γ is a discrete time constant with 0 < γ < 1, A indicates the amplitude of each spike, and c(1) = c1 + As(1), with c1 an initial condition for the calcium concentration. Our fluorescence observation vector y , can be written as y (t) = c(t) + b + εt , where b is the baseline concentration and εt ∼ N (0, σ2 ) is some random Gaussian noise. Our goal is to estimate the spiking vector s given the observation vector y , which we assume is normalized in the interval [0, 1]. Approximate MAP methods [3] have been shown to perform well under high SNR assumptions. However in the low SNR regime their performance degrades, and the parameter estimation becomes more challenging. To overcome these limitations we introduce a block-Gibbs sampler that produces samples from the joint posterior distribution of the spikes and model parameters given y . Let G ∈ RT ×T and v ∈ RT de fined respectively as         . . . 0 . . . 1 . . . . . . . . . −γ 1 γ ... γ T −1 1 −γ ... 0 0 0 ... 1 G = , v = . sT G−T G−1 s + p(sπ)p(π α, β )dπ = By denoting θ = [A, b, c1 ]T , the likelihood can be written as A2 sT G−T y(cid:19) , p(y s, θ , σ2 ) ∝ exp (cid:18)− A σ2 2σ2 with y = y − b1T − c1v (1T denotes a vector of ones of length T). We here place an i.i.d. Bernoulli process prior on the spike trains so the probability of a spike in any timebin is π . Under this uniform spiking assumption the discrete time constant γ can be estimated robustly from the autocovariance function of y . For the prior probability π we set a hyper-prior π ∼ Beta(α, β ). At each iteration we update the parameters α, β using empirical Bayes [6]: For a spiking vector s the evidence function [7] can be written as (α/β )1T p(sα, β ) = Z 1 s T (1 + α/β )T . 0 The evidence function is constant for a fixed ratio r = α/β and T s). To find distinct values is maximized for r = 1T T s/(T − 1T for α and β we place a flat hyperprior on β , which yields an exponential posterior β π , r ∼ Exp(− log(π)r − log(1 − π)). For the rest of parameters θ we assume a joint half-normal (nonnegative) distribution, θ ∼ N (µ, Σ)1{θ∈R3 +} . The param- eters µ and Σ can also be learned from the data, but we selected general values that assume little prior knowledge and model a wide prior. Other prior choices, e.g. exponential, yield practically the same results. Finally, for the noise variance σ2 we set an inverse Gamma prior σ2 ∼ InvGamma(1, 0.1)1 , which is a weak and relatively flat prior for both high and low-SNR regions for traces y normalized to the [0, 1] interval. Under these assumptions, the block Gibbs sampler proceeds as follows to draw samples from the joint posterior: σ2 G−T y + l1T (cid:19) sT W s + sT (cid:18) A 1 2 r = 1T T s/(T − 1T T s) β π , r ∼ Exp(− log(π)r − log(1 − π)) α = rβ π s ∼ Beta(1T s + α, T − 1T s + β ) θ s, σ2 , y ∼ N (Λ(σ−2 S T y + Σ−1µ), Λ)1{θ∈R3 +} σ2 s, θ , y ∼ InvGamma(1 + T /2, 0.1 + ky − S θk2/2), log p(sπ , θ ,σ2 , y) ∝ − with W = A2 (G−T G−1 )/σ2 , l = log(π/(1 − π)), S is a T × 3 matrix that depends on the current state of s, given by S = (cid:2)G−1s, 1T , v(cid:3) , and Λ = (Σ−1 + σ−2S T S )−1 . We now turn to the main problem of sampling from the posterior of the spike vector s. I I I . D I SCRE T E B INARY SAM P L ER To sample s we take a Gibbs-sampling approach where at each iteration we sample s(i) given the current state of the rest of the entries. We use a Metropolized Gibbs (MG) sampler, 1 The variance σ2 can also be estimated from the autocovariance function via the Yulle-Walker equations but we favor a Bayesian approach here. (1) with the flip of each binary entry being the possible move [8]. If s(k) is the state of s(i) at the k -th sample, and de fine i 1:i−1 , s(k−1) scur = [s(k) ]T the current state of s. The sampling i:T algorithm flips s(k) i with probability P(scur ) (cid:19) P flip (i) = min (cid:18)1, P(sflipped) The ratio inside (1) can be more efficiently computed in the log-domain. By denoting α = A σ2 G−T y + l1T the log-ratio becomes )((2s(k−1) (1 − 2s(k−1) − 1)[W ]ii/2 − sT curW ei + a(i)), (2) i i with ei the i-th standard basis vector. It is important to note that the log-ratio of (2) can be computed efficiently in just O(1) time. The matrix G−1 is lower triangular and Toeplitz with [G−1 ]ij = γ i−j 1i≥j . Therefore it can be approximated by a banded Toeplitz matrix with bandwidth that depends on γ . Consequently W can again be approximated by a banded matrix and is also asymptotically Toeplitz for T → ∞. It curW ei can be computed in just follows that the products sT O(1) time (technically in O(m/ log(1/γ )) for m bits of accuracy). This gives a total O(T ) complexity per full sweep to obtain a new sample from p(sy , p, θ , σ2 ). For large T , the algorithm can also be parallelized – notice that the entries of s which are at least L = O(1/ log(1/γ )) timesteps apart, are approximately conditionally independent. s can be partitioned into chunks of length L, that can be sampled in parallel. It is obvious that instead of using the MG sampler presented here, one can also use a plain Gibbs sampler with similar O(T ) complexity. In practice we observed that the MG sampler led to faster mixing. IV. CONT INUOU S T IME SAM P L ER The spike samplers presented in section III operate on a discrete domain, with the length of each timebin set by the frame rate of the calcium imaging. While simple and effective, these methods can have several disadvantages. First, when the calcium signal is acquired by raster scanning, the frame rate is typically in the range of 10-30Hz. The length of each timebin is too large to assume that the underlying neuron can fire at most 1 spike per bin. In certain applications, higher resolution of the spike time is needed. A typical example is in “connectomics” where the order with which the various neurons fire spikes is crucial for determining their connect ivity [9], [10]. Moreover, the complexity of the discrete samplers scales with the number of observed timebins (i.e. with the temporal resolution). In experiments where this resolution is very fine discrete, binary samplers can become computation- ally expensive. Compounding this, neural activity may be very sparse and thus sampling at intervals where no spikes occur is uninformative. Instead, a method that scales with the number of spikes is more desirable. To address these issues we propose a sampler that samples directly the spike times in continuous time, given the calcium observations. The state space corresponds to the set of spike times, and transdimensional moves change the dimensionality of this state space. At every iteration we sample the new time of the i-th spike ti given the current location of the rest of the spikes and the hyperparameters. Since the number of spikes K is in general unknown, we also allow for spikes insertions and deletions at each iteration. Similar approaches have been proposed before, see e.g. [11] for the context of signals with finite rate of innovation, where however the number of “spike s” K was considered known. In the continuous time setup, the calcium evolution is described by the differential equation c(t) = − c(t) + As(t), d 1 τ dt where τ is the (continuous) time constant of the calcium indicator. For a timebin of size ∆ with ∆ ≪ τ the discrete and continuous time constants are related through γ = 1 − ∆/τ . If t1 , . . . , tk are the spike times, denoted by T, the calcium signal is given by e− (3) t−t k τ 1{t≥tk } , cT (t) = c1 e− t τ + A T Xk=1 where the subscript T indicates the dependance of the calcium signal on the set of spike times T, and the observations are interpreted as y (n) = c(n∆) + b + εn , giving a likelihood of the form (we ignore c1 for simplicity) 1 2σ2 ky − b1T − cTk2 , where cT = [cT (∆), . . . , cT (T ∆)]T . Finally, since the algo- rithm now operates in the continuous domain, we replace the i.i.d. Bernoulli prior for the spikes with a homogeneous continuous Poisson process prior with parameter λ. If we set a prior λ ∼ Gamma(α, β ) (with β denoting the rate), then log p(y T, θ , σ2 ) ∝ − λT ∼ Gamma(α + T, β + T ∆). p(Tλ)p(λα, β )dλ = For the hyper-parameters α, β we can compute again the evidence function: p(Tα, β ) = Z ∞ βαΓ(α + T) (β + T ∆)a+TΓ(α) 0 p(Tα, β ) is maximized for α, β → ∞ with β = α∆T /T. By choosing these values the posterior of λ puts all the mass in the maximum likelihood estimate λMLE = T/∆T . We can either use this deterministic value or assume a fixed finit e value for α and set β = α∆T /T at each iteration. We next describe the sampling of the spike times in more detail. . A. Move existing spikes We can update individual spike times either using random walk Metropolis-Hastings (MH) moves or a Gibbs algorithm. For the MH algorithm, we propose a move ti → t′ i , using a Gaussian density, centered at ti and with standard deviation equal to e.g. 10∆, as a proposal distribution. Since this proposal is symmetric, the proposed move is accepted with probability i ) = min (cid:18)1, P(ti → t′ p(Tnew y , θ , σ2 ) p(Tcur y , θ , σ2 ) (cid:19) . (4) Similarly to the discrete case, each spike contributes an exponentially decaying calcium trace (3), that has a temporally localized effect to the whole vector cT . By exploiting this, we can implement the spike shifting operation in just O(1) time, provided that we keep in memory the residual vector w ∈ RT with w = y − b1T − cT . After any operation Tcur → Tnew , the vectors c and w can be locally updated, and the ratio in (4) can be computed in just O(1) time. An alternative random walk MH method uses a local proposal density more tuned to the calcium data than is a Gaussian random walk proposal. For each spike time, we construct a distribution from the residual between the data and the current calcium signal, restricted to small time interval I centered on the current spike time, i.e., I = [ti − L, ti + L] with e.g., L = 10∆. The proposal distribution can then be expressed as log A(ti → t′ i ) ∝ − i } (n∆))2 , (y (n) − b − cT\i∪{t′ 1 2σ2 Xn:n∆∈I where T\i is the set of spike times excluding ti . Note that in this case, the proposal distribution is no longer symmetric (since the local intervals are different), and therefore the Hast- ings ratio is also included in the acceptance probability. How- ever, this can also be computed in O(1) time2 and thus this scheme remains efficient. Finally we note that this proposal distribution can be used to derive a Rao-Blackwellized scheme for updating the continuous time posterior spike distribution, instead of using the actual spike samples. This method works quite well empirically. Lastly, instead of the local random walk methods for moving spikes, we may use Gibbs sampling. We can sample each new location t′ i from the likelihood p(y T\i ∪ {t′ i}, θ , σ2) using e.g., rejection sampling. While this approach mixes with less samples than using MH (intuitively the spike can move to anywhere, instead of only locally), the computational cost for moving each spike is O(T ) which is undesirable. B. Adding and removing spikes The sampler over spike trains also requires transdimensional moves in order to sample over the number of spikes. To add or remove spikes, we follow a standard birth-death MH algorithm [12]. We choose a fixed probability z for proposing new spikes and uniform proposal densities for adding and removing spikes. Suppose we want to add a new spike tk+1 = ξ to the existing set of spikes T. This is accepted with probability with rbirth = P(T → T ∪ {ξ}) = min(1, rbirth ), p(T ∪ {ξ}y , θ , λ, σ2)(1 − z )qd (T ∪ {ξ}, {ξ}) p(Ty , θ , λ, σ2)z qb (T, {ξ}) where qd (T∪{ξ}, {ξ}) is the proposal probability for removing {ξ} from T ∪ {ξ}, and qb (T, {ξ}) is the proposal density for adding {ξ} to T. p(T ∪ {ξ}y , θ , λ, σ2 ) and p(Ty , θ , λ, σ2) , 2 Technically, this computation scales with the resolution at which we want to discretize the proposal distribution. In practice one can use a coarse resolution to choose a bin from the proposal distribution and then sample inside this bin uniformly. are posterior probabilities of the spike train given the data and the spiking prior respectively after and before the proposed moves. Under uniform proposals and T = K , we have qd (T∪ {ξ}, {ξ}) = 1/(K + 1), qb (T, {ξ}) = (∆T )−1 , giving rbirth = p(y T ∪ {ξ}, θ , σ2)(1 − z )λ∆T p(y T, θ , σ2)z (K + 1) Similarly the acceptance ratio for removing spike ti = η is . rdeath = = p(T\{η}y , θ , σ2)z qb (T\{η}, {η}) p(Ty , θ , σ2 )(1 − z )qd (T, {η}) p(y T\{η}, θ , σ2)zK p(y T, θ , σ2 )(1 − z )λ∆T . Following [13], a typical choice of the prior proposal prob- ability is z = 1/2, and we repeat the birth-death sampling process 10 times per iteration. Each iteration of the algorithm is schematically presented in Alg. 1. Algorithm 1 Schematic representation of each iteration for i = 1 to current number of spikes K do Sample ti T\i , λ, θ , σ2 using MH. for j = 1 to 10 do Propose addition of spikes. Update K . Propose removal of spikes. Update K . Sample parameters λ, θ , σ2 . Similarly to the case of shifting spikes, the addition and removal can also be implemented in just O(1) time. Using a similar argument it is also easy to show that θ can also be sampled in O(1) time using the method described in section II. As a result, the complexity of each iteration scales linearly with the number of spikes and not the number of timebins, rendering this algorithm particularly attractive for recordings that are very sparse and/or have a fine temporal resolution. V. COL LA P S ED G IBB S SAM P L ER It is possible to marginalize over the baseline b and initial value c1 and enhance the mixing rates of our algorithm [14]. We present this approach here for the discrete case and note that the continuous case can be treated similarly. For this part we assume σ2 to be fixed. By denoting the marginal prior for [b; c1 ] ∼ N (µb , Σb ), the marginal likelihood is computed as3 p(y s, A, p) = Z Zb,c1∈R2 A2 ∝ exp (cid:18)− sT G−T V G−1s + 2σ2 p(y s, p, θ )p(b, c1)db dc1 sT G−T V yb(cid:19) , A σ2 with yb = y − [1T , v ]µb V = I + [1T , v ](σ2Σ−1 b + [1T , v ]T [1T , v ])−1 [1T , v ]T Calcium Traces Data MAP MCMC Marginal spike probabilities 200 400 600 Spike Amplitude samples Timestep 800 1000 1200 1400 Prior spiking probability samples 100 200 300 500 400 Sample # Posterior PDF of baseline 600 700 800 0.09 0.08 0.07 0.06 0.05 400 300 200 100 200 100 300 500 400 Sample # Posterior PDF of initial condition 600 700 800 0.6 0.4 0.2 0 1 0.5 0 0.1 0.09 0.08 0.07 3000 2000 1000 0 0.041 0.0415 0.042 0.0425 0.043 0.0435 0 0.04 0.045 0.05 0.055 0.06 Fig. 1. Application of the collapsed, discrete Metropolized-Gibbs sampler on real spinal cord data (in-vitro). Top row: Real data (red dashed), MAP estimate (green) and mean from 800 Gibbs samples (blue). Second row: Marginal spike probabilities per timestep. The true spike times are also shown (purple circles). Third row: 800 samples from the spike amplitude (left) and the prior spiking probability (right). Bottom row: Rao-Blackwellized estimates of the posterior distributions for the baseline (left) and initial value (right). The collapsed MG sampler mixes fast and provides low variance estimates for the model parameters. 1 2 1 2 and we used the fact that for x ∼ N (0, Φ) and a symmetric matrix C such that Φ−1 − C is positive de finite, we have Ex he xT C x+bT x i = I − ΦC −1/2 e bT (Φ−1−C )−1 b . The marginalized likelihood has the same functional form and therefore s and A can again be sampled with the same methods. Moreover, by exploiting the structure of V it is easy to see that each multiplication of the form V G−1 s can be performed in O(1/ log(1/γ )) time and therefore the complexity of each full Gibbs sweep still scales as O(T ) and the algorithm remains efficient. After the initial burn- in period, the (Rao-Blackwellized) posterior of the marginalized parameters b, c1 can be approximated as M Xi=1 each and values sampled the are where s(i) , A(i) p(b, c1 s(i) , A(i) ) can be computed by conditioning p(θ s(i) ). Note that σ2 is treated as a known parameter here. If we assume an inverse Gamma prior then the posterior is no longer inverse Gamma. However σ2 can still be sampled with standard rejection sampling methods. Finally marginalization over A is also possible, however the marginalized posterior of s has no longer the nice quadratic form and the computational complexity of the algorithm increases. p(b, c1 s(i) , A(i) ), p(b, c1 s, A) ≈ 1 M 3 Technically this approach is approximate since [b, c1 ] have truncated normal priors but here are integrated over the whole R2 . However, in practice the posterior of [b, c1 ] puts very little (if not negligible) mass on negative values and the approximation is very tight. V I . RE SULT S We apply our methods to calcium imaging data from spinal cord neurons in-vitro, collected using the calcium indicator Data MCMC 1 0.8 0.6 0.4 0.2 0 Calcium Traces Marginal spike PDF 0 10 20 30 40 50 Time [sec] Sampled spike train 60 70 80 90 s e k i p s f o # 4 3 2 1 0 200 400 600 Timestep 800 1000 1200 1400 Fig. 2. Application of the collapsed, discrete Metropolized-Gibbs sampler on real spinal cord data (in-vitro). Top row: Real data (red dashed) and mean from 700 Gibbs samples (blue). Second row: Estimated marginal spike PDF in continuous time. Third row: A typical samples of the spike train in continuous time, binned in the resolution defined by the imag ing rate.The true spike times are also shown (purple circles). The continuous time sampler can assign multiple spikes at each timebin and provide better approximation of the spiking behavior in high-SNR conditions. GCaMP6s with a temporal resolution 15Hz [15]. The neurons were stimulated antidromically (so we know the spike times) and fired small bursts of spikes. In several timebins multipl e spikes were fired. Fig. 1 shows an application of the discrete algorithm with b, c1 marginalized out. 1000 samples were collected with the first 200 discarded (burn-in period). Thi s particular trace is of low SNR, however the algorithm predicts mosts of the spikes and provides low variance estimates of the model parameters. A limitation of the discrete algorithm is that it can only produce one spike per timebin. To resolve this issue we run the continuous time sampler to the same dataset. All the traces from imaged pixels that correspond to the same neuron were averaged to produce a high-SNR trace. The algorithm produced 500 samples after a burn in period of 200 samples. The results are shown in Fig. 2. Again the algorithm predicts well the produced calcium trace and provides estimates of the spike posterior in continuous time. By binning the produced sample in the original bin size (bottom row) we see that the algorithm can assign multiple spikes per timebin and better approximate the true spikes. V I I . CONCLU S ION S - FUTURE WORK We presented two classes of Bayesian methods for spike train inference from calcium imaging data. Our methods provide a principled approach for estimating the posterior distribution of the spike trains and provide robust estimates of the model parameters, especially in low-SNR conditions. We also derived collapsed Gibbs samplers that exhibit faster mixing with no significant computational cost per sample. In future work, we plan to explore the use of Hamiltonian Monte Carlo [16] and particle Markov chain Monte Carlo [17] methods for more efficient spike sampling, and extend our methods to allow for a slowly time-varying baseline, a phenomenon that is often observed in vivo experimental conditions. We also plan to scale up to a spatial setup where each measurement corresponds to a pixel that is part of a neuron (or is shared across a small number of neurons) [18], and in the case of compressive calcium imaging [19], where each measurement is formed by projecting the calcium activity of the whole imaged spatial onto a random vector. ACKNOW L EDGMENT We thank S. Chandramouli and D. Carlson for useful discussions, and T. Machado for sharing his spinal cord data. LP is supported from an NSF career award. This work is also supported by ARO MURI W911NF-12-1-0594. RE F ERENCE S [1] M. B. Ahrens, M. B. Orger, D. N. Robson, J. M. Li, and P. J. Keller, “Whole-brain functional imaging at cellular resolution us ing light-sheet microscopy,” Nature methods, vol. 10, no. 5, pp. 413–420, 2013. [2] J. Vogelstein, B. Watson, A. Packer, R. Yuste, B. Jedynak, and L. Panin- ski, “Spike inference from calcium imaging using sequentia l monte carlo methods,” Biophysical journal, vol. 97, no. 2, pp. 636–655, 2009. [3] J. Vogelstein, A. Packer, T. Machado, T. Sippy, B. Babadi, R. Yuste, and L. Paninski, “Fast non-negative deconvolution for spike tr ain inference from population calcium imaging,” Journal of Neurophysiology, vol. 104, no. 6, pp. 3691–3704, 2010. [4] B. F. Grewe, D. Langer, H. Kasper, B. M. Kampa, and F. Helmchen, “High-speed in vivo calcium imaging reveals neuronal netwo rk activity with near-millisecond precision,” Nature methods, vol. 7, no. 5, pp. 399– 405, 2010. [5] J. O nativia, S. R. Schultz, and P. L. Dragotti, “A finite r ate of innovation algorithm for fast and accurate spike detection from two-photon calcium imaging,” Journal of neural engineering, vol. 10, no. 4, p. 046017, 2013. [6] G. Casella, Biostatistics, vol. 2, “Empirical Bayes Gibbs sampling,” no. 4, pp. 485–500, 2001. [7] C. Bishop, Pattern Recognition and Machine Learning. Springer, 2006. [8] J. S. Liu, “Metropolized Gibbs sampler: an improvement, ” Technical report, Dept. Statistics, Stanford Univ, Tech. Rep., 1996. [9] Y. Mishchenko, J. Vogelstein, and L. Paninski, “A Bayesi an approach for inferring neuronal connectivity from calcium fluorescent i maging data,” The Annals of Applied Statistics, vol. 5, no. 2B, pp. 1229–1261, 2011. [10] S. Keshri, E. Pnevmatikakis, A. Pakman, B. Shababo, and L. Paninski, “A shotgun sampling approach for solving the common input pr oblem in neural connectivity inference.” 2013, arXiv preprint ar Xiv:1309.372. te rate [11] V. Y. F. Tan and V. K. Goyal, “Estimating signals with fini of innovation from noisy samples: A stochastic algorithm,” Signal Processing, IEEE Transactions on, vol. 56, no. 10, pp. 5135–5146, 2008. [12] J. Moller and R. P. Waagepetersen, Statistical inference and simulation for spatial point processes. CRC Press, 2003. [13] R. P. Adams, I. Murray, and D. J. MacKay, “Tractable nonp arametric bayesian inference in poisson processes with gaussian process intensi- ties,” in Proceedings of the 26th Annual International Conference on Machine Learning. ACM, 2009, pp. 9–16. [14] J. S. Liu, “The collapsed Gibbs sampler in Bayesian comp utations with applications to a gene regulation problem,” Journal of the American Statistical Association, vol. 89, no. 427, pp. 958–966, 1994. [15] T. Machado, L. Paninski, and T. Jessell, “Functional or ganization of motor neurons during fictive locomotor behavior revealed by large-scale optical imaging,” in SFN Neuroscience, 2013. [16] A. Pakman and L. Paninski, “Auxiliary-variable exact H amiltonian Monte Carlo samplers for binary distributions,” in Advances in Neural Information Processing Systems 26, 2013. [17] C. Andrieu, A. Doucet, and R. Holenstein, “Particle Mar kov chain Monte Carlo methods,” Journal of the Royal Statistical Society: Series B (Statistical Methodology), vol. 72, no. 3, pp. 269–342, 2010. [18] E. Pnevmatikakis, T. Machado, L. Grosenick, B. Poole, J. Vogelstein, and L. Paninski, “Rank-penalized nonnegative spatiotempo ral deconvo- lution and demixing of calcium imaging data,” in Computational and Systems Neuroscience Meeting COSYNE, 2013. [19] E. Pnevmatikakis and L. Paninski, “Sparse nonnegative deconvolution for compressive calcium imaging: algorithms and phase transitions,” in Advances in Neural Information Processing Systems 26, 2013.
1911.02364
2
1911
2019-11-08T03:51:16
Rapid Anxiety Reduction (RAR): A unified theory of humor
[ "q-bio.NC" ]
Here I propose a novel theory in which humor is the feeling of Rapid Anxiety Reduction (RAR). According to RAR, humor can be expressed in a simple formula: -d(A)/dt. RAR has strong correspondences with False Alarm Theory, Benign Violation Theory, and Cognitive Debugging Theory, all of which represent either special cases or partial descriptions at alternative levels of analysis. Some evidence for RAR includes physiological similarities between hyperventilation and laughter and the fact that smiles often indicate negative affect in non-human primates (e.g. fear grimaces where teeth are exposed as a kind of inhibited threat display). In accordance with Benign Violation Theory, if humor reliably indicates both a) anxiety induction, b) anxiety reduction, and c) the time-course over which anxiety is reduced, then the intersection of these conditions productively constrains inference spaces over latent mental states with respect to the values and capacities of the persons experiencing humor. In this way, humor is a powerful cypher for understanding persons in both individual and social contexts, with far-reaching implications. Finally, if humor can be expressed in such a simple formula with clear ties to phenomenology, and yet this discovery regarding such an essential part of the human experience has remained undiscovered for this long, then this is an extremely surprising state of affairs worthy of further investigation. Towards this end, I propose an analogy can be found with consciousness studies, where in addition to the "Hard problem" of trying to explain humor, we would do well to consider a "Meta-Problem" of why humor seems so difficult to explain, and why relatively simple explanations may have eluded us for this long. (Please note: RAR was conceived in 2008, and last majorly updated in 2012.)
q-bio.NC
q-bio
Rapid Anxiety Reduction (RAR): A unified theory of humor Adam Safron Indiana University Abstract Here I propose a novel theory in which humor is the feeling of Rapid Anxiety Reduction (RAR). According to RAR, humor can be expressed in a simple formula: -d(A)/dt. RAR has strong correspondences with False Alarm Theory, Benign Violation Theory, and Cognitive Debugging Theory, all of which represent either special cases or partial descriptions at alternative levels of analysis. Some evidence for RAR includes physiological similarities between hyperventilation and laughter and the fact that smiles often indicate negative affect in non-human primates (e.g. fear grimaces where teeth are exposed as a kind of inhibited threat display). In accordance with Benign Violation Theory, if humor reliably indicates both a) anxiety induction, b) anxiety reduction, and c) the time-course over which anxiety is reduced, then the intersection of these conditions productively constrains inference spaces over latent mental states with respect to the values and capacities of the persons experiencing humor. In this way, humor is a powerful cypher for understanding persons in both individual and social contexts, with far-reaching implications. Finally, if humor can be expressed in such a simple formula with clear ties to phenomenology, and yet this discovery regarding such an essential part of the human experience has remained undiscovered for this long, then this is an extremely surprising state of affairs worthy of further investigation. Towards this end, I propose an analogy can be found with consciousness studies, where in addition to the "Hard problem" of trying to explain humor, we would do well to consider a "Meta-Problem" of why humor seems so difficult to explain, and why relatively simple explanations may have eluded us for this long. (Please note: RAR was conceived in 2008, and last majorly updated in 2012.) Table of Contents Rapid Anxiety Reduction (RAR) .............................................................................................................. 2 Comparisons with other theories ........................................................................................................... 2 False Alarm Theory (FAT) ....................................................................................................................... 2 Benign Violation Theory (BVT) .............................................................................................................. 3 Cognitive Debugging Theory (CDT) ...................................................................................................... 3 Directions for further research............................................................................................................... 3 References ....................................................................................................................................................... 4 Rapid Anxiety Reduction (RAR) RAR can be stated simply as follows: Humor is the feeling of rapid anxiety reduction. Considered as a continuous function, the moment-to-moment experience of humor can be quantified as the negative first (or possibly second4) derivative of anxiety with respect to time: Humor = -d(A)/dt Less simply, both reduction of anxiety as well as rapidity of change are necessary and sufficient conditions for producing the subjective experience of humor. According to RAR, humor is experienced as pleasurable and reinforcing because a negative affective state is removed5. This release will be experienced as especially pleasurable and reinforcing if the change occurs rapidly, both because the contrast will be enhanced between high and low anxiety states, and because such changes will be more effective at producing greater magnitude reward-prediction errors, with concomitant phasic increases in neuromodulators such as dopamine6,7. RAR is a unified model for understanding all humorous phenomena, without exception. For jokes that do not appear to involve rapid releases of anxiety, these seeming exceptions are predicted to be less humorous, and thus constitute further evidence for the theory. For example, most people think that puns are only moderately funny, but puns are also only moderately effective at inducing anxiety, and most people are only moderately effective at engaging in cognitive reframing capable of rapidly releasing that anxiety. This formulation is not intended reduce all humorous experiences to the rapid reduction of anxiety in an eliminative sense. Rather, humor is deployed in numerous and flexible ways, with each of these likely having substantial impacts on the specific ways that humor is experienced. Yet across all of these cases, the core emotion8 may be one of rapid anxiety reduction. Comparisons with other theories Several competing and complimentary theories of humor have been proposed, with RAR having the most in common with False Alarm Theory (FAT)9, Benign Violation Theory (BVT)10, and Cognitive Debugging Theory (CDT)11. False Alarm Theory (FAT) FAT states that humor involves a gradual build-up of expected threat, followed by a sudden non-threatening change in expectation, such that anticipation of threat is reduced. Laughter would serve as a means of informing conspecifics that they need not orient to the false alarm, and this ethological function is proposed to be the primary selective pressure leading to the origin of humor. RAR can be considered to be a more precise formulation of this model, with FAT providing an evolutionary account of selective pressures leading to the -- likely brainstem mediated12 -- 14 -- laughing reflex. However, RAR proposes humor itself may have originally evolved as a byproduct of selection for more fundamental capacities15. More specifically, an organism will respond with pleasure to threat reduction to the extent that associated anxiety is experienced as aversive, can be rapidly released, and thereby produces negative reinforcement. In this way, humorous experience could have evolved regardless of the social signaling value of laughter, since the underlying mechanisms are essential for basic goal-oriented behavior16,17. To the extent that humor is uniquely developed in humans, this may be explainable by the fact that we can so easily experience anxiety from both subtle and complex patterns, and also rapidly reduce that experience via cognitive reframing. Benign Violation Theory (BVT) BVT states that humor involves the perception of a violation that is simultaneously perceived as benign. RAR can be considered to be a broader principle that explains BVT, with the addition of two crucial details: 1. The humorous effect of benign violation is mediated by anxiety reduction. 2. Humor is inversely proportional to the time it takes for the anxiety to be reduced. To the extent that there seem to be exceptions to BVT, it is because either a) the violation fails to produce anxiety, or b) the benign reframing fails to occur rapidly. Cognitive Debugging Theory (CDT) CDT states that humor is an evolved process that helps humans to maintain the data- integrity of world knowledge by rewarding them for detecting errors. Again, RAR can be considered to be a broader explanatory principle for which CDT is a special case. That is, cognitive debugging events are humorous to the extent that they involve a rapid reduction of anxiety. However, this is not the primary reason that humor evolved. As previously described, RAR considers humor to be a byproduct of organisms that a) experience anxiety as aversive, b) experience pleasure accompanying anxiety reduction, and c) are capable of manipulating their cognitive states in ways that can rapidly reduce anxiety. Rather than a specific mechanism being needed for cognitive debugging, this process of error-detection may happen automatically as part of a basic cortical algorithm that allows the brain to function as a predictive memory system18,19. Humorous pleasure may help to enhance debugging -- and hence represent an additional selective pressure for humorous experience, particularly for organisms that depend on both precise and complex cognition20 -- but this would be just one of many dynamics that contribute to shaping humor sensitivity. Directions for further research RAR appears to be the first unified theory capable of explaining the full range of humorous phenomena. The theory can be falsified if a single example is found wherein either a) substantial humor can be demonstrated without rapid reductions in anxiety, or b) rapid releases in anxiety can be demonstrated without the experience of humor. Granted, this falsifiability is limited by the facts that a) it is difficult to prove the absence of phenomena, and b) the amount of requisite rapidity may vary from person to person. Event-related fMRI could potentially be used to identify particular neural systems associated with humorous experience, some of which may include brainstem, amygdala21, anterior insula22,23, and anterior cingulate cortex24. However, given the time-dependency of humor, better data may be obtained with the superior temporal resolution -- albeit inferior spatial resolution -- of electroencephalography or magnetoencephalography. More compelling support may involve using different theories to construct jokes, and seeing which jokes are most reliably perceived as humorous. For example, emphasizing anxiety and timing should produce superior jokes as compared with BVT. If humor is such a universal experience, and if it can be explained by such a simple formula, then why do so few people think of humor in this way? This question may be answered by the very thing that allows us to take pleasure in humor in the first place. That is, if people are negatively reinforced by the reduction of anxiety, and if contemplating anxiety makes people anxious, then people will be shaped in ways that make them less likely to contemplate anxiety-based models. Further, contemplating anxiety will make jokes less humorous if this activity either a) reduces the amount of induced anxiety, or b) reduces the degree to which that anxiety can be rapidly reduced. For these reasons, many people are likely to find RAR to be counter-intuitive, or at least until they become comfortable with the idea 25,26. References 1. Graziano, M. S. A. The Spaces Between Us: A Story of Neuroscience, Evolution, and Human Nature. (Oxford University Press, 2018). 2. Robinson, R. Stimulating the Cingulum Relieves Anxiety During Awake Neurosurgery: What Is the Therapeutic Potential? Neurol. Today 19, 27 (2019). 3. Chalmers, D. J. The Meta-Problem of Consciousness. in (2018). 4. Joffily, M. & Coricelli, G. Emotional valence and the free-energy principle. PLoS Comput. Biol. 9, e1003094 (2013). 5. Thorndike, E. L. The fundamentals of learning. (Ams PressInc, 1932). 6. Schultz, W. Predictive reward signal of dopamine neurons. J. Neurophysiol. 80, 1 -- 27 (1998). 7. Schultz, W. Dopamine signals for reward value and risk: basic and recent data. Behav. Brain Funct. BBF 6, 24 (2010). 8. Parrott, W. G. Ur-Emotions and Your Emotions: Reconceptualizing Basic Emotion. in (2010). doi:10.1177/1754073909345547. 9. Ramachandran, V. S. The neurology and evolution of humor, laughter, and smiling: the false alarm theory. Med. Hypotheses 51, 351 -- 354 (1998). 10. McGraw, A. P. & Warren, C. Benign violations: making immoral behavior funny. Psychol. Sci. 21, 1141 -- 1149 (2010). 11. Hurley, M. M., Dennett, D. C., Jr, R. B. A. & more, & 1. Inside Jokes: Using Humor to Reverse-Engineer the Mind. (The MIT Press, 2013). 12. Couderq, C. et al. [Pathological laughter after the brainstem infarction]. Rev. Neurol. (Paris) 156, 281 -- 284 (2000). 13. Lal, A. P. & Chandy, M. J. Pathological laughter and brain stem glioma. J. Neurol. Neurosurg. Psychiatry 55, 628 -- 629 (1992). 14. Parvizi, J., Anderson, S. W., Martin, C. O., Damasio, H. & Damasio, A. R. Pathological laughter and crying A link to the cerebellum. Brain 124, 1708 -- 1719 (2001). 15. Gould, S. J. & Lewontin, R. J. The spandrels of San Marco and the Panglossian paradigm: a critique of the adaptationist programme. Proc. R. Soc. Lond. Ser. B Contain. Pap. Biol. Character R. Soc. G. B. 205, 581 -- 598 (1979). 16. Panksepp, J. & Burgdorf, J. 'Laughing' rats and the evolutionary antecedents of human joy? Physiol. Behav. 79, 533 -- 547 (2003). 17. Panksepp, J. Cross-species affective neuroscience decoding of the primal affective experiences of humans and related animals. PloS One 6, e21236 (2011). 18. Friston, K. The free-energy principle: a unified brain theory? Nat. Rev. Neurosci. 11, 127 -- 138 (2010). 19. Hawkins, J. & Blakeslee, S. On Intelligence. (Times Books, 2004). 20. Tomasello, M. THE HUMAN ADAPTATION FOR CULTURE. Annu. Rev. Anthropol. 28, 509 -- 529 (1999). 21. Pessoa, L., Japee, S., Sturman, D. & Ungerleider, L. G. Target visibility and visual awareness modulate amygdala responses to fearful faces. Cereb. Cortex N. Y. N 1991 16, 366 -- 375 (2006). 22. Craig, A. D. B. How do you feel--now? The anterior insula and human awareness. Nat. Rev. Neurosci. 10, 59 -- 70 (2009). 23. Medford, N. & Critchley, H. D. Conjoint activity of anterior insular and anterior cingulate cortex: awareness and response. Brain Struct. Funct. 214, 535 -- 549 (2010). 24. Magno, E., Foxe, J. J., Molholm, S., Robertson, I. H. & Garavan, H. The anterior cingulate and error avoidance. J. Neurosci. Off. J. Soc. Neurosci. 26, 4769 -- 4773 (2006). 25. Viral Tainment. Orangutan finds magic trick hilarious. 26. mandkyeo. Emerson - Mommy's Nose is Scary! (Original).
1610.09431
1
1610
2016-10-29T00:21:31
Attention acts to suppress goal-based conflict under high competition
[ "q-bio.NC" ]
It is known that when multiple stimuli are present, top-down attention selectively enhances the neural signal in the visual cortex for task-relevant stimuli, but this has been tested only under conditions of minimal competition of visual attention. Here we show during high competition, that is, two stimuli in a shared receptive field possessing opposing modulatory goals, top-down attention suppresses both task-relevant and irrelevant neural signals within 100 ms of stimuli onset. This non-selective engagement of top-down attentional resources serves to reduce the feedforward signal representing irrelevant stimuli.
q-bio.NC
q-bio
Attention acts to suppress goal-based conflict under high competition Author names: Claflin, O.1,2 Former Author Affiliations: 1Departments of Psychiatry & Psychology, University of California, Los Angeles 90095, USA. 2Department of Bioengineering, University of California, San Francisco, 94158, USA. 3Departments of Physiology, Neurology and Psychiatry, and the Center for Integrative Neuroscience, University of California, San Francisco, 94158, USA. Number of figures: 3; Number of tables: 3 Word count: 1350; Abstract: 88; References: 28 Acknowledgements: Thanks to Joseph Hsiao and Anh Tranh for collecting the data. Thanks to arxiv.org for a medium to post this manuscript on graduate work completed in Adam Gazzaley's lab at UCSF. 1 Abstract It is known that when multiple stimuli are present, top-down attention selectively enhances the neural signal in the visual cortex for task-relevant stimuli, but this has been tested only under conditions of minimal competition of visual attention. Here we show during high competition, that is, two stimuli in a shared receptive field possessing opposing modulatory goals, top-down attention suppresses both task-relevant and irrelevant neural signals within 100 ms of stimuli onset. This non-selective engagement of top-down attentional resources serves to reduce the feedforward signal representing irrelevant stimuli. It is well established that attention modulates visual processing in extrastriate cortex (Tsotsos et al., 1995). Strong evidence for the mutual competition theory, acting at the level of the receptive fields in the extrastriate cortex, suggests that local neuronal activity representing simultaneous stimuli result in suppression of these representations at the level of the receptive field (Moran and Desimone, 1985; Reynolds et al., 1999; Kastner and Ungerleider, 2001). This local suppression of local activity occurs automatically without top-down influences (Kastner, 1998). Many researchers have shown that top-down spatial attention serves to counteract the local suppressive influences on the stimuli of attentional interest present in the receptive field (Moran and Desimone, 1985; Luck et al., 1997; Kastner, 1998; Reynolds et al., 1999). This mechanism, acting at the level of the receptive field, spares the stimuli outside the field of directed spatial attention, leaving them suppressed, and thereby exerting selective enhancement on the stimuli of interest. However, this mechanism of attention does not address the co-presence of task-relevant and a task- irrelevant stimuli in a shared spatially attended field. Long-range enhancement mechanisms would also enhance the irrelevant stimuli, acting as informational noise to the goal at hand, also present in the same receptive field, to be also amplified with the task-relevant stimuli. Otherwise, forced to act at the resolution of receptive fields, simultaneously presented objects with opposing behavioral relevance that share the same receptive fields may engage top-down noise-suppression mechanisms thereby suppressing the goal-relevant signal in order to suppress the goal-irrelevant signal. However, if these top-down mechanisms are able to act independently of receptive fields, then they will selectively act on these objects without modulatory limitation such that either the enhancement of relevant events or the suppression of irrelevant events, or both, will occur under these high competition settings (sharing space and timing). 2 Attentional selection behavior in the extrastriate cortex has been shown to be strongly dependent on attentional rivalry; namely, presentation containing simultaneously relevant and irrelevant stimuli have elicited new attentional behavior not previously observed (Zhang and Luck, 2009). Studies examining biased competition mechanisms have typically examined the suppressive interactions among irrelevant stimuli on each other (Beck and Kastner, 2009) but omit how competing goal- modulation in the same receptive field or goal-based attentional modulation is handled. Other studies examine simultaneous irrelevant and relevant stimuli sharing the same space and timing (Rutman et al., 2010; Gazzaley, 2011) but use discrete stimuli rather than continuous events. It is possible the ongoing representation result in more real-time stimuli processing mechanisms and, in fact, when used, has been observed to evoke novel attention behavior in studies that have used continuous events (Zhang and Luck, 2009). Using electroencephalography (EEG) recordings and a continuous visuomotor task (modified from Anguera et al., 2013), we visualized stimulus competition affecting the earliest perceptual stream, as reflected by the P1 component, under different conditions of attentional rivalry (presence or absence of simultaneous relevant and irrelevant events; Table 1) and under different event timing overlaps (Event Onset Asynchrony (EOA); Table 1). A continuous visuomotor tracking task that required compensatory tracking adjustments was co-presented with a perceptual discrimination task that elicited an observable P1 component. Participants actively engaged with the tracking events (TID), the discrimination events (DIT), or both (MT). For our control condition, participants underwent an Ignore All (IA) condition where they simply attended to central fixation, while event onset asynchrony (EOA) between the two event types were jittered as in all the other conditions. In line with previous findings (Hillyard et al., 1998; Beck and Kastner, 2009), P1 amplitude was larger for relevant events than irrelevant events (MT/DIT > TID/IA, p < 0.028; Figure 3A), and larger for isolated stimuli (DO) versus visually competitive stimuli (DO > MT/DIT/TID/IA, p < 0.011; Figure 3A). Additionally, P1 amplitude for behaviorally irrelevant events in our passive control condition (IA) did not show significant attenuation across EOAs (p > 0.94; Table 3B). Using this passive view as a baseline for all other conditions, we observed an impact of event onset asynchrony on conditions with an active task component (MT/DIT/TID; p < 1.1 x 10-3). Thus, using the P1 wave as a neural index, event onset asynchrony affected goal-based attention modulation of early sensory information within 100 ms of stimulus onset. 3 For non-simultaneous event onset (EOAs > 0 ms), irrelevant events remained at baseline levels of modulation (TID; p > 0.3) while relevant events demonstrated increased amplitude (MT, DIT; p < 0.03). For relevant events, a decline in this enhancing modulation was observed as event onset asynchrony increased (Figure 3B: MT: F(1,19) = 16.21, p = 7.2 x 10-4; DIT: F(1,19) = 49.12, p = 1.13 x 10-6). Thus the ability to selectively enhance relevant events during stimulus competition in the face of decreasing event recovery time diminished even when both events were relevant. For conditions where competing attentional modulation was present (presence of relevant and irrelevant events; e.g. DIT, TID), simultaneous event onset resulted in a suppression of the P1 wave (Figure 3B: DIT, TID). Thus, in the case of simultaneous event onset (EOA 0 ms) and attentional rivalry, attention modulates the flow of early sensory information within 100 ms of stimulus onset by suppression of the event, regardless of its relevance. In contrast to previous competitive suppression findings, this suppression results in a signal strength below the passive view task (rather than just a relative suppression when compared to the single-task condition). This observed decrease in P1 amplitude was not a bottom-up effect, as no decrease was observed in the signal strength in the passive condition sharing the same stimulus presentation (IA300/IA600 = IA0, t(1,19) < 0.253, p > 0.8). Nor was the simple presence of top-down modulatory engagement explanatory as the dual-relevant condition did not exhibit this suppression (DIT0/TID0 < MT0, t(1,19) = 2.51, p < 0.05). Rather, an interaction revealed the presence of competing top-down influences (Attentional Rivalry * EOA; (F(2,38) = 11.3, p = 1.4 x 10-4)) explaining the pattern of P1 modulation in the setting of maximum stimulus competition (EOA 0 ms; Table 2: p < 0.029). However, as would be expected from the literature, post hoc t-tests revealed stimulus relevance was the explanatory factor affecting P1 modulation in less competitive settings (EOA 600 ms) (Table 2: p < 0.006), driving the stimulus relevance * EOA interaction (F(2,38) = 8.07, p = 1.1x10-3). These results suggest that attentional modulation responds to simultaneous goal rivalry between relevant and irrelevant events by suppressing early sensory representation for both events. Our experiment further demonstrates (1) that the competitive suppression between spatially-shared events is reflected in the visual signal, (2) that this competition is differentially alleviated by top- down enhancement of goal-relevant events, and (3) that this ability to selectively enhance linearly degrades with decreasing temporal spacing between events. This EOA-dependence of selective enhancement may represent a local neuronal competition effect in which local neurons representing 4 these receptive fields become more sensitive to long-range enhancing projections with more recovery time between event onsets. Previous research examining attentional modulation has shown attentional processing relies on general local suppression at the extrastriate cortex (Moran and Desimone, 1985; Reynolds et al., 1999; Beck and Kastner, 2005) accompanied by spatially-selective and feature-selective enhancement (Luck et al., 2000; Zhang and Luck, 2009), as well as a reliance on selective modulation at later stages of processing of object-relevant or irrelevant areas such as the FFA or PPA (Egner and Hirsch, 2005; Gazzaley et al., 2005). Considered together, these findings suggests that under conditions of high visual competition and opposing goal-relevance, top-down attention suppresses early stage signals to decrease feedforward signal from the irrelevant events and potentially rely on later processing stages of attentional selectivity for discrimination. These findings support a goal-rivalry-dependent suppression of both relevant and irrelevant events during the earliest sensory processing stages in shared receptive fields under maximum temporal overlap. We hypothesize that this pattern of activity serves to reduce the feedforward signal representing irrelevant stimuli for further stages of processing, acting as an independent top-down suppression mechanism in the visual cortex. The observed suppression may represent an optimal strategy for the early selective attention mechanism when relevant and irrelevant events cannot be discriminated either spatially or temporally. In this case, reducing the early processing of the irrelevant event may be more advantageous compared to trying to enhance the relevant information along with unrelated noise. It may be that this mechanism of attention reflects a meta-monitoring of frontal processes that engage when lower-level features cannot selectively discriminate (Botvinick et al., 1999; Kerns, 2004; Carter and van Veen, 2007). 5 References Al-Hashimi O, Zanto TP, Gazzaley A (2015) Neural sources of performance decline during continuous multitasking. Cortex 71:49–57. Anguera JA, Boccanfuso J, Rintoul JL, Al-Hashimi O, Faraji F, Janowich J, Kong E, Larraburo Y, Rolle C, Johnston E, Gazzaley A (2013) Video game training enhances cognitive control in older adults. Nature 501:97–101. Beck DM, Kastner S (2005) Stimulus context modulates competition in human extrastriate cortex. Nat Neurosci 8:1110–1116. Beck DM, Kastner S (2009) Top-down and bottom-up mechanisms in biasing competition in the human brain. Vision Res 49:1154–1165. Botvinick M, Nystrom LE, Fissell K, Carter CS, Cohen JD (1999) Conflict monitoring versus selection-for-action in anterior cingulate cortex. Nature 402:179–181. Carter CS, van Veen V (2007) Anterior cingulate cortex and conflict detection: an update of theory and data. Cogn Affect Behav Neurosci 7:367–379. Cavanagh JF, Frank MJ, Klein TJ, Allen JJB (2010) Frontal theta links prediction errors to behavioral adaptation in reinforcement learning. Neuroimage 49:3198–3209. Egner T, Hirsch J (2005) Cognitive control mechanisms resolve conflict through cortical amplification of task-relevant information. Nat Neurosci 8:1784–1790. Gazzaley A (2011) Influence of early attentional modulation on working memory. Neuropsychologia 49:1410–1424. Gazzaley A, Cooney JW, Rissman J, D'Esposito M (2005) Top-down suppression deficit underlies working memory impairment in normal aging. Nat Neurosci 8:1298–1300. Hillyard S a, Vogel EK, Luck SJ (1998) Sensory gain control (amplification) as a mechanism of selective attention: electrophysiological and neuroimaging evidence. Philos Trans R Soc Lond B Biol Sci 353:1257–1270. Kastner S (1998) Mechanisms of Directed Attention in the Human Extrastriate Cortex as Revealed by Functional MRI. Science (80- ) 282:108–111. Kastner S, Ungerleider LG (2001) The neural basis of biased competition in human visual cortex. Neuropsychologia 39:1263–1276. Kayser J, Tenke CE, Gates NA, Bruder GE (2007) Reference-independent ERP old/new effects of auditory and visual word recognition memory: Joint extraction of stimulus- and response- locked neuronal generator patterns. Psychophysiology 44:949–967. Kayser J, Tenke CE, Gates NA, Kroppmann CJ, Gil RB, Bruder GE (2006) ERP/CSD indices of impaired verbal working memory subprocesses in schizophrenia. Psychophysiology 43:237– 252. 6 Kerns JG (2004) Anterior Cingulate Conflict Monitoring and Adjustments in Control. Science (80- ) 303:1023–1026. Luck S, Woodman G, Vogel E (2000) Event-related potential studies of attention. Trends Cogn Sci 4:432–440. Luck SJ, Chelazzi L, Hillyard S a, Desimone R (1997) Neural mechanisms of spatial selective attention in areas V1, V2, and V4 of macaque visual cortex. J Neurophysiol 77:24–42. Martinez A, DiRusso F, Anllo-Vento L, Sereno M, Buxton R, Hillyard S (2001) Putting spatial attention on the map: timing and localization of stimulus selection processes in striate and extrastriate visual areas. Vision Res 41:1437–1457. Mishra J, Gazzaley a. (2012) Attention Distributed across Sensory Modalities Enhances Perceptual Performance. J Neurosci 32:12294–12302. Moran J, Desimone R (1985) Selective attention gates visual processing in the extrastriate cortex. Science (80- ). Perrin F, Pernier J, Bertrand O, Echallier JF (1989) Spherical splines for scalp potential and current density mapping. Electroencephalogr Clin Neurophysiol 72:184–187. Reynolds JH, Chelazzi L, Desimone R (1999) Competitive mechanisms subserve attention in macaque areas V2 and V4. J Neurosci 19:1736–1753. Rutman AM, Clapp WC, Chadick JZ, Gazzaley A (2010) Early top-down control of visual processing predicts working memory performance. J Cogn Neurosci 22:1224–1234. Tsotsos J, Culhane S, Wai WK, Lai Y (1995) Modeling visual attention via selective tuning. Artif Intell 3702. Woldorff M (1993) Distortion of ERP averages due to overlap from temporally adjacent ERPs: analysis and correction. Psychophysiology:98–119. Zanto TP, Toy B, Gazzaley A (2010) Delays in neural processing during working memory encoding in normal aging. Neuropsychologia 48:13–25. Zhang W, Luck SJ (2009) Feature-based attention modulates feedforward visual processing. Nat Neurosci 12:24–25. 7 Supplementary Methods, Results, Figures, Tables and Legends Methods In the current study, we explored the hypothesis that selective attention mechanisms are limited in their ability to modulate under temporally, spatially and goal competitive conditions. We examine early signal modulation during a continuous paradigm with competing stimuli in a shared visual receptive field. We examined performance and neural measures while multitasking using a jittered design and concurrent electroencephalography (EEG) data collection to assess neural markers of perceptual processing. The paradigm involved variable event onset asynchronies (EOAs) between two tasks to assess the relationship between task timing and neural markers of early visual processing. We used a continuous visuomotor tracking task to generate continuous visual engagement and a forced choice perceptual discrimination task as a punctuated interruptor ('NeuroRacer': Anguera et al., 2013) to generate reliable event related potential (ERP) markers. The paradigm has already been shown to generate interference costs on the behavioral and neural markers of the discrimination task (Anguera et al., 2013; Al-Hashimi et al., 2015), although the impact of EOA is unknown. We hypothesized that as the time between events in the visuomotor task (road turns) and events in the discrimination task (sign onset) was reduced, behavioral and early perceptual neural measures in the discrimination task would show greater interference costs, revealing how early stimulus processing is modulated by the complex demands of goal-related attentional modulation under increasingly temporally competitive conditions. Under these conditions we predicted that the effects of EOA and attentional goal on visual processing attentional modulation would interact to reveal limitations, or a new mode, of goal-related attentional processing. Materials and Methods Participants. Twenty healthy young, right-handed adults (mean age, 24.8 years; range, 20-29 years; 12 females) gave informed consent to participate in the study approved by the Committee on Human Research at the University of California in San Francisco. All participants had normal or corrected- to-normal vision as examined using a Snellen chart. Additionally, all participants were considered to be non-video game players, as defined by having less than 2 hours of any type of video-game usage per month in the past two years. 8 Stimuli and experimental procedure. Stimuli and tasks were presented using a custom designed video game ('NeuroRacer'; Anguera et al., 2013) on a Dell Optiplex GX620 with a 22" Mitisubishi Diamond Pro 2040U CRT monitor. Participants were seated with a chin rest in a dark room 80 cm from the monitor using a Logitech game controller to control tracking (left thumb) and responding to sign types (left & right index fingers). Each experimental run lasted 180 seconds with a few seconds of a self-paced break available to participants after every run. The visuomotor tracking task (Figure 1A) is one of two tasks available in the 'NeuroRacer' environment (Anguera et al., 2013; Al-Hashimi et al., 2015), and involves keeping a car within a target box drawn on a continuously moving road (when the car was not within the target box, the fixation cross would shake to indicate poor performance). A pseudo-randomized, counterbalanced selection of road segments (that is, right/left turns & inclining/declining hills) formed the tracks, with turn/hill severity being either mild or severe. Transitions between road segments were treated as events in the visuomotor task, with each segment lasting 2000, 2500 or 3000 ms. This task involved continuous visuomotor tracking to an always inclining/declining and turning road. Tracking events involved transitions between these road segments where the inclination and/or turn direction and severity was changed suddenly. The point where the car crossed the union of these two road segments was treated as an event that required additional tracking correction to stay within the tracking zone. The discrimination task (Figure 1A) involves responding to visual stimuli presented for 400 ms, less than two degrees above a fixation cross. Discrimination involved a 2-alternative forced choice task where subjects responded to green circles (33% frequency) with a right key press and all other colored shapes with a left key press. Real-time feedback was indicated by a 100 ms color change of the fixation cross one second after stimulus presentation (green for correct, red for incorrect) for the sign task and by a shaking of the fixation cross when the car was outside the target-tracking zone. Thresholded performance. Prior to the experimental runs, participants underwent an adaptive thresholding procedure to assess perceptual discrimination and visuomotor tracking abilities performed in isolation (Figure 1A). For discrimination, a staircase algorithm changed the time window allowed for a correct response for each 120 second run (48 signs) over nine runs. For the visuomotor tracking task, the speed of the road was thresholded with a similar staircase algorithm 9 over twelve 60 second runs. The algorithm increased the difficulty of each task when performance level was over 80% on the previous run and decreased the difficulty when performance was less than 80% (for more details, see Anguera et al., 2013). Upon completion of each thresholding block, difficulty levels were interpolated for 80% performance to estimate an individual discrimination difficulty level and tracking difficulty level for that participant. The difficulty of the experimental tasks were set to these individualized levels for each participant so that individuals engage each condition in their own ability level following thresholding procedures, thus facilitating a fairer comparison across overall differences in perceptual discrimination abilities. Conditions. Following the driving and sign thresholding procedures (Figure 1A), participants performed four perceptually-matched conditions and a fifth sign-only isolation condition randomly counterbalanced across participants (each condition performed three times in a pseudo-randomized fashion). Participants were cued to the upcoming condition before each run: 1: Multitasking (MT; Both tasks relevant); 2: Discrimination Ignore Tracking (DIT; Signs relevant); 3. Tracking Ignore Discrimination (TID; Road relevant); 4. Ignore All (IA; Neither task relevant, attend to fixation); 5. Discrimination Only (DO; Signs task present, on a black background). The road was present in all conditions (Figure 1B) except for Discrimination Only; signs were present for all conditions. These conditions included combinations of attend and ignore goals to the two tasks (Figure 1D), such that for the Discrimination Ignore Tracking and Ignore All conditions, the road task was instructed as irrelevant, driving feedback was turned off and the car was placed on 'auto pilot' for the duration of the run (Figure 1B). For the Tracking Ignore Discrimination and Ignore All condition, the sign task was made irrelevant through similar instruction and disabling of feedback and response mechanisms for the sign. In Multitasking condition, participants were told to respond to the signs as fast and accurately as possible and continue to drive as accurately as possible. Feedback was given at the end of each run as the proportion correct to all signs presented for the perceptual discrimination task and percent time spent in the tracking zone for the driving performance. Prior to the start of the subsequent run, participants were informed as to which condition would be engaged in next, and made aware of how many experimental runs were remaining. Event onset asynchrony. Road events always preceded sign events (except in the case of 0 ms where they coincided) in the multitasking and distraction conditions. Event onset asynchronies of 0, 300 and 600 ms buffered the discrimination task from the road event (see Figure 1C) to observe the temporal impact on attentional modulation. This design was done not only for the multitasking 10 condition (MT), but also for the conditions where both tracking and sign stimuli are present (TID, DIT, IA). Note that tracking events were not always preceded by sign events so that in this paradigm, tracking events an unreliable predictor for a discrimination event. Participants were instructed to fixate at the center of the screen at all times, respond to signs as quickly as possible and keep the car centered in the tracking box at all times. Feedback was given at the end of each run as per their percentage correct on the discrimination task and percentage time spent within the tracking box. Correct responses to the appropriate signs within the thresholded response time window were categorized as hits; non-responses, late responses or mismatched key presses to stimuli were counted as incorrect for purposes of feedback. In the multitasking condition (MT), feedback was given for both tasks. Both tasks were equally emphasized in the instructions and briefing. Performance costs. While both tasks were equally emphasized in instruction to the participants, we were focused on the performance effects on the discrimination task in our paradigm since it typically bears the brunt of multitasking and distraction costs (Anguera et al., 2013; Al-Hashimi et al., 2015). Driving-related metrics related for the visuomotor task are not discussed here. As an additional level of individualization beyond the individualized thresholding procedure, performance costs were compared for each participant against their undistracted single task performance. Distraction costs and multitasking costs were calculated by subtracting the participant's individual single task performance from their multitasking performance (MT - DO) and their distraction performance (DIT - DO) in order to individually baseline each participant's performance in these conditions. Response times (RTs) and accuracy were analyzed using ANOVAs, with a Greenhouse-Geisser correction when appropriate and post hoc t-test comparisons performed when statistically called for. EEG data acquisition. Data were recorded during nine runs (three conditions, three runs each). Electrophysiological signals were recorded with a BioSemi ActiveTwo 64-channel EEG acquisition system in conjunction with BioSemi ActiView software (Cortech Solutions). Signal were amplified and digitized at 1024 Hz with a 24-bit resolution. All electrode offsets were maintained between +20 mV. EEG data analysis. Preprocessing and further ERP analyses was conducted in Analyzer 2.0 (Brain Vision, LLC). Raw EEG data were digitally re-referenced offline to the average of all electrodes. Eye artifacts were removed through independent component analyses by excluding components 11 consistent with topographies for blinks and eye movements. Data were high-passed filtered at 1 Hz to exclude slow DC drifts. An ERP analysis at posterior electrodes was conducted to assess markers of early visual processing for each sign presented. All ERP analyses were time-locked to the onset of each sign stimuli, yielding 216 epochs of data for each condition (and 72 epochs per EOA). Signals were averaged in 1000 ms epochs with a 200 ms prestimulus interval used as baseline. Epochs that exceeded a voltage threshold of +100 uV were rejected. The P100 and N200 component were evaluated at electrode clusters during the peak latency intervals of 100-180 ms for P100 and 140-260 ms for N200. To minimize electrode bias, activity at posterior electrodes was collapsed as follows: left-lateralized (P9, P7, PO7), right-lateralized (P1000, P8, PO8) and posterior midline (Pz, POz, Oz). Similar to the approach taken in other studies (Zanto et al., 2010; Mishra and Gazzaley, 2012), we used the greatest amplitude of the P100 and N200 in an electrode group when collapsed across all conditions to guide subsequent ERP analyses. This approach led to the posterior midline electrode group being selected for P100 interrogation and the right electrode group for N200 analysis. To analyze the effects of Stimulus Relevance (Relevant, Irrelevant) and Attentional Rivalry (Present: [TID, DIT], Absent: [MT, IA]) (Figure 1B), across the three event onset asynchronies (0, 300, 600 ms), we used a 2x2x3 three-way ANOVA. ERP peak amplitudes were baselined against the control task amplitude (the IA peak was subtracted from peaks of the other conditions) for each participant to reduce intra-subject bias. The appropriate event onset asynchronies (IA0, IA300, IA600) were used to baseline all conditions with event onset asynchronies, and the subject's grand-averaged IA amplitude was used to baseline the amplitudes when analyzing the amplitudes of conditions collapsed across event onset asynchrony. ERP amplitude and latencies were analyzed using ANOVAs, with a Greenhouse-Geisser correction when appropriate and post hoc t-test comparisons performed when statistically called for. Current Source Density (CSD) With respect to the current study goals, it has been suggested that early visual ERPs like the P100 may benefit from current source density (CSD) filtering by removing volume conduction effects (Martinez et al., 2001). Resolving two discrete ERP events from stimuli that are presented closely in time has inherent technological issues (cf. Woldorff, 1993). CSD (Perrin et al., 1989) reduces redundant contributions due to volume conduction providing sharper topographies compared to 12 those of scalp potentials (Kayser et al., 2006, 2007). Reducing volume conduction with CSD transformation may highlight local electrical activity at the expense of diminishing representation of distal activities (Cavanagh et al., 2010). Thus, we applied a CSD filter following preprocessing to reveal subtle visual cortical effects. Results - Behavioral data Performance costs The influence of the EOA manipulation on multitasking costs (response times for the discrimination task during MT – response times for the discrimination task during DO) and distraction costs (response times for the discrimination task during DIT – response times for the discrimination task during DO) on the discrimination task was assessed using a 2 X 3 repeated-measures ANOVA with factors of Condition (MT, DIT) and EOA (0, 300, 600 ms). This analysis revealed a significant main effect of condition (F(1,19) =95.7, p < 7.5x10-9), event onset asynchrony (F(2,38) =38.7, p < 6.8x10-10), but no interaction (F(1,19) =1.2, p = 0.33). The condition main effect indicated that RT costs were greater during MT (48.4 ms  3.4) versus DIT (21.8 ms  1.2; t(19) = 3.17, p = 0.005). In terms of the EOA effect, as the time between stimuli was reduced, costs became greater for both conditions (MT0 >MT300 >MT600: t(19)>3.31, p < 0.004 for each comparison; DIT0 >DIT300 >DIT600: t(19)≥ 1.99, p ≤ 0.062 for each comparison; see Table 1 and Figure 2). Reaction times were subtracted from individual's single-task performance in order to baseline the impact of the additional stimuli on the active task for each participant. Thus, the presence of a tracking event (road turn) (visuomotor task) relative to the time of sign onset (discrimination task) differentially impacted discrimination speed, regardless of whether the participant was actually tracking the road or the road was passively viewed. A similar analysis of accuracy costs revealed a main effect of condition (F(2,38) = 8.9, p = 0.008), but neither a significant main effect of EOA (F(2,38) = 2.5, p = 0.11) nor an interaction (F(2,38) = 0.51, p =0.61). Follow-up analyses revealed participants to be less accurate when multitasking (84% + 2) versus single-tasking with a road background (87%  2; t(19) = 3.35, p = 0.003). Thus, the presence of an active secondary task (that is, visuomotor tracking) reduced discrimination accuracy compared to single-tasking, regardless of EOA. 13 Neural data P100: To examine the influence of attentional modulation on the early visual processing of the discrimination task, we examined ERP P100 amplitude and latency time-locked to the onset of the signs. For P100 peak amplitude, a 4 X 3 repeated-measures ANOVA with the factors of Condition (MT, DIT, TID, IA) and EOA (0 ms, 300 ms, and 600 ms) revealed a main effect of EOA (F(2,38) = 33.6, p < 4x10-9) and a significant 2-way interaction (F(4,76) = 6.3, p < .00019), but no main effect of condition (F(1,19) = 2.11, p = 0.14). Pair-wise comparisons of EOA revealed an amplitude-reducing effect of decreasing EOA on P100 amplitudes for MT, DIT and TID, such that as sign presentation occurred closer to a road event (either a motoric road event (MT, TID) or a passive road event (DIT), the lower the P100 amplitude for that sign (DIT0 <DIT300 <DIT600: t(19)>= 2.37, p <= 0.029 for each comparison; MT0 <MT300 =MT600: t(19)>=2.38, p <= 0.028; TID0 <TID300 =TID600: t(19)>= 3.25, p <= 0.004; see Table 2 and Figure 3). Additionally, pair-wise comparisons of conditions by EOA revealed MT and DIT to have greater P100 amplitudes at 600 milliseconds (see Table 2 and Figure 3). P100 latency analysis revealed a main effect of EOA (F(2,38) = 7.09, p = 2.4x10-3), but no main effect of condition (F(3,57) =0.788, p = 0.51) or interaction (F(6,114) = 1.61, p = 0.152). Post-hoc tests revealed that decreasing EOA, collapsed across condition, resulted in increased latency of P100 peaks (600: 151.7 ms; 300: 158.5 ms; 0: 164 ms; see Table 3). The use of Ignore All condition to baseline the P100 ERPs when examining the impact of individual event onset asynchronies on peak amplitude and latency was appropriate as an ANOVA revealed no modulation on P100 amplitudes for the Ignore All condition (F(2,38) = 0.44, p = 0.96) nor for the IA latency (F(2,38) = 0.327, p = 0.723). One-way ANOVA's on unbaselined P100 amplitudes showed significant event onset asynchrony effects for all the other analyzed conditions (MT: F(2,38) = 8.26, p = 0.001; TID: F(2,38) = 9.4, p = 0.004; DIT: F(2,38) = 30.4, p = 1.3x10-8). To examine the effects of attentional rivalry and stimulus competition, a 2 X 2 X 3 repeated- measures ANOVA with the factors of Stimulus Relevance (Relevant: MT, DIT; Irrelevant: TID, IA), Attentional Rivalry (Present: DIT, TID; Absent: MT, IA) and EOA (0 ms, 300 ms, and 600 14 ms) revealed no main effect of AR or SR (F(1,19) = 0.009, p = 0.927; F(1,19) = 3.47, p = 0.078) but a main effect of EOA (F(2,19) = 14.9, p = 1.6x10-6). Significant 2-way interaction for AR*EOA (F(2,38) = 11.3, p = 1.4 x 10-4), SR*EOA (F(2,38) = 8.07, p = 1.1x10-3), but not for AR*SR (F(1,19) = 0.001, p = 0.97). The 3-way interaction was not significant AR*SR*EOA (F(2,38) = 0.214, p = 0.81). Exploration of these EOA interactions reveals a modulation pattern dependent on the individual relevance of that stimuli (Stimuli Relevance) at greater event onset asynchronies (600 ms); a modulation pattern that is impacted by the presence of opposing attentional goals (Attentional Rivalry) at event overlap (0 ms) (see Table 2). If the greater multitasking amplitude over the single-tasking amplitudes in the competitive setting (MT0 > DIT0/TID0) was a result of cumulative enhancement of having two relevant events, we might expect to observe (DIT/TID > IA), which we do not (p < 0.218). Additionally, no detectable ERPs for road events were observed in our drive only (DO) condition ruling out an ERP contribution from the tracking task events. If selective enhancement was at work in our competitive setting (EOA 0 ms), we would observe a pattern where DIT > MT or DIT = MT but we actually observe the opposite (MT > SW, p = 0.027). Finally, we observe a significant ANOVA effect for attentional rivalry for 0 ms whereas we observe a significant main effect for stimulus relevance for 600 ms (Table 2A). In other words, in the setting where two events coincide, attentional rivalry (p < 0.029) rather than stimulus relevance (p > 0.97) becomes the determinant in the observed modulation patterns (Table 2). N200: We also assessed the N200 component, another early visual marker that has been associated with the identification or conscious processing of visual stimuli. Using the same analysis approach as for the P100, a main effect of EOA was seen for N200 amplitude (F(2,38) = 6.43, p = 0.0039), but no main effect of condition (F(4,76) = 26.05, p = 0.68) nor interaction (F(2,38)= 1.79, p = 0.18) were present. Post-hoc tests revealed the attenuation of N200 amplitude with more temporally overlapping stimuli (Table 1: 0<300, 0<600, 300=600; 0 ms: -9.24 uV, 300 ms: -11.7 uV, 600 ms: -13.1 uV). N200 latency analyses revealed a main effect of condition (F(1,19) = 10.65, p = 0.0041), but no main effect of EOA (F(2,38) = 1.35, p = 0.27) or interaction (F(2,38) = 2.12, p = 0.13). Post-hoc tests revealed N200 latency delay of multitasking (MT>DIT, Table 1: t(19) = 2.45, p = 0.024; MT: 219.1ms, DIT: 15 207.9ms). Thus, increasing temporal overlap resulted in N200 amplitude attenuation, but no difference between conditions, whereas the addition of active tracking resulted in a non-EOA- dependent N200 latency delay. Eye Movement. To ensure that the observed effects were not due to eye movement, electrooculographic data were analyzed (Anguera et al., 2013). Vertical (VEOG = FP2 – IEOG) and horizontal (HEOG = REOG – LEOG) difference waves were calculated from the raw data and baseline corrected to the mean prestimulus activity. The magnitude of eye movement was computed as follows: (VEOG2 – HEOG2)/2. The variance in the magnitude of eye movement was computed across trials at each time point between 0 and 200 ms post stimulus onset, which encompasses the P100 and N200, the ERP peaks of interest. The variance was compared between EOAs of 0, 300 and 600 via the two-tailed paired t-test. Uncorrected for multiple comparisons, no effects were observed at any time point tested regardless of EOA (p > 0.05), indicating that effects observed in the ERP are not due to eye movements. 16 Legends Table 1. Multitasking and Distracted-Tasking Response Costs for the Multitasking (MT) and Discrimination Ignore Tracking (TID) conditions. The response costs were calculated as individual response time differences for each Event Onset Asynchrony (EOA) by subtracting the Discrimination Only (DO) response time (MT RT – DO RT; DIT RT – DO RT for both conditions respectively). P-values are reported for the individual Event Onset Asynchrony within and between conditions. Table 2. Summary of the 2-way ANOVA analyses for P100 amplitude for the high temporal competition setting (EOA 0 ms) and for the low temporal competition setting (EOA 600 ms). A 2x2 ANOVA explored the effects of Attentional Rivalry (Present Competing Attentional Modulation: DIT, TID; Absent Competing Attentional Modulation: IA, MT) against Stimulus Relevance (Relevant sign events: MT, DIT; Irrelevant sign events: DI, IA). As expected, P100 amplitude for signs shows a main effect for stimulus relevance (MT, DIT > DIT, IA) in the low temporal competition setting. However, under high temporal competition, P100 amplitude is instead driven by the presence of top-down attentional rivalry rather than goal relevance. Post-hoc t-tests demonstrating the interaction across time and the four conditions are shown. Table 3. Event-Related Potentials (ERPs) for sign events during the Multitasking (MT), Sign Ignore Road (DIT), Ignore All (IA), and Tracking Ignore Discrimination (TID) conditions. P-values are reported for ANOVA interactions, main effects and t-tests involving both EOA and condition for P100/N200 latency and amplitude. As seen, only the P100 amplitude demonstrated an interaction of attentional modulation and event onset asynchrony. B. P100 Amplitude Interaction Analysis for all Conditions. P-values are reported for one-way ANOVAs, for each condition. Figure 1. NeuroRacer Tasks and Experimental Design. A. Screenshots of thresholding conditions. The Tracking Only and Discrimination Only conditions were used to establish titrated 'driving' and 'discrimination' levels in the setting of no distraction for each participant. These difficulty levels were used in the subsequent experimental conditions. B. Perceptually-matched Multitasking (MT), Sign Ignore Road (DIT), Tracking Ignore Discrimination (TID) and Ignore All (IA) conditions were specifically interrogated during this experiment. C. Jittered design illustrating the variable EOAs of 0, 300 and 600 ms for the presentation of each sign event with respect to each tracking event were 17 used in all conditions where both stimuli were present (conditions found in 1B). D. Schematic of active and passive task components by condition. Figure 2. Distraction and Multitasking Performance Costs with Event Onset Asynchrony. Means of response times (RT) in the DIT (blue) condition and MT (green) condition across each EOAs (0, 300, 600 ms) with standard error shown. Mean response times were baselined to the Discrimination Only (DO) condition for each participant to gauge distraction (DIT – DO) and multitasking costs (MT – DO). A 2 X 3 repeated-measures ANOVA with factors of Condition (MT, DIT) and EOA (0, 300, 600 ms) revealed a main effect of condition (F(1,19) = 95.7, p < .001), EOA (F(2,38) = 38.7, p < .001), but no interaction (F(1,19) = 1.2, p = 0.33). Follow-up t-tests indicated that RTs of the discrimination task were slower while multitasking versus distracted-tasking (t(19) = 3.17, p = 0.005) and that distracted-tasking was slower than the Discrimination Only condition (t(19) = 7.68, p = 3x10- 7) and that multitasking was slower than the Discrimination Only condition (t(19) = 6.36, p = 4.2x10- 6). Responses took longer for both conditions as the EOA approached complete temporal overlap between tasks (MT0 >MT3 >MT6; DIT0 >DIT3 >DIT6; see Table 1). Error bars represent standard error means. +p < 0.05 between temporal synchronies. Figure 3. Attentional Rivalry and Event Onset Asynchrony Influence on P100 amplitude. A. P100 amplitude means for stimuli in the Discrimination Only (DO) condition, Sign Ignore Road (DIT; blue) condition, Multitasking condition (MT; green), Tracking Ignore Discrimination condition (TID; red) condition and Ignore All (IA; black) condition. P100 amplitudes were baselined against our Ignore All (IA) condition where participants ignored the signs and roads. A one-way ANOVA revealed the conditions to have different P100 modulation (F(4,76) = 10.7, p < 6.5x10-7). B. P100 amplitudes for all four conditions containing EOA (MT, DIT, TID, IA) are baselined against the control condition (IA) for each EOA. A 2 X 2 X 3 repeated-measures ANOVA with factors of for the effect of Stimulus Relevance (SR), Attentional Rivalry (AR) and Event Onset Asynchrony (EOA; 0 ms, 300 ms, and 600 ms) on the selected electrode group of interest revealed a main effect of EOA (F(1,19) = 15.0, p < 1.6x10-5), but no main effect of Rivalry (F(1,19) = 0.55, p = 0.93) or Relevance F(1,19) = 3.5, p < 0.078). 2-way interactions were significant such that AR*EOA (F(2,38) = 11.3, p < 7.3x10-4), SR*EOA (F(2,38) = 8.1, p < 0.004) but not AR*SR (F(2,38) = 0.21, p < 0.81). Pair- wise comparisons of EOA indicated an attenuating effect of decreasing EOA on P100 amplitudes for DIT, such that the closer events were together, the lower the P100 amplitude (DIT0 <DIT300 <DIT600), with similar patterns observed for the multitasking (MT) condition and TID condition 18 (see Table 1). As the plot shows, the interaction is driven by the values at an event onset asynchrony of 0 ms, where TID and DIT exhibited diminished P100 amplitudes compared to baseline. At an EOA of 600 ms, DIT and MT exhibited enhancement of P100 amplitudes compared to baseline. *p < 0.05 between conditions, +p < 0.05 between temporal synchronies. C. Waveforms for ERPs. (i) Waveforms for single-task conditions (SO, TO) and the control condition (IA) demonstrating no ERP elicited from road events (TO). (ii) MT across all three synchronies (0, 300, 600 ms). (iii) ERPs for all four conditions (MT, IA, TID, DIT) at 0 ms. (iv) ERPs for all four conditions at 600 ms. 19 Illustration and Tables Table 1 – Summary of Response Cost analysis (p-values shown) T-TESTS Multitask (MT) Sign Ignore Road 0 vs 300 ms 300 vs 600 ms 0 vs 600 ms (DIT) 7.8x10-5 0.062 1.2x10-6 0.0037 0.0035 1.9x10-5 0 ms 300 ms 600 ms ALL EOAs MT vs DIT 6.7x10-7 4.9x10-9 1.12x10-8 0.005 20 Table 2 – Summary of P100 analysis (p-values shown) A. Stimulus Relevance Attentional Rivalry Stimulus Relevance * Attentional Rivalry ANOVA (p-values shown) 0 ms 0.97 0.029 0.75 600 ms 0.006 0.125 0.74 B. Attentional DIT MT Stimulus Relevance T-test comparisons (p-values shown) Attend 0 ms 600 ms Rivalry Present Absent Attend 0.027 0.297 TID Present Ignore IA MT IA Absent Ignore 0.053 0.352 Absent Attend Absent Ignore 0.854 0.015 DIT Present Attend TID Present Ignore 0.88 0.030 21 Table 3 – Summary of ERP analysis (p-values shown) A. ANOVA EOA Condition EOA * Condition P100 Latency P100 Amplitude N200 Latency N200 Amplitude 0.0054 0.27 0.3 3.44 x 10-8 0.94 0.0006 0.27 0.0041 0.13 0.0039 0.68 0.18 B. P100 Amplitude Interaction Analysis for all Conditions ANOVA EOA MT 1.1x10-3 IA 0.946 TID 4.6x10-4 DIT 1.3x10-8 22 Figure 1 23 Figure 2 24 C. Figure 3 25
1301.6357
2
1301
2013-05-12T06:31:51
Multi-command Tactile and Auditory Brain Computer Interface based on Head Position Stimulation
[ "q-bio.NC", "cs.HC" ]
We study the extent to which vibrotactile stimuli delivered to the head of a subject can serve as a platform for a brain computer interface (BCI) paradigm. Six head positions are used to evoke combined somatosensory and auditory (via the bone conduction effect) brain responses, in order to define a multimodal tactile and auditory brain computer interface (taBCI). Experimental results of subjects performing online taBCI, using stimuli with a moderately fast inter-stimulus interval (ISI), validate the taBCI paradigm, while the feasibility of the concept is illuminated through information transfer rate case studies.
q-bio.NC
q-bio
Multi-command Tactile and Auditory Brain Computer Interface based on Head Position Stimulation H. Mori1, Y. Matsumoto1, Z.R. Struzik2,3, K. Mori4, S. Makino1, D. Mandic5, and T.M. Rutkowski1,2 1Life Science Center of TARA, University of Tsukuba, Japan; 2RIKEN Brain Science Institute, Wako-shi, Japan; 3The University of Tokyo, Tokyo, Japan; 4Research Institute of National Rehabilitation Center for Persons with Disabilities, Tokorozawa, Japan; 5Imperial College London, London, UK Correspondence: T.M. Rutkowski, Life Science Center of TARA, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, Ibaraki, Japan. E-mail: [email protected] Abstract. We study the extent to which vibrotactile stimuli delivered to the head of a subject can serve as a platform for a brain computer interface (BCI) paradigm. Six head positions are used to evoke combined somatosensory and auditory (via the bone conduction effect) brain responses, in order to define a multimodal tactile and auditory brain computer interface (taBCI). Experimental results of subjects performing online taBCI, using stimuli with a moderately fast inter-stimulus interval (ISI), validate the taBCI paradigm, while the feasibility of the concept is illuminated through information transfer rate case studies. Keywords: EEG, P300, somatosensory evoked potentials, auditory evoked potentials, tactile BCI 1. Introduction State of the art BCI relies mostly on mental visual and motor imagery paradigms, which require long training and non-impaired vision of the subjects. Recently, alternative solutions have been proposed to utilize spatial auditory [Halder et al., 2010; Schreuder et al., 2010] or tactile (somatosensory) modalities [Muller-Putz et al., 2006; van der Waal et al., 2012] in order to enhance brain-computer interface comfort, or to boost the information transfer rate (ITR) achieved by users. The concept described in this paper, of utilizing brain somatosensory (tactile) modality, opens up the attractive possibility of targeting the tactile sensory domain, which is not as demanding as vision during the operation of robotic interfaces (wheelchair, prosthetic arm, etc.) or visual computer applications. The first successful trial to utilize somatosensory modality to create a BCI [Muller-Putz et al., 2006] targeted a very low stimulus frequency in a range of 20-31 Hz to elucidate the subject’s attentional modulation of steady-state responses. A very recent report [van der Waal et al., 2012] proposed using a Braille stimulator with a 100 ms long static push stimuli delivered to six fingers to evoke somatosensory response related P300. Very encouraging results were obtained with 7.8 bit/min on average and 27 bit/min for the best subject. Here we propose to combine the two above-mentioned modalities in a taBCI paradigm, which relies on a P300 response evoked by the audio and tactile stimuli delivered simultaneously via the vibrotactile exciters attached to positions on the head, thus benefiting from the bone-conduction effect for audio, which could help ALS-TLS (Amyotrophic Lateral Sclerosis / Total Locked-in State) patients with compromised vision and hearing due to weakened blinking and middle ear effusion/negative pressure, respectively. This offers a viable alternative for individuals lacking somatosensory responses from the fingers. 2. Material and Methods In the experiments reported in this paper, eleven BCI-naïve subjects took part (mean age 21.82 years, with a standard deviation of 0.87). All the experiments were performed at the Life Science Center of TARA, University of Tsukuba, Japan. The psychophysical and online EEG taBCI paradigm experiments were conducted in accordance with the WMA Declaration of Helsinki - Ethical Principles for Medical Research Involving Human Subjects. The subjects of the experiments received a monetary gratification. The 100 ms long stimuli in the form of 350 Hz sinusoidal waves were delivered to areas of the subjects' heads via the tactile exciters HiWave HIAX19C01-8 working in the range of 300-20,000 Hz. The vibrotactile stimulators were arranged as follows. The pairs of exciters were attached on both sides of the forehead, chin, and behind the ears respectively. During the online taBCI experiments the EEG signals were captured with an eight dry electrodes portable wireless EEG amplifier system, g.MOBllab+ & g.SAHARA by g.tec. The electrodes were attached to the following head locations Cz, CPz, P3, P4, C3, C4, CP5, and CP6, as in the 10/10 extended international system (see topographic plot in Fig. 1). The ground and reference electrodes were attached behind the left and right ears respectively. In order to limit electromagnetic interference, the subjects' hands were additionally grounded with armbands connected to the amplifier ground. No electromagnetic interference was observed with the vibrotactile exciters attached to the head. The recorded EEG # e d o r t c e l e 2 4 6 8 # e d o r t c e l e 2 4 6 8 0.56 0.55 0.54 0.53 0.52 0.51 0.5 0.49 0.48 0.47 Averaged non(cid:239)targets [µV] 2 0 (cid:239)2 Targets vs. non(cid:239)targets AUC scores 2 0 (cid:239)2 signals were processed by an in-house enhanced BCI2000 application using a linear discrimination analysis (LDA) classifier with features drawn from 0-600 ms event related potential (ERP) intervals. The sampling rate was set to 256 Hz, high pass filter at 0.1 Hz, and low pass filter at 40 Hz. The ISI was 400 ms and each stimuli length was 100 ms. The subjects were instructed to spell six-digit random sequences of numbers 1-6, which were represented by the exciters in each session. Each target was presented five times in a single spelling trial, and the averages of five ERPs were later used for the classification. AUC at 480 ms Averaged targets [µV] 2 4 6 8 0 600 500 400 200 100 300 time [ms] Figure 1. The averaged ERP responses of the eleven BCI-naïve subjects taking part in the experiment. The left panel presents a topographical plot with the target vs. non-target ERP area under the curve (AUC) for ERP discriminative features extraction [Schreuder et al., 2010] plotted at the maximum value at 480 ms. The right panel presents the averaged ERPs (11 subjects; averaged six-digit sequences spelled in three sessions by each subject) for targets in the top panel; non-targets in the middle; and AUC coefficients at the bottom. The EEG electrode, represented by numbers on the vertical axes is shown in the right panels. The electrode order is as follows Cz, CPz, P3, P4, C3, C4, CP5, CP6. # e d o r t c e l e 0.55 0.5 3. Results and Discussion The averaged ERP responses from the eleven BCI-naïve subjects are presented in Fig. 1 for the target and non- target digits separately, together with the area under the curve (AUC) [Schreuder et al., 2010] discrimination coefficient plots marking the most separable latencies. A topographic plot of the AUC coefficient distributions is also presented in Fig. 1, supporting the choice of the eight dry EEG electrodes covering the parietal cortex. The results of online BCI interfacing sessions are summarized in Table 1 in the form of mean accuracies above the theoretical chance level of 16.6%. In our experiments, only one BCI-naïve subject obtained 100%, and likewise one obtained 0% for the six-digit sequence spelling accuracy with the 5-trials averaging procedure. The preliminary, yet encouraging results presented are a step forward in the search for new BCI paradigms for ALS-TLS patients with compromised vision and hearing symptoms. Table 1. Summary of the online taBCI interfacing results with the 11 naïve subjects (one subject scored 100% accuracy reaching 12.9 bit/min and one 0% with 0 bit/min). Accuracy standard deviation Mean ITR ERP averages Mean accuracy ITR standard deviation 5 50% 27% 3.22 bit/min 3.6 bit/min Acknowledgements This research was supported in part by the Strategic Information and Communications R&D Promotion Programme no. 121803027 of The Ministry of Internal Affairs and Communication in Japan, and by KAKENHI, the Japan Society for the Promotion of Science grant no. 12010738. We also acknowledge the technical support of YAMAHA Sound & IT Development Division in Hamamatsu, Japan. References [Muller-Putz et al., 2006] Muller-Putz, G., Scherer, R., Neuper, C., and Pfurtscheller, G. (2006). Steady-state somatosensory evoked potentials: suitable brain signals for brain-computer interfaces? Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 14(1):30–37. [Halder et al., 2010] Halder, S., Rea, M., Andreoni, R., Nijboer, F., Hammer, E., Kleih, S., Birbaumer, N., and Kübler, A. (2010). An auditory oddball brain–computer interface for binary choices. Clinical Neurophysiology, 121(4):516 – 523. [Schreuder et al., 2010] Schreuder, M., Blankertz, B., and Tangermann, M. (2010). A new auditory multi-class brain-computer interface paradigm: Spatial hearing as an informative cue. PLoS ONE, 5(4):e9813. [van der Waal et al., 2012] van der Waal, M., Severens, M., Geuze, J., and Desain, P. (2012). Introducing the tactile speller: an ERP-based brain– computer interface for communication. Journal of Neural Engineering, 9(4):045002.
1703.06270
3
1703
2017-03-25T13:19:14
SIM-CE: An Advanced Simulink Platform for Studying the Brain of Caenorhabditis elegans
[ "q-bio.NC", "cs.NE", "q-bio.QM", "stat.ML" ]
We introduce SIM-CE, an advanced, user-friendly modeling and simulation environment in Simulink for performing multi-scale behavioral analysis of the nervous system of Caenorhabditis elegans (C. elegans). SIM-CE contains an implementation of the mathematical models of C. elegans's neurons and synapses, in Simulink, which can be easily extended and particularized by the user. The Simulink model is able to capture both complex dynamics of ion channels and additional biophysical detail such as intracellular calcium concentration. We demonstrate the performance of SIM-CE by carrying out neuronal, synaptic and neural-circuit-level behavioral simulations. Such environment enables the user to capture unknown properties of the neural circuits, test hypotheses and determine the origin of many behavioral plasticities exhibited by the worm.
q-bio.NC
q-bio
SIM-CE: An Advanced Simulink Platform for Studying the Brain of Caenorhabditis elegans Ramin M. Hasani 1 Victoria Beneder 2 Magdalena Fuchs 1 David Lung 1 Radu Grosu 1 Abstract We introduce SIM-CE, an advanced, user- friendly modeling and simulation environment in Simulink for performing multi-scale behavioral analysis of the nervous system of Caenorhabditis elegans (C. elegans). SIM-CE contains an imple- mentation of the mathematical models of C. ele- gans's neurons and synapses, in Simulink, which can be easily extended and particularized by the user. The Simulink model is able to capture both complex dynamics of ion channels and additional biophysical detail such as intracellular calcium concentration. We demonstrate the performance of SIM-CE by carrying out neuronal, synaptic and neural-circuit-level behavioral simulations. Such environment enables the user to capture un- known properties of the neural circuits, test hy- potheses and determine the origin of many be- havioral plasticities exhibited by the worm. 1. Introduction C. elegans is most likely the world's best-understood ani- mal (Ardiel & Rankin, 2010). However, the fundamental principles underlying its behavior are yet to be understood. Its relatively simple nervous system is constructed from 302 neurons, hardwired by means of about 8000 chemi- cal and electrical synapses (Varshney et al., 2011). De- spite its simplicity, it has shown remarkable complex be- havioral features which make it an attractive model system. Multi-scale behavioral analyses have been conducted on the worms nervous system (Kato et al., 2015; Roberts et al., 2016), and global attempts on modeling its emergent be- havior have been commenced (Szigeti et al., 2014). In this regard, there is a high demand for a comprehensive neuron- by-neuron model platform which incorporates many elec- trophysiological properties of neurons while avoiding large parameter space (Roberts et al., 2016; Hasani et al., 2016). 1Vienna University of Technology, Austria 2University of Nat- ural Resources and Life Sciences, Vienna, Austria. Correspon- dence to: Ramin M. Hasani <[email protected]>. In the present study, we construct a modular simulation platform, SIM-CE, for investigating the fundamental prin- ciples underlying physiological processes within the neural circuits of the brain of C. elegans, in Simulink. There are three key components in the process of modeling the brain dynamics; 1) The choice of a suitable model for imitating the desired behavior and analysis. 2) Deciding the level of abstraction and biophysical details to be considered. 3) selecting a proper model for the connectivity of the net- work. Accordingly, we design Simulink models of single- compartmental neuron models, as previously developed in (Kuramochi & Iwasaki, 2010), and synapses, and employ them in the implementation of neural circuits. In particular, we exemplify the performance of our modeling platform by creating the tap-withdrawal (TW) neural circuit, a circuit which mediates a reflexive response to mechanical stimu- lation of touch sensory neurons (Wicks & Rankin, 1997). We show that SIM-CE can capture behavioral features of the neural circuits in multi-scale details from biophysics of neurons and synapses. Moreover, SIM-CE can presumably test hypotheses and provide valid predictions. For instance one can define the structure of the wiring diagram, reveal the working principles of numerous unknown neural cir- cuits such as the central pattern generator (CPG), determine the synaptic polarities and find the origin of many associa- tive and non-associative learning modalities (Hasani et al., 2017). We structure the paper as follows. In Section 2, we step- by-step explain the design procedures of SIM-CE plat- form. We recapitulate the mathematical modeling methods and show how to add various ion channel dynamics to the neuron model, and complete the Simulink design in Sec- tion 2.1. We also illustrate how to design the connectiv- ity among the neurons in Section 2.2. We then, in Section 2.3, demonstrate the capability of the designed platform to construct neural circuits and correspondingly simulate their dynamics. We provide a useful discussion on the attributes of SIM-CE platform in Section 3 and conclude our work in Section 4. SIM-CE: The Simulink C. elegans 2. Methods In this section we introduce the methodologies utilized for implementation of Simulink models of the neurons and synapses. We revisit the single-compartmental neu- ron modeling techniques initially introduced by Hodgkin & Huxley (1952) and Kuramochi & Iwasaki (2010). The model comprises several ordinary differential equations (ODE) describing dynamics of the cell. ODEs are imple- mented by using the integrator block in Simulink. For in- stance, the equation dV dt = −60 − V , is constructed as shown in Figure 1. The input to the integrator block is the derivative term dv/dt. The output is V (t), which is multiplied by −1, summed up with the constant and is fed back to the integrator's input. The initial condition of the variable V is defined inside the integrator block. The equation is then solved by using the ode45 solver of MATLAB which realizes the Dolman-Prince (Dormand & Prince, 1980), a numerical method for solving ODEs. Figure1. ODE implementation in Simulink 2.1. Neuron Model The overall current passing through the membrane of a neu- ron is maintained by the sum of inward and outward cur- rents passing through it. Such currents are generated as a result of propagation of ions across the cell membrane. Fig- ure 2 shows the membrane of a C. elegans neuron where various channels such as voltage-gated potassium chan- nels, calcium-gated potassium channels, voltage-gated cal- cium channels and leakage channels together with a cal- cium pump establish the membrane potential of the neuron. Accordingly, the dynamics of the membrane potential can be written in the form of the following differential equa- tion: Cm dV dt = −(ICa + IK + Isk + ILeak)+ Σ(ISyn + Igap + Istimuli), (1) where Cm corresponds to the membrane capacitance and V is the membrane potential. ICa, IK, Isk and ILeak rep- Figure2.Representation of a single neuron of C. elegans (Paint- ing by: suppressed for anonymity, reproduced from (Kuramochi & Iwasaki, 2010)). resent the calcium current, potassium current, intercellular calcium-gated potassium channel current and leakage cur- rent respectively (Kuramochi & Iwasaki, 2010). Istimuli, Isyn and Igap are currents which are injected into the neuron from a stimulus to a sensory neuron, a chemical synapse and an electrical synapse (gap junction) respec- tively. For modeling the ion conductance channels we use the clas- sic Hodgkin-Huxley formalism (Hodgkin & Huxley, 1952) where the conductance of a channel, Gion, is defined by the product of the maximum conductance, the probability of activation function, m, and inactivation function, h, of the channel as follows: Gion = Gmaxmphq. (2) For the calcium and potassium channels we only consider the activation variables as proposed by Hodgkin & Huxley (1952) and Engel et al. (1999). We define α and β, the vari- ables of the gate-rate-functions m for each channel by an- alyzing the steady-state of the function, m∞, and the time constant, τm, of the gate relaxation (Hodgkin & Huxley, 1952). m∞ = αm/(αm + βm) τm = 1/(αm + βm) (3a) (3b) Once we determined the gate-rate-functions for each chan- nel, we describe the straightforward way of implementing each channel and consequently the neuron itself, as below. x (-1) Product −60 Constant 1 s V(t) Integrator + + Sum dV/dt The Model SK Channel Leak Channel Stimulus 𝐾+ Channel 𝐼6 Ca2+ PumpCa2+ Channel 𝐼89𝐼:9;𝐼<=> Presynaptic Neuron Chemical Synapse Ca23 𝐼76 𝐼?@9A 𝐼7BCDEFC 𝐶𝑎2++𝐵 ⇌𝐶𝑎𝐵 𝑘𝑓𝑘𝑏 Calcium Buffering Postsynaptic Neuron SIM-CE: The Simulink C. elegans calcium concentration (considered constant), respectively (Kuramochi & Iwasaki, 2010). Later on, we employ three different mechanisms in order to set the value of the con- centration of calcium inside the neuron, [Ca2+]in. As we discussed before, the dynamics of the voltage depen- dent gating variable, mCa, can be derived from equations 3a and 3b: dmCa = αCa(V )(1 − mCa) − βCa(V )mCa, (5a) dt where and αCa(V ) = −0.5(V + 20) − 1 βCa(V ) = 0.23e −(v+29) e −(V +20) 10 7 . f (5b) (5c) We construct the behavior of mCa based on equation 5a in Simulink as shown in Figure 3A. The variation of the acti- vation function of the calcium conductance as a function of the membrane potential is plotted in Figure 3B. 2.1.2. POTASSIUM CURRENT In almost all animal cells, potassium channels are respon- sible for setting the resting potential and bringing a neuron from an excited state back to its equilibrium state (Katz, 1966). The concentration of potassium ions inside the cell is higher than outside, therefore, potassium tends to flow outward. The current generated by the outflow of potas- sium ions, can be abstracted as follows: IK = GKn4 K(V − EK), (6) where GK and EK are the maximum potassium channel conductance and potassium equilibrium potential, respec- tively. The kinetics of the potassium gating variable, nK, is given by Equation 7a, which is directly derived from Equa- tions 3a and 3b (Hodgkin & Huxley, 1952): dnK dt = αK(V )(1 − nK) − βK(V )nK. (7a) where αK(V ), the gating variable which is changing with the presynaptic membrane potential, is calculated as: αK(V ) = 0.6e v+40 20 and βK(V ) is computed as follows: βK(V ) = 0.01(V + 12) 23 − 1 V +12 e (7b) (7c) . We implement the described potassium-conductance acti- vation function in Simulink as illustrated in Figure 4A and show its dynamics as a function of the membrane poten- tial in Figure 4B. Note that the activation of the potassium channel is set to be slightly slower than that of the calcium channels. Figure3.Calcium gate-rate-function. A) Simulink implementa- tion of calcium gate-rate-function. B) Activation function of the calcium conductance. 2.1.1. CALCIUM CURRENT Calcium ions are key elements in signalling in the ner- vous systems. They are responsible for several functions such as synaptic vesicle release, activation of ionic chan- nels and excitatory currents across the membrane of neu- rons (Hobert, 2013). This is notable since C. elegans lacks voltage-gated sodium channels and hence it does not gener- ate sodium-based action potentials (Goodman et al., 1998). The current passing through the voltage-sensitive calcium channel can be derived from the Goldman-Hodgkin-Katz equation (Engel et al., 1999) as follows: RT CaPCa nCaF V ICa = m2 [Ca2+]in − [Ca2+]out e −2F V , (4) where PCa, nCa, F , R, T and [Ca2+]out are calcium permeability, calcium charge valance, Faraday constant, gas constant, temperature (in Kelvin) and extracellular 1 − e −2F V RT RT 0 -100 -50 0 Vm (mV) 50 Ca Ca B 1 0.5 mCa A SIM-CE: The Simulink C. elegans 2.1.3. SMALL-CONDUCTANCE POTASSIUM CHANNEL'S CURRENT As regulators, these channels significantly influence the excitability of the neurons (Salkoff et al., 2005). Small- conductance calcium-activated potassium channels are also included as it is shown in Equation 8. IsK = GsKmsK(V − EsK), (8) where GsK and EsK show the maximum channel conduc- tance and the corresponding Nernst potential, respectively. Activation of these channels is purely controlled by the in- tracellular calcium concentration in the form of the follow- ing equation (Engel et al., 1999): dmsK dt = [Ca2+]in ψsK (1 − msK) − KsK ψsK msK, (9) where ψsK as the time constant rate and KsK as the potas- sium concentration are used in order to calculate the gated variable rate functions α and β, in a classic Hudgkin- Huxley scheme (Kuramochi & Iwasaki, 2010). We de- sign the Simulink model of the small-conductance potas- sium channel as it is depicted in Figure 5A and show its Figure5.Small-conductance potassium channel (SK) gate-rate- function. A) Simulink model of the SK B) Activation function of the potassium conductance as a function of the intracellular calcium concentration. dynamics as a function of the inner calcium concentration in Figure 5. 2.1.4. LEAKAGE CURRENT Leakage channels are the mediators of the cells where ions can freely move along them. For a neuron we model the leakage current passing through a leakage channel with a constant conductance, GLeak, and an average Nernst po- tential of ELeak, as follows: ILeak = GLeak(V − ELeak). (10) Figure 6 shows the Simulink model of the leakage channel. Figure4.Potassium channel gate-rate-function. A) Simulink im- plementation B) Activation function of the potassium conduc- tance. Figure6. Simulink model of the leakage channel 0 -100 0 -50 Vm (mV) 50 K K B 1 0.5 mK A 5 [Ca2+]in ( M) 10 0 0 0.5 msk A B 1 SK SK SIM-CE: The Simulink C. elegans Figure7.Simulink model of the inner calcium concentration 2.1.5. INTRACELLULAR CALCIUM CONCENTRATION DYNAMICS During the last years the use of calcium imaging resulted in significant advancements in the understanding of the phys- iology of neurons and neural circuits of C. elegans (Kato et al., 2015). The dynamics of the intracellular calcium concentration for a neuron represents the dynamics of the whole cell. There are several mechanisms involved in de- termining the inner calcium level; Inflow of the calcium current through voltage-gated calcium channels is the main source of an increase in its level. Calcium ions inside the cell bind to specific types of proteins called binding pro- teins. Such chemical process varies the amount of freely available intracellular calcium ions (Kuramochi & Iwasaki, 2010). Moreover, a calcium pump gets activated when the amount of intracellular calcium exceeds a certain thresh- Figure9.Simulation of the deterministic response of the neuron to an input current stimulus. old and transports calcium out of the cell. In C. elegans, calcium currents play the excitatory role that sodium cur- rents have in mammalian cells. Dynamics of the inner cal- cium concentration based on the described mechanisms is derived in the Equation 11a where d is the calcium diffu- sion length, kb, kf are the coefficients of the bidirectional chemical process of the binding proteins [B], with calcium ions which create the molecule, [CaB] and Gpump is the conductance of the pump (Kuramochi & Iwasaki, 2010). Equation 11b quantitatively represents the dynamics of the chemical process of the calcium ions and binding proteins (Kuramochi & Iwasaki, 2010). d[Ca2+]in dt 1 2F d ICa + kb[B][CaB] = − − kf [Ca2+]in[B](1 − [CaB]) − [Ca2+]in + Kpump Gpump[Ca2+]in (11a) Figure8. Simulink model of a C. elegans neuron d[CaB] dt = −kb[CaB] + kf [Ca2+]in(1 − [CaB]) (11b) Intracellular calcium concentration kinetics is designed as A B C D SIM-CE: The Simulink C. elegans a part of the neuron model and shown in Figure 7. All the building blocks of the neuron model shown in Fig- ure 2 are designed and put together, forming a single neu- ron block. Figure 8 represents the schematic of such de- tailed neuron model where all the parameters are conve- niently adjustable. The model enables the user to input external stimuli from the environment or from a presynap- tic synapse in real time while monitoring the physiological behavior of the membrane potential, intracellular calcium concentration and the ion channel currents. Note that there are several types of ion channels determined in the neurons of C. elegans (Salkoff et al., 2005). One can easily im- plement models of any arbitrarily chosen ion channel and include it in the neuron dynamics. Table 1 summarizes the parameters of the neuron and synapse model together with the range of their values. Table1.Parameter-space of the model of neuron and synapse Figure10.Simulation of the stochastic response of the neuron to an intrinsic random current-pulse generator. A) The simulation setup. B) Membrane potential C) Calcium concentration. PARAMETER VALUE-RANGE PARAMETER VALUE-RANGE havior of the cells and neural circuits (Roberts et al., 2016). Cm PCa nCa F R T [Ca2+]out GK EK GsK EsK ψsK KsK J J M.V µM.mV.cm2 15 − 100 µF cm2 nA 1 − 3 2 9.6 × 104 8.3 M.°K 293.1 °K 2 − 2.4 mM 1 − 15 mS/cm2 −95 to − 60 mV 0.2 − 0.3 mS/cm2 −95 to − 60 mV 2.5 − 3 µM.ms 0.1 − 0.9 µM GLeak ELeak d kb [B] kf Kpump Gpump Esyn Gsyn Vshif t Vrange Ggap 1 ms.µM 0.04 − 0.1 mS cm2 −74mV 1 − 4 µm 0.3 1 ms 15 − 30 µM 0.1 0.3 − 1 µM 1 − 10 µM −90 − 0 mV 0.1 − 1 mS cm2 −10 to − 40 mV 3 − 6 mV 0.1 − 1 mS cm2 ms We now simulate a single neuron and observe its response to an external stimulus. Figure 9 demonstrates the deter- ministic response of the designed neuron to an external in- put stimulus shown in Figure 9B. The membrane potential response together with the intracellular calcium concentra- tion is plotted in Figure 9C and 9D, respectively. Calcium dynamics follows the excitability of the membrane poten- tial demonstrating how calcium imaging can be representa- tive of the dynamics of the whole cell. In addition to reproducing the deterministic behavior of the neuron, we include stochastic properties to the system. We apply a Gaussian white noise on every ion channel con- ductance and also consider the intrinsic random activation of the neuron by stimulating the neuron with a random current-pulse generator. The results of a transient simula- tion are shown in Figure 10 where the membrane potential together with the calcium concentration of a single neuron is plotted in Figure 10B and 10C, respectively. The calcium dynamics is similar to that of biological neurons captured in calcium imaging experiments with a reasonable degree of accuracy (Kato et al., 2015). In this way, the model addi- tionally allows the user to study the intrinsic stochastic be- 2.2. Wiring Synapses are the points of communication of neurons. In- formation transfer happens as a result of neurotransmitters release at a chemical synaptic port of a neuron. There are two major types of synapses, chemical synapses and elec- trical synapses (gap junctions). Chemical synapses are the elements at which the signaling from a presynaptic neuron to a postsynaptic cell occurs through secretion of neuro- transmitters. Depending on the type of neurotransmitters Figure11.Simulink model of synapses. A) Chemical excitatory and inhibitory synapse model B) Model of the gap junction. 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 Time (ms) 500 -40 -50 -60 Vm(mV) B 0.05 0.03 0.01 0 [Ca2+]in(mM) C A By changing 𝐸𝑐ℎ𝑒𝑚 we can set a synapse to be excitatory or inhibitory A Excitatory Synapse Inhibitory Synapse B Gap Junction SIM-CE: The Simulink C. elegans Figure12. Tap-withdrawal neural circuit implemented in the Simulink platform. available at the presynaptic neuron, a chemical synapse can be excitatory or inhibitory for the postsynaptic neuron. Dy- namics of the chemical synapses depends on several fac- tors, such as the concentration of the available neurotrans- mitters, the state of the activation/inactivation of the pre- vious neuron, the probability of the binding of neurotrans- mitters to the postsynaptic receptors and the amount of the receptors available at the postsynaptic cell (Schutter, 2009). A static mathematical representation of such synapses can be formulated as follows: ISyn = nij GSyn −(Vpre−Vshif t ) Vrange 1 + e (ESyn − Vpost), (12) where, nij is the number of synaptic connections from neu- ron i to neuron j, Gsyn is the maximum conductance of the synapse, Vshif t and Vrange stand for the equilibrium po- tential and the steepness of the synaptic activation, respec- tively (Koch & Segev, 1998). ESyn, represents reversal potential of the synapse where its variations results in exci- tatory or inhibitory behavior of the synapse. In our model it is set to zero for an excitatory synapse and to −90mV for an inhibitory synapse. Figure 11A represents the designed Simulink model of chemical synapses. Electrical synapses (gap junctions) can be modeled simply by employing Ohm's law as follows: Igap = ngapGgap(Vpre − Vpost), (13) where ngap and Ggap, represent the number gap junctions between two neurons and the gap junction conductance, re- spectively. Figure 11B depicts the implementation of the gap junction in Simulink. 2.3. Neural Circuits Implementation In this section we illustrate how our proposed Simulink platform can enable users to construct arbitrarily chosen neural circuits from the nervous system of the worm and simulate its behavior while fully observing the dynamics of the circuit in neuronal and synaptic level. Pulse Generator1 Signal Generator1 IinPLM Signal Generator Pulse Generator3 Iinput 0 Ichem Constant12 Igap PLM V Cain ICa IK V-PLM Isk Ileak Pulse Generator2 0 Constant10 Vpre Vpost n Ichem 4 50 Vpre Vpost n Ichem 2-1 25 Vpre Vpost n Igap 2 120 Iinput Ichem Igap V Cain ICa IK Isk Ileak ALM Vpre Vpost n Igap 2-2 120 Vpre Vpost n Ichem 5 62.5 Constant2 Constant7 Constant1 Constant3 Constant13 V-AVM V Cain ICa IK Isk Ileak Iinput 0 Ichem Constant11 Igap AVM V-Sensory IinAVM V-Interneuron Iin-PVC YAVA YAVB Y Cain Y-AVD Y Cain Y-PVC Y Cain Y-AVA Y Cain Y-AVB Vpre Vpost n Ichem 6 75 Vpre Vpost n Ichem 1 12 Vpre Vpost n Igap 2-3 120 Constant14 Constant15 Constant16 Vpre Vpost n Igap 2-4 120 Vpre Vpost n Ichem 8 100 Vpre Vpost n Ichem 10 125 Constant17 Constant18 Constant19 0 Iinput Constant6 Ichem Igap V Cain ICa IK V-PVC Isk Ileak PVC Vpre Vpost n Ichem 28 700 V Cain ICa IK Isk Ileak V-LUA 0 Iinput Constant4 Ichem Igap LUA V-AVD 0 Iinput Constant5 Ichem Igap AVD V Cain ICa IK Isk Ileak Vpre Vpost n Ichem 17 425 Vpre Vpost n Igap 10-2 600 Vpre Vpost n Igap Vpre Vpost n Ichem 2-5 120 1-2 25 Vpre Vpost n Ichem 13 325 Constant20 Constant21 Constant22 Constant23 Constant24 Constant25 Iin-AVD Vpre Vpost n Igap 2-6 120 Vpre Vpost n Ichem 3 37.5 Vpre Vpost n Ichem 20 250 Vpre Vpost n Igap 1-3 60 Vpre Vpost n Ichem 10-3 125 Constant26 Constant27 Constant28 Constant29 Constant30 0 Iinput Constant8 Ichem Igap AVA V Cain ICa IK Isk Ileak V Cain ICa IK Isk Ileak 0 Iinput Constant9 Ichem Igap AVB Vpre Vpost n Ichem 2-i 25 Vpre Vpost n Ichem Vpre Vpost n Ichem Vpre Vpost n Ichem 2-ii 25 4-i 50 5-i 62 Vpre Vpost n Igap 1-9 60 Vpre Vpost n Igap 10-3 600 Constant36 Constant37 Constant38 Constant39 Constant40 Constant41 IinAVA VAVA IinAVB VAVB CainAVA CainAVB Vpre Vpost n Ichem 6-i 75 Vpre Vpost n Ichem 3-i 37.5 Constant42 Constant43 Vpre Vpost n Ichem 56 1400 Vpre Vpost n Ichem 1-4 25 Vpre Vpost n Ichem 1-1 25 Vpre Vpost n Ichem Vpre Vpost n Igap 3-1 75 2-7 120 Constant31 Constant32 Constant33 Constant34 Constant35 SIM-CE: The Simulink C. elegans An example of a well-understood neural circuit within the nervous system of C. elegans, is the tap-withdrawal (TW) circuit. The circuit modulates a reflexive response to a me- chanical stimulus subjecting the petri dish in which the worm crawls. TW circuit comprises 8 neurons optimally hardwired through almost 360 synapses. A mechanical stimulus excites specific sensory neurons (PLM, AVM or ALM) and results in the activation of the corresponding command neurons (AVA or AVB) through a group of in- terneurons (PVC, LUA, AVD). AVA activates the motor neuron responsible for reversal movement while AVB com- mand neuron initiates forward locomotion. Depending on the strength of the stimulation on the sensory neurons, AVA or AVB gets activated and results in a reflexive motion (Wicks & Rankin, 1997). Here, we construct the model of the neural circuit in our platform. Figure 12, represents the designed TW circuit. The architecture of the circuit is taken from the initial C. elegans connectome data (White et al., 1986). The polarity of the synaptic connections is set based on the evaluations provided in (Wicks & Rankin, 1997). We then perform a simulated experiment in which we stim- ulate the AVM sensory neuron and show how the com- mand neurons behave accordingly. Figure 13 shows the behavior of the command neurons AVA and AVB to the de- scribed stimulation. Figure 13B and 13C, show the mem- brane potential of the command neurons and Figure 13D and 13E depict their intracellular calcium concentrations. We observe that the response of the model correctly imi- tates the expected behavior where an anterior stimulation of the worm results in the activation of the reversal com- mand neuron and therefore initiation of a backward reflex- ive movement. Note that the forward-movement command neuron, AVB, is suppressed due to the dynamics of the cir- cuit. 3. Discussion Over the past years a considerable number of behavioral analyses on the nervous system of C. elegans have been conducted. However, the origin of many behavioral plas- ticities have not yet been fully determined. SIM-CE plat- form, provides researchers with the opportunity to explore beyond the observable dynamics in the brain of the worm where multi-scale details of the neural circuits can be con- structed and decoded, precisely in a user-friendly fash- ion. SIM-CE permits us to study attractive behavioral fea- tures of C. elegans such as gene modification effects, non- associative and associative learning in various levels by tuning parameters (?). Moreover, one can perform numer- ous simulated experiments within the platform and provide multi-scale predictions on the physiology of the neural cir- cuits. Figure13.Responses of the tap-withdrawal neural circuit A) In- puts to the sensory neurons B) Output of the AVA command neu- ron C) Output of the AVB comannad neuron D) Intracellular cal- cium concentration of the AVA neuron E) Intracellular calcium concentration of the AVB neuron The top-level block diagram representation in SIM-CE considerably simplifies the investigation of basic behav- ioral primitives within the C.elegans neural circuits. More- over, SIM-CE is a suitable base for performing large-scale block level simulations. This enables our platform to be flexible regarding simulation of large-scale networks (ide- ally the entire nervous system). However, simulation of such networks would essentially require much hardware computational power. 4. Conclusions We introduced SIM-CE, a Simulink-based modeling and simulation platform for investigating the nervous system of C. elegans. The platform offers detailed mathematical models of the neurons and synapses and enables the user to conveniently perform simulated experiments in neuronal, synaptic and neural circuit level. We comprehensively ex- plained the design-flow of single neurons and synapses. We then exemplified the performance of the platform by per- forming simulated experiments of single neurons in deter- ministic and stochastic modes, and a neural circuit. For future work, we aim to explore the dynamics of more neural circuits in multi-scale physiological conditions, pro- vide valuable predictions about the unknown properties of such circuits and ultimately include the dynamics of the en- tire nervous system of the worm. A B D C E SIM-CE: The Simulink C. elegans Eduardo J, Faumont, Serge, Lindsay, Rebecca A, Brit- ton, Matthew Cale, Pokala, Navin, et al. A stochas- tic neuronal model predicts random search behaviors at multiple spatial scales in c. elegans. Elife, 5:e12572, 2016. Salkoff, L, Wei, AD, Baban, Beravan, Butler, Alice, Fawcett, G, Ferreira, Gonzalo, and Santi, Celia M. Potassium channels in c. elegans. 2005. Schutter, Erik De. Computational modeling methods for neuroscientists. The MIT Press, 2009. Szigeti, Bal´azs, Gleeson, Padraig, Vella, Michael, Khayrulin, Sergey, Palyanov, Andrey, Hokanson, Jim, Currie, Michael, Cantarelli, Matteo, Idili, Giovanni, and Larson, Stephen. Openworm: an open-science approach to modeling caenorhabditis elegans. Frontiers in compu- tational neuroscience, 8:137, 2014. Varshney, Lav R, Chen, Beth L, Paniagua, Eric, Hall, David H, and Chklovskii, Dmitri B. Structural properties of the caenorhabditis elegans neuronal network. PLoS Comput Biol, 7(2):e1001066, 2011. White, JG, Southgate, E, Thomson, JN, and Brenner, S. The structure of the nervous system of the nematode caenorhabditis elegans: the mind of a worm. Phil. Trans. R. Soc. Lond, 314:1–340, 1986. Wicks, Stephen R and Rankin, Catharine H. Effects of tap withdrawal response habituation on other withdrawal be- haviors: the localization of habituation in the nematode caenorhabditis elegans. Behavioral neuroscience, 111 (2):342, 1997. References Ardiel, Evan L and Rankin, Catharine H. An elegant mind: learning and memory in caenorhabditis elegans. Learn- ing & Memory, 17(4):191–201, 2010. Dormand, John R and Prince, Peter J. A family of embed- ded runge-kutta formulae. Journal of computational and applied mathematics, 6(1):19–26, 1980. Engel, Jutta, Schultens, Howard A, and Schild, Detlev. Small conductance potassium channels cause an activity-dependent spike frequency adaptation and make the transfer function of neurons logarithmic. Biophysical journal, 76(3):1310–1319, 1999. Goodman, Miriam B, Hall, David H, Avery, Leon, and Lockery, Shawn R. Active currents regulate sensitivity and dynamic range in c. elegans neurons. Neuron, 20(4): 763–772, 1998. Hasani, Ramin M, Esterle, Lukas, and Grosu, Radu. In- vestigations on the nervous system of caenorhabditis el- egans. Current AI Research in Austria (CAIRA) Work- shop, 39th German Conference on Artificial Intelligence (KI2016), 2016. Hasani, Ramin M, Fuchs, Magdalena, Beneder, Victoria, and Grosu, Radu. Non-associative learning representa- tion in the nervous system of the nematode caenorhabdi- tis elegans. arXiv preprint arXiv:1703.06264, 2017. Hobert, Oliver. The neuronal genome of caenorhabditis elegans. WormBook, 2013:1–106, 2013. Hodgkin, Alan L and Huxley, Andrew F. A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of phys- iology, 117(4):500, 1952. Kato, Saul, Kaplan, Harris S, Schrodel, Tina, Skora, Su- sanne, Lindsay, Theodore H, Yemini, Eviatar, Lockery, Shawn, and Zimmer, Manuel. Global brain dynamics embed the motor command sequence of caenorhabditis elegans. Cell, 163(3):656–669, 2015. Katz, Bernard. Nerve, muscle, and synapse. 1966. Koch, Christof and Segev, Idan. Methods in neuronal mod- eling: from ions to networks. MIT press, 1998. Kuramochi, Masahiro and Iwasaki, Yuishi. Quantitative modeling of neuronal dynamics in c. elegans. In Neural Information Processing. Theory and Algorithms, pp. 17– 24. Springer, 2010. Roberts, William M, Augustine, Steven B, Lawton, Kristy J, Lindsay, Theodore H, Thiele, Tod R, Izquierdo,
1507.07270
1
1507
2015-07-27T00:39:34
Searching for behavioral homologies: Shared generative rules for expansion and narrowing down of the locomotor repertoire in Arthropods and Vertebrates
[ "q-bio.NC" ]
We use immobility as an origin and reference for the measurement of locomotor behavior; speed, the direction of walking and the direction of facing as the three degrees of freedom shaping fly locomotor behavior, and cocaine as the parameter inducing a progressive transition in and out of immobility. In this way we expose and quantify the generative rules that shape fruit fly locomotor behavior, which consist of a gradual narrowing down of the fly's locomotor freedom of movement during the transition into immobility and a precisely opposite expansion of freedom during the transition from immobility to normal behavior. The same generative rules of narrowing down and expansion apply to vertebrate behavior in a variety of contexts, Recent claims for deep homology between the vertebrate basal ganglia and the arthropod central complex, and neurochemical processes explaining the expansion of locomotor behavior in vertebrates could guide the search for equivalent neurochemical processes that mediate locomotor narrowing down and expansion in arthropods. We argue that a methodology for isolating relevant measures and quantifying generative rules having a potential for discovering candidate behavioral homologies is already available and we specify some of its essential features.
q-bio.NC
q-bio
Searching for behavioral homologies: Shared generative rules for expansion and narrowing down of the locomotor repertoire in Arthropods and Vertebrates Alex Gomez-Marina1, Efrat Oronb, Anna Gakamskyb, Dan Valentec, Yoav Benjaminid and Ilan Golanib1 aChampalimaud Neuroscience Programme, Lisbon, Portugal; bDepartment of Zoology, Faculty of Life Sciences and Sagol School of Neuroscience, Tel Aviv University, Israel; cCold Spring Harbor laboratory, USA; dDepartment of Statistics and Sagol School of Neuroscience, Tel Aviv University. 1To whom correspondence should be addressed: [email protected], [email protected] Abstract (Running title: A common mobility gradient across phyla) We use immobility as an origin and reference for the measurement of locomotor behavior; speed, the direction of walking and the direction of facing as the three degrees of freedom shaping fly locomotor behavior, and cocaine as the parameter inducing a progressive transition in and out of immobility. In this way we expose and quantify the generative rules that shape fruit fly locomotor behavior, which consist of a gradual narrowing down of the fly's locomotor freedom of movement during the transition into immobility and a precisely opposite expansion of freedom during the transition from immobility to normal behavior. The same generative rules of narrowing down and expansion apply to vertebrate behavior in a variety of contexts, Recent claims for deep homology between the vertebrate basal ganglia and the arthropod central complex, and neurochemical processes explaining the expansion of locomotor behavior in vertebrates could guide the search for equivalent neurochemical processes that mediate locomotor narrowing down and expansion in arthropods. We argue that a methodology for isolating relevant measures and quantifying generative rules having a potential for discovering candidate behavioral homologies is already available and we specify some of its essential features. Significance Statement Formulating the generative rules that shape behavior across species and phyla is an essential step in the establishment of a comparative study of behavior. Invariant generative rules provide an architectural plan (bauplan) that guides research and defines and validates the intrinsic constituents of that bauplan. Examining whether the same rules shape arthropod and vertebrate locomotor repertoire requires solutions to questions such as whether a homology should be based first and foremost on structure or on function, on generative rules or on 1 common descent or on both, and how can the invariant core be distinguished from the adaptive envelope of behavior. The significance of this study lies in the new solutions it offers to these old questions: defining an intrinsic origin and measuring the transition from simple to complex and from complex to simple behavior; using actively managed kinematic quantities; suspending judgment about common descent; using the generative rules as a heuristic search image for the discovery of equivalent behaviors and the neural processes that mediate them; and only then offering the established behavioral bauplan as candidate for the status of a Darwinian homology, based on demonstrated common descent. Introduction The establishment of homologies is an indispensable goal in evolutionary biology. In pre- Darwinian comparative anatomy, a homologue has been defined as "The same organ in different animals under every variety of form and function"1. Based on this definition, anatomists compared skeletons using validated distinctions like a forelimb, a humerus, a radius and an ulna, and compared brains using validated structures such as thalamus, cortex and striatum. These structures gained their identity and validity as homologues by demonstrating that they occupied the same relative position, and had the same connectivity across a wide array of taxonomic groups2 and structures sharing the same name and the same morphogenetic history. The validity of these structures has been indispensable for establishing a rigorous science of anatomy. To the same extent, the comparative study of behavior requires the identification of distinct elementary processes, much like skeletal segments and neural structures, which would be established across a wide variety of taxonomic groups on the basis of their connectivity3, and their moment-to-moment generative history. In the present study we analyze the morphogenesis of cocaine-induced fruit fly behavior, implementing a strategy and tools that aim at exposing generative rules having the potential of becoming universal right from the very first description of a newly studied behavior. The accumulation of detailed descriptions of arthropod and vertebrate movement makes the issue of shared principles of organization in the behavior of these taxonomic groups increasingly accessible for comparison. An opportunity for such comparison is offered by the report that, when treated with the dopamine reuptake inhibitor cocaine, Drosophila melanogaster performs a sequence of stereotyped behavior patterns leading in and out of immobility including locomotion and circling, apparently similar to the sequence observed in rodents4,5. This led the researchers who discovered the phenomenon to suggest that the behavior was homologous in the two phyla. That 2 same behavior has been, however, portrayed as aberrant, unusual, and uncontrolled6 by other researchers, who used a description that highlighted impairment in the functionality of the behavior. Here we analyze the structure of the very same behavior both in its own right, but also using a strategy and tools that aim at obtaining a description that will entice a cross phyletic comparison. Using our structural analysis we derive from the fly's seemingly aberrant behavior the generative rules that shape a substantial component of both arthropod and vertebrate locomotor behavior. Following comparative anatomy we term the set of invariant relations or generative rules that we look for, the architectural plan, or bauplan7-10 of the behavior. Given that a rigorous comparative study of behavior requires the establishment of behavioral homologies, how is it that cross phyletic behavioral homologies were hardly documented? In this study we suggest that the seed for obtaining a cross phyletic perspective (or for missing it) is sown in the initial measurement phase. The choices that are made at that stage reflect age old controversies on whether to i) define homologies on the basis of generative rules11,12, common descent13,14, or both15, ii) measure the function or the structure of behavior16, iii) focus on the behavioral level first or mix behavioral, neural and genetic levels from the start15, iv) use any reference frame or use intrinsic frames of reference for measuring the behavior17, v) suspend or even ignore judgement about common descent17,18, and vi) more generally distinguish between the invariant skeleton of a behavior and its adaptive envelope19. All these choices are part and parcel of the process of observation, and are unavoidable; they are made, explicitly or implicitly, by each and every observer of behavior, determining from the start the potential for the universality of the obtained description. Since they are inescapable, they might as well be made deliberately, on the basis of one's aims. Our aim, of establishing a rigorous comparative study of behavior, requires the discovery of the behavioral bauplan that will have the potential of becoming a bona fide behavioral homology. The choices we made are derived from this aim (implying that other aims might justify other choices). One strategy in deciphering the organization of anatomical structure is following the process of its morphogenetic differentiation from inception to full blown form. The transition from simple to complex provides a view that is not available by studying the final product. We therefore focus on studying fly behavior in a situation involving differentiation from simple to complex, and decay from complex to simple, teasing apart in this way the particulate processes that add on top of each other to compose full blown behavior17,20 (or are eliminated one after the other to full decay). 3 While any selection of variables might be useful (informative), only a selection of the key variables that are actually managed actively by the fly has the potential of defining a behavioral bauplan that will subsequently prove to be cross phyletic. Conversely, variables or kinematic quantities that prove universal across phyla are more likely to represent perceptual quantities that are actually managed by the brain. A judicious selection of the key variables that describe the behavior is therefore necessary for discovering cross phyletic homologies. Recent claims for deep homology between the arthropod central complex and the vertebrate basal ganglia21 provide an opportunity to examine whether the bauplan we discovered can be supplemented with a historical perspective. If cocaine induced behavior, which is mediated by these centers, is the same in the two phyla, then the claim for a behavioral homology would be supported by both a generative claim for a common bauplan and a historical claim for common descent. Moreover, the shared generative rules of the behavior can provide a specification of the demand17,22 on the neural activity and network connectivity within and between substructures of the central complex and the basal ganglia. For example, the expansion of a vertebrate's locomotor repertoire, which has been recently attributed to dopaminergic feedforward loops operating in the basal ganglia23,24, can guide a study of the relations between the arthropod transition out of immobility and the arthropod's central complex. Results Behavior in and out of immobility: narrowing down of the path's spatial spread and its build up to normal behavior. Figure 1 presents the path traced by the center of mass of a single fly walking in the experimental arena throughout the whole 90 minute session. Upon cocaine administration, a complex dynamics of fly movement leads to immobility (marked by the red dot) which is followed by full recovery of movement. The path segment leading to immobility is colored in blue, and the path leading out of it is colored in green. The path traced in Figure 1A unfolds in time in Figure 1B, highlighting immobility as the origin to which motion converges and from which motion unfolds. As shown in Figure 1C, the fly first traces relatively straight paths, which become increasingly more curved culminating with immobility. In Figure 1D, transition out of immobility starts with highly curved paths involving many very small circles followed by increasingly straighter paths. The progressive narrowing down of the locomotor path into immobility, and the progressive buildup of the path into normal behavior is quantified for all flies in Figures 1E and F. 4 Figure 1. The locomotor path of cocaine treated flies narrows down into immobility and builds up to spread-out normal behavior. (A) Fly path during the whole 90 minutes session in a circular arena. Red dot indicates location of immobility. Blue colored path depicts transition in and green colored path transition out of immobility. (B) Same path as in (A) unfolded in time. (C) The transition into immobility is marked by the performance of straight, then increasingly more curved paths, narrowing down the animal’s locomotor repertoire. (D) The transition out of immobility is marked by the performance of curved, then increasingly straighter paths, widening up the animal’s locomotor repertoire. (E) By rescaling time in reference to immobility (marked by the vertical red lines), we demonstrate the narrowing of the locomotor path for all flies during transition into immobility, and (F) the expansion of the locomotor path for all flies during transition out of immobility. Behavior in and out of immobility: narrowing down of the fly's locomotor repertoire and its build up to normal behavior. Defining the three degrees of freedom of locomotor behavior: speed, direction of walking, and body orientation. The trajectory traced by an animal on a substrate can be characterized by the x-y coordinates as a function of time (Figure 2A), from which the velocity vector is calculated. The absolute value of the velocity vector is the speed of the animal (Figure 2B). In addition to speed, we can calculate the path curvature (Figure 2C). Taking into account the orientation of the longitudinal axis of the fly’s body, we can access a third degree of freedom (Figure 2D), distinguishing and quantifying in what direction the animal is moving, how fast it is progressing and in what direction it is facing. The left column of Figure 2 corresponds to these degrees of freedom evolving in time for a short trajectory segment, while the right column illustrates the same degrees of freedom in space. The centroid path is presented in Figure 2E. It is color coded according to its curvature in Figure 2F. Disk size is proportional to instantaneous speed in Figure 2G. Arrows depict the facing direction of the animal in Figure 2H. In the illustrated trajectory segment, a burst in speed as the animal traces a relatively straight path is followed by low speed at very high curvature, while the animal rotates in place, at approximately 360 degrees per 5 second. Figure 2. The three degrees of freedom exercised by a fly at the trajectory level. Time evolution of XY position (A), centroid speed (B), path curvature (C), and body orientation (D), during a 30 second time segment. Same degrees of freedom presented in space: black line represents fly path (E), circle diameter depicts speed (G), color intensity depicts curvature (F), and arrows depict body orientation (H). Small red dot marks the beginning of the trajectory. Increase in speed is followed by very low speed at high curvature while the fly vigorously rotates in place. Figure 3 shows the three degrees of freedom for the whole dynamics of transition into and recovery from immobility: speed (Figure 3A), curvature (Figure 3B) and rotation (Figure 3C). Using extended immobility as a reference we trace the behavior that precedes it, starting with the inflow of cocaine into the arena, and ending with full recovery of the fly marked by the absence of high rotation in place (high changes in body orientation at low translational speed) and the performance of straight paths (high speed at low curvature). As shown, the fly was totally immobile for 10 minutes (grey shaded area). Recovery from immobility starts with very fast whole-body rotations in place. Each diagonal line in magenta stands for a full 360 degrees rotation. The fly performs approximately 50 full rotations in 10 minutes with almost zero translational velocity and extremely high path curvature. The high frequency of rotations gradually decreases, and so does path curvature. At that time, the animal resumes normal forward progression involving relatively straight paths and high velocity. Transition into immobility starts with normal locomotion marked by high speed, low curvature and absence of extensive body rotations. Next, we observe bursts of high velocity followed by medium and then high curvature, which concur with the setting in of rotations at very low translation speeds, culminating with immobility. We can summarize the moment-to- 6 moment dynamics of this fly as the following sequence: predominance of translation, then high curvature, and finally extensive rotation in place, ceasing in immobility, from which the same sequence is performed in reverse. In other words, forward translation is eliminated from the fly's repertoire first, and rotation last, before the onset of immobility (shutdown; movie S1) rotation is added to the repertoire first, and forward translation last, in the transition out of immobility (warmup; movie S2). A representation of the progressive narrowing down of degrees of freedom for movement is provided in movie S3, and of progressive expansion in movie S4. Figure 3. Moment-to-moment dynamics of the three kinematic degrees of freedom of a fly during a whole session. The shaded area marks the period of immobility which is used as a reference for the events that precede and follow it. (A) The session starts with bursts of speed that progressively decrease towards zero and, after a 10 minutes period of complete immobility, very low speed is followed by normal speed. The green shaded area highlights small but non-zero velocity components at high-curvature during rotation in place, to be studied in depth in Figure 4. (B) Straight path is followed by bursts of high curvature until immobility, from which the fly resumes its movement with very high curvature (of the order of a 360 degrees turn in a millimeter) monotonically decreasing to straight paths again. (C) Extensive body rotations precede and follow immobility, proceeding from low to high frequency and from high to low frequency. Red shaded areas represent time segments when the fly touches the walls of the arena and body orientation is not tracked. Each diagonal line represents a 360 degrees body rotation. All in all, the session starts with extensive translation, then increasing curvature concurring with frequent body rotations leading into immobility. After immobility, extensive rotation in place concurring with very high path curvature is followed by forward progression along straight paths. Behavior out of immobility: walking on highly curved paths at low speed during fast rotation in place near immobility. A closer look at path dynamics (Figure 4A) reveals that high curvature emerges from small, fast and alternating oscillations in orthogonal components of the velocity vector (Figure 4B), that betray the minute circles traced by the animal as it is rotating in place. In particular, during the transition out of immobility, curvature shows a globally monotonical decrease from extremely high values to practically zero curvature. This illustrates that the three independent kinematic degrees of freedom are coordinated. To what extent are they coupled? Mathematically, 7 speed is independent of curvature and of body orientation, and body orientation is independent of path curvature (see Methods). Physically, they can be partly constrained in their relative magnitudes; for example, there is a limit to how fast an organism can go given a certain path curvature, and it is difficult to proceed along a straight line while rotating at high frequency. Biologically, it is an empirical relevant question whether and how these three degrees of freedom are actively managed by the animal. Flies show a range of different velocity components while walking at higher speeds (Figure 4C) implying at least partial independence. The freedom in the velocity components (thus, speed and curvature degrees of freedom) is constrained to synchronized oscillations at typical speeds of 1mm/s (Figure 4D) which correspond to the fly’s body rotating in place. Figure 4. Detailed dynamics of the transition out of immobility based on velocity components, which determine the dynamics and coordination of speed and curvature. (A) Transition out of immobility is illustrated by plotting the x and y positions as a function of time. Starting from absolute immobility, the fly performs tiny but fast oscillations in the x and y positions, reflecting fast rotations in place, which progressively slow down, and finally change to large displacements, corresponding to normal progression. (B) The early stage of transition out of immobility is characterized by very small speeds, whose perpendicular components (vx and vy) alternate in anti-phase oscillatory dynamics, corresponding to very high curvature. Speed along the x and y directions shows the coordinated circling as the fly gets out of immobility. Total speed (v) does not capture the subtleties of rotation in place. (C) Phase-plot of speeds along x and y directions, containing low and also high speed progression segments in all directions in a later stage of transition out of immobility. (D) Zoom in of the plot in (C) showing only the velocity components during the time interval from minute 2 to minute 14. Behavior in and out of immobility: the direction a fly walks and the direction it faces alternate in who leads and who follows. The relationship between walking direction and facing direction is versatile in flies. For example, flies can walk in any direction while keeping their body orientation fixed (Figure 5C), or else facing every which way while proceeding in a specific direction. In intact 8 flies, the direction of progression changes first, and the direction the fly faces then converges to the new direction set by progression, facing lagging behind by a small angular interval that is quickly closed. The same order of leading and following is exhibited with cocaine, but facing direction lags behind by a much larger angular interval taking a much longer time to close the interval25. Here we show that during the stage of transition out of immobility, as the fly rotates along highly curved paths, the two directions of progression and of facing tend to converge to the same values (Figure 5A-C). The further away from immobility, the less tight the coupling between these two degrees of freedom progressively gets (Figure 5D). Furthermore, the direction of progression leads first while body orientation follows and conversely, in later occasions as the fly regains its freedom of movement away from immobility, body orientation leads and direction of progression follows (Figure 5E). In other words, the angular interval between the direction of progression and body orientation is actively managed in two opposite ways, and modulated dynamically with respect to immobility. Figure 5. Active management of facing and walking directions as reflected in the dynamics of curvature and rotation. (A) Centroid speed time course as the fly transitions out of immobility. (B) Walking and facing directions during intense rotation in place and high-curvature dynamics. A two minutes segment of the top plot is amplified in the middle plot; and a one minute segment of the middle plot is amplified in the bottom plot, where the tight, but not perfect, coordination between both angles (facing and 9 walking directions) can be appreciated. (C) Phase-plot of facing-walking direction value combinations across a whole session (black), and selected segment showing that close to immobility the coordination is tight (red). (D) Cosine of the difference of facing and walking angles, progressively decreasing from 1 (zero difference), to 0.7 (45 degree difference), and reaching below 0 (more than 90 degree difference), quantifying the transition from tighter to looser coupling of rotational and curvature degrees of freedom as the fly gets out of immobility. (E) Quantification of the difference in angle between the degree of freedom that leads and the one that follows. Facing leads close to immobility, with walking direction lagging behind up to 30 degrees. Walking direction leads at later stages. Behavior in and out of immobility: the generative rules common to all fly sessions. Having discussed the relationship between curvature and rotation, we can now concentrate on the relationship between translation and rotation. Since the timescale we examine comprises the whole session dynamics (which can last for more than an hour), we calculate next the cumulative translation by integrating velocity in time and cumulative rotation by unwrapping the body angle (see Figure 6A). After smoothing the cumulative measures (see Methods) we calculate their derivative, and in this way we obtain the changes in rotation and in translation for the whole session (see Figure 6B). Again we use immobility as a reference and measure the global peaks of activity before and after immobility, along each of the two degrees of freedom, in order to quantify the sequence in which they unfold. This procedure reveals that before immobility the maximal peak of translation (T*) is exhibited before the maximal peak of rotation (R*). After immobility, it is the maximal peak of rotation that is exhibited before the maximal peak of translation. To quantify the relative order of global peaks we calculate the difference between the time of maximal peak of rotation and the time of maximal peak of translation, namely, t(R*)-t(T*). This procedure shows that translation precedes rotation before immobility, and that rotation precedes translation after immobility (see Figure 6C). While there is a very high variability in the time intervals across animals, all follow the same sequential order of global peaks. 10 Figure 6. Quantification of synchronic and diachronic dynamics of translation and rotation. (A) Cumulative body rotation (magenta) and translation (cyan) reveals the global sequence of changes in the rotational and translational degrees of freedom. The shaded area marks the period of immobility which is used as a reference for measuring the events that precede and follow it. (B) Global changes in speed of progression and body orientation are obtained from the absolute time derivative of the curves in (A). As shown, a global peak in translation followed by a global peak in rotation precede immobility, and a global peak in rotation followed by a global peak in translation follow immobility (TRIRT). To characterize the strength of the reciprocal relations between global peaks we calculate the ratios between global peak of one degree of freedom, R(R*), and its value at the global peak of the other degree of freedom, R(T*). (C) Difference between the time of maximal peak of translation and the time of maximal peak of rotation, t(T*)- t(R*), is negative for transitions into immobility (p=0.004, Sign test) and positive for transitions out of immobility (p=0.004, Sign test). Absolute time differences are greater for transitions out of immobility than for transitions into immobility (p=0.0352, Sign test). Each line connecting dots represents the same animal. Mean and standard deviation in blue. (D) Quantification of the strength of reciprocity in the value of global peaks of rotation and translation during transitions in and out of immobility. The ratio T(R*)/T* is smaller than 1 for transitions in and out of immobility (p<0.004, Sign test). Dots represent the score for individual flies and lines connect the results for the same individual. The mean and standard deviation are colored in blue. Next, in order to quantify the relative strength of the reciprocal relationship between global peaks, we calculate translation at its peak, T(T*), and compare it with translation when rotation is at its peak T(R*). Thus we measure the amount of reduction in translation by the time that rotation reaches its peak. The smaller the ratio T(R*)/T(T*), the stronger the phenomenon of reciprocity between translation and rotation (Figure 6D). Note that this relationship is by no means reciprocal at all times: in the graph presented there are cases where both rotation and translation increase, and also where both decrease together. In other words, rotation and translation are globally, not locally, 11 reciprocal. Characterizing the relative strength of rotation and translation peaks by means of the above ratio is invariant to time rescaling and to absolute values of rotation and translation. This is necessary for capturing the invariance in the sequence and strength across individual animals, and particularly useful given the large variability in the timescales of unfolding of the phenomenon (some flies take minutes, others take hours) and in the rotation values (some flies perform ten full body rotations, others perform hundreds; and they do so at different rates). On the whole, flies follow the same sequence of transition into immobility, involving, for each dimension separately, an enhancement, a reduction and then elimination of that degree of freedom, thus progressively narrowing down the fly's locomotor repertoire, and the same but opposite sequence of transition out of immobility, involving, for each dimension separately, an enhancement, and then subsiding to normal of that degree of freedom, thus progressively widening up the fly's locomotor repertoire. Such invariance can be summarized by the acronym TRIRT (Translation, Rotation, Immobility, Rotation, Translation). Behavior in and out of immobility: the rate of switching between directions of rotation. One predominant effect of cocaine is the high rate of repetition of full body rotations and their rotational speed. During rotation the animal may switch between clockwise (CW) and counterclockwise (CCW) rotations. It is now possible to examine the dynamics of switching in reference to immobility, in the context of the animal’s freedom of movement. In asking this question, we focus on how frequently the fly changes the direction of rotation and how biased successive rotations are. Globally, there are long-term predominant biases to rotate in a particular direction. In particular, transitions out of immobility start by very long rotations in the same overall direction. However, the flies do not rotate monotonically in one direction but rather alternate between large amplitude rotations in one preferred direction and low amplitude rotations in the other direction (Figure 7B- C). Locally, the fly alternates between CW and CCW rotations. As shown in Figure 7, the switching rate decreases when going into immobility and increases when going out of immobility. As found for the synchronic relationship between translation and rotation (TRIRT), this diachronic pattern for rotational switching also exhibits mirror symmetry between the process leading to immobility and out of it (Figure 7E-F), and can be summarized by the acronim SsIsS (high Switching, reduced switching ability, Immobility, and the reverse). 12 Figure 7. Switching between clockwise and counterclockwise rotation decreases into immobility and increases out of immobility. (A) Cumulative body orientation as a function of time for a single fly across the session. The grey shaded area marks the period of immobility. The blue shaded area marks a zoomed in time interval presented in (B) and (C). As shown, the general tendency of this fly is to rotate clockwise both before immobility and after immobility. Note that during transition into immobility the fly performs of the order of 50 full body rotations in 5 minutes, and during transition out of immobility, as much as 100 full body rotations in 5 minutes. (B) Zoomed in time segment in (A) of cumulative body rotation. As illustrated, while the fly generally rotates clockwise, it keeps switching between clockwise and counterclockwise rotations. Each transition is marked by a small black dot on the curve and a corresponding vertical bar on the transitions raster plot below. (C) Time derivative of the fly’s momentary body orientation. Zero crossings in the rotational speed mark switching, corresponding to the vertical bars in (B). As shown, there is an overall drift in one direction, concurrent with a decrease in the number of transitions as a function of time. (D) Global changes in body orientation obtained from the absolute time derivative of the curve in (A). To examine the change in switching across the session, we partition the period leading to immobility into the segments preceding and following maximal rotation. We do the same for the period leading out of immobility. (E) Switching rate, calculated as the number of transitions per second, decreases in the interval preceding immobility (left panel) and increases in the interval following immobility (right panel). Dots represent the score for individual flies and lines connect the results for the same individual. (F) Switching rate decreases when going into immobility (left panel, p<0.039, Sign test) and increases when going out of immobility (right panel, p<0.039, Sign test). Dots represent the switching rate change for every animal. Discussion The generative rules that shape fruit fly cocaine induced locomotor behavior during the transition out and into immobility. Using immobility as a reference for the measurement of behavior and cocaine as the parameter 13 inducing a behavioral gradient we found that the flies exhibit a progressive expansion of their locomotor repertoire starting from immobility, and a progressive narrowing down of the repertoire leading into immobility. During expansion, for each key variable separately (fig. 6), the fly exhibits first enhancement and then reduction to normal values of movement along that variable: first, of body rotation in the horizontal plane, then of path curvature and then of speed of translation. The extents of movement across the key variables show reciprocal relations: when rotation is at its peak translation is low, and when translation is at its peak rotation is low, and path curvature is partly coupled to rotation. Transition into immobility from rich normal locomotor behavior unfolds by narrowing down of the repertoire in the opposite sequential order, also showing reciprocal relations between the extents of the same variables. Quantification of the generative rules of this behavior, based on the temporal sequence of global peaks of extent (fig. 6 C,D) and their reciprocal values provides a summary of the bauplan of this arthropod behavior, allowing a comparison to the rules reported in previous studies for movement into and out of immobility in vertebrates, where the behavior has been termed The mobility gradient17. Vertebrates and fruit flies share the same generative rules. Warmup: In infant mice transition out of novelty-induced immobility consists of side-to-side head movements that increase in amplitude, gradually recruiting the forequarters, and then the hindquarters to extensive rotation in place around the hindquarters. Only after exhausting the horizontal plane by rotating around the hindquarters, forward stretching and subsequently forward translation appears, first along curved and then along straight paths (movie S5). The same sequence is exhibited in amphibians26, Fish27, insectivores (movie S6), and carnivores28. Head raising, forequarters raising and finally rearing on the hind legs are exhibited next29,30. The same sequence is exhibited both during moment-to-moment behavior and in ontogeny, during recovery from lateral hypothalamic damage31. Later in development, during, for example, play and ritualized fighting interactions, the inferior animal exhibits the less mobile portion of the sequence, culminating it with rearing and rotating around the hindquarters, whereas the superior may rear and rotate both around the hind legs and around the forelegs, exhibiting an expanded freedom of movement in both the horizontal and the vertical dimension28,32; for a review see17,33 (movie S7). Shutdown: the opposite sequence, proceeding from rich normal behavior to enhanced, then reduced, then annulled movement, first of rearing, then of translation along straight, and then along curved paths, then of rotation, culminating in relative immobility is exhibited in rats following the administration of several dopamine agonists4,5,34,35,36,37,38 (Movie S8). This mobility gradient17,39, which is composed of warmup and shutdown sequences29,30 shares with the mobility gradient 14 exposed in fruit flies the same origin (immobility), knob (dopaminergic stimulation; but in vertebrates also novelty and proximity to a rival) and bauplan. In the absence of fossils of behavior we temporarily suspend judgement about common descent. From the vantage point of comparative anatomy, bauplans are the core, or skeleton that resisted adaptation and around which there is a variable adaptive component. As shown by ethology, the core/adaptive distinction also applies to behavior19,40. Searching for equivalent neurochemical substrates mediating the expansion and narrowing down of the animal's locomotor repertoire. Recently, it has been claimed that the vertebrate basal ganglia and the arthropod central complex are deeply homologous. In both, comparable systems of dopaminergic neurons, their projections, and dopaminergic receptor activities are involved in the modulation and maintenance of behavior21. Dopamine systems are also key players in generating and regulating the mobility gradient23,24,34-38, and dopaminergic stimulation of specific substructures of the basal ganglia and central complex induce specific components of the mobility gradient respectively in rodents41-45 and in flies46. The neurochemical processes mediating the vertebrate expansion of locomotor repertoire have been recently attributed by Cools and co-workers to dopaminergic feedforward loops operating in the basal ganglia23,24. The equivalence between the mobility gradient core phenomena and the feedforward loops exposed in the basal ganglia can be used as a search image or working hypothesis for studying the relations between the arthropod mobility gradient and the central complex. It might be useful to examine whether feedforward loops also mediate the expansion of locomotor behavior functions in the central complex. Methodological considerations. We found the following choices useful for uncovering common cross-phyletic generative rules of behavior In the study of behavior, we use the methodology practiced by comparative anatomy. Ethology's search for behavioral homologies was fulfilled only in closely related species. Founded by comparative anatomists, ethology aspired to provide a description of the fixed core of behavior, 15 conceived to be resistant to past environmental pressures as well as to current experimental manipulations40. Two main discoveries that were claimed by classical ethology were that (i) behavior was an extension of brain anatomy, and therefore its structure must reflect the connectivity of the brain, and that (ii) homologies were as applicable to behavior as to anatomy47,48. The comparative study of behavior made a clear-cut distinction between structure and function, examining them from the point of view of a fixed core (e.g., grasping with foot in canaries) used in a variety of contexts (e.g., for nest building and for feeding19). Lorenz accepted his share of the Nobel Prize with the reflection that his “most important contribution to science” has been his discovery “that the very same methods of comparison, the same concepts of analogy and homology, are as applicable to characteristics of behavior as they are to those of morphology” 19. In particular, to the same extent that anatomists use bones to establish the concept of skeletal homology and brain circuitries to refer to neural homologues, ethologists need to isolate the particulate processes of behavior in order to establish behavioral homologues. We prioritize the use of continuous kinematic variables over the use of ad hoc discrete patterns. The fundamental building block of classical ethology has been the so-called “behavior pattern”, a species-specific coordination involving movements of some or all the parts of the animal’s body, and performed one at a time by the animal6,4950. These expert-established units were based, however, on form, which has not been part of the definition of anatomical homologies. For example, there is no similarity in the respective form and relative size of the bone identified as the radius in the forelimb of a human and of a whale, yet these bones are homologous. As already proposed in the 19th century, anatomical homology must be defined, not by its form, but by the relative positions and spatial relations between the elements of a structure, (“principle of connections” of St. Hillaire; “equivalence under transformation”2,7,8,9). The radii of the human and the whale thus derive their identity from the same relative position they occupy and their identical connectivity to the other bones of the respective forelimbs. Whenever we call a radius a radius we implicitly or explicitly imply a forelimb, which in turn implies a vertebrate's skeleton. In much the same way, the striatum of a bird and of a reptile, although very different in size and form, is identified as striatum because of its connectivity to other brain structures51. By parsing fly behavior prematurely into discrete patterns such as circling and locomotion4, rather than representing them in terms of speed of translation, path curvature, and body orientation, the opportunity to derive the generative rules that constrain these variables is lost, and with it is lost the opportunity to discover the universality of the rules. Lorenz and followers studying duck display could not demonstrate homological sequences of patterns because the sequence was much too variable even across these 16 closely related species, disallowing the claim that a pattern occupied the same relative position in two corresponding sequences performed by different species3. Ethology did succeed to expose behavioral homologies across closely related species (e.g.,52,18), but not across phyla. We describe the progressive unfolding of behavior from simple to complex and vice versa in reference to an origin presumably used by the organism itself. As quantified in figures 5-7, demonstrating the invariant sequence could be accomplished by proceeding away from immobility in both directions so as to capture the serial emergence of first global peaks along each of the degrees of freedom in warm up and the serial performance of the last global peaks in shutdown. As degrees of freedom are added (or, respectively, eliminated) the sequence becomes increasingly unpredictable in warm up, or respectively stereotyped in shutdown, and the invariant sequence is lost. This applies both to the fly in this study and the mouse30. This is why sequential invariance (and hence homology) cannot be demonstrated in full blown behavior, which becomes increasingly differentiated, i.e., unpredictable, across performance. We prioritize structure over function. The question whether homology should be based first and foremost on structure or on function engaged comparative anatomists for almost two hundred years16. While prioritizing structure has become common practice for almost a century with regard to brain and skeletal anatomy13,14, behavioral neuroscientists have been fluctuating between prioritizing function over structure and vice versa to this day15. The "structure first" advocates3,18,53 were aware of the need for data acquisition and analysis tools not available at the time, but missed the need for a descriptive technology highlighting the quantities that are actively managed by the organism (which was already available at the time54. Current adherents of "function first", influenced by experimental psychology, Neo-Darwinism, and medical diagnosis, focus on the adaptive components of behavior, involving learning, habit formation, and normal locomotor performance, eg.,21. The descriptions they use are useful as long as the animal exhibits adaptive behavior, but fail with drug- or lesion- or genetically modified behavior. This is perhaps why the warm up and shutdown sequences (figure 1) were described as abnormal, aberrant, unusual and uncontrolled6, and dopamine-manipulated behavior was described as ataxic, hypo- or hyperkinetic21, missing the highly organized dynamics (figures 2-7). While a "function first" methodology thus considered warmup and shutdown as conserved4 though aberrant6 relics, a "structure first" methodology would treasure it as a crystalized manifestation of a cross phyletic invariance exposing the generative rules of a substantial component of arthropod and vertebrate locomotor behavior. In Goethe's words55 "Organisms can transform their behavior into 17 misshapen things not in defiance of law but in conformity with it, while at the same time, as if curbed with a bridle , they are forced to acknowledge its inevitable dominion". The dominion belongs to the generative rules. We prioritize generative rules over common descent explanations of behavior. The tension arising between these two ways of making sense of the diversity and unity of species that has been revealed through evolution has animated many biological debates for the last 180 years2,17,26,56,57 and is currently experiencing a revived interest (https://royalsociety.org/events/2015/03/nervous- system/). The mainstream approach in neuroscience relies on historical explanations (common descent), embryonic lineages, and genetic programming21. A less common way, implemented in the current study, is based on demonstrating equivalence of structure by exposing the common cross phyletic generative rules that shape a diversity of forms through transformational unity. Controversy prevails, not only with regard to the primacy of structure but also with regard to how dependent is a structural definition of homology on the demonstration of a common descent. While Beer argues that "Clearly there has to be a way of judging homology independently of evolutionary considerations3, and Wagner13,14 separates homology from phylogeny by defining homology in terms of structure and development, Hodos and Atz maintain that behavioral homology must include a statement of the underlying structural elements that at least in principle can be traced to specific ancestral precursors18,53. The consistency of the phenotype (e.g., behavior) that species show despite much turnover in their gene pool58) is another argument for independent characterization of the phenotype. Characterization of homology on the basis of equivalence of generative rules is implemented by Goodwin9 with regard to the morphogenesis of anatomical structures and by Golani17 with regard to Kinematics. Here we address the common architectures while suspending judgement about common descent, which, once demonstrated, will endow the bauplan with the status of a Darwinian homology. This procedure implements the heuristic potential of a map that sensitizes and guides the researcher to incipient, disguised, or missing parts of a structure59,16 both at the neural and the behavioral levels. We apply the “serial homology” concept to situations where the same structure serves different functions in the same species. Using our generative rules as a search image, not only can we predict essential sameness in the brain/behavior interface of flies and rodents, but also anticipate that the same rules might underlie apparently different functional behaviors in the same species. This is similar to identifying serial homology in anatomy, where, e.g., hand and foot are considered 18 homologous because they share the same set of developmental constraints, caused by locally acting self-regulatory mechanisms of differentiation14. Reviewing the fruit fly larval behavior literature with a search image for low and high mobility one's attention is immediately captured by the abnormally high amounts of turning behavior exhibited by larvae with mutations in scribbler in the absence of food60. These appear to be respective manifestations of the high and low ends of the mobility gradient. The four key features characterizing low mobility in the cocaine treated fly – low speed of translation, highly curved path, high body rotation and immobility – exhibit a full correspondence to the features of the "abnormal crawling pattern" exhibited by scribbler larvae: low speed, curved paths, high turning rate, and long pauses61. The "knob" precipitating this behavior could be, as Sokolowski and co- workers suggest, the absence of food, or else, given our search image, the stress brought about by the absence of food, and even its presence in hungry flies62,63. Equivalent differences in mobility, expressed by pivoting and/or rearing on hind legs and forelegs, reported between vertebrate partners engaged in interactions32,28, should be looked for in fruit fly courtship and agonistic interactions. Conclusion. Homology is studied at all levels of biological organization from molecules to behavior64. Its nature is elusive 13,14 , even more so with regard to behavior, where the distinction between the invariant conserved core and the adaptive component has been hardly described yet. There is a consensus with regard to the need for a comprehensive description of the equivalence of relations across all levels of the hierarchy all the way down to the molecular level and all the way back to common descent. But there is little awareness of the absence of cross phyletic homologues of behavior, the final common pathway of the hierarchy specifying the demand17,65 on all the other levels. Homologues are indispensable because there cannot be a comparative science without having validated particulate processes that combine to exhibit the overall structure. Homology validates the partitioning of the whole into parts, based on the demonstrated equivalence under transformation. The state of the art with regard to the comparative study of behavior is similar to that of comparative anatomy 200 years ago: in pre-Darwinian times baupläne were useful heuristic tools for searching for "missing" bones59,16. The same methodology can now be applied in phyletic and cross phyletic comparisons of behavior without reducing in any possible way the search for underlying equivalent connectivity at the neural, genetic and molecular levels, in the pursuit of common descent. The generative rules derived from fruit fly cocaine-induced seemingly aberrant behavior are rules 19 of differentiation and decay, exhibited on the three time scales of phylogeny66, ontogeny29,30, and actual genesis. Because they represent managed perceptual quantities they also define the organism's operational world, its umwelt67. In this way baupläne and umwelten are represented within the same phenomenological frame. Given the foundational role of behavior in neuroscience68, it is perhaps time to start relating the various levels of the biological hierarchy to the behavioral baupläne they support, turning them -where possible- into Darwinian homologies. Materials and Methods Fly stocks. Drosophila cultures were maintained at 24°C on a standard cornmeal-molasses medium in 12 hour light-dark cycle at 60% humidity. The experiments were performed on three-day-old flies of the wild-type laboratory strain Canton-S. Nine male flies were tested in a low-throughput high-content data approach. Behavioral arena. The experimental setup for observing and tracking the flies was a 15 cm diameter circular arena with 0.7cm height wall and glass ceiling. The arena was illuminated from above with a 40W bulb. A camera placed above the arena recorded the fly's behavior. During the experiment there was a continuous airflow through the arena, through two small wall openings allowing also the introduction of the volatilized drug into the arena during the experiment25. Animal preparation. Neither food nor water was supplied to the fly during the entire experiment. All experiments were performed during the 12 hours light period, and on one fly at a time. Each fly was transferred to the arena and allowed to habituate there for one hour. Upon exposure to cocaine, the fly behavior was uninterruptedly recorded including full sedation (immobility) and the process of recovery (regaining normal locomotor behavior). Behavioral tracking. The fly locomotor behavior was recorded at 25 frames per second using a CCD camera. Following video acquisition, the centroid position of the fly and its body axis direction were tracked with FTrack69, a custom-made software written in Matlab (Mathworks). Raw trajectory data were corrected for tilt and rotation of the camera. Data segments during which it was not possible to assess the fly's orientation (fly located on the wall or jumping) were excluded from analysis. 20 Behavioral analysis. Quantitative analysis of the animal’s behavior was based on the dynamics of three main degrees of freedom: centroid speed, centroid change in direction per unit of distance (path curvature) and body rotation (change in trunk orientation in the horizontal plane). Distinguishing between the direction of progression (centroid-based), the speed of progression, and the change in direction of rotations (body-based; where its longitudinal axis is facing) - the animal can, e.g., walk in one direction while facing with its rigid body any which way - north and then turn left, while still having the freedom to face north- allows to study the active management of where it is facing, where it is going, and how fast it is going25,70,71. Changes in the direction of progression are calculated per unit of progression and as a function of time in order to have a geometric curvature72,73 in kinematic terms74,75. Switching between clockwise and counterclockwise body rotation was assessed via the zero-crossings of the instantaneous time derivative of the body angle, removing artifacts during arrests by pruning out rotations smaller than 12 degrees. In order to produce reliable estimates of local and global variables (as we study phenomena at different timescales: from sub-second small oscillations during rotation in place, to near-hour recovery trends) we use the variable-window smoothing LOWESS method76. Immobility, defined as the longest time interval of complete arrest (no translation nor rotation) across the whole session, allows to transform chronological time into activity, revealing dynamical invariants despite animal- to-animal behavioral variability. Acknowledgements. This study was funded by the Portuguese Foundation for Science and Technology (FCT grant No SFRH/BPD/97544/2013 to AGM) and by the Israel Science Foundation (grant No 760/08 to IG and YB). We thank the participants of the “Homology in NeuroEthology” course held at the Champalimaud Neuroscience Programme for fruitful discussions. We acknowledge the gifts of fly stocks from Dani Segal. We thank Gonçalo Lopes, Eduardo Dias-Ferreira, Andre Brown, Lauren McElvain, Troy Shirangi, Eyal Gruntman and Ehud Fonio for valuable feedback on the manuscript. Author contributions: I.G., Y.B., and A.G.M. conceived the research; E.O. performed the experiments; D.V. developed the tracking software; A.G. preprocessed the data; Y.B. and A.G.M designed the quantitative analysis; A.G.M. analyzed the data and made the figures; A.G.M, Y.B, and I.G. wrote the paper. 21 22 References 1. Owen, R. (1843). "Lectures on the Comparative Anatomy and Physiology of the Invertebrate Animals, Delivered at the Royal College of Surgeons, in 1843." Longman, Brown, Green, and Longmans, London. 2. 3. Saint-Hilaire, E. G. (1830). Principes de philosophie zoologique, discutés en mars 1830, au sein de l’Académie Royale des Sciences. Beer, C. G. (1980) Beer, C. G. (1980). Perspectives on animal behavior comparisons. In Comparative methods in psychology (pp. 17-64). Erlbaum Hillsdale, NJ. 4. McClung, C. & Hirsh, J. (1998). Stereotypic behavioral responses to free-base cocaine and the development of behavioral sensitization in Drosophila. Curr. Biol. 8, 109–112 5. Bainton, R. J., Tsai, L. T., Singh, C. M., Moore, M. S., Neckameyer, W. S., & Heberlein, U. (2000). Dopamine modulates acute responses to cocaine, nicotine and ethanol in Drosophila. Current Biology, 10(4), 187-194. 6. Rothenfluh, A. & Heberlein, U. (2002) Drugs, flies, and videotape: the effects of ethanol and cocaine on Drosophila locomotion. Curr. Opin. Neurobiol. 12, 639–645. 7. Woodger, J. H. (2014) Biological principles: A critical study. Routledge. 8. 9. Goodwin, B. C. (1982) Development and evolution. J. Theor. Biol. 97, 43–55. Goodwin, B. (1993) Homology and a generative theory of biological form. Acta Biotheor. 41, 305–314. 10. Inglis, W. G. (1966) The observational basis of homology. Syst. Zool. 219–228. 11. Goodwin, B. C. (1984). Changing from an evolutionary to a generative paradigm in biology. In "Evolutionary Theory. Paths into the Future" (J. W. Pollard, ed.), pp. 99-120 12. Webster, G., & Goodwin, B. (1996). Form and transformation: generative and relational principles in biology. Cambridge University Press. 13. Wagner, G. P. (1989a). The origin of morphological characters and the biological basis of homology. Evolution (Lawrence, Kansas) 43, 1157-1171. 14. Wagner, G. P. (1989b). The biological homology concept. Annu. Rev. Ecol Syst. 20, 51-69. 23 15. Lauder, G. V. (2012). Homology, form, and function. Homology: The Hierarchial Basis of Comparative Biology, 151. 16. Appel, T. A. (1987). Cuvier-geoffroy Debate: French Biology in Decades Before Darwin. Oxford University Press. 17. Golani, I. (1992). A mobility gradient in the organization of vertebrate movement: the perception of movement through symbolic language. Behavioral and Brain Sciences, 15(02), 249-266. 18. Atz, J. (1970). The application of the idea of homology to animal behavior. In "Development and Evolution of Behavior: essays in honor of T. C. Schneirla" (L. Aronson, E. Tobach, D. S. Lehrman, and J. S. Rosenblatt, eds.), pp. 53-74. Freeman, Sam Francisco. 19. Lorenz, K. (1957) The nature of Instinct. Trans. Clair H. Schiller. In Instinct: the development of a modern concept. Ed. . C.H. Schiller pp. 129-175. London: Methuen. 20. Thom R. (1992). A Mobility Gradient in the Organization of Vertebrate movement. Behavioral and Brain Sciences 15 (2):289-289. 21. Strausfeld, N. J. & Hirth, F. (2013). Deep homology of arthropod central complex and vertebrate basal ganglia. Science 340, 157–161. 22. 23. Powers, W. T. (1973) Behavior: The control of perception. (Aldine Chicago). Ikeda, H., Saigusa, T., Kamei, J., Koshikawa, N. & Cools, A. R. (2013). Spiraling dopaminergic circuitry from the ventral striatum to dorsal striatum is an effective feed-forward loop. Neuroscience 241, 126–134. 24. Cools, A. R. (1992) Striatal structures, dopamine and the mobility gradient model. Behav. Brain Sci. 15, 271–272. 25. Gakamsky, A., Oron, E., Valente, D., Mitra, P. P., Segal, D., Benjamini, Y., & Golani, I. (2013). The Angular Interval between the Direction of Progression and Body Orientation in Normal, Alcohol-and Cocaine Treated Fruit Flies. PloS one,8(10), e76257. 26. Coghill, G. E. (2015). Anatomy and the Problem of Behaviour. Cambridge University Press. 27. Eaton R. C. (1992) Eshkol-Wachman movement notation and the evolution of locomotor patterns in vertebrates. Behavioral and Brain Sciences 15 (2):272-274 (1992) 24 28. Golani, I., Moran, G. (1983). A motility-immobility gradient in the behavior of the" inferior" wolf during" ritualized fighting. Advances in the study of mammalian behavior. The American Society of Mammalogists, Special Publication, (7), 65- 94. 29. Golani, I., Bronchti, G., Moualem, D. & Teitelbaum, P. (1981). ‘ Warm-up’ along dimensions of movement in the ontogeny of exploration in rats and other infant mammals. Proc. Natl. Acad. Sci. 78, 7226–7229 30. Eilam, D. & Golani, I. (1988). The ontogeny of exploratory behavior in the house rat (Rattus rattus): the mobility gradient. Dev. Psychobiol. 21, 679– 710. 31. Golani, I., Wolgin, D. L., & Teitelbaum, P. (1979). A proposed natural geometry of recovery from akinesia in the lateral hypothalamic rat. Brain research,164(1), 237-267. 32. Yaniv, Y. & Golani, I. (1987). Superiority and inferiority: A morphological analysis of free and stimulus bound behaviour in honey badger (Mellivora capensis) interactions. Ethology 74, 89–116. 33. Golani, I. (2012). The developmental dynamics of behavioral growth processes in rodent egocentric and allocentric space. Behav. Brain Res. 231, 309–316. 34. Szechtman, H., Ornstein, K., Teitelbaum, P. & Golani, I. (1985). The morphogenesis of stereotyped behavior induced by the dopamine receptor agonist apomorphine in the laboratory rat. Neuroscience 14, 783–798. 35. Adani, N., Kiryati, N. & Golani, I. (1992). The description of rat drug-induced behavior: kinematics versus response categories. Neurosci. Biobehav. Rev. 15, 455–460. 36. Kafkafi, N., Levi-Havusha, S., Golani, I. & Benjamini, Y. (1996). Coordination of side-to-side head movements and walking in amphetamine-treated rats: a stereotyped motor pattern as a stable equilibrium in a dynamical system. Biol. Cybern. 74, 487–495. 37. Eilam, D., Golani, I. & Szechtman, H. (1989). D 2-agonist quinpirole induces perseveration of routes and hyperactivity but no perseveration of movements. Brain Res. 490, 255–267. 25 38. Golani, I., Einat, H., Tchernichovski, O. & Teitelbaum, P. (1997). Keeping the body straight in the unconstrained locomotion of normal and dopamine- stimulant-treated rats. J. Mot. Behav. 29, 99–112. 39. Golani, I. The natural geometry of a behavioral homology. (1992). Behav. Brain Sci. 15, 291–308. 40. Lehrman, D. S. (1953). A critique of Konrad Lorenz's theory of instinctive behavior. Quarterly Review of Biology, 337-363. 41. Kelly, P. H., Seviour, P. W. & Iversen, S. D. (1975) Amphetamine and apomorphine responses in the rat following 6-OHDA lesions of the nucleus accumbens septi and corpus striatum. Brain Res. 94, 507–522. 42. Kelley, A. E., Gauthier, A. M. & Lang, C. G. (1989) Amphetamine microinjections into distinct striatal subregions cause dissociable effects on motor and ingestive behavior. Behav. Brain Res. 35, 27–39. 43. Kelly, P. H. & Moore, K. E. (1976) Mesolimbic dopaminergic neurones in the rotational model of nigrostriatal function.. 44. Swerdlow, N. R. & Koob, G. F. (1989) Norepinephrine stimulates behavioral activation in rats following depletion of nucleus accumbens dopamine. Pharmacol. Biochem. Behav. 33, 595–599. 45. Staton, D. M. & Solomon, P. R. (1984) Microinjections of d-amphetamine into the nucleus accumbens and caudate-putamen differentially affect stereotypy and locomotion in the rat. Physiol. Psychol. 12, 159–162. 46. Martin, J. R., Raabe, T., & Heisenberg, M. (1999). Central complex substructures are required for the maintenance of locomotor activity in Drosophila melanogaster. Journal of Comparative Physiology A, 185(3), 277-288.‏ 47. Whitman, C. O. & Carr, H. A. (1919) The behavior of pigeons. 3, (Carnegie Inst.). 48. Lorenz, K. (1974) Analogy as a source of knowledge. In: American Association for the Advancement of Science,. 49. Baerends, G. P. (1957) The ethological analysis of fish behavior. Physiol. Fishes 2, 229–269. 50. Barlow, G. W. (1996) Ethological units of behavior. Found. Anim. Behav. Class. Pap. Comment. 138–153. 26 51. Jarvis, E. D., Güntürkün, O., Bruce, L., Csillag, A., Karten, H., Kuenzel, W., & Butler, A. B. (2005) Avian brains and a new understanding of vertebrate brain evolution. Nature Reviews Neuroscience, 6(2), 151-159. 52. Kavanau, J. L. (1990). Conservative behavioural evolution, the neural substrate. Anim. Behav. 39, 758-767. 53. Campbell, C. B. G., & Hodos, W. (1970). The Concept of Homology and the Evolution of the Nervous System (Part 2 of 2). Brain, Behavior and Evolution,3(5-6), 361-367. 54. Eshkol, N., & Wachman, A. (1958). Movement Notation (Weidenfeld and Nicholson, London). ‏ 55. Goethe, J.W. (1830). Principes de Philosophic zoologiques discutes en Mars, 1830, au sein de l'Academie Royale des Sciences. Natnrwiss. Schriften, Werke VII, 189. 56. Hall, B. K. (2007) Homoplasy and homology: dichotomy or continuum? J. Hum. Evol. 52, 473–479. 57. Hall, B. K. (2012) Homology: The Hierarchial Basis of Comparative Biology. (Academic Press). 58. Mayr, E., et al., (1963). Animal species and evolution(Vol. 797). Cambridge, Massachusetts: Belknap Press of Harvard University Press. 59. Wells, G. A. (1967). Goethe and the intermaxillary bone. The British Journal for the History of Science, 3(04), 348-361. 60. Yang, P., Shaver, S. A., Hilliker, A. J. & Sokolowski, M. B. (2000) Abnormal turning behavior in Drosophila larvae: identification and molecular analysis of scribbler (sbb). Genetics 155, 1161–1174. 61. Suster, M. L., Karunanithi, S., Atwood, H. L. & Sokolowski, M. B. (2004) Turning behavior in Drosophila larvae: a role for the small scribbler transcript. Genes Brain Behav. 3, 273–286. 62. Dethier, V. G. (1976) The hungry fly: A physiological study of the behavior associated with feeding.. 63. Bell, W. J. Searching behavior patterns in insects. (1990) Annu. Rev. Entomol. 35, 447–467. 27 64. Hall, B. K. (Ed.). (2012). Homology: The hierarchical basis of comparative biology. Academic Press. 65. William T. Powers . (1992). Testing for controlled variables. Open Peer Commentary Behav. Brain Sci. 15, 286–287. 66. Eilam D. (1992) The mobility gradient from a comparative phylogenetic perspective. Open Peer Commentary Behav. Brain Sci. 15, 274–275. 67. von Uexküll, J. (1992). A stroll through the worlds of animals and men: A picture book of invisible worlds. Semiotica, 89(4), 319-391.‏64. 68. Gomez-Marin, A., Paton, J. J., Kampff, A. R., Costa, R. M. & Mainen, Z. F. (2014) Big behavioral data: psychology, ethology and the foundations of neuroscience. Nat. Neurosci. 17, 1455–1462. 69. Valente, D., Golani, I. & Mitra, P. P. (2007) Analysis of the trajectory of Drosophila melanogaster in a circular open field arena. PloS One 2, e1083. 70. Gomez-Marin, A., Stephens, G. J. & Louis, M. (2011) Active sampling and decision making in Drosophila chemotaxis. Nat. Commun. 2, 441. 71. Gomez-Marin, A., Partoune, N., Stephens, G. J. & Louis, M. (2012) Automated tracking of animal posture and movement during exploration and sensory orientation behaviors. PloS One 7, e41642. 72. Dvorkin, A., Szechtman, H. & Golani, I. (2010) Knots: attractive places with high path tortuosity in mouse open field exploration. PLoS Comput. Biol. 6, e1000638. 73. Iino, Y. & Yoshida, K. (2009) Parallel use of two behavioral mechanisms for chemotaxis in Caenorhabditis elegans. J. Neurosci. 29, 5370–5380. 74. Kafkafi, N. & Elmer, G. I. (2005) Texture of locomotor path: a replicable characterization of a complex behavioral phenotype. Genes Brain Behav. 4, 431–443. 75. Gomez-Marin, A. & Louis, M. (2014) Multilevel control of run orientation in Drosophila larval chemotaxis. Front. Behav. Neurosci. 8,. 76. Hen, I., Sakov, A., Kafkafi, N., Golani, I. & Benjamini, Y. (2004) The dynamics of spatial behavior: how can robust smoothing techniques help? J. Neurosci. Methods 133, 161–172. 28
1904.11702
1
1904
2019-04-26T07:46:49
Voluntary phantom hand and finger movements in transhumeral amputees could be used to naturally control polydigital prostheses
[ "q-bio.NC", "eess.SP", "physics.med-ph", "q-bio.TO" ]
An arm amputation is extremely invalidating since many of our daily tasks require bi-manual and precise control of hand movements. Perfect hand prostheses should therefore offer a natural, intuitive and cognitively simple control over their numerous biomimetic active degrees of freedom. While efficient polydigital prostheses are commercially available, their control remains complex to master and offers limited possibilities, especially for high amputation levels. In this pilot study, we demonstrate the possibility for upper-arm amputees to intuitively control a polydigital hand prosthesis by using surface myoelectric activities of residual limb muscles (sEMG) associated with phantom limb movements, even if these residual arm muscles on which the phantom activity is measured were not naturally associated with hand movements before amputation. Using pattern recognition methods, three arm amputees were able, without training, to initiate 5-8 movements of a robotic hand (including individual finger movements) by simply mobilizing their phantom limb while the robotic hand was mimicking the action in real time. This innovative control approach could offer to numerous upper-limb amputees an access to recent biomimetic prostheses with multiple controllable joints, without requiring surgery or complex training; and might deeply change the way the phantom limb is apprehended by both patients and clinicians.
q-bio.NC
q-bio
Voluntary phantom hand and finger movements in transhumeral amputees could be used to naturally control polydigital prostheses Nathanaël Jarrassé, Caroline Nicol, Florian Richer, Amélie Touillet, Noël Martinet, Jean Paysant and Jozina B. De Graaf to master and offers Abstract -- An arm amputation is extremely invalidating since many of our daily tasks require bi-manual and precise control of hand movements. Perfect hand prostheses should therefore offer a natural, intuitive and cognitively simple control over their numerous biomimetic active degrees of freedom. While efficient polydigital prostheses are commercially available, their control remains complex limited possibilities, especially for high amputation levels. In this pilot study, we demonstrate the possibility for upper-arm amputees to intuitively control a polydigital hand prosthesis by using surface myoelectric activities of residual limb muscles (sEMG) associated with phantom limb movements, even if these residual arm muscles on which the phantom activity is measured were not naturally associated with hand movements before amputation. Using pattern recognition methods, three arm amputees were able, without training, to initiate 5-8 movements of a robotic hand (including individual finger movements) by simply mobilizing their phantom limb while the robotic hand was mimicking the action in real time. This innovative control approach could offer to numerous upper-limb amputees an access to recent biomimetic prostheses with multiple controllable joints, without requiring surgery or complex training; and might deeply change the way the phantom limb is apprehended by both patients and clinicians. I. INTRODUCTION [5] a for comparative A perfect hand prosthesis should offer amputees a natural, intuitive and cognitively simple control over numerous biomimetic active degrees of freedom (DoF) [1, 2]. But there is an evident gap between devices and control approaches. Efficient polydigital prostheses such as the iLimb hands by Touch Bionics [3] or the Michelangelo hand by Ottobock [4] (see already commercialized, but are only available with limited control approaches such as a switchable catalog of postures chosen by co-contraction impulses. Researchers have been working for numerous years in developing alternative approaches to enhance controllability over active prosthetics and to offer direct control of each active Dof. Approaches like Targeted Muscle Reinnervation (TMR) [6] or brain computer interfaces (BCI) [7] are promising solutions, but they require heavy surgery with still reduced performances, especially for BCI. review) are N. Jarrassé and F. Richer are with Institute of Intelligent Systems and Robotics, ISIR CNRS-UMR 7222, INSERM U1150 Agathe-ISIR, Sorbonne University, UPMC Univ (e-mail: [email protected]). France Paris, Paris 06, C. Nicol and J. De Graaf are with the Institute of Movement Sciences, UMR 7287 -- CNRS & Aix-Marseille University, Marseille, France (e-mails: caroline.nicol / [email protected]). A. Touillet, N. Martinet and J. Paysant are with the Louis Pierquin Centre of the Regional Institute of Rehabilitation, Nancy, France (e-mail: amelie.touillet / noel.martinet / [email protected]). Non-invasive approaches, based on the use of electrode arrays placed over the residual group of muscles and pattern recognition algorithms, have made huge progress within the last decades [8]. This last approach has been recently reported to offer forearm amputees a control over a polydigital hand prosthetics via decoding of finger muscle activities with an array of electrodes placed over the residual forearm [9]. Figure 1. Global view of the setup along with the myoelectric activity associated to the voluntary control of phantom hand (measured with 12 sEMG electrodes placed over the residual limb of one participant). Phantom hand movements are named the following way: TF stands for Thumb Flexion, LF for Little finger Flexion, PC for Pinch Closing, HC for Hand Closing and HO for Hand Opening. Although not often explicitly mentioned, such control over remaining muscles in the residual forearm might be related to the "mobile phantom limb" phenomenon. Most amputees report having a phantom limb, i.e., the perception that their lost limb is still present and attached to their body [10], and many of them describe the capacity of voluntary mobilization of this phantom limb (85% of arm amputees, according to a recent epidemiological study we performed on mobile phantom limbs). Since phantom movements cannot be observed, they are often confounded with motor imagery of the lost limb. Yet, recent studies show that movement control of the phantom limb is different from motor imagery [11,12], an important argument being that, contrary to motor imagery, muscle activity is always present in the residual limb during phantom movement execution, with muscle activation patterns that are not random but seem to depend on the type of phantom movement [13,14]. This is not so much surprising for forearm amputees for whom residual muscles were already involved in finger and wrist movements before the amputation. Yet, it was also found for transhumeral amputees for whom the residual limb muscles had never been involved in finger and wrist movements before amputation. Therefore, decoding voluntary phantom hand movements from residual upper limbs differs from that from residual forearms. In the present study, we evaluated the possibility to decode phantom fingers activity, to offer upper arm amputees a dexterous control of polydigital prosthetics by using the sEMG activities recorded over the arm residual limb while mobilizing the phantom limb. Our approach is based on classical sEMG pattern recognition methods used in the present context to classify phantom finger and hand movements in order to have these phantom movements in real-time mimicked by a polydigital hand prosthesis. We were recently able to classify online phantom movements of elbow, wrist and hand based on the associated sEMG recordings from the residual upper-arm muscles, with an average successful classification rate over 80%, and to offer participants a real-time control over a graphical user interface through limb mobilization [15]. These preliminary results pushed us to believe that such approach could be transferred to the decoding of phantom finger activity, and used to a novel control approach of polydigital prostheses for the misconsidered upper arm amputees. their phantom A. Participants II. METHODS Three volunteers participated in the study (2 men, 1 woman, respectively 33, 62 and 77 years old), recruited at the Institute of Physical Medicine and Rehabilitation. All patients were healthy except that they had one unilateral transhumeral amputation following a traumatic accident. The brachial plexus of one participant (P2) was affected by the accident preceding the amputation. Before the accident all patients were right-handed. The inclusion criteria were a good control ability of the mobile phantom limb as well as repeatable phantom limb movement of a minimal 20° amplitude, an absence of phantom limb pain, and the availability of the participant during the recording period. The study was approved by the Ethical commission of the Institute of Physical Medicine and Rehabilitation, and performed in accordance with the Declaration of Helsinki. The participants provided written informed consent to take part in the study. The actual voluntary mobilization of the phantom limb was precisely defined using a questionnaire that intended to make clear distinctions between sensations in the residual limb, phantom pain, phantom sensations as well as between mobility of the residual limb and of the phantom limb. Patients' demographic and clinical data are reported in the Table 3. The participant (P2) with an affected brachial plexus pre-amputation suffered from a prosthetic fitting failure. The absence and reduced control over his biceps and triceps, respectively, required the prothesist to place the electrodes of trapezoidal his classical myoelectric prosthesis on muscles. All to perform voluntarily movements of thumb, index and major fingers along with whole hand opening and closing actions. Table 4 presents the list of the hand movements that each subject was able to perform. the three participants were able B. Experimental setup We used the BCI2000 software suite to develop the global control architecture on a desktop computer running Windows 7 (Intel Core i5-4690K (3.5 GHz) with 16 Go DDR3). BCI2000 is a general-purpose software suite designed for brain-computer interface (BCI) and was used here to run in parallel three principal modules: one acquisition driver to acquire the sEMG data (at a frequency of 1kHz), one Matlab© classification algorithm script (executed every 128ms), and one driver (C++) to control the Prensilia IH2 Azzura robotic hand through a serial port. An electrophysiological recording system (Eegosports, ANT-Neuro©, the Netherland) with 24 bipolar channels and 24-bit resolution was used to record sEMG activity from the participant's residual limb muscles at a sample rate of 1 kHz. Because of the variety in residual limb length and muscle anatomy, the scheme of electrode placement had to be adapted for each participant. Twelve pairs of active sEMG electrodes (Ag/AgCl snap bipolar electrodes with a 1.25cm- diameter circular contact) were used for each participant to record activity on various parts of the residual muscles where slight contraction could be felt by hand during phantom movements. Placement for each participant is shown in Figure 2. Twelve couples of electrodes were placed over the residual biceps, triceps, deltoid and sometimes trapezoidal and grand pectoralis muscles. No specific skin preparation was used before placing the active electrodes over the residual limb. The recorded sEMG signals were then filtered with a [10 Hz ; 400 Hz] third-order bandpass Butterworth filter. Figure 2. 12 electrode pairs were placed on the residual limb of each of the three participants. Two views for each participant. In this study, we used a linear discriminant analysis (LDA) classifier [16] running on the Matlab© Engine within the BCI Toolbox suite. An LDA classifier was chosen because it is based on a simple and robust statistical and thus computationally efficient approach. Features were computed from the sEMG using a 512-ms sliding analysis window with a 128-ms overlap between successive windows. The first 4 autoregressive coefficients (AR), the root mean square (RMS) value and the sample entropy of the sEMG were extracted for each channel and used to create the feature vector. The confidence value of the classifier was used to filter the algorithm output: if classification confidence was below 90%, no order was sent to the robotic hand. Figure 3. Control of all possible phantom movements (Test 1).A: Number of repetitions performed during Test 1 for each possible type of phantom movement and the associated ratio of successful classification (in black). The results are shown for each participant individually (P1-P3). B: Confusions matrices for each of the three participants of our multi-feature set for Test 1.The confusion matrix color scale increases from white to black as a function of increasing classification rate. Remark: participant P1 was not able to perform MF, RF, RE, PC, PO; P2: LF and LE; P3 could not perform TE. We use a right 5-DoF Prensilia IH2 Azzurra robotic hand. Position control of the fingers is used to generate a velocity- like control: each time a movement class is detected with sufficient confidence, an angular position displacement (5°) is sent to the concerned finger or group of fingers. A specific filter is added to avoid punctual changes in the classification output to be sent to the hand (a given movement has to be by the robotic hand). C. Protocol The experiments are organized in two testing series: Test 1 with a control of all possible phantom movements and Test 2 with only preferred movements. Each test consisted of two steps: first training of the classification algorithm, and then testing of its performances with online control of the polydigital robotic hand by the participant. The overall duration of the experimental session was approximately 2 hours, including the placement of electrodes. Both tests were videotaped. The patient was comfortably seated in a specific dedicated chair, fitted with armrests and a head rest. The robotic hand was attached to one of the arm-rests of the chair, and placed in a chosen position to be as much as possible symmetric with the intact arm lying on the other armrest. The first test started by verifying the previous description of their phantom hand movements and the effort required for their execution in order to determine the phantom movement testing order. Yet, when during the training phase of Test 1 the participant mentioned that he/she was able to perform an extra gesture (which happened sometimes), the latter was added to the catalog of possible movements, even if the participant was not fully sure of its repetition capability. Once the electrodes were placed, Test 1 started by a training phase during which the sEMG activities of the residual limb were recorded. The participants were asked to perform 2 repetitions of each possible phantom movement with a few minutes of rest in-between. The experimenter was in charge of verbally asking the subject to perform tasks and therefore was in charge of the rhythm of the performance. No instruction was given about the velocity of the gesture or its amplitude. Then, the recordings were manually tagged offline with the movement type (from Thumb Flexion TF to Hand Opening HO) by the experimenter through a dedicated interface. This did not last more than 5 minutes. Once these recordings were obtained, the data were analyzed and assembled to constitute the training set of data to be used in the testing phase (online classification of the phantom limb activity and real-time mimicking by the robotic hand). Similar training was performed at the beginning of Test 2 with only some preferred movements. On the average, the inter-test time delay was 10 to 15 minutes. During the testing phase, the participants were asked by the experimenter to perform a list of randomized movements of their phantom limb (all interspersed with pauses) while the classification algorithm was these to movements (thanks training parameters previously obtained) and to reproduce them in real-time with the robotic hand. trying the set of identify to III. RESULTS We performed experiments on three upper-arm amputees with voluntary motor control over their phantom limb. The specificity of our protocol was that the participants were neither trained nor used to mobilize their phantom limb (apart from P1 who reported regularly mobilizing his phantom little finger to release unpleasant contractions in the residual limb). This resulted sometimes in surprise, fatigue and tensions, and for-the-participants-unknown even mobilization capabilities in Methods). Therefore, the number and types of possible phantom movements varied between the participants (see [17] for a characterization of phantom movement kinematics). (see Protocol revealed some Twelve electrode pairs were placed over the residual limb (i.e., the upper arm). First, a very short training phase of the pattern recognition algorithm was done based on only two repetitions of each possible phantom movement (see Protocol in Methods).Then, for the subsequent control tests, a polydigital robotic hand was attached to the armrest of their chair at the approximate location of a normal prosthetic hand (see Fig. 1). The two tests consisted of the experimenter asking them to perform a pseudo-random series of given phantom movements among the catalog of their possible movements. In the first test (see Methods) the participants were asked to perform a pseudo-random sequence of phantom movements among their individually defined catalog of possible movements. Fig.3A shows, for each participant, the variation in the number of repetitions of each type of phantom movement as well as the associated ratio of successfully mimed movements by the robotic hand. When considering the whole range of phantom movements (i.e., 9-12 different movements), the obtained individual rates varied from 57.1 to 71.6% of successful reproductions. Yet, some movements being never correctly recognized by the classifier, practically, the participants were able to control 5 to 8 different gestures with this control approach. It can be seen that the catalog of possible movements as well as the number of repetitions for a given movement varied between the participants. Also, the rate of successfully mimed movements by the robotic hand was rather variable over the different types of movements (Fig. 3A). The size of the catalog of possible phantom limb movements is probably influencing the recognition by the classifier of the phantom limb actions. The bigger the catalog is, the more frequent the confusion between gestures can be for the pattern recognition algorithm (see Fig.3 B). The interesting result is that even if the distinction between finger actions is not perfectly clear, the distinction between flexion/closing and extension/opening is generally obtained. Figure 4. Control of preferred phantom movements (Test 2).A: Number of repetitions performed for each movement among the catalog of preferred movements, and associated ratio of classification success. B: Confusion matrices for each of the 3 participants for our multi-feature set for Test 2 (preferred /easiest movements). The confusion matrix color scale increases from white to black as a function of increasing classification rate. Since all phantom movements were not performed with the same facility, we conducted a second test with a reduced set of preferred/easiest movements (i.e., 4 to 6 gestures for each participant). The inter-individual variations between the repetition numbers of each movement are shown in Fig.4A along with the associated distribution of successfully mimed movements by the robotic hand. Confusion matrices between the phantom limb gestures performed by each subject (asked by the experimenter) and the movement mimed by the robotic hand are shown in Figure 4.B. This reduction in the size of the movement catalog to only the preferred movements led to a clear increase of the successful reproduction rates for two participants (P2 increasing from 57.1 to 92.3 % and P3 from 63.2 to 75 %, see Table 2). Figure 5. Classification performance as a function of the number of repetitions. This figure was created by only considering the movements that were performed a total of at least 5 times in the two tests. The color scale increases from white to black as a function of increasing classification rate. The global results suggest that two phenomena may interfere during the test: practice and fatigue effects. For all three participants, when the control of a phantom limb movement was easy for them, the sEMG activation pattern had a high probability to remain stable along time, and thus to be successfully mimed for several repetitions by the robotic hand. Indeed, the control of some movements was found to be improved across trials or at least stable among task repetitions (see Fig. 5). On the other hand, the activation patterns of phantom finger movements that hadn't been executed regularly and/or for a long time were more effortful and probably changed during the intensive testing series such that the classifier could not recognize the movement anymore. IV. DISCUSSION The first hypothesis, concerning the high control dimensionality allowed by this control method, is well supported by the present results. The control dimensionality of this phantom limb associated muscle activity is high since subjects were able to initiate 4 to 8 different movements, with a successful classification rate of 78%. Such approach is definitively promising for offering arm amputees a full and functional control over commercially available polydigital hand prosthetics, similarly to what is being studied in research laboratories amputees. Classification performances based on sEMG activity of forearm amputees are generally slightly higher (80%) [9], but the sEMG signals are recorded by arrays of electrodes placed over the partly preserved forearm muscle groups whereas in the present study we recorded over upper arm muscles. Moreover, the literature reports a reduced catalogue of 4 to 6 phantom wrist and hand (opening/closing) movements [18] and rarely of individual finger movements as we do in the present study. For upper- arm amputees who want to increase the number of active for forearm prosthetic DoF, our approach actually constitutes the only alternative to approaches requiring surgery like TMR. Indeed, similar to TMR, the mapping between the input signals and the output controlled joints is direct and intuitive: the phantom joint movement is simply mapped into similar joints on the robotic hand. Our second hypothesis was that the present prosthetic control approach (real-time mimicking of the phantom hand by the robot) is intuitive and doesn't need prior learning. This hypothesis is confirmed by our preliminary results showing the possibility to control, without any training or surgery, dexterous prosthetics based on sEMG muscle activities associated to voluntary phantom limb movements. It is fundamental to underline that the sEMG activities were measured over the residual limb, and thus on muscles that were not involved in wrist, hand and finger movements before the amputation. Furthermore, even if this approach relies on the measurement of sEMG signals, it does not require a fine control of the EMG activation levels but, instead, the control of a movement, which is much more natural for humans. Indeed, human motor control is dedicated to movement and force control rather than to selective control of individual muscle activity [19]. This is why even simple and generic myoelectric control approaches for hand prosthetic control based on bicep/triceps sEMG activations require long training in order for the patient to be able to generate single-muscle activation and avoid "automatic" co-activations. Although our control approach clearly does not need any prior learning, practicing phantom movements seems to stabilize the associated sEMG patterns, which eases the recognition by the classifier. This is indeed supported by the improvements for 2 participants when they performed their preferred movements that generally are the ones that they practice to relax their residual limb muscles or just "to play with the phantom". Thus, while patients do not have to learn a new mapping between muscle contractions and robotic degrees of freedom, regular practicing phantom movements might still the associated classification performances. Therefore, patients and the medical community should go beyond the usual skepticism concerning phantom movements and practice them regularly. their control and improve The third hypothesis was that, as long as the patient is able to mobilize the phantom limb, the condition of the residual limb as well as the age of participants are no limiting factors. Elderly persons often show a decline in physical and cognitive capacities, limiting their learning possibilities. Yet, encouraging results were obtained with the oldest participant (P3, 77 years, amputated for 14 years), which underline the simplicity of this intuitive control approach. Moreover, patients with affected plexus bracchi (P2 in our study) generally have to be fitted with conventional myoelectric prosthetics because of residual limb muscle paralysis and/or altered and weak sEMG signals. Yet, their partly-paralyzed residual limb still exhibits myoelectric activity patterns encoding the phantom limb movements. For instance, while P2 had been to control his conventional myoelectric hand prosthetics by activating his trapezoid muscles (due to the partial paralysis of the biceps and triceps muscles), he was able to intuitively control the polydigital hand by the sEMG activity recorded from his trouble trained biceps and triceps muscles, even if he couldn't contract them voluntarily other than by mobilizing his phantom limb. individual muscle activation The major limiting factor on the control dimensionality remains the controllability of the phantom limb: if only a very limited number of movements can be performed, the interest of our control approach is reduced. Nevertheless, executing a phantom movement remains more natural than learning to control levels. Moreover, training of the phantom movements could, in addition to reducing phantom limb pain (PLP) [20], help unlocking extra phantom limb movements, as we observed in the present study. So even with a reduced set of possible phantom limb movements, since patients tend generally to perform at least phantom hand opening and closing, this approach could even be a good alternative to control simple conventional 1 or 2 DoF myoelectric hands. In conclusion, the present results are very encouraging, especially since current solutions for transhumeral amputees who want to control a polydigital hand is to undergo either TMR surgery or an intensive training in order to learn an indirect mapping between muscle contractions and different prosthetic joint actions, and/or to explore a catalog of postures with co-contraction signals. The present intuitive control approach could deeply impact the use and appropriation of upper-limb prosthetics and might offer transhumeral amputees access to recent technological advances by: • offering new perspectives to amputees with already varied and selective voluntary control of their phantom limb: they can, without either surgery or heavy training and cognitive load, intuitively obtain a more selective control leading to an increased number of active prosthetic DoF, even of individual fingers; • simplifying the control for other patients: upper-arm amputees with reduced phantom limb mobility and fitted with simple prosthesis (opening/closing of the hand and active prono/supination of the wrist) could benefit from this control mode and obtain a more intuitive and non-sequential control. conventional myoelectric hand • offering amputees with a much deteriorated residual limb anatomy and function (affected brachial plexus) a supplementary chance fitted with myoelectric prostheses. to be Finally, these results could deeply change the way the phantom limb condition is apprehended by both patients and clinicians. Indeed, the phantom limb still makes both patients and clinicians feel uncomfortable [21], leading to completely ignored mobile phantom limbs [22]. So far, the mobility of the phantom limb has no use for rehabilitation of upper-arm amputee and is often considered by the amputee and/or the family as a sign of non-acceptance of the limb loss. Also, interference between phantom limb movements and conventional prosthesis control is often evoked. Yet, since the prostheses will be controlled by the phantom limb, this will clearly not be a problem anymore. And the important positive co-effect of the present approach will be the acceptance, and even the encouragement to consider as useful for daily life, the mobility of the phantom limb by patients, their family and rehabilitators. movements in upper limb amputees are slow but naturally controlled movements. Neuroscience (2015) [18] Smith, L. H. & Hargrove, L. J. Comparison of surface and intramuscular emg pattern recognition for simultaneous wrist/hand motion classification. In Engineering in Medicine and Biology Society (EMBC), 2013 35th Annual International Conference of the IEEE, 4223-4226 (IEEE, 2013). [19] Sanes, J. N., Donoghue, J. P., Thangaraj, V., Edelman, R. R. &Warach, S. Shared neural substrates controlling hand movements in human motor cortex. Science 268, 1775-1777 (1995) [20] Ortiz-Catalan, Max, et al. "Treatment of phantom limb pain (PLP) based on augmented reality and gaming controlled by myoelectric pattern recognition: a case study of a chronic PLP patient." Frontiers in neuroscience 8 (2014). [21] Crawford, Cassandra S. Phantom Limb: Amputation, Embodiment, and Prosthetic Technology. NYU Press, 2014. [22] Bouffard, Jason, et al. "Interactions between the phantom limb sensations, prosthesis use, and rehabilitation as seen by amputees and health professionals." JPO: Journal of Prosthetics and Orthotics 24.1 (2012): 25-33. V. CONCLUSION A conclusion section is not required. Although a conclusion may review the main points of the paper, do not replicate the abstract as the conclusion. A conclusion might elaborate on the importance of the work or suggest applications and extensions. ACKNOWLEDGMENT We warmly thank the participants who so willingly gave us several hours of their time. We are also very grateful to the medical staff and the Prosthetics Department of the IRR Nancy. This project was supported by the CNRS (project Subilmaof Défi AUTON 2016; PEPS INS2I JCJC MOFACO 2015), the ANR (project PhantoMovControl ANR-15-CE19- 0008-02), the Labex SMART (ANR-11-LABX-65) and by the Region Provence-Alpes-Cote d'Azur (Project ExplorAmp 2012, no 2012-07072). REFERENCES [1] Datta, D., Selvarajah, K. & Davey, N. Functional outcome of patients with proximal upper limb deficiency-acquired and congenital. Clinical rehabilitation 18, 172-177 (2004). [2] Plettenburg, D. H. Basic requirements for upper extremity prostheses: the wilmer approach. In Engineering in Medicine and Biology Society, 1998. Proceedings of the 20th Annual International Conference of the IEEE, vol. 5, 2276-2281 (1998). [3] Touch bionics web site. Mansfield (MA): TouchBionicsInc; 2015. URL http://www.touchbionics.com/ [4] Luchetti, Martina, et al. "Impact of michelangelo prosthetic hand: findings from a crossover longitudinal study." Journal of rehabilitation research and development 52.5 (2015): 605. [5] Belter, J. T., Segil, J. L., Dollar, A. M. & Weir, R. F. Mechanical design and performance specifications of anthropomorphic prosthetic hands: a review. JRehabil Res Dev 50, 599-618 (2013). [6] Kuiken, T. A. et al. Targeted reinnervation for enhanced prosthetic arm function in a woman with a proximal amputation: a case study. The Lancet 369, 371-380 (2007). [7] Hochberg, L. R. et al. Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature 442, 164-171 (2006). [8] Hakonen, M., Piitulainen, H. &Visala, A. Current state of digital signal related applications. processing Biomedical Signal Processing and Control 18, 334-359 (2015) in myoelectric interfaces and [9] Cipriani, C. et al. Online myoelectric control of a dexterous hand prosthesis by transradial amputees. Neural Systems and Rehabilitation Engineering, IEEE Transactions on 19, 260-270 (2011) [10] Ramachandran, V. S. &Hirstein, W. The perception of phantom limbs. the do hebb lecture. Brain 121, 1603-1630 (1998). [11] Raffin, E., Giraux, P. & Reilly, K. T. The moving phantom: motor execution or motor imagery? Cortex 48, 746-757 (2012). [12] Raffin, E., Mattout, J., Reilly, K. T. &Giraux, P. Disentangling motor execution from motor imagery with the phantom limb. Brain 135, 582- 595 (2012). [13] Gagne, M., Reilly, K., Hetu, S. & Mercier, C. Motor control over the phantom limb in above-elbow amputees and its relationship with phantom limb pain. Neuroscience 162, 78-86 (2009). [14] Reilly, K. T., Mercier, C., Schieber, M. H. &Sirigu, A. Persistent hand motor commands in the amputees' brain. Brain 129, 2211-2223 (2006). [15] Jarrassé, N, et al. "Classification of Phantom Finger, Hand, Wrist and Elbow Voluntary Gestures in Transhumeral Amputees with sEMG." IEEE Transactions on Neural Systems and Rehabilitation Engineering (2016). [16] Englehart, K., Hudgins, B., Parker, P. A. & Stevenson, M. Classification of the myoelectric signal using time-frequency based representations. Medical engineering & physics 21, 431-438 (1999). [17] De Graaf, J. B., Jarrassé, N., Nicol, C., Touillet, A., Coyle, T., Maynard, L., Martinet, N. and Paysant, J. Phantom hand and wrist
1706.06208
1
1706
2017-06-19T23:13:54
Using deep learning to reveal the neural code for images in primary visual cortex
[ "q-bio.NC", "cs.CV" ]
Primary visual cortex (V1) is the first stage of cortical image processing, and a major effort in systems neuroscience is devoted to understanding how it encodes information about visual stimuli. Within V1, many neurons respond selectively to edges of a given preferred orientation: these are known as simple or complex cells, and they are well-studied. Other neurons respond to localized center-surround image features. Still others respond selectively to certain image stimuli, but the specific features that excite them are unknown. Moreover, even for the simple and complex cells-- the best-understood V1 neurons-- it is challenging to predict how they will respond to natural image stimuli. Thus, there are important gaps in our understanding of how V1 encodes images. To fill this gap, we train deep convolutional neural networks to predict the firing rates of V1 neurons in response to natural image stimuli, and find that 15% of these neurons are within 10% of their theoretical limit of predictability. For these well predicted neurons, we invert the predictor network to identify the image features (receptive fields) that cause the V1 neurons to spike. In addition to those with previously-characterized receptive fields (Gabor wavelet and center-surround), we identify neurons that respond predictably to higher-level textural image features that are not localized to any particular region of the image.
q-bio.NC
q-bio
Using deep learning to reveal the neural code for images in primary visual cortex William F. Kindel1,*, Elijah D. Christensen1, and Joel Zylberberg1,2 1Department of Physiology and Biophysics, University of Colorado School of Medicine 2Learning in Machines and Brains Program, Canadian Institute for Advanced Research *[email protected] Abstract Primary visual cortex (V1) is the first stage of cortical image processing, and a major effort in systems neuroscience is devoted to understanding how it encodes information about visual stimuli. Within V1, many neurons respond selectively to edges of a given preferred orientation: these are known as simple or complex cells, and they are well-studied. Other neurons respond to localized center-surround image features. Still others respond selectively to certain image stimuli, but the specific features that excite them are unknown. Moreover, even for the simple and complex cells– the best-understood V1 neurons– it is challenging to predict how they will respond to natural image stimuli. Thus, there are important gaps in our understanding of how V1 encodes images. To fill this gap, we train deep convolutional neural networks to predict the firing rates of V1 neurons in response to natural image stimuli, and find that 15% of these neurons are within 10% of their theoretical limit of predictability. For these well predicted neurons, we invert the predictor network to identify the image features (receptive fields) that cause the V1 neurons to spike. In addition to those with previously-characterized receptive fields (Gabor wavelet and center-surround), we identify neurons that respond predictably to higher-level textural image features that are not localized to any particular region of the image. 1 Introduction Our ability to see arises because of the activity evoked in our brains as we view the world around us. Ever since Hubel and Wiesel mapped the flow of visual information from the retina to thalamus and then cortex, understanding how these different regions encode and process visual information has been a major focus of visual systems neuroscience. In the first cortical layer of visual processing– primary visual cortex (V1)– Hubel and Wiesel identified neurons that respond to oriented edges within image stimuli. These are called simple or complex cells, depending on how sensitive the neurons' responses are to shifts in the position of the edge [1]. The simple and complex cells are well studied [2, 3, 4]. However, many V1 neurons are neither simple nor complex cells, and the classical models of the simple and complex cells often fail to predict how those neurons will respond to naturalistic stimuli [5]. Thus, much of how V1 encodes visual information remains unknown. We use deep learning to address this longstanding problem. Recent advances in neural recording technology and machine learning have put solving the V1 neural code within reach. Experimental technology for simultaneously recording from large populations of neurons– such as multielectrode arrays– has opened the door to studying how the collective behavior of neurons encodes sensory information. Moreover, methods of machine learning, inspired by the anatomy of the mammalian visual system, known as convolutional neural networks, have achieved impressive success in increasingly difficult image classification tasks [6, 7]. Recently, these artificial Figure 1: Experimental data collection and processing. A) Neural activity was recorded in monkeys' V1 as they were shown a series of images. B) The image set contains 270 circularly cropped natural images. C) The response of a single neuron over repeated presentations of an image. Ticks indicate the neuron's spiking; each row corresponds to a different image presentation trial. During the response window, the firing rate is computed and then averaged over trials to yield the average response An,i used in our analysis. D) The neuron responds to image stimuli with a ∼ 50 ms latency from the image onset at t = 0, as seen in the peri-stimulus time histogram (firing rate plotted against time, averaged over all 270 images). neural networks have been used to study the visual system [8], setting the state of the art for predicting stimulus-evoked neural activity in the retina [9] and inferior temporal cortex (IT) [10]. Despite these successes, we have not yet achieved a full understanding of how V1 represents natural images. In this work, we present a convolutional neural network that predicts V1 activity patterns evoked by natural image stimuli. We use this network to predict the activity of 355 individual neurons in macaque monkey V1, and in doing so, this network represents the neural visual code for many neurons regardless of cell type. On held out validation data, the network predicts firing rates that are highly correlated (rmodel = 0.56 ± 0.02) with the neurons' actual firing rates. For 15% of these neurons, the firing rates are predicted to within 10% of the theoretical limit set by the trial-to-trial variation in the neural responses. To advance our understanding of the visual processing that takes place in V1, we invert the network to identify visual features that cause individual cells to spike. In the process, we identify novel functional cell types in monkey V1. 2 Methods 2.1 Experimental data We use publicly-available multielectrode recordings from macaque V1 [11]. In these experiments, macaque monkeys were anesthetized and then presented with a series of images, while the experi- menters simultaneously recorded the spiking activity of a population of neurons in V1 (Fig. 1A,B) with a multielectrode array. These recordings were conducted in 10 experimental sessions with 3 different animals, resulting in recordings from a total of 392 well-isolated neurons. A full description of the data and experimental methods can be found in Coen-Cagli et al. [11]. We use 37 of these neurons from one session to determine how we construct our network (its hyper-parameters), and the remaining 355 neurons to evaluate its performance. For each neuron n, we calculate the mean firing rate An,i evoked by each image i, by averaging its firing rate across the 20 repeated presentations of that image. The firing rates are calculated over a window from 50 to 100 ms after the image was presented, to account for the signal propagation delay from retina to V1. 2 V1ABCAn,iD Figure 2: The optimized architecture of the deep convolutional neural network model. The network's inputs are the pixel values of an image, and each output unit gives the predicted firing rate of a single neuron in monkey V1. We analyze the responses to 270 natural images that are circularly cropped to the size of the retinal visual field (Fig. 1B). The full dataset contains responses to natural and artificial stimuli both full- sized and cropped. We use only natural images because we are interested in the real-world behavior of the visual system, and we use only the cropped images because these images have the same visual field as the grating stimuli that we use to characterize the neurons as either orientation selective or not. Prior to training the neural network, we downsample the images using a non-overlapping 2 × 2 window and crop them to a size of 33 × 33 pixels. 2.2 Deep neural network model To construct our predictive network, we use a convolutional neural network whose input is an image and whose output is the predicted firing rates of every neuron in a given recording session. As shown in Fig. 2, the network consists of a series of linear-nonlinear layers. The first layer(s) performs local feature extraction on the image by sweeping banks of convolutional filters over the image, and then applying a maximum pooling operation. These local features are then globally combined at the all-to-all layer(s) to generate the predicted firing rate for every neuron in that data session. The number of each type of layer (convolutional with maximum pooling, or all-to-all), and the details about each layer (number of units, convolution stride, etc.), are optimized to maximize the accuracy of the neural activity predictions on the 37 neurons recorded in the second data session. We do this using a combination of manual and automated searches, where the results of our manual search inform the range of the hyper-parameter space for an automated random search [12]. Using the optimal parameters (Table 1) we train and evaluate our network with the remaining 9 datasets. We train our network using a cross-validation procedure where we randomly subdivide a given dataset into a training subset (80% of the images and corresponding V1 activity patterns) and an evaluation subset (20% of the images). We then train all layers of our network using the TensorFlow Python package with the gradient-descent optimizer. We attribute a loss i(yn,i − An,i)2 vari (An,i ) (cid:80) Ln = (1) to each neuron (indexed by n), where i is the image index, An,i is the measured response, and y is the network's predicted response. The neurons' losses are summed yielding the total loss used by the optimizer. To ensure the performance generalizes, the training data is subdivided into data used by the optimizer to train the weights, and another small subset (14% of the images) to stop the training when accuracy stops improving (early stopping). To quantify the performance of the predictor, we compare the network's predicted firing rates to the neurons' measured firing rates using a held-out evaluation set. This set is neither used to determine the hyper-parameters, nor to train the weights in our neural network. We calculate the Pearson correlation coefficient rmodel between the predicted firing rates and the measured rates, for each neuron. To enable comparison with other work, we also calculate how much of the variance in the 3 convolution and max pool layersinput: imageall-to-all layeroutput: firing ratesV1 neurons Table 1: Hyper-parameters of the optimized network. Conv. layer(s) All-to-all layer(s) 2 Dropout keep rate 1 0.55 Conv. 1 Conv. 2 All-to-all Num. filters Conv. kernel Maxpool stride Maxpool kernel 16 Num. filters 7 Conv. kernel 2 Maxpool stride 3 Maxpool kernel 32 Num. elem. 7 2 3 300 neurons' mean firing rates (over all test images) could be explained by the network's predictions: the fraction of variance explained (FVE). 2.3 Benchmarking the performance measures Because of the trial to trial variability in neural activity, no predictor could achieve rmodel = 1. To understand how well our network can predict the V1 neurons' firing rates, we compare its predictability to a theoretical maximum value set by the variability in the neural responses. To compute this maximum, we generate fake data by drawing random numbers from Gaussian distributions with the same statistics as the measured neural data. For each neuron and image, we average over 20 of these such values to obtain a simulated prediction. We then compute the correlation between these simulated predictions and the neurons' actual mean firing rates to find the maximum correlation rmax possible given the variability. 2.4 Characterizing the selectivity of cells To interpret the breadth of our results, we group the cells into functional classes, and look at how well the firing rates from each class can be predicted by our neural network model. We classify cells by how selective they are to specific natural images, and by their selectivity to specific orientations of grating stimuli. The selectivity of each neuron to specific natural images is quantified by the image selectivity index = , (2) (cid:18) N − ((cid:80) i Ai)2(cid:80) i(A2 i ) (cid:19) 1 N − 1 where Ai is the cell's firing rate indexed i over the set of N images [13]. This index has a value of zero for neurons that fire equally to all images, and a value of 1 for cells that spike in response to only one of the images. The orientation selectivity is measured by the cell's circular variance = 1 − (cid:80) (cid:80) θ Aθei2θ θ Aθ , (3) where Aθ is the neuron's firing rate in response to a grating oriented at angle θ. The circular variance is less sensitive to noise than the more commonly-used orientation selectivity index [14]. Following the results of Mazurek et al. [14], we use the thresholds: circular variance < 0.6 to define orientation selective cells (the simple and complex cells according to the Hubel and Wiesel convention), and circular variance > 0.75 for non-orientation-selective cells. We omit all other cells from these two groupings. 2.5 Identifying visual features that cause the neurons to spike To use our model as a tool to investigate the functions of individual neurons, we use DeepDream- like [15] techniques to identify the visual features that cause each cell to spike. We invert our network by finding input images that cause a given cell to spike at a pre-specified level. To do this, we first take the fully trained network, and set Gaussian white noise images as the input. We then use backpropagation to modify the pixel values of the input image to push the chosen neuron's predicted 4 Figure 3: The performance of the network model. A) A histogram of the Pearson correlation coefficient between the network predictions and the actual firing rates of all 355 neurons. These values are obtained using held-out data not used in training the network. B) The average performance of the predictor rmodel (+) and the linear model rlinear (o) for: 1) all neurons; 2) the orientation selective neurons; 3) non-orientation selective neurons. The limits on predictability rmax are shown in dashed lines. C) Scatter plot of how well the predictor can predict each neuron's firing rates (vertical axis) against the neuron's circular variance (horizontal axis). D) The predictability plotted against image selectivity index. For C and D, each data point corresponds to a single neuron. firing rate towards the pre-specified level. Thus, we find an input image that induces the pre-specified response. We applied this procedure to several different neurons, and at several different target firing rates. 3 Results With our optimal network (Table 1), the predicted firing rates are highly correlated with the measured neuron's firing rates for most neurons (Fig. 3A) when evaluated on held-out data. The correlation between the predictions and actual neural firing rates is rmodel = 0.56 ± 0.02 (FVEmodel = 0.36 ± 0.01) averaged over all 355 neurons in the evaluation set (Fig. 3B). Given the noise in the neural responses, a perfect predictor would achieve a correlation rmax = 0.851. Therefore, our predictor achieves 66% of the maximum possible performance. Moreover, for many neurons, the predictability approaches this theoretical maximum: for 54 of the 355 cells, the predictability was within 10% of the theoretical maximum. To show that our predictor performs complex, nonlinear processing on the images, we compare our results to a linear model that predicts firing rates based on each pixel value of the input images. The linear model's predictions have a correlation of rlinear = 0.008 ± 0.003 to the measured firing rates, which is substantially lower than the network model (Fig. 3B). Because simple and complex cells have been extensively studied, we are motivated to compare the predictability of simple and complex cells to the predictability of the other neurons in the dataset. Grouping the cells into orientation selective– simple and complex-like cells– and non- 5 Figure 4: Using the network model to reveal the visual features that drive individual neurons. For each neuron, we synthesize images that drove the predicted firing rates to the specified target values. These target firing rates are chosen to be different percentiles of the neurons' firing rate distribution. Cells A and B respond to localized image features, whereas cells C and D respond to more abstract image features. orientation-selective cells (see Methods), we find that our network predicts non-orientation-selective cell responses with rmodel = 0.49 ± 0.02, and orientation selective cell responses with rmodel = 0.60 ± 0.04 (Fig. 3B). Therefore, our model predicts the firing rates of both cell types, and performs slightly better on the simple and complex-like cells than the non-orientation-selective cells. Given that some neurons' firing rates are well predicted by the network while others are not, we are motivated to ask what distinguishes predictable from unpredictable cells. To answer this question, we quantify the cells' orientation selectivity, and their image selectivity (see Methods). Comparing the predictability of each cell's firing rates with its respective image selectivity index (Fig. 3D), and circular variance (Fig. 3C), we find that the predictability depends only weakly on these characteristics of the cells. Regardless of these values, some neurons' firing rates are well predicted while others are not. Thus, the orientation selectivity and image selectivity are only minor factors in determining how well our model performs. 3.1 Identifying visual features that cause the neurons to spike By construction, our convolutional neural network assumes nothing about the neurons' receptive fields. They are learned exclusively from the training data. Therefore, we can use our network to determine the response properties of both well-characterized neurons (simple and complex cells) and poorly understood neurons in an unbiased manner. To do this, we invert the network and identify visual features that evoke specified responses in several of the well predicted neurons. Based on the measured firing rate distribution, we repeat this procedure for different target firing rates, from low (at the 20th percentile of the neuron's firing rate distribution) to high (80th percentile). As seen in Fig. 4, this method allows us to classify several different types of cells, and identify novel response features. Cells A (rmodel = 0.88) and B (rmodel = 0.89) appear to function like previously characterized cells. Cell A responds to a center-surround image feature, and cell B's receptive field 6 center-surround Gaborwaveletabstractfeaturescell A cell Bcell C cell D is a Gabor wavelet. In contrast, cells C (rmodel = 0.91) and D (rmodel = 0.90) respond to more abstract image features that are not well represented by simple localized image masks. 4 Discussion We train a deep convolutional neural network to predict the firing rates of neurons in macaque V1 in response to natural image stimuli. We find that the firing rates of both orientation-selective, and non-orientation-selective neurons can be predicted with high accuracy. Moreover, we find that the network can identify the image features that cause the neurons to spike. This procedure reveals both canonical localized neural receptive fields (such as Gabor wavelets and center-surrounds), as well as abstract image features (previously uncharacterized in V1 neural receptive fields) that were not localized to a single region of the image. Our results have implications for developing new computer vision algorithms as well as studying the visual centers of the brain. 4.1 Comparisons to other work Studying visual processing in V1 we find that the optimal architecture of our convolutional neural network is relatively shallow compared to recent results by Yamins et al. [10]. Deep neural networks require large training datasets to generalize [16]. With only 270 images to train and evaluate our network, the optimal architecture that we find is likely a balance between the architecture that truly represents V1 encoding given infinite data and one that generalizes well. To probe whether a deeper network could more truly represent V1, a much larger dataset is required; this highlights the persistent limitation that small datasets impose on deciphering the neural code. Although, it is difficult to fairly compare the performances of published results for a variety of factors, we predict neural activity with performance that is comparable to the state of the art. Over all neurons, the correlation between our network predictions and the actual neural firing rates is rmodel = 0.56 ± 0.02. For comparison, Lau et al. [17] achieved predictability of r = 0.45 for simple cells and r = 0.31 for complex cells, and Prenger et al. [18] achieved r = 0.24 averaged over all cells. Lehky et al. [2] achieved r = 0.78 and Willmore et al. [19] achieved a predictability of FVE = 0.4. However, some contextual factors confound direct comparison to these results. Specifically, Lehky et al. selected neurons that are easier to predict by specifically choosing neurons that responded strongly to the presentation of bars of light, and Willmore et al. adjusted their image to match the respective field of each neuron they predicted. Despite methodology differences, we report comparable performance or better to recent published results. 4.2 Implications for machine learning While supervised learning methods lead to impressive performance on image categorization tasks [6, 7], the trained networks are easily fooled by imperceptible image manipulations [20], and they require large amounts of training data to achieve high levels of performance. Primates, however, are not so easily fooled. They can learn to perform classification tasks given only small numbers of training examples. Thus, it may be possible to improve the deep networks used for computer vision by building on the primate brain's representation. Deep neural networks – like the one presented here – could be pre-trained to predict primate V1 firing patterns, and subsequently trained to perform object recognition tasks. Thus, combining our approach with traditional learning could lead to more robust and data-efficient algorithms. 4.3 Implications for neuroscience and medicine By inverting our network (Fig. 4), we show that we can use the network as a tool to investigate the neurons' response properties. As demonstrated by distinguishing Gabor wavelet (cell B) from center-surround (cell A) receptive fields, this tool can identify and classify functional cell types. Going forward, this tool shows promise for characterizing the response properties of more cells in V1, and precisely defining functional cell types that were previously overlooked. Looking beyond V1, these methods could be applied to understanding higher level cortical processing, such as visual encoding in V2. By finding the features that elicit a response in V2 neurons, this tool could help fill the visual encoding knowledge gap [21] that exists between the abstract encoding of IT and V4 and the low-level encoding of the retina and V1. 7 Additionally, our results have taken a key step towards cracking the neural code for how visual stimuli are translated into neural activity in V1. This would be a major step forward in sensory neuroscience, and would enable new technologies that could restore sight to the blind. For example, cameras could continuously feed images into networks that would determine the precise V1 activity patterns that correspond to those images: a camera to brain translator. Brain stimulation methods like optogenetics [22] could then be used to generate those same activity patterns within the brain, thereby restoring sight. To master neural encoding, we propose closing the loop between conducting the experiments and performing analysis. By using the network to generate visual stimuli hypothesized to evoke particular patterns of neural activity, experiments could directly probe the neural code, and in doing so pave the way for a new class of neuroscience experiments. Acknowledgments We thank Adam Kohn and Ruben Coen-Cagli for providing the experimental data, and we thank Gidon Felson, John Thompson, Adam Kohn and Ruben Coen-Cagli for providing invaluable feedback. This research is supported by the National Institutes of Health under award numbers T15 LM009451 and T32 GM008497, the Canadian Institute for Advanced Research (CIFAR) Azrieli Global Scholar Award, and the Google Faculty Research Award. References [1] D. H. Hubel and T. N. Wiesel. Receptive fields of single neurones in the cat's striate cortex. The Journal of Physiology, 148(3):574–91, 1959. [2] S. R. Lehky, T. J. Sejnowski, and R. Desimone. Predicting responses of nonlinear neurons in monkey striate cortex to complex patterns. Journal of Neuroscience, 12(9), 1992. [3] Stephen V. David, William E. Vinje, and Jack L. Gallant. Natural Stimulus Statistics Alter the Receptive Field Structure of V1 Neurons. Journal of Neuroscience, 24(31), 2004. [4] Jorrit S. Montijn, Guido T. Meijer, Carien S. Lansink, and Cyriel M.A. Pennartz. Population- Level Neural Codes Are Robust to Single-Neuron Variability from a Multidimensional Coding Perspective. Cell Reports, 16(9):2486–2498, 2016. [5] Bruno A. Olshausen and David J. Field. How Close Are We to Understanding V1? Neural Computation, 17(8):1665–1699, 2005. [6] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. Advances In Neural Information Processing Systems, pages 1–9, 2012. [7] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015. [8] Daniel L. K. Yamins and James J. DiCarlo. Using goal-driven deep learning models to under- stand sensory cortex. Nature Neuroscience, 19(3):356–365, 2016. [9] 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 29, pages 1369–1377. 2016. [10] Daniel L. K. Yamins, Ha Hong, Charles F Cadieu, Ethan A. Solomon, Darren Seibert, and James J. DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences of the United States of America, 111(23):8619–24, 2014. [11] Ruben Coen-Cagli, Adam Kohn, and Odelia Schwartz. Flexible gating of contextual influences in natural vision. Nature Neuroscience, 18(11):1648–1655, 2015. [12] James Bergstra and Yoshua Bengio. Random Search for Hyper-Parameter Optimization. Journal of Machine Learning Research, 13:281–305, 2012. [13] Joel Zylberberg and Michael Robert DeWeese. Sparse coding models can exhibit decreasing sparseness while learning sparse codes for natural images. PLOS Computational Biology, 9(8):1–10, 2013. 8 [14] Mark Mazurek, Marisa Kager, and Stephen D. Van Hooser. Robust quantification of orientation selectivity and direction selectivity. Frontiers in Neural Circuits, 8:92, 2014. [15] Aravindh Mahendran and Andrea Vedaldi. Understanding Deep Image Representations by Inverting Them. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. [16] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15:1929–1958, 2014. [17] Brian Lau, Garrett B. Stanley, and Yang Dan. Computational subunits of visual cortical neurons revealed by artificial neural networks. Proceedings of the National Academy of Sciences of the United States of America, 99(13):8974–9, 2002. [18] Ryan Prenger, Michael C.-K. Wu, Stephen V. David, and Jack L. Gallant. Nonlinear V1 responses to natural scenes revealed by neural network analysis. Neural Networks, 17(5):663– 679, 2004. [19] Ben D. B. Willmore, Ryan J. Prenger, and Jack L. Gallant. Neural Representation of Natural Images in Visual Area V2. Journal of Neuroscience, 30(6):2102–2114, 2010. [20] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 427–436, 2015. [21] Corey M. Ziemba, Jeremy Freeman, J. Anthony Movshon, and Eero P. Simoncelli. Selectivity and tolerance for visual texture in macaque V2. Proceedings of the National Academy of Sciences, 113(22):E3140–E3149, 2016. [22] Baris N. Ozbay, Justin T. Losacco, Robert Cormack, Richard Weir, Victor M. Bright, Juliet T. Gopinath, Diego Restrepo, and Emily A. Gibson. Miniaturized fiber-coupled confocal fluores- cence microscope with an electrowetting variable focus lens using no moving parts. Optics Letters, 40(11):2553, 2015. 9
1604.01308
1
1604
2016-04-05T15:53:28
Natural movement with concurrent brain-computer interface control induces persistent dissociation of neural activity
[ "q-bio.NC" ]
As Brain-computer interface (BCI) technology develops it is likely it may be incorporated into protocols that complement and supplement existing movements of the user. Two possible scenarios for such a control could be: the increasing interest to control artificial supernumerary prosthetics, or in cases following brain injury where BCI can be incorporated alongside residual movements to recover ability. In this study we explore the extent to which the human motor cortex is able to concurrently control movements via a BCI and overtly executed movements. Crucially both movement types are driven from the same cortical site. With this we aim to dissociate the activity at this cortical site from the movements being made and instead allow the representation and control for the BCI to develop alongside motor cortex activity. We investigated both BCI performance and its effect on the movement evoked potentials originally associated with overt execution.
q-bio.NC
q-bio
1Imperial College London, Bioengineering, UK; 2Bernstein Centre & Faculty of Biology & BrainLinks-BrainTools, Univ. of Freiburg, Germany; 3Bioengineering, Ctr. for Sensorimotor Neural Eng., 4Dept. of Neurolog. Surgery, Ctr. for Sensorimotor Neural Eng., Univ. of Washington, USA. *Corresponding author LB: [email protected] Natural movement with concurrent brain-computer interface control induces persistent dissociation of neural activity Luke BASHFORD1, 2,*, Jing WU3, Devapratim SARMA3, Kelly COLLINS4, Jeff OJEMANN4, Carsten MEHRING2 Introduction As Brain-computer interface (BCI) technology develops it is likely it may be incorporated into protocols that complement and supplement existing movements of the user [1]. Two possible scenarios for such a control could be: the increasing interest to control artificial supernumerary prosthetics [2,3], or in cases following brain injury where BCI can be incorporated alongside residual movements to recover ability [4]. In this study we explore the extent to which the human motor cortex is able to concurrently control movements via a BCI and overtly executed movements. Crucially both movement types are driven from the same cortical site. With this we aim to dissociate the activity at this cortical site from the movements being made and instead allow the representation and control for the BCI to develop alongside motor cortex activity. We investigated both BCI performance and its effect on the movement evoked potentials originally associated with overt execution. Methods Patient's undergoing epilepsy monitoring with subdural Electrocorticography (ECoG) signed informed consent to participate in the studies. All experiments were approved by the ethics committees at the University of Washington and Seattle Children's Hospital. The experiment contained three parts; preceding the BCI task all subjects performed a 'pre' movement screening. Subjects then performed either a '1D BCI' task and/or the 'Concurrent BCI' task, followed by a 'post' movement screening identical to the 'pre'. We present data from 4 subjects; 1 who performed concurrent only, 1 who performed 1D only, and 2 who performed both on separate days. Pre/Post Screening: Subjects were given a visual cue to perform a repeated movement. The first cohort of subjects performed a force matching task, squeezing a force sensor to match a predefined value. The second cohort of subjects performed a self- paced tapping of the index finger. In each case subjects were cued to make the movement over 6s, repeated 20 times with a rest period of no movement lasting 2-4s (randomly selected to avoid anticipation) in between each trial. This task was performed identically before and after each BCI session to measure a baseline for movement only evoked activity, allowing us to determine changes caused by performing the BCI task. In an initial screening this task was used to select the single channel to control the BCI. The channel selected was that which had the largest 70-90Hz amplitude. The position of the electrode was also confirmed to be over motor cortex using MRI reconstructions of electrode locations where possible. 1D BCI: Brain activity was recorded using a TDT RZ5D processor and a PZ5 ADC (Tucker-Davis Technologies, Alachua, Florida, USA), or g.USBAmps (g.tec medical engineering GmbH, Schiedlberg, Austria) sampled at 1220 Hz and 1200 Hz, respectively. The vertical cursor movement was controlled by the normalized output of an auto- regressive filter using the spectral power in the 70-90Hz band of the recorded signal. Two targets were visible at the top and the bottom of a workspace. Subjects were instructed that movement imagery, with no overt execution, would move the cursor upward, while rest would move the cursor downward. Concurrent BCI: Subjects were required to move a cursor from the centre of the workspace to one of 8 targets presented around a circle. The vertical component of cursor movement was controlled via the BCI (identical to the 1D task above). The horizontal component of the movement was controlled by repeatedly pressing a computer keyboard key with the index finger contralateral to the BCI control electrode location. Continuous tapping was required otherwise the cursor would automatically move back towards the centre of the workspace at a constant rate. The cursor position was updated at 30Hz, at each update the cursor position moved (key press distance * number of presses) – return distance, where key press distance is 0.87% and return distance is 0.026% of the total workspace 0-1(AU). Subjects had to achieve a concurrent control to reach the off vertical targets. Subjects completed 48 trials, of equal presentation of the 8 targets. This run was repeated as many times as possible per session. Subjects had 10s to complete each trial before a time out, followed by a 4s inter trial interval. Results Subjects were able to gain control using our concurrent BCI task (Fig1A top). As the control site used was also activated by hand movement execution, an increased gamma power produced by the movement execution should drive the cursor upwards; this is evident in the example 'miss' trajectories (Fig1A bottom). In the successful trials therefore dissociation of the activity at the control electrode site from finger movements was required to move Fig1. A) Top; Cursor trajectories and individual subject target accuracy (row per subject) for 3 subjects during concurrent BCI. Bottom; Example 'miss' trajectories. B) STFT averaged over all successful trials of an exemplary session for one subject. the cursor in both vertical directions independently of the requirement to tap the finger for horizontal control (Fig1B). To validate that the movement execution and BCI control occurred independently, the Y position of the cursor (Fig2A), under BCI control, should not be influenced by movements. We made a best fit linear model between the change in Y position around the key presses and used this to recreate the cursor Y trajectories from the key presses (Fig2B), the recreated trajectories were then scaled up to match the endpoint standard deviation of the real trajectories. If the control was not independent we would expect a high correlation between real and fit trajectories, however we found a correlation within the chance distribution produced from a 10000 fold shuffle fit per run (Sub1; R=0.49, 92.1%ile. Sub2; R=0.78, 98.2%ile. Sub3; R=0.48, 77.6%ile. Fit and percentile in shuffle distribution respectively). Furthermore the fitted trajectories were not able to capture the endpoint separability seen in the real data. We fitted two thresholds to the data that maximised endpoint accuracy classification over a 10000 sample bootstrapping. The distribution of accuracies produced showed significantly lower endpoint accuracies in the model fit trajectories vs the real Fig2. Rows show 3 subjects (Sub1 – first row, etc.). Column A) Smoothed real cursor Y position with endpoint distribution. B) Modelled cursor Y position with endpoint distribution. C) Accuracy of endpoint separability based on fitting 2 thresholds. (Sub1 p<0.0005, Sub2&3 p<0.01) (Fig2C). We further hypothesised that the dissociation of the BCI control activity from the finger movement induces a persistent neural reorganisation. To investigate this hypothesis, we correlated the 70-90Hz (control frequency band) amplitude envelope of the neural signal between the control and other channels and computed the change of the correlation between 'pre' and 'post' screening. Our results showed a clear reduction in the correlation from 'pre' to 'post' between the control channel and adjacent channels (distance of 1cm away) compared to nonadjacent channels (distance >1cm) only for concurrent but not for 1D BCI sessions (pooled data p=0.01, p=0.3 respectively, Wilcoxon rank sum test, Holm- Bonferroni corrected) (Fig3A inlay). Comparing between the groups demonstrated a significant difference in correlation also at adjacent electrodes only (p=0.013) (Fig3A). This effect decreases with increasing distance. We further compared the signal correlation to adjacent channels across different frequency bands and show this effect was also unique to the control frequency band (p=0.013) (Fig3B). This suggests that it is not general BCI use but rather the concurrent control that induces a dissociation of neural activity specifically from the control site that persists beyond the concurrent control task. Discussion We demonstrated that human subjects are able to gain an independent and concurrent control of a BCI and overt movement execution. Fig3. A) Correlation change between 'pre' and 'post' Furthermore we revealed that this concurrent control, unlike typical BCI movement screening at the control frequency for only paradigms, enforces dissociation between the neural control signal electrodes at different distances from the control and the overt movement evoked neural activity. Moreover, we showed channel. Inlay shows pooled data within the group. that this dissociation of neural activity persists beyond the concurrent B) Correlation change for concurrent and 1D BCI at control task. We therefore propose that it reflects a change in the adjacent electrodes across different frequencies. cortical organisation indicating that a distinct mapping or representation Error bars show SEM. of BCI control can develop amongst concurrently active cortical areas. This provides a novel extension of previous ideas of map formation in BCI use [5]. The potential for concurrently controlled BCI and movement execution has until now only been demonstrated in primates [6,7], we extend this to the human case and give a novel demonstration of the physiological changes at the control site due to concurrent control. This framework demonstrates the potential for the control of BCIs in addition to natural movements, for example a future 'third arm' BCI control. References: [1] J. Wolpaw, et. al, Brain-Computer Interfaces: Something New under the Sun. In: Wolpaw JR, Wolpaw EW. Brain- Computer Interfaces: Principles and Practice, Oxford University Press, USA, 2012. [2] E. Abdi, et. al, PLoS ONE 10 (2015) 0134501 [3] F. Parietti, et. al, 2014 IEEE Int. Conf. Robot. Autom. ICRA, 2014, pp. 141–148. [4] C. Brunner, et. al, Brain-Comput. Interfaces 2 (2015) 1–10. [5] K. Ganguly, et. al, PLoS Biol. 7 (2009) e1000153–e1000153. [6] A.L. Orsborn, et. al, Neuron 82 (2014) 1380– 1393. [7] I. Milovanovic, et. al, Brain-Comput. Interfaces 0 (2015) 1–12. Acknowledgements: This work was funded by Bernstein Focus Neurotechnology Fr/Tu, BrainLinks-BrainTools Cluster of Excellence and Center for Sensorimotor Neural Engineering, University of Washington.
1204.0710
1
1204
2012-04-03T15:26:53
Optogenetic control of genetically-targeted pyramidal neuron activity in prefrontal cortex
[ "q-bio.NC", "q-bio.CB" ]
A salient feature of prefrontal cortex organization is the vast diversity of cell types that support the temporal integration of events required for sculpting future responses. A major obstacle in understanding the routing of information among prefrontal neuronal subtypes is the inability to manipulate the electrical activity of genetically defined cell types over behaviorally relevant timescales and activity patterns. To address these constraints, we present here a simple approach for selective activation of prefrontal excitatory neurons in both in vitro and in vivo preparations. Rat prelimbic pyramidal neurons were genetically targeted to express a light-activated nonselective cation channel, channelrhodopsin-2, or a light-driven inward chloride pump, halorhodopsin, which enabled them to be rapidly and reversibly activated or inhibited by pulses of light. These light responsive tools provide a spatially and temporally precise means of studying how different cell types contribute to information processing in cortical circuits. Our customized optrodes and optical commutators for in vivo recording allow for efficient light delivery and recording and can be requested at www.neuro-cloud.net/nature-precedings/baratta.
q-bio.NC
q-bio
! " # $ % & ' $ ( ) " * !"#$%&'("!'" Optogenetic control of genetically-targeted pyramidal neuron activity in prefrontal cortex Michael V. Baratta1, Shinya Nakamura1, Peter Dobelis1, Matthew B. Pomrenze1, Samuel D. Dolzani1 & Donald C. Cooper1* !"#$%&'()"*'$)+,'"-*".,'*,-()$%"/-,)'0"-,1$(&2$)&-(" &#")3'"4$#)"5&4',#&)6"-*"/'%%")6.'#")3$)"#+..-,)")3'")'7.-,$%" &()'1,$)&-(" -*" '4'()#" ,'8+&,'5" *-," #/+%.)&(1" *+)+,'" ,'#.-(#'#9" !"7$:-," -;#)$/%'" &(" +(5',#)$(5&(1" )3'" ,-+)&(1" -*" &(*-,7$)&-(" $7-(1" .,'*,-()$%" ('+,-($%" #+;)6.'#" &#" )3'" &($;&%&)6" )-" 7$(&.+%$)'" )3'" '%'/),&/$%" $/)&4&)6" -*" (cid:74)(cid:72)(cid:81)(cid:72)(cid:87)(cid:76)(cid:70)(cid:68)(cid:79)(cid:79)(cid:92)(cid:3) (cid:71)(cid:72)(cid:191)(cid:81)(cid:72)(cid:71)(cid:3) (cid:70)(cid:72)(cid:79)(cid:79)(cid:3) (cid:87)(cid:92)(cid:83)(cid:72)(cid:86)(cid:3) (cid:82)(cid:89)(cid:72)(cid:85)(cid:3) (cid:69)(cid:72)(cid:75)(cid:68)(cid:89)(cid:76)(cid:82)(cid:85)(cid:68)(cid:79)(cid:79)(cid:92)(cid:3) (cid:85)(cid:72)(cid:79)(cid:72)(cid:89)(cid:68)(cid:81)(cid:87)(cid:3) (cid:87)(cid:76)(cid:80)(cid:72)(cid:86)(cid:70)(cid:68)(cid:79)(cid:72)(cid:86)(cid:3) (cid:68)(cid:81)(cid:71)(cid:3) (cid:68)(cid:70)(cid:87)(cid:76)(cid:89)(cid:76)(cid:87)(cid:92)(cid:3) (cid:83)(cid:68)(cid:87)(cid:87)(cid:72)(cid:85)(cid:81)(cid:86)(cid:17)(cid:3) (cid:55)(cid:82)(cid:3) (cid:68)(cid:71)(cid:71)(cid:85)(cid:72)(cid:86)(cid:86)(cid:3) (cid:87)(cid:75)(cid:72)(cid:86)(cid:72)(cid:3) (cid:70)(cid:82)(cid:81)(cid:86)(cid:87)(cid:85)(cid:68)(cid:76)(cid:81)(cid:87)(cid:86)(cid:15)(cid:3) (cid:90)(cid:72)(cid:3) (cid:83)(cid:85)(cid:72)(cid:86)(cid:72)(cid:81)(cid:87)(cid:3) (cid:75)(cid:72)(cid:85)(cid:72)(cid:3) (cid:68)(cid:3) (cid:86)(cid:76)(cid:80)(cid:83)(cid:79)(cid:72)(cid:3) (cid:68)(cid:83)(cid:83)(cid:85)(cid:82)(cid:68)(cid:70)(cid:75)(cid:3) (cid:73)(cid:82)(cid:85)(cid:3) (cid:86)(cid:72)(cid:79)(cid:72)(cid:70)(cid:87)(cid:76)(cid:89)(cid:72)(cid:3) (cid:68)(cid:70)(cid:87)(cid:76)(cid:89)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3) (cid:68)(cid:81)(cid:71)(cid:3) (cid:86)(cid:76)(cid:79)(cid:72)(cid:81)(cid:70)(cid:76)(cid:81)(cid:74)(cid:3) (cid:82)(cid:73)(cid:3) (cid:86)(cid:83)(cid:72)(cid:70)(cid:76)(cid:191)(cid:70)(cid:3) (cid:83)(cid:82)(cid:83)(cid:88)(cid:79)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:86)(cid:3) -*" .,'*,-()$%" '0/&)$)-,6" ('+,-(#" &(" ;-)3" &("4&),-" $(5" &("4&4-" .,'.$,$)&-(#9" <$)" .,'%&7;&/" .6,$7&5$%" ('+,-(#" =','" 1'(')&/$%%6" )$,1')'5" )-" '0.,'##" $" %&13)>$/)&4$)'5" (-(#'%'/)&4'" /$)&-(" /3$(('%?" /3$(('%,3-5-.#&(>@?" -," $" %&13)>5,&4'(" &(=$,5"/3%-,&5'".+7.?"3$%-,3-5-.#&(?"=3&/3"'($;%'5" )3'7" )-";'" ,$.&5%6"$(5" ,'4',#&;%6"$/)&4$)'5" (cid:82)(cid:85)(cid:3) (cid:76)(cid:81)(cid:75)(cid:76)(cid:69)(cid:76)(cid:87)(cid:72)(cid:71)(cid:3) (cid:69)(cid:92)(cid:3) (cid:83)(cid:88)(cid:79)(cid:86)(cid:72)(cid:86)(cid:3) (cid:82)(cid:73)(cid:3) (cid:79)(cid:76)(cid:74)(cid:75)(cid:87)(cid:17)(cid:3) (cid:55)(cid:75)(cid:72)(cid:86)(cid:72)(cid:3) (cid:79)(cid:76)(cid:74)(cid:75)(cid:87)(cid:16)(cid:85)(cid:72)(cid:86)(cid:83)(cid:82)(cid:81)(cid:86)(cid:76)(cid:89)(cid:72)(cid:3) (cid:87)(cid:82)(cid:82)(cid:79)(cid:86)(cid:3) (cid:83)(cid:85)(cid:82)(cid:89)(cid:76)(cid:71)(cid:72)(cid:3) (cid:68)(cid:3) (cid:86)(cid:83)(cid:68)(cid:87)(cid:76)(cid:68)(cid:79)(cid:79)(cid:92)(cid:3) (cid:68)(cid:81)(cid:71)(cid:3) (cid:87)(cid:72)(cid:80)(cid:83)(cid:82)(cid:85)(cid:68)(cid:79)(cid:79)(cid:92)(cid:3) (cid:83)(cid:85)(cid:72)(cid:70)(cid:76)(cid:86)(cid:72)(cid:3)(cid:80)(cid:72)(cid:68)(cid:81)(cid:86)(cid:3) -*" #)+56&(1" 3-=" 5&**','()" /'%%" )6.'#" /-(),&;+)'" )-" &(*-,7$)&-(" .,-/'##&(1" &(" /-,)&/$%" /&,/+&)#9" A+," /+#)-7&2'5" -.),-5'#"$(5"-.)&/$%"/-77+)$)-,#"*-,"&("4&4-(cid:3)(cid:85)(cid:72)(cid:70)(cid:82)(cid:85)(cid:71)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:79)(cid:79)(cid:82)(cid:90)(cid:3)(cid:73)(cid:82)(cid:85)(cid:3)(cid:72)(cid:73)(cid:191)(cid:70)(cid:76)(cid:72)(cid:81)(cid:87)(cid:3)(cid:79)(cid:76)(cid:74)(cid:75)(cid:87)(cid:3)(cid:71)(cid:72)(cid:79)(cid:76)(cid:89)(cid:72)(cid:85)(cid:92)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:85)(cid:72)(cid:70)(cid:82)(cid:85)(cid:71)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:70)(cid:68)(cid:81)(cid:3) ;'",'8+'#)'5"$)"===9('+,->/%-+59(')B($)+,'>.,'/'5&(1#B;$,$))$9 membrane proteins for photocontrol of speci"c groups of neurons. We present here in vivo recordings from excitatory prelimbic cortex (PL) neurons transduced with either the light-activated cation channel, channelrhodopsin-2 (ChR2), or the light-driven chloride pump, halorhodopsin (NpHR) in rat. !" #" !e ability to activate or silence a speci"c cell type within a neural circuit in a temporally precise fashion is critical for understanding how the prefrontal cortex (PFC) processes the di#erent types of information underlying emotion and cognition. Experimental control over PFC activity has largely relied on traditional neuroscience loss-/gain-of-function tools (e.g., electrical stimulation, pharmacological modulation, surgical ablation) that do provide the required resolution for controlling speci"c populations of PFC neurons, either on a temporal or spatial scale. Although a temporally fast tool, electrical stimulation, even within a small volume of neural tissue, indiscriminately targets multiple classes of cells, "bers of passage, and terminals within the electrical "eld. In addition, the use of electrical stimulation does not allow for loss-of- function behavioral experiments as inhibition is not readily possible with this technique. Pharmacological manipulations, as well, lack cell-type speci"city and may exert their e#ects on a timescale that does re$ect behaviorally-relevant neuronal activity. Given that almost all cells in the PFC have some degree of overlapping electrical and pharmacological response pro"les, an ideal tool for regulating neuronal activity would utilize a physical stimulus that does not a#ect native neuronal function. Light is an ideal stimulus because most neurons do not express photoreceptors and many of its parameters (wavelength, intensity, temporal pattern, duration) can be brought under experimental control. tools !e recent development of optogenetic provides an exciting prospect for studying the complex and diverse functions of the PFC by enabling bidirectional control over the electrical activity of genetically-targeted cell populations with the use of light. Although it has been known for decades that certain microbial proteins react to light with ion $uxes, only recently have neuroscientists used genetic and viral manipulations to express these light-sensitive 1 Department of Psychology and Neuroscience, Institute for Behavioral Genetics, University of Colorado, Boulder, CO 80303 USA *Correspondence should be sent to [email protected] Page 1 of 2 ! ! ! ! Fig 1. Channelrhodopsin-2 (ChR2) induces temporally precise activity in rat prelimbic (PL) cortex. a) In vitro schematic (le!) showing blue light delivery and whole-cell patch-clamp recording of light-evoked activity from a $uorescent CaMKllα::ChR2-EYFP expressing PL pyramidal neuron (right) in an acute brain slice. b) In vivo schematic (le!) showing blue light (473 nm) delivery and single-unit recording. (bottom le!) Coronal brain slice showing expression of CaMKllα::ChR2-EYFP in the PL. Light blue arrow shows tip of the optical "ber ; black arrow shows tip of the recording electrode (le!). White bar, 100 microns. (bottom right) In vivo light recording of PL neuron in a transduced CaMKllα::ChR2-EYFP rat showing light-evoked spiking to 20 Hz delivery of blue light pulses (right). Inset, representative light-evoked single-unit response. "# Attribution 3.0 2011 Nature Precedings RESULTS Before recording from intact rodents, it was critical to con"rm that functional ChR2 could be expressed and in the targeted PL excitatory neurons. An adeno-associated viral (AAV) vector carrying the fusion protein ChR2-enhanced yellow $uorescent protein (ChR2-EYFP) under the control of the excitatory neuron-speci"c promoter CaMKllα. CaMKllα::ChR2-EYFP was stereotactically delivered to the rat PL, where brain slices of the transduced region were made 3 weeks later for whole-cell patch clamp experimentation. Optical stimuli were delivered from a multimode optical "ber (200 µm core diameter, 0.48 NA; !orlabs) coupled to a 100 mW blue laser. !e laser light was positioned directly above the recording site and maximal light intensity failed to produce any detectable postsynaptic responses in nontransduced tissue (data not shown). Under whole-cell current clamp, ChR2-expressing PL neurons exhibited high "delity spiking in response to illumination with blue light pulses (excitation wavelength, λex, was 473 nm). It should be noted that ChR2-expressing neurons (n=6) were able to reliably follow photostimulation trains up to 20 Hz with 100% "delity (Fig. 1a). !is was true for both blue (473 nm) and green (532 nm) laser light pulses. Fig 2. Halorhodopsin (NpHR) rapidly and reversibly silences spontaneous activity in vivo in rat prelimbic (PL) cortex. (Top le!) Schematic showing in vivo green (532 nm) light delivery and single- unit recording of a spontaneously active CaMKllα::eNpHR3.0- EYFP expressing PL pyramidal neuron. (Bottom) Example trace showing that continuous 532 nm illumination inhibits single-unit activity in vivo. Inset, representative single unit event; Green bar, (532 nM) is 10 s. *Note silenced responses similar to that depicted in Fig 2 were observed throughout the dorsal/ventral extent of the PFC (n=6). For in vivo recording, a tungsten electrode (~1.5 M) attached to an optical "ber (200 µm core diameter, 0.48 NA; center-to-center distance between electrode tip and optical "ber tip ~500 µm) that was coupled to either a blue (473 nm) or green (532 nm) diode laser was lowered into the PL to record single unit (n=6) activity in a head-"xed preparation. !e light intensity exiting the "ber was approximately ~200 mW/mm2. As with the in vitro recordings, in vivo delivery of blue light-pulse trains into the PL generated single-units that were capable of following trains of 20 Hz with perfect "delity (Fig. 1b). "# Attribution 3.0 2011 Nature Precedings ! " # $ % & ' $ ( ) " * in vivo photoinhibition of We next explored spontaneous activity in rat PL using an enhanced version of the light-gated third-generation chloride pump NpHR under the CaMKllα promoter (CaMKllα::eNpHR3.0-EYFP). In contrast to what was observed with ChR2, NpHR mediated complete silencing of spontaneous PL activity that was time-locked to the continuous light delivery (λex = 532 nm; Fig. 2). DISCUSSION !e application of optogenetics opens up an exciting prospect for "ne-scale functional analysis of rodent PFC circuitry in which the experimental manipulation is commensurate with the read-out measure (e.g., in vivo electrophysiology). Because these photosensitive proteins are genetically targetable, these tools can be used to study the causal function of a variety of de"ned cell types distributed within intact heterogenous tissue such as the PFC. METHODS All procedures were performed with the approval of the University of Colorado, Boulder Institutional Animal Care and Use Committee. For information on the adeno-associated viral (AAV) vectors used in this study visit www.neuro-cloud. net/nature-precedings/baratta. !e "nal viral concentration was 3 X 1012 infectious particles/mL. Viruses were delivered to adult Sprague-Dawley rat PL using a 10 µl syringe and a thin 33 gauge metal needle with a beveled tip (Hamilton Company). !e injection volume (1 µl) and $ow rate (0.1 µl/ min) were controlled with a microinjection pump (UMP3- 1; World Precision Instruments). Following injection, the needle was le( in place for an additional 10 minutes to allow for virus di#usion. For in vivo recording, a tungsten electrode (~1.5 M) attached to an optical "ber (200 m core diameter, 0.48 NA; center-to-center distance between electrode tip and optical "ber tip ~500 m) was coupled to either a blue (473 nm) or green (532 nm) diode laser. Custom optrodes and laser couplings can be requested www.neuro-cloud.net/ nature-precedings/baratta. For expanded methods visit www. neuro-cloud.net/nature-precedings/baratta. PROGRESS AND COLLABORATIONS To see up-to-date progress or if you are interested in collaborating with us visit www.neuro-cloud.net/nature-precedings/baratta ACKNOWLEDGEMENTS !is work was supported by National Institute on Drug Abuse grant R01-DA24040 (to D.C.C.), NIDA K award K-01DA017750 (to D.C.C.) AUTHOR CONTRIBUTIONS MVB, SN, PD, MBP, SDD, DCC performed the experiments, DCC, MVB, SN, PD designed the experiments, DCC and MVB wrote the manuscript. !"# $"#%"#&'()*+,#-"#./0+1,#$"#&023*41#*5#06",#705#7*84'9 :;<#=#>?@,#!ABC#>ADDE@" A"# -"#./0+1,#F"#G"#H0+1,#I"#&408+*4#*5#06",#70584*#JJB# >K!CB@,#BCC#>ADDK@ Page 2 of 2
1709.08591
3
1709
2018-10-17T15:50:50
Evaluating performance of neural codes in model neural communication networks
[ "q-bio.NC" ]
Information needs to be appropriately encoded to be reliably transmitted over physical media. Similarly, neurons have their own codes to convey information in the brain. Even though it is well-known that neurons exchange information using a pool of several protocols of spatio-temporal encodings, the suitability of each code and their performance as a function of network parameters and external stimuli is still one of the great mysteries in neuroscience. This paper sheds light on this by modeling small-size networks of chemically and electrically coupled Hindmarsh-Rose spiking neurons. We focus on a class of temporal and firing-rate codes that result from neurons' membrane-potentials and phases, and quantify numerically their performance estimating the Mutual Information Rate, aka the rate of information exchange. Our results suggest that the firing-rate and interspike-intervals codes are more robust to additive Gaussian white noise. In a network of four interconnected neurons and in the absence of such noise, pairs of neurons that have the largest rate of information exchange using the interspike-intervals and firing-rate codes are not adjacent in the network, whereas spike-timings and phase codes (temporal) promote large rate of information exchange for adjacent neurons. If that result would have been possible to extend to larger neural networks, it would suggest that small microcircuits would preferably exchange information using temporal codes (spike-timings and phase codes), whereas on the macroscopic scale, where there would be typically pairs of neurons not directly connected due to the brain's sparsity, firing-rate and interspike-intervals codes would be the most efficient codes.
q-bio.NC
q-bio
Evaluating performance of neural codes in model neural communication networks Chris G. Antonopoulos1, Ezequiel Bianco-Martinez2 and Murilo S. Baptista3 July 8, 2021 1Department of Mathematical Sciences, University of Essex, Wivenhoe Park, UK 2Data Science Studio - IBM Netherlands, Amsterdam, The Netherlands 3Department of Physics (ICSMB), University of Aberdeen, SUPA, Aberdeen, UK Abstract Information needs to be appropriately encoded to be reliably transmitted over physical media. Similarly, neurons have their own codes to convey information in the brain. Even though it is well-known that neurons exchange information using a pool of several protocols of spatio-temporal encodings, the suitability of each code and their performance as a function of network parameters and external stimuli is still one of the great mysteries in neuroscience. This paper sheds light on this by modeling small-size networks of chemically and electrically coupled Hindmarsh-Rose spiking neurons. We focus on a class of temporal and firing-rate codes that result from neurons' membrane-potentials and phases, and quantify numerically their performance estimating the Mutual Information Rate, aka the rate of information ex- change. Our results suggest that the firing-rate and interspike-intervals codes are more robust to additive Gaussian white noise. In a network of four interconnected neurons and in the absence of such noise, pairs of neurons that have the largest rate of information exchange using the interspike-intervals and firing-rate codes are not adjacent in the network, whereas spike-timings and phase codes (temporal) promote large rate of information exchange for ad- jacent neurons. If that result would have been possible to extend to larger neural networks, it would suggest that small microcircuits would preferably exchange information using temporal codes (spike-timings and phase codes), whereas on the macroscopic scale, where there would be typically pairs of neurons not directly connected due to the brain's sparsity, firing-rate and interspike-intervals codes would be the most efficient codes. 1 Introduction The main function of the brain is to process and represent information, and mediate decisions, behaviors and cognitive functions. The cerebral cortex is responsible for internal representations, maintained and used in decision making, memory, motor control, perception, and subjective ex- perience. Recent studies have shown that the adult human brain has about 86 × 109 neurons [1], which are connected to other neurons via as many as 1015 synaptic connections. Neurophysiology has shown that single neurons make small and understandable contributions to behavior [2, 3, 4]. However, most behaviors involve large numbers of neurons, which are often organized into brain regions, with nearby neurons having similar response properties, and are distributed over a number of anatomically different structures, such as the brain-stem, cerebellum, and cortex. Within each of these regions, there are different types of neurons with different connectivity-patterns and typical responses to inputs. The coexistence of segregation and integration in the brain is the origin of neural complexity [5]. Connectivity is essential for integrating the actions of individual neurons and for enabling cognitive processes, such as memory, attention, and perception. Neurons form a network of con- nections and communicate with each other mainly by transmitting action potentials, or spikes. To this end, the mechanism of spike-generation is well understood: spikes generate a change in the membrane potential of the target neuron, and when this potential surpasses a threshold, a spike 1 might be generated [6]. Brain regions show significant specialization with higher functions such as integration, abstract reasoning and consciousness, all emerging from interactions across distributed functional neural networks. At the local level, the function of individual neurons is relatively well understood. However, the full understanding of the information processing in networks of spiking neurons, the so-called "neural code", is still elusive. A neural code is a system of rules and mechanisms by which a signal carries information, with coding involving various brain structures. It is clear that neurons do not communicate only by the frequency of their spikes (i.e. by a rate code) [7], since part of the infor- mation can also be transmitted in the precise timing of individual spikes (i.e. temporal code) [8]. Also, it is known that some parts of the brain use rate codes (especially motor systems, matching to the slower muscles) and some use timing codes. In some cases, oscillations are very important (e.g. in sniffing), while in others may not be that much [9]. There is still a debate as to which neural code is used in which brain region, and how much of the potential timing, information is actually used [9]. Interactions at different timescales might be related to different types of processing, and thus, understanding information processing requires examining the temporal dynamics among neurons and their networks. Precise spike-timing would allow neurons to communicate more information than with random spikes. Different types of neural coding, including temporal and spatial coding, may also coexist on different time scales [10]. The scientific evidence so far supports the argument that we are still lacking full understanding on the codes used by neurons to carry and process information, as well as on which neural code is used in which brain region. What emerges from the scientific evidence so far suggests that fast systems and responses use fast spike-timings coding. For example, the human visual system has been shown to be capable of performing very fast classification [11], where a participating neuron can fire at most one spike. The speed by which auditory information is decoded, and even the generation of speech also suggest that most crucial neural systems of the human brain operate quite fast. For example, human fingertip sensory neurons were found to support this by demonstrating a remarkable precision in the time- to-first spikes from primary sensory neurons [12]. Thus, investigating the fundamental properties of neural coding in spiking neurons may allow for the interpretation of population activity and, for understanding better the limitations and abilities of neural computations. In this paper, we study neural coding and introduce four neural codes. We quantify and compare the rate of information exchange for each code in small-size networks of chemically and electrically coupled Hindmarsh-Rose (HR) spiking neurons [13, 14]. We do not deal with spatial codes, but only with temporal and firing-rate codes. For each neuron in the network, we record the temporal courses of its membrane-potential and phase. We construct a suitable map representation of these variables and compute the rate of information exchange for each pair of neurons, aka the Mutual Information Rate (MIR) [15], as a function of connectivity and synaptic intensities. We consider the precise spike-timings of neural activity (i.e. a temporal code), the maximum points of the phase of neural activities (i.e. neural phase), considering all oscillatory behaviors with arbitrary amplitude, including the high-frequency spiking and low-frequency bursting oscillations, the interspike intervals, and the firing-rate (i.e. ratio of spiking activity over a specific time interval). For the first three codes, we assume that all measurements are performed with respect to the ticks of a local master "clock" [16], meaning relative to the activity produced by one of the participating neurons in the network. This choice is arbitrary in the sense that the activity of any single neuron in the network can be used as the "clock". This allows for the estimated mutual information rates to reflect a measure between "synchronous" events that occur within a reasonable short-time window. Thus, our estimations provide the strength with which information is exchanged without any significant time-delay, and therefore reflecting a non-directional, non-causal estimation. In relation to the estimations for the MIR of the neural codes, it would be possible to refine them, considering finer spatial partitions, for example finer than the binary ones considered in this work. These refinements would correspond to the search for a generating, higher-order, Markov partition [17]. However, here, we study whether looking at the codes based on the interspike intervals would provide one with more information than the instantaneous spike-timings code. This is motivated by the question: in a time-series of events, what does carry more information? A code based on the times between events or a code based on the precise times of the occurrence of the events? We control our estimations by comparing them with a theoretical upper bound for MIR to verify the plausibility of the analysis. 2 Our main findings are summarized as follows: in the simplest case of a single pair of coupled HR spiking neurons, we find that they exchange the largest amount of information per unit of time when the neural code is based on the precise spike-timings. If observable (additive Gaussian white) noise is present, firing-rates are able to exchange larger rates of information than those based on temporal codes and together with the interspike-intervals code are the most robust to noise. In the case of four chemically and electrically coupled HR neurons as in Fig. 3, the largest rate of information exchange can be attributed to the neural codes of the maximum points of the phases (mod 2π, i.e. to a code dependent on the period of neurons' oscillations) and of the interspike intervals. Surprisingly, pairs of neurons with the largest rate of information exchange using the interspike-intervals and the firing-rate codes are not adjacent in the network, with the spike-timings and phase codes (temporal) promoting large rate of information exchange for adjacent neurons in the network. The latter is also backed by the results in Fig. 4 where connectivity (chemical and electrical connections) is swapped. These results provide evidence for the non-local character of firing-rate codes and local character of temporal codes in models of modular dynamical networks of spiking neurons. When neurons form a multiplex network of 20 HR neurons arranged in two equal-size modules in a bottleneck configuration, communication between pairs of neurons in the two modules is mostly efficient when using either the spike-timings or the maximum points of their phases codes. 2 Materials and Methods 2.1 The Hindmarsh-Rose Neural Model We simulate the dynamics of each "neuron" by a single Hindmarsh-Rose neuron system. Namely, following [13, 14], we endow each node (i.e. neuron) in the network with the dynamics [18] p = q − ap3 + bp2 − n + Iext, q = c − dp2 − q, n = r[s(p − p0) − n], (1) where p is the membrane potential, q the fast ion current (either Na+ or K+), and n the slow ion current (for example Ca2+). The parameters a, b, c, d, which model the function of the fast ion channels, and s, p0 are given by a = 1, b = 3, c = 1, d = 5, s = 4 and p0 = −8/5, respectively. Parameter r, which modulates the slow ion channels of the system, is set to 0.005, and the external current Iext that enters each neuron is fixed to 3.25. For simplicity, all neurons are submitted to the same external current Iext. For these values, each neuron can exhibit chaotic behavior and the solution to p(t) exhibits typical multi-scale chaos characterized by spiking and bursting activity, which is consistent with the membrane potential observed in experiments made with single neurons in vitro [18]. We couple the HR system (1) and create an undirected dynamical network (DN) of Nn neurons connected by electrical (linear diffusive) and chemical (nonlinear) synapses [13] pi = qi − ap3 i + bp2 i − ni − gn(pi − Vsyn) BijS(pj) Nn(cid:88) Nn(cid:88) j=1 − gl GijH(pj) + Iext, i − qi, qi = c − dp2 ni = r[s(pi − p0) − ni], φi = qipi − piqi p2 i + q2 i , i = 1, . . . , Nn, j=1 (2) where φi is the instantaneous angular frequency of the i-th neuron [19, 20], φi is the phase defined 3 by the fast variables (pi, qi) of the i-th neuron, H(p) = p and [13] S(p) = 1 1 + e−λ(p−θsyn) . (3) Our work intents to study the transmission of information in models of small-size neural networks by treating them as communication systems, for which we measure and evaluate the rates at which information is exchanged among neurons. We are also not interested in considering realistic biological models for the function S. Instead, we consider a biologically inspired function S of a sigmoid type as in Eq. (3) [13]. The remaining parameters θsyn = −0.25, λ = 10, and Vsyn = 2 are chosen so as to yield an excitatory DN [13]. The synaptic coupling behaves as a short delta- function and carries other features required for synaptic coupling. Particularly, Vsyn can be tuned to reproduce excitatory or inhibitory behavior, and S(p) has θsyn to allow for the disconnection of pre-synaptic neurons that have not reached an activation level. For λ = 10, S(p) is a continuous, sigmoid function that behaves similarly to a "binary process", either 0 or 1, a fundamental property necessary to use in analytical works when networks with neurons connected simultaneously by electrical and chemical means become synchronous [13]. Naturally, this is mimicking the democratic fashion with which chemical synapses behave, where a large community of activated pre-synaptic neurons needs to be activated to induce a relevant response in the post-synaptic neuron(s). This further allows us to study the system knowing the domain of parameters for which we can obtain oscillatory behavior. The choices for the couplings and network topologies in this work are purely abstract, not guided by realistic physiological reasons, and so does time t in Eqs. (1) and (2). The parameters gn and gl denote the coupling strength of the chemical and electrical synapses, respectively. The chemical coupling is nonlinear and its functionality is described by the sigmoid function S(p), which acts as a continuous mechanism for the activation and deactivation of the chemical synapses. For the chosen parameters, pi < 2, with (pi − Vsyn) being always negative for excitatory networks. If two neurons are connected via an excitatory synapse, then if the pre- synaptic neuron spikes, it might trigger the post-synaptic neuron to spike. We adopt only excitatory chemical synapses here. G accounts for the way neurons are electrically (diffusively) coupled and is represented by a Laplacian matrix [13] G = K − A, matrix of A, leading to(cid:80)Nn and, therefore, its diagonal elements are equal to 0, thus(cid:80)Nn where A is the binary adjacency matrix of the electrical connections and K the degree identity j=1 Gij = 0 as Gii = Kii and Gij = −Aij for i (cid:54)= j. By binary we mean that if there is a connection between two neurons, then the entry of the matrix is 1, otherwise it is 0. B is a binary adjacency matrix and describes how neurons are chemically connected [13] j=1 Bij = ki, where ki is the degree of the i-th neuron. ki represents the number of chemical links that neuron i receives from all other j neurons in the network. A positive off-diagonal value in both matrices in row i and column j means that neuron i perturbs neuron j with an intensity given by glGij (electrical diffusive coupling) or by gnBij (chemical excitatory coupling). Therefore, the adjacency matrices C are given by (4) (5) i , qi = −7.32183132+ For each neuron i, we use the following initial conditions: pi = −1.30784489+ηr i , ni = 3.35299859 + ηr ηr i is a uniformly distributed random number in [0, 0.5] for all i = 1, . . . , Nn (see [14] for details). These initial conditions place the trajectory on the attractor of the dynamics quickly, reducing thus the computational time in the simulations. i and φi = 0, where ηr C = A + B. 2.2 Numerical Simulations and Upper Bound for MIR We have integrated numerically Eqs. (2) using Euler's first order method with time-step δt = 0.01 to reduce the numerical complexity and CPU time to feasible levels. A preliminary comparison for trajectories computed for the same parameters (i.e. δt, initial conditions, etc.) using integration methods of order 2, 3 and 4 (e.g. the Runge-Kutta method) produced similar results. The numerical (2) was performed for a total integration time of tf = 107 units and the integration of Eqs. computation of the various quantities were computed after a transient time tt = 300 to make sure 4 that orbits have converged to an attractor of the dynamics. Thus, the sample size used in the estimation of the MIR for the various neural codes is large enough and amounts to 999,970,000 data points (excluding the transient period that corresponds to the first 30000 points). After Shannon's pioneering work [21] on information, it became clear [22, 23] that it is a very useful and important concept as it can measure the amount of uncertainty an observer has about a random event and thus provides a measure of how unpredictable it is. Another concept related to Shannon entropy that can characterize random complex systems is Mutual Information (MI) [21], a measure of how much uncertainty one has about a state variable after observing another state variable in the system. In [24], the authors have derived an upper bound for the MIR between two nodes or groups of nodes of a complex dynamical network that depends on the two largest Lyapunov exponents l1 and l2 of the subspace formed by the dynamics of the pair of nodes. Particularly, they have shown that MIR ≤ Ic = l1 − l2, l1 ≥ l2, (6) where l1, l2 are the two finite-time and -size Lyapunov exponents calculated in the 2-dimensional observation space of the dynamics of the pair of nodes [24, 25]. Typically, l1, l2 approach the two largest Lyapunov exponents λ1, λ2 of the dynamics of the DN (2) if the network is connected and the time to calculate l1, l2 is sufficiently small [24]. Ic is an upper bound for MIR between any pair of neurons in the network, where MIR is measured in the 2-dimensional observation space. It can be estimated using mainly two approaches: (i) the expansion rates between any pair of neurons, taking the maximal value among all measurements [24], which is tricky and difficult to compute [24] and (ii) the two largest positive Lyapunov exponents λ1, λ2 of the DN. Here, we use the second approach, since the equations of motion of the dynamics are available (Eqs. (2)) and the full spectrum of Lyapunov exponents can be calculated [26]. Particularly, we estimate Ic by Ic = λ1 − λ2 (assuming that l1 ≈ λ1 and l2 ≈ λ2) which will stand for an approximation to the upper bound for the MIR in the network. The phase spaces of the dynamical systems associated to the DNs are multi-dimensional and thus, estimating an upper bound for MIR using λ1 and λ2, reduces considerably the complexity of the calculations. Besides, parameter changes that cause positive or negative changes in MIR are reflected in the upper bound Ic with the same proportion [24]. 2.3 Estimation of MIR for Maps There is a huge body of work on the estimation of MI in dynamical systems and in neuroscience, for example [24, 27, 28, 29, 30, 31, 32, 23, 17]. In this work, we follow the method introduced recently in [33], to estimate MI between pairs of time-series X(t) and Y (t). This methodology is the same one used in [17] to estimate MI using refinements or generating Markov partitions. Below, we explain how we estimate MIR by using the estimated values for MI. Pairs X(t) and Y (t) can represent a mapping of any two variables, as in the neural codes introduced in this work. Particularly, we estimate MI by considering binary symbolic dynamics that encode each time-series X(t) and Y (t) into the symbolic trajectory represented by (α, β). N sequentially mapped points of X(t) and Y (t) are encoded into the symbolic sequences α = α1, α2, α3, . . . , αN and β = β1, β2, β3, . . . , βN , each composed of N elements. The encoding is done by firstly normalizing the time-series X(t) and Y (t) to fit the unit interval. Both αi and βi can assume only two values, either "0", if the value is smaller than 0.5 , or "1", otherwise. The Mutual Information, MI(L), between X(t) and Y (t) is thus estimated by the MI between the two symbolic sequences α and β by (cid:88) (cid:88) k l P (X(L)α k , Y (L)β l ) log MIXY (L) = k , Y (L)β l ) k )P (Y (L)β l ) , P (X(L)α P (X(L)α (7) k , Y (L)β where P (X(L)α l ) is the joint probability between symbolic sequences of length L observed simultaneously in α and β, and P (X(L)α l ) are the marginal probabilities of symbolic sequences of length L in the sequences α and β, respectively. The subindices k and l vary from 1 up to the number of symbolic sequences of different lengths L observed in α and β, respectively. k ) and P (X(L)β 5 MIR is estimated by the slope of the curve of the MI for symbolic sequences of length L ∈ [2, 5] with respect to L, which amounts to grid sizes of smaller and smaller cells as L increases. Consequently, this MIR can be considered as an estimation of the increase of MI per time interval. More details can be found in [33, 15]. The L interval considers sequences starting from L = 2 to L = 5 bits. The reason is that MI behaves linearly with L in this interval, allowing the calculation of MIR. Particularly, the built-in correlations in the time-series and the fact that the chosen partition is likely not the best possible, suggests the exclusion of L = 1, since correlations would start appearing for larger symbolic sequences. Also, we do not consider L > 5 as we would run into numerical problems and would introduce under-sampling effects because of the time-series length. For example, if we would assume L = 5, the analysis for 2 neurons would effectively deal with symbolic sequences of length 2L = 10 and there would be 210 different sequences of length L = 10. A significant trajectory length would then have to be larger than 10 210 = 10240 trajectory points. Due to the ergodic property of chaotic systems, the probability of observing a given symbolic sequence of length L in one neuron and another of the same length in another neuron is equivalent to the probability of finding trajectory points in a cell of the phase space. The larger L is, the smaller the cell is, and thus, it contains more information about the state of the pair of neurons. Our MIRii estimator in Subsec. 2.4.3 for the interspike intervals is similar to the work in [34]. In [34], the authors encode the time-signal by making a time partition, where temporal bins are defined, and a binary encoding is done, by associating 0's to bins without spiking and 1's to bins with spikes. Our encodings in Sec. 2.4 are based on partitions of the space created by the two time-series X(t) and Y (t). In our approach, we have not sought to maximize MI and search for the generating Markov partition as in [17]. However, we have dealt with biasing, when we compare our MIR estimations of the neural codes with Ic estimated by the difference of the two maximal Lyapunov exponents λ1 and λ2. All our MIR estimations in Sec. 2.4 are bounded by the mathematical upper bound Ic for MIR, except for three cases on which we elaborate later. It is worth it to note that the parameters and initial conditions in Eqs. (2) give rise to chaotic behavior with positive Lyapunov exponents. Thus, chaos is responsible for generating the probabil- ities necessary for the estimation of MIR in the 2-dimensional spaces of the data from the encoding of the trajectories of pairs of neurons [33]. Chaotic behavior in turn gives rise to uncertainty and production of information. Information is then transmitted through the various nodes in the neural network through the electrical and chemical connections (see Eqs. (2)). 2.4 Neural Codes Here, we introduce four neural codes and their methodologies to quantify the rate of information exchange between pairs of neurons. The first uses the spike-timings of neural activity (temporal code), the second the maximum points of the phase of neural activities (neural phase), the third the interspike intervals and the fourth, the firing-rates (ratio of spiking activity over a specific time interval). For the first three, we assume that all recordings are done with respect to the ticks of a local master "clock" [16], relative to the activity produced by a single neuron. This choice can be arbitrary in the sense that the activity of any single neuron can be used. The purpose is to obtain MIR values that can be interpreted as being the current rate of information exchanged between any two neurons, and not any time-delay mutual information. For the estimation of the MIR of the neural codes, we integrate numerically the system of Eqs. (2) as discussed in Subsec. 2.2 to obtain the numerical solutions to pi, qi, ni, φi, i = 1, . . . , Nn as a function of time. We then use these solutions (time-series) to construct the pairs of time-series X(t) and Y (t) to estimate the MIR for each particular neural code as explained below. Our coupling and topology choices are abstract and inspired by current research in multilayer networks [35, 36, 37, 38]. We seek to study whether looking at the instantaneous spike-timings provides less information than the codes based on the interspike intervals. Particularly, our study is motivated by the question: having a time-series of events, what does carry more information? A code based on the times between events, or a code based on the exact times of the occurrence of the events? 6 2.4.1 Neural Code Based on spike-timings: MIRst Here, we explain how we estimate the amount of information exchanged per unit of time between neurons i, j based on the spike-timings of the first neuron, MIRst, where st stands for spike-timings. Particularly, we assume that the first neuron plays the role of the "clock" and we record pi, pj from Eqs. (2) at times when p1 of the first neuron attains its local maxima. This allows us to construct a time-series of events Xi(t), Yj(t) by transforming the continuous dynamics of variables pi, pj into a time-series of discrete-time spike events Xi, Yj. We then use Xi, Yj to compute the rate of information exchanged between neurons i and j as explained in Subsec. 2.3. We divide the rate of information exchanged by the mean of the interspike times of the spike activity of the first neuron. We call this quantity, MIRst of the pair of neurons i, j. 2.4.2 Neural Code Based on Phase: MIRmφ Next, we explain how we estimate the amount of information exchanged per unit of time between neurons i, j based on the maximum points of the time evolution of the phase variables φi, φj, what we denote by MIRmφ, where mφ stands for maximum phase φ. We assume that the first neuron plays the role of the "clock" and record in time Φi ≡ mod(φi, 2π) and Φj ≡ mod(φj, 2π) from Eqs. (2) at times when Φ1 of the first neuron attains local maxima as a function of time t. This allows us to construct a time-series of events Xi(t), Yj(t) by transforming the continuous dynamics of the phase variables of both neurons into a time-series of discrete time events Xi, Yj. We then use Xi, Yj to estimate the rate of information exchanged between neurons i, j. We divide the rate of information exchanged by the mean of the time intervals for Φ1 of the first neuron to attain its local maxima. We call this quantity, MIRmφ of pair i, j. 2.4.3 Neural Code Based on interspike Intervals: MIRii Here, we show how we estimate the amount of information exchanged per unit of time between neurons i, j based on the interspike intervals of their pi, pj variables, denoted by MIRii, where ii stands for interspike intervals. Each neuron can produce a series of different interspike intervals in the course of time. When measuring MI between two interacting systems, we need to specify two correlated relevant events occurring at roughly the same time, if no time-delays are to be considered. These events need to match, i.e., one event happening for one neuron needs to be correlated to one event happening to the other neuron. In order to relate two such time-series with matched pairs of events, we introduce the notion of a relative "clock". An interspike interval in neuron i will be matched to the interspike interval of neuron j, if neuron j spikes after neuron i. Notice that by doing this, we neglect several spikes happening for both neurons, however, we produce a discrete two-dimensional variable that is meaningfully correlated, and therefore, producing a meaningful MI. Another cumbersome approach would be to find an appropriate time-interval within which the two neurons spike, and then correlate their spike-timings intervals or a method as complicated as to calculate the MI considering interspike intervals occurring at different time-delays. This analysis would be more complicated than the one adopted in this work, and would not be necessary. This allows us to construct a time-series of interspike events Xi(t), Yj(t) from the continuous trajectories of both neurons. We then use Xi, Yj to compute the rate of information exchanged between neurons i and j, dividing this by the mean of the time intervals constructed as the difference between the spike-timings of neuron j and those of neuron i, given that the spike of neuron j occurred after that of neuron i. We call this quantity, MIRii of pair i, j. 2.4.4 Neural Code Based on firing-rates: MIRf r Lastly, we show how we estimate the amount of information exchanged per unit of time between neurons i, j for the firing-rates of the pi, pj variables, MIRf r, where f r stands for firing-rate. Here, we divide the time window between the first and last recorded spike-timing of neuron i into 1.5×106 equal-size time windows, and compute the firing-rates for both neurons in these time windows. By firing-rate, we mean the ratio between the number of spikes in a given time interval divided by the length of the time interval. This allows us to construct a time-series of firing-rate events Xi(t), Yj(t). We then use these time-series to compute the rate of information exchanged between 7 neurons i, j, dividing it by the length of the equal-size time windows. We call this quantity, MIRf r of pair i, j. 2.5 The link between interspike-intervals and firing-rate codes time window t. Now, t =(cid:80)N represent the interspike intervals. Defining the average interspike interval by (cid:104)τ(cid:105) =(cid:80)N There is a link between interspike-intervals and firing-rate codes that goes back to Kac's lemma [39], which relates return-time intervals, first Poincar´e returns of the trajectory recurring to a region in phase space, with the probability measure of the trajectory returning to a region in phase space. The firing rate is calculated by f = N/t, where N denotes the number of spikes in the i τi, where τi represents the first Poincar´e returns, which could also i τi/N , one can see that f = 1/(cid:104) τ(cid:105). The last equation relates firing rates with the average spike times, though in a statistical sense. 3 Results 3.1 Neural Codes for the Communication of Two Neurons We study the four neural codes introduced, in the simplest case of a pair of chemically and bidirec- tionally connected HR neurons (see Fig. 1a), in the absence of noise as we consider its effect in the next section. Our goal is to understand which neural code can maximize the rate of information exchange between the two neurons, considering them as a communication system. We are also interested in finding the chemical coupling strengths gn this is happening. This is motivated by the question how would two neurons exchange information when disconnected from a network, acting as a single pair. We note that we are not interested in the directionality of the information flow but only in the rate of information exchanged pair-wise. Particularly, in Fig. 1, we calculate the amount of MI per unit of time exchanged between the two neurons, aka their MIR, for the four neural codes and for different chemical coupling strengths gn. Before we proceed with a detailed analysis, we summarize the main results of this figure as we increase the synaptic coupling strength from 0.1 (panels c-f), to 0.48 (panels g-j) and to 1 (panels k-n). The time-series of the membrane potential is not synchronous for gn = 0.1, it becomes strongly synchronous for gn = 0.48, and weakly synchronous for gn = 1. The scenario is similar to the amplitudes of the spikes, in panels d, h and l. There is no localization in panel d for gn = 0.1, total localization in h for gn = 0.48, and partial localization in panel l. In fact, Fig. 1l is the most interesting, as it shows there is a low-dimensional attractor associated to the membrane potential, something usually observed when coupled systems are generalized-synchronous. This scenario changes considerably when observing the behavior of the phases (panels e, i, and m) and the spike-timings (panels f, j, and n). Even though, there is weak phase-synchronization in panel e for gn = 0.1, there is no apparent synchronization in panels i and m for gn = 0.48 and gn = 1, respectively. Figure 1c shows that the membrane potentials (p1 and p2) are mainly asynchronous in time with epochs (time intervals) of synchronicity already visible in this small interval depicted in the panel. This is indeed happening in the full time-series of the numerical simulations (not shown in Fig. 1) and is depicted in Fig. 1e for the phases which are defined in Eqs. (2) as a function of p1, p2, q1, q2 and their derivatives. However, the interspike intervals become weakly synchronous in panels f and n, for gn = 0.1 and gn = 1, respectively, and strongly synchronous in j for gn = 0.48. These differences in the intensity of how neurons exchange information will be explored further in the following. We first see in Fig. 1b that MIRst and MIRmφ are bigger than MIRii and MIRf r in certain regions of intermediate and large enough chemical coupling strengths gn. Almost all MIR quantities are smaller than the upper bound for MIR, Ic, except for three chemical coupling strengths, ranging from smaller to larger values. This intriguing result is due to the fact that when calculating Ic using the Lyapunov exponents λ1 and λ2 as discussed previously, Ic is an approximation to the real upper bound for MIR. Consequently, when comparing this upper bound to any lower bound estimations for the MIR, as the neural codes in our work, it might happen these lower bounds be larger than the estimated upper bounds. We note that for gn values larger than about 1.3, the 8 dynamics becomes quasi-periodic and thus, there is no production of information. The reason is that in such cases, the largest Lyapunov exponent of the dynamics is negative, and consequently there is no chaos, but quasi-periodic behavior which gives rise to predictability, lack of uncertainty and thus no production of information. In Fig. 1b, we present MIR based on the spike-timings of both neurons (MIRst), the MIR for the maximum values of the phase (MIRmφ), the MIR of the interspike intervals (MIRii) and the MIR of the firing-rates (MIRf r). We also plot the upper bound for MIR, i.e. Ic = λ1 − λ2 [24]. We focus on three characteristic cases: the first corresponds to the case where MIRmφ >MIRst for chemical coupling strength gn = 0.1. The second, to a case where MIRii > Ic, MIRmφ and MIRst for gn = 0.48 (one of the three distinct cases where the computed MIR is bigger than the upper bound Ic), and the last one to a case where MIRst >MIRmφ for gn = 1. In the first case, the two neurons communicate more efficiently by exchanging larger amounts of information per unit of time using their phases whereas in the third case by exchanging information by the precise spike-timings. In the second case, the two neurons communicate more efficiently by encoding their information in their interspike activity. To appreciate the performance of the four codes, we first focus on the case of gn = 0.1 for which MIRmφ >MIRst. We plot in Fig. 1c the time evolution of p1, p2, the data used to compute MIRst in panel d, the plane of the phase variables of both neurons (Φ1, Φ2) in panel e, in which the computation of MIRmφ is based, and in panel f, the data used to compute MIRii, where τi, i = 1, 2 are the interspike intervals of both neurons. We observe in panel c that the spike times of both neurons are different. Fig. 1c shows that the membrane potentials of the two neurons (p1 and p2) are mainly asynchronous in time with epochs (time intervals) of synchronised activity already visible in this small interval depicted in the panel. This is characterized by the two neurons having a phase shift of about π. The displacement between the phases causes a time shift in the spike trains, thus making time-discrete variables such as the spike-timings asynchronous. The phase reflects the continuous oscillatory behavior of the trajectory, and is connected to the zero Lyapunov exponent. The spike trains reflect its timely character, and is connected to the positive Lyapunov exponent. In fact, both spike trains and phases are asynchronous. However, when discretization filters out the continuous oscillatory behavior of the trajectory, such as those producing the spike-timings, it is expected that asynchronous behavior is more noticeable from the timing variables due to its connection to the positive Lyapunov exponents. Particularly, panel d shows that when the first neuron spikes, the second usually remains silent as there is a high density of p2 values around -1, with spikes occurring around p2 ≈ 1.9. This behavior is due to the second neuron which is actually in its quiescent period when the first is spiking. In contrast, when observing the plane of phases in panel e, it becomes apparent that there are two regions of high phase-synchronicity (i.e. stripes of high concentration) and the rest of the region with considerably smaller concentration of phase points. This behavior indicates that the two neurons communicate by chaotically adapting their phases. For the same gn, panel f indicates that the interspike activity of both neurons is well spread in the plane with a high concentration of points occurring close to the origin. Moreover, MIRf r is seen to attain the smaller value with respect to all other quantities. In panels g to j we study the second case, for gn = 0.48, for which MIRii > Ic. We note that this apparent violation comes about because we estimate Ic by the Lyapunov exponents and not by the expansion rates. Since MIR is estimated by a mesh grid of finite resolution, an upper bound for MIR calculated for this grid would require the calculation of expansion rates using the very same grid resolution. Ic estimated by Lyapunov exponents is smaller than the bound estimated by expansion rates (see Supplementary material in [24]). Therefore, Ic in this case could not be a true upper bound for MIR. Here, we also observe that MIRst >MIRmφ (see panel b), a result that shows that the two neurons communicate mostly by exchanging information by their precise spike-timings and less by their phases. This can be appreciated in panel g where both p variables attain approximately similar amplitudes during their time evolution. It becomes evident in panel h where the second neuron spikes when the first neuron spikes and that both attain approximately the same amplitudes in their p time-evolution. This behavior is highly localized. In contrast, panel i shows that their phases actually spread all over [0, 2π] × [0, 2π] and that there is no localization of points as it happened for gn = 0.1 in which the two neurons communicate by exchanging the largest amount of information per unit of time by their phases. Here, panel j indicates that the 9 Figure 1: Results for the neural communication channel and the code used between two chemically, bidirectionally connected, non-noisy HR neurons. Panel a: the pair of chemically connected neurons, where gn is the strength of the chemical coupling. Panel b: Ic, the MIR of spike-timings MIRst, MIR of the maxima of the phases MIRmφ, MIR of the interspike intervals MIRii and MIR of the firing-rates MIRf r, respectively. Panels c to f: p1, p2 as a function of time in panel c, the plane of phase variables (Φ1, Φ2) (panel d) and, the data used to compute MIRii (panel f), where τi, i = 1, 2 are the interspike intervals of both neurons. Panels g to j: similarly for gn = 0.48 and panels k to n for gn = 1. In panel b, gn = 0.1 that corresponds to a case where MIRmφ >MIRst, gn = 0.48 to a case where MIRii > Ic, MIRmφ and MIRst, and the case for gn = 1 that corresponds to MIRst >MIRmφ. 10 interspike activity of both neurons is well localized in two regions with high concentration closer to the origin and on the right upper part of the plot. Moreover, MIRf r is seen to attain the smallest value for this particular chemical coupling strength. Finally, we focus on the third characteristic case in which MIRst >MIRmφ for gn = 1. The situation here is quite different. Indeed, panel k reveals a phenomenon in which the spike times and quiescent periods of both neurons are actually similar. Particularly, panel l reveals that most of the times, either when the first neuron spikes, the second spikes or when the first is in its quiescent period, so is the second, showing a higher density of points in the upper right corner of the plot (spike activity) and a smaller one in its lower left corner (quiescent period). In contrast, the plane of phases in panel m reveals there is no phase synchronization in their activity, as there are no dense regions as in the first case in which MIRmφ >MIRst. These results show that the two neurons communicate by their spike-timings, i.e. they use a temporal neural code in which the time of each spike conveys information that is transmitted to other neurons. Lastly, panel n exhibits an interspike activity mostly concentrated in the lower left corner of the plot and less in the other three, a situation completely different to the behavior in panel j of the second case. MIRf r is seen here to attain the smallest value, similarly to the first case. 3.2 Neural Codes for the Communication Between Two Noisy Neurons We now study the same problem in the presence of noise. We consider the effect of additive Gaussian white noise in the performance of the neural codes introduced in Subsec. 2.4. We want to understand which neural code is more robust to the increase of the noise strength σ, a case which is more close to realistic neural behavior [40, 41]. Particularly, in the neural activity of variable p of each neuron, we add white Gaussian noise with standard deviation σ to obtain its noisy signal ¯p: (8) where N (0, 1) is the Gaussian distribution of zero mean and standard deviation equal to 1. We then use such noisy data to estimate the MIR of the different neural codes for different chemical coupling strengths and noise strengths σ. The dynamics is chaotic and comes from the deterministic system in Eqs. (2). ¯p = p + σN (0, 1), We demonstrate these results in Fig. 2. Particularly, we plot the MIR between the two neurons in Fig. 1a, for different chemical couplings and three noise strengths. Figure 2a shows the same MIR quantities of Fig. 1b but for σ = 0.4, panel b for σ = 0.8 and panel c for σ = 1.5. As σ increases from zero, all MIR quantities start decreasing, except MIRf r and MIRii, which remain practically unaffected by the increase of noise strength. Figure 2 reveals that even though for small noise strengths, MIRst and MIRmφ are larger than MIRf r, they are nevertheless considerably affected by the increase of the noise strength. As we demonstrate in panels d, e and f, MIRf r and MIRii prove to be consistently robust with respect to the increase of σ, even for values as high as 1.5. This underlines the importance of firing-rate against temporal codes, such as spike-timings or phase codes, which prove to be prone to noise contamination and to the transmission of smaller amounts of information per unit of time with the increase of noise strength. Comparing Fig. 1b and Fig. 2, it can be seen that in the presence of Gaussian additive noise (8), the various MIR quantities drop below Ic around gn = 0.48. Also, the region where MIRst >MIRmφ disappears. Particularly, with the noisy strength increasing, MIRf r and MIRii became dominant as they are larger than MIRmφ and MIRst, respectively, except for some singular values. Our findings suggest that the firing-rate and interspike-intervals codes are more robust to readout noise. 3.3 Neural Codes in a Communication System of Four Neurons Here, we extend our study to the case of four bidirectionally connected non-noisy HR neurons, which are chemically and electrically coupled as shown in Fig. 3a. The first neuron is chemically connected with the third, whereas the first with the second and, the third with the fourth, are electrically connected. The strengths of the electrical and chemical connections are given by gl and gn, respectively. The four neurons in Fig. 3a are arranged in a typical configuration when one wants to infer topology from information-theoretical quantities. The open-ring topology offers 11 Figure 2: Results for the neural code used between two chemically connected noisy HR neurons. Panel a: the MIR values of the different neural codes for noise strength σ = 0.4. Panel b is for σ = 0.8 and panel c for σ = 1.5. Panels d, e and f are similar to a, b and c for MIRii and MIRf r only and same σ = 0.4 (panel d), σ = 0.8 (panel e) and σ = 1.5 (panel f) noise strengths as in the first three panels. We also plot Ic in all panels to guide the eye. Notice that MIRii and MIRf r in panels d, e and f, remain unaffected by the increase of the noise strength. 12 a way to test whether adjacent (non-adjacent) neurons share higher (lower) rates of information exchange. We aim to understand which neural code is best suited for the maximization of the rate of information-exchange for different coupling strengths and also, for which pairs this is so. The first and third neurons are the intermediates that facilitate the communication between the second and fourth. We consider the setup of Fig. 3a as a communication system in which, information is transmitted through the connections and reaches the neurons. In what follows and for each pair of coupling strengths, we estimate the MIR of the neural codes, and for each of them, we find its maximum MIR value and the corresponding pair of neurons that produces it. Then, for each coupling pair, we plot that maximum value and the corresponding pair of neurons. Figure 3: Topology and parameter spaces for the neural codes for four, non-noisy, HR neurons connected by 2 electrical and 1 chemical connection. Panel a: the network of connections of the four neurons, where gn, gl are the strengths of the chemical and electrical couplings, respectively. Panels b and c: the parameter spaces for MIRst for the two nodes that provide the largest MIR value and for the links that maximizes it, respectively. Panels d and e: similarly for MIRmφ. Panels f and g: similalry for MIRii. Panels h and i: similarly for MIRf r. In all cases, the notation i ↔ j indicates the bidirectional transfer of information between neurons i, j. In the following, we study the four neural codes for the model of four non-noisy neurons in Fig. 3a. In panels b and c, we plot the parameter spaces (gn, gl) for the MIRst of the spike-timings and 13 for the links that maximize it, respectively. The orange spots in panel b correspond to couplings that produce the largest amounts of MIRst whereas blue to regions with the smallest MIRst. The former occurs for relatively big chemical and electrical couplings whereas the latter for very small electrical and, small to large chemical couplings. Panel c reveals that, depending on the couplings, the largest amounts of MIRst are transmitted between different pairs of neurons, giving rise to a complicated pattern in the parameter space (see Fig. 3). The pattern is mainly characterized by the pair of neurons 3,4 (red) for small chemical and small to large electrical coupling strengths, by pair 1,2 (black) for comparatively small to large chemical and small to large electrical coupling strengths, and by many smaller-sized regions of different colors, such as blue, magenta, green and yellow that correspond to the remaining pairs of neurons. We decided to use as a "clock" the first neuron as it is one of the two mediators that facilitate the transmission of information in this network (with the other one being neuron 3). This choice however is relative in the sense that for every pair of neurons we want to estimate MI, we should choose a "clock". In this sense, there is no universal "clock", but several ones can be used. This choice also intends to maximize the amount of MI measured between any two pairs of neurons. The parameter space for MIRmφ in panel d is mainly dominated by red (that corresponds to comparatively large values), a smaller blue region of moderately very low values and a smaller orange region, for high chemical and electrical couplings, that corresponds to the highest observed MIRmφ values in the parameter space. Similarly to panel c (for MIRst), panel e for the pairs of neurons that maximize MIRmφ shows that, depending on the coupling values, the largest amounts of MIRmφ are transmitted between different pairs of neurons, giving rise to a complicated pattern in the parameter space, dominated mainly by the pair of neurons 3,4 (red) for small chemical and small to large electrical coupling strengths, by pair 1,2 (black) for comparatively small to large chemical and small to large electrical coupling strengths, and by many smaller-sized regions of different colors, (i.e. blue, magenta, green and yellow) that correspond to the remaining pairs of neurons. The situation changes slightly in panel f for MIRii where almost all the parameter space is dominated by red (of moderately large MIRii values) with a few orange spots (very large values) and blue spots (of very low MIRii values). The blue regions are considerably smaller in size than the blue region in panel d. The case for MIRii is also different with respect to the pairs of neurons for which it is maximal. The parameter space in panel g reveals completely different structural properties than in panels c and e. Interestingly, the largest amounts of MIRii occur for all pairs except 3,4 and, less for 1,2, implying that the first and third neurons play mainly the role of the facilitators in the transmission of information in the system. A similar situation is happening for MIRf r, with the parameter space in panel h looking uni- formly covered by red of moderately high MIRf r values and with a few quite small blue spots of very low values. MIRf r is less dependent on the coupling strengths. The parameter space for the links that maximize this quantity looks quite similar to that of MIRii, in the sense that the largest amounts of MIRf r occur for all pairs except 3,4 (red) and, less between 1 and 2 (black). This implies again that the first and third neurons play mainly the role of the facilitators in the transmission of information in the system. A comparison of the parameter spaces in Fig. 3 shows that the highest rate of information ex- change can be attributed to the neural codes of the maximum points of the phase MIRmφ and to the interspike intervals, MIRii. Moreover, MIRf r is practically unaffected by the coupling strengths, even though its maximum values are smaller than the maximum values of the neural codes based on the maximum points of the phase and interspike intervals. This result is in agreement with its performance in the case of the two neurons in Sec. 3.1, where it attained the lowest values of all other codes. Interestingly, the pair of nodes more likely to exchange the largest amount of information per unit of time using the interspike-intervals and firing-rate codes are not adjacent in the network, whereas the spike-timings and the phase codes promote large exchange of infor- mation from adjacent nodes in the network. This provides evidence for the non-local character of firing-rate codes and local character of precise, spike-timings, codes. The latter result on the character of the codes is also backed by the results in Fig. 4 where the role of chemical and electrical connections has been swapped (compare Figs. 3a and 4a). In particular, comparing panels c, e and g, i in Fig. 4, one can again deduce that temporal codes (MIRmφ and MIRii) perform optimally for adjacent neurons in the network whereas MIRf r and 14 MIRii for non-adjacent neurons. It becomes thus clear that the type of neural code with largest information transmission rate depends on network adjacency. Figure 4: Topology and parameter spaces for the neural codes for four, non-noisy, HR neurons connected by 2 chemical and 1 electrical connection. Panel a: the network of connections of the four neurons, where gn, gl are the strengths of the chemical and electrical couplings, respectively. Panels b and c: the parameter spaces for MIRst for the two nodes that provide the largest MIR value and for the links that maximizes it, respectively. Panels d and e: similarly for MIRmφ. Panels f and g: similalry for MIRii. Panels h and i: similarly for MIRf r. In all cases, the notation i ↔ j indicates the bidirectional transfer of information between neurons i and j. 3.4 Neural Codes in a Network of Twenty Neurons in a Bottleneck Configuration Finally, we study the neural codes in an extended model of two identical clusters of 10 HR, non- noisy, neurons each. For simplicity, both clusters have the same small-world structure [44] and their neurons are internally coupled with electrical connections of strength gl. This construction is interesting as it resembles a bottleneck, in which the two clusters communicate via the only link between the first and the eleventh neuron in the two clusters. The bottleneck is represented by a 15 single, chemical link with strength gn that connects the two clusters. We used one interconnection as this is the simplest case in which information travels from one cluster to the other through the only chemical link. Moreover, it allows to draw interesting conclusions with regard to the neural codes for different coupling strengths. The topology in Fig. 5a is an example of how two neural networks would interact via a connection which implements a bottleneck. Again, the network is undirected and for each pair of coupling strengths, we estimate the MIR of the four neural codes. For each code, we find its maximum MIR and the corresponding pair of neurons that produces it. Then, for each coupling pair, we plot the maximum value. The network in Fig. 5a is motivated by the modular organization of the brain in which neurons are linked together to perform certain tasks and cognitive functions, such as pattern recognition, data processing, etc. Modular processors have to be sufficiently isolated and dynamically differentiated to achieve independent computations, but also globally connected to be integrated in coherent functions [45, 46, 4]. The structure in Fig. 5a helps us understand which neural code in modular neural networks is best suited for the transmission of the largest amount of information per unit of time and for which coupling strengths it occurs. Again, we treat the model in Fig. 5a as a communication system in which, information is transmitted through the links and reaches out to its different parts. Figure 5: Topology with a bottleneck configuration and parameter spaces for the neural codes between two identical small-world, chemically connected, non-noisy clusters. Panel a: the two identical clusters of electrically connected neurons with coupling strength gl and chemical strength gn. Panel b: the parameter space for MIRst, Panel c: similarly for MIRmφ, Panel d: similarly for MIRii and panel e: for MIRf r. The colors indicate the maximal MIR value that any two nodes exchange using a particular neural code. In Fig. 5a, we study the four neural codes. Panels b and c show the parameter space (gn, gl) for MIRst and MIRmφ, respectively. Orange corresponds to couplings that produce the largest amounts of MIR values whereas blue or black to regions with the smallest values. Red is for intermediate MIR values. Panel b is for MIRst and reveals that the highest values can be achieved for large chemical and intermediate electrical coupling strengths. For example, for zero chemical 16 coupling (i.e. gn = 0), MIRst is considerably smaller than for gn around 1.4. This underlines the importance of chemical connections among the clusters as they help the system transmit larger rates of information when neurons exchange information by the precise spike-timings (temporal code). In contrast, MIRmφ seems to perform more consistently in the sense that the parameter space in panel c is more uniformly red with a few orange spots of large MIR values. Interestingly, this quantity becomes maximal for large chemical and moderate electrical coupling strengths, similarly to MIRst. The situation is similar for MIRii, where again it becomes maximal for large chemical and moderate electrical coupling strengths. We note that the maximum MIRii values of the orange spots in panel d are bigger by one order of magnitude than those in panels b and c. Finally, MIRf r still shows the same dependence on the coupling strengths to achieve its maximum, even though these maximum values are smaller by one or two orders of magnitude than those of the other three neural codes. Lastly, for MIRf r, there are blue regions of very small values, distributed evenly in the parameter space. Comparing the behavior of the various neural codes, the firing-rate seems to be less advanta- geous with respect to the maximum amounts of transmitted information for the rest. Our results suggest that it is more prominent for neurons to use temporal codes or the maximum points of their phases to communicate the maximal rate of information in modular neural networks, for chemical coupling strengths twice as that of the electrical coupling. 4 Discussion In this paper we sought to study how information is encoded in neural activity as it is crucial for understanding the computations underlying brain functions. Information is encoded by pat- terns of activity within neural populations responsible for similar functions and the interest in studying them is related to how the "neural code" can be read, mainly to understand how the brain processes information to accomplish behavior and cognitive functions. Thus, investigating the fundamental properties of neural coding in networks of spiking neurons may allow for the in- terpretation of population activity and, for understanding better the limitations and abilities of neural computations. To this end, we studied numerically neural coding in small-size networks of chemically and electrically coupled Hindmarsh-Rose spiking neurons. We have introduced four codes and have quantified the rate of information exchange for each code. The quantity used to measure the level of information exchanged is the Mutual Information Rate. The latter is by definition a symmetric quantity and cannot be used to infer the directionality of information flow. Therefore, our analysis cannot infer the direction of information exchange, only its intensity. In the simplest case of pairs of spiking neurons we have found that they exchange the largest amount of information per unit of time by opting for a temporal code in which the time of each spike conveys information which is transmitted to the other participating neuron. Our findings suggest that the firing-rate and interspike-intervals codes are more robust to additive Gaussian white noise. We have also studied four, chemically and electrically, coupled neurons and found that the largest rates of information exchange are attributed to the neural codes of maximum points of their phases and interspike intervals. In this network and in the absence of noise, pairs of nodes that are likely to exchange the largest amount of information per unit of time using the interspike- intervals and firing-rate codes are not adjacent in the network, whereas the spike-timings and phase codes promote large rate of information exchange for adjacent neurons in the network. This finding is also backed by similar results obtained for the same network with the role of chemical and electrical connections swapped. Our results provide evidence for the non-local character of firing- rate codes and local character of precise spike-timings, temporal, codes in modular dynamical networks of spiking neurons. It becomes thus clear that the type of neural code with largest information transmission rate depends on network adjacency. This result, if possible to extend to larger neural networks, would suggest that small microcircuits of fully connected neurons, also known as cliques [4], would preferably exchange information using temporal codes (spike-timings and phase codes), whereas on the macroscopic scale, where typically there will be pairs of neurons not directly connected due to the brain's sparsity, the most efficient codes would be the firing-rate and interspike-intervals codes. 17 For a relatively larger network of 20 neurons arranged in two equal-size small-world modules that form a bottleneck, our work shows that neurons choose a temporal code or the maximum points of their phases to transmit the maximal rate of information for chemical coupling strengths twice as that of the electrical coupling. Our estimations of the Mutual Information Rate are based on the symbolic encoding of tra- jectories, and thus, depending on the encoding, similar results can be obtained with the standard binary code [34, 47]. Particularly, if the chosen time-window for the binary code is close to the average interspike-intervals, MIRii would produce similar values with the binary code, as 1's would encode spikes and 0's would typically encode relaxation in the neural activity. Another possibility would be to use refinements in the estimation of the Mutual Information Rate, aiming at obtaining its true value for each. These refinements would correspond to the search for a generating Markov partition of higher order as in [17]. Since this is out of the scope of the present paper, we leave it for a future publication. In fact, we sought to study whether looking at the instantaneous spike-timings would provide less information than the codes based on the interspike intervals. Our decision was driven by the question: in a time-series of events, what does carry more information? A code based on the times between events, or a code based on the precise times of the occurrence of the events? Here, we have used chemical and electrical synapses with identical coupling strengths among all model neurons. As such, it is a limited study of relatively simple dynamical model neurons, small-size networks and equal synaptic connectivity. This choice was made for simplification. A similar study using unequal coupling strengths and larger networks would allow for more general results and would add more value from a neurophysiological perspective. Lastly, we have shown the importance of firing-rate and interspike-intervals codes against the spike-timings code and those based on phases. The latter codes prove to be more prone to noise contamination and to the transmission of smaller amounts of information per unit of time with the increase of noise intensity. 5 Acknowledgments This work was performed using the Maxwell high performance and ICSMB computer clusters of the University of Aberdeen. All authors acknowledge financial support provided by the EPSRC Ref: EP/I032606/1 grant. C. G. A. contributed to this work while working at the University of Aberdeen and then, while working at the University of Essex. References [1] S. Herculano-Houzel, The remarkable, yet not extraordinary, human brain as a scaled-up primate brain and its associated cost, Proc. Natl. Acad. Sci. USA 109 (2012) 10661. [2] D. H. Perkel, T. H. Bullock, Neural coding, Neurosciences researchs symposium summaries. The MIT Press 3 (1969) 405 -- 527. [3] E. Schneidman, M. J. Berry, R. Segev, W. Bialek, Weak pairwise correlations imply strongly correlated network states in a neural population, Nature 440 (2006) 1007 -- 12. [4] M. W. Reimann, M. Nolte, M. Scolamiero, K. Turner, R. Perin, G. Chindemi, P. D(cid:32)lotko, R. Levi, K. Hess, H. Markram, Cliques of neurons bound into cavities provide a missing link between structure and function, Frontiers in Computational Neuroscience 11 (2017) 48. [5] O. Sporns, Networks of the brain, Cambridge MA: The MIT Press, 2011. [6] E. Kandel, J. Schwartz, T. Jessell, Principles of Neural Science (third ed.), Elsevier/North- Holland, Amsterdam, London, New York, 1991. [7] W. Gerstner, A. K. Kreiter, H. Markram, A. V. M. Herz, Neural codes: Firing rates and beyond, Proc. Natl. Acad. Sci. USA 94 (1997) 12740 -- 12741. 18 [8] S. M. Both, The evidence for neural information processing with precise spike-times: A survey, Natural Computing 3 (2004) 195 -- 206. [9] P. M. DiLorenzo, J. D. Victor, Spike timing: Mechanisms and function, CRC Press, 2013. [10] M. Oram, D. Xiao, B. Dritschel, K. Payne, The temporal resolution of neural codes: does response latency have a unique role?, Trans. R. Soc. Lond. B 357 (2002) 987 -- 1001. [11] S. Thorpe, F. Fize, C. Marlot, Speed of processing in the human visual system, Nature 381 (1996) 520 -- 522. [12] R. Johansson, I. Birznieks, Fist spikes in ensembles of human tactile afferents code complex spatial fingertip events, Nature Neurrosci. 7 (2004) 170 -- 177. [13] M. S. Baptista, F. M. Kakmeni, C. Grebogi, Combined effect of chemical and electrical synapses in Hindmarsh-Rose neural networks on synchronization and the rate of information, Phys. Rev. E 82 (2010) 036203. [14] C. G. Antonopoulos, S. Srivastava, E. d. S. S. Pinto, M. S. Baptista, Do brain networks evolve by maximizing their information flow capacity?, PLoS Comput. Biol. 11 (8) (2015) e1004372. [15] E. Bianco-Martinez, N. Rubido, C. G. Antonopoulos, M. S. Baptista, Successful network inference from time-series data using mutual information rate, Chaos: An Interdisciplinary Journal of Nonlinear Science 26 (4) (2016) 043102. [16] J. J. Hopfield, Pattern recognition computation using action potential timing for stimulus representation, Nature 376 (1995) 33. [17] O. N. Rubido, C. Grebogi, M. S. Baptista, Entropy-based generating Markov partitions for complex systems, Chaos (in press). [18] J. L. Hindmarsh, R. M. Rose, A model of neuronal bursting using three coupled first order differential equations, Proc. R. Soc. London Ser. B 221 (1984) 87 -- 102. [19] T. Pereira, B. M. S., J. Kurths, Phase and average period of chaotic oscillations, Phys. Let. A 362 (2007) 159 -- 165. [20] T. Pereira, B. M. S., J. Kurths, General framework for phase synchronization through localized sets, Phys. Rev. E 75 (2007) 026216. [21] C. E. Shannon, A mathematical theory of communication, The Bell System Technical Journal 27 (1948) 379. [22] A. Borst, F. E. Theunissen, Information theory and neural coding, Nat. Neurosc. 2 (1999) 947 -- 957. [23] M. Wibral, R. Vicente, J. T. Lizier, Directed information measures in Neuroscience, Springer- Verlag Berlin Heidelberg, 2014. [24] M. S. Baptista, R. M. Rubinger, E. R. Viana, J. C. Sartorelli, U. Parlitz, C. Grebogi, Mutual information rate and bounds for it, PLoS One 7 (2012) 10:e46745. [25] C. G. Antonopoulos, E. Bianco-Martinez, M. S. Baptista, Production and transfer of energy and information in Hamiltonian systems, PLoS ONE 9 (2) (2014) e89585. [26] G. Benettin, L. Galgani, A. Giorgilli, J.-M. Strelcyn, Lyapunov characteristic exponents for smooth dynamical systems and for Hamiltonian systems: A method for computing all of them. Part 1: Theory and Lyapunov characteristic exponents for smooth dynamical systems and for Hamiltonian systems: A method for computing all of them. Part 2: Numerical application, Meccanica 15 (1980) 9 -- 20, 21 -- 30. [27] A. Kraskov, H. Stogbauer, P. Grassberger, Estimating mutual information, Phys. Rev. E 69 (2004) 066138. 19 [28] O. Sporns, D. R. Chialvo, M. Kaiser, C. C. Hilgetag, Organization, development and function of complex brain networks, Trends in Cognitive Sciences 8 (2004) 418 -- 425. [29] M. Paulus, V. Komarek, T. Prochazka, Z. Hrncir, K. Sterbova, Synchronization and informa- tion flow in EEGs of epileptic patients, IEEE Engineering in Medicine and Biology Magazine 20 (5) (2001) 65 -- 71. [30] L. Paninski, Estimation of entropy and mutual information, Neural Computation 15 (6) (2003) 1191 -- 1253. [31] R. Steuer, J. Kurths, C. O. Daub, J. Weise, J. Selbig, The mutual information: Detecting and evaluating dependencies between variables, Bioinformatics 18 (suppl 2) (2002) S231 -- S240. [32] A. G. Dimitrov, A. A. Lazar, J. D. Victor, Information theory in neuroscience, Journal of computational neuroscience 30 (2011) 1 -- 5. [33] E. Bianco-Martinez, M. S. Baptista, Space-time nature of causality, Chaos 28 (2018) 075509. [34] 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. [35] M. S. Baptista, R. M. Szmoski, R. F. Pereira, S. E. d. S. Pinto, Chaotic, informational and synchronous behaviour of multiplex networks, Scientific Reports 6 (2016) 22617. [36] R. Sevilla-Escoboza, I. Sendina Nadal, I. Leyva, R. Guti´errez, J. Buld´u, S. Boccaletti, Inter- layer synchronization in multiplex networks of identical layers, Chaos 26 (2016) 065304. [37] I. Leyva, R. Sevilla-Escoboza, I. Sendina Nadal, R. Guti´errez, J. M. Buld´u, S. Boccaletti, Inter-layer synchronization in non-identical multi-layer networks, Scientific Reports 7 (2017) 45475. [38] C. G. Antonopoulos, M. S. Baptista, Maintaining extensivity in evolutionary multiplex net- works, PLoS ONE 12 (2017) 1 -- 14. [39] M. Kac, On the notion of recurrence in discrete stochastic processes, Bull. Amer. Math. Soc. 53 (1947) 1002?1010. [40] M. N. Shadlen, W. T. Newsome, Noise, neural codes and cortical organization, Current Opin- ion in Neurobiology 4 (4) (1994) 569 -- 579. [41] R. B. Stein, E. R. Gossen, K. E. Jones, Neuronal variability: Noise or part of the signal?, Nat. Rev. Neurosci. 6 (5) (2005) 389 -- 397. [42] M. Baptista, F. M. Kakmeni, G. Del Magno, M. Hussein, How complex a complex network of equal nodes can be?, arXiv preprint arXiv:0805.3487. [43] M. S. Baptista, J. X. De Carvalho, M. S. Hussein, Finding quasi-optimal network topologies for information transmission in active networks, PloS one 3 (10) (2008) e3479. [44] D. J. Watts, S. H. Strogatz, Collective dynamics of small-world networks, Nature 393 (1998) 440 -- 442. [45] G. Zamora-L´opez, C. S. Zhou, J. Kurths, Cortical hubs form a module for multisensory inte- gration on top of the hierarchy of cortical networks, Front. Neuroinform. 4 (2010) 1. [46] D. Meunier, R. Lambiotte, E. T. Bullmore, Modular and hierarchically modular organization of brain networks, Front. Neurosci. 4 (2010) 200. [47] M. S. Baptista, C. Grebogi, R. Koberle, Dynamically multilayered visual system of the mul- tifractal fly, Phys. Rev. Lett. 97 (2006) 178102. 20
1507.00327
1
1507
2015-07-01T19:39:51
Comparative Connectomics: Mapping the Inter-Individual Variability of Connections within the Regions of the Human Brain
[ "q-bio.NC" ]
The human braingraph, or connectome is a description of the connections of the brain: the nodes of the graph correspond to small areas of the gray matter, and two nodes are connected by an edge if a diffusion MRI-based workflow finds fibers between those brain areas. We have constructed 1015-vertex graphs from the diffusion MRI brain images of 395 human subjects and compared the individual graphs with respect to several different areas of the brain. The inter-individual variability of the graphs within different brain regions was discovered and described. We have found that the frontal and the limbic lobes are more conservative, while the edges in the temporal and occipital lobes are more diverse. Interestingly, a "hybrid" conservative and diverse distribution was found in the paracentral lobule and the fusiform gyrus. Smaller cortical areas were also evaluated: precentral gyri were found to be more conservative, and the postcentral and the superior temporal gyri to be very diverse.
q-bio.NC
q-bio
Comparative Connectomics: Mapping the Inter-Individual Variability of Connections within the Regions of the Human Brain Csaba Kerepesia, Bal´azs Szalkaia, B´alint Vargaa, Vince Grolmusza,b,∗ aPIT Bioinformatics Group, Eotvos University, H-1117 Budapest, Hungary bUratim Ltd., H-1118 Budapest, Hungary Abstract The human braingraph, or connectome is a description of the connections of the brain: the nodes of the graph correspond to small areas of the gray mat- ter, and two nodes are connected by an edge if a diffusion MRI-based workflow finds fibers between those brain areas. We have constructed 1015-vertex graphs from the diffusion MRI brain images of 392 human subjects and compared the individual graphs with respect to several different areas of the brain. The inter- individual variability of the graphs within different brain regions was discovered and described. We have found that the frontal and the limbic lobes are more conservative, while the edges in the temporal and occipital lobes are more di- verse. Interestingly, a "hybrid" conservative and diverse distribution was found in the paracentral lobule and the fusiform gyrus. Smaller cortical areas were also evaluated: precentral gyri were found to be more conservative, and the postcentral and the superior temporal gyri to be very diverse. 1. Introduction Large co-operative research projects, such as the Human Connectome Project [1], produce high-quality MRI-imaging data of hundreds of healthy in- dividuals. The comparison of the connections of the brains of the subjects is a challenging problem that may open numerous research directions. In the present work we map the variability of the connections within different brain areas in 392 human subjects, in order to discover brain areas with higher variability in their connections or other brain regions with more conservative connections. The braingraphs or connectomes are the well-structured discretizations of the diffusion MRI imaging data that yield new possibilities for the compari- son of the connections between distinct brain areas in different subjects [2, 3] ∗Corresponding author Email addresses: [email protected] (Csaba Kerepesi), [email protected] (Bal´azs Szalkai), [email protected] (B´alint Varga), [email protected] (Vince Grolmusz) 1 or for finding common connections in distinct cerebra [4], forming a common, consensus human braingraph. Here, by using the data of the Human Connectome Project [1], we describe, by their distribution functions, the inter-individual diversity of the braingraph connections in separate brain areas in 392 healthy subjects of ages between 22 and 35 years. Since every brain is unique, the workflow that produces the braingraphs consists of several steps, including a diffeomorphism [5] of the brain atlas to the brain-image processed. After the diffeomorphism, corresponding areas of different human brains are pairwise identified through the atlas and, conse- quently, can be compared with one another. The braingraphs, with nodes in the corresponded brain areas, are prepared from the diffusion MRI images of the individual cerebra through a workflow detailed in the "Methods" section. Every braingraph studied contains 1015 nodes (or vertices). The vertices correspond to the subdivision of anatomical gray matter areas in cortical and subcortical regions. For the list of the regions and the number of nodes in each region, we refer to Table S1 and Figure S1 in the Appendix. Next, we describe the variability, or the distribution of the graph edges in each brain region, and also in each lobe. Figure 1 contains a simplified example on three small graphs (1,2,3) each with only two regions (A & B). The example clarifies the method, the way the results are presented through a distribution function, and the diagrams describing these functions. For any fixed brain area, and for any x : 0 ≤ x ≤ 1, let F (x) denote the fraction of the edges1 in the fixed area2 that are present in at most the fraction x of all braingraphs, (for a more exact definition of F (x) we refer to the "Methods" section). We note that F (x) is a cumulative distribution function [6] of a random variable described in the "Methods" section. 2. Results and Discussion Table 1 summarizes the edge diversity results for the 392 graphs for the lobes of the brain, described by the distribution functions F (x). The last column contains the data for the whole brain with 1015 nodes and 70,652 edges. The sum of the edges of the lobes in Table 1 is 30,326: these edges have both endpoints in the same lobe. More than forty thousand edges are present and accounted for only in the last column, because these edges connect nodes from different lobes. Therefore, the values in the last column cannot be derived from the other columns, since that column contains the contribution of edges that do not contribute to any other columns. We want to find out which brain areas are more conservative and which are more diverse than the others. We suggest to designate an area as "conservative" 1i.e., the number of the edges in question, divided by the number of all edges in the fixed area; 2i.e., with both vertices in the fixed area; 2 Figure 1: A simple example of computing the edge distribution between brain areas. In the example, there are three "braingraphs", each with two areas: A and B. We intend to count the edges that are present in all three graphs, only in two graphs and only in a single graph, respectively (between the same nodes, but in different graphs). For example, the copies of edge e are present in all three A areas, copies of edge h in all three B areas, copies of edge g in two B areas and edge f is present only in B1. The edges crossing the boundary of A and B (colored green) are ignored when counting the edge distribution within the areas A and B. In area A, two edges are present once, two edges twice and also two edges (including edge e) exactly three times. In area B, two edges (including f ) are present once, four edges (including g) twice and one edge – h – three times. In the diagram on the bottom, we give the F (x) distribution functions for both areas. On axis x, the fractions of the graphs are given, 1/3 correspond to one graph, 2/3 for two and 1.0 for all three graphs. F (x) is defined as the fraction of the edges in the fixed area that are present in at most the fraction x of all braingraphs. Data points corresponding to area A are on the same blue line (1/3, 2/3, 1) and those, corresponding to area B are on the broken, red line (2/7, 6/7, 1). We remark that if all three graphs are the same, then the data points are (0,0,1) (the extremely conservative case, orange line). Similarly, if no two graphs have the same edges, the data points are (1,1,1) (that is the extremely diverse case, green line). This type of diagram is used for the presentation of the results of the distribution of the edges in separate areas of the brain: The faster the line reaches the top F (x) = 1 value, the more diverse is the edge set in the corresponding brain area. We also note that in the diagram the lines connect the data points corresponding to the discrete values on axis x, and do not describe the step-function F (x) between the data points: we have chosen this visualization method because of its clarity even if a higher number of areas are shown (c.f. Figures 2 and 3 with numerous crossing lines). if for most x values, its F (x) distribution function is less than the F (x) of the all brain, given in the last column. We also suggest to designate an area as "diverse" if for most x values, its F (x) distribution function is greater than the F (x) of the all brain, given in the last column. The most conservative lobes are the smallest ones: the brainstem, the tha- 3 lamus and the basal ganglia contain only 1, 2 and 8 nodes, resp., and most of the edges in those regions are present in almost all braingraphs. If we take the average number of the braingraphs containing an edge from those regions, we get 316, 390 and 213 graphs, resp. It is much more interesting to review the diversity of the connections in larger areas. The frontal and the limbic lobes are conservative for most values of x (i.e., their F (x) values are less than that of the last column), while the temporal and the occipital lobes are diverse for larger x's. The distribution of the edges in the fusiform gyrus is particularly interesting: more than 10% of the graphs contain 46% of the edges which means this is a conservative brain area in that parameter domain, compared to the other lobes. The fusiform gyrus remains conservative for x = 0.2 and even for x = 0.3, but more than 50% of the graphs contain only 0.7% of the edges. That means that some edges of the fusiform gyrus are well conserved, and some parts are very diverse. The paracentral lobule has a very similar distribution. Table 1: The number of nodes, the number of edges and the diversity of the edges in different lobes, measured by the distribution function F (x). The list includes some brain areas that usually are not counted as lobes: like the fusiform gyrus, basal ganglia, and the paracentral lobule. The lobes, whose columns reach the value 1 faster (i.e. have more 1's at the bottom) have higher diversity. For example, the frontal and the limbic lobes are more conservative, while the temporal and the occipital lobes are more diverse. The distribution of the edges in the fusiform gyrus is particularly interesting: more than 10% of the graphs contain 46% of the edges which means this is a conservative brain area in that parameter domain, compared to the other lobes. The fusiform gyrus remains conservative for x = 0.2 and even for x = 0.3, but more than 50% of the graphs contain only 0.7% of the edges. Therefore, some edges of the fusiform gyrus are well conserved, and some other parts are very diverse. The paracentral lobule has a similar distribution. The data are also visualized on Figure 2 and an interactive figure http://uratim.com/diversity/Figure_2.html Table 2 summarizes the diversity results for those cortical areas which have more than 222 edges (see Table S2 in the Appendix for the edge numbers). 4 Figure 2: The diversity of the edges in different lobes, measured by the distribution function F (x). Only the areas with more than 10 nodes and F (x) values of more than 0.8 are visualized. The lobes, whose lines faster (i.e., with smaller x) reach value 1, have higher diversity. The fusiform gyrus and the paracentral lobule clearly moves from the bottom to the top of the diagram, relative to the other lines: this observation suggests that some of their edges are very conservative, and other areas have high diversity. An interactive version of this figure can be found at http://uratim.com/diversity/Figure_2.html 3. Methods We have worked with a subset of the anonymized 500 Subjects Release pub- lished by the Human Connectome Project [1]: (http://www.humanconnectome. org/documentation/S500) of healthy subjects between 22 and 35 years of age. Data were downloaded in October, 2014. We have applied the Connectome Mapper Toolkit [7] (http://cmtk.org) for brain tissue segmentation, partitioning, tractography and the construction of the graphs. The fibers were identified in the tractography step. The pro- gram FreeSurfer was used to partition the images into 1015 cortical and sub- cortical structures (Regions of Interest, abbreviated: ROIs), and was based on the Desikan-Killiany anatomical atlas [7](see Figure 4 in [7]). Tractography was performed by the Connectome Mapper Toolkit [7], using the MRtrix process- ing tool [8] and choosing the deterministic streamline method with randomized seeding. The graphs were constructed as follows: the 1015 nodes correspond to the 1015 ROIs, and two nodes were connected by an edge if there exists at least one fiber connecting the ROIs corresponding to the nodes. 5 Table 2: The diversity of the edges in different cortical areas, measured by the distribution function F (x). The abbreviation "ctx-lh" stands for "cortex left-hemisphere", "ctx-rh" for "cortex right-hemisphere". The areas, whose columns reach the value 1 faster (i.e., have more 1's at the bottom) have higher diversity. As in Table 1, the frontal regions are relatively more conservative, while the parietal regions are more diverse. Both precentral gyri are also conservative, and the postcentral and the superiortemporal gyri are more diverse. The last row contains the expected number of the graphs which contain a randomly chosen edge from the brain area indicated. Large expected number implies a conservative area, a small value implies a more diverse area. The data for the left hemisphere are also visualized on Figure 3 and on an interactive figure http://uratim.com/diversity/Figure_3.html Figure 3: The diversity of the edges in different cortical areas of the left hemisphere, measured by the distribution function F (x). The areas, whose lines faster (i.e., with smaller x) reach value 1, have higher diversity. An interactive version of this figure can be found at http: //uratim.com/diversity/Figure_3.html 6 3.1. The distribution function The variability of the edges in regions or lobes are described by cumulative distribution functions (CDF) (also called just the "distribution function") of the edges [6]. The general definition of the CDF is as follows: Definition 1. Let Y be a real-valued random variable. Then F (x) = P r(Y ≤ x) defines the cumulative distribution function of Y for real x values. For example, if a is the maximum value of Y then F (a) = 1, and if b is less than the minimum value of Y , then F (b) = 0. CDFs are used the following way: Suppose that our cohort consists of n persons' braingraphs (in the present work n = 392). For a given, fixed brain area, our random variable Y takes on values Y = u/n, u = 0, 1, . . . , n. The equation Y = u/n corresponds to the event that a uniformly, randomly chosen edge is in exactly u graphs from the n possible one, and the probability P r(Y = u/n) gives the probability of this event. Or, in other words, the equation Y = u/n corresponds to the set of edges - with both nodes in the fixed brain area - which are present in exactly u braingraphs, and the probability P r(Y = u/n) gives the fraction of the edges that are present in exactly u braingraphs. Therefore, F (x) = P r(Y ≤ x) gives the fraction (i.e., the probability) of the edges that are present in at most of a fraction x of all the graphs. The number of nodes and edges in each brain regions are given in supporting Tables S1 and S2 in the Appendix. We remark that we counted the edges without multiplicities: that is, if an edge e was either present in, say, 42 copies or just 1 copy of the braingraph, in both cases we counted it only once. The distributions were computed by counting the number of appearances of each edge in all the 392 braingraphs. Then the distribution of these numbers were evaluated in lobes and smaller cortical areas. 4. Conclusions: By our knowledge for the first time, we have mapped the inter-individual variability of the braingraph edges in different cortical areas. We have found more and less conservative areas of the brain: for example, frontal lobes are conservative, superiortemporal and the post-central gyri are very diverse. The fusiform gyrus and the paracentral lobule have shown both conservative and diverse distributions, depending on the range of the parameters. Data availability: The unprocessed and pre-processed MRI data are available at the Human Connectome Project's website: http://www.humanconnectome.org/documentation/S500 [1]. 7 The assembled graphs that were analyzed in the present work can be accessed and downloaded at the site http://braingraph.org/download-pit-group-connectomes/. Acknowledgments Data were provided in part by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University. References [1] Jennifer A. McNab, Brian L. Edlow, Thomas Witzel, Susie Y. Huang, Himanshu Bhat, Keith Heberlein, Thorsten Feiweier, Kecheng Liu, Boris Keil, Julien Cohen-Adad, M Dylan Tisdall, Rebecca D. Folkerth, Han- nah C. Kinney, and Lawrence L. Wald. The Human Connectome Project and beyond: initial applications of 300 mT/m gradients. Neuroimage, 80: 234–245, Oct 2013. doi: 10.1016/j.neuroimage.2013.05.074. URL http: //dx.doi.org/10.1016/j.neuroimage.2013.05.074. [2] Madhura Ingalhalikar, Alex Smith, Drew Parker, Theodore D. Satterth- waite, Mark A. Elliott, Kosha Ruparel, Hakon Hakonarson, Raquel E. Gur, Ruben C. Gur, and Ragini Verma. Sex differences in the structural connec- tome of the human brain. Proc Natl Acad Sci U S A, 111(2):823–828, Jan 2014. doi: 10.1073/pnas.1316909110. URL http://dx.doi.org/10.1073/ pnas.1316909110. [3] Bal´azs Szalkai, B´alint Varga, and Vince Grolmusz. Graph theoretical anal- ysis reveals: Women's brains are better connected than men's. PLOS One, 10(7):e0130045, July 2015. doi: doi:10.1371/journal.pone.0130045. URL http://dx.plos.org/10.1371/journal.pone.0130045. [4] Bal´azs Szalkai, Csaba Kerepesi, B´alint Varga, and Vince Grolmusz. The Budapest Reference Connectome Server v2. 0. Neuroscience Letters, 595: 60–62, 2015. [5] Morris Hirsch. Differential Topology. Springer-Verlag, 1997. ISBN 978-0- 387-90148-0. [6] Willliam Feller. An introduction to probability theory and its applications. John Wiley & Sons, 2008. [7] Alessandro Daducci, Stephan Gerhard, Alessandra Griffa, Alia Lemkad- dem, Leila Cammoun, Xavier Gigandet, Reto Meuli, Patric Hagmann, and Jean-Philippe Thiran. The connectome mapper: an open-source processing 8 pipeline to map connectomes with MRI. PLoS One, 7(12):e48121, 2012. doi: 10.1371/journal.pone.0048121. URL http://dx.doi.org/10.1371/ journal.pone.0048121. [8] J Tournier, Fernando Calamante, Alan Connelly, et al. Mrtrix: diffusion International Journal of Imaging tractography in crossing fiber regions. Systems and Technology, 22(1):53–66, 2012. Appendix Abbreviations: ctx-rh: cortex right-hemisphere ctx-lh: cortex left- hemisphere Area name No. Of nodes ctx-lh-superiorfrontal ctx-rh-superiorfrontal ctx-rh-precentral ctx-lh-precentral ctx-lh-postcentral ctx-rh-postcentral ctx-lh-superiorparietal ctx-rh-superiorparietal ctx-rh-rostralmiddlefrontal ctx-lh-superiortemporal ctx-lh-rostralmiddlefrontal ctx-rh-inferiorparietal ctx-rh-superiortemporal ctx-rh-lateraloccipital ctx-rh-precuneus ctx-lh-lateraloccipital ctx-lh-precuneus ctx-lh-inferiorparietal ctx-lh-supramarginal ctx-rh-supramarginal ctx-rh-middletemporal ctx-lh-fusiform ctx-rh-lateralorbitofrontal ctx-rh-fusiform ctx-rh-lingual ctx-lh-insula ctx-lh-lingual ctx-lh-inferiortemporal ctx-rh-insula ctx-rh-inferiortemporal ctx-lh-middletemporal ctx-lh-lateralorbitofrontal 9 45 42 36 35 31 30 29 29 27 26 26 26 25 23 23 23 22 22 21 20 19 18 17 17 17 17 17 16 16 16 16 16 ctx-lh-caudalmiddlefrontal ctx-rh-paracentral ctx-lh-paracentral ctx-rh-caudalmiddlefrontal ctx-rh-medialorbitofrontal ctx-lh-medialorbitofrontal ctx-lh-parsopercularis ctx-lh-posteriorcingulate ctx-rh-posteriorcingulate ctx-rh-parsopercularis ctx-rh-parstriangularis ctx-rh-cuneus ctx-rh-pericalcarine ctx-lh-cuneus ctx-lh-pericalcarine ctx-lh-isthmuscingulate ctx-lh-parstriangularis ctx-rh-parahippocampal ctx-lh-bankssts ctx-rh-caudalanteriorcingulate ctx-rh-isthmuscingulate ctx-lh-parahippocampal ctx-rh-bankssts ctx-lh-rostralanteriorcingulate ctx-lh-caudalanteriorcingulate ctx-rh-parsorbitalis ctx-lh-transversetemporal ctx-lh-parsorbitalis ctx-rh-rostralanteriorcingulate ctx-lh-entorhinal ctx-lh-temporalpole ctx-rh-temporalpole ctx-rh-transversetemporal ctx-lh-frontalpole ctx-rh-entorhinal ctx-rh-frontalpole Left-Thalamus-Proper Left-Amygdala Right-Hippocampus Right-Amygdala Right-Putamen Right-Accumbens-area Left-Hippocampus Left-Pallidum Right-Pallidum Right-Thalamus-Proper Left-Putamen 10 13 12 11 11 11 10 10 9 9 9 8 8 8 7 7 7 7 6 6 6 6 6 6 5 5 4 4 4 4 3 3 3 3 2 2 2 1 1 1 1 1 1 1 1 1 1 1 Right-Caudate Left-Caudate Left-Accumbens-area Brain-Stem 1 1 1 1 Sum of nodes Table S1: The number of nodes in each ROI. 1015 Figure S1: The number of nodes in ROIs and lobes. The interactive figure can be viewed at http://uratim.com/diversity/Figure_S1-Krona.html 'all' 'ctx-lh-superiorfrontal' 'ctx-rh-superiorfrontal' 'ctx-rh-precentral' 'ctx-lh-precentral' 'ctx-rh-rostralmiddlefrontal' 'ctx-rh-inferiorparietal' 'ctx-lh-rostralmiddlefrontal' 70652 910 774 500 448 352 340 331 11 'ctx-lh-superiorparietal' 'ctx-rh-superiorparietal' 'ctx-lh-postcentral' 'ctx-rh-postcentral' 'ctx-rh-lateraloccipital' 'ctx-lh-lateraloccipital' 'ctx-lh-superiortemporal' 'ctx-lh-inferiorparietal' 'ctx-rh-superiortemporal' 'ctx-rh-precuneus' 'ctx-lh-precuneus' 'ctx-lh-supramarginal' 'ctx-rh-supramarginal' 'ctx-rh-middletemporal' 'ctx-lh-fusiform' 'ctx-rh-lateralorbitofrontal' 'ctx-lh-inferiortemporal' 'ctx-rh-insula' 'ctx-lh-lingual' 'ctx-rh-fusiform' 'ctx-rh-inferiortemporal' 'ctx-lh-lateralorbitofrontal' 'ctx-lh-insula' 'ctx-lh-middletemporal' 'ctx-rh-lingual' 'ctx-lh-caudalmiddlefrontal' 'ctx-rh-paracentral' 'ctx-rh-caudalmiddlefrontal' 'ctx-lh-paracentral' 'ctx-rh-medialorbitofrontal' 'ctx-lh-parsopercularis' 'ctx-lh-medialorbitofrontal' 'ctx-lh-posteriorcingulate' 'ctx-rh-parsopercularis' 'ctx-rh-posteriorcingulate' 'ctx-rh-parstriangularis' 'ctx-rh-cuneus' 'ctx-rh-pericalcarine' 'ctx-lh-cuneus' 'ctx-lh-pericalcarine' 'ctx-lh-isthmuscingulate' 'ctx-lh-parstriangularis' 'ctx-lh-bankssts' 'ctx-rh-caudalanteriorcingulate' 'ctx-lh-parahippocampal' 'ctx-rh-parahippocampal' 'ctx-rh-isthmuscingulate' 12 317 314 305 273 263 254 250 242 228 227 222 209 206 176 157 144 135 131 131 130 130 127 125 119 114 91 76 65 64 59 55 54 45 45 43 36 35 35 28 28 28 28 21 21 21 20 20 'ctx-rh-bankssts' 'ctx-lh-rostralanteriorcingulate' 'ctx-lh-caudalanteriorcingulate' 'ctx-rh-parsorbitalis' 'ctx-lh-parsorbitalis' 'ctx-rh-rostralanteriorcingulate' 'ctx-lh-transversetemporal' 'ctx-lh-entorhinal' 'ctx-rh-transversetemporal' 'ctx-lh-temporalpole' 'ctx-rh-entorhinal' 'ctx-rh-temporalpole' 'Left-Thalamus-Proper' 'Left-Amygdala' 'ctx-lh-frontalpole' 'Right-Hippocampus' 'Right-Amygdala' 'ctx-rh-frontalpole' 'Right-Putamen' 'Right-Accumbens-area' 'Left-Hippocampus' 'Left-Pallidum' 'Right-Pallidum' 'Right-Thalamus-Proper' 'Left-Putamen' 'Right-Caudate' 'Left-Caudate' 'Left-Accumbens-area' 'Brainstem' 20 15 15 10 10 10 8 6 5 4 3 3 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Table S2: The number of edges in each ROI. 13
1909.08071
1
1909
2019-09-17T20:06:40
Longitudinal structural connectomic and rich-club analysis in adolescent mTBI reveals persistent, distributed brain alterations acutely through to one year post-injury
[ "q-bio.NC", "q-bio.QM" ]
The diffused nature of mild traumatic brain injury (mTBI) impacts brain white-matter pathways with potentially long-term consequences, even after initial symptoms have resolved. To understand post-mTBI recovery in adolescents, longitudinal studies are needed to determine the interplay between highly individualised recovery trajectories and ongoing development. To capture the distributed nature of mTBI and recovery, we employ connectomes to probe the brain's structural organisation. We present a diffusion MRI study on adolescent mTBI subjects scanned one day, two weeks and one year after injury with controls. Longitudinal global network changes over time suggests an altered and more 'diffuse' network topology post-injury (specifically lower transitivity and global efficiency). Stratifying the connectome by its back-bone, known as the 'rich-club', these network changes were driven by the 'peripheral' local subnetwork by way of increased network density, fractional anisotropy and decreased diffusivities. This increased structural integrity of the local subnetwork may be to compensate for an injured network, or it may be robust to mTBI and is exhibiting a normal developmental trend. The rich-club also revealed lower diffusivities over time with controls, potentially indicative of longer-term structural ramifications. Our results show evolving, diffuse alterations in adolescent mTBI connectomes beginning acutely and continuing to one year.
q-bio.NC
q-bio
Longitudinal structural connectomic and rich-club analysis in adolescent mTBI reveals persistent, distributed brain alterations acutely through to one year post-injury Ai Wern Chung1*, Rebekah Mannix2, Henry A. Feldman3, P. Ellen Grant1, Kiho Im1* 1Fetal Neonatal Neuroimaging and Developmental Science Center, Division of Newborn Medicine, Boston 2Division of Emergency Medicine, Brain Injury Center, Boston Children's Hospital, Harvard Medical School, MA, Children's Hospital, Harvard Medical School, MA, USA 3Institutional Centers for Clinical and Translational Research and Division of Newborn Medicine, Boston USA Children's Hospital, Harvard Medical School, MA, USA *Corresponding authors: [email protected]; [email protected] Abstract The diffused nature of mild traumatic brain injury (mTBI) impacts brain white-matter pathways with potentially long-term consequences, even after initial symptoms have resolved. To understand post-mTBI recovery in adolescents, longitudinal studies are needed to determine the interplay between highly individualised recovery trajectories and ongoing development. To capture the distributed nature of mTBI and recovery, we employ connectomes to probe the brain's structural organisation. We present a diffusion MRI study on adolescent mTBI subjects scanned one day, two weeks and one year after injury with controls. Longitudinal global network changes over time suggests an altered and more 'diffuse' network topology post-injury (specifically lower transitivity and global efficiency). Stratifying the connectome by its back-bone, known as the 'rich-club', these network changes were driven by the 'peripheral' local subnetwork by way of increased network density, fractional anisotropy and decreased diffusivities. This increased structural integrity of the local subnetwork may be to compensate for an injured network, or it may be robust to mTBI and is exhibiting a normal developmental trend. The rich-club also revealed lower diffusivities over time with controls, potentially indicative of longer-term structural ramifications. Our results show evolving, diffuse alterations in adolescent mTBI connectomes beginning acutely and continuing to one year. Keywords: adolescents, mild traumatic brain injury, concussion, network theory, rich club 1 Introduction Mild traumatic brain injury (mTBI) or concussion poses an immense public health burden, particularly in adolescents. Adolescents accounted for greater than 750/100,000 of Emergency Department visits between 2002-2006 in the US1. While deemed "mild", concussed adolescents remain symptomatic for more than two weeks in 50% of cases and up to one month in 30% of children after injury2. Despite this large public health burden, clinicians have few objective tools to guide diagnosis and management of this common adolescent injury. As such, there has been increasing interest in the use of neuroimaging for objective markers of injury to improve patient monitoring and to better understand the mechanisms of concussion on a neurobiological level3 -- 5. However, most efforts have focused on adults, and such findings may not be translatable to adolescents due to differences in neurobiological properties such as maturity of tissue (myelination), greater vulnerability to impact (weaker neck musculature and support), increased water content and propensity towards cerebral edema6,7. In addition, neural developmental processes are ongoing in the adolescent brain, alongside the capacity for recovery through compensatory or developmental mechanisms7,8. While symptoms of neurologic dysfunction post- concussion are transient, diffused mechanical injury may have longer-lasting implications on brain structure given that the time-course for physiological recovery exceeds current clinical clearance9. For these reasons, we present an adolescent study that is longitudinal in design, vital for understanding the longer-term impact of concussion in this age group. Advanced diffusion-weighted MRI and its related techniques have revealed localised, microstructural white-matter injury in mTBI3 but the relationship of focal findings to overall brain function as an organised system remains poorly characterised. Structural neuroimaging analyses are shifting towards a global interrogation of the brain as an inter- connected system using network theory10 -- 13. In network theory, the cerebral cortex is parcellated into regions that are defined as nodes. Nodes are connected by edges defined by tractography reconstructed white-matter fibres, where edges are weighted by features of connectivity strength such as the number of fibres between nodes14. This graph of the brain incorporates localised, structural alterations and includes their contribution towards global, whole brain measures that capture the network's organisation or 'topology'15. Given the distributed nature of concussive injury and the ensuing biochemical cascade (focal versus diffused injuries, primary versus secondary responses and injuries over time)6,16 analysing diffusion-based, structural data with a gestalt approach afforded by network theory may be appropriate to identify disorganisation or aberrations in the mTBI brain. There are few diffusion network theory studies on children and adolescents with strictly mild TBI4,17. To date, one study in children found significant diffusion network differences between mTBI in 2 the acute stage (MRI within 72 hours post-injury) and controls, specifically a reduction in global integration and greater regional segregation in the network18. For children in the chronic stage of TBI (from complicated mild through to severe injuries), significant network differences similar to acute mTBI18 have been found up to nine years post-injury19 -- 21. In addition to global network theory measures, rich-club based analysis is a complementary investigation of subnetworks in relation to a collection of highly inter-connected hub brain regions. Hubs are critical for integrating distributed network domains and possess characteristics (such as having a high number of connections and high metabolic demand) potentially reflective of their importance for efficient neuronal signalling and communication in the brain22 -- 24. Densely connected hubs form a "rich-club" subnetwork that has been consistently identified in studies across ages in both normal and patient populations25 -- 31. The rich-club is a high-cost network in terms of wiring and metabolic usage, making them equally vulnerable in a number of disorders13,23. It has exhibited impaired structural connectivity in adolescent chronic moderate to severe TBI19 and increased functional connectivity in adult subacute and chronic mild to severe TBI32,33. To date, the characteristics of rich-club connectivity in adolescent mild TBI have not been studied, certainly not longitudinally, and is one of the primary contributions of this work. Brain structure in adolescence continues to change into adulthood34, leading to brain networks that differ from adults35. Network theoretical studies in children and adolescent concussion have been cross-sectional18 -- 21. Our pilot study investigates longitudinal changes of global structural network organisation in adolescents with mTBI assessed at the acute (< 3 days post-injury), subacute (two weeks post-injury) and chronic stages (one year post- injury), in relation to controls. We sought to further understand any observed changes in network topology alongside its rich-club organisation to identify subnetworks which may explain an altered system in adolescent concussion over time. Material and Methods Subjects This study was approved by the Institutional Review Board at Boston Children's Hospital. Informed consent was obtained from all participants and/or their legal guardians and research was performed in accordance with relevant guidelines/regulations. The patient group presented to the Emergency Department within 24 hours of injury (n=9). Each mTBI subject was scanned longitudinally within 72 hours (acute), two weeks (subacute) and one 3 year (chronic) post-injury. mTBI was defined as a blunt, sports-related injury to the head resulting in either (1) alteration in mental status (including loss of consciousness, disorientation, or amnesia) or (2) any of the following symptoms that started within four hours of injury and were not present before the injury: headache, nausea, vomiting, dizziness/balance problems, fatigue, drowsiness, blurred vision, memory difficulty or difficulty concentrating. Patients were excluded from the study if they presented to the Emergency Department with Glasgow Coma Scale<14, focal symptoms or other indications for head imaging or intracranial hemorrhage seen when imaging was obtained, orthopaedic fracture, co-existing intra-abdominal or intra-thoracic trauma, or spinal-cord injury, or an underlying neurologic disorder or psychiatric illness requiring medications. A group of healthy controls were also recruited and MRI scanned once (n=9), matched for age and sex with the patient group at the acute time-point. Controls were recruited without current neurological complaints or recent head trauma at least a year prior to scanning. MRI Acquisition T1- and diffusion-weighted imaging (DWI) data were acquired on a 3T Siemens Tim Trio system, maximal gradient strength 40mT/m (Erlangen, Germany). T1-weighted motion mitigated multi-echo MPRAGE36 parameters were: TR=2520ms; TE=1.74, 3.54, 5.34 and 7.14ms; inversion time=1350ms; FOV=240mm2; voxel size=1mm3. DWI simultaneous multi- slice, echo-planar imaging37 parameters were: 63 non-collinear gradient direction volumes acquired at b=3000s/mm2; 4 b=0 s/mm2; TR=5800ms; TE=119ms; FOV=240mm2, voxel size=2mm3. Additional non-diffusion weighted volumes acquired in the anterior-posterior and posterior-anterior direction were obtained (one for each phase-encoding direction) for susceptibility artefact correction in the pre-processing stage later. Image Processing T1-Weighted Data - were pre-processed with the Freesurfer 'recon-all' pipeline (https://surfer.nmr.mgh.harvard.edu) and output were visually assessed. The T1-weighted cortical surface was parcellated into cortical regions as defined by the Desikan-Killiany atlas38. A transform between native skull-stripped b0 to T1-weighted space was computed via a rigid and affine-registration (NiftyReg, http://cmictig.cs.ucl.ac.uk/wiki/index.php/NiftyReg39). The inverse of this transform, t', was used to register the Freesurfer cortical parcellation to b0 space. t' was also applied on Freesurfer's whole brain white-matter mask to map it to b0 space for tractography. DWI Data - were visually assessed for artefacts and up to three corrupt gradient volumes were removed for each dataset. Remaining DWI data were corrected for susceptibility 4 artefact, subject motion and eddy current distortion with topup and eddy in FSL40,41. Gradient b-vectors were rotated following eddy. Linear least squares diffusion model fitting and 2nd order Runga-Kutta tractography was achieved using Diffusion Toolkit and visualised with TrackVis42. Tractography was seeded from the centre of each white-matter voxel in the brain, with a tracking step-size of 0.1mm and angular threshold of 45°. Tracts with length between 20mm and 200mm were retained. The cortical regions reached by each tract's endpoint was recorded. The number of streamlines connecting any two regions formed the corresponding entry in the connectome. Network nodes included 68 cortical regions, and edge weights were the number of streamlines connecting pair-wise nodes. Connectivity matrices were normalised by the total number of tracts in the same matrix30. The following imaging and network features were calculated for all mTBI subjects' time-point data and for controls. Whole Brain Diffusion Measures Analyses We determine the potential and extent of whole brain diffusion measures and volume change during this developmental phase of our subjects to assess whether these changes may impact the networks constructed. Whole brain white-, cortical and deep grey-matter tissue volumes were calculated using T1-weighted images from their respective Freesurfer masks. Whole brain tissue volume was calculated by combining all three masks. Each tissue mask was mapped to native diffusion space (with the t' transform, as described in 'Image Processing) and the FA and mean diffusivity (MD) values within each mask were averaged. Network Analyses Global Network Theory Analyses Analyses presented in this manuscript were performed on weighted networks. Network theoretical measures calculated (https://sites.google.com/site/bctnet) were: network density, global transitivity, efficiency, node betweenness-centrality, characteristic path length, modularity, clustering coefficient and the small-world coefficient15. We generated 1000 random realisations of the observed network while preserving network size, degree distribution and density43. A network measure is then normalized by dividing it by the mean corresponding measure computed across all random realisations to assess if the determined differences are random. We computed these null-network normalised measures for transitivity, characteristic path length, global efficiency and clustering coefficient. This is commonly done to improve the comparability of determined network measures between groups and subjects14,15,18,19. Figure 1 shows the stages from network construction onwards. To investigate the effect of the diffusion model on network theory measures we repeated this 5 analysis on connectomes computed from QBall reconstructed data (see Supplementary Analysis S1 for full Methods). Rich-Club Connectivity Analyses Rich-club (RC) nodes were defined as ten a priori regions on the Desikan-Killiany Freesurfer atlas: the superior frontal and parietal gyri, precuneus, posterior cingulate and insular bilaterally26,44 -- 47. These regions have been consistently established as key brain hubs across age, pathologies and species13,22,28,48,49. To confirm that these a priori regions were reasonable for our population, traditional RC analysis was also performed on controls and the patient group at each time-point, on both weighted and binary networks (see Supplementary Analysis S2). Normalised RC coefficients50,51 were computed on group-level connectomes, revealing significant values greater than 1 for a range of degrees (Figure S2.1). The majority of RC nodes extracted based on these analyses intersected with all the above a priori regions (Tables S2.1-S2.3). For the experiment presented for the remainder of this manuscript, three subnetworks were defined by grouping edges according to their association with the a priori RC nodes: 'rich-club' (connecting RC nodes only), 'feeder' (connecting an RC and a non-RC node) and 'local' (connecting non-RC nodes only) subnetworks. Figure 1 illustrates how a model network is divided into subnetworks in relation to rich-club nodes, and plots the connectomes of the three subnetworks for a control subject. The network density (number of connected edges, NoE), average number of reconstructed streamlines per edge (NoS), average FA, MD, axial and radial diffusivities (AD and RD) were calculated for each subnetwork for statistical analyses. 6 Figure 1 - Overview of network construction and rich-club subnetwork definitions. Top panel: depicts general post-processing steps from tractography. Middle panel: Using a priori definitions of ten bilateral rich-club nodes (SFG = superior frontal gyrus, INS = insular, PCC = posterior cingulate, Prec = precuneus, SPG = superior parietal gyrus), edges and remaining nodes in the network are subsequently grouped into rich-club, feeder and local subnetworks according to their relationship to the rich-club nodes. Bottom panel: Given the a priori rich-club nodes (in cyan), the axial and sagittal projections of each subnetwork is plotted for an example Control subject. Statistical Analyses Fisher's exact test was used to compare sex differences between control and mTBI groups, and Mann-Whitney U-tests to compare their ages. The following statistical analyses were performed on brain volume and diffusion measures, network theoretical measures and diffusivity measures from RC analyses, using SPSS (IBM, SPSS Statistics, 2012). ANCOVA covarying for age at time of MRI was performed for cross-sectional comparisons between 7 controls and mTBI subjects at the acute and chronic time-point. For longitudinal mTBI group analyses, a repeated-measures ANCOVA model was performed with the following equation fitting the outcome measure yit for subject i at time-point t=[1,2,3] : yit = a + (b × agei) + ct + eit, with main effect time-point representing acute, subacute and chronic stages, and eit is the error term with mean 0 and variance δ2 t (specific to time). Time-point was treated as a 3- level factor to enable nonlinear trends with time. Analyses were covaried by subject's age at baseline MRI (acute time-point for mTBI or time of MRI for control subjects). An unstructured covariance matrix was employed to allow for arbitrary correlations for a given subject's dependent variable over time. This model enabled the use of all data points by accounting for missing time-points. In the interest of understanding longitudinal effect of concussion, we limit our report of results to significance found with time-effect. Where only main factor 'time- point' is significant, the observed changes are significant irrespective of a subject's age. If both 'time-point' and 'age at baseline MRI' factors are significant, then the age at time of concussion has an effect on the MRI measure but not on the observed change with time. As such, results with only significant age-effect are reported in Supplementary Materials. We took p <= 0.05 as the criterion for statistical significance. Due to the exploratory nature of our pilot study, p-values are uncorrected for multiple comparisons correction. Results Demographics Table 1 contains group demographics of subjects analysed and the MRI timings for each mTBI time-point. Following visual inspection of the MRI data, two mTBI subjects in the subacute stage and one control were removed from analyses due to excessive motion during acquisition. Another two mTBI subjects were only scanned at a single time-point. This resulted in five mTBI subjects with full longitudinal data available at all three time-points, two subjects with data at both acute and chronic time-points, and two subjects with data from a single time-point (one subject acquired at the acute stage, and another subject at the subacute stage). The sex and age between control and mTBI groups were non-significant (all p>0.05). Every effort was made to recruit controls age-matched to mTBI patients in the acute time-point, with Mann-Whitney U-tests revealing a closer match of controls to patients at the acute time-point (control vs. acute (p-value = 0.879); control vs. chronic (p = 0.152)). On average, patients were scanned 1, 17 and 418 days post-injury. See Supplementary Table S3 for all ages and data analysed at each time-point. 8 Table 1 - Demographics and time of MRI Controls mTBI Patients Number of subjects [Acute/Subacute/Chronic] Gender F/M [Acute][Subacute][Chronic] 8 1/7 Median (IQR) [range] age at time of scan 8/6/7 [1/7],[1/5],[1/6]# 13.46(1.62) [11.97 - 15.84] Controls mTBI Patients: Acute Subacute Chronic Mean (stdev) [range] time of scan post-injury Acute (days) Subacute (days) Chronic (years) 13.34(1.54) [11.51-20.29]* 13.32 (3.50) [12.42-20.33]* 14.36 (2.36) [12.65-21.41]* 1.1 (0.7) [0-2] 16.8 (2.2) [5-20] 1.15 (0.06) [1.10-1.26] #Fisher's test on sex differences between groups all p> 0.05. *All p> 0.1 for age comparisons between Controls versus each mTBI time-point (Mann-Whitney U test). IQR = Inter-quartile range. Whole Brain Diffusion Measures Analyses Full statistical results and figures of volume and diffusion measures for each whole brain segmentation can be found in Supplementary Figure S3 and Table S4. Cross-sectionally, there was a significant increase in deep grey-matter FA for mTBI chronically compared to controls (Figure 2) (MD was conversely significantly lower, Figure S3). Although not significant, mTBI subjects in the acute stage exhibited, on average, greater FA than controls in white- and deep grey-matter (Figure 2a and 2c). White-matter volume was significantly larger at chronic time-point versus controls. Longitudinally, mTBI subjects had a significant change in deep grey-matter FA with time (F(2, 6.85) = 8.806, p = 0.013) with greater FA at the chronic stage compared to previous time-points (Figure 2c). Consistent with FA increases, a significant decrease in MD was also found in deep grey-matter (p = 0.001, Figure S3, with age-effect also significant). All remaining whole brain regions did not show significant FA, MD and tissue volume changes over time (all p > 0.05). 9 Figure 2 - Box plots of FA in whole brain tissue masks for each mTBI time-point and Controls, with individual data points plotted as black circles. Significant statistical result for cross-sectional t-tests between mTBI stage and controls are denoted in blue. Longitudinal general linear model results for mTBI are represented in red brackets, where significant post-hoc differences between time-points are shown when the corresponding ANCOVA analysis is significant. Age at MRI was not a significant confounder. Outliers are denoted at 1.5 times the inter-quartile range by black diamonds. Outliers are similarly defined in all further boxplot figures. Global Network Theory Analyses Network theoretical measures for each group, full statistical results and box plots are in Supplementary Figure S4 and Table S5. Cross-sectionally, there was a significant decrease in global efficiency (p = 0.05) in chronic mTBI subjects compared to controls (Figure 3b, in blue). No significance was reached for other global networks between controls and mTBI. More significant results were found longitudinally within the mTBI cohort, specifically in transitivity (F(2, 2.865) = 21.884, p = 0.018), global efficiency (F(2, 6.263) = 14.764, p = 0.004) and network degree (F(2, 3.975) = 11.232, p = 0.023) (Figure 3, in red). Transitivity and global efficiency decreased significantly over time: transitivity was significantly greater in the acute stage compared to both remaining time-points (Figure 3a); and global efficiency was significantly lower in the chronic time-point compared to earlier time-points (Figure 3b). Degree increased significantly over time, being significantly lower in the acute stage compared to later time-points (Figure 3c). Normalised measures of transitivity and global 10 efficiency were similarly significant and trending (Supplementary Figure S4, and Table S5). Although not significant, it is of interest to note qualitatively the 'normalisation' of several network measures over time longitudinally when comparing acute with chronic time-points, ending with values within the ranges of controls (small-world coefficient, shortest path measures, see Supplementary Figure S4). Figure 3 - Box plots of global network theoretical measures transitivity, global efficiency and degree at each mTBI stage and for controls. Longitudinal general linear model for mTBI are represented in red, where significant post-hoc differences between time-points are shown when the corresponding ANCOVA analysis is significant. Repeating the analysis on multi-fibre model, QBall, revealed largely similar network measure trends over time and between Controls and mTBI as with DTI (namely betweenness centrality, transitivity, global efficiency, modularity and degree. See Analysis S1 and Figure S1.1). Cross-sectional comparisons between mTBI and Controls were largely non-significant and transitivity and global efficiency measures exhibited significant changes over time for mTBI subjects in both models (Table S1.1). It is possible that in future work with larger sample sizes, the significance of these trends will be more consistent between models. Rich-Club Connectivity Analyses Figure 4 shows box plots of each network and diffusion measure grouped by cohort and subnetwork. See Supplementary Materials Table S6 for full statistical results. Longitudinally, mean NoS in the RC subnetwork is significantly altered over time (Figure 4a, in red, F(2, 4.701) = 8.283, p = 0.029). However, most significant longitudinal changes were observed in 11 the local subnetwork, namely in network density (NoE) (F(2, 5.662) = 5.315, p = 0.05), FA (F(2, 5.264) = 9.658, p = 0.017), MD (F(2, 7.130) = 4.990, p = 0.044) and RD (F(2, 6.147) = 7.424, p = 0.023) (MD and RD were also significant for age-effect) with most differences occurring between the acute stage and remaining two time-points (Figure 4 and Table S6). Although not significant, mean NoS was greater in the mTBI cohort at all stages compared to controls in the RC networks (Figure 4a). This increase in NoS would most likely be due to the greater whole brain FA (compared to controls, see Figure 2) leading to more tracts being reconstructed. FA is also observed within each subnetwork (Table S6) to significantly increase over time in the local subnetwork, and remained relatively consistent in RC and feeder networks. These diffusion changes in the network may explain the rise in connected edges in more 'peripheral' subnetworks (NoE, feeder F(2, 3.509) = 8.250, p = 0.047, local F(2, 5.662) = 5.315, p = 0.05). An interesting observation are the non-linear 'elbow' trends in many of the measures and subnetworks from acute to chronic time-points (Figure 4 and Supplementary Material Table S6), most being significant in the local subnetworks. Cross-sectionally, MD was significantly lower in acute versus controls in the RC subnetwork (p = 0.045, Figure 4c) which was primarily driven by decreasing RD (for both acute and chronic time-points versus controls at p = 0.039 and 0.027, respectively, Figure 4e), Although not significant, for all three subnetworks, mean FA was greater in mTBI acute subjects compared to controls, with all remaining diffusivity measures (MD, AD and RD) behaving similarly in the opposite direction as expected (lower diffusivity in mTBI compared to controls). 12 Figure 4 - Boxplots of mean diffusion measures for each subject group, in each subnetwork (rich-club, feeder and local subnetworks). Significant statistical result for cross-sectional t-tests between controls and mTBI stages are denoted in blue. Longitudinal general linear model for mTBI are represented in red, where post-hoc differences between time-points are shown when the corresponding ANCOVA analysis is significant. NoS = Number of streamlines, NoE = Number of edges. 13 Discussion To our knowledge, this is the first network theoretical and rich-club investigation on the longitudinal trajectory of brain structural organisation in concussed adolescents. Patients were scanned acutely (< 3 days), subacutely (2 weeks) and chronically (1 year) after injury. Our work also addressed the lack of mTBI diffusion MRI studies in adolescents, with the aim to utilise a global model of analysis offered by connectomics to understand the distributed nature of concussion injury. Compared with controls, the rich-club was the only subnetwork with significantly lower diffusivity measures in mTBI subjects at acute and chronic time- points. Longitudinally, we found alterations in global network topology indicative of lower efficiency over time that may be due to diffusion changes in the local subnetwork of the brain. Global Network Theory Analyses We found few significant difference in network metrics between mTBI and control groups. Unlike Yuan et al, we did not observe significant change in the acute stage compared to controls, although comparing medians suggests trends similar to theirs (namely increased small-world coefficient and transitivity)18. With the addition of more subjects in our pilot, these trends could become significant. With respect to differences in controls and chronic mTBI, we found significant decrease in global efficiency without a significant impact on other network metrics. Our chronic time-point of one year post-injury is earlier than two other studies in adolescent mTBI, which ranged from one to nine years post-injury20,21. These studies found an increase in small-world coefficient and also largely no significant network metric changes20. It may be that beyond a year after injury, global network features are similar to controls in adolescent mTBI, unlike in moderate to severe cases where network alterations have been documented from one to six years post-injury19,21. Interestingly, in our longitudinally analysis, patients revealed overall decreased transitivity and global efficiency, suggestive of a reduction in both network segregation and integration over time. Without other longitudinal network analyses in adolescent mTBI, we can only discern that while significant differences with controls are weak, we are able to detect within-subject network alterations up to a year post-injury. Whether these changes remain beyond a year remains to be studied. Rich-Club Connectivity Analyses We found significant structural changes in all three subnetworks in contrast with controls and also longitudinally within our mTBI cohort. Only the rich-club subnetwork revealed significant differences with controls - with primarily lower diffusivity in mTBI soon after injury and one 14 year later. Longitudinally significant changes also occurred in the subnetworks, namely the rich-club subnetwork in terms of mean NoS and the local subnetwork exhibiting the most alterations over time (in number of edges, FA, MD and RD). Our analyses suggest that significant FA increase in the local subnetwork drove the increase in network density observed in this same subnetwork. A rich-club analysis in adolescent moderate to severe TBI found differential patterns of change in connectivity strength between subnetworks in relation to the rich-club, when compared to controls19. Two years post-injury, Verhelst et al. identified increased connectivity in the local subnetwork which they suggest may be a compensatory effect to their observed decrease in rich-club connectivity. However, the subjects in Verhelst et al. presented with distributed diffuse axonal injury on MRI at time of injury with Glasgow Coma Scale 3 to 7, were comatose for >24 hours to 7 weeks and had encephalomalacia in the majority of cases. Interestingly, we found alterations in rich-club based subnetworks for our cohort which suffered from a milder form of brain injury. Theoretically the rich-club is the backbone of a network and is therefore by nature of its high inter-connectivity robust to network failures (or injury) as it provides alternative pathways to maintain network stability24. This characteristic makes this subnetwork rigid and constrained with lower 'evolvability'52, or biologically with respect to the brain, may possess lower plasticity, or too high a cost for repair. As the only subnetwork with enough sensitivity to display significant difference with controls, the rich-club's response to mTBI may also be a gradual process, affecting recovery or normal development beyond that of a year (compared to the more flexible peripheral subnetworks). This lack of sufficient time to detect recovery or developmental changes may explain why we did not find significant longitudinal trends in the rich-club for concussion subjects. Taken all together, our results suggest subnetworks are altered with changes found as early as three days after concussion, with potential aberrations through to one year particularly in the local subnetwork. Furthermore, the 'diffuse' organisation observed from our global network analyses may be explained by a shift in "importance" from the rich-club to the local subnetwork as measured by their diffusion properties. Subnetwork Analyses and the Biomechanics of Concussion But what does it mean for a local subnetwork to be altered? In contrast to the high metabolic and wiring cost of the rich-club, the more 'peripheral' local subnetwork requires less up- keep23 and perhaps its strengthening is indicative of a mechanism to compensate for an affected rich-club. What is yet to be established is not only a biomechanical explanation for the differential changes in strength and connectivity between subnetworks in the brain, 15 particularly over time after injury, but simply the role of feeder and local subnetworks for communication in the brain and their corresponding neuro-anatomy. Identifying a node as a rich-club member is dependent on how connected it is, which is in turn associated with its corresponding edges possessing high FA to drive tractography. Thus the rich-club subnetwork typically consists of high FA tracts that are highly myelinated for effective action potential and information propagation, whereas feeder and local subnetworks are predominantly comprised of lower FA pathways23. The high rotational acceleration forces from concussion on the brain have been extensively modelled to include both FA and the directional information from the diffusion tensor to model tissue 'stiffness' and capture the strain from impact53 -- 56. Such models not only show deformation to be widely distributed in white-matter throughout the brain, but suggest that regions with high fiber directionality (therefore high FA) have greater maximal principle stresses than regions with lower FA54,56. It is reasonable to hypothesise from this that the local subnetwork may be less likely (compared to the rich-club) to be directly affected following an impact. This may also explain why the rich-club was the most sensitive subnetwork in exhibiting significant diffusivity changes when compared to controls. Moreover, perhaps the local subnetwork was less affected and is presenting significant changes that are expected in normal development (increased FA, decreased MD) over the course of a year in our adolescent population. Whether our observed changes to the local subnetwork are simply due to normal development or are compensatory in nature remains to be investigated with a longer follow- up period and neuropsychological assessment. Significant differences between controls and mTBI in diffusion measures of FA and MD are commonly reported in cross-sectional studies employing localised region-of-interest or tract- based image analyses. Typically, greater FA at the acute stages and lower FA in the chronic stage is found when compared to controls3,4,17. These group differences in specific regions do not appear to manifest as a global deficit in terms of global network measures when compared with controls. Even so, global network measures exhibit sensitivity on an individual level by way of greater longitudinal significance, demonstrating the importance of longitudinal analyses and assessment in mTBI. Interestingly, analysing the network by rich- club associated subnetworks revealed similar diffusivity trends between controls and mTBI in the acute and chronic stages as in other region-based DTI studies3,4. In addition, we observed non-linear trends across mTBI time-points in these subnetworks, possibly capturing a period of normalisation in our network and diffusion measures, potentially returning to pre-injury values or transiently passing through normal57 -- 59. 16 Technical Considerations Our mTBI cohort exhibited the general trend of increased white-matter volume and FA, and decreased MD that typifies adolescent development60, however, none of these changes reached significance (see Whole Brain Diffusion Measures Analyses). Only deep grey- matter FA increase and MD decrease reached significance with grey-matter tissue volumes remaining unchanged over time. The longitudinal stability of white-matter measures suggests that normal developmental changes over one year in our cohort are unlikely to impact tractography results and the networks computed (particularly as we excluded deep grey-matter regions from our network). In particular, our observations with age are unlikely to be driven by significantly larger white-matter volumes or greater FA producing more streamlines. To further account for normal developmental changes in our analyses, we normalised our networks according to total number of streamlines for each subject and statistically factored for age to account for the age range in our mTBI group. In addition to the complexity of comparing concussion studies with different types of injury and stages of recovery when measurements are made, technical differences arising from network theoretical analysis must also be considered. There is no consensus on network construction in the neuroscience community to date with differences in diffusion modelling, tract reconstruction, choice of nodes, edge weights, network normalisation procedure and rich-club node definition all contributing toward variations in TBI network theoretical findings in literature61. One variation we account for is by using a priori regions to define the rich club, regions that have been repeatedly reproduced in literature and similarly in our cohorts (Supplementary Analysis S2). The rich club regions identified largely overlapped between weighted and binary networks and between cohorts, with greatest variation between weighted and binary networks, demonstrating the ability of tract information to restrict the rich club to a small number of similar regions that are biologically plausible23,24,62. For an in- depth discussion on the effect of weighted versus binary networks on the rich-club, see Supplementary Analysis 2.3. Using these a priori RC nodes ensures consistency when comparing across our groups and time-points, and also with other studies. Fixing the rich club to a set of a priori regions is also particularly important in our subsequent analyses, such as for feeder and local subnetwork analysis. A point of consideration is our use of the diffusion tensor model as it is constrained to model diffusivity in a single fibre population, therefore unable to model multiple fibre populations with more complex configurations. By repeating our 'Global Network Theory Analyses' on connectomes derived from a multi-fibre diffusion model (Qball) we found largely similar network measure trends over time as the DTI model (Figure S1.1). There were fewer significant results in the QBall analysis, although measures that were significant were similar to DTI (transitivity, global efficiency). Future work 17 with larger sample sizes will determine the significance of the observed trends and their consistency for and between models. The tensor has the advantage of extracting the most probable connections, thus limiting the inter-subject variability and reconstruction of false- positive tracts found in more complex models63,64. We also acquired high diffusion sensitisation MRI data at b=3000 s/mm2 with other studies acquiring at b=750 to 1200 s/mm2 18 -- 21. High b-value imaging captures slow diffusing water molecules such as those associated with myelinated white-matter with greater sensitivity, enabling the detection of more subtle alterations in diffusion properties than lower b-values65 -- 67. Limitations The lack of significant differences compared to the control group may be due to the number of subjects in our pilot study being too small to overcome inter-subject variance in our groups. It should be noted that the network theoretical studies in children and adolescents discussed above are also of modest size, between 16 and 23 subjects18 -- 21. Future work with larger sample sizes will confirm the findings reported here, elucidating whether subnetwork changes are from normal development or compensatory and whether our observations at one year after injury have normalised or are transient, alongside a longitudinal control cohort. Conclusion We investigated the longitudinal impact of mTBI in adolescent structural network organisation from the acute, to subacute and finally chronic stages after concussion. The evolution of mTBI networks revealed global changes in network specialisation and integration over time. Rich-club analysis suggested these global alterations may be driven by changes to the structural integrity of the local, 'peripheral', subnetwork. This may be a compensatory response to injury, or reflective of normal developmental maturation in the local subnetwork. Furthermore, the rich-club was the only subnetwork to have significantly lower diffusivities when compared to controls. Overall, our study points towards mTBI having a diffused and distributed effect on brain structural network organisation up to a year after injury. When these neurophysiological alterations resolve and the longer-term neurological implications remain to be determined. However our early investigation suggests continual patient observation may be necessary during this period of ongoing development. 18 Acknowledgements The authors wish to thank the families that participated and our colleagues at Boston Children's Hospital for their help. This study was partially funded by the Harvard Catalyst, and the American Heart Association and Children's Heart Foundation 19POST34880005. References 1. Faul, M., Xu, L., Wald, M. M. & Coronado, V. G. Traumatic brain injury in the United States: emergency department visits, hospitalizations, and deaths. (Atlanta: Centers for Disease Control and Prevention, 2010). 2. Zemek, R. et al. Clinical Risk Score for Persistent Postconcussion Symptoms Among Children With Acute Concussion in the ED. JAMA 315, 1014 (2016). 3. Eierud, C. et al. Neuroimaging after mild traumatic brain injury: Review and meta-analysis. NeuroImage Clin. 4, 283 -- 294 (2014). 4. Mayer, A. R. et al. Advanced Biomarkers of Pediatric Mild Traumatic Brain Injury: Progress and Perils. Neurosci. Biobehav. Rev. (2018). doi:10.1016/j.neubiorev.2018.08.002 5. Wintermark, M. et al. Imaging Evidence and Recommendations for Traumatic Brain Injury: Conventional Neuroimaging Techniques. J. Am. Coll. Radiol. 12, e1 -- e14 (2015). 6. Adelson, P. D. & Kochanek, P. M. Head injury in children. J. Child Neurol. 13, 2 -- 15 (1998). 7. Kochanek, P. M. Pediatric traumatic brain injury: quo vadis? Dev. Neurosci. 28, 244 -- 255 (2006). 8. Giza, C. C., Mink, R. B. & Madikians, A. Pediatric traumatic brain injury: not just little adults. Curr. Opin. Crit. Care 13, 143 -- 152 (2007). 9. Kamins, J. et al. What is the physiological time to recovery after concussion? A systematic review. Br. J. Sports Med. 51, 935 -- 940 (2017). 10. Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat. Rev. Neurosci. 10, 186 -- 198 (2009). 11. Chung, A. W. et al. Characterising brain network topologies: A dynamic analysis approach using heat kernels. NeuroImage 141, 490 -- 501 (2016). 12. Fornito, A., Zalesky, A. & Breakspear, M. The connectomics of brain disorders. Nat. Rev. Neurosci. 16, 159 -- 172 (2015). 13. Crossley, N. A. et al. The hubs of the human connectome are generally implicated in the anatomy of brain disorders. Brain J. Neurol. 137, 2382 -- 2395 (2014). 14. Fornito, A., Zalesky, A. & Breakspear, M. Graph analysis of the human connectome: Promise, progress, and pitfalls. NeuroImage 80, 426 -- 444 (2013). 15. Rubinov, M. & Sporns, O. Complex network measures of brain connectivity: Uses and interpretations. NeuroImage 52, 1059 -- 1069 (2010). 16. Giza, C. C. & Hovda, D. A. The Neurometabolic Cascade of Concussion. J. Athl. Train. 36, 228 -- 235 (2001). 17. Dennis, E. L., Babikian, T., Giza, C. C., Thompson, P. M. & Asarnow, R. F. Diffusion MRI in pediatric brain injury. Childs Nerv. Syst. ChNS Off. J. Int. Soc. Pediatr. Neurosurg. 33, 1683 -- 1692 (2017). 18. Yuan, W., Wade, S. L. & Babcock, L. Structural connectivity abnormality in children with acute mild traumatic brain injury using graph theoretical analysis. Hum. Brain Mapp. 36, 779 -- 792 (2015). 19. Verhelst Helena, Vander Linden Catharine, De Pauw Toon, Vingerhoets Guy & Caeyenberghs Karen. Impaired rich club and increased local connectivity in children with traumatic brain injury: Local support for the rich? Hum. Brain Mapp. 0, (2018). 20. Yuan, W., Treble-Barna, A., Sohlberg, M. M., Harn, B. & Wade, S. L. Changes in Structural Connectivity Following a Cognitive Intervention in Children With Traumatic Brain Injury: A Pilot Study. Neurorehabil. Neural Repair 31, 190 -- 201 (2017). 19 21. Königs, M. et al. The structural connectome of children with traumatic brain injury. Hum. Brain Mapp. (2017). doi:10.1002/hbm.23614 22. van den Heuvel, M. P. & Sporns, O. Network hubs in the human brain. Trends Cogn. Sci. 17, 683 -- 696 (2013). 23. Collin, G., Sporns, O., Mandl, R. C. W. & van den Heuvel, M. P. Structural and functional aspects relating to cost and benefit of rich club organization in the human cerebral cortex. Cereb. Cortex N. Y. N 1991 24, 2258 -- 2267 (2014). 24. Heuvel, M. P. van den, Kahn, R. S., Goñi, J. & Sporns, O. High-cost, high-capacity backbone for global brain communication. Proc. Natl. Acad. Sci. 109, 11372 -- 11377 (2012). 25. Heuvel, M. P. van den & Sporns, O. Rich-Club Organization of the Human Connectome. J. Neurosci. 31, 15775 -- 15786 (2011). 26. van den Heuvel, M. P. et al. Abnormal rich club organization and functional brain dynamics in schizophrenia. JAMA Psychiatry 70, 783 -- 792 (2013). 27. Ball, G. et al. Rich-club organization of the newborn human brain. Proc. Natl. Acad. Sci. 111, 7456 -- 7461 (2014). 28. Grayson, D. S. et al. Structural and functional rich club organization of the brain in children and adults. PloS One 9, e88297 (2014). 29. Daianu, M. et al. Rich club analysis in the Alzheimer's disease connectome reveals a relatively undisturbed structural core network. Hum. Brain Mapp. 36, 3087 -- 3103 (2015). 30. Schirmer, M. D. & Chung, A. W. Structural subnetwork evolution across the life-span: rich-club, feeder, seeder. in Connectomics in Neuroimaging Workshop 134 -- 143 (LNCS, Springer, 2018). doi:https://doi.org/10.1007/978-3-030-00755-3_15 31. Schirmer, M. D., Chung, A. W., Grant, P. E. & Rost, N. S. Network structural dependency in the (2019). Neurosci. life-span. Netw. human doi:10.1162/netn_a_00081 connectome across the 1 -- 30 32. Hillary, F. G. et al. The Rich Get Richer: Brain Injury Elicits Hyperconnectivity in Core Subnetworks. PLOS ONE 9, e104021 (2014). 33. Antonakakis, M., Dimitriadis, S. I., Zervakis, M., Papanicolaou, A. C. & Zouridakis, G. Altered Rich- Club and Frequency-Dependent Subnetwork Organization in Mild Traumatic Brain Injury: A MEG Resting-State Study. Front. Hum. Neurosci. 11, (2017). 34. Lebel, C. & Beaulieu, C. Longitudinal Development of Human Brain Wiring Continues from Childhood into Adulthood. J. Neurosci. 31, 10937 -- 10947 (2011). 35. Richmond, S., Johnson, K. A., Seal, M. L., Allen, N. B. & Whittle, S. Development of brain networks and relevance of environmental and genetic factors: A systematic review. Neurosci. Biobehav. Rev. 71, 215 -- 239 (2016). 36. Tisdall, M. D. et al. Volumetric navigators for prospective motion correction and selective reacquisition in neuroanatomical MRI. Magn Reson Med 68, 389 -- 99 (2012). 37. Setsompop, K. et al. Improving diffusion MRI using simultaneous multi-slice echo planar imaging. Neuroimage 63, 569 -- 80 (2012). 38. Fischl, B. FreeSurfer. NeuroImage 62, 774 -- 781 (2012). 39. Ourselin, S., Roche, A., Subsol, G., Pennec, X. & Ayache, N. Reconstructing a 3D structure from serial histological sections. Image Vis. Comput. 19, 25 -- 31 (2001). 40. Andersson, J. L. R., Skare, S. & Ashburner, J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. NeuroImage 20, 870 -- 888 (2003). 41. Andersson, J. L. R. & Sotiropoulos, S. N. An integrated approach to correction for off-resonance effects and subject movement in diffusion MR imaging. NeuroImage 125, 1063 -- 1078 (2016). 42. Wang, R., Benner, T., Sorensen, A. & Wedeen, V. Diffusion Toolkit: A Software Package for Diffusion Imaging Data Processing and Tractography. in International Society for Magnetic Resonance in Medicine 3720 (2007). 43. Maslov, S. & Sneppen, K. Specificity and stability in topology of protein networks. Science 296, 910 -- 913 (2002). 20 44. Collin, G., de Nijs, J., Hulshoff Pol, H. E., Cahn, W. & van den Heuvel, M. P. Connectome organization is related to longitudinal changes in general functioning, symptoms and IQ in chronic schizophrenia. Schizophr. Res. 173, 166 -- 173 (2016). 45. Collin, G. et al. Impaired Rich Club Connectivity in Unaffected Siblings of Schizophrenia Patients. Schizophr. Bull. 40, 438 -- 448 (2014). 46. Wierenga, L. M. et al. A multisample study of longitudinal changes in brain network architecture in 4-13-year-old children. Hum. Brain Mapp. 39, 157 -- 170 (2018). 47. Markett, S. et al. Serotonin and the Brain's Rich Club-Association Between Molecular Genetic Variation on the TPH2 Gene and the Structural Connectome. Cereb. Cortex N. Y. N 1991 27, 2166 -- 2174 (2017). 48. Harriger, L., Heuvel, M. P. van den & Sporns, O. Rich Club Organization of Macaque Cerebral Cortex and Its Role in Network Communication. PLOS ONE 7, e46497 (2012). 49. de Reus, M. A. & van den Heuvel, M. P. Rich club organization and intermodule communication in the cat connectome. J. Neurosci. Off. J. Soc. Neurosci. 33, 12929 -- 12939 (2013). 50. Opsahl, T., Colizza, V., Panzarasa, P. & Ramasco, J. J. Prominence and control: the weighted rich- club effect. Phys. Rev. Lett. 101, 168702 (2008). 51. Colizza, V., Flammini, A., Serrano, M. A. & Vespignani, A. Detecting rich-club ordering in complex networks. Nat. Phys. 2, 110 -- 115 (2006). 52. Csermely, P., London, A., Wu, L.-Y. & Uzzi, B. Structure and dynamics of core/periphery networks. J. Complex Netw. 1, 93 -- 123 (2013). 53. Ji, S. et al. Group-wise evaluation and comparison of white matter fiber strain and maximum principal strain in sports-related concussion. J. Neurotrauma 32, 441 -- 454 (2015). 54. Giordano, C., Zappalà, S. & Kleiven, S. Anisotropic finite element models for brain injury prediction: the sensitivity of axonal strain to white matter tract inter-subject variability. Biomech. Model. Mechanobiol. 16, 1269 -- 1293 (2017). 55. Colgan, N. C., Gilchrist, M. D. & Curran, K. M. Applying DTI white matter orientations to finite element head models to examine diffuse TBI under high rotational accelerations. Prog. Biophys. Mol. Biol. 103, 304 -- 309 (2010). 56. Giordano, C., Cloots, R. J. H., van Dommelen, J. a. W. & Kleiven, S. The influence of anisotropy on brain injury prediction. J. Biomech. 47, 1052 -- 1059 (2014). 57. Mayer, A. R. et al. A prospective diffusion tensor imaging study in mild traumatic brain injury. Neurology 74, 643 -- 50 (2010). 58. Messé, A. et al. Structural integrity and postconcussion syndrome in mild traumatic brain injury patients. Brain Imaging Behav. 6, 283 -- 292 (2012). 59. Beek, L. V., Vanderauwera, J., Ghesquière, P., Lagae, L. & Smedt, B. D. Longitudinal changes in mathematical abilities and white matter following paediatric mild traumatic brain injury. Brain Inj. 29, 1701 -- 1710 (2015). 60. Tamnes, C. K., Roalf, D. R., Goddings, A.-L. & Lebel, C. Diffusion MRI of white matter microstructure development in childhood and adolescence: Methods, challenges and progress. Dev. Cogn. Neurosci. (2017). doi:10.1016/j.dcn.2017.12.002 61. Welton, T., Kent, D. A., Auer, D. P. & Dineen, R. A. Reproducibility of Graph-Theoretic Brain Network Metrics: A Systematic Review. Brain Connect. 5, 193 -- 202 (2015). 62. Griffa, A. & Van den Heuvel, M. P. Rich-club neurocircuitry: function, evolution, and vulnerability. Dialogues Clin. Neurosci. 20, 121 -- 132 (2018). 63. Maier-Hein, K. H. et al. The challenge of mapping the human connectome based on diffusion tractography. Nat. Commun. 8, 1349 (2017). 64. Zalesky, A. et al. Connectome sensitivity or specificity: which is more important? NeuroImage 142, 407 -- 420 (2016). 65. Baumann, P. S. et al. High b-value diffusion-weighted imaging: a sensitive method to reveal white matter differences in schizophrenia. Psychiatry Res. 201, 144 -- 151 (2012). 21 66. Chung, A. W., Seunarine, K. K. & Clark, C. A. NODDI reproducibility and variability with magnetic field strength: A comparison between 1.5 T and 3 T. Hum. Brain Mapp. 37, 4550 -- 4565 (2016). 67. Dudink, J. et al. High b-Value Diffusion Tensor Imaging of the Neonatal Brain at 3T. AJNR Am J Neuroradiol 29, 1966 -- 1972 (2008). Author Contributions R.M. and P.E.G. designed and supervised data collection. A.W.C. and K.I. were responsible for imaging analysis design. A.W.C. conducted the experiments and analysed the results. H.A.F., R.M. and K.I. provided statistical analysis guidance. A.W.C. and P.E.G. wrote the paper, all authors reviewed the manuscript. 22
1507.08817
3
1507
2016-02-26T12:48:11
Seeking for a fingerprint: analysis of point processes in actigraphy recording
[ "q-bio.NC" ]
Motor activity of humans displays complex temporal fluctuations which can be characterized by scale-invariant statistics, thus documenting that structure and fluctuations of such kinetics remain similar over a broad range of time scales. Former studies on humans regularly deprived of sleep or suffering from sleep disorders predicted change in the invariant scale parameters with respect to those representative for healthy subjects. In this study we investigate the signal patterns from actigraphy recordings by means of characteristic measures of fractional point processes. We analyse spontaneous locomotor activity of healthy individuals recorded during a week of regular sleep and a week of chronic partial sleep deprivation. Behavioural symptoms of lack of sleep can be evaluated by analysing statistics of duration times during active and resting states, and alteration of behavioural organization can be assessed by analysis of power laws detected in the event count distribution, distribution of waiting times between consecutive movements and detrended fluctuation analysis of recorded time series. We claim that among different measures characterizing complexity of the actigraphy recordings and their variations implied by chronic sleep distress, the exponents characterizing slopes of survival functions in resting states are the most effective biomarkers distinguishing between healthy and sleep-deprived groups.
q-bio.NC
q-bio
Seeking for a fingerprint: analysis of point processes in actigraphy recording Ewa Gudowska-Nowak∗ Marian Smoluchowski Institute of Physics and Mark Kac Center for Complex Systems Research, Jagiellonian University, ul. (cid:32)Lojasiewicza 11, 30 -- 348 Krak´ow, Poland Malopolska Center of Biotechnology, Jagiellonian University, ul. Gronostajowa 7, 30 -- 348 Krak´ow, Poland ∗ Contributing author: [email protected] Jeremi K. Ochab Marian Smoluchowski Institute of Physics and Mark Kac Center for Complex Systems Research, ul. (cid:32)Lojasiewicza 11, 30 -- 348 Krak´ow, Poland Katarzyna Ole´s Marian Smoluchowski Institute of Physics and Mark Kac Center for Complex Systems Research, ul. (cid:32)Lojasiewicza 11, 30 -- 348 Krak´ow, Poland Ewa Beldzik Malopolska Center of Biotechnology, Jagiellonian University, ul. Gronostajowa 7, 30 -- 348 Krak´ow, Poland Dante R. Chialvo CONICET, Buenos Aires, Argentina Aleksandra Domagalik Malopolska Center of Biotechnology, Jagiellonian University, ul. Gronostajowa 7, 30 -- 348 Krak´ow, Poland Magdalena F¸afrowicz Malopolska Center of Biotechnology, Jagiellonian University, ul. Gronostajowa 7, 30 -- 348 Krak´ow, Poland Tadeusz Marek Malopolska Center of Biotechnology, Jagiellonian University, ul. Gronostajowa 7, 30 -- 348 Krak´ow, Poland Maciej A. Nowak 2 Marian Smoluchowski Institute of Physics and Mark Kac Center for Complex Systems Research, Jagiellonian University, ul. (cid:32)Lojasiewicza 11, 30 -- 348 Krak´ow, Poland Halszka Ogi´nska Malopolska Center of Biotechnology, Jagiellonian University, ul. Gronostajowa 7, 30 -- 348 Krak´ow, Poland Jerzy Szwed Marian Smoluchowski Institute of Physics and Mark Kac Center for Complex Systems Research, Jagiellonian University, ul. (cid:32)Lojasiewicza 11, 30 -- 348 Krak´ow, Poland Jacek Tyburczyk Marian Smoluchowski Institute of Physics and Mark Kac Center for Complex Systems Research, Jagiellonian University, ul. (cid:32)Lojasiewicza 11, 30 -- 348 Krak´ow, Poland Abstract. Motor activity of humans displays complex temporal fluctuations which can be characterized by scale-invariant statistics, thus documenting that structure and fluctuations of such kinetics remain similar over a broad range of time scales. Former studies on humans regularly deprived of sleep or suffering from sleep disorders predicted change in the invariant scale parameters with respect to those representative for healthy subjects. In this study we investigate the signal patterns from actigraphy recordings by means of characteristic measures of fractional point processes. We analyse spontaneous locomotor activity of healthy individuals recorded during a week of regular sleep and a week of chronic partial sleep deprivation. Behavioural symptoms of lack of sleep can be evaluated by analysing statistics of duration times during active and resting states, and alteration of behavioural organization can be assessed by analysis of power laws detected in the event count distribution, distribution of waiting times between consecutive movements and detrended fluctuation analysis of recorded time series. We claim that among different measures characterizing complexity of the actigraphy recordings and their variations implied by chronic sleep distress, the exponents characterizing slopes of survival functions in resting states are the most effective biomarkers distinguishing between healthy and sleep-deprived groups. PACS numbers: 05.40.Fb, 05.10.Gg, 02.50.-r, 02.50.Ey, 05.70.-a KEYWORDS: complexity measures, universality, biological locomotion, sleep depri- vation SUBJECT AREA: mathematical physics, theory of complex systems, biomathematics 3 1. Introduction Despite numerous studies indicating anomalous temporal statistics and scaling in spontaneous human activity and interhuman communication [1 -- 6], there is much on- going discussion on the origin and universality of observed statistical laws. Behavioural processes are frequently conveniently characterized in terms of stimulus-response approach [3,7], by adapting the same systematically repeated external sensory protocol, which allows to estimate the statistics of subject's responses. In a more general approach, in which brains are conceived as information processing input-output systems [8, 9], the observed self-similar temporal patterns of non-stimulated spontaneous neuronal activity can be determined by analysing spatiotemporal statistics of location and timing of neural signals. Similar to scale-free fluctuations detected in psychophysical time series, also dynamics of collective neuronal activity at various levels of nervous systems exhibit power-law scalings. Remarkable scale-free fluctuations and long-range correlations have been detected on long time scales (minutes and hours) in data recorded with magneto- and electroencephalography [9], and have been attributed to the underlying dynamic architecture of spontaneous brain activity discovered with functional MRI (fMRI) and defined by correlated slow fluctuations in blood oxygenation level-dependent (BOLD) signals. On the other hand, negative deflections in local field potentials recorded at much shorter time scales (milliseconds) have been shown to form spatiotemporal cascades (neuronal avalanches) of activity, whose size (amplitude) and lifetime distributions are again well described by power laws. These power-law scaling behaviours and fractal properties of neuronal long-range temporal correlations and avalanches strongly suggest that the brain operates near a critical, self-organized state [8] with neuronal interactions shaping both, temporal correlation spectra and distribution of signal intensities. It seems thus plausible to further investigate timing, location and amplitudes of such cascades to gain information about underlying patterns of brain rhythms and to identify characteristics of stochastic spatial point processes which can serve as reliable models of the governing dynamics. Also, if the behaviour can be described and interpreted as resulting interface between brain dynamics and the environment, scale invariant features may be expected to emerge in human cognition [10] and human motion [6, 11 -- 14]. Some neurological and psychopathic diseases such as Parkinson's disease, vascular dementia, Alzheimer's disease, schizophrenia, chronic pain and even sleep disorders and depression are related to abnormal activity symptoms [15 -- 17]. So far, there are many non-unique evaluative measures used in clinical practice to determine severity of these disorders or the effect of applied drugs. The challenge thus remains to what extent 4 correlations during resting state (spontaneous) activity are altered in disease states and whether a set of characteristic parameters can be classified unambiguously to describe statistics of healthy versus unhealthy mind states and spatiotemporal organization of such disrupted brain dynamics. An objective measure of differences in sleep durations and effects of sleep deficit on restlessness and impulsivity in human activity is based on actigraph recording. The instrument is used in sleep assessment to discriminate between stages of sleep and wake through documented body movements [14]. The analysis of actigraphy data is frequently considered a cost-effective method of first choice, used both in clinical research and practice. times [18 -- 20]. The survival probability C(a) ≡ P rob(T ≥ a) =(cid:82) ∞ The aim of this study is to apply various complexity measures for large activity datasets and to quantify use of advanced statistical methods for activity data. Our analysis is based on assumption that a typical actigraph time series that monitor records of movements per fixed time period of known duration can be analysed as a stochastic point process. We establish a time-inhomogeneity of such a renewal process and describe its fractional character by analysing counting number distributions. The renewal process is specified by probability laws describing waiting times or inter-arrival (inter-event) a PT (τ )dτ is a common measure that relates the (mean) time of discharge (or escape time) to the probability density function PT (t). In contrast to the Poisson case, for which a characteristic mean time of escape exists, fractional point processes may possess C(a) and PT (τ ) functions which do not decay exponentially but algebraically. As a consequence of the power-law asymptotics, the process is then non-Markovian with a long memory, which may result in a lack of characteristic time-scales for relaxation. The paper is organized as follows: In subsequent sections the method of data collection is presented (Section II), followed by a brief description of a stochastic point process (Section III) which serves as a model for further statistical analysis. Section IV demonstrates our results obtained by use of different complexity measures. The findings are concluded in Section V. 2. Experiment: Actigraphy recordings An actigraph is a watch-like device worn on the wrist that uses an accelerometer to measure any slight body movement over the investigated period of time. This instrument applies simple algorithms (time-above-threshold, zero-crossing, and digital integration) to summarize the overall intensity of the measured movement within consecutive time- periods (usually 1-2 minutes long), so that the events recorded represent activity counts. The most general approach to analyse such a recording is to reduce the time series of measurements to a summary statistic such as sleep/wake ratios, sleep time, wake after sleep onset, and ratios of night-time activity to daytime or total activity [14 -- 16]. While such summary measures allow for hypothesis testing using classic statistical methods [14, 21 -- 23], large amounts of relevant information are disclosed in 5 those complex data when more advanced tools of quantification (cf. Section 3), like fractal analysis, are applied [11, 13, 15, 17, 24, 25]. In particular, actigraphy studies by Sun et al. [15] indicated that the scaling exponent of the power law detected in temporal autocorrelation of activity significantly correlates with the severity of Parkinson's disease symptoms. Similarly, universal scaling laws have been found in locomotor activity periods of humans suffering from major depressive disorders [12] and individuals subject to sleep debt [24, 25]. The disruption of the characteristic universality classes of such laws has been further addressed by Proekt et al. [7] in studies on dynamics of rest and activity fluctuations in light and dark phases of the circadian cycle. Figure 1. Time series for one-week recording of actigraphy in the ZCM mode for a typical subject in RW condition. Upper panel: colours in the graph reflect daytime (black) and night (red) recordings following the circadian rhythm. Lower panels depict normalized frequency histograms of the number of movements in a minute derived from the raw actigraphy recordings. Unlike histograms of activities at night, which contain a considerable amount of zero-recordings (no movement) within 1 min intervals, histograms of awake activity reflect higher dispersion and multi-modality. Here, the actigraphy measurements were performed on 18 healthy individuals over one week of their normal life with either full recovery sleep according to individual needs, i.e., at rested wakefulness (RW or control group), or with a daily partial sleep deprivation (SD), with a two-week gap in between the two conditions. During the sleep deficit week, the participants were asked to shorten their sleep by 33% (by 2 hours 45 minutes on average) of their "ideal sleep" by delaying bed-time and using an alarm clock in the morning. The precise length of the restricted sleep was calculated individually for each participant. Half of the subjects began with the RW phase followed by the SD phase, while the other half had the order reversed. Movement tracking was recorded with 6 Micro Motionlogger SleepWatch (Ambulatory Monitoring, Inc., Ardsley, NY), worn on the participant's non-dominant wrist. The data were collected in 1-min epochs in the Zero-Crossing Method (ZCM) mode, which counts the number of times per epoch that the activity signal level crosses a zero threshold [24]. An example of the raw time series, as well as resulting activity histograms are displayed in Fig. 1. The histograms for RW and SD groups are compared in Fig. 2. Both, night-time and daytime measurements were analysed and the data collected are provided in a supplementary material (see website of the Department of the Theory of Complex Systems, Jagiellonian University, http://cs.if.uj.edu.pl/jeremi/Actigraphy2013/). Figure 2. Raw actigraphy recordings: Figure displays a normalized histogram of activity in the ZCM mode averaged over all subjects (error bars are standard deviations). Horizontal axis corresponds to the number of activity counts within one minute intervals, while vertical axis to normalized frequencies of counts. Black bars reflect frequency values for the control group (individuals in rested wakefulness; RW), while red bars the frequencies for subjects undergoing sleep deprivation (SD). The activity counts show a clear bimodal distribution with peaks characterizing maxima of low- and high-activity subdistributions (cf. Fig. 1). The circadian cycle of both groups differs substantially: while RW individuals have relatively long "nights" and short "days", members of the SD group are characterized by a reversed motif of longer "days" and shorter "nights", which clearly influences their activity/rest patterns. To overcome this problem we normalized the "days" and "nights" of both groups to the same length. The resulting time series have been statistically analysed as described previously in [24]. In addition, in this study, complexity analysis has been extended to describe the statistics of events' counts as a function of the observation time. 3. Model: Spiking patterns and measures of deviation from uniformity Many pulse signals observed in nature can be described in terms of stochastic point processes, i.e., sequences of random points {ti} arriving in time [26]. The simplest realization of such signals can be a "spike train" composed of delta peaks arriving at δ(t − ti). The average value of such a signal is given by a spike rate r(t), or the firing probability density (firing probability per unit of time): random times Y (t) = (cid:80) M(cid:88) 1 ti (cid:90) t+∆t/2 lim ∆t→0 M ∆t i=1 t−∆t/2 dt(cid:48)Yi(t(cid:48)) ≡ (cid:104)Y (t)(cid:105) = r(t). (1) Here Yi(t) stands for a given realization of the firing process. The firing rate can be thus obtained by observing at least one spike in the time interval [t − ∆t/2, t + ∆t/2], dividing this probability by ∆t and taking the ensemble average (over M realizations). Accordingly, correlation function of the spike train is defined as (cid:104)Y (t1)Y (t2)(cid:105) ≡ C(t1, t2) = dt(cid:48) (cid:90) t1+∆t/2 M(cid:88) 1 (cid:90) t2+∆t/2 M ∆t = lim ∆t→0 (2) and for a stationary process it depends on the time difference t1−t2 only [20]. Vanishing width of impulses allows one to construct another descriptor of the spike train, i.e., a counting process N (t) which gives the count of impulses observed in a given time window t2−∆t/2 t1−∆t/2 i=1 1 dt(cid:48) 2Yi(t(cid:48) 1)Yi(t(cid:48) 2) (cid:90) t N (t) = dt(cid:48)Y (t(cid:48)). 7 (3) (4) (5) (6) If the spike sequence depends only on the homogeneous rate 0 r = lim ∆t→0 1 ∆t P rob{∆N (t, t + ∆t) = 1}, and the events appear independently, the resulting distribution of counts can be described by the probability P (N, t) = pn(t)δ(N − n) satisfying a balance equation pn = rpn−1 − rpn. When pn(0) = 0 for n ≥ 1 and p0(0) = 1, the recurrence leads to n! e−rt. In this scenario, the probability density function the general solution pn(t) = (rt)n (PDF) of inter-event times has the exponential form PT (τ ) = re−rτ which is a signature of spikes following each other on average at (cid:104)τ(cid:105) = 1/r intervals. Since dispersion of τ is finite in this case, very long inter-event times are exponentially rare. Passing the δ-pulse signal of intensity σ through a linear filter τ dX dt = −X + σY (t) (cid:88) i results in "shot noise" with exponentially decaying pulses X(t) = σ τ Θ(t − ti) exp[−t − ti ]. τ The above equations represent a basic model [27, 28] for synaptic noise with X(t) (cid:80) standing for a neuron's membrane potential and τ denoting relaxation time of the activated neuron. In more general terms one can consider a stochastic process X(t) = j cjxj(ωj(t−tj)) resulting from superposition of many statistically identical stochastic signals xj transmitted from independent sources whose intensity cj, frequencies ωj and "initiation" times tj may be also considered random variables [29]. Such superposition 8 models are salient in systems carrying information traffic and well qualify long- range dependence, power-law scaling, 1/f noise and anomalous relaxation [30]. Since extensive research shows that also brain dynamics -- from cultured single neurons to whole brains -- exhibit spatiotemporal self-similarity [31], it is tempting to assume a direct relation between the scaling laws of behavioural (macroscopic) and neuronal (microscopic) fluctuations [9, 31] and to approach them theoretically by similar models. Poissonian randomization of the parameter set {(cj, ωj, tj)} on the half space (−∞,∞) × (−∞,∞) × (0,∞) establishes a mapping of the microscopic signal to the output "macroscopic" process X(t). As discussed in [29, 30], if the Poissonian intensity follows algebraic Pareto laws, the aggregate "output" signal can exhibit amplitude-universal and temporally universal statistics independent of the pattern of generic signals xi(t). The emerging output patterns may then display inter-event time PDFs characterized -- unlike their Poisson (regular) counterparts -- by power-law tails PT (τ ) ∝ τ−α . This in turn implies "burstiness" with short time intervals of intense spiking separated by long quiescent intervals, contrasting strongly with a uniform spiking pattern of a Poisson processes. A general type of clustering processes stemming from the Poisson class has been introduced by Neyman [2, 32, 33]. In short, the Neyman-Scott process is a mixture of a Poisson (parental) process of cluster centres with a characteristic intensity r(t) with another, statistically independent (daughter) stochastic process which governs distribution of points within the cluster. The Laplace transform (moment generating i X(i), conditioned on Poisson distribution of random variable N takes then the form of = ΦN (t)(ΦXi(s)). It is easy to check that dispersion (or so called function) of a random variable representing a compound signal XN (t) ≡ (cid:80)N (t) ΦXN (t)(s) ≡(cid:68) (cid:69) X s N (t) Fano factor), defined as (cid:68) N (t) −(cid:10)XN (t) (cid:10)XN (t) (cid:11) X 2 (cid:11)2(cid:69) F F ≡ for such a distribution is greater than 1, the value characteristic of a simple Poisson distribution with uncorrelated events. Accordingly, dispersion of experimental distributions, derived by use of Eq.(7), serves as an indicating measure for clustering. Here, the point process is defined as an activity crossing a predefined threshold (see the description of zero-crossing mode in Sec. 2). This threshold is preset by the producer of the actigraphs; we do not have access to it. It should not be mistaken for the threshold used to define activity and rest periods in Sec. 4.4. Note that since the events are recorded within 1-minute bins, we do not have access to the time variable below the bin length. In order to detect variability of the difference in the number of events occurring in successive intervals t, another measure -- the Allan variance -- is frequently used (cid:10)[XN (t)+1 −(cid:10)XN (t) (cid:11) (cid:10)2XN (t) (cid:11)]2(cid:11) AF ≡ (7) (8) 9 In case of the fractal-rate stochastic point processes, for which the times between adjacent events are random variables drawn from fractal distributions, one observes a hierarchy of clusters of different durations [2, 33 -- 35] and both Fano and Allan factors deviate from unity. In what follows, we consider measures suitable to discriminate basic properties of the point process representing events' count in actigraphy recording. First, we analyse frequency histograms PX(x) of signal intensity and statistics of inter-event intervals. To further detect character of the event count process, we evaluate variations of rate r(t) in expanding time windows checking for event clustering. Analysis of variance of counts, based on derivation of Fano and Allan factors allows to detect non-Poissonian character of the counting process and identify presence of correlations in recorded sequences. 4. Results of statistical analysis 4.1. Frequency histograms for distributions of activity increments In order to detect dynamics of fluctuations in the intensity X(t), (cf. Fig.1), we have analysed increments ∆X = X(t + 1) − X(t) for RW and SD subjects separately, and computed density distributions P∆X presented in Fig. 3. The plots have been standardised and compared to a Gaussian density with the same mean and variance. There is a clear discrepancy between the Gaussian statistics and the observed densities of activity fluctuations. They display much heavier tails in intensity distribution than their Gaussian counterparts, indicating that large fluctuations in intensity of the signal are more frequent than predicted from a standard (classical) central limit theorem. 4.2. Timing of events: Rate analysis The degree of non-homogeneity of the event counting process N (t) has been approached by considering cumulative number of events as a function of time. The fluctuations in the number of events N (t) crossing a predefined activity threshold Xth has been analysed utilising Fano (FF, Eq. 7) and Allan (AF, Eq. 8) factors depicted in Figs. 4-7. The averages in the definitions of FF and AF have been calculated by taking consecutive, non-overlapping windows and disregarding the last one, if its length was smaller than the counting time T . As can be inferred from the displayed data sets, the quantities: FF measuring the event-number variance divided by event-number mean and AF reflecting discrepancy of counts in consecutive time windows, both strongly deviate from unity during daytime (left panels in Fig. 5 and 7), thus indicating a super-Poisson statistics N (t) with clustering of events in the observation time. Since the obtained fractal exponents are close to one, which is the maximal exponent value that FF can detect, AF seems better fitted for the analysis, as it can detect exponent values up to three. Both factors increase as T d within counting time T ∈ [1, 500]. Reshuffling time stamps in the series breaks down inter-event correlations and results in reduced and 10 Figure 3. Time series of activity increments ∆X for a typical subject. Lower panel depicts density distribution P∆X ( ∆X S.D. ) for RW and SD condition. The dashed line is a Gaussian density with the same variance. uniformly distributed values of FF and AF, as expected [2, 4]. Separate analysis performed for daytime and night-time data does not discriminate between FF and AF values derived for RW and SD groups. Also, no significant difference in estimated d exponents for those two groups is observed when the analysed time series are all standardized to the length covering 4h of sleep followed by 14h of daytime activity (alternatively, 14h of daytime activity followed by 4h of sleep). The FF curves for raw time series of the whole week of recording (Fig. 4) were fitted with least-squares linear regression on logarithmic data in the range indicated by the dashed lines (17 data points, with window sizes: 1-222; the fitting range was determined by dropping the rightmost points until the adjusted R2 coefficient stabilised). A leave-one-out cross-validation yields sample standard deviation of the fitted slope: 0.024 (0.030) for RW (SD), which is less than 10% larger than the standard error of the fit of the subject-averaged curve (Fig. 4, left panel). This ensures that the fitting regions were chosen reliably. The fitted lines are not significantly different in terms of slopes (the values of d together with the standard error of the fit are reported in the figure; the difference is below 1 SE), but after subtracting the common slope they are different in terms of the y-intercept (98.2 ± 1.1 and 82.5 ± 1.1 for RW and SD, respectively; a two-sided t-test for the difference between the intercepts yields p = 0.015). The same 11 Fano factor (FF) of the 7-day raw time series for the control (black) Figure 4. and sleep deprived (red) group. Each point is an average FF over 18 subjects; the error bars and shaded regions indicate standard deviations. The range and result of linear fit is indicated by the dashed lines. The lower horizontal curves show FF for time-reshuffled data. Left: untrimmed 24h days; the slopes of linear fits of the two curves are not significantly different, but the y-intercepts are. Right: days trimmed to 4h sleep followed by 14h daytime activity; there is no significant difference between the intercepts, which implies that the difference visible for 24h-days can be accounted for solely by the different lengths of day/night activity. Reshuffling the data within day/night compartments, shows that the clustering of activity indicated by FF can be introduced by day/night cycle. We scrutinize this effect in Fig. 5. Figure 5. Fano factor as in Fig. 4 controlled for time of day. Left: FF for 14h daytime; the power-law behaviour remains, although with a markedly smaller exponent. Right: FF for 4h night-time; significantly non-random clustering of activity can be observed only for very short windows. There is no significant difference between control and SD curves. vertical shift is present between time-reshuffled data. Hence, the difference between the intercepts can be accounted for solely by the different lengths of day/night activity. To avoid that, the data can be trimmed to the same day/night lengths, which reduces the shift as in Fig. 4 (right). 12 To assess the inter-subject variability, the standard deviation of slopes fitted for individual subjects (averaged over days) was calculated yielding ≤ 0.030 for both untrimmed and trimmed time-series. The intra-subject variability, can be described by standard deviation of the fitted slopes for a given subject but for different days; it yields ≤ 0.062. This means that the measurements across subjects and across days in the experiment can be expected to vary with similar magnitude; as a consequence, the subject-averaged curves should be indicative of the whole sample and, for clarity, we do not show the individual subject values. As a cautionary note, let us remark that the slopes fitted for individual subjects are different between the RW and SD groups under a paired t-test, but this difference disappears after the trimming as well. A paired t-test in such cases, however, may produce a false positive conclusion, because it does not take into account error of the fit of the slopes. Slope of an averaged curve thus provides more conservative test conclusions. Figure 6. Allan factor (AF) of the 7-day raw time series for the control (black) and sleep deprived (red) group, as in Fig. 4. Left: untrimmed 24h days; the y- intercepts of linear fits of the two curves are not significantly different, but the slopes are. Right: days trimmed to 4h sleep followed by 14h daytime activity. There is no significant difference between control and SD curves, which implies that the differences for 24h-days can be accounted for solely by the different lengths of day/night activity, as previously observed for the FF measure. Reshuffling the data within day/night compartments, shows that the day/night cycle introduces clustering of activity between windows longer than 10 min. We scrutinize this effect in Fig. 7. The lines fitted for AF curves (Fig. 6; 20 data points, with the window sizes ranging 1-404) are not different in terms of their intercepts (5.4 ± 1.1 for RW and 6.1 ± 1.1 for SD, where the numbers correspond to the fitted intercept and standard error of the fit; the difference between RW and SD is around 1 SE), but after subtracting the common intercept they are different in terms of the slope (see values in the figure; a two-sided t-test for the difference between the adjusted slopes yields p = 0.035). The lower horizontal curves in Fig. 6 show AF after randomly reshuffling time stamps of the raw data, resulting in a more pronounced difference between the curves. Again, those differences can be accounted for solely by different time-lengths of day/night activity, 13 Figure 7. Allan factor as in Fig. 6 controlled for time of day. Left: AF for 14h daytime; the power-law behaviour remains, although with a slightly smaller exponent. Right: AF for 4h night-time; clustering of activity can be observed only between very short successive windows. There is no significant difference between control and SD curves. so that data trimming is necessary. Observation of AF curves starting to separate only for large window sizes, with the FF and reshuffled AF curves separated for all window sizes, can be explained by the fact that AF measure is based on increments between neighbouring windows. This implies that AF curves would separate only when the windows are big enough to often contain the "edge" between a day and a night. 4.3. Detrended fluctuation analysis and Hurst exponent (cid:113) 1 N original time series X(t), t = 1, ..., N have been integrated Z(k) = (cid:80)k where X =(cid:80)N In order to quantify time-correlations in recorded time series, we have used the DFA method [36, 37], frequently applied in analysis of physiological data [25, 38 -- 41]. The t=1[X(t) − X], t=1 X(t)/N is the data mean and k is an integer number k ∈ {1, ..., N}. Subsequently, the series Z(k) have been divided into boxes of equal length n and, by linear regression analysis, the linear trend of data in each window has been obtained. (cid:80)N In the next step, the series Z(k) have been detrended by subtracting the local trend Zn(k) from the data and the root-mean-square fluctuation has been calculated F (n) = k=1[Z(k) − Zn(k)]2. The procedure has been repeated for different box sizes (n = 8 − 840 min for days and n = 7 − 220 min for nights, respectively) and the relationship F (n) ∝ nH has been established with H being the Hurst exponent. The complexity measure H provides information [36, 37] about the memory of investigated dynamic system: when H = 1/2, the variations in the time series are random and uncorrelated with each other. For 0 < H < 1/2, variations are likely to be antipersistent, i.e., increases of recorded values will be followed by subsequent decreases, whereas for 1/2 < H < 1 persistent behaviour of increment variations is most likely observed with 14 Figure 8. DFA measure F (n) plotted against the box size n. The red (brown) lines represent day-time data and blue (green) lines are characteristic for nocturnal actigraphy recordings. The F (n) curves are averaged over 18 individuals and over 6 days. The linear fits have been performed in the ranges 8 − 150 for days and 7 − 60 for nights. Results for healthy sleepers: Hd = 0.9568 ± 0.0019 for day activity and Hn = 0.7599 ± 0.0095 for nights. Analogous measures for individuals deprived from sleep: Hd = 0.9760 ± 0.0015, Hn = 0.7681 ± 0.0065. increases (decreases) followed by increases (decreases), respectively. Figures 8-9 illustrate results of the detrended fluctuation analysis (DFA) performed on the data. As can be inferred from Fig. 8, the H exponent derived for the mean value of 18 subjects analyzed over 6 days of recordings is greater than the value 1/2, typical for time-series representing an uncorrelated white noise. All extracted H values are close to 1, thus signaling persistent time-correlations in subsequent intervals. This observation stays in line with the findings based on the other methods: The exponent d ≈ 1, as obtained from FF and AF measures or derived from the power spectrum [24], implies the expected value H = (d + 1)/2 ≈ 1. Moreover, results of analysis presented in Fig.8 suggest a more pronounced difference between the groups of healthy sleepers and individuals deprived of sleep in day-activity recordings: Hd = 0.9568 ± 0.0019 versus Hd = 0.9760 ± 0.0015 for RW and SD, respectively. Note, that reported values ±0.0019 and ±0.0015 refer to a standard error (SE) of the fit. Further analysis of day-time recordings for 18 subjects shows relatively low In Figure 9 each box represents results variability of the evaluated Hd parameters: of DFA performed over all subjects yielding one exponent for each separate day. In contrast to Fig.8, the comparison of daily exponents (F (n) still being averaged over subjects) excludes clear differentiation between the groups (no evidence of statistically significant difference at the 95% level of the two-sided two-sample t-test, p-value 0.850). Unlike the derived Hd values, the complexity exponent Hn obtained for nocturnal 15 Figure 9. Boxplot illustrating variability of the DFA exponents for RW and SD groups over 6 consecutive days (the lines within boxes are medians; the upper and lower box edges represent values of the third and the first quartiles, respectively). The daytime Hurst exponents are labelled: D,RW and D,SD; the exponents for night: N,RW and N,SD. The subjects have lower value of the mean DFA exponent for nocturnal activity. No difference between the groups is found. recordings, averaged over all subjects and over 6 subsequent days, reflects lower values and larger variation: Hn = 0.7681± 0.0065 (SE) for SD and Hn = 0.7599± 0.0095 (SE) for RW, which already precludes discrimination between the two groups (see Fig.8). The high variability of the daily derived Hn exponents (see Fig.9) obscures those findings even more, and significance testing indicates no clear difference between SD and RW groups (test as above, p-value 0.684). 4.4. Survival probability in active and inactive states To assess the statistics of rare events in tails of waiting time PDF's, as the main measure of the discussed phenomena, we have constructed the (complementary) cumulative distribution C(a) of durations a (cid:90) ∞ C(a) ≡ P rob(T ≥ a) = PT (τ )dτ, (9) a which represents the survival probability for the system to stay in a given state up to the time a. For a renewal process with the counting function N (t) indicating the effective number of events before or at instant of time t: N (t) ≡ max{ktk ≤ t, k = 0, 1, 2, ...}, the survival function can be identified as C(a) = P rob(N (a) = 0). For a stationary time series following a Poisson renewal process the survival probability C(a) is expected to have a characteristic scale (a relaxation time τrel) related to the probability P rob(N (t) = 0) = e−t/τrel = e−rt. In contrast, for non-stationary complex systems with inhomogeneous distributions of decay times τrel, the resulting complementary distribution C(a) may be scale invariant [2, 9, 24]. 16 To further asses this point, we have assumed that the human subjects can be in two states (active or resting) with some dwell time distributions characterizing duration of active and resting phases. We have defined a subject to be in the active state if the experimental time series X(t) > 85, and to be at rest otherwise. The choice of an appropriate threshold is argued in Supplementary Material to [24]. The extracted estimates of cumulative distributions have been fitted with two formulae: a power-law C(a) = a−γ for rest periods and a stretched exponential C(a) = exp(cid:0)−αaβ(cid:1) for activity periods. The fitting has been performed using log-log or log-linear data, respectively, in order to account for the tails in the distributions. The fitted parameters α, β, and γ have been then compared for the RW and SD conditions. Complementary cumulative distribution of (a) activity and (b) rest Figure 10. periods for the control (black) and sleep deprived (red) group. Each curve corresponds to a distribution gathered from all the valid subjects on a single day (consisting of 4h sleep followed by 14h daytime activity). The activity and rest periods were defined as the times where ZCM was respectively above or below 85. Dashed lines in (a) show fitted stretched exponentials; solid lines in (b) show the average (over 6 days) fitted slope, and extend across the maximal range of fitting. i.e., each individual Each cumulative distribution, line shown in Fig. 10 has been constructed from rest/activity periods collected over a single day from actigraph recordings of all 18 participants under RW or SD conditions. Consequently, as shown in Fig. 11, for each day one data point characterizing the distribution of the whole group in RW (control) or SD condition has been obtained. The parameters characterizing a single cumulative distribution of rest/activity periods have been further averaged over the whole week. In order to compare RW and SD groups, such means of parameters fitted for several consecutive days have been compared with the two-tailed t-tests performed at 95% confidence level, preceded by a set of tests for equal variances, as explained elsewhere [24]. In the case of activity periods no clear difference between the RW and SD individuals has been observed. The overall fit (averaged over days and individuals) yields for the RW sample ¯α = 0.31 ± 0.04 and ¯β = 0.49± 0.03. Similar values for the SD mode are not significantly different (p > 0.05; two-tailed t-test). 17 Figure 11. The power-law exponents γ fitted to the complementary cumulative distributions from Fig. 10b for each consecutive day (4h sleep followed by 14h daytime activity). The error bars indicate the standard error of the fit. Unlike these findings, the cumulative distributions C(a) of rest period durations, averaged over all participants indicate a significant difference in behavioural motifs between the control group and sleep deprived individuals. Specifically, the means ¯γ of exponents for RW and SD subjects are different at 5% level, with p-value 0.012 (for 4-hour sleep followed by 14-hour daytime activity); additional leave-one out cross- validation was performed with the same conclusion. As mentioned before in Sec. 4.2, the reverse scheme of 14-hour daytime activity followed by 4-hour sleep (the length of time series stays as before, but different parts of data are cut away) was checked as well. In this instance, the means ¯γ do not differ at 5% level, with p-value 0.065. The higher (average) coefficient ¯γ = 0.85± 0.03 extracted for sleep-deficient individuals emphasizes the fact that the pattern of their resting times consists of more short periods and, respectively, fewer longer inactivity time intervals than in the control group. 5. Discussion and conclusions Vulnerability to chronic sleep deprivation is a complex phenomenon in which performance, subjective and neural measures indicate distinct features, possibly related to chronotype, somnotype and gender [12, 21 -- 23, 25]. Although not used routinely as a measurement of choice for evaluating the severity of locomotor disorders related to sleep deprivation in patients, the actigraphy records are considered by many researchers and medical doctors a valuable tool for demonstrating behavioural disturbances related to sleep deficit. Information detected in actigraph data provide measures for sleep and wake motifs in healthy individuals and in people with certain sleep problems and can be used to establish links between duration of sleep and fatigue and performance In particular, it seems that a strategy of assessment based on use effectiveness [23]. 18 of actigraphy recordings may serve as an objective tool to discriminate behavioural patterns associated with shortness of sleep and insomnia [11, 17, 22, 25]. In this paper properties of actigraph time-records have been analysed by use of multiple complexity measures. Presented model of a non-uniform random point process had successfully explained the main properties of analysed time-series indicating clustering of event counts as measured by Fano and Allan factors. Provided estimates allow to detect correlations in inter-event periods typical for a fractal stochastic process [4, 19, 29, 42]. Both factors increase with the counting time window as T d suggesting variability of the response (activity counts) characterized by the autocorrelation function: The linear growth in T of the event count variance is characteristic of a diffusive process. In contrast, the FF value increasing as a power function of the counting window indicates that fluctuations in the rate are observable on many time scales with a self-similar exponent d. The Fano factors derived in our study are always larger than 1, documenting that the rate of events in the data is an inhomogeneous process with aggregated event-counts. This super-Poissonian behaviour is observed for both -- nocturnal and day-time recordings. Similarly, the Allan factor which allows to detect the average variation in the pattern of adjacent counts, also points to the time correlations in analysed series. Nevertheless, variations of extracted FF and AF coefficients in both groups of examined subjects exclude drawing inference in hypothesis testing and do not allow to discriminate between the degree of correlation and patterns of bursting in the counts of events observed for the control (RW) or sleep-deprived (SD) individuals. In parallel to these findings, the complexity scores based on DFA analysis stay in conformity with the observed self-affinity of activity signals. However, while series of nocturnal recordings display a higher median of the Hurst exponent for the SD group with respect to the RW subjects, applied significance tests show no clear difference between the control group and the group with sleep deficiency. That observation limits practical use of DFA analysis in the identification of reduced sleep dynamics, in contrast to results on sleep impaired subjects diagnosed with acute insomnia [25]. To further identify the best mathematical model behind measured activity signals, they can be tested for stationarity and ergodicity [43]. Verification of stationarity of the time series can be approached by quantile lines test, followed by checking ergodic behaviour [44] to validate anomalous diffusion models of the data. These can be complemented by tests based on power spectra of FARIMA model [45]. While so far we have based our conclusions on fairly simple properties of the experimental data, there are indications of their non-stationarity. These issues are a starting point for our further research. Altogether, our investigations of long-range correlations present in the analysed time-series provide evidence that the best measure of discrimination between healthy sleepers and sleep-deficient group is obtained by examination of the cumulative (survival) distribution of resting periods. Even though period durations are limited by the time resolution of an actigraph and the night length, the survival distribution in a resting state exhibits approximately a power-law scaling which is robust over two 19 decades. Significantly higher values of the exponent γ for sleep-deprived subjects signal less heavy tails of waiting-time distributions in an immobile (resting) state than in an analogous distribution for the control group and can be associated with restlessness/inquietude and increased variability (burstiness) of activity in recorded time series. Consequently, those alterations of locomotor pattern can be informative about mood/behaviour disturbances related to sleep deficiency and possibly, used as a valuable diagnostic fingerprint discriminating between healthy and depressed/disordered individuals. Acknowledgments This work was supported in part by National Science Center (ncn.gov.pl) grants No. DEC-2011/02/A/ST1/00119 (JO) and DEC-2014/13/B/ST2/020140 (EGN), and the Polish Ministry of Science and Higher Education grant No. 7150/E-338/M/2014 (KO). The project communicated to Unsolved Problems of Noise 2015, July 14-17, Barcelona, Spain. References [1] A. Vazquez, J.G. Oliveira, Z. Dezso, K-I. Goh, I. Kondor, A-L. Barabasi, Modeling bursts and heavy tails in human dynamics Phys. Rev. E 73 036127 (2006). [2] D.R. Bickel, Estimating the intermittency of point processes with applications to human activity and viral DNA Physica A 265 634 (1999). [3] A. Gomez-Marin, J. J. Paton, A. R. Kampff, R. M. Costa, Z. F. Mainen Big behavioural data: psychology, ethology and the foundations of neuroscience Nature Neuroscience 17 1455 (2014). [4] C. Anteneodo, R.D. Malmgren, D.R. Chialvo Poissonian bursts in e-mail correspondence Eur. Phys. J. B 75 389 (2010). [5] A. Haimovici, E. Tagliazucchi, P. Balenzuela, D.R. Chialvo Brain organization into resting state networks emerges at criticality on a model of the human connectome, Phys. Rev. Lett. 110 178101 (2013). [6] D.R. Chialvo, A.M.G. Torrado, E. Gudowska-Nowak, J.K. Ochab, P. Montoya, M.A. Nowak, E. Tagliazucchi How we move is universal arXiv:1506.06717v1 [q-bio.NC] [7] A. Proekt, J.R. Banavar, A. Maritan, D. W. Pfaff Scale invariance in the dynamics of spontaneous behaviour Proc. Natl. Acad. Sci. USA 109 10564 (2012). [8] E. Tagliazucchi, P. Balenzuela, D. Fraiman, D.R. Chialvo Criticality in large-scale brain fMRI dynamics unveiled by a novel point process analysis Frontiers in Physiology 3 15 (2012). [9] J.M. Palva, A. Zhigalov, J. Hirvonen, O. Korhonen, K. Linkenkaer-Hansen, S. Palva Neuronal long-range temporal correlations and avalanche dynamics are correlated with behavioural scaling laws Proc. Natl. Acad. Sci. USA 110 3585 (2013). [10] C.T. Kello, G.D. Brown, R. Ferrer-I-Cancho, J.G. Holden, K. Linkenkaer-Hansen, T. Rhodes, G.C. Van Orden Scaling laws in cognitive sciences Trends Cogn. Sci. 14 223 (2010). [11] T. Nakamura, K. Kiyono, K. Yoshiuchi, R. Nakahara, Z. R. Struzik, Y. Yamamoto Universal scaling law in human behavioural organization Physical Review Letters, 99138103 (2007). [12] T. Nakamura, T. Takumi, A. Takano, N. Aoyagi, K. Yoshiuchi, Z.R. Struzik, Y. Yamamoto Of Mice and Men - Universality and Breakdown of Behavioural Organization PLoS One 3 e2050 (2008). 20 [13] P. Wohlfahrt, J.W. Kantelhardt, M. Zinkhan, A.Y. Schumann, T. Penzel, I. Fietze, F. Pillmann, A. Stang Transitions in effective scaling behaviour of accelerometric time series across sleep and wake Europhys. Lett. 103 68002 (2013). [14] M.H. Teicher Actigraphy and motion: new tools for psychiatry Harv. Review Psychiatry 3 18 (1995). [15] W. Pan, K. Ohashi, Y. Yamamoto, S. Kwak Power-law temporal autocorrelation of activity reflects severity of Parkinsonism Movement Disorders 22 1308 (2007). [16] W. Pan, S. Kwak, F. Li et al., Actigraphy monitoring of symptoms in patients with Parkinson's disease Physiology and Behavior 119 156 (2013). [17] J. Kim, T. Nakamura, H. Kikuchi, T. Sasaki, Y. Yamamoto Co-variation of depressive mood and locomotor dynamics evaluated by ecological momentary assessment in healthy humans PLoS ONE 8 e74979 (2013). [18] D.R. Cox Renewal Theory Methuen, London (1964). [19] F. Mainardi, R. Gorenflo, E. Scalas A fractional generalization of the Poisson processes Vietna, J. Math. 32 53 (2004). [20] W.A. Gardiner Introduction to random processes with applications to signals and systems McGraw Hill Inc. New York 1991. [21] H. Oginska, J. Mojsa-Kaja, M. Fafrowicz, T. Marek Measuring individual vulnerability to sleep loss the CHICa scale Journal of Sleep Research 23 341 (2013). [22] Y. Sun, C. Wu, J. Wang, W. Pan Quantitative Evaluation of Movement Disorders by Specified Analysis According to Actigraphy Recors Int. J. Integrative Medicine 1 1 (2013). [23] L. Matuzaki, R. Santos-Silva, E. C. Marqueze, C. R. de Castro Moreno, S. Tufik, L. Bittencourt Temporal Sleep Patterns in Aduls Using Actigraph Sleep Science 7 152 (2014). [24] J.K. Ochab, J. Tyburczyk, E. Beldzik, D. R. Chialvo, A. Domagalik, M. F¸afrowicz, E. Gudowska- Nowak, T. Marek, M.A. Nowak, H. Ogi´nska, J. Szwed Scale-free fluctuations in behavioural performance: Delineating changes in spontaneous behaviour of humans with induced sleep deficiency PLoS One 9 e107542 (2014). [25] P. M. Holloway, M. Angelova, S. Lombardo, A.S.C. Gibson, D. Lee, and J. Ellis Complexity analysis of sleep and alterations with insomnia based on non-invasive techniques J. R. Soc. Interface 93 20131112 (2014). [26] F. Rieke, D. Warland, R. de Ruyter van Steveninck, W. Bialek Spikes: Exploring the neural code Bradford Book, MIT Press 1999. [27] N. Brunel, F.S. Chance, N. Fourcaud, L.F. Abbott Effects of synaptic noise and filtering on the frequency response of spiking neurons Phys. Rev. Lett. 86 2186 (2001). [28] M. Rudolph, A. Destexhe A multichannel shot noise approach to describe synaptic background activity in neurons Eur. Phys. J. B 52 125 (2006). [29] I. Eliazar, J. Klafter, A unified and universal explanation for L´evy laws and 1/f noises Proc. Natl. Acad. Sci. USA 106 12251 (2009). [30] I. Eliazar, J. Klafter, A probabilistic walk up power laws Physics Reports 511 143 (2012). [31] G. Werner, Fractals in the nervous system: Conceptual implications for theoretical neuroscience Frontiers in Physiology, 1 1 (2010). [32] J. Neyman, E.L. Scott, Spatial distribution of galaxies, analysis of the theory of fluctuations Proc. Natl. Acad. Sci. USA 40 873 (1954). [33] M.O. Vlad, J. Ross, F.W. Schneider L´evy diffusion in a force field, Huber relaxation kinetics, and nonequilibrium thermodynamics: H theorem for enhanced diffusion with L´evy white noise Phys. Rev. E 62 1743 (2000). [34] B. Kaulakys, M. Alaburda Modeling scaled processes and 1/f β noise using nonlinear stochastic differential equations J. Stat. Mech. Theor. Exp. P02051 (2009). [35] L. Laurson, X.Illa, M.J. Alana The effect of thresholding on temporal avalanche statistics J. Stat. Mech. Theor. Exp. P01019 (2009). [36] C.K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E. Stanley, A.L. Goldberger Mosaic 21 organization in DNA nucleotides Phys. Rev. E 49 1685 (1994). [37] P. Talkner, R.O. Weber Power spectrum and detrended fluctuation analysis: Application to daily temperatures Phys. Rev. E 62 150 (2000). [38] E. Rodriguez, J.C. Echeverria, J. Alvarez-Ramirez, Detrended fluctuation analysis of heart interbeat dynamics Physica A 384 429 (2007). [39] D. Makowiec, A. Rynkiewicz, R. Galaska, J. Wdowczyk-Szulc, M. Zarczynska-Buchowiecka Reading multifractal spectra: Aging by multifractal analysis of heart rate Europhys. Lett. 94 68005 (2011). [40] J. Gieraltowski, D. Hoyer, F. Tetschke, S. Nowack, U. Schneider, J. Zebrowski Development of multiscale complexity and multifractality of fetal heart rate variability Auton Neurosci. 178 29 (2013). [41] G. Sahin, M. Erenturk, A. Hacinliyan Detrended fluctuation analysis in natural languages using non-corpus parametrization Chaos, Solitons and Fractals 41 198 (2009). [42] U.T. Eden, M.A. Kramer Drawing inferences from Fano factor calculations J. Neurosci. Methods 190 149 (2010). [43] M. Magdziarz, A. Weron Anomalous diffusion: Testing ergodicity breaking in experimental data Phys. Rev. E 84 051138 (2011). [44] K. Burnecki, E. Kepten, J. Janczura, I. Bronshtein, Y. Garini, A. Weron Universal algorithm for identification of fractional Brownian motion. A case of telomere subdiffusion Biophysical Journal 103 1839 (2012). [45] P. Preuss, M. Vetter Discriminating between long-range dependence and non-stationarity, Electronic Journal of Statistics 7 2241 (2013).
1809.05023
1
1809
2018-09-13T15:51:31
Foraging as an evidence accumulation process
[ "q-bio.NC", "physics.bio-ph" ]
A canonical foraging task is the patch-leaving problem, in which a forager must decide to leave a current resource in search for another. Theoretical work has derived optimal strategies for when to leave a patch, and experiments have tested for conditions where animals do or do not follow an optimal strategy. Nevertheless, models of patch-leaving decisions do not consider the imperfect and noisy sampling process through which an animal gathers information, and how this process is constrained by neurobiological mechanisms. In this theoretical study, we formulate an evidence accumulation model of patch-leaving decisions where the animal averages over noisy measurements to estimate the state of the current patch and the overall environment. Evidence accumulation models belong to the class of drift diffusion processes and have been used to model decision making in different contexts. We solve the model for conditions where foraging decisions are optimal and equivalent to the marginal value theorem, and perform simulations to analyze deviations from optimal when these conditions are not met. By adjusting the drift rate and decision threshold, the model can represent different strategies, for example an increment-decrement or counting strategy. These strategies yield identical decisions in the limiting case but differ in how patch residence times adapt when the foraging environment is uncertain. To account for sub-optimal decisions, we introduce an energy-dependent utility function that predicts longer than optimal patch residence times when food is plentiful. Our model provides a quantitative connection between ecological models of foraging behavior and evidence accumulation models of decision making. Moreover, it provides a theoretical framework for potential experiments which seek to identify neural circuits underlying patch leaving decisions.
q-bio.NC
q-bio
Foraging as an evidence accumulation process Jacob D. Davidson1, 2, 3, ∗ and Ahmed El Hady4, 5, † 1Department Ecology and Evolutionary Biology, Princeton University 2Department Collective Behavior, Max Planck Institute for Ornithology 3Department of Biology, University of Konstanz 4Princeton Neuroscience Institute 5Howard Hughes Medical Institute (Dated: September 14, 2018) Abstract A canonical foraging task is the patch-leaving problem, in which a forager must decide to leave a current resource in search for another. Theoretical work has derived optimal strategies for when to leave a patch, and experiments have tested for conditions where animals do or do not follow an optimal strategy. Nevertheless, models of patch-leaving decisions do not consider the imperfect and noisy sampling process through which an animal gathers information, and how this process is constrained by neurobiological mechanisms. In this theoretical study, we formulate an evidence accumulation model of patch-leaving decisions where the animal averages over noisy measurements to estimate the state of the current patch and the overall environment. Evidence accumulation models belong to the class of drift diffusion processes and have been used to model decision making in different contexts especially in cognitive and systems neuroscience. We solve the model for conditions where foraging decisions are optimal and equivalent to the marginal value theorem, and perform simulations to analyze deviations from optimal when these conditions are not met. By adjusting the drift rate and decision threshold, the model can represent different "strategies", for example an increment-decrement or counting strategy. These strategies yield identical decisions in the limiting case but differ in how patch residence times adapt when the foraging environment is uncertain. To account for sub-optimal decisions, we introduce an energy- dependent utility function that predicts longer than optimal patch residence times when food is plentiful. Our model provides a quantitative connection between ecological models of foraging behavior and evidence accumulation models of decision making. Moreover, it provides a theoretical framework for potential experiments which seek to identify neural circuits underlying patch leaving decisions. 8 1 0 2 p e S 3 1 ] . C N o i b - q [ 1 v 3 2 0 5 0 . 9 0 8 1 : v i X r a ∗ [email protected] † [email protected] 1 INTRODUCTION In systems and cognitive neuroscience, decision-making has been extensively studied us- ing the concept of evidence accumulation [1, 2]. Evidence accumulation has been implicated for example in social decisions [3], sensory decisions [4, 10, 11], economic decisions [5], gam- bling decisions [6], memory decisions [7], visual search decisions [8], and value decisions [9]. Such processes of evidence accumulation can be well described quantitatively with a drift-diffusion model, providing a moment by moment estimate of the animal's internal ac- cumulation process. This quantitative modeling has given the experimenter the opportunity to investigate a myriad of neuronal mechanisms underlying these processes, for example the contributions of different cortical areas that differentially accumulate evidence over time [11 -- 19]. Although this work has revealed a detailed account of the neural mechanisms asso- ciated with decision-making, an outstanding question remains as to how these mechanisms have been shaped by selection forces in the animal's environment [20, 21]. Foraging is one of the most ubiquitous behaviors that animals exhibit, as search for food is essential for survival [22, 23]. From a cognitive perspective, foraging comprises aspects of learning, statistical inference, self-control, and decision-making, thus providing the opportunity to understand how these processes have been shaped by natural selection to optimize returns in the face of environmental and physiological constraints and costs [21]. There is an increased interest to study foraging behavior within a neuroscience context and link neural signals to relevant foraging parameters [24 -- 29]. For example, during a visual foraging task with non-human primates (Macaca mulatta), the activity in the dorsal anterior cingulate cortex (dACC) region was found to increase while a patch depletes until a threshold, after which the animal switches patches [24]. Other work has found that neurons in primate posterior cingulate cortex (PCC) signal decision salience during visual foraging, and thus relate to disengagement from the current patch [30]. A canonical foraging task is the patch leaving problem where an animal must decide when to leave a resource to search for another. Ecological models, such as the well-known marginal value theorem (MVT) [31], describe patch-leaving decision rules that an animal should use to optimize its food intake. Deviations from optimal decisions may be due to internal state-dependence or environmental characteristics [32]. Studies that link cognitive biases to environmental structure highlight the importance of studying the decision-maker in their natural environment, by framing decision making in terms of "ecological rationality" (as opposed to "economic rationality") [33 -- 36]. The MVT and related work provides a quantitative basis for understanding patch decisions, but does not give a mechanistic account of the animal's internal decision process and how it uses its experience to reach patch-leaving decisions. In this work we formulate a mechanistic model of patch leaving decisions by linking 2 ecological models of the patch leaving problem with models of evidence accumulation that are used in systems neuroscience. We call this model the foraging drift-diffusion model (FDDM). This model builds on previous mechanistics models of patch leaving decisions [37 -- 42]. In our model, patch-leaving decisions are described by a drift-diffusion process [43, 44], which represents the noisy process through which an animal accumulates evidence (by finding food) and uses its experience to decide when to leave the patch. Evidence accumulation and decisions within a patch are coupled to a moving average process that keeps track of the average rate of energy intake available from the environment. We solve for conditions where the model yields optimal foraging decisions according to the MVT, and perform simulations to analyze deviations from optimal when these conditions are not met. We then consider different decision "strategies", which are adaptive to different environmental conditions, and introduce a utility function into the model in order to account for sub-optimal foraging. More importantly, our model generates testable predictions about the different decision strategies an animal may employ in an uncertain environment. The model provides a quantitative connection between foraging behavior and experiments that seek to understand the neural basis of patch leaving decisions. For example, the model may be used to investigate how the neural activity of a particular brain area is tuned to the decision variable, or how different brain areas process other aspects of the decision making process detailed by the FDDM. RESULTS To present the model, we first define the governing equations then describe the dynam- ics of patch depletion by introducing parameters to represent the patch characteristics in the foraging environment. We then solve a simplified form of the model to establish condi- tions where the foraging drift-diffusion model yields identical decisions to the marginal value theorem. We show that optimal decisions can be represented in the model using different decision "strategies", including an increment-decrement mechanism, where receiving food reward makes the forager more likely to stay in the patch, and a counting mechanism, where receiving food reward makes the forager more likely to leave. Following this, we perform simulations with the general form of the model, to show how noise in the patch decision process and discrete food rewards affect energy intake and patch residence times. We then consider different configurations of the foraging environment, and present approximate so- lutions along with simulation results to show how the different decision strategies can be adaptive, depending on the uncertain versus known information about the foraging envi- ronment. Finally, we introduce a utility function into the model, and show how this can quantitatively account for the salient experimental observation that patch residence times tend to be longer than optimal. 3 Energy and patch decision variables E Estimated environment energy rate τE Timescale for updates of E r(t) Current gross rate of energy (food) intake s Constant cost x Decision state for when to leave a patch τ Timescale for updates of x α Drift rate η Threshold for decision to leave a patch. σ Noise for patch decisions W (t) Wiener process TABLE I. Variable definitions for the coupled model formulation in Eqs. 1-2. Foraging drift-diffusion model (FDDM) The model that we term foraging drift-diffusion model (FDDM) is a drift diffusion process that includes two coupled variables. The first calculates the energy intake available from the environment (E) by taking a moving average over a timescale τE. The second uses a drift-diffusion process for patch leaving decisions, which we represent by a patch decision variable x. Upon entering a patch x = 0, and changes in x occur with evidence accumulation from a constant drift α and a time-dependent reward function r(t). The forager decides to leave the patch when the threshold of x = η is reached. There is a constant cost of s, so that the net rate of energy gain while in a patch is r(t) − s, and while traveling between patches it is −s. The two equations are defined as: 1. Net energy intake rate τEdE = (r(t) − s − E) dt 2. Decision to leave a patch τ dx = (α − r(t)) dt + σdW (t), (1) (2) Fig 1 shows a schematic of the model, an example for the probability density of x when in a patch, and example traces of E and x across multiple patches (see Section S1 for details on the numerical simulation of the probability density of the patch decision variable). Table I lists the quantities defined in the governing equations. Patch characteristics We formulate equations to represent patch depletion, and incorporate a parameter that interpolates between continuous and discrete food rewards. The function r(t) describes the 4 FIG. 1. Foraging-Drift-Diffusion Model. (A) Schematic showing the patch leaving problem: A forager estimates the average rate of reward from the environment, and the decision to leave a patch occurs when the internal decision variable reaches a threshold. Travel time between patches is Ttr, and patches are described by the parameters ρ0, A, and c (see Table II). (B) Evolution of the probability density of the patch decision variable (x) while in a single patch, along with the time-dependent probability that the decision to leave the patch has been made. Blue arrows denote the receipt of food rewards. (C) Energy estimate coupled with the patch decision variable over multiple patches. (D) Patch depletion with discrete rewards, showing examples of the food reward received and the time-dependent in-patch food density for different values of the food chunk size (c). rate of food reward that the animal receives while in a patch, and ρ(t) is the density of food in the current patch. The initial density of food in the patch is ρ0, and when a forager finds and eats a piece of food, the total amount of food remaining in the patch decreases. To formalize this, we consider that patches have an area of a and that food is uniformly scattered within a patch in chunk sizes of c. If the forager searches at a rate of v, the probability of finding k chunks of food in a time interval ∆t is given by a Poisson distribution with event rate of ρ(t)v∆t/c: (cid:18) ρ(t)v∆t (cid:19) c Pk = Poisson , k . (3) When food is found, the total amount of food remaining is reduced by an amount kc. On 5 AB1.00.50.00.51.00123456780.00.20.40.60.81.0Time in patch (normalized units)Probability of leavingcurrent patch: P(t)Expected patchresidence time: Tx(t)/(t)DTime in patch (normalized units)01234Food reward:r(t)dtc=2c=4051015200.02.55.07.510.0Patch fooddensity: ρ(t)c=2c=4(t)Total time (normalized units)C012EnergyOptimum EE(t) energy estimateE (actual avg)Minimum to surviveFood received0510152025301.00.50.00.51.0x(t)/(t)In patchρ0, τP, cTtrTtr Initial food density Patch variables and parameters ρ(t) Time-dependent food density in the current patch ρ0 A Patch size c Food chunk size Ttr Travel time between patches TABLE II. Variables and parameters used to describe patch quality and depletion. average, the total amount of food, aρ(t), changes according to a(cid:104)ρ(t + 1)(cid:105) → a(cid:104)ρ(t)(cid:105) − (cid:104)k(cid:105) c, (4) where (cid:104)·(cid:105) denotes an ensemble average. Using Eq. 3, the ensemble average for the number of pieces of food found in one time step is (cid:104)k(cid:105) = ρ(t)v∆t/c, and using this, average change in density follows a linear differential equation [45]: d(cid:104)ρ(cid:105) dt A = −(cid:104)ρ(cid:105) , (5) where A = a/v is the effective time constant of the patch. Without loss of generality, we set v = 1, i.e. the forager explores one unit area per unit time, and therefore refer to A as the "patch size", with units of time. Average patch depletion (as well as the average rate of reward received) follows a simple exponential decay: (cid:104)ρ(t)(cid:105) = (cid:104)r(t)(cid:105) = ρ0e−t/A, (6) where t is the time spent in the current patch. Fig 1D shows example time traces of patch density and food received for different values of the chunk size c. With larger chunk size there is larger variability in the food rewards found per unit time. In limit of zero chunk size, food reward is continuous and the food reward rate and patch density are equal to the average density from Eq. 6: limc→0 r(t) = limc→0 ρ(t) = (cid:104)ρ(t)(cid:105). Optimal foraging decisions and model equivalence to marginal value theorem We solve the model to establish conditions on the drift rate α and the decision threshold η which lead to optimal patch residence times. To do this, we consider E = (cid:104)E(cid:105) = const., (the estimated value of energy is constant and equal to the actual average), σ = 0, (no noise on the patch decision variable), and c = 0 (food reward is received continuously). We establish an equivalence between the FDDM and the marginal value theorem for this case, then relax these assumptions and use simulations to show that the derived rules lead to approximately optimal patch decisions over a wide range of parameter values and configurations of the foraging environment. 6 First, we rewrite the marginal value calculation for patch residence time using the above notation. If there is a travel time between patches of Ttr, then the average rate of energy intake is given by a weighted sum of intakes during time spend in patches and traveling between patches. Taking the derivative of the average energy intake rate, setting to zero, and re-arranging, yields the well-known condition to solve for the optimal time Topt to stay in a patch: r(Topt) − s = (cid:124) (cid:123)(cid:122) (cid:125) marginal in-patch rate r(t)dt − s ∗ (Ttr + Topt) Ttr + Topt (cid:104)E(cid:105)=average energy rate (cid:123)(cid:122) (cid:125) (cid:82) Topt (cid:124) 0 (7) Eq. 7 can be written more compactly in the form r(T ) − s = (cid:104)E(cid:105), where (cid:104)E(cid:105) is the average energy rate from the environment. Using the average patch depletion dynamics (Eq. 6) in Eq. 7, we can solve for the optimal time to remain in a patch: Topt = A ln ρ0 (cid:104)E(cid:105) + s . (8) Integrating the patch decision variable (Eq. 2) to the threshold and inserting Eq. 8 for the optimal patch residence time yields a relationship between the threshold, drift rate, energy, and patch parameters: (cid:18) (cid:19) (cid:18) ρ0 (cid:104)E(cid:105) + s η = A α ln (cid:19) − ρ0 + (cid:104)E(cid:105) + s . (9) If Eq. 9 is satisfied, optimal decisions can be obtained with different values of the drift rate α. To determine a valid range for α values, in Section S3 we solve for conditions on the drift rate such that there is only a single threshold crossing up to the time Topt. In addition to this, we omit the small range where α and η have opposite signs. With these conditions, the valid range for the drift rate is α ≤ 0, or α ≥ ρ0 − (cid:104)E(cid:105) − s ln ρ0 (cid:104)E(cid:105)+s , (10) Using this range, and also substituting E, the energy estimate, for (cid:104)E(cid:105), we highlight the following different "strategies": Density-adaptive: αD = ρ0 Size-adaptive: ηS = 0 ρ0 − E − s αS = Counting: αC = 0 Robust counting: αR < 0. 7 ln ρ0 E+s (11) For the density-adaptive, counting, and robust counting strategies, η is defined by Eq. 9 with the corresponding value of α, and substituting E instead of (cid:104)E(cid:105). These strategies are illustrated in Fig 2 using zero noise and discrete rewards. The density-adaptive strategy (Fig 2A) uses an increment-decrement mechanism [39, 46] which in previous work has been suggested as adaptive for the case when the forager does not initially know the number of expected reward items on the patch [47]. In the FDDM framework, this is implemented using η > 0 and α > 0, so that finding food makes the forager more likely to stay in the patch, but otherwise the drift brings the x towards the threshold. We show in Section S4 that using αD is optimal to adapt PRTs to uncertain food density within each patch. The size-adaptive strategy also uses an increment/decrement mechanism, but with different values of α and η that are optimal to adapt PRTs with respect to uncertainty in the size of each patch (Section S4). This strategy uses a threshold of zero, such that x first decreases below zero and then rises back to the threshold. Because x both starts and ends at zero, the size-adaptive strategy is sensitive to noise and randomness in the timing of rewards received. We therefore illustrate this strategy in Fig 2B by choosing a value α > αS, which yields η > 0. The counting strategy has zero drift rate and a negative value of the threshold. With this strategy, finding food makes the forager more likely to leave the patch, and the forager leaves only after a set amount of food reward has been received (Fig 2C). Since the choice of α = 0 in the counting strategy can cause PRTs to become infinite if patches do not contain as much food as expected, we define an additional strategy termed 'robust counting' which has a nonzero drift α < 0. With this, there is still drift towards the threshold in the absence of food reward; thus, in contrast to the density-adaptive or size-adaptive strategies, both drift and receiving food reward bring x closer to the threshold (Fig 2D). The size-adaptive and counting strategies represent limiting cases of η = 0 and α = 0, respectively, and this makes these choices sensitive to noise. We therefore focus our analysis on the density-adaptive and robust counting strategies, which have (α > 0, η > 0) and (α < 0, η < 0), respectively. Patch decisions using these strategies are exactly equivalent to the marginal theorem for the case of E = (cid:104)E(cid:105), σ = 0, and c = 0. In the next section we investigate model behavior and compare simulated patch decisions to optimal for a range of parameter values in the general case of E (cid:54)= (cid:104)E(cid:105), σ > 0, and c > 0. Parameter dependence: noisy decisions and discrete food rewards In the general case, individual patch decisions will be noisy, food may come in discrete chunks, the estimate of available energy in the environment will vary as the forager explores and obtains food rewards, and patches may vary in quality and distribution. We investigate both a range of environmental configurations and patch parameters as well as different patch decision strategies. 8 FIG. 2. Patch leaving decision strategies. Different strategies are represented with different choices of the drift rate (α) and the threshold (η) (Eq. 11). (A) The choice α = ρ0 is optimal for uncertainty in patch food density; this represents an "increment-decrement" mechanism for patch decisions. (B) A threshold of zero is optimal for uncertainty in patch size. Since η = 0 is sensitive to noise, we choose a small value η > 0 to illustrate. (C) The counting strategy uses zero drift, so that the forager leaves after a set amount of food rewards (D) The robust counting strategy uses α < 0 so that there is still drift towards the threshold. Each plot shows the patch decision variable along with the time-dependent patch decision threshold that changes with receipt of food reward due to updates of energy estimate. To simplify model analysis, we use τ as the unit of time, and s as the unit of energy, and set τE = 50τ to represent that the energy estimate occurs at a longer time scale that individual patch decisions. We illustrate dominant trends by choosing an intermediate range for characteristics of the foraging environment: E = 2s, A = 5τ , and Ttr = 5τ . Simulation results for a range of different configurations defined in Section S2 are shown in Figs S1 and S2. Fig 3A shows that for small increases of noise on the patch decision variable, both the mean energy intake and mean patch residence time stay near optimal values, but the variance 9 x(t)η(t) (threshold)In patchFood receivedADensity-adaptive (α=ρ0)BSize-adaptive (η→0)CCounting (α=0)DRobust counting (α<0)x(t), η(t)x(t), η(t)x(t), η(t)x(t), η(t)201000200510152025303540Total time(normalized units)402003020100 FIG. 3. Noisy patch decisions and discrete food rewards. Shown are the average and standard deviation of the energy intake and patch residence times, simulated using intermediate values of the patch parameters: A = 5, Ttr = 5, and Eopt = 2 (or equivalently, ρ0 = 9.439). The filled blue curves use the density-adaptive strategy, and the filled orange curves use the robust counting strategy. The robust counting strategy simulations use α = −0.2ρ0. (A) Simulation results when the noise on the patch decision variable (σ) is increased. (B) Simulation results when the food chunk size (c) is increased. of patch residence time increases. Even with zero noise, the mean simulated PRTs are slightly lower than optimal; this is due to the finite time scale for the moving average estimate of E. Because E increases within a patch and then decreases outside of a patch, E tends to be slightly higher than the actual average energy when the agent leaves the patch (see Fig 1C). The increase in E causes the threshold to decrease in magnitude before the forager leaves the patch (see Fig 2), which is why the average simulated PRTs are slightly less than optimal. With higher values of σ, the simulated average energy intake decreases, and the effect is larger for the robust-counting (RC) strategy compared to the density-adaptive (DA) strategy. With the DA strategy, the variance of patch residence time increases with noise, but the average stays nearly the same. With the RC strategy, the variance increases more strongly with noise, and for large values of σ, the average patch residence time is longer than optimal. Fig 3B shows average energy intake and patch residence time when the food chunk size (c) increases. For both strategies, larger chunk sizes increase the variance of PRTs without much effect on the mean. However, the two strategies show opposite trends for average energy intake: with the DA strategy, average energy decreases for large chunk size, but with the RC strategy, average energy increases for large c, to values that are higher than 10 Noisy patch decisions012Energy (units ofs)0.000.250.50Noise: /0051015PRT (units of )0510cFood chunk size: Topt<T>Density-adaptive strategyRobust counting strategyFood in discrete chunksEopt<E>012Energy (units ofs)051015PRT (units of )ABEopt<E>Topt<T> the optimum determined by the marginal value theorem. This is because with a counting strategy, food reward makes the forager more likely to leave the patch, and therefore patch leaving decisions tend to occur immediately after receipt of a food reward, instead of after a certain amount of time in the patch (Fig 2). With large chunk sizes, the number of food chunks per patch will be small, and there- fore instantaneous food intake and leaving decisions are not well described by a 'rate', as expressed with the MVT. The optimum number of food chunks obtained per patch is Nopt = (ρ0 − E − s) . A c (12) For example, using parameter values from Fig 3, a chunk size of c = 8 leads to Nopt = 4.02. In this case it is difficult to assess current food density, which is why average energy intake with the DA strategy is less than optimal. For extreme cases where Nopt < 1, which occurs for example in small patch size, short inter-patch travel times, and low available energy in the environment, the DA strategy performs poorly, while the RC strategy yields average energy intake rates that are higher than MVT optimal (Fig S2). Patch uncertainty and adaptive decisions To this point we have considered cases where patch quality and inter-patch travel times are the same for all patches; we now ask how the different strategies perform when aspects of the foraging environment are uncertain and may vary from patch to patch. The MVT predicts that foragers should stay longer in high quality patches, and shorter in low quality patches. However, this assumes that as they enter a patch, the forager recognizes the 'type' of the patch and therefore adjusts their expectation of food rewards. We instead consider that the forager only knows the average patch quality in the environment, and must use this along with the estimate of E and its current experience of food rewards to determine when to leave a patch. We first consider the case that patch quality is uncertain, by varying the initial food density of each patch. This is simulated by drawing the initial density (ρ0) from a Gaussian distribution with mean ¯ρ0 and standard deviation ∆ρ0. If the density of each patch is known, the MVT predicts that the forager should adjust its PRT according to Eq. 8. For the FDDM, we show in Section S4 that changes in patch residence time (T ) in response to a small change in patch food density, ρ0 = ¯ρ0 + δρ0, are approximated by T ≈ Topt + ¯A (− ¯ρ0 + E + s) ¯ρ0(−α + E + s) δρ0. (13) With the DA strategy, foragers stay longer in higher quality patches (i.e. patches with higher ρ0) and shorter in lower quality patches (i.e. lower ρ0), and changes from patch to patch 11 FIG. 4. Different foraging environments with associated patch decision strategies. Shown are simulation results with the density-adaptive and robust-counting strategies in two dif- ferent foraging environments. The left column illustrates the foraging environment for a given case, the middle column shows average energy and patch residence time when a particular strategy is used in that environment, and the right column shows simulation results compared to optimal strategies in each environment. All simulations use a noise level of σ = 0.3 ¯ρ0 and a patch size of A = 5, and the robust counting strategy is implemented by setting α = −0.2ρ0. (A) Uncertainty in patch food density. Patches have a Gaussian distribution for initial food density with mean of ¯ρ0 = 9.439 and a standard deviation of ∆ρ0 = 0.3 ¯ρ0, and rewards are received continuously (c = 0). Travel time between patches is constant at Ttr = 5. (B) Scattered patches with discrete rewards. Food reward is received in discrete chunks (c = 8) and each patch has the same initial food density of ρ0 = 9.439. Travel time between patches is drawn from an exponential distribution with mean ¯Ttr = 5. asymptotically follow optimal adjustments. In contrast, the RC strategy yields the opposite trend: patch residence time decreases with patch quality (Fig 4A). Simulations with added patch decision noise agree well with Eq. 13 for small changes in ρ0 about ¯ρ0, but for the RC strategy there are deviations from the linear trend for large changes. This demonstrates that for an environment where patch food density varies, the DA strategy yields an average energy intake and PRT close to optimal, while using the RC strategy yields an energy intake lower than optimal due to PRTs that are higher than optimal (Fig 4A). We next consider a different configuration of the foraging environment: food is received in discrete chunks, patches are randomly distributed about the landscape, but the quality of each patch is the same. Because each patch contains the same amount of food, an optimal strategy is to leave a patch after a certain amount of food reward is received, and thus a 'counting' strategy is expected to be optimal. We represent randomly scattered patches by drawing inter-patch travel times from an exponential distribution with mean ¯Ttr. Simulations with noise show that in this environment, the RC strategy leads to a higher average energy intake than the DA strategy (Fig 4B). This is because the distribution of number of food 12 Strategy 1:Density adaptiveStrategy 2:Robust countingA010200.000.050.100.15P(0)0Uncertain patch food density0123EnergyAverage energyobtainedMVT optimal valueStrategy 1: Density adaptiveStrategy 2: Robust counting051015PRT01020Patch food density (0)051015PRTOptimalEq. 13Simulation01020Patch food density (0)OptimalEq. 13Simulation12Strategy0123Energy12Strategy051015PRT012345678Chunks taken/patch0.00.20.40.6Probability012345678Chunks taken/patch0510150.00.10.2P(Ttr)TtrBScattered patches with discrete rewardsAverage patchresidence timePatch food density (0)Travel time (Ttr)OptimalOptimal items per patch is sharply peaked near the optimal value for the RC strategy, while the distribution is broader with the peak skewed from optimal for the DA strategy (Fig 4B). Similar to Fig 3B, Fig 4B shows that the RC strategy leads to mean energy intakes that are higher than the optimum predicted by the MVT, because patch leaving decisions tend to occur immediately following the receipt of food reward. Another type of patch uncertainty can come from patches that vary in size. The size- adaptive strategy defined in Eq. 11 yields adjustments to PRTs based on the size of each patch that follow, in the limiting case of zero noise, the optimal times given by Eq. 8. How- ever, because the size-adaptive strategy has a threshold of zero, it is very sensitive to noise. In simulations with added noise, using a strategy close to the size-adaptive strategy with a small but nonzero threshold yields similar or slightly lower average energy intakes com- pared to the density-adaptive strategy when patch size is uncertain (Fig S3). This suggests that while a forager with an appropriate strategy can nearly optimally adapt individual patch residence times to uncertainty in patch food density, it is more difficult to use a noisy sampling process to adapt individual patch residence times to uncertainty in patch size. Sub-optimal behavior: satisficing In the previous section we showed that the FDDM can represent different decision strate- gies that are appropriate to optimize the response to uncertainty in different environments. However, with the exception of the RC strategy in an environment where patch quality is un- certain, simulations yield average PRTs that are near or slightly lower than optimal. Many studies have examined patch residence times in comparison to MVT predictions; the most common trend is that animals tend to stay longer in patches than predicted by the MVT [32]. In this section we introduce a change to the model to account for this observation. An animal's perception of a reward, and subsequent foraging decisions, depend on their internal state. One way to capture this is by using a utility function approach, borrowed from behavioral economics [48 -- 50]. Conceptually this is also related to 'satisficing' [51 -- 56], defined as the process by which animals do not seek to maximize food intake, but instead seek to maintain food intake above a threshold. If food is plentiful, then the utility of increasing intake is small; in this case, an animal will likely be more concerned with, for example, avoiding threats than leaving a current patch in search of higher returns. Conversely, if food is scarce, then survival depends on maximizing the rate of food rewards. The utility concept can represent these scenarios. We model this by introducing a utility function u(E), Which depends on available energy in the environment. The utility function modifies patch decision dynamics by changing the drift rate and the impact of receiving food: τ dx =(cid:0)αu(E) − r(t)u(E)−sgn(η)(cid:1) dt + σdW (t). (14) 13 Using this form, the utility function decreases the rate of drift towards the threshold, and either increases or decreases the change in x with food reward depending on whether the threshold is positive or negative. To set a form for u(E), first recall that the animal must obtain energy E > 0 in order to survive. In the limit E → 0, we therefore expect that an animal will adopt a foraging strategy that maximizes energy intake; this is set by u(0) = 1. For high values of E, we expect that the animal cares less about maximizing food intake rate, and therefore u should decrease. We consider two functions to represent this: uexp(E) = (1 − A)e−βE + A 1 − βE if 1 − βE ≥ A if 1 − βE < A ulin = A , (15) (16) where β > 0 is a parameter that determines how fast the utility changes with energy We choose these forms for u(E) to investigate the model response, and note that other functional forms can be used. Integrating Eq. 14 using either Eq. 15 or 16, setting σ = 0 and E = (cid:104)E(cid:105), and com- bining with Eqs. 7-8 yields an approximate correction for how the utility function affects both average intake energy intake and patch residence time. We compare this approximate solution with simulation results in Fig 5. The simulations use the density adaptive strategy in an environment where patch food density is uncertain (i.e. the same configuration as Fig 4A). The results show that using either form of the utility function leads to patch deci- sions that approach optimal when energy is low, but deviate from optimality when energy is high. Patch residence times are longer than optimal when the available energy is high, in particular for the larger value of β show in Fig 5. Although both forms of the utility function demonstrate longer than optimal patch residence times, the change of PRTs with energy levels depends on whether the exponential or threshold linear form is used. In both cases, simulation results agree reasonably well with the approximate solution. Analogous results for the robust counting strategy, and for an environment with discrete rewards and uncertain travel times, are shown in Fig S4. DISCUSSION In this study, we developed a foraging drift-diffusion model (FDDM) to describe how an animal accumulates evidence over time in the form of food rewards and uses this experience to decide when to leave a foraging patch. We solved for conditions where the FDDM yields identical decisions to the marginal value theorem, and performed simulations to show how deviations from optimality are affected by noisy patch decisions and discrete versus contin- uous food rewards. By adjusting the drift rate and the threshold for patch decisions, the 14 FIG. 5. Sub-optimal behavior. The utility of receiving additional food reward may depend on the current rate of energy intake. We consider two possible functions: (A) Exponential decreasing utility, shown using A = 0 in Eq. 15. (B) Threshold linear decreasing utility (Eq. 16), shown here using a threshold of 0.65. Each form of the utility function has a parameter β that sets how fast the utility decreases with energy. Simulation results using the exponential utility function are shown in (C), and corresponding results using the threshold linear utility function in (D). For each case of the utility function, the average energy intake and patch residence time are shown for two different values of β. Solid lines are an approximate solution to the governing equations and points are the mean and standard deviation of simulation results. Both (C) and (D) use the density-adaptive strategy, and the environmental configuration and other parameters from Fig 4A. Analogous results for the robust counting strategy, and an additional environmental configuration, are shown in Fig S4. model can represent different decision strategies, including an increment-decrement mecha- nism, where finding food makes the animal more likely to stay in the patch, and a counting mechanism, where finding food makes the animal more likely to leave the patch. We ob- tained approximate solutions in addition to model simulations to demonstrate how these different strategies are adaptive, depending on the known and unknown aspects of the for- aging environment. We then showed that incorporating a utility function into the model can quantitatively account for the common experimental observation that patch residence times tend to be longer than optimal. Our model links ecological models of patch foraging with drift-diffusion models of decision making, and by representing both MVT-optimal behaviors and sub-optimal deviations, it provides a mechanistic description which yields predictions for future experimental work. The FDDM model builds on a body of previous work that has considered statistics of patch depletion [45], averaging mechanisms to estimate available energy [57, 58], and "patch 15 AB024Energy (units ofs)0.00.51.0uexp(E)Exponential utility function=0=0.1=0.30123456Energy (units ofs)Avg. energy intake010203040PRT (units of )Patch residence times024Eopt024Eopt=0.1=0.3E=Eopt=0.1=0.3Topt024Energy (units ofs)0.00.51.0ulin(E)Threshold linear utility function=0=0.1=0.30123456Energy (units ofs)Avg. energy intake010203040PRT (units of )Patch residence times024Eopt024Eopt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3ToptCDApproximate solution Simulation leaving potentials" or other mechanistic descriptions of when to leave a patch [37 -- 42]. Re- wards that come stochastically in discrete chunks have been treated in [45, 59], for example the case of Poisson distributed rewards we considered as well as more general scenarios [45]. The moving average of energy in the environment that we use is similar to a recent experience-driven model, in that the timescale for updates of energy is finite [58]. Other models have also considered that the forager creates a moving average of the average prof- itability of the environment [57]. Previous mechanistic models of patch leaving decisions have proposed that a forager has a patch potential, which declines in the absence of food and increases when food is found, and then the forager leaves when the potential crosses zero [37, 38]. This model has been used to analyze the behavior of parasitoid wasps seeking hosts, where it is referred to as the incremental-decremental model [39]. In this model, find- ing a successful host increases the potential, and thus the probability of staying, and in the absence of a finding a successful host the potential continues to move towards the thresh- old. The incremental/decremental model of parasitoid host decisions has been extended to consider different types of encounters, i.e with parasitized or unparasitized individuals [40] The increment/decrement mechanism has been modeled in a similar way by a considering a leaving potential, which reflects how likely the forager is to leave [39, 41]. Another simi- lar related model, based on the concept of increment-decrement, has been used to analyze bumblebee foraging [42]. A different class of models have considered patch leaving rules where the forager "counts", for example leaving after a single food item has been found [60]. By combining these different concepts into a single model with a tractable analytical form, we provide a framework for future experiments that seek to understand different de- cision strategies that may depend on environmental characteristics, neural dynamics, and state-dependence of the animal. The utility-function approach represents foraging decisions which lead to sub-optimal energy intake and longer than optimal patch residence times. This formulation relates to the mechanisms of satisficing and temporal discounting. Satisficing refers when animals do not seek to maximize food intake, but instead to maintain food intake above a threshold [51 -- 56]. Temporal discounting refers to when the animal values current rewards more than expected future rewards [61]. This can be called other things as well, for example impulsivity, failure to delay gratification, and delay discounting [62, 63], and studies have examined this phenomena in the context of inter-temporal choice tasks [64 -- 66]. Various models of cognitive biases have been formulated to represent staying in a patch longer than optimal (which is also referred to as "over-harvesting" the patch). One way is to define a subjective cost that approximates the aversion to leave the patch [65, 67]. Another method is discounting of future rewards, so that for example an expected large reward in a new patch is discounted because of the time delay until which it is available [68]. An alternative interpretation uses a decreasing marginal utility function, such that an expected large reward in a new patch is 16 not viewed as proportionally better than the current low rate of reward in an almost-depleted patch, for example due to costs associated with switching patches [69]. In our model, the utility function describes an explore-exploit tradeoff; if food is plentiful, then the relative utility of leaving the current patch to find a new patch with possibly higher rewards is small, and therefore the animal strays longer in the current patch. The reason for this could be that the animal is satisfied with its current rate of food intake, or that due to other factors (e.g. risks involving with continued search), it values receiving smaller, certain rewards in the present moment instead of leaving to obtain uncertain but possibly larger rewards. We investigated two examples for the form of the utility function in Fig 5, and note that an interesting area for future work is to ask how an animal's perception of the value or utility of a reward depends on internal state and external environmental conditions. Foraging decisions differ from common models of economic choice in a key aspect: de- cisions are sequential, instead of between discrete alternatives [73]. Experiments with the self-control preparation, where an animal must choose between two alternatives, and the patch preparation, which is a sequential foraging preparation, have seen behavioral differ- ences even though from an economic standpoint the setups are equivalent [62]. Additionally, when rats are required to physically move to perform foraging, the observed behavior differs from tasks that simulate foraging by presenting sequential choices or that consider visual search [65]. These studies highlight the importance of state-dependence and context to understand decision-making processes. In future work it will be interesting to understand the neural basis for why these treatments differ, and how this contributes to state- and environment-dependent decision biases. In this study we modeled a single forager acting independently. Often times a more realistic situation involves others agents who simultaneously exist in the environment, which leads to competitive and/or collective foraging. If foragers are competing for resources, the ideal free distribution theory describes an optimal way to distribute multiple agents at different food sources in relation to the quality of food sources and the density of competition [74]. In other cases a group may forage together collectively, leading to individual decisions that incorporate both non-social and social information (e.g. [75]). Patch-leaving decisions will then depend on the group reaching consensus. The drift-diffusion modeling framework has been extended to represent coupled decision-makers who share information to collectively reach a decision [76], and this approach could be used to extend the FDDM to multiple agents who make decisions as a group. We considered that the forager knows the average patch food density ( ¯ρ0) and the average patch size ( ¯A), and uses these to set an optimal decision "strategy" by choosing values of the drift rate (α) and threshold (η). Other models have considered the process of learning about the environment during foraging using reinforcement learning [77]. Reinforcement learning (RL) is a framework to represent how an agent that receives information about the 17 state of the world along with a scalar valued reward signal learns to select actions which maximize the long run accrued reward. Kolling and Akam [77] reframed the MVT rule as an average reward RL algorithm, which estimates relative values of staying and leaving using a particular assumption about the patch's reward rate dynamics. To incorporate RL into the FDDM, one possibility is that the agent has to learn the patch characteristics ( ¯ρ0, ¯A), and then uses these learned values to set α and η. Another possibility is that the agent could adjust α and η directly, based on feedback from the amount of reward received. Bayesian foraging theories have considered how patch foraging decisions should be based on a prior estimate of the distribution of patches and expected reward in the environment [78 -- 81]. For example, if you know the variability of the environment, you should adjust the strategy. If you know patches contain a set number of reward items, then finding a prey item should decrease your probability of staying at the patch (i.e. you should choose a counting strategy). Conversely, if you know that patches vary in their quality, finding a food item should increase the probability that you stay in the patch (i.e. the density-adaptive strategy is a good choice). Experimental work has shown that bumblebees make exactly this adjustment to their patch-leaving strategies [82], but bluegill fish do not [83]. Other studies have considered the effect of reward uncertainty (e.g. [84, 85]), suggesting that foragers may not follow optimal rules when patch quality is uncertain [86]. From our simulation results, one possible explanation for sub-optimal decisions when the foraging environment is uncertain is adopting the "wrong strategy" (Fig 4). In the FDDM, the forager has memory of its previous foraging experience through the estimate of available energy. A related question is how foraging decisions are affected when the environment changes over time, which for example can lead to biases from contrast effects [85]. The speed of environmental fluctuations affects which strategy is optimal [87], and the relative importance of taking different adaptive strategies depends on the dynamics and predictability of the environment [59]. Spatio-temporal autocorrelation is a common feature of natural environments, and this may have driven certain observed decision biases [88]. Related to this, work has shown that patch time allocation is influenced by recent experiences of travel time [89 -- 91], and patch quality [92 -- 94]. In summary, in this work we developed a mechanistic model of a natural behavior (for- aging), with a mathematical form inspired by models used in systems neuroscience. The model considers an agent that calculates a moving average of the available energy in the environment, and makes noisy patch decisions according to the receipt of food rewards and a decision "strategy", which can be adapted to optimize for the characteristics of the foraging environment. This work provides a step towards establishing a unifying framework tying concepts from systems neuroscience, ecology and behavioral economics to study naturalistic decision making. With the advent of functional imaging [95 -- 97] and wireless eletrophysio- logical techniques in freely moving animals [98 -- 102], one can monitor different brain areas 18 simultaneously along with the detailed movement and postural dynamics of the animal [103 -- 108], with the aim to map the involvement of both neurobiological and biomechanical mechanisms that relate to certain aspects of behavior. Additionally, recent advancements in closed loop techniques allow precise perturbations of neural systems that depend on the state and current behavior of the animal [109 -- 111]. The proposed model provides a moment- by-moment estimate of the evolution of the decision process, which enables future work to map brain activity to quantitative behavioral variables using neural recordings and targeted perturbations. AUTHOR CONTRIBUTIONS JDD and AEH have both conceived the project, formulated the model conceptually, solved the model and wrote the manuscript. ACKNOWLEDGMENTS We would like to thank the Marine Biological Laboratory and course organizers Mark Goldman and Michale Fee, where these ideas were first developed during the Methods in Computational Neuroscience summer school. We would especially like to thank Sylvia Guil- lory for initial work on the topic and helpful discussions. REFERENCES [1] Carlos D Brody and Timothy D Hanks. Neural underpinnings of the evidence accumulator. Current opinion in neurobiology, 37:149 -- 157, 2016. [2] Timothy D Hanks and Christopher Summerfield. Perceptual decision making in rodents, monkeys, and humans. Neuron, 93(1):15 -- 31, 2017. [3] Ian Krajbich, Dingchao Lu, Colin Camerer, and Antonio Rangel. The attentional drift- diffusion model extends to simple purchasing decisions. Frontiers in psychology, 3:193, 2012. [4] William T Newsome, Kenneth H Britten, and J Anthony Movshon. Neuronal correlates of a perceptual decision. Nature, 341(6237):52, 1989. [5] Sebastian Gluth, Jorg Rieskamp, and Christian Buchel. Deciding when to decide: time- variant sequential sampling models explain the emergence of value-based decisions in the human brain. Journal of Neuroscience, 32(31):10686 -- 10698, 2012. 19 [6] Jerome R Busemeyer and James T Townsend. Decision field theory: a dynamic-cognitive approach to decision making in an uncertain environment. Psychological review, 100(3):432, 1993. [7] Roger Ratcliff. A theory of memory retrieval. Psychological review, 85(2):59, 1978. [8] Braden A Purcell, Richard P Heitz, Jeremiah Y Cohen, Jeffrey D Schall, Gordon D Lo- gan, and Thomas J Palmeri. Neurally constrained modeling of perceptual decision making. Psychological review, 117(4):1113, 2010. [9] M. Milosavljevic, J. Malmaud, A. Huth, C. Koch, and A. Rangel. The drift diffusion model can account for the accuracy and reaction time of value-based choices under high and low time pressure. Judgment and Decision Making, 5(6):437449, 2010. [10] Bingni W. Brunton, Matthew M. Botvinick, and Carlos D. Brody. Rats and humans can optimally accumulate evidence for decision-making. Science, 340(6128):95 -- 98, 2013. [11] Timothy D. Hanks, Charles D. Kopec, Bingni W. Brunton, Chunyu A. Duan, Jeffrey C. Erlich, and Carlos D. Brody. Distinct relationships of parietal and prefrontal cortices to evidence accumulation. Nature, advance online publication, January 2015. [12] G. Deco, E.T. Rolls, L. Albantakis, and R. Romo. Brain mechanisms for perceptual and reward-related decision-making. Prog. in Neurobiol., 103:194 -- 213, 2013. [13] K. Doya and M.N. Shadlen (Eds). Decision making. Curr. Opin. Neurobiol., 22 (6), 2012. Special issue. [14] B.A. Purcell, R.P. Heitz, J.Y Cohen, J.D. Schall, G.D. Logan, and T.J. Palmeri. Neurally constrained modeling of perceptual decision making. Psychol. Rev., 117(4):1113 -- 1143, 2010. [15] J.N. Kim and M.N. Shadlen. Neural correlates of a decision in the dorsolateral prefrontal cortex of the macaque. Nat. Neurosi., 2:176 -- 185, 1999. [16] G.D. Horwitz and W.T. Newsome. Representation of an abstract perceptual decision in macaque superior colliculus. J. Neurophysiol., 91:2281 -- 2296, 2004. [17] L. Ding and J.I. Gold. Caudate encodes multiple computations for perceptual decisions. J. Neurosci., 30:15747 -- 15759, 2010. [18] L. Ding and J.I. Gold. Neural correlates of perceptual decision making before, during, and after decision commitment in monkey frontal eye field. Cereb. Cortex, 22:1052 -- 1067, 2012. [19] L. Ding and J.I. Gold. Separate, causal roles of the caudate in saccadic choice and execution in a perceptual decision task. Neuron, 75:865 -- 874, 2012. [20] John W. Krakauer, Asif A. Ghazanfar, Alex Gomez-Marin, Malcolm A. MacIver, and David Poeppel. Neuroscience needs behavior: Correcting a reductionist bias. Neuron, 93(3):480490, Feb 2017. [21] Dean Mobbs, Pete C Trimmer, Daniel T Blumstein, and Peter Dayan. Foraging for founda- tions in decision neuroscience: insights from ethology. neuroscience, 13(18):19, 2018. [22] David W Stephens and John R Krebs. Foraging theory. Princeton University Press, 1986. 20 [23] David W. Stephens, Joel S. Brown, and Ronald C. Ydenberg. Foraging: Behavior and Ecology. University of Chicago Press, Sep 2008. ISBN 978-0-226-77265-3. [24] Benjamin Y Hayden, John M Pearson, and Michael L Platt. Neuronal basis of sequential foraging decisions in a patchy environment. Nature neuroscience, 14(7):933, 2011. [25] Amitai Shenhav, Mark A Straccia, Jonathan D Cohen, and Matthew M Botvinick. Anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value. Nature neuroscience, 17(9):1249, 2014. [26] Adam J Calhoun, Sreekanth H Chalasani, and Tatyana O Sharpee. Maximally informative foraging by caenorhabditis elegans. Elife, 3, 2014. [27] Adam J Calhoun and Benjamin Y Hayden. The foraging brain. Current Opinion in Behav- ioral Sciences, 5:24 -- 31, 2015. [28] Benjamin Y Hayden and Mark E Walton. Neuroscience of foraging. Frontiers in neuroscience, 8:81, 2014. [29] Fang Li, Mingbo Li, Wenyu Cao, Yang Xu, Yanwei Luo, Xiaolin Zhong, Jianyi Zhang, Ruping Dai, Xin-Fu Zhou, Zhiyuan Li, et al. Anterior cingulate cortical lesion attenuates food foraging in rats. Brain research bulletin, 88(6):602 -- 608, 2012. [30] David L Barack, Steve WC Chang, and Michael L Platt. Posterior cingulate neurons dy- namically signal decisions to disengage during foraging. Neuron, 96(2):339 -- 347, 2017. [31] Eric L Charnov. Optimal foraging, the marginal value theorem. Theoretical population biology, 9(2):129 -- 136, 1976. [32] Peter Nonacs. State dependent behavior and the marginal value theorem. Behavioral Ecology, 12(1):71 -- 83, 2001. [33] Tim W Fawcett, Benja Fallenstein, Andrew D Higginson, Alasdair I Houston, Dave EW Mallpress, Pete C Trimmer, John M McNamara, et al. The evolution of decision rules in complex environments. Trends in cognitive sciences, 18(3):153 -- 161, 2014. [34] Daniel G Goldstein and Gerd Gigerenzer. Models of ecological rationality: the recognition heuristic. Psychological review, 109(1):75, 2002. [35] Peter M Todd and Gerd Gigerenzer. Environments that make us smart: Ecological rational- ity. Current directions in psychological science, 16(3):167 -- 171, 2007. [36] Peter M Todd and Gerd Gigerenzer. Ecological rationality: Intelligence in the world. OUP USA, 2012. [37] Jeffrey K Waage. Foraging for patchily-distributed hosts by the parasitoid, nemeritis canescens. The Journal of Animal Ecology, pages 353 -- 371, 1979. [38] John McNamara. Optimal patch use in a stochastic environment. Theoretical Population Biology, 21(2):269288, Apr 1982. [39] Gerard Driessen and Carlos Bernstein. Patch departure mechanisms and optimal host ex- ploitation in an insect parasitoid. Journal of Animal Ecology, 68(3):445 -- 459, 1999. 21 [40] Jean-S´ebastien Pierre, Joan Van Baaren, and Guy Boivin. Patch leaving decision rules in parasitoids: do they use sequential decisional sampling? Behavioral Ecology and Sociobiology, 54(2):147 -- 155, 2003. [41] Patsy Haccou, Sake J De Vlas, Jacques JM Van Alphen, and Marcel E Visser. Information processing by foragers: effects of intra-patch experience on the leaving tendency of leptopilina heterotoma. The Journal of Animal Ecology, pages 93 -- 106, 1991. [42] Dale E Taneyhill. Patch departure behavior of bumble bees: rules and mechanisms. Psyche: A Journal of Entomology, 2010, 2010. [43] Philip L Smith and Roger Ratcliff. Psychology and neurobiology of simple decisions. Trends in Neurosciences, 27(3):161 -- 168, March 2004. [44] Roger Ratcliff, Philip L. Smith, Scott D. Brown, and Gail McKoon. Diffusion Decision Model: Current Issues and History. Trends in Cognitive Sciences, 20(4):260 -- 281, April 2016. [45] H. Rita and E. Ranta. Stochastic patch exploitation model. Proceedings of the Royal Society of London B: Biological Sciences, 265(1393):309315, Feb 1998. [46] Eric Wajnberg, Xavier Fauvergue, and Odile Pons. Patch leaving decision rules and the marginal value theorem: an experimental analysis and a simulation model. Behavioral Ecol- ogy, 11(6):577 -- 586, 2000. [47] Yoh Iwasa, Masahiko Higashi, and Norio Yamamura. Prey distribution as a factor deter- mining the choice of optimal foraging strategy. The American Naturalist, 117(5):710 -- 723, 1981. [48] Leslie Real and Thomas Caraco. Risk and foraging in stochastic environments. Annual Review of Ecology and Systematics, 17(1):371 -- 390, 1986. [49] Herbert A Simon. Theories of decision-making in economics and behavioral science. The American economic review, 49(3):253 -- 283, 1959. [50] Sandeep Mishra. Decision-making under risk: Integrating perspectives from biology, eco- nomics, and psychology. Personality and Social Psychology Review, 18(3):280307, Aug 2014. [51] Herbert A Simon. A behavioral model of rational choice. The quarterly journal of economics, 69(1):99 -- 118, 1955. [52] Herbert Simon. Reason in human affairs. Stanford University Press, 1990. [53] Herbert A Simon et al. An empirically-based microeconomics. Cambridge Books, 2009. [54] David Ward. The role of satisficing in foraging theory. Oikos, pages 312 -- 317, 1992. [55] Peter Nonacs and Lawrence M Dill. Is satisficing an alternative to optimal foraging theory? Oikos, pages 371 -- 375, 1993. [56] Yohay Carmel and Yakov Ben-Haim. Info-gap robust-satisficing model of foraging behavior: Do foragers optimize or satisfice? The American Naturalist, 166(5):633 -- 641, 2005. [57] Jane F Ward, Roger M Austin, and David W Macdonald. A simulation model of foraging behaviour and the effect of predation risk. Journal of Animal Ecology, 69(1):16 -- 30, 2000. 22 [58] Feng Zhang and Cang Hui. Recent experience-driven behaviour optimizes foraging. Animal behaviour, 88:13 -- 19, 2014. [59] Sigrunn Eliassen, Christian Jrgensen, Marc Mangel, and Jarl Giske. Quantifying the adaptive value of learning in foraging behavior. The American Naturalist, 174(4):478489, Oct 2009. [60] John McNamara and Alasdair Houston. A simple model of information use in the exploitation of patchily distributed food. Animal Behaviour, 33(2):553 -- 560, 1985. [61] John H Kagel, Leonard Green, and Thomas Caraco. When foragers discount the future: Constraint or adaptation? Animal Behaviour, 34:271 -- 283, 1986. [62] David W Stephens. Decision ecology: foraging and the ecology of animal decision making. Cognitive, Affective, & Behavioral Neuroscience, 8(4):475 -- 484, 2008. [63] Benjamin Y Hayden. Time discounting and time preference in animals: a critical review. Psychonomic bulletin & review, 23(1):39 -- 53, 2016. [64] Tommy C Blanchard and Benjamin Y Hayden. Monkeys are more patient in a foraging task than in a standard intertemporal choice task. PloS one, 10(2):e0117057, 2015. [65] Andrew M Wikenheiser, David W Stephens, and A David Redish. Subjective costs drive overly patient foraging strategies in rats on an intertemporal foraging task. Proceedings of the National Academy of Sciences, page 201220738, 2013. [66] Taiki Takahashi, Hidemi Oono, and Mark HB Radford. Psychophysics of time perception and intertemporal choice models. Physica A: Statistical Mechanics and its Applications, 387 (8-9):2066 -- 2074, 2008. [67] Evan C Carter and A David Redish. Rats value time differently on equivalent foraging and delay-discounting tasks. Journal of Experimental Psychology: General, 145(9):1093, 2016. [68] Tommy C Blanchard, John M Pearson, and Benjamin Y Hayden. Postreward delays and systematic biases in measures of animal temporal discounting. Proceedings of the National Academy of Sciences, page 201310446, 2013. [69] Sara M Constantino and Nathaniel D Daw. Learning the opportunity cost of time in a patch-foraging task. Cognitive, Affective, & Behavioral Neuroscience, 15(4):837 -- 853, 2015. [70] Melissa Bateson and Alex Kacelnik. Rate currencies and the foraging starling: the fallacy of the averages revisited. Behavioral Ecology, 7(3):341 -- 352, 1996. [71] C Randy Gallistel and John Gibbon. Time, rate, and conditioning. Psychological review, 107 (2):289, 2000. [72] Alex Kacelnik. Normative and descriptive models of decision making: time discounting and risk sensitivity. Characterizing human psychological adaptations, 208:51 -- 66, 1997. [73] Alex Kacelnik, Marco Vasconcelos, Tiago Monteiro, and Justine Aw. Darwins tug-of-war vs. starlings horse-racing: how adaptations for sequential encounters drive simultaneous choice. Behavioral Ecology and Sociobiology, 65(3):547558, Mar 2011. 23 [74] David W Stephens, Joel S Brown, and Ronald C Ydenberg. Foraging: behavior and ecology. University of Chicago Press, 2008. [75] Ariana Strandburg-Peshkin, Damien R. Farine, Iain D. Couzin, and Margaret C. Crofoot. Shared decision-making drives collective movement in wild baboons. Science, 348(6241): 13581361, Jun 2015. [76] Vaibhav Srivastava and Naomi Ehrich Leonard. Collective decision-making in ideal net- works: The speed-accuracy tradeoff. IEEE Transactions on Control of Network Systems, 1 (1):121132, Mar 2014. [77] Nils Kolling and Thomas Akam. (reinforcement?) learning to forage optimally. Current Opinion in Neurobiology, 46(Supplement C):162169, Oct 2017. [78] John McNamara and Alasdair Houston. The application of statistical decision theory to animal behaviour. Journal of Theoretical Biology, 85(4):673690, Aug 1980. [79] David C. Krakauer and Miguel A. Rodrguez-Girons. Searching and learning in a random environment. Journal of Theoretical Biology, 177(4):417429, Dec 1995. [80] John M. McNamara, Richard F. Green, and Ola Olsson. Bayes theorem and its applications in animal behaviour. Oikos, 112(2):243251, Feb 2006. [81] Ola Olsson and Joel S. Brown. The foraging benefits of information and the penalty of ignorance. Oikos, 112(2):260 -- 273, 2006. [82] Jay M Biernaskie, Steven C Walker, and Robert J Gegear. Bumblebees learn to forage like bayesians. The American Naturalist, 174(3):413 -- 423, 2009. [83] Elizabeth A Marschall, Peter L Chesson, and Roy A Stein. Foraging in a patchy environment: prey-encounter rate and residence time distributions. 1989. [84] Frederic Bartumeus, Daniel Campos, William S Ryu, Roger Lloret-Cabot, Vicen¸c M´endez, and Jordi Catalan. Foraging success under uncertainty: search tradeoffs and optimal space use. Ecology letters, 19(11):1299 -- 1313, 2016. [85] John M McNamara, Tim W Fawcett, and Alasdair I Houston. An adaptive response to uncertainty generates positive and negative contrast effects. Science, 340(6136):1084 -- 1086, 2013. [86] Alan C Kamil, Robin L Misthal, and David W Stephens. Failure of simple optimal foraging models to predict residence time when patch quality is uncertain. Behavioral Ecology, 4(4): 350 -- 363, 1993. [87] Andrew D. Higginson, Tim W. Fawcett, Pete C. Trimmer, John M. McNamara, and Alas- dair I. Houston. Generalized optimal risk allocation: Foraging and antipredator behavior in a fluctuating environment. The American Naturalist, 180(5):589603, Nov 2012. [88] Tommy Blanchard, Andreas Wilke, and Benjamin Hayden. Hot-hand bias in rhesus monkeys. Journal of Experimental Psychology: Animal Learning and Cognition, 40:280, Jul 2014. 24 [89] Alejandro Kacelnik and Ian A Todd. Psychological mechanisms and the marginal value theorem: effect of variability in travel time on patch exploitation. Animal Behaviour, 43(2): 313 -- 322, 1992. [90] Innes C. Cuthill, Alejandro Kacelnik, John R. Krebs, Patsy Haccou, and Yoh Iwasa. Starlings exploiting patches: the effect of recent experience on foraging decisions. Animal Behaviour, 40(4):625 -- 640, October 1990. [91] Andra Thiel and Thomas S. Hoffmeister. Knowing your habitat: linking patch-encounter rate and patch exploitation in parasitoids. Behavioral Ecology, 15(3):419 -- 425, May 2004. [92] Mark L. Wildhaber, Richard F. Green, and Larry B. Crowder. Bluegills continuously update patch giving-up times based on foraging experience. Animal Behaviour, 47(3):501 -- 513, March 1994. [93] Yannick Outreman, Anne Le Ralec, Eric Wajnberg, and Jean-Sbastien Pierre. Effects of within- and among-patch experiences on the patch-leaving decision rules in an insect para- sitoid. Behavioral Ecology and Sociobiology, 58(2):208 -- 217, June 2005. [94] Andra Thiel and Thomas Hoffmeister. Selective information use in parasitoid wasps. Animal Biology, 56(2):233 -- 245, April 2006. [95] Jason ND Kerr and Axel Nimmerjahn. Functional imaging in freely moving animals. Current opinion in neurobiology, 22(1):45 -- 53, 2012. [96] Fritjof Helmchen, Winfried Denk, and Jason ND Kerr. Miniaturization of two-photon mi- croscopy for imaging in freely moving animals. Cold Spring Harbor Protocols, 2013(10): pdb -- top078147, 2013. [97] Kartikeya Murari, Ralph Etienne-Cummings, Gert Cauwenberghs, and Nitish Thakor. An integrated imaging microscope for untethered cortical imaging in freely-moving animals. In Engineering in medicine and biology society (EMBC), 2010 annual international conference of the IEEE, pages 5795 -- 5798. IEEE, 2010. [98] Ming Yin, David A Borton, Jacob Komar, Naubahar Agha, Yao Lu, Hao Li, Jean Lau- rens, Yiran Lang, Qin Li, Christopher Bull, et al. Wireless neurosensor for full-spectrum electrophysiology recordings during free behavior. Neuron, 84(6):1170 -- 1182, 2014. [99] Tobi A Szuts, Vitaliy Fadeyev, Sergei Kachiguine, Alexander Sher, Matthew V Grivich, Mar- garida Agrochao, Pawel Hottowy, Wladyslaw Dabrowski, Evgueniy V Lubenov, Athanas- sios G Siapas, et al. A wireless multi-channel neural amplifier for freely moving animals. Nature neuroscience, 14(2):263, 2011. [100] Reid R Harrison, Haleh Fotowat, Raymond Chan, Ryan J Kier, Robert Olberg, Anthony Leonardo, and Fabrizio Gabbiani. Wireless neural/emg telemetry systems for small freely moving animals. IEEE transactions on biomedical circuits and systems, 5(2):103 -- 111, 2011. [101] Laszlo Grand, Sergiu Ftomov, and Igor Timofeev. Long-term synchronized electrophysiolog- ical and behavioral wireless monitoring of freely moving animals. Journal of neuroscience 25 methods, 212(2):237 -- 241, 2013. [102] Vikash Gilja, Cindy A Chestek, Paul Nuyujukian, Justin Foster, and Krishna V Shenoy. Autonomous head-mounted electrophysiology systems for freely behaving primates. Current opinion in neurobiology, 20(5):676 -- 686, 2010. [103] Alexander B Wiltschko, Matthew J Johnson, Giuliano Iurilli, Ralph E Peterson, Jesse M Katon, Stan L Pashkovski, Victoria E Abraira, Ryan P Adams, and Sandeep Robert Datta. Mapping sub-second structure in mouse behavior. Neuron, 88(6):1121 -- 1135, 2015. [104] Matthew Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference. In Advances in neural information processing systems, pages 2946 -- 2954, 2016. [105] Jeffrey E Markowitz, Winthrop F Gillis, Celia C Beron, Shay Q Neufeld, Keiramarie Robert- son, Neha D Bhagat, Ralph E Peterson, Emalee Peterson, Minsuk Hyun, Scott W Linderman, et al. The striatum organizes 3d behavior via moment-to-moment action selection. Cell, 2018. [106] Gordon J Berman, Daniel M Choi, William Bialek, and Joshua W Shaevitz. Mapping the stereotyped behaviour of freely moving fruit flies. Journal of The Royal Society Interface, 11 (99):20140672, 2014. [107] David J Anderson and Pietro Perona. Toward a science of computational ethology. Neuron, 84(1):18 -- 31, 2014. [108] Talmo D Pereira, Diego E Aldarondo, Lindsay Willmore, Mikhail Kislin, Samuel S-H Wang, Mala Murthy, and Joshua W Shaevitz. Fast animal pose estimation using deep neural net- works. bioRxiv, page 331181, 2018. [109] Ahmed El Hady. Closed loop neuroscience. Academic Press, 2016. [110] Steve M Potter, Ahmed El Hady, and Eberhard E Fetz. Closed-loop neuroscience and neu- roengineering. Frontiers in neural circuits, 8:115, 2014. [111] Logan Grosenick, James H Marshel, and Karl Deisseroth. Closed-loop and activity-guided optogenetic control. Neuron, 86(1):106 -- 139, 2015. 26 SUPPLEMENTARY MATERIAL FIG. S1. Full simulation results with added patch decision noise. Shown are the average and standard deviation of the energy intake (left grid) and patch residence times (right grid), for the density adaptive strategy (top) and the robust counting strategy (bottom), when the noise on the patch decision variable (σ) is increased. The robust counting strategy is implemented by setting α = −0.2ρ0 for each case. Each grid of 9 plots contains simulation results with different values of the travel time and the optimal available energy in the environment: columns correspond to values of Ttr = (1, 5, 10)τ , and rows correspond to values of Eopt = (0.5, 2, 5)s. For each plot, the filled blue curve uses a patch size of A = 1.5τ , the filled red curve uses a patch size of A = 5τ , and solid line is the optimal energy or patch time. 27 Density-adaptive strategy02468Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )01234Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )0.000.250.50Noise: /00.00.51.01.52.0Energy (units ofs)0.000.250.50Noise: /0Eopt<E>EA=1.5A=5Patch sizeTopt<T>TA=1.5A=5Patch size0.000.250.50Noise: /00.000.250.50Noise: /00.02.55.07.510.012.515.0PRT (units of )0.000.250.50Noise: /00.000.250.50Noise: /0Ttr=1Ttr=5Ttr=10Ttr=1Ttr=5Ttr=10Robust counting strategy02468Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )01234Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )0.000.250.50Noise: /00.00.51.01.52.0Energy (units ofs)0.000.250.50Noise: /00.000.250.50Noise: /00.000.250.50Noise: /00.02.55.07.510.012.515.0PRT (units of )0.000.250.50Noise: /00.000.250.50Noise: /0Ttr=1Ttr=5Ttr=10Ttr=1Ttr=5Ttr=10Eopt=0.5Eopt=2Eopt=5Eopt=0.5Eopt=2Eopt=5 FIG. S2. Full simulation results with discrete food rewards. The organization of the grid of plots and other parameters are the same as Fig S1, but show here are simulation results when the food chunks size (c) is increased. 28 02468Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )01234Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )0510Chunk size:c0.00.51.01.52.0Energy (units ofs)0510Chunk size:c0510Chunk size:c0510Chunk size:c0.02.55.07.510.012.515.0PRT (units of )0510Chunk size:c0510Chunk size:cDensity-adaptive strategyEopt<E>EA=1.5A=5Patch sizeTopt<T>TA=1.5A=5Patch sizeTtr=1Ttr=5Ttr=10Ttr=1Ttr=5Ttr=1002468Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )01234Energy (units ofs)0.02.55.07.510.012.515.0PRT (units of )0510Chunk size:c0.00.51.01.52.0Energy (units ofs)0510Chunk size:c0510Chunk size:c0510Chunk size:c0.02.55.07.510.012.515.0PRT (units of )0510Chunk size:c0510Chunk size:cRobust counting strategyTtr=1Ttr=5Ttr=10Ttr=1Ttr=5Ttr=10Eopt=0.5Eopt=2Eopt=5Eopt=0.5Eopt=2Eopt=5 FIG. S3. Uncertain patch size and adaptive strategies. Shown are simulations in an en- vironment where the patch size is uncertain. The size of individual patches (A) is drawn from a Gaussian distribution with mean ¯A = 5 and standard deviation ∆A = 0.3 ¯A. The average energy and patch residence times, and the distribution of individual patch residence times, are shown for three strategies: the density adaptive and robust counting strategies are implemented in the same manner as in Fig 4, and also an approximate size-adaptive strategy with α = 1.05αS. Other parameters are set corresponding to Fig 4: Ttr = 5, Eopt = 2 (or equivalently, ρ0 = 9.439), c = 0, and σ = 0.3ρ0. The bottom three plots show patch residence times for each strategy along with the optimal relationship from Eq. 8, and the approximate adjustment to PRTs calculated in Eq. S22 according to the value of α for each strategy. 29 MVT optimalStrategy 1: Density adaptiveStrategy 2: Robust countingStrategy 3: Size adaptive0510A0.00.10.2P(A)AUncertain patch size123012Energy123051015PRTAverage energyobtainedStrategyStrategyAverage patchresidence time0.02.55.07.510.0Patch size (A)051015PRT0.02.55.07.510.0Patch size (A)0.02.55.07.510.0Patch size (A)Strategy 1:Density adaptiveStrategy 2:Robust countingOptimalEq. S22SimulationOptimalEq. S22SimulationOptimalEq. S22SimulationStrategy 3:Size adaptive FIG. S4. Full simulation results with different strategies and forms of the utility func- tion. Analogous results to Fig 5C-D are shown here for both the density-adaptive strategy (left grid) and the robust counting strategy (right grid), each in the two environments from Fig 4: un- certain patch food density (top row), and scattered patches with discrete reward (bottom row). Simulation parameters correspond to the analogous cases in Fig 4, except for the available energy in the environment, which is varied here by changing the value of ¯ρ0 in the simulations. For each case of the utility function, the average energy intake and patch residence time are shown for two different values of β. Solid lines are an approximate solution to the governing equations and points are the mean and standard deviation of simulation results. (A) Results using the exponential utility function (see Fig 4A). (B) Results using the linear threshold utility function (see Fig 4B). 30 0123456Energy (units ofs)Avg. energy intake010203040PRT (units of )Patch residence times0123456Energy (units ofs)Avg. energy intake010203040PRT (units of )Patch residence times024Eopt0123456Energy (units ofs)024Eopt010203040PRT (units of )024Eopt0123456Energy (units ofs)024Eopt010203040PRT (units of )Exponential utility function0123456Energy (units ofs)Avg. energy intake010203040PRT (units of )Patch residence times0123456Energy (units ofs)Avg. energy intake010203040PRT (units of )Patch residence times024Eopt0123456Energy (units ofs)024Eopt010203040PRT (units of )024Eopt0123456Energy (units ofs)024Eopt010203040PRT (units of )Environment A: Uncertain patch food densityThreshold linear utility functionEnvironment B: Scattered patcheswith discrete rewardsEnvironment A: Uncertain patch food densityEnvironment B: Scattered patcheswith discrete rewards=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3Topt=0.1=0.3E=Eopt=0.1=0.3ToptABDensity-adaptive strategyRobust counting strategyDensity-adaptive strategyRobust counting strategyApprox. solution SimulationApprox. solution Simulation S1. Fokker-Planck formulation and numerical solution for probability density Consider the master equation for the patch decision variable, rewritten here for clarity: τ dx = (α − r(t)) dt + σdW (t). (S1) We will formulate this as a Fokker-Plank equation and solve for the probability density via the finite element method. To do this, we first define a normalized patch decision variable with and take the differential: y = τ x(t) η(t) , dy = τ ≈ τ dx(t) η(t) dx(t) η(t) − τ x(t) η(t)2 dη(t) , (S2) (S3) where the approximation is used, because the threshold η(t) changes slowly compared to the patch decision variable. Now we can write a new master equation with this change of variables: dy = (αy(t) − ry(t)) dt + σy(t)dW (t), (S4) where αy(t) ≡ α/η(t), ry(t) ≡ r(t)/η(t), and σy(t) ≡ σ/η(t), and the decision threshold occurs at y = 1. Note that since we consider strategies where α and η are either zero or have the same sign, αy(t) will always be either zero or positive, setting a drift towards the threshold. For food rewards, if η > 0 then ry > 0, and from Eq. S4 food reward will decrease y, i.e. lowering it away from the threshold of y = 1. If η < 0, then ry < 0, and food will increase y towards the threshold. Thus, the normalized formulation with the threshold at y = 1 can represent the different decisions strategies without any other further modifications. The Fokker-Plank equation corresponding to Eq. S4 is ∂G ∂t = − (αy(t) − ry(t)) ∂G ∂y + σy(t)2 2 ∂2G ∂y2 , (S5) where G(y, t) is the time-dependent probability density for the normalized decision variable y. We keep the terms αy(t) and ry(t) separate, because the former is a continuous function while the latter is defined by discrete inputs via a Poisson process when food rewards are received in chunks. To solve this using the finite element method, first let G = Ni(y)gi(t), where Ni(y) are the shape functions and gi(t) are the nodal variables. Summation notation applies over the indices i and j. After writing the weak form of the equation and setting the integral of the residual to zero, we obtain the finite element matrix equation: Mij dgj dt = − (αy(t) − ry(t)) Bijgj + σy(t)2 2 Aijgj, (S6) 31 where Mij is the mass matrix, Aij is a second-derivative matrix operator, and Bij is a first-derivative matrix operator. We consider the solution over a domain of [−L, 1], and choose the lower value of the domain as sufficiently low to encompass the full range of the probability distribution of y. The upper boundary of y = 1 is absorbing, and therefore has the condition G(1, t) = 0. We define the lower boundary as reflecting: ∂G(−L, t)/∂t = 0. The mass matrix is defined by integrating the shape functions: Mij = NiNjdy. (S7) To define Aij, which is the second derivative matrix operator, we will use integration by parts so that only a first derivative remains (and thus we will only need to use linear shape functions). Writing out the integral, and then integrating by parts, we have (cid:90) 1 −L (cid:90) 1 Aij = Ni −L = Ni ∂Nj ∂y (cid:90) 1 (cid:90) 1 ∂2Nj ∂y2 dy − (cid:12)(cid:12)(cid:12)(cid:12)1 (cid:12)(cid:12)(cid:12)(cid:12)1 − −L −L ∂Nj ∂y ∂Ni ∂y ∂Nj ∂y dy ∂Ni ∂y ∂Nj ∂y −L dy, = Ni (S8) where the last equality uses the zero-flux reflecting boundary condition at y = −L. For all elements the 2nd term in Eq. S8 yields 1/dy((−1, 1), (1,−1)), where dy is the size of each element. The absorbing boundary at y = 1 leads to a nonzero flux, and therefore must be included in the global matrix calculation. To do this, consider the last element in the mesh. Evaluating the boundary term yields an element matrix of 1/dy((0, 0), (1,−1)), which must also be included in the calculation of Aij to enforce the boundary condition. The first derivative matrix operator, Bij, is also defined by integrating by parts: (cid:90) 1 Bij = Ni −L ∂Nj ∂y = NiNj1−L − = NiNj−L − dy (cid:90) 1 (cid:90) 1 −L −L ∂Ni ∂y ∂Ni ∂y Njdy Njdy (S9) where the last equality applies the absorbing boundary condition of G(1, t) = 0. The reflecting boundary condition at y = −L adds an additional contribution of ((1, 0), (0, 0)) to the first element of the mesh, which must also be included in the calculation of Bij. To solve these equations numerically, the discrete food rewards are treated separately from the drift and diffusion of the probability density. Therefore, in the code, we solve the equation Mij dgj dt = −αy(t)Bijgj + 32 σy(t)2 2 Aijgj, (S10) and add an extra statement to shift the probability distribution when discrete food rewards ry(t) are received. We use the simulation to determine the flux through the upper boundary and the time- dependent probability P (t) that a decision to leave the patch has been made. Flux through the upper boundary can occur from either drift, diffusion, or the receipt of food reward. We calculate P (t) by integrating over the probability density: P (t) = 1 − G(y, t)dy. (S11) (cid:90) 1 −L (cid:90) tmax For the simulations shown in Fig 1C, we coupled patch decisions with the estimate of the energy in the environment by using the expectation value of the decision time: ¯T = T (cid:48)P (T (cid:48))dT (cid:48). (S12) where tmax is a sufficiently large time value. 0 33 S2. Simulations for full parameter range In the main text we focused on the case A = 5τ , Ttr = 5τ , and E = 2s. To investigate the full parameter dependence of the model, we consider scenarios that represent different configurations of the environment: 1. Low, medium, and high available energy rates. The animal needs to obtain energy E > 0 to survive. We therefore consider three regimes of the amount of energy surplus available from the environment: low (E = 0.5s), medium (E = 2s), and high (E = 5s). 2. Short, medium, and long inter-patch travel times. We consider this by using three values for travel times: short (Ttr = τ ), medium (Ttr = 5τ ), and long (Ttr = 10τ ) 3. Small vs large patches. A small patch will be depleted quickly, and a large patch will be depleted slowly. We consider small patches with A = 1.5τ , and larger patches with A = 5τ . In all simulations, we set the energy level by using Eqs. 7-8 to solve for the value of ρ0 that leads to a certain optimal energy level, given the values of the other parameters. 34 S3. Range for drift rate values Here we determine the values of the drift rate α that lead to valid model behavior, i.e. where there is only a single threshold crossing during the time 0 < t < Topt. Consider the value α = αS in Eq. 11, which yields a threshold of η = 0. For this case, the patch decision variable will start at x = 0, decrease, and then increase again to reach the threshold at zero. However, when α < αs, which yields η < 0, the patch decision variable will start at zero and will at first decrease, crossing the threshold at an early time t < Topt, then staying below the threshold before reaching it again at time Topt. Therefore, for some range of value αcrit < α < αS, there will be two threshold crossings, one at t < Topt and one at t = Topt, while outside of this range there is only a single threshold crossing at t = Topt. We solve for the critical value of the drift rate, αcrit, by considering the derivative of the patch decision variable at t = Topt. The critical value is when the derivative changes signs from positive to negative, i.e. (cid:2)αcrit − ρ0e−T /A(cid:3) = αcrit − E − s = 0, (S13) T =Topt which yields αcrit = E + s. For drift values in the range αcrit < α < αS, there will be two threshold crossings, and therefore a simulation would need an extra rule to "ignore" the first crossing in order to obtain optimal decisions. We therefore restrict drift values to be outside of this range. In our analysis, we make a further restriction to simply results by additionally neglecting the range 0 < α < αcrit, because in this range α and η have opposite signs. Note that when α is near the boundaries of this range, we can expect patch decisions to be very sensitive to the addition of noise on the patch decision variable, uncertainty in patch characteristics, and/or if rewards come in discrete chunks. 35 S4. Drift and threshold choices for optimal patch residence times with patch un- certainty The values of the drift rate, α, and the threshold, η, are defined using the average values of the patch characteristics in the environment: the average initial patch density ¯ρ0, and the average patch size ¯A. In this section we derive expressions for α and η to consider two possible cases: to optimally adjust patch residence times for uncertainty in patch density, or to optimally adjust patch residence times for uncertainty in patch size. Eq. 8 is the optimal form for patch residence time as a function of patch density and patch size; we repeat it here for clarity, using E = (cid:104)E(cid:105): Topt = A ln ρ0 E + s . We consider changes of patch residence time of the form T = Topt + δT. (S14) (S15) First consider a small change in patch density about an average value by using the expansion ρ0 = ¯ρ0 + δρ0. Plugging this into Eq. S14, expanding to first order terms, and comparing with Eq. S15 yields the optimal first order changes in patch residence time as function of changes in individual patch density: δTopt = ¯A ¯ρ0 δρ0. (S16) Similarly, considering a change in patch size of the form A = ¯A + δA yields an optimal first order change in patch residence time with changes in patch size: δTopt = ln ¯ρ0 E + s δA. (S17) We derive values for the drift rate and threshold so that either Eq. S16 or Eq. S17 are satis- fied; these represent two different strategies that an animal may use to adapt to uncertainty in an environment. In doing so, we demonstrate that both Eqs. S16 and S17 cannot be satisfied; the strategies represented by these cases represent a tradeoff between optimally adapting to uncertainty in patch density versus optimally adapting to uncertainty in patch size. We start with the integral of the patch decision variable equation (Eq. 2) with zero noise, using the average patch depletion function from Eq. 6. Integrating up to a time T when the threshold is reached yields (S18) Applying the condition that the threshold is reached at the optimal patch residence time in Eq. S14 yields a relationship between the threshold and the drift rate: η = ¯A α ln − ¯ρ0 + E + s , (S19) η = αT + ρ0A(cid:0)e−T /A − 1(cid:1) (cid:18) (cid:18) ¯ρ0 (cid:19) (cid:19) E + s 36 where we note that this is the same form as Eq. 9, except that here the average patch parameters ¯A and ¯ρ0 are used. We now combine Eq. S18 and S19, plug in expansions for T = Topt + δT and ρ0 = ¯ρ0 + δρ0, expand to first order in δT , and solve for the first-order changes in patch residence times: δT = ¯A (− ¯ρ0 + E + s) δρ0 ¯ρ0(−α + E + s) + δρ0(E + s) ≈ ¯A (− ¯ρ0 + E + s) ¯ρ0(−α + E + s) δρ0, (S20) where the approximation uses a series expansion in δρ0 to first order terms. Comparing this with Eq. S16 leads a value of α which satisfies optimal adaption to randomness in patch density, which is simply αD = ¯ρ0. (S21) We use an analogous process to calculate values of the drift rate and threshold for optimal adaption to uncertainty in patch size. Again we combine Eq. S18 and S19, then plug in expansions for T = Topt + δT and A = ¯A + δA, expand to first order in δT , and solve for the first-order changes in patch residence times: δT = ¯ρ0 (cid:0) ¯A + δA(cid:1) −(cid:0) ¯ρ0 ≈ (E + s)(cid:0)ln(cid:0) ¯ρ0 ¯A+δA(cid:0)(E + s) ¯A + ¯ρ0δA(cid:1) (cid:1) ¯ρ0 − α(cid:0) ¯ρ0 (cid:1) (cid:1) + 1(cid:1) − ¯ρ0 δA, ¯A+δA E+s E+s ¯A ¯A E+s −α + E + s (S22) (S23) (S24) where the approximation uses a series expansion in δA to first order terms. Comparing this with Eq. S17 and solving for α yields the drift rate that satisfies optimal adaption to randomness in patch size: E+s Using this in Eq. S19 yields the threshold value of ¯ρ0 − e − s ln(cid:0) ¯ρ0 (cid:1) . αS = ηS = 0. Thus, for optimal adaptation to patch size, the decision variable will start at zero, decrease to negative values as the animal finds food, and then increase back to zero for a decision to leave the patch. 37 S5. Optimal energy when patches vary in quality To investigate the model dependence on parameters and environmental characteristics, we perform simulations by choosing a value of ρ0 such that optimal energy return of the environment has a certain value. However, when patches vary in quality, we must consider the distribution to calculate the optimal energy. In the simulation we consider patches where Gaussian noise is added to the initial patch density (ρ0) and/or the patch size (A). These distributions are defined by mean parameters ¯ρ0 and ¯A and standard deviation parameters ∆ρ0 and ∆A. In this section we calculate a correction to the optimal energy that depends on the standard deviation of initial patch food density ∆ρ0. First, we write the average energy in the environment, following Eq. 7: (cid:68)(cid:82) Topt 0 (cid:69) − s ∗ (Ttr + (cid:104)Topt(cid:105)) r(t)dt (cid:104)E(cid:105) = (S25) where the average over the environment, denoted (cid:104)·(cid:105), must be evaluated over the distribution of patches. Using Gaussian probability distributions for these, we have Ttr + (cid:104)Topt(cid:105) , P (ρ0) = Cρ0e − (ρ0− ¯ρ0)2 2∆ρ2 0 P (A) = CAe − (A− ¯A)2 2∆A2 , (S26) (S27) where Cρ0 and CA are normalization factors. The average of some quantity z over these probability distributions is zP (ρ0)P (A)dρ0dA, (S28) where the lower end of the integral should be restricted to zero, because patches cannot have negative density or size. We first use this to evaluate the average return from patches by using the optimal time from Eq. 8: (cid:90) ∞ (cid:90) ∞ 0 0 (cid:104)z(cid:105) = (cid:28)(cid:90) Topt 0 (cid:29) r(t)dt = (cid:104)ρ0A − A ((cid:104)E(cid:105) + s)(cid:105) ≈ ¯ρ0 ¯A − ¯A ((cid:104)E(cid:105) + s) , (S29) where the approximation uses an evaluation of the Gaussian probability distribution over a full range, instead of restricting to positive values as expressed in Eq. S28. This approxi- mation holds well for ∆ρ0/ ¯ρ0 (cid:28) 1 and ∆A/ ¯A (cid:28) 1. To evaluate the average optimal patch residence time, we use the same approximation for the distribution of A, but evaluate the distribution of ρ0 over the restricted range due to the nonlinear form: (cid:104)Topt(cid:105) = (cid:28) ≈ ¯A (cid:29) A ln (cid:90) ∞ ρ0 (cid:104)E(cid:105) + s ρ0 ln (cid:104)E(cid:105) + s 0 38 P (ρ0)dρ0. (S30) The solution to this integral can be expressed in closed form using special functions; we performed this calculation using Mathematica. We use this to calculate a correction to the optimal energy, which was used to plot results in Fig 4, 5, and Fig S4. For example, using ¯ρ0 = 9.439 and ∆ρ0 = 0 leads to Eopt=2, while using ¯ρ0 = 9.439 and ∆ρ0 = 0.3 ¯ρ0 (which was used Fig 4), leads to Eopt = 2.077. To apply this to the cases shown in Figs 5 and S4, we found that the solution for the correction to the optimal energy, using Eqs. S25 and S30 can be approximated as a linear function function of the "uncorrected" energy, E0, using Eopt ≈ 0.0256307 + 1.02563E0. 39
1703.06264
3
1703
2017-03-25T13:24:29
Non-Associative Learning Representation in the Nervous System of the Nematode Caenorhabditis elegans
[ "q-bio.NC", "cs.NE", "q-bio.QM" ]
Caenorhabditis elegans (C. elegans) illustrated remarkable behavioral plasticities including complex non-associative and associative learning representations. Understanding the principles of such mechanisms presumably leads to constructive inspirations for the design of efficient learning algorithms. In the present study, we postulate a novel approach on modeling single neurons and synapses to study the mechanisms underlying learning in the C. elegans nervous system. In this regard, we construct a precise mathematical model of sensory neurons where we include multi-scale details from genes, ion channels and ion pumps, together with a dynamic model of synapses comprised of neurotransmitters and receptors kinetics. We recapitulate mechanosensory habituation mechanism, a non-associative learning process, in which elements of the neural network tune their parameters as a result of repeated input stimuli. Accordingly, we quantitatively demonstrate the roots of such plasticity in the neuronal and synaptic-level representations. Our findings can potentially give rise to the development of new bio-inspired learning algorithms.
q-bio.NC
q-bio
Non-Associative Learning Representation in the Nervous System of the Nematode Caenorhabditis elegans Ramin M. Hasani*, Magdalena Fuchs*, Victoria Beneder** and Radu Grosu* *Cyber-Physical-Systems Group, Vienna University of Technology, Austria **University of Natural Resources and Life Sciences, Vienna, Austria {ramin.hasani, radu.grosu}@tuwien.ac.at [email protected] [email protected] March 25, 2017 Abstract Caenorhabditis elegans (C. elegans) illustrated remarkable behavioral plasticities including complex non-associative and associative learning representations. Understanding the principles of such mecha- nisms presumably leads to constructive inspirations for the design of efficient learning algorithms. In the present study, we postulate a novel approach on modeling single neurons and synapses to study the mechanisms underlying learning in the C. elegans nervous system. In this regard, we construct a precise mathematical model of sensory neurons where we include multi-scale details from genes, ion channels and ion pumps, together with a dynamic model of synapses comprised of neurotransmitters and receptors kinetics. We recapitulate mechanosensory habituation mechanism, a non-associative learning process, in which elements of the neural network tune their parameters as a result of repeated input stimuli. Ac- cordingly, we quantitatively demonstrate the roots of such plasticity in the neuronal and synaptic-level representations. Our findings can potentially give rise to the development of new bio-inspired learning algorithms. 1 Learning the Learning Mechanisms Origin of many deep learning techniques can be traced back to nature [LeCun et al., 2015]. For instance, convolutional neural nets [LeCun et al., 1990], are inspired by the concept of simple and complex cells in the cat's visual cortex [Hubel and Wiesel, 1962]. In this study, we further investigate the sources of learning within the nervous system of small species such as C. elegans, by deriving detailed mathematical models of the elements of the nervous system. C. elegans is a 1mm-long, bodily transparent worm which dwells in soil and is considered the world's best understood organism [Ardiel and Rankin, 2010]. An adult's stereotypic nervous system comprises exactly 302 neurons connected through 8000 chemical and electrical synapses [White et al., 1986]. The worm, at first, seemed to be a hardwired model organism that can simply crawl forward and backward. Later, it proved otherwise by illustrating fantastic behavioral plasticity. Evidence of expressing well-founded non-associative and associative learning behaviors has been observed in C. elegans. Examples include habituation (short term and long term memory) to mechanical stimulation [Wicks and Rankin, 1997, Kindt et al., 2007, Cai et al., 2009], in the context of non-associative learning, and decision making over various favorable environmental stimuli such as food, temperature and oxygen in the form of associative learning [Saeki et al., 2001, Mohri et al., 2005, Tsui and van der Kooy, 2008]. 1 Model of Learning in C. elegans C. elegans is highly responsive to experience. This indicates that learning can be mediated from every sensory modality. For instance, mechanosensory habituation is a plasticity for the tap-withdrawal (TW) neural circuit (a neural network responsible for generating a reflexive response as a result of a mechanical tap on the petri dish in which the worm swims [Islam et al., 2016]), where the amplitude and the speed of the reflex can be modified as a result of periodic tap stimulation [Wicks and Rankin, 1997]. Detailed studies on such non-associative type of learning in the TW circuit, reveal fundamental notes underliying learning principles in the C. elegans, such as: • Sensory (input) neurons within the network are subjected to a mediation during the non-associative training process (repeated tap stimulation) [Kindt et al., 2007]. • A single gene's function during learning can interact with multiple postsynaptic connections. Such genes can potentially induce synaptic reconfiguration in several layers during memory consolidation [Stetak et al., 2009]. • Effects of certain genes are local within a neuron; They do not influence the synaptic pathways dis- tributed from that sensory neuron [Ohno et al., 2014]. • During the learning process, change in the amount of neurotransmitter release is notable and is an indicator of learning and memory [Jin et al., 2016]. • Within a neural circuit, only some of the interneurons are proposed to be the substrate of memory [Sugi et al., 2014]. Biological experiments on the principles of learning in C. elegans, from high-level behavioral features, to the dynamics of neural circuits, to the gene functional effect on the characteristic of ion channels in a neuron, have demonstrated the remarkable capability of C. elegans for learning and memory [Ardiel and Rankin, 2010]. In the present study, we aim to mathematically model the fundamental roots of learning mechanisms within the brain of the worm. Our findings in such attempt presumably enable machine-learning experts to get grounding inspirations towards development of novel learning algorithms. We utilize SIM-CE [Hasani et al., 2017], our simulation platform for performing physiological analysis on neurons and synapses, to explore the non-associative learning principles originated from autonomous gene modifications, variation of neurotransmitters concentrations and mediation of the ion channels activation and deactivation rates. Accordingly, we hypothesize the computational origins of such class of learning within two paradigms of neuronal habituation and synaptic plasticities. 2 Neuronal Habituation In neurons, electrical signals are being generated and propagate as a result of transportation of ions through ion channels [Salkoff et al., 2005]. Figure 1A graphically represents a simplified structure of a neuron of C. elegans in which we include voltage-gated calcium channel, calcium pump together with two different potassium channels and a leak channel to be the main mechanism responsible for the dynamics. Neurons can get excited by external sensory stimulus or chemical and electrical synapses from a presynaptic neuron. Considering the total membrane current, one can compose kinetics of the membrane potential and current as follows: Cm dV dt = −(ICa + IK + Isk + ILeak) + Σ(ISyn + Igap + Istimuli), (1) where Cm is the membrane capacitance and V represents the membrane potential. ICa, IK, Isk and ILeak stand for the calcium current, potassium current, intercellular calcium-gated potassium channel current and leakage current respectively [Kuramochi and Iwasaki, 2010]. Each ion channel's dynamics is precisely modeled with an inspiration from Hodgkin-Huxley channel modeling technique where the current-voltage relationship is developed based on Ohm's law, Iion = Gion(Eion − V ), while conductance of the ion channel, 2 Model of Learning in C. elegans Figure 1: Learning the learning representation. A) Structure of a neuron B) Habituation of a real sensory neuron (reprinted with permission from [Kindt et al., 2007]) C) Model of a neuronal habitation D) Response of an interneuron habituated and dishabituated with synaptic plasticities. Gion is considered to be the product of a constant maximum conductance with a dynamic parameter which is a function of the membrane potential [Hodgkin and Huxley, 1952]. In order to model the gene modification effects and the corresponding learning, we consider the maximum conductance of some ion channels to be a dynamic variable as a function of the input stimulus maximum value in time. In order to evaluate our assumption, we compare our result to the real sensory neuron habituation. A real sensory neuron habituates to repeated mechanical pulses on the body of the worm. The effect appears in the form of a reduction in the level of intracellular calcium concentration shown in Figure 1B. In our simulation such behavior is captured by modifying the maximum potassium channel conductance to be a dynamic function of the input tap-stimulus and time (Figure 1C). Key finding 1: The state of the activation of an input neuron can be modified by repeatedly induced stimuli through an intrinsic parameter which is dynamically altered based on the input amplitude and frequency. 2.1 Synaptic Plasticity Fundamentally, many learning mechanisms originate from the alternation of the amount of neurotransmitters concentration. Information propagates from a postsynaptic neuron to a presynaptic cell through chemical and electrical synapses. The information transportation process depends on several circumstances; the amount of available neurotransmitters, Gmax, at the presynaptic end, the state of the activation of the presynaptic cell which causes the neurotransmitter secretion tone G(Vpre), number of available receptors at the postsynaptic neuron and the probability of the neurotransmitters binding to the postsynaptic receptors, S(t). A detailed model of a chemical synapse's current therefore can be derived as follows: ISyn = GmaxG(Vpre)S(t)(ESyn − Vpost), (2) where ESyn represents the reversal potential of a synapse [Schutter, 2009]. Variables in Eq. 2 are dynamically modeled by partial differential equations where the steady state and the time-constant of each variable can be tuned in order to capture synaptic plasticity in C. elegans. Rankin and Wicks [2000] showed that the glutamate concentration considereably modifies the habituation responses in touch sensory neurons. Such 3 A Stimulus ICa ISK IK Ileak Postsynaptic neuron Ca2+ Pump C Gap junction Chemical synapses Presynaptic neuron B D Model of Learning in C. elegans behavior is captured as a property of a synapse where simultaneous alternation of the parameters inside Gmax and S(t), results in the similar tap-withdrawal response (See Figure 1D). Key finding 2: Synapses with an internal state can autonomously modify the behavior of the system. Synapses can presumably be involved in completion of a habituation process, dishabituation process (See Figure 1D the second calcium spike train) and propagation of a neuronal habituation to the rest of the neural circuit. 3 Final Note Learning is an essential attribute for survival. C. elegans demonstrates optimal adaptive behavior to react to the environmental circumstances for survival. Learning how the animal learns and adapts not only is a notable step forward towards decoding the human brain's principles but can also lead us to the development of better learning algorithms.1 References E. L. Ardiel and C. H. Rankin. An elegant mind: learning and memory in caenorhabditis elegans. Learning & Memory, 17(4):191–201, 2010. S.-Q. Cai, Y. Wang, K. H. Park, X. Tong, Z. Pan, and F. Sesti. Auto-phosphorylation of a voltage-gated k+ channel controls non-associative learning. The EMBO journal, 28(11):1601–1611, 2009. R. M. Hasani, V. Beneder, M. Fuchs, D. Lung, and R. Grosu. Sim-ce: An advanced simulink platform for studying the brain of caenorhabditis elegans. arXiv preprint arXiv:1703.06270, 2017. A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of physiology, 117(4):500, 1952. D. H. Hubel and T. N. Wiesel. Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. The Journal of physiology, 160(1):106–154, 1962. M. A. Islam, Q. Wang, R. M. Hasani, O. Bal´un, E. M. Clarke, R. Grosu, and S. A. Smolka. Probabilis- In High Level Design tic reachability analysis of the tap withdrawal circuit in caenorhabditis elegans. Validation and Test Workshop (HLDVT), 2016 IEEE International, pages 170–177. IEEE, 2016. X. Jin, N. Pokala, and C. I. Bargmann. Distinct circuits for the formation and retrieval of an imprinted olfactory memory. Cell, 164(4):632–643, 2016. K. S. Kindt, K. B. Quast, A. C. Giles, S. De, D. Hendrey, I. Nicastro, C. H. Rankin, and W. R. Schafer. Dopamine mediates context-dependent modulation of sensory plasticity in c. elegans. Neuron, 55(4): 662–676, 2007. M. Kuramochi and Y. Iwasaki. Quantitative modeling of neuronal dynamics in c. elegans. In Neural Information Processing. Theory and Algorithms, pages 17–24. Springer, 2010. B. B. LeCun, J. Denker, D. Henderson, R. Howard, W. Hubbard, and L. Jackel. Handwritten digit recognition with a back-propagation network. In Advances in Neural Information Processing Systems. Citeseer, 1990. Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436–444, 2015. A. Mohri, E. Kodama, K. D. Kimura, M. Koike, T. Mizuno, and I. Mori. Genetic control of temperature preference in the nematode caenorhabditis elegans. Genetics, 169(3):1437–1450, 2005. 1A comprehensive version of this work, will be published soon elsewhere. 4 Model of Learning in C. elegans H. Ohno, S. Kato, Y. Naito, H. Kunitomo, M. Tomioka, and Y. Iino. Role of synaptic phosphatidylinositol 3-kinase in a behavioral learning response in c. elegans. Science, 345(6194):313–317, 2014. C. H. Rankin and S. R. Wicks. Mutations of the caenorhabditis elegansbrain-specific inorganic phosphate transporter eat-4affect habituation of the tap–withdrawal response without affecting the response itself. Journal of Neuroscience, 20(11):4337–4344, 2000. S. Saeki, M. Yamamoto, and Y. Iino. Plasticity of chemotaxis revealed by paired presentation of a chemoat- tractant and starvation in the nematode caenorhabditis elegans. Journal of Experimental Biology, 204 (10):1757–1764, 2001. L. Salkoff, A. Wei, B. Baban, A. Butler, G. Fawcett, G. Ferreira, and C. M. Santi. Potassium channels in c. elegans. 2005. E. D. Schutter. Computational modeling methods for neuroscientists. The MIT Press, 2009. A. Stetak, F. Horndli, A. V. Maricq, S. van den Heuvel, and A. Hajnal. Neuron-specific regulation of associative learning and memory by magi-1 in c. elegans. PloS one, 4(6):e6019, 2009. T. Sugi, Y. Ohtani, Y. Kumiya, R. Igarashi, and M. Shirakawa. High-throughput optical quantification of mechanosensory habituation reveals neurons encoding memory in caenorhabditis elegans. Proceedings of the National Academy of Sciences, 111(48):17236–17241, 2014. D. Tsui and D. van der Kooy. Serotonin mediates a learned increase in attraction to high concentrations of benzaldehyde in aged c. elegans. Learning & Memory, 15(11):844–855, 2008. J. White, E. Southgate, J. Thomson, and S. Brenner. The structure of the nervous system of the nematode caenorhabditis elegans: the mind of a worm. Phil. Trans. R. Soc. Lond, 314:1–340, 1986. S. R. Wicks and C. H. Rankin. Effects of tap withdrawal response habituation on other withdrawal behaviors: the localization of habituation in the nematode caenorhabditis elegans. Behavioral neuroscience, 111(2): 342, 1997. 5
1603.05659
7
1603
2017-08-27T21:08:09
Unveiling causal activity of complex networks
[ "q-bio.NC", "nlin.AO", "physics.bio-ph", "physics.data-an" ]
We introduce a novel tool for analyzing complex network dynamics, allowing for cascades of causally-related events, which we call causal webs (c-webs), to be separated from other non-causally-related events. This tool shows that traditionally-conceived avalanches may contain mixtures of spatially-distinct but temporally-overlapping cascades of events, and dynamical disorder or noise. In contrast, c-webs separate these components, unveiling previously hidden features of the network and dynamics. We apply our method to mouse cortical data with resulting statistics which demonstrate for the first time that neuronal avalanches are not merely composed of causally-related events.
q-bio.NC
q-bio
epl draft Unveiling causal activity of complex networks(a) Rashid V. Williams-Garc´ıa(b), John M. Beggs, and Gerardo Ortiz Department of Physics, Indiana University, Bloomington, Indiana 47405, USA PACS 87.19.lc – Noise in the nervous system PACS 64.60.av – Cracks, sandpiles, avalanches, and earthquakes PACS 87.19.lj – Neuronal network dynamics Abstract –We introduce a novel tool for analyzing complex network dynamics, allowing for cas- cades of causally-related events, which we call causal webs (c-webs), to be separated from other non-causally-related events. This tool shows that traditionally-conceived avalanches may contain mixtures of spatially-distinct but temporally-overlapping cascades of events, and dynamical disor- der or noise. In contrast, c-webs separate these components, unveiling previously hidden features of the network and dynamics. We apply our method to mouse cortical data with resulting statis- tics which demonstrate for the first time that neuronal avalanches are not merely composed of causally-related events. 7 1 0 2 g u A 7 2 ] . C N o i b - q [ 7 v 9 5 6 5 0 . 3 0 6 1 : v i X r a In systems consisting of many interacting elements, a variety of methods (e.g., transfer entropy or Granger causality) are often used to reveal hidden dynamical causal links between them. This naturally leads to a complex net- works description [2], raising interesting questions. For ex- ample, what fraction of the activity in such a network can be attributed to the hidden causal dynamics, and what fraction is produced by other processes, such as noise? Here we describe a new approach to this problem and demonstrate its utility on neural networks. Over the past twenty years, there have been a number of theoretical [3–15] and experimental [16–28] attempts to connect activity in living neural networks to critical avalanches like those seen in the Bak-Tang-Wiesenfeld (BTW) sandpile model [29, 30]. It has been hypothe- sized that homeostatic mechanisms might tune the brain, a complex neural network, towards optimality associated with a critical point [31] which separates ordered ("su- percritical") and disordered ("subcritical") phases, where cascades of activity are amplified or damped, respectively [15,24,27,32,33]. In the BTW model, grains of "sand" are dropped one at a time at random lattice locations; sites which reach a threshold height topple their grains to their neighboring sites, potentially inducing further topplings, together forming an emergent cascade of events called an (a)The original version of this article was uploaded to the arXiv on March 17th, 2016 [1] (b)E-mail: [email protected]. Current affiliation: Departments of Neurobiology and Mathematics, University of Pittsburgh, Pitts- burgh, Pennsylvania 15260, USA avalanche. Successive topplings are thus causally related, with each new toppling having been induced by another which happened before. The sandpile model eventually reaches a steady state in which the probability distribution of avalanche sizes follows a power law, a potential indica- tor of criticality. It is important to note that the grains are dropped at an infinitesimally slow rate such that the relaxation timescale, i.e., the duration of the avalanches, is much shorter than the time between grain drops. This separation of timescales is essential to this concept of self- organized criticality (SOC) [3]. In real systems, however, this separation is not always achieved. As we will see, closer inspection of experi- mental neuronal avalanche data reveals potential conflicts with the SOC approach. For example, temporally distinct avalanches could be concatenated by sporadic events oc- curring between them, or two spatially distinct avalanches could be concatenated if they occurred synchronously. These confounding situations highlight the need for a method which clearly separates causally-related from in- dependent activity. In our recent work, we developed a framework in which there is a mixing of timescales, as opposed to a separation of timescales [34]. Using this framework, it was demonstrated that access to a critical point depends on the coupling of the concerned network to an external environment, resulting in a non-zero sponta- neous activation probability ps; the higher the probability, the further from criticality, and thus the further from op- timality. p-1 R. V. Williams-Garc´ıa et al. These spontaneous activation events could be caused by some unobserved influence, such as long-range innervation (in which case the network has been undersampled), acti- vations which occurred prior to the start of an experiment, or the intrinsic properties of the network elements, such as those neurons which have a propensity to fire spon- taneously or are tonically active, e.g., as in the case of pacemaker cells or some inhibitory neurons [35, 36]. In neuronal cultures, spatiotemporal localization of cellular noise can lead to spontaneous avalanche production [37]. Vanishing spontaneous activity (in combination with brief synaptic delays) is effectively equivalent to a separation of timescales, as described in SOC. Thus, achieving optimal- ity by operating at a critical point may not be feasible for a living, open neural network, according to the quasi- criticality hypothesis introduced in [34], although a rela- tive optimality may still be achieved along a nonequilib- rium Widom line [34]. The accessibility of this relatively- optimal quasicritical region will then likely depend on the character of the environment and the fundamental prop- erties of the neural network itself. Our ability to apply the nonequilibrium Widom line framework and test the qua- sicriticality hypothesis, however, hinges on the ability to identify spontaneous activity in a living neural network. In this paper, we introduce a new method to disentan- gle spontaneous activity-which we define as activations occurring without an established causal link to a prior activation-from that which is causally-related and pri- marily governed by the network structure and dynamics. To this end, we next introduce the notion of causal webs (or c-webs for short), as a new emergent cascade of cor- related events, whose properties contrast and complement those of standard avalanches. Whereas the latter are de- fined as spatiotemporal patterns of activation spanning a number of adjacent time steps framed by time steps lacking activity, c-webs explicitly depend on the effective network and the temporal delays associated with the con- nections therein, thus accommodating the potential non- Markovian dynamics of complex networks. In the effective network, connections from source elements to target ele- ments (which are analogous to presynaptic and postsynap- tic neurons in physical neural networks) are established based on the predictive information that the activation of the source element provides of the activation of the target element; spontaneous events are those activations of ele- ment which occur in the absence of an established causal link to a prior activation of a source element. Knowledge of the network structure and delay information is key, as it allows to distinguish between different spatiotemporal patterns of activation in a way which is not possible with avalanches (see fig. 1). Let us formalize the concept of c-webs in the context of neural networks. We label individual events by x = (i, t), representing the activation of neuron i at time t, or fol- lowing the notation used in [34], equivalently zi(t) = 1 (zi(t) = 0 meant quiescence and zi(t) > 1 corresponded to refractory states, which we do not consider to be acti- Fig. 1: (Color online) Causal webs (c-webs) are distinct from avalanches in that they rely on network structure and connec- tion delays. A. This network produces a variety of spatiotem- poral activity patterns (numbers correspond to neuron ID i). B. A time raster showing the activity patterns in panel A; indi- vidual activations are labeled xµ. Whereas only two neuronal avalanches are detected, a richer structure is revealed when spontaneous events (blue annuli) are separated from c-webs (orange disks). Acceptance windows Wij(t) are shaded light- orange, where i and j correspond to different neuron indices, for example, W12(2) = [3, 5]. Notice that because of their con- trasting definitions, c-webs and avalanches will generally have different statistical properties, e.g., c-webs may occur over a longer span of time due to connection delays, as shown here. vation events). We write the set of all events A = {xµ}, e.g., in fig. 1B, A = {x1, x2, x3, x4, x5, x6, x7}. Formally, we define a c-web C as a set C = {pc}c=1,C (C being the cardinality of the set C) of correlated ordered pairs pc = (xµ, xν)c of events (i.e., spikes), which we call causal pairs; quiescent and refractory neurons are not included in the set. The first and second entries, xµ and xν, of the cth causal pair represent causally-related source and target neuron events, respectively; each activation event can be associated to at most one c-web. (Despite causal relations being made in a pairwise fashion, we emphasize that this does not preclude multivariate interactions, as multiple pairings can be made to a single event.) In the following, we show how to determine those causal pairs. A complete set of causal pairs X is constructed by tak- ing the Cartesian product of each event xµ with its cor- responding dynamic target neuron events U(xµ), i.e., xµ∈A xµ × U(xµ), where U(x) ≡ U(i, t) is the set X = (cid:83) given by U(i, t) = {(j, t(cid:48)) j ∈ N (i) and t(cid:48) ∈ Wij(t)}. (1) N (i) refers to the set of all target neurons j of neuron i, and Wij(t) = [t+dij−∆ij, t+dij +∆ij] is a predetermined dynamical acceptance window: if a target neuron j is active within the acceptance window, then a causal link p-2 ABNeuron ID, (cid:1861)123412345678910Network structurePossible activity propagation patterns,all consistent with the network structure1234(cid:2206)(cid:2784)(cid:2206)(cid:2781)Avalanche 1Avalanche 2(cid:2206)(cid:2780)(cid:2206)(cid:2783)(cid:2206)(cid:2782)Spontaneous EventsCausal Webs(cid:2206)(cid:2779)(cid:2206)(cid:2778)Time steps, (cid:1872) between the activation events is inferred (see fig. 1B). The lower bound of the acceptance window is adjusted such that it is greater than t. We write the set of distinct events in X as A(X) ⊆ A. Connection delays dij associated with the connection from a source neuron i to a target neuron j, are allowed to have some uncertainty ∆ij due to variability in the tar- get neuron spike timing. We will later present a method by which this information can be determined from data; for the moment, we assume it is given. In fig. 1B, connection delays and their uncertainties are given for the connec- tions in fig. 1A: d12 = 2, d14 = 4, d31 = 2, and d42 = 1, with ∆12 = 1, ∆14 = 0, ∆31 = 1, and ∆42 = 1. This in- formation can be used to determine causal pairs, e.g., the event x1 = (1, 2) in fig. 1B has U(x1) = {x3, x4}, resulting in the causal pairs x1 × U(x1) = {(x1, x3), (x1, x4)}. The complete set of causal pairs for the spacetime graph 1B is X = {(x1, x3), (x1, x4), (x5, x6)} and so in fig. A(X) = {x1, x3, x4, x5, x6}. A causal web represents the connected components of a directed graph whose vertices and edges are A(X) and X, respectively. The example in fig. 1B thus has two c-webs, C1 = {(x1, x3), (x1, x4)} and C2 = {(x5, x6)}. Note that spontaneous events initiate c-webs and may become part of ongoing c-webs; we call spontaneous events associated with a c-web C, its roots r(C), e.g., r(C1) = {x1} and r(C2) = {x5}. The size s(C) of a c-web is defined as the total number of distinct events within it. Defining that set as A(C), the size s(C) is then given by its cardinality: s(C) = A(C). Note that A(C) ⊆ A(X). For example, A(C1) = {x1, x3, x4} and A(C2) = {x5, x6} in fig. 1B, with s(C1) = 3 and s(C2) = 2, respectively. The duration D(C) of a c-web C can be defined in terms of its chord. The chord of a c-web K(C) is the se- quence of distinct time steps for which there are events be- longing to that c-web, arranged in ascending order in time, with no repeated elements. That is, K(C) = (t1, t2, ..., tn), where t1 and tn are the times of the first and last events, respectively. In contrast to the definition of duration for avalanches, the length of a c-web's chord is not equal to the c-web duration. Instead, we define the duration of a c-web as a measure of its chord plus one, i.e., D(C) = 1 + λ(K(C)), where λ(K(C)) = tn − t1. The chords of the c-webs in fig. 1B, for example, are K(C1) = (2, 4, 6) and K(C2) = (7, 8), with durations D(C1) = 5 and D(C2) = 2. Finally, we define the branching fraction b(C) of a c-web C as the average number of target neuron events associated with each source neuron event: (cid:88) C(cid:88) δxα,xµ , (2) b(C) = 1 s(C) xα∈A(C) c=1 where δ is the Kronecker delta. The first sum is evaluated over all elements xα of A(C), while the second one is over all its causal pairs pc = (xµ, xν)c, where xµ represents the source neuron event of the pair pc ∈ C. For example, in fig. 1B, b(C1) = 2/3 and b(C2) = 1/2. Unveiling causal activity of complex networks We performed tests of our method using simulations of the cortical branching model (CBM) [34]. In the CBM, spontaneously activated nodes i initiate cascades of ac- tivity which spread to neighboring nodes j with activ- ity transmission probabilities (i.e., connection weights) Pij depending on their states zj ∈ {0, 1, 2, ..., τr}, where zj = 0 corresponds to quiescence, zj = 1 to activation, and zj ∈ {2, ..., τr} correspond to refractory states. Fol- lowing activation at time step t, zj(t) = 1, a node becomes refractory in the very next time step, zj(t + 1) = zj(t) + 1, and iterates until zj(t) = τr, after which the node returns to quiescence, zj(t + τr) = 0. Only nodes which are qui- escent at one time step can be activated in the following time step. There are no inhibitory nodes in the CBM. Fig. 2: (Color online) Simulated avalanche and causal web size distributions coincide in the case of a separation of timescales (blue), but are substantially different under conditions corre- sponding to a mixing of timescales (orange/black). Simulations were performed in a data-inspired network of N = 243 nodes. As we have stated earlier, neuronal avalanches and c- webs should coincide as emergent cascades of correlated events in the limit ps → 0 for all nodes and dij = 1 for all pairs of nodes (i, j). We simulated 106 avalanches on a net- work of N = 243 nodes, whose structure and connection weights were inspired by experimental data; all connec- tion delays were set to a single time step, i.e., dij = 1. To simulate the ps → 0 limit (a separation of timescales), we initiated avalanches at single, random nodes, only start- ing a new avalanche when the previous one had finished; no spontaneous events or concurrent avalanches were al- lowed. The resulting avalanche and c-web size probability distributions were identical, as expected (see fig. 2). In order to examine these distributions under conditions cor- responding to a mixing of timescales, we allowed for nodes to become active spontaneously. Spontaneous activation probabilities for each node were drawn from a Gaussian distribution with mean and standard deviation of 5×10−3; negative values were set to zero. As a result, the two dis- p-3 100101102103104Cascadesize,s10-810-610-410-2100Cascadesizedistribution,P(s)Avalanchesandcausalwebsatps=0Avalanchesatps=5#10!3Causalwebsatps=5#10!3 R. V. Williams-Garc´ıa et al. tributions differ greatly; most notably, c-webs better cap- ture the abundance of isolated spontaneous events (s = 1 c-webs) than avalanches do. In another test, we constructed a random network of N = 360 nodes, each with an in-degree of kin = 3, as in [34]. The network was not strongly connected (i.e., it was reducible, thus contained subgraphs) and had a spec- tral radius (i.e., Perron-Frobenius eigenvalue) of κ = 0.23. Connection delays (in time steps) were drawn from a uni- form distribution of integers in a closed interval, dij ∈ [1, 16]. Spontaneous activation probabilities for each node were drawn from a Gaussian distribution with mean and standard deviation of 10−4; negative values were again set to zero. The simulation was performed over 3.6 × 106 time steps and refractory periods of all nodes were set to a single time step (or, in the language of [34], τr = 1). Spontaneous events detected by our method were used to construct a new spontaneous activation probability distri- bution, which we compared with the initial distribution using a Kolmogorov-Smirnov (KS) test at a 5% signifi- cance level: the distributions were in agreement with a p-value of 0.996 [38]. We note that as the overall connec- tivity of the network (which we quantify by κ, as in [34]) is increased, spontaneous events become less prominent as c-webs begin to dominate the dynamics, leading to more driven activations (and, when τr > 1, refractory nodes), thus preventing spontaneous events: neural network dy- namics present a fluctuating bottleneck to the influence of an environment. To determine the c-webs, we have so far assumed knowl- edge of the network structure and delay information (i.e., the dij and ∆ij's), however in practice, this information must be learned from experimental data. We now describe a method, based on delayed transfer entropy (TE) [39,40], by which this information can be established from high temporal resolution multiunit timeseries data. The use of TE is not absolutely necessary; alternative measures of ef- fective connectivity and causal effect, such as information flow, which also compute conditional probabilities directly from data, could in principle be used [41, 42]. The notion of c-webs is completely independent of any one measure of effective connectivity. For a particular pair of neurons (i, j), TE is cal- culated at various delays d, peaking at the appropri- ate d = dij, with a width of ∆ij [39]. At a given is given by d, the TE from neuron i to neuron j, (cid:18) p(zj(t)zj(t − 1), zi(t − d)) p(zj(t)zj(t − 1)) (cid:19) , (3) Ti→j(d) = p(zj(t − 1), zj(t), zi(t − d)) log2 (cid:88) zi→j (d) the binary variables zj(t − 1), zj(t), and zi(t − d). Joint and conditional probabilities in eq. (3) are estimated from spike-sorted data, as in [39]. For every pair of neu- rons (i, j), Ti→j(d) is computed over a range of values d ∈ [1, 16]; the peak value Ti→j(dij) represents the puta- tive connection from neuron i to neuron j. The spike data is then randomly shuffled to establish a rejection thresh- old; TE values below this threshold are not considered to be significant and are set to zero [39, 43]. The remain- ing TE values are then converted to activity transmission probabilities, as described in the supplemental materials of [44], which assumes that spiking activity is Poisson- distributed. Note that the c-webs method could poten- tially be further improved by stochastically establishing causal pairs depending on those probabilities. Spurious connections, such as those due to common drive and tran- sitive connections, are removed by considering delays of the significant connections; these results "were valid over a wide range of values of the rejection threshold" [45]. We use only the delays dij of the significant connections to establish causal pairs. We next demonstrate the utility of our method when applied to experimental data (see fig. 3). For our demon- stration, we have used ten data sets from [46], which were collected in vitro from organotypic cultures of mouse so- matosensory cortex using a 512-microelectrode array with a 60 µm electrode spacing and a 20 kHz sampling rate Fig. 3: (Color online) One minute of neural network activity recorded from somatosensory cortex, after processing to sepa- rate spontaneous events (dark blue) from c-webs (orange) using the c-webs method. Note that tonically-active neurons mainly produce spontaneous events. where zi→j(d) = {zj(t − 1), zj(t), zi(t − d)} indicates that the sum is performed over all possible configurations of p-4 50100150200(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)0102030405060(cid:9)(cid:10)(cid:11)(cid:2)(cid:12)(cid:1)(cid:13)(cid:14)(cid:15)024(cid:2)(cid:13)(cid:1)(cid:15)101102103(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:3)(cid:7)(cid:8)(cid:6)(cid:9)(cid:1)10-610-510-410-310-210-1100(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:3)(cid:7)(cid:8)(cid:6)(cid:10)(cid:11)(cid:12)(cid:13)(cid:2)(cid:13)(cid:7)(cid:14)(cid:7)(cid:15)(cid:16)(cid:5)(cid:7)(cid:3)(cid:15)(cid:11)(cid:7)(cid:13)(cid:17)(cid:15)(cid:7)(cid:12)(cid:18)(cid:9)(cid:2)(cid:19)(cid:1)(cid:20)(cid:21)(cid:22)(cid:23)(cid:9)(cid:24)(cid:25)(cid:21)(cid:2)(cid:26)(cid:2)(cid:14)(cid:2)(cid:18)(cid:4)(cid:27)(cid:6)(cid:3)(cid:28)(cid:24)(cid:25)(cid:9)(cid:29)(cid:30)(cid:29)(cid:4)(cid:2)(cid:17)(cid:3)(cid:2)(cid:14)(cid:31)(cid:6)(cid:13)(cid:3)101102(cid:1)100101102(cid:1)(cid:3)(cid:2)(cid:19)(cid:1)(cid:20)(cid:32)(cid:2)(cid:15)(cid:2)(cid:33)(cid:34)(cid:1)(cid:2)(cid:19)(cid:1)(cid:20)(cid:3)(cid:1)(cid:1)(cid:23)(cid:4)(cid:24)(cid:35)(cid:2)(cid:19)(cid:1)(cid:20)(cid:3)(cid:1)(cid:1)(cid:24)(cid:4)(cid:30)(cid:25)A B Unveiling causal activity of complex networks Fig. 4: (Color online) Cascade size (A) and duration (B) probability distributions from simulations, which used the effective connectivity, delays, and times of spontaneous events extracted from in vitro data using TE and c-webs analyses. Dashed lines represent guides to the eye over regions in which maximum likelihood estimates were performed. Fig. 5: (Color online) Cascade size (A) and duration (B) probability distributions from in vitro data. Avalanches (black) and c-webs (orange) exhibit different statistical properties due to a mixing of timescales (panel A inset). Compare with the simulated predictions in fig. 4; similarly, dashed lines represent guides to the eye. over hour-long recordings [47, 48]. Applying our method to a data set containing N = 243 neurons, we extract spontaneous events (highlighted in dark blue) and c-webs (in orange) to illustrate their qualitative differences; note that spontaneous events may initiate or contribute to c- webs as in fig. 1B. In fig. 3, an activity time raster (top panel) is presented along with the corresponding time- series of the activity (bottom panel), on which we have performed a moving average with a ∆t = 100 ms window: i=1 δzi(t),1 [49]. t(cid:48)=0 x(t−t(cid:48))/∆t, where x(t) =(cid:80)N y(t) =(cid:80)∆t−1 We then performed simulations of the CBM using in- formation extracted from the experimental data with TE and from applying the c-webs method, i.e., the activity transmission probabilities (i.e., connection weights), con- nection delays, and spontaneous events. Using a data set which contained N = 435 neurons, the connection weights were adjusted by a factor κ (to manipulate the Perron- Frobenius eigenvalue of the corresponding adjacency ma- trix) over the range [0.20, 1.40] in steps of ∆κ = 0.05 and performed simulations at each value of κ-the raw data had a Perron-Frobenius eigenvalue of κ ≈ 0.31. Instead of stochastically initiating activity cascades using a fixed value of ps (as in [34]), we used the spontaneous events identified from the data to initiate activity; as a result, p-5 100101102103(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:3)(cid:7)(cid:8)(cid:6)(cid:9)(cid:1)10-710-610-510-410-310-210-1100(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:3)(cid:7)(cid:8)(cid:6)(cid:5)(cid:7)(cid:3)(cid:10)(cid:11)(cid:7)(cid:12)(cid:13)(cid:10)(cid:7)(cid:14)(cid:15)(cid:9)(cid:2)(cid:16)(cid:1)(cid:17)(cid:18)(cid:19)(cid:20)(cid:9)(cid:21)(cid:22)(cid:19)(cid:2)(cid:23)(cid:2)(cid:24)(cid:2)(cid:15)(cid:4)(cid:25)(cid:6)(cid:3)(cid:26)(cid:20)(cid:22)(cid:9)(cid:20)(cid:26)(cid:20)(cid:4)(cid:2)(cid:13)(cid:3)(cid:2)(cid:24)(cid:27)(cid:6)(cid:12)(cid:3)(cid:2)(cid:16)(cid:1)(cid:17)(cid:1)(cid:1)(cid:1)(cid:28)(cid:3)(cid:29)(cid:26)(cid:2)(cid:16)(cid:1)(cid:17)(cid:1)(cid:1)(cid:1)(cid:28)(cid:3)(cid:21)(cid:20)A100101102103(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:5)(cid:7)(cid:8)(cid:2)(cid:9)(cid:10)(cid:11)(cid:12)(cid:13)(cid:1)10-710-610-510-410-310-210-1100(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:5)(cid:7)(cid:8)(cid:2)(cid:9)(cid:10)(cid:11)(cid:12)(cid:5)(cid:10)(cid:3)(cid:9)(cid:8)(cid:10)(cid:14)(cid:7)(cid:9)(cid:10)(cid:11)(cid:12)(cid:13)(cid:2)(cid:15)(cid:1)(cid:16)(cid:17)(cid:18)(cid:19)(cid:13)(cid:20)(cid:21)(cid:18)(cid:2)(cid:22)(cid:2)(cid:23)(cid:2)(cid:12)(cid:4)(cid:24)(cid:6)(cid:3)(cid:25)(cid:19)(cid:21)(cid:13)(cid:19)(cid:25)(cid:19)(cid:4)(cid:2)(cid:7)(cid:3)(cid:2)(cid:23)(cid:26)(cid:6)(cid:14)(cid:3)(cid:2)(cid:15)(cid:1)(cid:16)(cid:1)(cid:1)(cid:1)(cid:27)(cid:3)(cid:28)(cid:17)(cid:2)(cid:15)(cid:1)(cid:16)(cid:1)(cid:1)(cid:1)(cid:28)(cid:3)(cid:25)(cid:28)B.100101102103(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:3)(cid:7)(cid:8)(cid:6)(cid:9)(cid:1)10-710-610-510-410-310-210-1100(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:3)(cid:7)(cid:8)(cid:6)(cid:5)(cid:7)(cid:3)(cid:10)(cid:11)(cid:7)(cid:12)(cid:13)(cid:10)(cid:7)(cid:14)(cid:15)(cid:9)(cid:2)(cid:16)(cid:1)(cid:17)(cid:18)(cid:19)(cid:20)(cid:9)(cid:21)(cid:22)(cid:18)(cid:2)(cid:23)(cid:2)(cid:24)(cid:2)(cid:15)(cid:4)(cid:25)(cid:6)(cid:3)(cid:26)(cid:21)(cid:22)(cid:9)(cid:27)(cid:28)(cid:27)(cid:4)(cid:2)(cid:13)(cid:3)(cid:2)(cid:24)(cid:29)(cid:6)(cid:12)(cid:3)101102(cid:1)100101102(cid:1)(cid:3)spont(cid:2)(cid:16)(cid:1)(cid:17)(cid:30)(cid:2)(cid:10)(cid:2)(cid:31)(cid:32)(cid:1)(cid:2)(cid:16)(cid:1)(cid:17)(cid:3)(cid:1)(cid:1)(cid:20)(cid:4)(cid:18)(cid:18)(cid:2)(cid:16)(cid:1)(cid:17)(cid:3)(cid:1)(cid:1)(cid:21)(cid:4)(cid:28)(cid:22)A100101102103(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:5)(cid:7)(cid:8)(cid:2)(cid:9)(cid:10)(cid:11)(cid:12)(cid:13)(cid:1)10-710-610-510-410-310-210-1100(cid:1)(cid:2)(cid:3)(cid:4)(cid:2)(cid:5)(cid:6)(cid:5)(cid:7)(cid:8)(cid:2)(cid:9)(cid:10)(cid:11)(cid:12)(cid:5)(cid:10)(cid:3)(cid:9)(cid:8)(cid:10)(cid:14)(cid:7)(cid:9)(cid:10)(cid:11)(cid:12)(cid:13)(cid:2)(cid:15)(cid:1)(cid:16)(cid:17)(cid:18)(cid:19)(cid:13)(cid:20)(cid:21)(cid:17)(cid:2)(cid:22)(cid:2)(cid:23)(cid:2)(cid:12)(cid:4)(cid:24)(cid:6)(cid:3)(cid:25)(cid:20)(cid:21)(cid:13)(cid:26)(cid:27)(cid:26)(cid:4)(cid:2)(cid:7)(cid:3)(cid:2)(cid:23)(cid:28)(cid:6)(cid:14)(cid:3)(cid:2)(cid:15)(cid:1)(cid:16)(cid:1)(cid:1)(cid:1)(cid:20)(cid:3)(cid:17)(cid:25)(cid:2)(cid:15)(cid:1)(cid:16)(cid:1)(cid:1)(cid:1)(cid:20)(cid:3)(cid:27)(cid:19)B R. V. Williams-Garc´ıa et al. each node exhibited a vastly different number of sponta- neous activations. Refractory periods for each node were set to 1 time step and simulations were run over 3.6× 106 time steps. The resulting avalanche (defined in 1 ms bins) and c-web size and duration probability distributions for simulations performed at κ = 0.80 are plotted in fig. 4 using a logarithmic binning of 1.1 (for a description of log-binning, refer to Appendix E of [30]). In fig. 5, we plot size and duration probability dis- tributions (again with logarithmic binning of 1.1) using avalanches and c-webs identified directly from the data; as predicted in fig. 1, maximum c-web durations were longer than those of avalanches in all ten data sets. Maximum likelihood estimation of the putative power law exponents was performed over the regions indicated by the dashed lines in fig. 5 using the methods described in [50]. The dashed lines simply serve as guides to the eye. This anal- ysis produced putative exponents of 2.06 and 1.33 with log-likelihoods −3.47 and −4.05 for avalanche and c-web sizes, respectively (see fig. 5A). The avalanche and c-web duration distributions (fig. 5B) featured exponents of 2.37 and 2.01 with log-likelihoods of −1.70 and −3.96, respec- tively. Additionally, we note the emergence of isolated spontaneous activation events (s = 1 c-webs) and domi- nance over larger c-webs, which does not fit the traditional picture of avalanche criticality. This demonstrates that while neuronal avalanches may exhibit apparent power- law scaling, thus suggesting potential underlying critical behavior, the corresponding c-web distributions do not; this is evident from both a qualitative and quantitative examination. Although we have not determined whether the data feature quasicritical dynamics, operating at or near the nonequilibrium Widom line, these results strongly suggest non-critical dynamics. To further illustrate how avalanches confound important dynamical information in the data, we plotted the average number of spontaneous events (cid:104)nspont(cid:105)(s), contained in avalanches as a function of their size s, and compared this to the expected result from a situation with a separation of timescales (which we verified using simulations), where SOC would be applica- ble (see fig. 5A inset). In other words, our results put the application of the SOC framework to complex neural network dynamics into question. In summary, we introduced here a novel spatiotemporal cascade of causally related activity, c-webs, to describe network dynamics in ways which contrast and comple- ment conventional avalanches. This allows us to separate causally-related from unrelated events in complex network dynamics. Whereas avalanches strongly depend on the choice of temporal binning [17,18,51], c-webs only depend on the accuracy of the methods used to determine them- in this case, transfer entropy (TE). While TE provides good, model-free estimates of the information flow in a network (i.e., the effective connectivity), it suffers from a number of limitations; for instance, TE requires establish- ing a suitable threshold for accepting putative connections and, in neural networks, it does not identify connections as excitatory or inhibitory [39]. Additional experimental ma- nipulations not performed here could possibly determine connection types when used in conjunction with TE (see, for example [52]). Because of this, we have validated our c-webs method on an Izhikevich model network, with 80 excitatory regular spiking neurons and 20 inhibitory fast spiking neurons-using the sample network from [39]- uniformly-distributed random synaptic delays between 1 ms and 20 ms for excitatory synapses, and 1-ms delays for inhibitory synapses. Individual neurons were stimu- lated with 10-Hz Poisson processes; activations triggered by Poisson inputs were marked as true spontaneous events. Under these conditions, the c-webs method was able to identify 71.3% of the true spontaneous events, with a false positive rate of 18.2%, demonstrating (1) the utility of the method when both excitatory and inhibitory neurons are present, and (2) that the ability to identify sponta- neous events is limited by the ability of TE to identify true connections and their associated delays. Spurious connec- tions detected by TE have been minimized here using re- cent methods [45]. We have used the cortical branching model (CBM) to compare statistical properties of c-webs with those of neuronal avalanches, finding that they en- tirely coincide in the absence of spontaneous events, i.e., when ps = 0-a situation which corresponds with a sepa- ration of timescales-and differ significantly when ps > 0, i.e., when there is a mixing of timescales. Notably, the c-web size distribution does not appear to be scale-free- a result inconsistent with self-organized criticality (SOC). Indeed, application of our c-webs method to mouse cor- tical data demonstrated that neuronal avalanches are not merely composed of causally-related events (cf. fig. 5). Potential further applications for c-webs in neural net- works are numerous. For instance, c-webs could play a role in the characterization of different classes of neurons, net- work structures, and dynamical states. Because inhibitory neurons exhibit different firing patterns from excitatory neurons, e.g., fast-spiking and tonic activation [35,36], dis- tributions of spontaneous neuronal events may help iden- tify inhibitory neurons, complementing established meth- ods [53]. Hidden network cycles might be discovered by employing our approach in conjunction with popu- lation coupling methods [54]. Additionally, c-webs may be used to distinguish recurrent from feed-forward net- work dynamics, by searching their structure for driven re- activations. Moreover, different dynamical network states and neurological disorders may be characterized by the prevalence of spontaneous events-a sort of signal-to-noise ratio. Finally, c-webs allow for a more careful examination of the quasicriticality hypothesis and practical application of the nonequilibrium Widom line framework [34] to dis- tinguish quasicriticality from criticality in living neural networks. Similar applications could be envisioned for complex networks in general. For example, financial networks could be decomposed into agents that directly interact through exchanges as well as exogenous factors like weather or in- p-6 Unveiling causal activity of complex networks [25] Shriki O. et al., J. Neurosci., 33 (2013) 7079. [26] Yu S. et al., Front. Sys. Neurosci., 2013 (7) . [27] de Arcangelis L. et al., J. Stat. Mech., 2014 (2014) P03026. [28] Shew W. L. et al., Nat. Phys., 11 (2015) 659. [29] Bak P., Tang C., and Wiesenfeld K., Phys. Rev. Lett., 59 (1987) 381. [30] Christensen K. and Moloney N. R., Complexity and Criticality (Imperial College Press, London) 2005. [31] Nishimori H. and Ortiz G., Elements of Phase Tran- sitions and Critical Phenomena (Oxford University Press, New York) 2011. [32] Hsu D. and Beggs J. M., Neurocomputing, 69 (2006) 1134. [33] Pearlmutter B. A. and Houghton C. J., Neural Com- put., 21 (2009) 1622. [34] Williams-Garc´ıa R. V., Moore M., Beggs J. M., and Ortiz G., Phys. Rev. E, 90 (2014) 062714. [35] Freund T. F. and Buzs´aki G., Hippocampus, 6 (1996) 347. [36] Hu H. et al., Science, 345 (2014) 1255263. [37] Orlandi J. G. et al., Nat. Phys., 9 (2013) 582. [38] James F., Statistical Methods in Experimental Physics (World Scientific, Singapore) 2006. [39] Ito S. et al., PLoS ONE, 6 (2011) e27431. [40] Wibral M. et al., PLoS ONE, 8 (2013) e55809. [41] Ay N. and Polani D., Advances in Complex Systems, 11 (2008) 17. [42] Lizier J. T. and Prokopenko M., Eur. Phys. J. B, 73 (2010) 605. [43] Beggs J. M. and Plenz D., J. Neurosci., 24 (2004) 5216. [44] Friedman N., Ito S., Brinkman B. A. W., Shimono M., DeVille R. E. L., Dahmen K. A., Beggs J. M., and Butler T. C., Phys. Rev. Lett., 108 (2012) 208102. [45] Nigam S. et al., J. Neurosci., 36 (2016) 670. [46] Ito S., Yeh F. C., Timme N. M., Hottowy P., Litke A. M., and Beggs J. M. (2016), Spontaneous spiking ac- tivity of hundreds of neurons in mouse somatosensory cor- tex slice cultures recorded using a dense 512 electrode array. CRCNS.org, http://dx.doi.org/10.6080/K07D2S2F [47] Litke A. M. et al., IEEE Transactions on Nuclear Sci- ence, 51 (2004) 1434. [48] Ito S. et al., PloS ONE, 9 (2014) e105324. [49] Oppenheim A. V. et al., Discrete-Time Signal Processing (Prentice Hall, Englewood Cliffs) 1989. [50] Marshall N. et al., Front. Physiol., 7 (2016) 250. [51] Plenz D. et al., Criticality in Neural Systems (Wiley- VCH, Weinheim) 2014. [52] Orlandi J. G. et al.,, PloS ONE, 9 (2014) e98842. [53] Barth´o P. et al., J. Neurophysiol., 92 (2004) 600. [54] Okun M. et al., Nature, 521 (2015) 511. [55] Runge J. et al., Nat. Commun., 6 (2015) 8502. [56] Anderson R. M. and May R. M., Nature, 1979 (280) . [57] Ratkiewicz J., Conover M. D., Meiss M., Gonc¸alves B., Patil S., Flammini A., and Menczer F., in Proceed- ings of the 20th International Conference Companion on World Wide Web (ACM, New York) 2011, p. 249. flation. In climate research, identifying causal connections between different geographic regions is important for un- derstanding the impact of localized events on global cli- mate [55]. In models of disease spreading, such as the SIRS model, c-webs could differentiate between sources of infection [56]. Such an approach is likely to be useful whenever considering interacting units, whether they are people in social networks, species in ecological webs, or protein molecules in a stochastic environment. A specific application in social media could involve the detection of Twitterbots and astroturfing [57]. ∗ ∗ ∗ The authors would like to thank Hadi Hafizi, Emily B. Miller, Benjamin Nicholson, and Zachary C. Tosi for valu- able discussions, as well as Shinya Ito, Alan M. Litke, and Fang-Chin Yeh for providing their in vitro data. R. V. Williams-Garc´ıa is presently supported by a National In- stitutes of Health Ruth L. Kirschstein National Research Service Award, T32 NS086749. REFERENCES [1] Williams-Garc´ıa R. V., Beggs J. M., and Ortiz G., arXiv:1603.05659v1 [q-bio.NC] (2016). [2] Gros C., Complex and Adaptive Dynamical Systems (Springer-Verlag, Berlin) 2013. [3] Bak P., How Nature Works: The Science of Self-Organized Criticality (Springer-Verlag, New York) 1996. [4] Bertschinger N. and Natschlager T., Neural Comput., 16 (2004) 1413. [5] Chialvo D. R., Physica A, 340 (2004) 756. [6] Haldeman C. and Beggs J. M., Phys. Rev. Lett., 94 (2005) 058101. [7] Kinouchi O. and Copelli M., Nat. Phys., 2 (2006) 348. [8] Levina A., Herrmann J. M., and Geisel T., Nat. Phys., 3 (2007) 857. [9] Kitzbichler M. G. et al., PLoS Comput. Biol., 5 (2009) e1000314. [10] Chen W. et al., BMC Neurosci., 11 (2010) 3. [11] Larremore D. B. et al., Phys. Rev. Lett., 106 (2011) 058101. [12] Mora T. and Bialek W., J. Stat. Phys., 144 (2011) 268. [13] Pei S. et al., Phys. Rev. E, 86 (2012) 021909. [14] Manchanda K. et al., Phys. Rev. E, 87 (2013) 012704. [15] Rybarsh M. and Bornholdt S., PLoS ONE, 9 (2014) e93090. [16] Worrell G. A. et al., Neuroreport, 13 (2002) 2017. [17] Beggs J. M. and Plenz D., J. Neurosci., 23 (2003) 11167. [18] Pasquale V. et al., Neuroscience, 153 (2008) 1354. [19] Poil S. S., van Ooyen A., and Linkenkaer-Hansen K., Hum. Brain Mapp., 29 (2008) 770. [20] Shew W. L. et al., J. Neurosci., 29 (2009) 15595. [21] Ribeiro T. L. et al., PLoS ONE, 5 (2010) e0014129. [22] Shew W. L. et al., J. Neurosci., 31 (2011) 55. [23] G. Solovey et al., Front. Integr. Neurosci., 6 (2012) 44. [24] Priesemann V. et al., PLoS Comput. Biol., 9 (2013) e1002985. p-7
1904.10399
1
1904
2019-04-21T02:45:09
Spike-Based Winner-Take-All Computation: Fundamental Limits and Order-Optimal Circuits
[ "q-bio.NC", "cs.DC" ]
Winner-Take-All (WTA) refers to the neural operation that selects a (typically small) group of neurons from a large neuron pool. It is conjectured to underlie many of the brain's fundamental computational abilities. However, not much is known about the robustness of a spike-based WTA network to the inherent randomness of the input spike trains. In this work, we consider a spike-based $k$--WTA model wherein $n$ randomly generated input spike trains compete with each other based on their underlying statistics, and $k$ winners are supposed to be selected. We slot the time evenly with each time slot of length $1\, ms$, and model the $n$ input spike trains as $n$ independent Bernoulli processes. The Bernoulli process is a good approximation of the popular Poisson process but is more biologically relevant as it takes the refractory periods into account. Due to the randomness in the input spike trains, no circuits can guarantee to successfully select the correct winners in finite time. We focus on analytically characterizing the minimal amount of time needed so that a target minimax decision accuracy (success probability) can be reached. We first derive an information-theoretic lower bound on the decision time. We show that to have a (minimax) decision error $\le \delta$ (where $\delta \in (0,1)$), the computation time of any WTA circuit is at least \[ ((1-\delta) \log(k(n -k)+1) -1)T_{\mathcal{R}}, \] where $T_{\mathcal{R}}$ is a difficulty parameter of a WTA task that is independent of $\delta$, $n$, and $k$. We then design a simple WTA circuit whose decision time is \[ O( \log\frac{1}{\delta}+\log k(n-k))T_{\mathcal{R}}). \] It turns out that for any fixed $\delta \in (0,1)$, this decision time is order-optimal in terms of its scaling in $n$, $k$, and $T_{\mathcal{R}}$.
q-bio.NC
q-bio
Spike-Based Winner-Take-All Computation: Fundamental Limits and Order-Optimal Circuits Lili Su Computer Science & Artificial Intelligence Laboratory Massachusetts Institute of Technology [email protected] Chia-Jung Chang Brain and Cognitive Sciences Massachusetts Institute of Technology [email protected] Nancy Lynch Computer Science & Artificial Intelligence Laboratory Massachusetts Institute of Technology [email protected] April 24, 2019 Abstract Winner-Take-All (WTA) refers to the neural operation that selects a (typically small) group of neurons from a large neuron pool. It is conjectured to underlie many of the brain's fundamental computational abilities. However, not much is known about the robustness of a spike-based WTA network to the inherent randomness of the input spike trains. In this work, we consider a spike-based k -- WTA model wherein n randomly generated input spike trains compete with each other based on their underlying statistics, and k winners are supposed to be selected. We slot the time evenly with each time slot of length 1 ms, and model the n input spike trains as n independent Bernoulli processes. The Bernoulli process is a good approximation of the popular Poisson process but is more biologically relevant as it takes the refractory periods into account. Due to the randomness in the input spike trains, no circuits can guarantee to successfully select the correct winners in finite time. We focus on analytically characterizing the minimal amount of time needed so that a target minimax decision accuracy (success probability) can be reached. We first derive an information-theoretic lower bound on the decision time. We show that to have a (minimax) decision error ≤ δ (where δ ∈ (0, 1)), the computation time of any WTA circuit is at least ((1 − δ) log(k(n − k) + 1) − 1)TR, 9 1 0 2 r p A 1 2 ] . C N o i b - q [ 1 v 9 9 3 0 1 . 4 0 9 1 : v i X r a where TR is a difficulty parameter of a WTA task that is independent of δ, n, and k. We then design a simple WTA circuit whose decision time is (cid:19) (cid:19) (cid:18)(cid:18) (cid:18) 1 (cid:19) δ O log + log k(n − k) TR . It turns out that for any fixed δ ∈ (0, 1), this decision time is order-optimal in terms of its scaling in n, k, and TR. 1 Introduction Humans and animals can form a stable perception and make robust judgments under ambiguous condi- tions. For example, we can easily recognize a dog in a picture regardless of its posture, hair color, and 1 whether it stands in the shadow or is occluded by other objects. One fundamental feature of brain com- putation is its robustness to the randomness introduced at different stages, such as sensory representations [KK01, HW59], feature integration [KTA+03, MCM07], decision formation [PG99, SN01], and motor plan- ning [HW98, LCG+15]. It has been shown that neurons encode information in a stochastic manner in the brain [BAB+97, KRR00, MA09, FDMM18]; even when the exact same sensory stimulus is presented or when the same kinematics are achieved, no deterministic patterns in the spike trains exist. Facing environmental ambiguity, humans and animals adaptively refine their behaviors by incorporating prior knowledge with their current sensory measurements [FSW08, KP04, SS06, EB02, KW04]. Nevertheless, it remains relatively unclear how neurons carry out robust computation facing ambiguity. Sparse coding is a common strategy in brain computation; to encode a task-relevant variable, often only a small group of neurons from a large neu- ron pool are activated [OF04, POMT+02, HDZ08, QKKF08, KF08, RPG99]. Understanding the underlying neuron selection mechanism is highly challenging. Winner-Take-All (WTA) is a hypothesized mechanism to select proper neurons from a competitive net- work of neurons, and is conjectured to be a fundamental primitive of cognitive functions such as attention and object recognition [RP99, IKN98, YG98, Maa00]. Among these studies, it is commonly assumed that neurons transmit information with a continuous variable such as the firing rate. This assumption, however, ignores how temporal coding may additionally contribute to cortical computations. For example, some neurons in the auditory cortex will respond to auditory events with bursts at a fixed latency [GKvHW96, Nel04]. This phase- locking property is also observed in the hippocampus as well as the prefrontal cortex [SLW05, HSM06, BC95]. Another feature that has been neglected in a rate-based model is the inherent noise in the inputs. Although some studies used additive Gaussian noise [KCF17, LLW13, LIKB99, RV06] to account for input random- ness, such WTA circuits are very sensitive to noise and could not successfully select even a single winner unless extra robustness strategy such as an additional nonlinearity is introduced into the dynamics [KCF17]. Last but not least, neurons have a refractory period, which prevents spikes from back propagating in axons [BIM98], and such a feature is usually neglected in the rate-based models. In contrast, a spike-based model may capture these neglected features. Nevertheless, how WTA computation can be implemented and its algorithmic characterization remains relatively under-explored. In this paper, we study a spike-based k-WTA model wherein n randomly generated input spike trains are competing with each other with their underlying statistics, and the true winners are the k input spike trains whose underlying statistics are higher than others. More precisely, we slot the time evenly with each time slot of length 1 ms. We assume that these n input spike trains are generated by n independent Bernoulli processes with different rates. An abstract example is depicted in Figure 1. We use Bernoulli processes to capture the randomness in the input spike trains rather than using the popular Poisson processes because a Bernoulli process can be viewed as the time-slotted version of a refractory-period-modified Poisson process; it is well-known that due to the existence of refractory periods, a neuron cannot spike twice within 1 ms. We focus on analytically characterizing the minimal amount of time needed so that a target minimax decision accuracy (success probability) can be reached. We first derive a lower bound on the decision time for a given decision accuracy. We show that no WTA circuits can have a computation time strictly less than ((1 − δ) log(k(n − k) + 1) − 1)TR, (1) where TR is a parameter defining the difficulty to distinguish between two spike trains with different statistics in a WTA task, n is the number of input spike trains, k is the number of winners, and δ is the given target decision accuracy. In many practical settings we care about the sparse coding region where k (cid:28) n. Our lower-bound is obtained by an information-theoretic argument, and holds for all WTA circuits without restricting their circuit architectures and their adopted activation functions. Throughout this paper, we are interested in the decision time's scaling in n, k, and TR, while treating δ ∈ (0, 1) as a fixed small constant. Not surprisingly, the above lower bound grows with the network size n when other parameters are fixed. This is because the larger n, the noisier the WTA competition. Similarly, when n and k are fixed, the easier to distinguish two spike trains with different statistics (i.e., the smaller TR), the shorter the necessary decision time is. We construct a simple circuit whose decision time is (cid:18)(cid:18) (cid:18) 1 (cid:19) δ O log + log k(n − k) TR . (cid:19) (cid:19) 2 Figure 1: In this figure, n randomly generated input spike trains are fed to a WTA circuit as the circuit input. Clearly, no deterministic patterns can be read off from these spike trains. Here, we do not specify the output of a WTA circuit because the detailed specifications of the circuits' outputs might vary with the corresponding applications. It turns out that for any fixed δ ∈ (0, 1), this decision time is order-optimal in terms of its scaling in n, k, and TR, i.e., its decision time matches the lower-bound in (1) up to a constant multiplicative factor. In our circuit, each output neuron is a thresholded accumulator unit whose threshold is determined by δ, k, n, and TR, and the circuit's output is the first group of k output neurons that spike in the same time. The typical dynamics under our circuit are: the number of output neurons that spike simultaneously (i.e., spike at the same time) is monotonically increasing until exactly k output neurons spike simultaneously. The simultaneous spikes of these k output neurons cause strong inhibition of other output neurons; in particular, no other output neuron can spike within a sufficiently long period Ω(cid:0)(cid:0)log(cid:0) 1 In addition, our results also give a set of testable hypotheses on neural recordings and humans'/animals' behaviors in decision-making. For instance, given the number of input spike trains and the number of true winners, our results can provide an estimate of the minimum decision time needed, which can provide some insights on the efficiency of a WTA circuit in terms of decision time. As another example, when two animals are involved in the same experiment, if both animals reach the same accuracy in discriminating two objects, does the animal that decides faster have more heterogeneous distributions of input spiking activities, i.e., smaller TR? Our results provide partial answers to this question. (cid:1) + log k(n − k)(cid:1) TR(cid:1). δ 2 Computational Model: Spiking Neuron Networks In this section, we provide a general description of the computation model used. There is much freedom in choosing the detailed specification of the model. In particular, in Section 5 we provide a circuit construction (for solving the k -- WTA competition) under this computation model. 2.1 Network Structure A spiking neuron network (SNN) N = (U, E) consists of a collection of neurons U that are connected through synapses E. We assume that a SNN can be conceptually partitioned into three non-overlapping layers: input layer Nin, hidden layer Nh, and output layer Nout; the neurons in each of these layers are referred to as input neurons, hidden neurons, and output neurons, respectively. The synapses E are essentially directed edges, i.e, E := {(ν, ν(cid:48)) : ν, ν(cid:48) ∈ U}. For each ν ∈ U , define PREν := {ν(cid:48) : (ν(cid:48), ν) ∈ E} and POSTν := {ν(cid:48) : (ν, ν(cid:48)) ∈ E}. Intuitively, PREν is the collection of neurons that can directly influence neuron ν; similarly, POSTν is the collection of neurons that can be directly influenced by neuron ν. 1 We assume that the input neurons cannot be influenced by other neurons in the network, i.e., PREν = ∅ for all ν ∈ Nin. Each edge (ν, ν(cid:48)) in E has a weight, denoted by w(ν, ν(cid:48)). The strength of the interaction between neuron ν and neuron ν(cid:48) is captured 1In the languages of computational neuroscience, the incoming neighbors and outgoing neighbors are often referred to as pre-synaptic units and post-synaptic units. 3 a WTA circuitInput spike train 1. . . Input spike train 2Input spike train noutputk selected winners Figure 2: A SNN consists of three layers: the input layer, the output layer, and the hidden layer. The hidden neurons might connect to both the input neurons and the output neurons to assist the computation of the neuron network. Neurons are connected through synapses. WTA circuits is a family of SNNs in which the number of output neurons equals the number of the input neurons. as w(ν, ν(cid:48)). The sign of w(ν, ν(cid:48)) indicates whether neuron ν excites or inhibits neuron ν(cid:48): In particular, if neuron ν excites neuron ν(cid:48), then w(ν, ν(cid:48)) > 0; if neuron ν inhibits neuron ν(cid:48), then w(ν, ν(cid:48)) < 0. The set E might contain self-loops with w(ν, ν) capturing the self-excitatory/self-inhibitory effects. An example of SNNs can be found in Figure 2. Generic network structure for WTA circuits The family of WTA circuits under consideration is rather generic. We only assume that Nin = Nout = n the numbers of the input neurons and of the output neurons are equal. For ease of exposition, denote Nin = {u1,··· , un} , and Nout = {v1,··· , vn} . The hidden neuron subset Nh can be arbitrary. The output neurons and the hidden neurons may be connected to each other in an arbitrary manner. 2.2 Network State In a SNN, the communication among neurons is abstracted as spikes. We assume each neuron ν has two local variables: spiking state variable S(ν) and memory state variable M (ν). Nevertheless, for input neurons, we only consider their spiking states, assuming that their memory states are not influenced by the dynamics of the spiking neuron network under consideration. We slot the time evenly with each time slot of length 1 ms. Let t = 1, 2,··· be the indices of the time slots. Henceforth, by saying time t, we mean the time interval [t− 1, t) ms. For t ≥ 1, let St(ν) ∈ {0, 1} be the spiking state of neuron ν at time t indicating whether neuron ν spikes at time t or not. By convention, S0(ν) := 0. For a non-input neuron ν and for t ≥ 1, let Mt(ν) be the memory state of neuron ν at time t summarizing the cumulative influence caused by the spikes of the neurons in PREi during the most recent m times, i.e., times t − 1, t − 2,··· , t − m. Concretely, let Vt(ν) be the charge of (non-input) neuron ν at time t (for t ≥ 1) defined as (cid:88) ν(cid:48)∈PREν Vt(ν) := w(ν(cid:48), ν)St(ν(cid:48)). Clearly, V0(ν) = 0. Let V ν t be the sequence of length m such that := [Vt(ν),··· , Vt−m+1(ν)] , V ν t and let St(ν) be the sequence of length m such that t := [St(ν),··· , St−m+1(ν)] . Sν 4 Input layerHidden layer......Output layer By convention, when 0 ≤ t ≤ m, let and := [Vt(ν),··· , V0(ν), 0,··· , 0] V ν t t := [St(ν),··· , S0(ν), 0,··· , 0] . Sν For t ≥ 1, define the memory variable Mt(ν) as a pair of vectors Sν t−1 and V ν t−1, i.e., Mt(ν) :=(cid:0)Sν t−1, V ν t−1 (cid:1) . By convention, let M0(ν) := (0, 0), where 0 is the length m zero vector. At time t + 1, the memory variable Mt+1(ν) is updated by shifting the two sequences forwards by one time unit -- fetching in St(ν) and Vt(ν), respectively, and removing St−m(ν) and Vt−m(ν), respectively. The memory state Mt(ν) is known to neuron ν only, and it can influence the probability of generating a spike at time t through an activation function φν, i.e., St(ν) = φν (Mt(ν)) ,∀ t ≥ 0. (2) Notably, φν might be a random function. In most neurons, the synaptic plasticity time window is about 80 -120 msec, but could also vary across In a sense, the brain regions, and vary across different time scales under different behavioral contexts. synaptic plasticity time window is closely related to m. As can be seen in Section 5, our order-optimal WTA circuit construction requires m to be sufficiently high. Nevertheless, this does not exclude the application of our WTA circuit to the contexts where m is small. This is because the memory variable can be implemented by a chain of hidden neurons near neuron ν. The detailed implementation of the local memory does not affect the order optimality of our WTA circuit. 3 Minimax Decision Accuracy/Success Probability 3.1 Random Input Spike Trains We study the k -- WTA model, wherein n randomly generated input spike trains are competing with each other, and, as a result of this competition, k out of them are selected to be the winners. In contrast, most existing works [VRP+18, Maa97, LMP16] assume deterministic input spike trains. Recall that time is slotted into intervals of length 1 ms. We assume that the n input spike trains are generated from n independent Bernoulli processes with unknown parameters p1,··· , pn, respectively. We refer to p = [p1,··· , pn] as a rate assignment of the WTA competition. For example, suppose there are 2 input spike trains with rates 0.6 and 0.8, respectively, i.e., n = 2 and p = [0.6, 0.8]. In each time, with probability 0.6 the first input spike train has a spike independently from whether the second input spike train has a spike or not; similarly for the second input spike train. Note that in the most general scenario the spikes of the input neurons might be correlated; see Section 6 for detailed comments. We would like to explore the more general input spikes in our future work. 3.2 Minimax Performance Metric We adopt the minimax framework [Wu17] (in which the circuit designer and nature play games against each other) to evaluate the performance (decision accuracy versus decision time) of a WTA circuit. Let R ⊆ [c, C] be an arbitrary but finite set of rates where c and C are two absolute constants such that 0 < c < C < 1. A rate assignment p is chosen by nature from Rn for which there exists a subset of [n] := {1,··· , n}, denoted by W(p), such that W(p) = k, and pi > pj ∀ i ∈ W(p), j /∈ W(p) (3) 5 -- recall that · is the cardinality of a set. We refer to set W(p) as the true winners with respect to the rate assignment p. For example, suppose n = 5, k = 2, and p = [p1 = 0.2, p2 = 0.1, p3 = 0.2, p4 = 0.8, p5 = 0.85] . Here the true winners are 4 and 5, i.e., W(p) = {4, 5}. In this paper, we consider the following collection of rate assignments, denoted by AR: AR := {p : p ∈ Rn, & ∃W(p) ⊆ [n] s.t. W(p) = k, and pi > pj ∀ i ∈ W(p), j /∈ W(p)} . (4) For each of reference, we refer to an element in AR as an admissible rate assignment. Recall that the input of a WTA circuit is a collection of n independent spike trains. For a given rate assignment p, let {St(ui)}T t=1 denote the spike train of length T at input neuron ui. The circuit designer wants to design a WTA circuit that outputs a good guess/estimate (cid:100)win of W(p) for any choice of rate assignment p in AR. Note that the estimate (cid:100)win is independent of p. Here S is used with a little abuse of notation as this notation hides (cid:104){St(u1)}T t=1 ,··· ,{St(un)}T conditioning on its connection with T and the rate parameter p. 2 Later, we use the same notation to denote the n spike trains with random rate assignment, i.e., where p is randomly generated. Nevertheless, this abuse of notation significantly simplifies the exposition without sacrificing clarity. In particular, we will specify the underlying rate assignment when it is not clear from the context. S := (cid:105) , t=1 Under minimax framework, we are interested in the minimax error probability 3 For a given deterministic WTA circuit (cid:100)win (i.e., the activation functions used are deterministic), the prob- (cid:111) ability in P(cid:110)(cid:100)win (S) (cid:54)= W(p) for a randomized WTA circuit (cid:100)win (i.e., the activation functions are stochastic), in addition to the aforemen- tioned source of randomness, the probability in P(cid:110)(cid:100)win (S) (cid:54)= W(p) (cid:111) P(cid:110)(cid:100)win (S) (cid:54)= W(p) (cid:111) the activation functions. In (5), the performance metric of a WTA circuit is the worst-case error probability is taken w.r.t. the randomness in the stochastic spikes of each input neuron; is also taken w.r.t. the randomness in . max p∈AR (5) Essentially, the statistical inference problem can be viewed as a game between the circuit designer and nature. 4 Information-Theoretic Lower Bound on Decision Time In this section, we provide a lower bound on the decision time for a given decision accuracy. The lower bounds derived in this section hold universally for all possible network structures (including the hidden layer), synapse weights, and the activation functions. One observation is that the decision time is naturally lower bounded by the sample complexity, which is closely related to the Kullback-Leibler (KL) divergence4 between two Bernoulli distributions. The KL divergence between Bernoulli random variables with parameters r and r(cid:48), respectively, is defined as P(cid:110)(cid:100)win (S) (cid:54)= W(p) (cid:111) . min(cid:100)win max p∈AR d(r (cid:107) r(cid:48)) := r log (cid:17) (cid:16) r (cid:104){St(u1)}T + (1 − r) log r(cid:48) t=1 , · · · , {St(un)}T (cid:19) (cid:18) 1 − r 1 − r(cid:48) (cid:105) , (6) 2A more rigorous notation should be S(T, p) := . We use S for S(T, p) for ease of exposition. 3In the following expression, the min should really be an inf, but we abuse notation here for ease of exposition. In addition, t=1 the max really is a max as the set R under consideration is of finite size. 4The Kullback-Leibler (KL) divergence gauges the dissimilarity between two distributions. 6 0 := 0. Notably, d(· (cid:107) ·) is not symmetric in r and r(cid:48). In addition, when r (cid:54)= 0 where, by convention, log 0 and r(cid:48) = 0 or 1, d(r (cid:107) r(cid:48)) = ∞. Recall that set R is an arbitrary but finite set that are contained in the It holds that d(r (cid:107) r(cid:48)) < ∞ for all r, r(cid:48) ∈ R. For the more general interval [c, C], where c, C ∈ (0, 1). distributions over a common discrete alphabet A, say distributions P and Q, the Kullback-Leibler (KL) divergence between P and Q is defined as follows. Definition 1 (KL-divergence). Let A be a discrete alphabet (finite or countably infinite), and P and Q be two distributions over A. Then define where 0 · log(cid:0) 0 0 D(P (cid:107) Q) := (cid:1) = 0 by convention. (cid:88) a∈A P (a) log (cid:18) P (a) (cid:19) Q(a) , Note that D(P (cid:107) Q) ≥ 0 and D(P (cid:107) Q) = 0 if and only if P = Q almost surely. Similar to d(· (cid:107) ·), D(P (cid:107) Q) is not symmetric in P and Q. In this paper, we choose the base to be 2. 5 Recall that the set of admissible rate assignments AR is defined in (4). Lemma 2. Fix a finite set R. Let p = [p1,··· , pn] and q = [q1,··· , qn] be two rate assignments in AR. Let PS and QS be the distributions of the n spike sequences of the input neurons under rate assignments p and q, respectively. Then, n(cid:88) D(PS (cid:107) QS) = T d(pi (cid:107) qi). Lemma 2 is proved in Appendix B. For the given R, define task complexity TR as i=1 TR := max r1,r2∈R s.t. r1(cid:54)=r2 1 d(r2 (cid:107) r1) + d(r1 (cid:107) r2) . (7) It is closely related to the smallest KL divergence between two distinct statistics in R. The task complexity TR kicks in due to the adoption of minimax decision framework (5). The following lemma is used in the proof of our information-theoretic lower bound. This is a technical supporting lemma, and the choice of the specific rate assignments is due to some technical convenience in proving Theorem 4. Lemma 3. For any finite set R, let r1, r2 ∈ R such that r1 (cid:54)= r2. Let p0 =(cid:2)p0 (cid:3) be 1,··· , p0 n p0 (cid:96) = if (cid:96) = 1,··· , k; otherwise. (8) r1, r2, (cid:40) p0 (cid:96) , p0 j , p0 i , For i = 1,··· , k and j = k + 1,··· , n, define rate assignment pij as if (cid:96) (cid:54)= i,(cid:54)= j; if (cid:96) = i; if (cid:96) = j. pij (cid:96) = Let Xp be a random rate assignment. If Xp is uniformly distributed over {p0} ∪(cid:8)pij : i = 1,··· , k, & j = k + 1,··· , n(cid:9) , then the mutual information I(Xp; S) satisfies the following: I(Xp; S) ≤ T (d(r2 (cid:107) r1) + d(r1 (cid:107) r2)) . 5Note that any base would work, see [PW14, Chapter 1.1]. 7 See Appendix A for definition of I(· ; ·). The proof of Lemma 3 can be found in Appendix C. It turns out that if the input spike train length T is not sufficiently large (specified in Theorem 4), no matter how elegant the design of a WTA circuit is (no matter which activation function we choose, how many hidden neurons we use, and how we connect the hidden neurons and output neurons), its actual decision accuracy is always lower than the target decision accuracy 1 − δ. Theorem 4. For any 1 ≤ k ≤ n − 1 and any set R and any δ ∈ (0, 1), if T ≤ ((1 − δ) log(k(n − k) + 1) − 1) TR, then min(cid:100)win max p∈AR P(cid:110)(cid:100)win (S) (cid:54)= W(p) (cid:111) ≥ δ, where the min is taken over all possible WTA circuits with different choices of activation functions and circuit architectures. WTA circuit is greater than δ, i.e., maxp∈AR P(cid:110)(cid:100)win (S) (cid:54)= W(p) Theorem 4 says that if T < ((1 − δ) log(k(n − k) + 1) − 1) TR, the worst case probability error of any > δ. Theorem 4 is proved in Appendix (cid:111) E. Remark 5 (Tightness of the lower bound in Theorem 4). Following our line of argument, by considering a richer family of critical rate assignments in Lemma 3, we might be able to obtain a tighter lower bound. Nevertheless, the constructed WTA circuit in Section 5 turn out to be order-optimal -- its decision time matches the lower bound in Theorem 4 up to a multiplicative constant factor. This immediately implies that the lower bound obtained in Theorem 4 is tight up to a multiplicative constant factor. 5 Order-Optimal WTA Circuits In Section 2.1, we provided a general description of the computation model we are interested in. In this section, we construct a specific WTA circuit under this general computation model. This WTA circuit turns out to be order-optimal in terms of decision time -- the decision time of our WTA circuit matches the lower bound in Section 4 up to a multiplicative constant factor. To do that, we need to specify (1) the network structure, including the number of hidden neurons, the collection of synapses (directed communication links) between neurons, and the weights of these synapses; (2) the memorization capability of each neuron, i.e., the magnitude of m; and (3) φν -- the activation function used by neuron ν. 5.1 Circuit Design In our designed circuit, there are four parameters R, m, b, and δ, where R ⊆ [c, C] 6 is a finite set from which the pi's of the input spike trains are chosen, m is the memory range and b is the bias at the non-input neurons, and (1 − δ) is the target decision accuracy (i.e., success probability). Here, we assume that every non-input neuron has the same bias, i.e., bν = b for all non-input neurons ν. The four parameters R, m, b, and δ can be viewed as some prior knowledge of the WTA circuit; they might be learned through some unknown network development procedure which is outside the scope of this work. In Sections 5.1.1, 5.1.3, and 5.1.4, we present the network structure and the activation functions adopted, and the requirement on m. For completeness, we specify the local memory update (in particular the vector V ) separately in Section 5.1.2. The dynamics of our WTA circuit is summarized in Section 5.1.5. 5.1.1 Network structure: We propose a WTA circuit with the following network structure: 6Recall that c, C ∈ (0, 1) are two absolute constants, i.e., they do not change with other parameters of the WTA circuit such as n, k, and δ. 8 • All output neurons are connected to each other by a complete graph. That is, (vi, vj) ∈ E for all vi, vj ∈ Nout such that vi (cid:54)= vj; • Each edge from an input neuron to an output neuron has weight 1, i.e., w(ui, vi) = 1 for all ui ∈ Nh, vi ∈ Nout. • All edges among the output neurons have weights − 1 k . That is, w(vi, vj) = − 1 k for all vi, vj ∈ Nout such that vi (cid:54)= vj. • There are no hidden neurons, i.e., Nh = ∅; 5.1.2 Update local charge vector: With the above choice of network structure, the charge Vt−1(vi) at the output neuron vi at time t − 1 is Notably, Vt−1(vi) ∈ [−1, 1] for all t ≥ 1 and output neuron vi. When k = 1, the above update becomes Vt−1(vi) = St−1(ui) − 1 k (cid:88) Vt−1(vi) = St−1(ui) − (cid:88) St−1(vj). j:1≤j≤n,& j(cid:54)=i St−1(vj). j:1≤j≤n,& j(cid:54)=i which can be viewed as a spike model counterpart of the potential update under the traditional continuous rate model [KCF17, MM07] with lateral inhibition. k − 1, i.e., at time t − 1, input neuron ui spikes, and fewer than k − 1 other output neurons spike. It is easy to see the following claims hold. For brevity, their proofs are omitted. t − 1, more than k other output neurons spike. Claim 6. For t ≥ 1 and for i = 1,··· , n, Vt−1(vi) > 0 if and only if St−1(ui) = 1 and(cid:80) Claim 7. For t ≥ 1 and for i = 1,··· , n, Vt−1(vi) ≤ −1 only if(cid:80) Note that(cid:80) suppose(cid:80) (1) St−1(ui) = 1 and(cid:80) (2) St−1(ui) = 0 and(cid:80) Claim 8. For t ≥ 1 and for i = 1,··· , n, if Vt−1(vi) = 0, one of the following holds: other output neurons spike; j:1≤j≤n,& j(cid:54)=i St−1(vj) = k, i.e., at time t − 1, input neuron ui spikes, and exactly k j:1≤j≤n,& j(cid:54)=i St−1(vj) = 0, i.e., at time t − 1, input neuron ui does not spike, and j:1≤j≤n,& j(cid:54)=i St−1(vj) ≥ k is not a sufficient condition to have Vt−1(vi) ≤ −1. To see this, j:1≤j≤n,& j(cid:54)=i St−1(vj) = k and St−1(ui) = 1. In this case it holds that Vt−1(vi) = 0. j:1≤j≤n,& j(cid:54)=i St−1(vj) ≥ k, i.e., at time j:1≤j≤n,& j(cid:54)=i St−1(vj) ≤ no other output neurons spike. 5.1.3 Activation functions: There are many different choices of activation functions; see [wik] for a detailed list. In our construction, we use a simple threshold activation function, i.e., if (b − 1)1{St−1(vi)=1} +(cid:2)(cid:80)m r=1 1{Vt−r(vi)>0} − m(cid:80)m r=1 1{Vt−r(vi)≤−1}(cid:3) ≥ b; + (cid:40) 1, St(vi) = 0, otherwise, [·]+ = max [·, 0], and b > 0 is the bias at neuron vi for i = 1,··· , n. It is easy to see that this activation function falls under the general form given by (2). Remark 9. If the output neuron vi does not spike at time t − 1, i.e., St−1(vi) = 0, then in order for vi to spike at time t, the following needs to hold: (cid:34) m(cid:88) m(cid:88) 1{Vt−r(vi)>0} − m 1{Vt−r(vi)≤−1} r=1 r=1 9 (cid:35) ≥ b. + In contrast, if the output neuron vi does spike at time t − 1, i.e., St−1(vi) = 1, then (cid:34) m(cid:88) m(cid:88) 1{Vt−r(vi)>0} − m 1{Vt−r(vi)≤−1} r=1 r=1 (cid:35) ≥ 1 + is enough for vi to spike at time t. That is, under our activation rule, St−1(vi) = 1 makes the activation of vi much easier in the next round. However, if there exists r ∈ {1, 2,··· , m} such that then Thus, m(cid:88) r=1 1{Vt−r(vi)>0} − m 1{Vt−r(vi)≤−1} = 1, m(cid:88) r=1 1{Vt−r(vi)≤−1} ≤ m(cid:88) 1{Vt−r(vi)>0} − m r=1 ≤ m − m = 0. (cid:34) m(cid:88) (b − 1)1{St−1(vi)=1} + = (b − 1)1{St−1(vi)=1} + 0 ≤ b − 1 < b, r=1 1{Vt−r(vi)>0} − m m(cid:88) r=1 1{Vt−r(vi)≤−1} (cid:35) + i.e., the output neuron vi does not spike at time t. In other words, as long as there exists r ∈ {1, 2,··· , m} such that 1{Vt−r(vi)≤−1} = 1, the activation of vi is inhibited at time t. 5.1.4 Local memorization capability: In our proposed circuit, we require that m satisfies the following: (cid:18) (cid:18) 3 (cid:19) m ≥ 8C 2(1 − c) c2(1 − C) log δ + log k(n − k) (cid:19) TR := m∗ (9) for target decision accuracy 1 − δ ∈ (0, 1). In addition, we set b = cm∗. Recall that c, C ∈ (0, 1) are two absolute constants that are lower bound and upper bound of any R, respectively. Intuitively, when other parameters are fixed, the higher the desired accuracy (i.e., the smaller δ) , the larger m∗, i.e., the more memory is needed for selecting the winners in our WTA circuit. Similarly, the easier to distinguish two spike trains with different statistics (i.e., the lower TR), the smaller m∗. Interesting, with other parameters fixed, m∗ depends on k as follows: m∗ is increasing in k when k ∈(cid:8)1,··· ,(cid:98) n is decreasing in k when k ∈(cid:8)(cid:100) n 2(cid:99)(cid:9), and m∗ 2(cid:101),··· , n − 1(cid:9). In many practical settings we care about the region where k (cid:28) n. Besides, with the choice of bias b = cm∗, the larger m∗ also implies longer time is needed for our WTA circuit to declare k winners; details can be found (1) in Theorem 10. On the other hand, in most neurons the synaptic plasticity time window is about 80-120 ms, and it is unclear whether (9) can be immediately satisfied or not. Fortunately, even if (9) is not immediately satisfied by a neuron due to its local bio-plausibility, it is possible that its local memory might be realized using a chain of hidden neurons. 5.1.5 Algorithm 1 The dynamics of our WTA circut is summarized in Algorithm 1, which is fully determined by what has been described in Sections 5.1.1, 5.1.2, 5.1.3, and 5.1.4. For Algorithm 1, we declare the first k output neurons that spike simultaneously to be winners. 10 Algorithm 1: k -- WTA 1 Input: R, m, b, and δ. 2 for t ≥ 1 do if (b − 1)1{St−1(vi)=1} +(cid:2)(cid:80)m At output neuron vi for i = 1,··· , n: Vt−1(vi) ← St−1(ui) − 1 Vt−1(vi) ← [Vt−1(vi), Vt−2(vi),··· , Vt−m(vi)]; St−1(vi) ← [St−1(vi), St−2(vi),··· , St−m(vi)]; Mt(vi) ← (Vt−1(vi), St−1(vi)); St(vi) ← 1. St(vi) ← 0. r=1 1{Vt−r(vi)>0} − m(cid:80)m else (cid:80) r=1 1{Vt−r(vi)≤−1}(cid:3) k 3 4 5 6 7 8 9 10 j:1≤j≤n,&j(cid:54)=i St−1(vj); ≥ b then + 5.2 Circuit Performance Recall that W(p) and m∗ are defined in (3) and (9), respectively. Theorem 10. Fix δ ∈ (0, 1], and 1 ≤ k ≤ n − 1. Choose m ≥ m∗ and b = max{cm∗, 2}. Then for any admissible rate assignment p, with probability at least 1 − δ, the following hold: (1) There exist k output neurons that spike simultaneously by time m∗. (2) The first set of such k output neurons are the true winners W(p). (3) From the first time in which these k output neurons spike simultaneously, these k output neurons spike consecutively for at least b times, and no other output neurons can spike within b times. The proof of Theorem 10 can be found in Appendix F. The first bullet in Theorem 10 implies that our WTA circuit can provide an output (a selection of k output neurons) by time m∗; the second bullet in Theorem 10 says that the circuit's output indeed corresponds to the k true winners; and the third bullet says that the k simultaneous spikes of the selected winners are stable -- the k selected winners continue to spike consecutively for at least b times. The proof of Theorem 10 essentially says that with high probability, under Algorithm 1, the number of output neurons that spike simultaneously is monotonically increasing until it reaches k. Upon the simultaneous spike of k output neurons, by our threshold activation rule, we know that the other output neurons are likely to be inhibited. In particular, if these k output neurons are the first k output neurons that spike simultaneously, then the activation of the other output neurons are likely to be inhibited for at least b times. Remark 11 (Controlling stability). As can be seen from the proof of Theorem 10, in the activation function of Algorithm 1 (cid:35) ≥ b + (cid:34) m(cid:88) m(cid:88) (b − 1)1{St−1(vi)=1} + 1{Vt−r(vi)>0} − m 1{Vt−r(vi)≤−1} r=1 r=1 the first term (b − 1)1{St−r(vi)=1} is crucial in achieving (3) in Theorem 10. In fact, we can increase the stability period by introducing a stability parameter s such that 1 < s ≤ m and modifying the activation rule. Details can be found in Algorithm 2. It is easy to see that the activation function falls under the general form in (2). In the new activation function in Algorithm 2, for output neuron vi, once it spikes, it continues to spike for at least s times. Following our line of analysis in the proof of Theorem 10, it can be seen that the declared k winners, from the first time they spike simultaneously, continue to spike consecutively for at least s times. Remark 12 (Order-optimality). The decision time performance stated in (1) of Theorem 10 matches the information-theoretical lower bound in Theorem 4 up to a multiplicative constant factor both (a) when δ is sufficiently small and does not depend on n, k, TR, c, and C, and (b) when δ decays to zero at a speed at most (k(n−k))c0 where c0 > 0 is some fixed constant. The detailed order-optimality argument is given next. 1 11 Algorithm 2: k -- WTA 1 Input: R, m, b, δ, and s where 1 < s ≤ m. 2 for t ≥ 1 do k j:1≤j≤n,&j(cid:54)=i St−1(vj); (cid:80) r=1 1{Vt−r(vi)>0} − m(cid:80)m At output neuron vi for i = 1,··· , n: Vt−1(vi) ← St−1(ui) − 1 Vt−1(vi) ← [Vt−1(vi), Vt−2(vi),··· , Vt−m(vi)]; St−1(vi) ← [St−1(vi), St−2(vi),··· , St−m(vi)] ; Mt(vi) ← (Vt−1(vi), St−1(vi)). St(vi) ← 1. if St−1(vi) = 1 and ∃ r ∈ {2,··· , s} such that St−r(vi) = 0 then r=1 1{Vt−r(vi)≤−1}(cid:3) if (cid:2)(cid:80)m else ≥ b then + St(vi) ← 1. St(vi) ← 0. else 3 4 5 6 7 8 9 10 11 12 13 14 (cid:19) (cid:33) Suppose that δ is sufficiently small and does not depend on n, k, TR, c, and C Here, for ease of exposition, we illustrate the order-optimality with a specific choice of δ. In fact, the order-optimality holds generally for constant δ ∈ (0, 1) as long as it does not depend on n, k, TR, c, and C. Suppose the target decision accuracy is 1 − δ = 0.9, i.e., δ = 0.1. Then as long as n ≥ 31, for any 1 ≤ k ≤ n − 1, m∗ = 8C 2(1 − c) c2(1 − C) log 3 0.1 + log k(n − k) TR ≤ 16C 2(1 − c) c2(1 − C) log k(n − k)TR. (cid:18) On the other hand, recall from Theorem 4 that to have δ = 0.1, the decision time is no less than ((1 − δ) log(k(n − k) + 1) − 1) TR ≥ 1 2 log(k(n − k) + 1)TR ≥ 1 2 log k(n − k)TR where the first inequality holds as long as n ≥ 8. Thus, when n ≥ 31, in order to achieve the decision accuracy 1− δ = 0.9, the decision time of our WTA circuit is on the same order of the information-theoretic lower bound in Theorem 4. Suppose δ decays to zero at a moderate speed The decision time of our WTA circuit is order-optimal even for diminishing decision error δ as long as δ = Ω( (k(n−k))c0 ) where c0 > 0 -- it does not decay to zero "too fast" in k(n − k). To see this, let δ = (k(n−k))c0 for some constant c0 > 0. We have 3 3 8C 2(1 − c) c2(1 − C) log 3 3 (k(n−k))c0 + log k(n − k) TR = 8C 2(1 − c)(c0 + 1) c2(1 − C) log k(n − k)TR. (10) (cid:32) (cid:32) (cid:33) Resetting circuit when the input spike trains become quiescent In Algorithm 1, if the input spike trains become quiescent, then the corresponding circuits also become quiescent despite some delay in this response. Lemma 13. If all input neurons are quiescent at time t0, and remain to be quiescent for all t ≥ t0, then Vt(vi) = 0 and St(vi) = 0 for any t > t0 + m. Lemma 13 is proved in Appendix D. 12 6 Discussion In this paper, we investigated how k-WTA computation is robustly achieved in the presence of inherent noise in the input spike trains. In a spike-based k-WTA model, n randomly generated input spike trains are competing with each other, and the top k neurons with highest underlying statistics are the true winners. Given the stochastic nature of the spike trains, it is not trivial to properly select winners among a group of neurons. We derived an information-theoretic lower bound on the decision time for a given decision accuracy. Notably, this lower bound holds universally for any WTA circuit that falls within our model framework, regardless of their circuit architectures or their adopted activation functions. Furthermore, we constructed a circuit whose decision time matches this lower bound up to a constant multiplicative factor, suggesting that our derived lower bound is order-optimal. Here the order-optimality is stated in terms of its scaling in n, k, and TR. 6.1 Comparison to previous WTA models Randomness is introduced at different stages of brain computation and the stochastic nature of the spike trains are well observed [BAB+97, KRR00, MA09, FDMM18]. In our work, we focused on how to robustly achieve k-WTA computation in face of the intrinsic randomness in the spike trains. A common WTA model assumes that neurons transmit information by a continuous variable such as firing rate [DA01], which ignores the intrinsic randomness in spiking trains. Although some studies used additive Gaussian noise [KCF17, LLW13, LIKB99, RV06] in their rate-based WTA circuits to account for input randomness, these circuits are usually very sensitive to noise and could not successfully select even a single winner unless additional non-linearity is added [KCF17]. In fact, a neuron with a second non-linearity is similar to an output neuron in our constructed WTA circuit in that they both integrate their local inputs. Unfortunately, only simulation results were provided in [KCF17]; a theoretical justification of why such second non-linearity makes their WTA circuit robust to input noise is lacking. Though we focused on spike-based model, we hope our results can provide some insights for the rate-based model as well. On top of that, a rate-based model would require a high communication bandwidth, yet communication bandwidth is limited in the brain. Our spiking neural network model captures this feature by having a low communication cost, since it broadcasts 1 bit only. However, we did not try to model every biologically relevant feature. In several studies using spiking network models, individual units are often modeled with details like ion channels and specific synaptic connectivity. Though more biologically relevant than our spiking neuron network model, those details sig- nificantly complicate the analysis. In fact, it could be challenging and intricate to move beyond computer simulation to characterize the model dynamics (such as the spiking nature of each unit, the time it takes to stabilize, etc.) analytically. 6.2 Potential applications for physiological experiments Our work further provided testable hypotheses on how network size, similarities between input spike trains, and synaptic memory capacity would affect this lower bound. For example, in behavioral experiments using electrolytic lesions or pharmacological inhibition [CMA+03, HDS06, YLS13, KYPH16], the changes in performance are often highly variable and nonlinear. One possible reason comes from the difficulty of precisely manipulating network size as well as a lack of theoretical description of the relationship between network size and performance. With our analytical characterization, one might be able to estimate changes in the effective network size given performance in a decision-making task. Besides the effect of network size, the distribution of feature representations (i.e., different set Rs of different individual animals) could be used to account for between-subject variability in decision making. Consider a random-dot coherent motion task where animals need to decide which of two directions the majority of dots are moving [SN01]. In this task, performance accuracy and reaction time vary across animals. If we perform neural recordings in their visual cortex (i.e., to record their Rs), we might be able to decode their reaction time or accuracy, given population representations of dot motion in these cortical neurons [SN96, JM06]. For example, an animal whose stimulus-evoked responses are more heterogeneous in the visual cortex might be able to react faster given the same accuracy, governed by our derived lower-bound. 13 Last but not least, our work also offered predictions on how local memory capacity could affect perfor- mance in decision-making. For example, when there is more ambiguity in input representations, to obtain the same performance (both accuracy and decision time), a larger time window for memory storage in synapses [KPS10] is required. From previous experimental work [BMG+17], we know that synaptic plasticity has time scale ranging from milliseconds to seconds across different brain regions, and such plasticity could efficiently store entire behavioral sequences within synaptic weights. Combining with our analytical characterization, when performance accuracy changes over time, assuming other parameters such as input statistics, decision time and network size are fixed, one might be able to predict how synaptic plasticity changes. Overall, our work not only provided a theoretical framework, but also provided a set of testable hypotheses on neural recordings and behaviors in decision-making under ambiguity. 6.3 Limitations and extensions When δ is a constant, our lower bound is order-optimal in terms of its scaling in n, k, and TR. Nevertheless, the scaling of the derived lower bound in terms of δ is not tight. It would be interesting to know the optimal scaling in δ when other parameters (n, k, and TR) are fixed. We leave it as one future direction. To simplify complexity, our model posed a few assumptions that ignored some features in the brain. One of these assumptions is that each input neuron is independent. However, various degrees of average noise correlations between cortical neurons have been reported. For example, average noise correlations in primary visual cortex could be close to 0.1 [SSB+15], 0.18 [SK08], or even much larger as 0.35 [GD08]. Similarly, noise correlations have been observed in other sensory brain regions [CK11]. In our work, we ignored correlations between these neurons, but it would be interesting as a future direction to extend in our spiking network model. Second, our model used a threshold activation function by assuming the synaptic transmission is basically noise-free and that the only noise source comes from the input in this paper. However, synaptic transmission is highly unreliable in biological networks [AS94, FSW08, Bor10], and a deterministic activation function would fail to capture this feature compared to a stochastic activation function. Moreover, failure in synaptic transmission could serve a computational role [BS09, Maa97]. Another assumption in our circuit is that the output neurons can inhibit each other. In common scenar- ios, an output neuron is usually excitatory, and does not inhibit other neurons directly without recruiting inhibitory cells. We incorporate stability in these output neurons by assuming they can inhibit each other in our circuit implementation. For a model where an output neuron is limited to be excitatory only, we can add a chain of inhibitory neurons to achieve stability WTA computation. Last but not least, in our k-WTA circuit, the number of output neurons that spike simultaneously increases monotonically until there are exactly k output neurons that spike simultaneously. We acknowledge that this might not be biologically plausible in most cases in the brain. From large-scale neural recordings, we know that the number of neurons that spike simultaneously is usually variable, so this could be a future direction to construct a circuit that better matches experimental observations. Acknowledgement We would like to thank Christopher Quinn at Purdue University and Zhi-Hong Mao at University of Pitts- burgh for the helpful discussions and references. References [AS94] Christina Allen and Charles F Stevens. An evaluation of causes for unreliability of synaptic transmission. Proceedings of the National Academy of Sciences, 91(22):10380 -- 10383, 1994. 14 [BAB+97] Roland Baddeley, Larry F Abbott, Michael CA Booth, Frank Sengpiel, Tobe Freeman, Ed- ward A Wakeman, and Edmund T Rolls. Responses of neurons in primary and inferior tempo- ral visual cortices to natural scenes. Proceedings of the Royal Society of London B: Biological Sciences, 264(1389):1775 -- 1783, 1997. 2, 13 14 [BC95] Gyorgy Buzs´aki and James J Chrobak. Temporal structure in spatially organized neuronal ensembles: a role for interneuronal networks. Current opinion in neurobiology, 5(4):504 -- 510, 1995. 2 [BIM98] Michael J Berry II and Markus Meister. Refractoriness and neural precision. In Advances in Neural Information Processing Systems, pages 110 -- 116, 1998. 2 [BMG+17] Katie C Bittner, Aaron D Milstein, Christine Grienberger, Sandro Romani, and Jeffrey C Science, Behavioral time scale synaptic plasticity underlies ca1 place fields. Magee. 357(6355):1033 -- 1036, 2017. 14 [Bor10] [BS09] [CK11] J Gerard G Borst. The low synaptic release probability in vivo. Trends in neurosciences, 33(6):259 -- 266, 2010. 14 Tiago Branco and Kevin Staras. The probability of neurotransmitter release: variability and feedback control at single synapses. Nature Reviews Neuroscience, 10(5):373, 2009. 14 Marlene R Cohen and Adam Kohn. Measuring and interpreting neuronal correlations. Nature neuroscience, 14(7):811, 2011. 14 [CMA+03] Luke Clark, Facundo Manes, Nagui Antoun, Barbara J Sahakian, and Trevor W Robbins. The contributions of lesion laterality and lesion volume to decision-making impairment following frontal lobe damage. Neuropsychologia, 41(11):1474 -- 1483, 2003. 13 [DA01] [EB02] Peter Dayan and Laurence F Abbott. Theoretical neuroscience: computational and mathemat- ical modeling of neural systems. 2001. 13 Marc O Ernst and Martin S Banks. Humans integrate visual and haptic information in a statistically optimal fashion. Nature, 415(6870):429, 2002. 2 [FDMM18] Ulisse Ferrari, Stephane Deny, Olivier Marre, and Thierry Mora. A simple model for low variability in neural spike trains. arXiv preprint arXiv:1801.01362, 2018. 2, 13 [FSW08] A Aldo Faisal, Luc PJ Selen, and Daniel M Wolpert. Noise in the nervous system. Nature reviews neuroscience, 9(4):292, 2008. 2, 14 [GD08] Diego A Gutnisky and Valentin Dragoi. Adaptive coding of visual information in neural pop- ulations. Nature, 452(7184):220, 2008. 14 [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. 2 [HDS06] [HDZ08] [HSM06] [HW59] [HW98] Timothy D Hanks, Jochen Ditterich, and Michael N Shadlen. Microstimulation of macaque area lip affects decision-making in a motion discrimination task. Nature neuroscience, 9(5):682, 2006. 13 Tom´as Hrom´adka, Michael R DeWeese, and Anthony M Zador. Sparse representation of sounds in the unanesthetized auditory cortex. PLoS biology, 6(1):e16, 2008. 2 Thomas TG Hahn, Bert Sakmann, and Mayank R Mehta. Phase-locking of hippocam- pal interneurons' membrane potential to neocortical up-down states. Nature neuroscience, 9(11):1359, 2006. 2 David H Hubel and Torsten N Wiesel. Receptive fields of single neurones in the cat's striate cortex. The Journal of physiology, 148(3):574 -- 591, 1959. 2 Christopher M Harris and Daniel M Wolpert. Signal-dependent noise determines motor plan- ning. Nature, 394(6695):780, 1998. 2 15 [IKN98] [JB67] [JM06] [KCF17] [KF08] [KK01] [KP04] [KPS10] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254 -- 1259, 1998. 2 I Jacobs and E Berlekamp. A lower bound to the distribution of computation for sequential decoding. IEEE Transactions on Information Theory, 13(2):167 -- 174, 1967. 21 Mehrdad Jazayeri and J Anthony Movshon. Optimal representation of sensory information by neural populations. Nature neuroscience, 9(5):690, 2006. 13 Birgit Kriener, Rishidev Chaudhuri, and Ila Fiete. How fast is neural winner-take-all when deciding between many options? bioRxiv, page 231753, 2017. 2, 9, 13 Mattias P Karlsson and Loren M Frank. Network dynamics underlying the formation of sparse, informative representations in the hippocampus. Journal of Neuroscience, 28(52):14271 -- 14281, 2008. 2 Masaharu Kinoshita and Hidehiko Komatsu. Neural representation of the luminance and bright- ness of a uniform surface in the macaque primary visual cortex. Journal of neurophysiology, 86(5):2559 -- 2570, 2001. 2 David C Knill and Alexandre Pouget. The bayesian brain: the role of uncertainty in neural coding and computation. TRENDS in Neurosciences, 27(12):712 -- 719, 2004. 2 Andreas Knoblauch, Gunther Palm, and Friedrich T Sommer. Memory capacities for synaptic and structural plasticity. Neural Computation, 22(2):289 -- 341, 2010. 14 [KRR00] 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. 2, 13 [KTA+03] Zoe Kourtzi, Andreas S Tolias, Christian F Altmann, Mark Augath, and Nikos K Logothetis. Integration of local features into global shapes: monkey and human fmri studies. Neuron, 37(2):333 -- 346, 2003. 2 [KW04] Konrad P Kording and Daniel M Wolpert. Bayesian integration in sensorimotor learning. Nature, 427(6971):244, 2004. 2 [KYPH16] Leor N Katz, Jacob L Yates, Jonathan W Pillow, and Alexander C Huk. Dissociated functional significance of decision-related activity in the primate dorsal stream. Nature, 535(7611):285, 2016. 13 [LCG+15] Nuo Li, Tsai-Wen Chen, Zengcai V Guo, Charles R Gerfen, and Karel Svoboda. A motor cortex circuit for motor planning and movement. Nature, 519(7541):51, 2015. 2 [LIKB99] Dale K Lee, Laurent Itti, Christof Koch, and Jochen Braun. Attention activates winner-take-all competition among visual filters. Nature neuroscience, 2(4):375, 1999. 2, 13 [LLW13] Shuai Li, Yangming Li, and Zheng Wang. A class of finite-time dual neural networks for solv- ing quadratic programming problems and its k-winners-take-all application. Neural Networks, 39:27 -- 39, 2013. 2, 13 [LMP16] Nancy Lynch, Cameron Musco, and Merav Parter. Computational tradeoffs in biological neural networks: Self-stabilizing winner-take-all networks. arXiv preprint arXiv:1610.02084, 2016. 5 [MA09] [Maa97] Gaby Maimon and John A Assad. Beyond poisson: primate parietal cortex. Neuron, 62(3):426 -- 440, 2009. 2, 13 increased spike-time regularity across Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10(9):1659 -- 1671, 1997. 5, 14 16 [Maa00] Wolfgang Maass. On the computational power of winner-take-all. Neural computation, 12(11):2519 -- 2535, 2000. 2 [MCM07] Najib J Majaj, Matteo Carandini, and J Anthony Movshon. Motion integration by neurons in macaque mt is local, not global. Journal of Neuroscience, 27(2):366 -- 370, 2007. 2 [MM07] [Nel04] [OF04] [PG99] Zhi-Hong Mao and Steve G Massaquoi. Dynamics of winner-take-all competition in recurrent neural networks with lateral inhibition. IEEE transactions on neural networks, 18(1):55 -- 69, 2007. 9 Israel Nelken. Processing of complex stimuli and natural scenes in the auditory cortex. Current opinion in neurobiology, 14(4):474 -- 480, 2004. 2 Bruno A Olshausen and David J Field. Sparse coding of sensory inputs. Current opinion in neurobiology, 14(4):481 -- 487, 2004. 2 Michael L Platt and Paul W Glimcher. Neural correlates of decision variables in parietal cortex. Nature, 400(6741):233, 1999. 2 [POMT+02] Javier Perez-Orive, Ofer Mazor, Glenn C Turner, Stijn Cassenaer, Rachel I Wilson, and Gilles Laurent. Oscillations and sparsening of odor representations in the mushroom body. Science, 297(5580):359 -- 365, 2002. 2 [PW14] Yury Polyanskiy and Yihong Wu. Lecture notes on information theory. Lecture Notes for ECE563 (UIUC) and, 6:2012 -- 2016, 2014. 7, 18, 20 [QKKF08] R Quian Quiroga, Gabriel Kreiman, Christof Koch, and Itzhak Fried. Sparse but not ?grandmother-cell?coding in the medial temporal lobe. Trends in cognitive sciences, 12(3):87 -- 91, 2008. 2 [RP99] Maximilian Riesenhuber and Tomaso Poggio. Hierarchical models of object recognition in cortex. Nature neuroscience, 2(11):1019, 1999. 2 [RPG99] Peter Redgrave, Tony J Prescott, and Kevin Gurney. The basal ganglia: a vertebrate solution to the selection problem? Neuroscience, 89(4):1009 -- 1023, 1999. 2 [RV06] [SK08] Nicolas P Rougier and Julien Vitay. Emergence of attention within a neural population. Neural Networks, 19(5):573 -- 581, 2006. 2, 13 Matthew A Smith and Adam Kohn. Spatial and temporal scales of neuronal correlation in primary visual cortex. Journal of Neuroscience, 28(48):12591 -- 12603, 2008. 14 [SLW05] Athanassios G Siapas, Evgueniy V Lubenov, and Matthew A Wilson. Prefrontal phase locking to hippocampal theta oscillations. Neuron, 46(1):141 -- 151, 2005. 2 [SN96] [SN01] [SS06] Michael N Shadlen and William T Newsome. Motion perception: seeing and deciding. Pro- ceedings of the national academy of sciences, 93(2):628 -- 633, 1996. 13 Michael N Shadlen and William T Newsome. Neural basis of a perceptual decision in the parietal cortex (area lip) of the rhesus monkey. Journal of neurophysiology, 86(4):1916 -- 1936, 2001. 2, 13 Alan A Stocker and Eero P Simoncelli. Noise characteristics and prior expectations in human visual speed perception. Nature neuroscience, 9(4):578, 2006. 2 [SSB+15] Marieke L Scholvinck, Aman B Saleem, Andrea Benucci, Kenneth D Harris, and Matteo Caran- dini. Cortical state determines global variability and correlations in visual cortex. Journal of Neuroscience, 35(1):170 -- 178, 2015. 14 17 [VRP+18] [wik] [Wu17] Stephen J Verzi, Fredrick Rothganger, Ojas D Parekh, Tu-Thach Quach, Nadine E Miner, Craig M Vineyard, Conrad D James, and James B Aimone. Computing with spikes: The advantage of fine-grained timing. Neural computation, 30(10):2660 -- 2690, 2018. 5 Activation function. https://en.wikipedia.org/wiki/Activation function. Accessed: 2018-08-08. 9 Yihong Wu. Lecture notes on information-theoretic methods for high-dimensional statistics. Lecture Notes for ECE598YW (UIUC), 2017. 5 [YG98] AL Yuille and D Geiger. The handbook of brain theory and neural networks, 1998. 2 [YLS13] Eric A Yttri, Yuqing Liu, and Lawrence H Snyder. Lesions of cortical area lip affect reach onset only when the reach is accompanied by a saccade, revealing an active eye -- hand coordination circuit. Proceedings of the National Academy of Sciences, 110(6):2371 -- 2376, 2013. 13 Appendices A Preliminaries In this section, we present some preliminaries on information measures and Fano's inequality. Interested readers are referred to [PW14] for comprehensive background. A.1 Information Measures Let X and Y be two random variables. The mutual information between X and Y , denoted by I(X; Y ), measures the dependence between X and Y , or, the information about X (resp. T ) provided by Y (resp. X). Definition 14 (Mutual information). Let X and Y be two random variables. I(X; Y ) := D(PXY (cid:107) PX PY ), D(P (cid:107) Q) := P (a) log P (a) Q(a) , (cid:88) a∈A where PXY denotes the joint distribution of X and Y , and PX PY denotes the product of the marginal distributions of X and Y . In the following, we use the notation X → Y to denote that Y is a (possibly random) function of X. Thus, W → X → Y → (cid:99)W means that X is a (possibly random) function of W ; Y is a (possibly random) function of X; and (cid:99)W is a (possibly random) function of Y . Fano's inequality: Theorem 15. [PW14, Corollary 5.1] Let T : Θ → [M ], and let θ → X → Y → (cid:98)T (θ) be an arbitrary Markov (cid:111) ≥ 1 − I(X; Y ) + 1 chain. Suppose both θ and T (θ) are uniformly distributed over a set of size M . Then Pe := P(cid:110) T (θ) (cid:54)= (cid:98)T (θ) . log M Theorem 16 (Chernoff Bound). Let X1,··· , Xn be i.i.d. with Xi ∈ {0, 1} and P{X1 = 1} = p. Set X =(cid:80)n i=1 Xi. Then • for any t ∈ [0, 1 − p], we have P{X ≥ (p + t) n} ≤ exp (−nd(p + t (cid:107) p)). • for any t ∈ [0, p], we have P{X ≤ (p − t) n} ≤ exp (−nd(p − t (cid:107) p)). 18 Proof of Lemma 2. Lemma 2 follows easily from the independence between input spike trains and the as- sumption that the spikes in each input spike train are i.i.d.. For completeness, we present the proof as follows. B Proof of Lemma 2 Recall that S := Let s = [s1,··· , sn] such that each component si is a binary sequence of length T , i.e., (cid:105) (cid:104){St(u1)}T si =(cid:2)bi t=1 ,··· ,{St(un)}T t=1 . (cid:3) ∈ {0, 1}T . 1,··· , bi T For each i = 1,··· , n, let PS({St(ui)}T t=1) be the marginal distributions of {St(ui)}T under joint distributions PS and QS respectively. Similarly, PS(St(ui)) and QS(St(ui)) are the corresponding two marginal distributions of St(ui). Thus, we have t=1) and QS({St(ui)}T t=1 where (cid:80) where we use (cid:80) entry fixed. 1,··· ,bi [bi T ] is the summation over all binary sequences of length T . In the last displayed equation, equality (a) follows from the definition of KL divergence; equality (b) is true because of independence of spikes; equality (c) follows from the fact that for any fixed bi t, (cid:88)  T(cid:89) 1,··· ,bi [bi T ]\{t} t(cid:48)=1&t(cid:48)(cid:54)=t  = 1, PS(St(cid:48)(ui) = bi t(cid:48)) 1,··· ,bi [bi T ]\{t} to denote the summation over all binary sequences of length T with the t -- th 19 t=1 =(cid:2)bi t=1 =(cid:2)bi 1,··· , bi 1,··· , bi T T PS({St(ui)}T QS({St(ui)}T (cid:3)) (cid:3)) t=1 PS(St(ui) = bi t) t=1 QS(St(ui) = bi t) PS(St(ui) = bi t) QS(St(ui) = bi t) PS(St(ui) = bi t) QS(St(ui) = bi t)  PS(St(ui) = bt) log PS(St(ui) = bi t) QS(St(ui) = bi t) T t=1 =(cid:2)bi (cid:17) t=1) (cid:107) QS({St(ui)}T (cid:3)) log t=1) 1,··· , bi (cid:33) (cid:81)T (cid:81)T (cid:33) T(cid:88) (cid:33) PS(St(cid:48)(ui) = bi PS(St(cid:48)(ui) = bi log log t(cid:48)) t(cid:48)) t=1 PS(St(cid:48)(ui) = bi t(cid:48)) log (cid:16) D (a) = PS({St(ui)}T t(cid:48)=0 t(cid:48)=0 1,··· ,bi T ] [bi 1,··· ,bi [bi T ] 1,··· ,bi T ] [bi PS({St(ui)}T (cid:88) (cid:32)T−1(cid:89) (cid:88) (cid:32)T−1(cid:89) (cid:88) (cid:32)T−1(cid:89) (cid:88) T(cid:88)  T−1(cid:89) (cid:88) T(cid:88) T(cid:88) (cid:88) (cid:18) T(cid:88) T(cid:88) 1,··· ,bi T ] [bi 1,··· ,bi T ] [bi pi log pi qi t(cid:48)=0 t=1 t=1 t=1 t=1 bi t t=1 (b) = = = = (c) = = = PS(St(cid:48)(ui) = bi t(cid:48)) t(cid:48)=0&t(cid:48)(cid:54)=t PS(St(ui) = bi t) log + (1 − pi) log PS(St(ui) = bi t) QS(St(ui) = bi t) 1 − pi 1 − qi (cid:19) d(pi (cid:107) qi) = T · d(pi (cid:107) qi). Similarly, we get D(PS (cid:107) QS) = = = n(cid:88) n(cid:88) i=1 (cid:88) (cid:16) D PS(S = s) log PS(S = s) QS(S = s) s=[s1,··· ,sn] PS({St(ui)}T T d(pi (cid:107) qi) = T t=1) (cid:107) QS({St(ui)}T n(cid:88) d(pi (cid:107) qi), t=1) (cid:17) i=1 i=1 proving the lemma. C Proof of Lemma 3 Proof of Lemma 3. Since mutual information can be viewed as distance to product distributions, by [PW14, Theorem 3.4], we have D(cid:0)PXp,S (cid:107) QXpQS (cid:1) . I(Xp; S) = min QXp QS where PXp,S is the joint distribution of Xp and S, and QXp and QS are any distributions of Xp and S, respectively. For any fixed QS, it holds that D(cid:0)PXp,S (cid:107) QXp QS min QXp (cid:1) = min D(cid:0)PSXpPXp (cid:107) QXp QS (cid:1) , ≤ D(cid:0)PSXpPXp (cid:107) PXpQS QXp (cid:1) where the equality follows from conditioning, and the inequality is true because the best choice over all QXp cannot be worse than any specific choice of QXp . Here S Xp denotes the n input spike trains conditioning on the choice of rate assignment. For any fixed QS, we have (cid:88) (cid:1) = PXp(Xp = p0) D(cid:0)PSXpPXp (cid:107) PXpQS (cid:88) n(cid:88) k(cid:88) (cid:88) PXp (Xp = pij) j=k+1 i=1 + 1 s s = k(n − k) + 1 PSXp=p0(S = s) log PSXp=pij (S = s) log (cid:20) (cid:34) (cid:20) 1 + k(n − k) + 1 PSXp=p0(S = s) log s k(cid:88) (cid:88) n(cid:88) D(cid:0)PSXp=p0 (cid:107) QS j=k+1 i=1 s (cid:1) + = 1 k(n − k) + 1 1 k(n − k) + 1 PSXp=pij (S = s) log (cid:21) (cid:35) PSXp=p0(S = s)PXp(Xp = p0) QS(S = s)PXp(Xp = p0) PSXp=pij (S = s)PXp (Xp = pij) QS(S = s)PXp(Xp = pij) (cid:21) PSXp=p0(S = s) QS(S = s) (cid:20) k(cid:88) PSXp=pij (S = s) (cid:21) D(cid:0)PSXp=pij (cid:107) QS QS(S = s) n(cid:88) (cid:1) , i=1 j=k+1 where(cid:80) s is summation over all possible n binary sequences of length T . Here PSXp=p0 is the distribution of S with the rate assignment p0, and PSXp=pij is the distribution of S with the rate assignment pij. Choosing QS to be the distribution of S with rate assignment p0 defined in (8), then for any i = 1,··· , k and j = k + 1,··· , n, we have (cid:16) D PSXpij (cid:107) QS (cid:17) = T (d(r2 (cid:107) r1) + d(r1 (cid:107) r2)). 20 Therefore, I (Xp (cid:107) S) ≤ 1 k(n − k) + 1 k(cid:88) n(cid:88) T (d(r2 (cid:107) r1) + d(r1 (cid:107) r2)) ≤ T (d(r2 (cid:107) r1) + d(r1 (cid:107) r2)). i=1 j=k+1 D Proof of Lemma 13 As all input neurons are quiescent at time t0 and remain to be quiescent for all t ≥ t0, it follows that (cid:16) 1{Vt0+m−r>0} − m1{Vt0+m−r≤−1} (cid:17) > b; By the activation rules in Algorithm 1, we know that (cid:40) St0+m = 1, if 0, otherwise. (b − 1)1{St0+m−1(vi)=1} +(cid:80)m (cid:16) m(cid:88) r=1 (b − 1)1{St0+m−1(vi)=1} + = (b − 1)1{St0+m−1(vi)=1} − m ≤ b − 1 < b. r=1 (cid:17) 1{Vt0+m−r>0} − m1{Vt0+m−r≤−1} m(cid:88) 1{Vt0+m−r≤−1} r=1 Thus, St0+m(vi) = 0 for all i = 1,··· , n. So we have Vt0+m+1(vi) = 0 for all i = 1,··· , n, which again implies that St0+m+1(vi) = 0 for all i = 1,··· , n. Therefore, we conclude that St(vi) = 0 and Vt(vi) = 0 for all t > t0 + m. E Proof of Theorem 4 Proof of Theorem 4. We prove this via a genie-aided argument [JB67] by assuming that there is a genie that can access the firing sequences of all the n input neurons. By assuming the existence of a genie, we are essentially considering the centralized setting. Clearly, if the error probability is high even in the centralized setting, then no SNNs (which are distributed algorithms) can achieve lower error probability. Suppose that T ≤ ((1 − δ) log(k(n − k) + 1) − 1) TR. By (7) there exists r1, r2 such that r1 (cid:54)= r2 and T ≤ ((1 − δ) log(k(n − k) + 1) − 1) 1 d(r2 (cid:107) r1) + d(r1 (cid:107) r2) . Without loss of generality, assume that r1 > r2. Consider the k(n − k) + 1 possible rate assignments defined in Lemma 3. Let P be the set of such rate assignments. By Yao's minimax principle, we know the minimax probability of error is always lower bounded by Bayes probability of error with any prior distribution: where Xp ∼ U nif (P) is uniformly distributed over set P. In addition, by Fano's inequality, we have max p∈ARk P(cid:110)(cid:100)win (S) (cid:54)= W(p) (cid:111) ≥ EXp∼U nif (P) (cid:104)P(cid:110)(cid:100)win (S) (cid:54)= W(Xp) (cid:104)P(cid:110)(cid:100)win (S) (cid:54)= W(Xp) (cid:111)(cid:105) (cid:111)(cid:105) ≥ 1 − I(Xp; S) + 1 . log(k(n − k) + 1) , EXp∼U nif (P) (11) 21 Applying Lemma 3, we get max p∈ARk P(cid:110)(cid:100)win (S) (cid:54)= W(p) (cid:111) ≥ 1 − I(Xp; S) + 1 ≥ 1 − T (d(r2 (cid:107) r1) + d(r1 (cid:107) r2)) + 1 ≥ δ. log(k(n − k) + 1) log(k(n − k) + 1) The last inequality holds as T ≤ ((1 − δ) log(k(n − k) + 1) − 1) TR. F Proof of Theorem 10 The proof of Theorem 10 uses the following technical fact and lemma. Fact 17. For any given p ∈ (0, 1) and b > 0, let fp,b : R → R, defined as: for all t > 0, (cid:18) −td (cid:18) b t (cid:19)(cid:19) (cid:107) p . fp,b(t) := exp Function fp,b(·) is increasing when t ∈ (0, b p ) and decreasing when t ≥ b p . This fact follows immediately from a simple algebra. Lemma 18. Assume u, v ∈ [c, C] ⊆ (0, 1). Then for any α ∈ (0, 1), d ((1 − α)u + αv (cid:107) u) ≥ α2c(1 − C) 2C(1 − c) (d (u (cid:107) v) + d (v (cid:107) u)) . Proof. Note that for any fixed q ∈ [c, C], d (x (cid:107) q) is a function of x, where x ∈ [c, C]. In addition, by simple algebra, we have d(cid:48) (x (cid:107) q) = log (1 − q)x q(1 − x) , and d(cid:48)(cid:48) (x (cid:107) q) = 1 x(1 − x) . (12) By Taylor expansion, we have d ((1 − α)u + αv (cid:107) u) = d (u (cid:107) u) + ((1 − α)u + αv − u) d(cid:48) (u (cid:107) u) ((1 − α)u + αv − u)2 + d(cid:48)(cid:48) (ξ (cid:107) u) , 2 where ξ ∈ [min{u, (1 − α)u + αv}, max{u, (1 − α)u + αv}]. By (12), d ((1 − α)u + αv (cid:107) u) = 0 + 0 + 1 ξ(1 − ξ) α2(u − v)2 2 ≥ α2(u − v)2 2C(1 − c) . On the other hand, since d (u (cid:107) v) + d (u (cid:107) v) is symmetric in u and v, without loss of generality, assume that u ≥ v. We have d (u (cid:107) v) + d (u (cid:107) v) = (u − v) log u(1 − v) v(1 − u) (cid:18) u − v v(1 − u) (cid:19) = (u − v) log 1 + u − v v(1 − u) (u − v)2 v(1 − u) ≤ (u − v) ≤ 2C(1 − c) c(1 − C)α2 d ((1 − α)u + αv (cid:107) u) , = ≤ (u − v)2 c(1 − C) proving the lemma. 22 Now we are ready to prove Theorem 10. Proof of Theorem 10. Without loss of generality, assume that p1 ≥ ··· ≥ pk > pk+1 ≥ ··· ≥ pn. For a given rate assignment p ∈ AR, define τ1, τ2,··· , τn as t : min{t,m∗}(cid:88) r=0 τi := inf t  , ∀ i = 1,··· , n. Sr(ui) ≥ b To show Theorem 10, it is enough to show that with probability 1 − δ, τi < τj ∀ i = 1,··· , k, and j = k + 1,··· , n; (13) and τi ≤ m∗ ∀ i = 1,··· , k.. (14) Before diving into proving (13) and (14) hold with probability at least 1− δ, let's check the sufficiency of (13) and (14). Let t0 := max1≤i≤k τi. Let E be the event on which (13) and (14) hold. Clearly, conditioning on event E, we have (cid:18) (cid:19) max 1≤i≤k τi E = t0 E ≤ m∗ − 1 ≤ m − 1, and (cid:18) (cid:19) t(cid:88) max 1≤i≤k τi E = t0 E < τj E ∀j = k + 1,··· , n. Notably, for any t ≤ t0 ≤ m − 1 and for i = 1,··· , n, 1{Vr(vi)>0} − m 1{Vr(vi)≤−1} (cid:35) ≤ t(cid:88) 1{Vr(vi)>0} ≤ t(cid:88) Sr(ui). (cid:34) t(cid:88) r=1 r=1 + r=1 r=1 Thus, conditioning on E, at most k−1 output neurons ever spike by time t0. So we have (1) 1{Vt(vi)≤−1} = 0, and (2) 1{Vt(vi)>0} = St(ui), for all i = 1,··· , n and for all t ≤ t0. In addition, we have for all t ≤ t0, (cid:35) 1{Vr(vi)≤−1} r=1 + (b − 1)1{St(vi)=1} + = (b − 1)1{St(vi)=1} + = (b − 1)1{St(vi)=1} + t(cid:88) (cid:34) t(cid:88) 1{Vr(vi)>0} − m t(cid:88) t(cid:88) 1{Vr(vi)>0} r=1 r=1 Sr(ui). r=1 By the activation rules in Algorithm 1, we know, conditioning on E, at time t0 + 1 ≤ m∗, output neurons v1,··· , vk spike simultaneously, and output neurons vk+1,··· , vn do not spike, proving (1) in Theorem 10. By the choice of t0, we know that, on E, t0 + 1 is the first time that k output neurons spike simultaneously, and no other k output neurons ever spike simultaneously, proving (2) in Theorem 10. By a simple induction argument, it can be shown that conditioning on E, in each of the time slot t such that t0 + 1 ≤ t ≤ m + 1, output neurons v1,··· , vk spike, and no other output neurons (i.e., output neurons vk+1,··· , vn do not spike). Let's consider the case when t = (m + 1) + 1. As among output neurons, only v1,··· , vk spike, and no other output neurons spike for any t(cid:48) ≤ m + 1, it follows that m 1{Vt−r(vi)≤−1} = 0, ∀ v1,··· , vk. m(cid:88) r=1 23 Thus, for these k output neurons, (cid:34) m(cid:88) r=1 1{Vt−r(vi)>0} − m m(cid:88) r=1 1{Vt−r(vi)≤−1} (cid:35) + 1{Vt−r(vi)>0} 1{Vt−1−r(vi)>0} + 1{Vt−1(vi)>0} − 1{Vt−1−m(vi)>0} (b − 1)1{St−1(vi)=1} + = (b − 1) + = (b − 1) + ≥ b − 2 + = b − 2 + r=1 m(cid:88) m(cid:88) m(cid:88) m(cid:88) r=1 r=1 r=1 1{Vt−1−r(vi)>0} Sr(ui) ≥ 2b − 2 ≥ b, (b − 1)1{St−1(vi)=1} + 1{Vt−r(vi)>0} − m 1{Vt−r(vi)≤−1} r=1 r=1 + m(cid:88) (cid:35) where the last inequality holds as long as b ≥ 2. For output neurons vk+1,··· , vn, we have (cid:34) m(cid:88) m(cid:88) r=1 (a) = ≤ m(cid:88) m(cid:88) m(cid:88) ≤ m(cid:88) r=1 r=1 = 1{Vt−r(vi)>0} 1{Vt−1−r(vi)>0} + 1{Vt−1(vi)>0} − 1{Vt−1−m(vi)>0} 1{Vt−1−r(vi)>0} − 1{Vt−1−m(vi)>0} 1{Vt−1−r(vi)>0} = 1{Vr(vi)>0} < b. r=1 r=1 Equality (a) follows because at time t − 1, output neurons v1,··· , vk spike, resulting in 1{Vt−1−r(vi)>0} = 0 for i (cid:54)= 1,··· , k. Thus, we know conditioning on event E, at time (m + 1) + 1, the output neurons v1,··· , vk spike, and no other output neuron spike. It can be shown by a simple induction that at each time t such that t0 + 1 ≤ t ≤ m + b, the output neurons v1,··· , vk spike, and no other output neurons spike. This proves (3) in Theorem 10. Next we prove (13) and (14). By definition of τj, we know that τj ≤ m∗ for all j = 1,··· , n. Thus, we only need to show that with probability 1 − δ, τi < τj ∀ i = 1,··· , k, and j = k + 1,··· , n, which is the focus of the remainder of our proof. Note that P{τi < τj, ∀i ∈ {1,··· , k},∀j ∈ {k + 1,··· , n}} = P{τi < τj, & τi < m∗, ∀i ∈ {1,··· , k},∀j ∈ {k + 1,··· , n}} ≥ 1 − k(cid:88) n(cid:88) P{τi ≥ τj, or τi = m∗} . For each term in the summation of (15), we have i=1 j=k+1 P{τi ≥ τj, or τi = m∗} = P{τi = m∗} + P{τi ≥ τj, & τi < m∗} , 24 (15) (16) which follows from the fact that P{A ∪ B} = P{A} + P{B − A} for any sets A and B. Note that m∗pi ≥ b. By Chernoff bound, the first term in (16) is bounded as (cid:41) (cid:18) Sr(ui) ≤ b ≤ exp −m∗ · d P{τi = m∗} = P For the second term in (16), we have (cid:19)(cid:19) (cid:18) b m∗ (cid:107) pi (cid:41) (cid:18) −t∗ · d . (17) (cid:19)(cid:19) (cid:18) b t∗ (cid:107) pk , P{τi ≥ τj and τi < m∗} = P Sr(uj) ≥ b, and τi < m∗ ≤ exp −t∗ · d + exp where t∗ ∈(cid:16) b pk+1 , b pk (cid:17) . Thus, (16) is upper bounded as P{τi ≥ τj, or τi = m∗} ≤ exp −m∗ · d + exp −t∗ · d ≤ exp Eq (15) is bounded as −t∗ · d (cid:19)(cid:19) (cid:18) b t∗ (cid:107) pk+1 (cid:19)(cid:19) (cid:18) b (cid:18) b (cid:19)(cid:19) m∗ (cid:107) pk+1 (cid:18) b (cid:19)(cid:19) t∗ (cid:107) pk+1 t∗ (cid:107) pk+1 (cid:18) (cid:18) + exp −t∗ · d + 2 exp −t∗ · d (cid:18) b (cid:19)(cid:19) (cid:18) b (cid:19)(cid:19) t∗ (cid:107) pk t∗ (cid:107) pk . r=0 (cid:40) m∗(cid:88) (cid:40) τi(cid:88) (cid:18) r=0 (cid:18) (cid:18) (cid:18) P{τi < τj, ∀i ∈ {1,··· , k},∀j ∈ {k + 1,··· , n}} P{τi ≥ τj, or τi = m∗} ≥ 1 − k(cid:88) ≥ 1 − k(cid:88) i=1 n(cid:88) n(cid:88) j=k+1 i=1 j=k+1 (cid:18) (cid:18) (cid:18) (cid:18) = 1 − k(n − k) exp −t∗ · d exp −t∗ · d (cid:18) b (cid:19)(cid:19) t∗ (cid:107) pk+1 (cid:18) b (cid:19)(cid:19) t∗ (cid:107) pk+1 (cid:18) (cid:18) − b + 2 exp + 2 exp (cid:19)(cid:19)(cid:19) (cid:19)(cid:19)(cid:19) . −t∗ · d −t∗ · d (cid:18) b (cid:18) t∗ (cid:107) pk (cid:18) b (cid:18) t∗ (cid:107) pk (cid:18) pk + pk+1 (cid:18) pk + pk+1 2 · d (cid:107) pk+1 (cid:107) pk (cid:19)(cid:19) (cid:19)(cid:19) , 2 · d = exp = 2 exp (pk + pk+1)/2 − b (pk + pk+1)/2 ≥ c(1 − C) 8C(1 − c) ≥ c(1 − C) 8C(1 − c) (d(pk+1 (cid:107) pk) + d(pk (cid:107) pk+1)) , (d(pk+1 (cid:107) pk) + d(pk (cid:107) pk+1)) . . (cid:19) Let t∗ = b exp 2 exp (cid:18) −t∗ · d −t∗ · d (cid:18) b (pk+pk+1)/2 , it holds that (cid:18) b (cid:18) t∗ (cid:107) pk+1 t∗ (cid:107) pk (cid:18) pk + pk+1 (cid:18) pk + pk+1 d 2 (cid:19)(cid:19) (cid:19)(cid:19) (cid:19) (cid:19) (cid:107) pk+1 (cid:107) pk d 2 By Lemma 18, we know and, Thus, we get P{τi < τj, ∀i ∈ {1,··· , k},∀j ∈ {k + 1,··· , n}} ≥ 1 − 3k(n − k) exp − 2b c(1 − C) 8C(1 − c) pk + pk+1 (cid:18) (d(pk (cid:107) pk+1) + d(pk+1 (cid:107) pk)) 25 Since b = 8C2(1−c) c(1−C) (cid:0)log 3 δ + log k(n − k)(cid:1) TR, we have (cid:18) 3k(n − k) exp − 2b pk + pk+1 c(1 − C) 8C(1 − c) (d(pk (cid:107) pk+1) + d(pk+1 (cid:107) pk)) (cid:19) ≤ δ. Thus, P{τi < τj, ∀i ∈ {1,··· , k},∀j ∈ {k + 1,··· , n}} ≤ 1 − δ. In addition, t∗ = 2b pk + pk+1 b = m∗ ≤ m, ≤ 1 c completing the proof of Theorem 10. 26
1209.6607
1
1209
2012-09-28T18:59:30
Evidence for an additive inhibitory component of contrast adaptation
[ "q-bio.NC" ]
The latency of visual responses generally decreases as contrast increases. Recording in the lateral geniculate nucleus (LGN), we find that response latency increases with increasing contrast in ON cells for some visual stimuli. We propose that this surprising latency trend can be explained if ON cells rest further from threshold at higher contrasts. Indeed, while contrast changes caused a combination of multiplicative gain change and additive shift in LGN cells, the additive shift predominated in ON cells. Modeling results supported this theory: the ON cell latency trend was found when the distance-to-threshold shifted with contrast, but not when distance-to-threshold was fixed across contrasts. In the model, latency also increases as surround-to-center ratios increase, which has been shown to occur at higher contrasts. We propose that higher-contrast full-field stimuli can evoke more surround inhibition, shifting the potential further from spiking threshold and thereby increasing response latency.
q-bio.NC
q-bio
DRAFT 5/19/09 Gaudry and Reinagel Evidence for an additive inhibitory component of contrast adaptation Kate S. Gaudry, PhD and Pamela Reinagel*, PhD. *to whom correspondence should be addressed: [email protected]; Section of Neurobiology, Division of Biological Sciences, University of California San Diego, La Jolla CA 92093. Summary The latency o f visual responses generally decreases as contrast increases. Recording in the lateral geniculate nucleus (LGN), we find that response latency increases with increasing contrast in ON cells for some visual stimuli. We propose that this surprising latency trend can be explained if ON cells rest further from thresho ld at higher contrasts. Indeed, while contrast changes caused a combination of mult iplicat ive gain change and addit ive shift in LGN cells, the addit ive shift predominated in ON cells. Modeling result s supported this theory: the ON cell latency trend was found when the distance -to-thresho ld shifted with contrast, but not when distance-to-thresho ld was fixed across contrasts. In the model, latency also increases as surround-to-center ratios increase, which has been shown to occur at higher contrasts. We propose that higher -contrast full-field st imuli can evoke more surround inhibit ion, shift ing the potential further from spiking thresho ld and thereby increasing response latency. Introduction Neurons in the early visual system must encode visual info rmation across a broad range o f temporal contrasts. Some neural properties are relat ively conserved across contrasts. For example, the pattern of individual spiking events within a stimulus is similar across contrasts. Other neural properties have been reported to change with contrast. One such property is the response latency. Studies have reported that it takes longer for a visual neuron to respond to an excitatory stimulus during low contrast condit ions (Baccus and Meister, 2002; Benardete and Kaplan, 1999; Chander and Chichilnisky, 2001; Kaplan and Benardete, 2001; Kim and Rieke, 2001; Levitt et al., 2001; Shapley and Victor, 1978; Solomon et al. , 2004; Victor, 1987; Zaghloul et al., 2005). In 1978, Shapley et al. reported that the first -order response became less sharply tuned during lower-contrast stimuli and the phase of the response relative to a periodic stimulus was delayed. More recently, neural responses during different contrasts have been fit to Linear-Nonlinear cascades, and reports have confirmed that both the dynamics and the sensit ivity o f neurons depend on contrast (Baccus and Meister, 2002; Chander and Chichilnisky, 2001; Kaplan and Benardete, 2001; Kim and Rieke, 2001; Zaghloul et al., 2005). In retinal ganglion cells, the consistent finding is that responses have shorter latency and faster dynamics when stimulus contrast is high. The effects o f contrast on response latency have not been examined in detail at the next stage of processing, in the LGN. Addit ionally, it is known that the ON and OFF pathways are characterized by numerous physio logical differences beyond a simply sign-inversion. For example, ON retinal ganglion cells can be characterized by a higher spontaneous firing rate, more linear light response, larger receptive fields, and faster kinet ics than OFF retinal ganglion cells (Chichilnisky and Kalmar, 2002; Cleland et al., 1973; Kaplan et al., 1987; Pa ssaglia Evidence for an additive inhibitory component of contrast adaptation 1 DRAFT 5/19/09 Gaudry and Reinagel et al., 2001; Troy and Robson, 1992; Zaghloul et al., 2003) . Addit ionally, contrast has been reported to different ially affect the sensit ivity of the two cell types in the retina. Further, while higher contrasts produced shorter latency responses for both cell types, this effect was more pronounced for OFF cells (Chander and Chichilnisky, 2001; Zaghloul et al., 2005). The interactions between the ON and OFF pathways provide further support for the differences between the pathways (Cohen, 1998; Renteria et al., 2006; Wassle et al., 1986; Zaghloul et al., 2003) . Therefore, we predicted that contrast may different ially affect latencies of LGN cells within the two pathways. In contrast to previous reports, we find that the response latency for cells o f the LGN can decrease at lower contrasts. This surprising trend is particularly evident in ON cells. We propose that the paradoxical trend o f shorter latencies at lower contrasts may be at least partially explained by a change in the distance-to-thresho ld with contrasts. A linear non-linear cascade analys is lends support to this hypothesis: at lower contrasts, spikes occur in response to a smaller generator potential ( eg., excitatory st imulus strength) than do spikes at higher contrasts. Further support is o ffered by a realist ic LGN spiking model that does not produce the unexpected latency effect when the thresho ld is fixed across contrasts but does when the distance to thresho ld is changed across contrasts. We show that our hypothesis can also predict the occurrence of low -contrast specific spiking events. Finally, Lesica et al. (2007) recently reported that the surround -to-center ratio of the spat ial-temporal receptive fields can increase as contrast increases. We use the LGN spiking model to show that this change in the relat ive inhibit ion can predict the unexpected latency effects that we observed. Results Latency of ON cells Increases with Contrast We recorded from 41 relay cells o f the LGN of anesthetized cats, while present ing full-field flickering binary white no ise stimuli with each of three contrasts: 100%, 33% or 11% (see Methods). Responses from one representative OFF cell and one representative ON cell are shown in Figure 1. The times o f the majority of spiking events are similar across contrasts, but we observe subtle changes in times of the spikes as contrast changes. For the OFF cell, for most of the spiking events, the response latency increases as the contrast decreases, consistent with previous results. Interestingly, the ON cell’s responses show the opposite trend: as contrast decreases, the response latency decreases. In order to quantify the latency during each contrast condit ion, we estimate the linear filter of the cell as the spike-triggered average. We then define the latency as the time of the first peak of the spike-triggered average. In general, the response latency o f ON cells increased as contrast increased, although some cell-by-cell variability was observed. In Figure 2A-F we show the spike-triggered averages during the high-, medium-, and low-contrast conditions for six ON LGN cells. Evidence for an additive inhibitory component of contrast adaptation 2 DRAFT 5/19/09 Gaudry and Reinagel Figure 1. Responses of one ON and one OFF cell as a function of temporal contrast. One representative OFF Y LGN cell and one representative ON X LGN cell were stimulated by a full -field binary white noise visual stimuli of different contrasts. First row: Raster plot for 64 repeats of the same stimulus at 100% contrast. Each point represents the time of an action potential within that trial. A 150-ms segment is shown from the middle of the 5 s trials. Second and third rows: Responses to 6 4 repeats of the same full-field binary stimulus sequence, which was scaled about the mean to 33% and 11% contrast, respectively. Fourth row: Time varying firing rate in 1ms time bins during 100% (red curve), 33% (blue curve), and 11% (gray curve) contrast stimuli, derived from data in the raster plots. Bottom row: Binary stimulus presented at each of the three contrasts. Evidence for an additive inhibitory component of contrast adaptation 3 DRAFT 5/19/09 Gaudry and Reinagel Figure 2A-B show two ON cells represent ing the most commonly obser ved latency trend amongst ON cells: the latency increased as contrast increased. That is, the time of the first peak of the spike-triggered average was later for high-contrast stimuli (red arrow) than for medium-contrast stimuli (blue arrow), and the t ime o f the first peak of the spike-triggered average was later for medium-contrast stimuli (blue arrow) than for low-contrast stimuli (gray arrow). Figure 2. Effects of contrast on latency. A-F, The average stimulus preceding spikes, normalized by the amplitude of the first peak, from six representative ON LGN neurons calculated in the 100% contrast (red), 33% contrast (blue) and 11% contrast (gray) conditions. The arrows indicate the latency of the cell at each of the three contrasts. G, Response latency across cells. Latency of responses at 100% contrast (horizontal axis) is compared to latency of the same cell at either 33% contrast (○) or 11% (∆) contrast. ON and OFF cells are represented by green and purple symbols, respectively. Open symbols indicate cells for which – for at least one of the contrast conditions from the comparison – over 20% of the responses were bursts. Symbols above the diagonal indicate a latency decrease with contrast. H, Latency differences across cells. Shapes and colors of symbols are as described for panel G. Bars indicate the average latency difference across either the ON or OFF cells for the indicated contrast comparison. Cells for which over 20% of the responses were bursts (open symbols in panel G) are not shown. Evidence for an additive inhibitory component of contrast adaptation 4 DRAFT 5/19/09 Gaudry and Reinagel in Resting Membrane Potential or For some ON cells, the latency increased as the contrast increased from medium contrast to high contrast (compare red and blue curves in Figure 2C-D) but showed minimal to no difference as the contrast increased from low contrast to medium contrast (compare blue and gray curves in Figure 2C-D). For a small fract ion of the ON cells, contrast had only negligible effects on latency (Figure 2E). Finally, for another small fraction of the ON cells, the latency during low -contrast stimuli was longer than that during high-contrast or medium-contrast stimuli. We note that in these condit ions, more than 20% of the low-contrast responses were bursts (see Discussion). Across cells, we found that the effect of contrast on response latency depended on whether the neuron was an ON or OFF neuron. For OFF cells, the latencies from the high-contrast responses are significantly sho rter than those from the medium- or low- contrast responses (purple symbo ls in Figure 2G-H; p < 0.001 for both comparisons). The decrease in latency from low to medium contrast was also significant (p<0.01). For ON cells, latency changed significant ly in the other direct ion (green symbo ls in Figure 2C). The increase in latency from medium to high contrast was significant (p<0.001). While most cells also increased latency from low to high contrast, some cells that were characterized by high burst rates at low contrast showed the opposite trend, such that the increased latency trend was not significant (p>0.05). Predicted Contrast-Dependent Change Threshold In order to explore the latency result s above, we propose that ON cells lie clo ser to their thresho ld at higher contrasts. We first consider a generator potential (convo lved stimulus) at two contrasts. We convo lve a segment of the high- and medium-contrast stimuli that were presented to the LGN cells with the spike-triggered average from the ON cell shown in Figure 1 to estimate the two generator potentials, gH(t) and gL(t). Figure 3A shows a segment of the generator potent ials gH(t) and gL(t) corresponding to the same segment of the stimulus shown in Figure 1. In a simple model, a cell fires a spike every t ime the generator potential g(t) crosses a thresho ld. The dotted line in Figure 3A indicates an estimated thresho ld, which was calculated by ident ifying the thresho ld producing the highest correlat ion between the observed t ime-varying probability of spiking and convo lved stimulus rectified at the thresho ld. For high contrast, the fluctuations in the st imulus are large. Therefore, the generator potential gH(t) may quickly cross the thresho ld before, for example, reaching the peak of any given fluctuation. Meanwhile, for the lower contrast, the fluctuations are smaller, such that the generator potential gL(t) may need to wait unt il the most excitatory part of the fluctuation to reach thresho ld. As explained above, the illustration o f Figure 3A predicts that the generator potential g(t) would cross the thresho ld later during low-contrast condit ions than high- contrast conditions, result ing in longer latencies during low contrasts. While this is consistent with results reported for retinal ganglion cells, the opposite trend is observed for the ON LGN cells (Figure 2G-H). Figure 3B illustrates the effect of reducing the distance-to-thresho ld for the lower- contrast condit ion. This can be accomplished either by adding a constant term to the generator potential gL(t) or by reducing the threshold for the lower-contrast condition. In Evidence for an additive inhibitory component of contrast adaptation 5 DRAFT 5/19/09 Gaudry and Reinagel this instance, we added a constant to the low-contrast generator potential gL(t) to obtain a modified generator potential gL’(t). (The constant was found by maximizing the correlat ion between the low-contrast time-varying firing rate and the modified generator potential gL’(t) rectified at the thresho ld). The same effect could be produced by increasing the distance-to-thresho ld for the higher-contrast condit ion. Figure 3. Effects of reducing distance-to-threshold of a generator potential. A, Generator potentials g(t) were calculated by convolving the stimulus at high - (red) or medium- (blue) contrast by the spike- triggered average of the ON cell from Figure 1B. The threshold (dotted line) was estimated as explained in Results. B, The modified generator potential (dotted line) was obtained by adding a constant value to the medium-contrast generator potential shown in A. C and D, the generator potentials shown in A and B are rectified at the threshold. E, The observed time-varying firing rate from the ON cell from Figure 1B for both contrast conditions. The fluctuations in the modified generator potential gL’(t) remain smaller than those of the high-contrast generator potential gH(t). However, the reduction of the distance-to-thresho ld enables the modified generator potential gL’(t) to cross the thresho ld Evidence for an additive inhibitory component of contrast adaptation 6 DRAFT 5/19/09 Gaudry and Reinagel earlier than would the unmodified low-contrast generator potential gL(t). A sufficient reduction of the distance-to-thresho ld can result in the modified low-contrast generator potential gL’(t) crossing thresho ld before the corresponding high-contrast generator potential gH(t) would cross thresho ld. We compared the rect ified high-contrast generator potential and either the rectified unmodified low-contrast generator potent ial (Figure 3C) or to the rectified modified low-contrast generator potential (Figure 3D) to the observed probability o f spiking (Figure 3E). Rectifying the generator potential g(t) is consistent with a mode l that assumes the probability o f spiking is zero until the generator potent ial g(t) crosses a thresho ld, after which the probability o f spiking is correlated with the generator potentia l g(t). Notably, the rect ified generator potential does not include refractory effects, which would truncate the spiking events and increase their reliability (Kara et al, 2000; Keat et al, 2001). Subt le differences between the rectified generator potential and the probability o f spiking are present, yet the rectified modified generator potential is able to produce similar latency changes observed in the neural responses of the LGN ON cell (blue peaks cross the thresho ld before the red peaks in Figure 3D). Meanwhile, the rectified unmodified generator potential cannot capture this latency trend (blue peaks cross the thresho ld after the red peaks in Figure 3C). Shift of Input-Output Functions Depends on Cell Type Above we speculate that the opposite latency effect observed in ON cells may be at least partially explainable by a decrease in the distance to thresho ld. We use a linear- nonlinear cascade to test this hypothesis in our LGN data. See Methods. Briefly, at each contrast, we estimate the linear filter o f the neuron by calculat ing the normalized spike - triggered average (Figure 4A). The generator potential, g(t), is defined as the st imulus convo lved by this filter. By comparing the generator potential with the observed probability o f firing across time bins, we estimate the Input -Output funct ion (Figure 4B). The Input-Output funct ion can be fit to a sigmoid. The amplitude of the sigmo id is related to the maximum probability o f firing within a time bin. The slope of the sigmo id is related to the gain or sensit ivity of the neuron: larger slopes indicate that the firing rate is more sensit ive to stimulus fluctuations. The horizontal o ffset of the sigmo id can be thought of as being related to the difference between the baseline membrane potential and the thresho ld of the neuron: offsets further to the left indicate that a less excitatory stimulus will cause the neuron to spike. Therefore, the hypothesized mechanism proposed to explain the latency result of the ON cells would also predict that the horizontal offset would change with contrast for the ON cells. The gain of neural responses is known to increase as st imulus contrast decreases (Benardete and Kaplan, 1997; Benardete et al., 1992; Brown and Masland, 2001; Chander and Chichilnisky, 2001; Jin et al., 2005; Kim and Rieke, 2001; Kremers et al., 2001; Shapley and Victor, 1978; Shou et al., 1996; Smirnakis et al., 1997; So lomon et al., 2004; Zaghloul et al., 2005) and we have previously reported the same trend for this data set (Gaudry and Reinagel, 2007a). Therefo re, we do not hold either the gain or horizontal offset constant across contrasts, such that the observed gain changes will not confound horizontal o ffset changes. Still, we find that the Input -Output functions o f the ON cell in Figure 4B differ by more than just the gain. This is demonstrated when we Evidence for an additive inhibitory component of contrast adaptation 7 DRAFT 5/19/09 Gaudry and Reinagel adjust the gain o f the medium-contrast Input-Output funct ion to match that at high contrast. The two nonlinear functions st ill differ in their horizontal o ffsets (Figure 4C and 4D). We found that the effect of contrast on horizontal offsets depended on the cell type. For ON cells, the horizontal o ffset of the Input -Output funct ion decreased as the contrast decreased (symbo ls below the diagonal in Figure 4E). A decrease in o ffset indicates that the input -output function shifted to the left, signifying that smaller generator potentials could evoke spikes at low contrasts than could at high contrasts. This shift is consistent with either a depolarization or a lower thresho ld at low contrasts. For OFF cells, the offset could increase, remain unchanged, or decrease (Figure 4F). Shift s in o ffsets for ON cells (average shift = -0.331 ± 0.324) differed significant ly (p< 0.0001) from OFF cells (average shift = -0.0120 ± 0.242), while there was no significant difference between X and Y cells (p=0.38). Many factors may interact to determine the effect that a change in the shift would have on latency. Nevertheless, we expected that cells characterized by large shifts in the offsets of the nonlinear funct ions would also show large changes in latency. Therefore, for each cell, we compared the difference in latency during two contrast condit ions to the difference in offsets. (We excluded from this analysis data points in which over 20% o f the responses were bursts, as bursts have different dynamics than tonic spikes). As shown in Figure 4G, the latency difference and offset difference were indeed correlated (R2=0.15, p<.001). Cells whose latency was longer at higher contrasts (symbo ls below horizontal y=0 line) were also described by nonlinear funct ions that shifted to the left as contrast decreased (symbo ls to the left of vertical x=0 line). In Figure 4I, we show that the observed differences in o ffsets can relate to substant ially different time-varying firing rates. We compare predicted low-contrast time-varying firing rates using the linear-nonlinear model with various o ffsets to the observed low-contrast time-varying firing rate (Figure 4I; gray curve) from the ON cell from Figure 1 and from Figures 4A-D. For each of three condit ions, we define the linear filter as the cell’s low-contrast spike-triggered average. Further, we define the gain o f the Input-Output funct ion as the low-contrast gain calculated for the cell. We then separately calculate the predicted firing rate using the o ffsets calculated for that cell for each o f the contrasts. Thus, differences in the predicted time -varying firing rates are ent irely due to the contrast dependence o f the o ffsets. The predicted time-varying firing rate associated with the low-contrast offset (black curve) provided a better match to the observed time - varying firing rate than either o f the t ime-varying firing rates associated with the medium-contrast offset (blue curve) or the high-contrast offset (red curve). In this instance, the predicted time-varying firing rate associated with the high-contrast offset barely predicted any spikes at all. Across cells, we calculated a set of first correlation coefficients between the predicted firing rates using the low-contrast offset and the observed low-contrast firing rate as well as a set of second correlat ion coeffic ients between the predicted firing rates using the high-contrast offset and the observed low-contrast firing rate. For OFF cells, the first and second sets of correlat ion coefficients were not significant ly different from each other (p=0.34). For ON cells, the first set of correlat ion coeffic ients was significant ly higher than the second set of correlation coefficients (p<0.01). Thus, for Evidence for an additive inhibitory component of contrast adaptation 8 DRAFT 5/19/09 Gaudry and Reinagel ON cells, contrast-dependent shift s in the offsets of the input -output funct ions can contribute to significant differences in t ime-varying firing rates. Evidence for an additive inhibitory component of contrast adaptation 9 DRAFT 5/19/09 Gaudry and Reinagel Figure 4. Shift in Input-Output functions with contrast. LGN responses were fit to a Linear -Nonlinear cascade at each contrast. Panels A-D illustrate the method for one Y OFF cell. A, Linear filters from a Linear-Nonlinear cascade model, fit to high- (red curve) and medium- (blue curve) contrast data. The filters were normalized by the amplitude of their first peak. B, Symbols indicate the observed probability of spiking versus the generator potential, g(t), for the same cell as in (A). The curves show the best-fit sigmoid functions. Based on these fits, the gain at the high contrast was 0.63 times that at medium contrast. C, Equivalently, we can scale the generator potential at medium contrast by a factor of 0.63 in order to set the gain of the two input-output functions to the value at high contrast. For this cell, the nonlinear functions also differed in their horizontal offsets. D, After shifting the blue curve horizontally to correct for the difference in the horizontal offset (which differed by 0.97 units), the same sigmoid function (black curve) can now describe both the medium contrast’s shifted and scaled nonlinear function (R2 = 0.996) and the nonlinear function at high contrast (red symbols; R2 = 0.991). E, Across all ON cells, we compare the offset between 100% and 33% contrast (○), 33% and 11% contrast (◊), and 100% and 11% (∆) contrast, where the gain at the higher contrast is always shown on the horizontal axis. If all differences in the nonlinearity were captured by changes in gain, offse ts would be constant across contrasts (x-y, thick, solid line). Symbols below the line indicate the nonlinear function was shifted to the left at lower contrasts. F, We compare changes in offset with contrast for all OFF cells. Symbols as in (E). G, Di fference in latency, shown in Figure 2C, compared to the difference in offsets from the nonlinear functions (shown in Figure 4E and 4F). Both differences are calculated as the variable at the lower contrast minus the variable at the higher contrast. Ther efore, symbols above the horizontal x=0 line indicate that latency increased as contrast decreased, and symbols to the left of the vertical y=0 line indicate that the nonlinear function at the lower contrast was shifted to the left of that at the higher contrast. Each symbol represents a comparison between two contrast conditions for one cell. Symbol shapes indicate whether the comparison was between 100% and 33% contrast (○), 33% and 11% contrast (◊), or 100% and 11% (∆). Symbol colors are as defined in (C). H, Anti-correlation of multiplicative and additive gain changes. Symbols as in (G). I, Time varying firing rate (light gray line) in 1ms time bins of the ON cell from Figure 1 and panels A -D during a low-contrast stimulus. Colored lines indicate fir ing rates predicted by the linear -nonlinear model using offsets from the high (red curve), medium (blue curve) or low (dark gray curve) contrasts. Anti-Correlation of Multiplicative and Additive Gain Change It is well known that the gain of an LGN neuron can change as the contrast changes, such that the neuron is more sensit ive to smaller fluctuations at lower contrasts. Typically, gain changes are referred to as a mult iplicat ive quality: it is thought that the neuron essentially amplifies or mult iplies a signal at low contrasts. This differs from the addit ive gain change illustrated in Figure 3. Only the addit ive gain change can account for the latency trends reported herein. However, we suggest that a combinat ion o f the mult iplicative and addit ive gain changes can occur. We hypothesize that cells that exhibit less contrast normalization (mult iplicat ive gain change) may exhibit a larger change in resting membrane potential or thresho ld (additive gain change). As described above, these properties can be separately measured for each neuron using the linear-nonlinear model. See Figure 4A-D. by comparing the change in We calculated the extent of contrast normalizat ion the neuron’s sensit ivity or gain to the change in contrast. If the ga in increased by the same factor that the contrast decreased, the contrast normalization is defined as 1, while if the gain did not change as contrast changed, the contrast normalizat ion is defined as 0. See Methods. We find a significant correlation between the contrast normalizat ion and the offset difference of the nonlinear funct ions described above (Figure 4H; R2 = 0.30, p<0.0001). Some neurons show litt le contrast normalizat ion and large offset differences (symbo ls in bottom left corner of Figure 4H), while others show large contrast normalizat ion and little Evidence for an additive inhibitory component of contrast adaptation 10 DRAFT 5/19/09 Gaudry and Reinagel offset differences (symbo ls in top right corner of Figure 4H). Still the majority o f neurons are characterized by a combinat ion of the mult iplicat ive and addit ive mechanisms. Latency Changes in LGN Spiking Model Above, we provide an illustration to show how changes in the generator potentia l may produce the opposite latency effects that we observe in ON and OFF cells in our experimental data (see Figure 3 and related Results sect ion). However, as ment ioned, the convo lved st imulus does not incorporate other factors known to influence spiking responses, such as refractoriness and no ise. Because ON cells’ firing rate increases with contrast more than OFF cells, it was possible that longer -latency responses at high contrasts are due to the more frequent ly induced abso lute or relat ive refractory periods. Therefore, we implemented a nonlinear model that has been shown to successfully reproduce responses o f LGN neurons to white no ise st imuli (Gaudry and Reinagel, 2007b; Keat et al., 2001). This model includes both noise terms and refractory terms (see Methods). A large populat ion of model cells (N=550) was created by varying the model parameters (see Methods). The same high- and medium-contrast stimuli that were presented to LGN neurons were used to generate model cell responses. In the first condit ion, all parameters remained fixed across the contrasts for each model cell. We have previously shown that even with fixed parameters, the model replicates the mult iplicat ive contrast normalization associated with contrast gain control (Gaudry and Reinagel, 2007a, 2007b). The latency was determined for each model cell at each contrast as described for Figure 2, by measuring the time of the first peak of the spike-triggered average. As shown in the contour density plot of Figure 5A and in the Figure 5. Latency trends in a spiking LGN model. The latency in milliseconds at high contrast was compared to that at medium contrast for each model cell. At each x-y point, the number of model cell results is indicated by color, where red corresponds to the highest density and dark blue to the lowest. The color is scaled as the log of the probability. A, Latency comparison when the threshold parameter for model cells was fixed across contrasts ( =1.6). B, Latency comparison when the threshold parameter at high contrast ( =1.6) was greater than that at medium contrast ( =0.4). C, Histograms of the probability of contrast-dependent latency differences. The dashed curve shows results from fixed- parameter data shown in (A), and the solid curve shows results from threshold -changing data shown in (B) Evidence for an additive inhibitory component of contrast adaptation 11 DRAFT 5/19/09 Gaudry and Reinagel dashed curve in Figure 5C, the only contrast-dependent latency changes observed across the populat ion of model cells were decreases in latency during higher contrast stimuli (the distribut ion indicated by the dashed curve is strict ly to the left of 0). This suggests that noise and refractoriness are not sufficient to explain the opposing latency increase observed in LGN ON cells. In the second simulat ion condit ion, the thresho ld was lowered during the lower-contrast condit ion. We note that in the model this is formally equivalent to depolarizing the rest ing potential relative to a fixed threshold. The change in thresho ld was suffic ient to produce both the increasing and decreasing latency trends with contrast, as shown in Figure 5B and the so lid curve in Figure 5C. Latency Variability within Individual Cells Above, we characterize the latency o f cells by calculat ing a single number: the time of the first-peak of the spike-triggered average. However, the spike-triggered average can be confounded by a variety o f factors, such as the number o f spikes elicited in response to a stimulus. Therefore, we also compared the times of peaks in the time- varying firing rate across contrasts. Peaks within the time-varying firing rates were ident ified as spiking events (see Methods). The times of spiking events can be conserved across contrasts, so a blind observer conservat ively ident ified which events were “shared” across contrasts. Because this analysis can require more data than the calculation o f a spike -triggered average, this analys is was performed for 8 of the ON cells and 8 of the OFF cells. Our previous spike-triggered average calculat ions indicated that ON cells decreased their latencies o f response at lower contrasts, opposite to the latency increase we observed in OFF cells and that others have reported in both cell types using other visual stimuli. This analysis similarly showed that the times of the spiking events at lower contrasts precede the times of the corresponding spiking events at higher contrasts (symbo ls are below the y=0 line in Figure 6). We also found that the latency difference varied among the spiking events within individual cells (symbo ls are scattered along the y-axis in Figure 6). This variability could be at least partially explained by differences in the slope o f the generator potential: the steeper the slope of the generator potential, the less an addit ive shift will affect the time o f thresho ld crossing. Above we theorize that for ON cells the baseline generator potential is closer to thresho ld when contrast is lower. This would imply that that the higher-contrast events will substant ially lag lower-contrast events when the slope o f the generator potential preceding the event is shallow (for example, t=350-390 ms in Figures 3B, 3D and 3E). Litt le to no difference in the event times is predicted when the slope o f the generator potential preceding the event is steep (for example, t=450-500 ms in Figures 3B, 3D and 3E). To quantify this predict ion, we calculated for each firing event the average slope of the most recent increasing segment of the generator potential preceding the event. As shown in Figure 6, when the slope o f the generator potential was shallow (closer to 0), the higher-contrast events are substant ially later than the corresponding lower -contrast events. When the slope o f the generator potential was steep, there is little to no difference in the event times. OFF cells also exhibited within-cell variability in the latency differences o f spiking events. Interestingly, responses o f most OFF cells included both spiking events Evidence for an additive inhibitory component of contrast adaptation 12 DRAFT 5/19/09 Gaudry and Reinagel for which high-contrast responses preceded lower-contrast responses and the converse (data not shown). Few cells showed a strong and significant correlat ion between generator potential slope and latency difference. Figure 6. Latency differences for spiking events for ON cells depend on generator potential slope. Each panel corresponds to one ON cell. Each symbol corresponds to one spiking event identified as an event shared across contrasts. For each event, an event time difference was identified as the time of the event at medium contrast minus the time of the corresponding event at high contrast. (Values below the y=0 line indicate that the medium-contrast event preceded the high -contrast event). These values were compared to the average slope of an increasing segment of the generator potential, g(t), preceding the event. Evidence for an additive inhibitory component of contrast adaptation 13 DRAFT 5/19/09 Gaudry and Reinagel Discussion Mechanistic Hypothesis to Exp lain Opposing Latency Effects Neurons of the early visual system have long been reported to exhibit longer - latency responses for low-contrast stimuli than fo r high-contrast stimuli. We report that LGN cells can under some condit ions exhibit the opposite t rend (Figures 1, 2 and 6). Both spike-triggered average analys is and spiking event -based analysis support this finding. We propose that at high contrast, an increase in inhibit ion causes an increase in the surround-to-center ratio, an increase in the distance-to-thresho ld, and therefore an increase in response latency. Although contrast adaptation is more o ften reported, there is precedent for an addit ive mechanism in neurons of cat visual cortex (Carandini and Ferster, 1997). Within the LGN, the tonic hyperpo larizing current through GABAA receptors is one candidate for providing the postulated addit ive shift in membrane potential (Cope et al., 2005). No addit ive adaptation has been described in the retinal ganglion cell inputs to the LGN, but comparable data do not exist to exclude this possibility. An increase in the strength of surround suppression could be attributed either to an increase in synapt ic inhibit ion, or withdrawal o f synapt ic excitation. Here we consider in more detail the possible role of synaptic inhibit ion at the level o f the LGN. LGN neurons receive direct or indirect inhibit ion through local feed -forward inhibitory pathway, a loop through the reticular nucleus, and a feedback loop through visual cortex. Feed forward inhibit ion in the LGN is both powerful and rapid. Inhibitory synapt ic input to the receptive field center has been shown to underlie other temporal computations underlying the visual control of burst ing (Wang et al 2007). In the case of contrast adaptation, we suggest that changes in temporal contrast result in changes in the spat ia l opponency o f the recept ive field, specifically, an increase in the inhibitory input to the opponent surround region. This hypothesis makes the experimental predict ion that the adaptation of spat ial antagonism should be stronger in the ON pathway than the OFF pathway, though no such asymmetry has yet been reported. Blitz and Regehr (2005) reported that many LGN relay neurons receive nonlocked inhibit ion from LGN interneurons, and suggest that such inhibit ion may account for receptive field surround inhibit ion properties. The nonlocked inhibit ion received by a specific LGN cell was elicited by activat ion of retinal ganglion cells that do not synapse direct ly onto the cell. This inhibit ion most likely reflects disynapt ic inhibit ion from retinal ganglion cells with the opposite light sensit ivity to the cell’s excitatory inputs (the “pull” of the push-pull receptive field). One possibility is that at high contrast, LGN relay cells receive nonlocked inhibitory input from a wider spatia l field o f LGN interneurons, increasing the suppressive surround of the LGN recept ive field. It is not clear, however, why such a mechanism would have an addit ive rather than mult iplicative effect. Evidence for an additive inhibitory component of contrast adaptation 14 DRAFT 5/19/09 Gaudry and Reinagel Although there is litt le evidence for ON-OFF asymmetries in LGN anatomy, the effect of feed-forward inhibit ion can be substantially different depending on whether the LGN neuron is in a burst mode or a tonic spiking mode (Blitz and Regehr, 2005). In Figure 2F-G and the associated text, we report that the probability o f burst responses increased to above 20% for three ON cells. The latency at low contrast for these three ON cells was greater than the latency at higher contrasts, in contrast to the genera l latency trend observed for ON cells. The difference in the effect of feed -forward inhibit ion may at least partially account for the differences in these cells’ latency trends. An alternat ive abstract model of the LGN has been presented that accounts for a wide range o f suppressive effects in terms of contrast gain control at the level o f neura l firing rate (Bonin et al, 2005). That model is not designed to make predict ions regarding spike latency, and does not incorporate an addit ive component of contrast adaptation, and thus we believe it addresses a dist inct component of contrast adaptation. It is unclear which inhibitory inputs contribute the mult iplicat ive suppression described by Bonon et al, and which underlie the addit ive effect we describe. Previously Reported Latency Trends The mechanism we hypothesize above can account for our data as well as apparent ly disparate previous reports. First, the majority o f the previous studies that reported a consistent latency decrease with higher contrast were studies o f ret inal neuro ns (eg., Baccus and Meister, 2002; Benardete and Kaplan, 1999; Chander and Chichilnisky, 2001; Kaplan and Benardete, 2001; Kim and Rieke, 2001; Shapley and Victor, 1978; Victor, 1987; Zaghloul et al., 2005) . Therefore, LGN interneurons would not yet be able to influence the responses. Alternat ively or in addit ion, the stimulus used in many previous studies differed from that of the current experiment. Full-fie ld stimuli may cause more nonlocked inhibit ion to an LGN relay neuron: the LGN relay neuron may receive nonlocked inhibit ion from a greater number of interneurons and/or the LGN relay neuron may receive more nonlocked inhibit ion from a single interneuron if many ganglion cells synapse onto the interneuron. Further, many previous studies used Gaussian st imuli or sum-o f-sinuso id st imuli, whereas the current study used binary st imuli. Our high- contrast stimulus may have thus been of substant ially higher contrast than other studies’ high-contrast stimulus. To determine the relat ive importance of the cell type and the stimulus, it will be o f crit ical importance to learn whether retinal ganglion cell responses also show latency advance under our stimulus condit ions. ON-OFF Cell Asymmetry We report two ON-OFF cell asymmetries: a contrast -dependent latency asymmetry (Figure 2) and a contrast-dependent shift in the nonlinear spiking funct ion (Figure 4). Both could be explained by a contrast -dependent change in distance to thresho ld in ON cells. We cannot at present specifically explain why ON and OFF cells show these asymmetries, although we consider that our reported asymmetries may relate to other previously reported differences between the pathways. We do not know if the opposing latency effect arises in the LGN, or whether this ON-OFF asymmetry may trace back to the retina. Evidence for an additive inhibitory component of contrast adaptation 15 DRAFT 5/19/09 Gaudry and Reinagel Although the average latency change for OFF cells was generally in the expected direct ion, individual spiking events of OFF cells could exhibit an opposite latency change to the general trend (Figure 6). If the addit ive shift is attributable to inhibitory input at the level of the LGN, it is possible that both the ON and OFF LGN relay neurons receive nonlocked inhibitory input from LGN interneurons but that the inhibitory input is more sustained for the ON pathway. We hypothesize that ON retinal ganglion cells provide more excitat ion to LGN interneurons than OFF retinal ganglion cells, thereby causing ON LGN relay cells to receive more sustained inhibit ion. Zaghloul et al. (2003) reported that the maintained firing rate for ON retinal ganglion cells (18.4 ± 3.8 Hz) is larger than that for OFF retinal ganglion cells (6.6 ± 1.3 Hz). Therefore, the ON retinal ganglion cells may excite a given LGN interneuron more than do OFF ret inal ganglion cells. Addit ionally, ON retinal ganglion cells have been reported to have larger recept ive fields than OFF ret inal ganglion cells (Chichilnisky and Kalmar, 2002). We consider the possibility that in the ON pathway, more retinal ganglion cells synapse onto an LGN interneuron, and/or that more LGN interneurons synapse onto an LGN relay cell. Blitz and Regehr (2005) reported that only a subset of LGN relay cells received nonlocked inhibit ion, and Lesica et al. (2007) reported that only a subset of the LGN cells showed an increase in the surround-to-center ratio with contrast. We would predict that more ON cells are included in these subsets than OFF cells. Low-Contrast Specific Spiking Events Predictions We have forwarded the hypothesis that the distance to thresho ld can change with contrast for LGN neurons. We note that this hypothesis makes the further prediction o f low-contrast specific firing events. If a change in the contrast merely scaled the generator potential without producing a change in the distance from the baseline generator potential to the thresho ld, all spiking events (ident ified by peaks in the t ime - varying firing rate) at low thresho ld should have corresponding spiking events at high contrast. For example, in Figure 3A every t ime the convolved low -contrast stimulus crosses the thresho ld, there is a corresponding instance that the convo lved high-contrast stimulus crosses the thresho ld. This is because every t ime the convo lved low-contrast stimulus crosses the thresho ld, the corresponding convo lved high-contrast stimulus is larger (as it is merely a scaled version of the convo lved low -contrast stimulus). Therefore, it necessarily also crosses thresho ld. On the other hand, if the distance-to-thresho ld is reduced at lower contrasts, low- contrast specific spiking events are predicted (eg., Figure 3D shows blue peaks without corresponding red peaks at approximately 80 and 445 ms). Consistent with our hypothesis, in LGN neurons we find that while many spiking events are conserved across contrasts, some spiking events are only present during low contrasts ( see Figure 3E). The times o f these low-contrast specific spiking events are the same as those predicted by the “depo larized” convo lved low-contrast stimulus (compare Figure 3D and 3E). Additive Shift versus Multiplicative Gain Change Figure 4E-G suggests that in the LGN, the magnitude of the contrast -dependent change of the distance-to-thresho ld (the shift in the nonlinear funct ion) varies across cells. Interestingly, we also find that cells exhibit ing strong mult iplicative gain changes are less likely to show a horizontal shift in their non-linear functions (see Figure 4H). Evidence for an additive inhibitory component of contrast adaptation 16 DRAFT 5/19/09 Gaudry and Reinagel Conceptually, both the gain changes and the addit ive shift may address the same problem: encoding a range o f st imuli using a fixed distribution o f firing rates. Encoding a range o f contrasts presents the difficulty o f avo iding firing rate saturation at high contrasts while also preserving sensit ivity to small fluctuations at low contrasts. Mult iplicat ive gain change indicates that a neuron is more sensit ive to changes in the stimulus at low contrasts than at high (eg., a scaling o f the generator potential). The addit ive shift indicates that all st imuli are closer to thresho ld, thereby improving the probability o f a resultant spike (eg., the addit ion of a constant to the generator potential). Thus, while neurons show variability both the extent of gain change and addit ive shift, either of these effects or a combination thereo f may serve the common purpose of improving encoding across a range o f contrast condit ions. Cell-to-cell variability in the extent of addit ive shift versus mult iplicat ive gain change may relate to bipo lar cells contributing to the response. Baccus and Meister (2002) reported that bipo lar cells showed either an addit ive or a mult iplicat ive gain change, but not both. Conclusion Under specific st imulus condit ions, some LGN neurons can exhibit shorter latency responses at low contrast than at high contrasts. Physio logical and model data supports the hypothesis that the distance-to-thresho ld is lower at low contrasts in those cells, which may account for the unexpected direct ion of the latency change. We hypothesize that ON LGN relay cells receive more inhibitory input at high contrasts, thereby producing a larger surround-to-center ratio, result ing in an addit ive form o f contrast adaptation that can be revealed by longer response latency. Experimental procedures Surgical preparation. Cats were init ially anesthetized with ketamine HCl (20 mg/kg, i.m.), fo llowed by sodium pentothal (2-4 mg * kg -1 * h-1, i.v., supplemented as needed). Cats were ventilated using an endotracheal tube. E lectrocardiogram, electroencephalogram, temperature, exp ired CO2, and oxygen in blood were continuously monitored. Surgical and experimental procedures were in accordance with Nat iona l Inst itutes of Health and United States Department of Agriculture guidelines and were approved by the UCSD Institutional Animal Care and Use Committee. Electrophysiology. We recorded from 41 LGN relay cells, sampled the four main cell classes: ON X, OFF X, ON Y, and OFF Y. Parylene-coated tungsten electrodes (AM Systems, Everett, WA) were inserted through a 0.5 cm diameter cranio tomy over the LGN. E lectrical recordings were amplified, filtered, and digit ized at 10kHz sampling rate (CED micro 1401 and Spike2, ver. 5.12a; Cambridge E lectronic Design, Cambridge, UK). Waveforms were analyzed offline iso late single unit responses (Fee et al., 1996). Visual stimulation. Stimuli were full-field and presented on a custom-built LED array. We began with a random binary st imulus of 125 frames/s for 10 s. The same binary st imulus was scaled about the mean to obtain three contrast condit ions (11%, 33%, and 100%, where contrast is defined as the standard deviat ion of the luminance over the mean). St imuli o f the three contrasts were interleaved and presented between 10 and 128 times each. We only analyzed the last five seconds of the response to each 10 -s stimulus. The mean luminance was constant across contrasts and was within the photopic range. Evidence for an additive inhibitory component of contrast adaptation 17 DRAFT 5/19/09 Gaudry and Reinagel Response Latency. Spike-triggered averages (STAs) were computed for each cell at each contrast. The latency was defined as the time to the shortest latency peak o f the STA. Because of no ise in our STA, we est imated the t ime o f the peak from a Gaussian fit as fo llows: we define tP as the time at which the abso lute value o f the STA reached it s maximum between -40 and 0 milliseconds before the spike. We determined the nearest time before and after tP (tA and tB) at which the STA crossed the mean st imulus value. The STA between tA and tB was fit to a Gaussian. The latency was defined as the time of the peak of the Gaussian. We repeated latency analysis by comparing the t imes to the second peak of the STA across contrasts. Results were unchanged. For 16 of the LGN cells (8 ON cells and 8 OFF cells), we also analyzed the latencies of individual spiking events. For each cell, at each contrast, we determined the time of the firing events (PSTH peaks). The times of the events were ident ified as in Berry et al., (1997), but we determined the event time as that for which the smoothed PSTH reached its first maximum within the event boundaries. A blind observer conservatively ident ified which of these events were shared across contrasts. Linear-Nonlinear Cascade. For each cell at each contrast, we first estimated the filter as the spike-triggered average. The filters fo r all contrasts were normalized by the amplitude of their first peaks. We convo lved the stimulus by the corresponding filter to create a generator potential, g(t) and then compared this generator potential to the observed probability o f spiking at each t ime bin. For each cell, we fit this observed input-output function at 100% contrast to the fo llowing sigmo id equation: Eq. 1 where A, G, and S describe the amplitude, slope, and horizontal offset of the nonlinear funct ion, respectively. We then fit the input -output funct ions o f the other two contrasts, ho lding the amplitude (A) constant for all contrast condit ions for each cell. Data sets were excluded from our analys is if there were less than 100 spikes observed in the response, or if the R2 value associated with the sigmo id fit was less than 0.90. For each cell at each contrast, we then define the Gain as G and the Offset as S from Eq. 1. Our results did not depend crit ically on how the filters were scaled; similar results wer e found by normalizing by the peak-to-peak amplitude of the spike-triggered average. In Figure 4I, we use the linear-nonlinear cascade model to predict the t ime- varying firing rate of an LGN neuron during the low-contrast stimulus condit ion. As described above, we first calculated the linear filter and the Gain G and Offset S of the input-output function for all three contrast condit ions. We then convo lved the low - contrast stimulus with the low-contrast linear filter. We defined three input -output funct ions, each of which used the Gain G from the low-contrast condit ion, but differed in that the Offset S was either from the high, medium or low -contrast condition. Based on the convo lved stimulus at each t ime po int, we could then estimate the probability o f spiking at a funct ion of time using each of these input -output functions. Contrast Normalization Index. The greater the contrast change, the larger gain (sensit ivity) change would be required to compensate. Therefore, we express the magnitude o f gain change relative to the magnitude of contrast change and call this the contrast normalization index : Evidence for an additive inhibitory component of contrast adaptation 18 )exp(SGxeAy DRAFT 5/19/09 Gaudry and Reinagel Eq. 2 where G is neural gain and C is st imulus contrast. LGN Spiking Model. Model cells were implemented as described in Keat et al., (2001). Briefly, the model first convo lves the st imulus with a linear filter. The generator potential is equal to the convo lved st imulus plus no ise; the amplitude and t ime -constant a and A, respect ively. A spike is of this no ise are defined by the model parameters generated whenever the generator signal crosses a thresho ld, . Each time a spike occurs, a negat ive after-potential is added to the generator potential, such that the thresho ld is crossed repeatedly during sustained excitatory st imuli. The amplitude, time-constant, and variability in the amplitude of the negative after -potential are defined by the mode l parameters B, P, and b, respect ively. With regard to the analysis presented in Figure 5, we generated a set of mode l cells using random combinat ions of the fo llowing first parameters: B = 3, 5, or 7; P = 20, b = 0.02, 0.15, or 0.28. The filter a = 0.01, 0.16, 0.31, or 0.61, 35, or 50; A = 20; 2π·exp(-(x- )2/(2· 1 2)) - 2π·exp(-(x- )2/(2· 2 2)). funct ion, F, was defined as F = 2 1 2 The parameters for F, were randomly chosen combinat ions from the fo llowing second parameter ranges: 2= [18,27], with the = [-70,-35]; 1= [10,18]; = [-45,-20]; restrict ion that F(x=0) be less than or equal to 5% of the maximum value of F. After the first and second parameters were chosen, data was generated for each contrast condit ion for two thresho lds: = 0.4 and 1.6. References Baccus, S.A., and Meister, M. (2002). Fast and slow contrast adaptation in retinal circuitry. Neuron 36, 909-919. Benardete, E.A., and Kaplan, E. (1997). The receptive field of the primate P retinal ganglion cell, II: Nonlinear dynamics. Vis Neurosci 14, 187-205. Benardete, E.A., and Kaplan, E. (1999). The dynamics o f primate M retinal ganglion cells. Vis Neurosci 16, 355-368. Benardete, E.A., Kaplan, E., and Knight, B.W. (1992). Contrast gain control in the primate retina: P cells are not X-like, some M cells are. Vis Neurosci 8, 483-486. Berry, M.J., Warland, D.K., and Meister, M. (1997). The structure and precision of retinal spike trains. Proc Natl Acad Sci U S A 94, 5411-5416. Blitz, D.M., and Regehr, W.G. (2005). Timing and specificity o f feed -forward inhibit ion within the LGN. Neuron 45, 917-928. Bonin, V., Mante, V., and Carand ini, M. (2005). The suppressive field of neurons in lateral geniculate nucleus. J Neurosci. 25,10844-56. Brown, S.P., and Masland, R.H. (2001). Spatial scale and cellular substrate of contrast adaptation by ret inal ganglion cells. Nat Neurosci 4, 44-51. Carandini, M. and Ferster, D. (1997) A tonic hyperpolarizat ion underlying contrast adaptation in cat visual cortex. Science. 276, 949-52. Chander, D., and Chichilnisky, E.J. (2001). Adaptation to temporal contrast in primate and salamander retina. J Neurosci 21, 9904-9916. Evidence for an additive inhibitory component of contrast adaptation 19 1/1/lowerhigherhigherlowerCCGG DRAFT 5/19/09 Gaudry and Reinagel Chichilnisky, E.J., and Kalmar, R.S. (2002). Functional asymmetries in ON and OFF ganglion cells of primate retina. J Neurosci 22, 2737-2747. Cleland, B.G., Levick, W.R., and Sanderson, K.J. (1973). Properties of sustained and transient ganglion cells in the cat retina. J Physio l 228, 649-680. Cohen, E.D. (1998). Interactions of inhibit ion and excitat ion in the light -evoked currents of X type retinal ganglion cells. J Neurophysio l 80, 2975-2990. Cope, D.W., Hughes, S.W., and Crunelli, V. (2005) GABAA Receptor-Mediated Tonic Inhibit ion in Thalamic Neurons. J Neurosci 25:4733-4742. Fee, M.S., Mitra, P.P., and Kleinfe ld, D. (1996). Automat ic sorting of mult iple unit neuronal signals in the presence of anisotropic and non-Gaussian variability. J Neurosci Methods 69, 175-188. Gaudry, K.S., and Reinagel, P. (2007a). Benefits of contrast normalizat ion demonstrated in neurons and model cells. J Neurosci 27, 8071-8079. Gaudry, K.S., and Reinagel, P. (2007b). Contrast adaptation in a non-adapting LGN model. J Neurophysio l. Jin, X., Chen, A.H., Gong, H.Q., and Liang, P.J. (2005). Information transmission rate changes of retinal ganglion cells during contrast adaptation. Brain Res 1055, 156-164. Kaplan, E., and Benardete, E. (2001). The dynamics of primate retinal ganglion cells. Prog Brain Res 134, 17-34. Kaplan, E., Purpura, K., and Shapley, R.M. (1987). Contrast affects the transmission of visual informat ion through the mammalian lateral geniculate nucleus. J Physio l 391, 267- 288. Kara, P., Reinagel, P., and Reid, R.C. (2000). Low response variability in simultaneously recorded retinal, thalamic, and cortical neurons. Neuron 27, 635-646. Keat, J., Reinagel, P., Reid, R.C., and Meister, M. (2001). Predict ing every spike: a model for the responses of visual neurons. Neuron 30, 803-817. Kim, K.J., and Rieke, F. (2001). Temporal contrast adaptation in the input and output signals of salamander retinal ganglion cells. J Neurosci 21, 287-299. Kremers, J., Silveira, L.C., and Kilavik, B.E. (2001). Influence o f contrast on the responses o f marmoset lateral geniculate cells to drift ing gratings. J Neurophysio l 85, 235-246. Lesica, N.A., Jin, J., Weng, C., Yeh, C.I., Butts, D.A., Stanley, G.B., and Alonso, J.M. (2007). Adaptation to Stimulus Contrast and Correlat ions during Natural Visual Stimulat ion. Neuron 55, 479-491. Levitt, J.B., Schumer, R.A., Sherman, S.M., Spear, P.D., and Movshon, J.A. (2001). Visual response properties of neurons in the LGN of normally reared and visually deprived macaque monkeys. J Neurophysio l 85, 2111-2129. Passaglia, C.L., Enroth-Cugell, C., and Troy, J.B. (2001). Effects of remote stimulat ion on the mean firing rate of cat retinal ganglion cells . J Neurosci 21, 5794-5803. Renteria, R.C., Tian, N., Cang, J., Nakanishi, S., Stryker, M.P., and Copenhagen, D.R. (2006). Intrinsic ON responses o f the retinal OFF pathway are suppressed by the ON pathway. J Neurosci 26, 11857-11869. Shapley, R.M., and Victor, J.D. (1978). The effect of contrast on the transfer properties of cat retinal ganglion cells. J Physio l 285, 275-298. Evidence for an additive inhibitory component of contrast adaptation 20 DRAFT 5/19/09 Gaudry and Reinagel Shou, T., Li, X., Zhou, Y., and Hu, B. (1996). Adaptation of visually evoked responses of relay cells in the dorsal lateral geniculate nucleus of the cat fo llowing prolonged exposure to drift ing gratings. Vis Neurosci 13, 605-613. Smirnakis, S.M., Berry, M.J., Warland, D.K., Bialek, W., and Meister, M. (1997). Adaptation of retinal processing to image contrast and spat ial scale. Nature 386, 69-73. Solomon, S.G., Peirce, J.W., Dhruv, N.T., and Lennie, P. (2004). Profound con trast adaptation early in the visual pathway. Neuron 42, 155-162. Troy, J.B., and Robson, J.G. (1992). Steady discharges of X and Y retinal ganglion cells of cat under photopic illuminance. Vis Neurosci 9, 535-553. Victor, J.D. (1987). The dynamics of the cat retinal X cell centre. J Physio l 386, 219-246. Wang X, Wei Y., Vaingankar, V., Wang, Q., Koepsell, K., Sommer, F.T. and Hirsch, J.A. (2007). Feedforward excitation and inhibit ion evoke dual modes of firing in the cat 's visual thalamus during naturalist ic viewing. Neuron 55:465-78. Wassle, H., Schafer-Trenkler, I., and Voigt, T. (1986). Analysis o f a glycinergic inhibitory pathway in the cat retina. J Neurosci 6, 594-604. Zaghloul, K.A., Boahen, K., and Demb, J.B. (2003). Different circuits for ON and OFF retinal ganglion cells cause different contrast sensitivit ies. J Neurosci 23, 2645-2654. Zaghloul, K.A., Boahen, K., and Demb, J.B. (2005). Contrast adaptation in subthresho ld and spiking responses of mammalian Y-type retinal ganglion cells. J Neurosci 25, 860- 868. Acknowledgements This work was supported by NEI R01-EY016856; KG was supported by an NSF graduate research fellowship. Evidence for an additive inhibitory component of contrast adaptation 21
1807.02103
2
1807
2018-09-30T14:28:34
Making Sense of Consciousness as Integrated Information: Evolution and Issues of IIT
[ "q-bio.NC" ]
The purpose of this article is to provide an overall critical appraisal of Integrated Information Theory(IIT) of consciousness. We explore how it has evolved and what problems are involved in the theory. IIT is a hypothesis that consciousness can be explained in terms of integrated information. It argues that a number of fundamental properties of experience can be properly analyzed and explained by physical systems' informational properties. Throughout the last decade, there have been many advances in IIT's theoretical structure and mathematical model. In addition, like all hypotheses in the field of science of consciousness, IIT has given rise to several controversies and issues. In this context, a critical survey for IIT is urgently needed. To this end, we first introduce fundamental concepts of IIT and related issues. Thereafter, we discuss major transitions IIT has been through and point out related intra-model issues. Finally, in the last section, some theoretical, extra-model issues involved in IIT's principles are presented. The article concludes by suggesting that, for the sake of future development, IIT should more seriously take metacognitive accessibility to experience.
q-bio.NC
q-bio
Making Sense of Consciousness as Integrated Information: Evolution and Issues of IIT Kyumin Moon1 and Hongju Pae2 1Dept. of Philosophy, Seoul National University [email protected] 2Interdisciplinary Program in Cognitive Science, Seoul National University [email protected] Abstract The purpose of this article is to provide an overall critical appraisal of Integrated Infor- mation Theory(IIT) of consciousness. We explore how it has evolved and what problems are involved in the theory. IIT is a hypothesis that consciousness can be explained in terms of integrated information. It argues that a number of fundamental properties of experi- ence can be properly analyzed and explained by physical systems' informational properties. Throughout the last decade, there have been many advances in IIT's theoretical structure and mathematical model. In addition, like all hypotheses in the field of science of conscious- ness, IIT has given rise to several controversies and issues. In this context, a critical survey for IIT is urgently needed. To this end, we first introduce fundamental concepts of IIT and related issues. Thereafter, we discuss major transitions IIT has been through and point out related intra-model issues. Finally, in the last section, some theoretical, extra-model issues involved in IIT's principles are presented. The article concludes by suggesting that, for the sake of future development, IIT should more seriously take metacognitive accessibility to experience. keywords: Integrated Information Theory, the science of consciousness, consciousness, ex- perience, qualia, panpsychism, metacognition 8 1 0 2 p e S 0 3 ] . C N o i b - q [ 2 v 3 0 1 2 0 . 7 0 8 1 : v i X r a 1 Introduction Integrated Information Theory of consciousness (IIT) is a hypothesis that consciousness can be explained in terms of integrated information. Among other theories, IIT might be one of the most interesting -- but also controversial -- hypotheses in the field of science of consciousness. IIT is sug- gested as a principled theoretical framework with the explanatory and predictive power. It argues that a number of fundamental properties of experience can be properly analyzed and explained by physical systems' informational properties. Further, IIT claims that this information-centered, mathematical framework would shed some light on many clinically difficult and ambiguous cases. 1 For this unique approach, IIT has consistently attracted considerable scholarly attention from neuroscientists, information theorists, and even physicists for over a decade. Throughout this period, there have been many advances in IIT's theoretical structure and mathematical model. Furthermore, like all hypotheses in the field of science of consciousness, IIT has also given rise to several controversies(Cerullo 2015; Horgan 2015). In this context, there is an urgent need for a critical survey for IIT that would address the following questions: What are the essentials of IIT? What has been changed and what has remained? And what problems can emerge against it? In the present article, we attempt to provide an overall critical appraisal of IIT. For this critical review, we will introduce core concepts, major transitions of IIT and related intra-model issues that are rooted in the mathematical formulation. Then, several theoretical extra-model issues involved in IIT's principles, which are not directly due to the mathematical model, are outlined.1 This article concludes by suggesting that, for the sake of future develop- ment, IIT should more seriously take metacognitive accessibility to experience. 2 Core Concepts of IIT Since IIT attempts to explain how conscious experience arises from physical substrates, there are several explanatory concepts describing this bottom-up process. While IIT has kept updating its version from 1.0 to 3.0(Tononi 2001, 2004, 2008, 2012; Balduzzi and Tononi 2008, 2009; Oizumi, Albantakis, and Tononi 2014), those core concepts remain to be fundamentals of the theory throughout all versions. Yet, despite their significant roles in the framework of IIT, the core concepts have not been clearly cashed out. To amend the situation, in what follows, we explain why those concepts are important in IIT and point out some related issues. While our characterization strongly reflects the view of the current version of IIT, providing a summary of IIT 3.0 is not the main concern in this section. Rather, following descriptions concern several central notions that persist regardless of versions. 2.1 Mechanisms, states, connections, and repertoires The central focus of IIT is on the physical substrates of experience and their causal structures. IIT analyzes candidate physical substrates of experience in a bottom-up manner; physical elements, which can causally interact with each other, are under consideration. Any set of elements can be considered as a mechanism. Furthermore, any set of mechanisms can be thought of as a higher- 1. One of the reviewers has expressed some worries about the general structure of the manuscript of this article. The reviewer has pointed out that considering the critical motivations of this article, the manuscript contains too many technical reviews, which is not necessarily needed to justify theoretical criticisms developed in the latter of this article. The reviewer further advised that to clarify the common thread of the article, it would be better to focus on critical points and reduce technical details that are not directly relevant to the purpose of the article. Although we tried to shorten and revise the technical review, we believe at least some of the presentations of technical details are required, since they are necessary to understand the very points we made. In order to trace the version updates of IIT and reveal some problematic consequences of those updates, reviews of technical details are somewhat inevitable. Moreover, this article is intended to be an overall critical appraisal, covering not only mathematical models but also theoretical backgrounds of IIT. Therefore, despite of lengthiness, we choose to stick to the initial structure of the article. 2 order mechanism or a system of mechanisms (in short, system). The system is composed of elements so that the system itself also can be a mechanism or a set of elements. On the other hand, causal structures of physical substrates are analyzed by two central notions of IIT; mechanisms, or systems, can be in a state, which corresponds to outputs of their elements. For instance, if three elements -- A, B, and C -- with the binary output 1 or 0 compose a mechanism, and these element's outputs are respectively 1, 0, and 0, the state of the mechanism ABC is represented as 100(see Oizumi, Albantakis, and Tononi 2014, Figure 1A). Further, such mechanism in a state can have a connection, which corresponds to a set of causal connections among elements of the mechanism(Balduzzi and Tononi 2009).2 For example, if causal connections c1, c2, c3, and c4 are given, there might be a set of connections, such as as {c1, c2}, {c1, c3}, {c1, c2, c3} or {c1, c2, c3, c4}, etc. Any causal relationships could be characterized as a connection, such as synapses between neurons, which could be ideally represented as logic gates with simple computational functions. From states and connections of the mechanism, one can have repertoires. A repertoire is defined as a probability distribution to possible states of the mechanism. In IIT, the causal struc- ture of the mechanism must be known a priori.3 When the state and connection of a mechanism are given at time t, one can infer which past or future states of which mechanism -- including the mechanism itself -- could be causes or effects of the given state of the mechanism, and how much probabilities would be distributed to each possible cause or future effect states. Therefore, these probability distributions are probabilistic expressions of how the mechanism's particular state could cause or be caused by a certain mechanism's past or future states. In this sense, the mechanism in the state specifies repertoires, or its possible causes and effects. The notions of mechanisms, states, connections, and repertoires are the very fundamentals in IIT. Without these concepts, calculating information from a mechanism's causal structure is not possible. As explained above, repertoires are derived from states and connections of the mechanism. Furthermore, as we will see in Section 2.2, the very concept of information is formally defined by repertoires and related notions. The concepts of mechanisms, states, connections, and repertoires tie causation and information together and enable us to calculate how much information is generated from the causal structure of the mechanism. In part, this is the reason why they survived several updates so far. These notions also provide IIT with a quite liberal view about possible physical substrates of consciousness. None of these notions tells about what kind of materials should be considered as a candidate for the physical base of experience. Therefore, when something has its state and connection and specifies repertories, it can be at least considered regarding if it produces experience. Given that mechanisms or systems in a state are not limited to biological substrates, 2. In (Balduzzi and Tononi 2009), the term 'submechanism' or 'mechanism' was originally used to refer to sets or subsets of causal connections among elements. However, this use of the term causes a serious confusion, as 'mechanism' is also used in IIT to refer to sets or subsets of elements that causally interact. In order to avoid possible confusions, in the present paper, we use the term 'connection' instead of 'submechanism' or 'mechanism'. 3. This a priori known causal structure can be mathematically described by the backward and forward Tran- sition Probability Matrix (TPM) of the set of elements under consideration. While this requirement of a priori given causal structure is usually not explicitly presented in literature, it is important and intrinsic to IIT. Thanks for the reviewer who clarified this point. 3 chemical structures such as silicon chips can be legitimate candidates for the physical base of consciousness. Thus, under the framework of IIT, the question "Is this cellular phone conscious?" is not a category-mistaken question that should be a priori rejected. As far as the cellular phone can be considered as a "system of mechanisms in a state", we can at least consider the possibility of its consciousness. In principle, anything that has its states and connections can be a mechanism, and any mechanism can be a possible candidate for a conscious mechanism(Tononi and Koch 2015). However, such liberalism comes at a price. While the notions of mechanism, states, con- nections, etc., do not limit the kinds of physical substrates of experience, they do not limit the levels of physical substrates either. Said differently, those basic concepts do not identify in which spatio-temporal grains we should find physical substrates of consciousness. Technically, there are elements, mechanisms, systems, state, and connections at each level of the grain; basic particles in microphysical interaction compose quantum mechanisms in a quantum state. Molecules in chemical bonding constitute chemical mechanisms in a chemical state. Neurons connected with synapses make neuronal mechanisms in a neuronal state. Among these levels, which mechanism should be taken as the origin of consciousness? The same applies to macro-levels. For exam- ple, in IIT, there appears to be no principled reason not to take China as a single mechanism in a state, composed of causally interacting Chinese people(Schwitzgebel 2012). Indeed, the problem of finding a proper spatio-temporal grain of consciousness has been admitted by IIT theorists themselves(Tononi 2008, 2012; Oizumi, Albantakis, and Tononi 2014). We think that the problem already lies in the center of the basic notions of IIT, leaving theoretical loose ends. It can be argued that the problem of spatio-temporal grains has already addressed in the current version of IIT.4 Applying the exclusion postulate introduced in IIT 3.0, proponents of the theory may argue that the appropriate spatio-temporal grains are ones that have maximum intrinsic cause-effect power, which is quantified by the highest value of integrated conceptual information(Tononi 2012; Tononi et al. 2016). It is nonetheless possible that there are multiple highest values of integrated conceptual information across different spatio-temporal levels. For instance, if a certain part of the cortico-thalamic system, which is at the macro-spatial level, and a few numbers of neurons in V1, which is at the micro-spatial level, produce the same highest values of integrated information at the same time, then which level should be chosen as the level where experience arises? Unless principled solution being suggested, the problem of proper spatio-temporal grains would remain. 2.2 Intrinsic and causal information According to IIT, an amount of information generated by a mechanism is calculated from reper- toires. This calculation is performed by measuring the distance between the unconstrained and constrained repertories. For the past or future state, IIT supposes the unconstrained repertoire as a probabilistic base. Given the system's causal structure, the repertoire is unconstrained in that such uncertainty is not constrained yet by the given state of the mechanism. Using Bayes' theorem, one can infer the constrained repertoire from the given state of the mechanism. It is 4. One of the reviewers reminded us that IIT theorists already dealt with this problem of proper spatio-temporal granularity in (Tononi et al. 2016). 4 this distance between unconstrained and constrained repertoires that is defined as information throughout all versions of IIT.5 The crucial point here is that those repertoires involved in information should be inferred from mechanisms within a considered system.6 To calculate repertoires specified by the mechanism in the state, one must consider past or future states of mechanisms within the system under consideration. No mechanism outside of the considered system should be taken into account. For example, to calculate the amount of information generated by the mechanism mentioned in Section 2.1, ABC in 100, one should consider mechanisms only within a considered system; suppose that with the mechanism of ABC, an element D constitutes a certain system under consideration. Other elements, such as E and F , are out of the considered system. Then, according to IIT, ABC in 100 cannot specify repertoires of mechanisms such as E, EF , or even AE, AF , ABE, ABF , ABCE, ABCF . It only specifies repertories of mechanisms A, B, C, AB, AC, BC, ABC, or AD, ABD, ABCD, BD, ACD, CD and so on. Those repertoires would represent possible causes or effects of ABC's being in 100 that are in the considered system with their probabilities. In short, mechanisms in a certain system under consideration only can specify repertoires of mechanisms within that system. In this specific sense, in IIT, repertoires specified by the mechanism in a state express intrinsic causal power of the mechanism. As repertoires represent the intrinsic causal power of the mechanism, information in IIT is Information generated by the mechanism is measured as the essentially intrinsic and causal. distance between repertoires. Of note, these repertoires involve nothing external to the system. They solely depend on possible causes or effects within the system. Therefore, information is intrinsic to the system in that it does not require anything external to the system. In addition, information has nothing to do with input/output signals that can be detected only by the external observer. Rather, it is about causes and effects that can be detected only from the system's own intrinsic perspective(Tononi 2008, 2012; Oizumi, Albantakis, and Tononi 2014). Moreover, given that repertoires specify possible causes or effects and their probabilities, information produced by the mechanism is causal. This is why IIT repeatedly emphasizes the notion of information as "differences that make a difference"(Bateson 1972). In IIT, for instance, the mechanism in a state specifies which past states of a certain mechanism ("differences") would likely to cause the mechanism's being in that state ("a difference"). This further implies that only something that can be selectively caused or cause can produce information. This intrinsic and causal notion of information is the hallmark of IIT, which distinguishes IIT from other information theories: anything informative has an intrinsic causal power, and anything intrinsically causal has information. This intrinsic and causal nature of information is directly inherited by the most central concept in IIT, integrated information. Although integrated information is defined in a sophisticated manner, in so far as it is information, it also should be intrinsic and causal. 5. For further detail on this calculation, see Section 3.2 and (Oizumi, Albantakis, and Tononi 2014). 6. As one of the reviewers clearly pointed out our inconsistent terminology in the original manuscript, we could correct this paragraph. The notion of "considered system" or "system under consideration" is explicitly introduced as candidate set in IIT 3.0. A candidate set is a set of elements under consideration; if certain elements are not included in the candidate set, those elements are considered as external noise, even if they still are part of the whole system. For further details on candidate set, see (Oizumi, Albantakis, and Tononi 2014), Figure 1A. 5 The intrinsic and causal information is fundamental to IIT in that it determines what kind of information the theory deals with. Although intrinsic and causal information constitutes one of the unique aspects of the theory, it also brings some problems concerning the function of consciousness. Simply put, intrinsic and causal information does not involve anything outside of the system. By definition, integrated information has nothing to do with causal inputs/outputs of the system either. This intrinsicness renders integrated information irrelevant to the functions of the system. In fact, as we will see in Section 4.2., IIT theoretically designs functional zombie systems, which share all the input- output relations with systems with highly integrated information. This suggests that integrated information is nearly irrelevant to functions of the system; integrating information has no nec- essary bearing on the system's functioning(Schwitzgebel 2014). In the sections below, we will see that IIT identifies integrated information and consciousness. If so, integrated information's functional irrelevancy would be directly transferred to consciousness. For instance, IIT implies that, at least in principle, there can be perfect functional equivalents of us that are unconscious. It is at least theoretically possible that we do whatever we are doing without consciousness. Such state makes the 'use' of consciousness as mysterious. Moreover, adaptive benefits of having ex- perience also become doubtable; whatever adaptive function experience provides, there is always a possible scenario that it might have been evolved without experience. This risk of functional irrelevancy of experience has been already rooted in the intrinsic nature of information in IIT. 2.3 Integrated information and complex The notion of integration first stems from the phenomenological aspects of experience: "[p]heno- menologically, every experience is an integrated whole, one that means what it means by virtue of being one, and which is experienced form a single point of view"(Tononi 2012, p.295). To be a physical underpinning of such integrated, unified experience, what should a mechanism be like? Here, IIT suggests one of its thought experiments: let's compare a highly informative, but unconscious mechanism and a conscious mechanism. For example, what is the difference between a conscious brain and an unconscious digital camera that consists of thousands of photodiodes? According to the IIT, the most significant difference is that while the former is causally integrated, the latter is not(Tononi 2012). Causal interactions within the brain are so highly integrated with each other that, once they are fragmented, the whole brain's performance might break down. This thought experiment on the camera model suggests that producing information is not sufficient for a mechanism to generate consciousness. Even if the mechanism is equipped with complicated connections and distinguishes vast repertoires, if its elements are not integrated into a single mechanism, the mechanism cannot give rise to experience. As a mechanism with a causal structure produces intrinsic information, one with integrated causal structure generates integrated information. The integrated information is integrated in the sense that, as a whole, the mechanism generates more information than the sum of its parts. Said differently, it is information produced only from the mechanism as a whole. By definition, the integrated information of the system is irreducible to its parts. Therefore, according to IIT, the amount of integrated information generated by the mechanism is calculated by partitioning the system by disconnecting the connections between the mechanisms. That is, if the information 6 disappears by partitioning, it would be the information generated by the mechanism as a whole, not by individual parts. The informational difference between the mechanism as a whole and the system's partitions' mechanism is defined as integrated information.7 Nonetheless, considering that there are many possible ways of how the mechanism is partitioned, it becomes crucial to decide which partition should be used in calculating integrated information. IIT chooses the partition which causes the least loss of information, which is called minimum information partition(MIP). Finally, depending on the level of calculation, the calculated values of integrated information are represented as Φ or φ. Based on the integrated information, a complex is defined: roughly put, parts of the system producing integrated information can be considered as complexes.8 As we will see in Section 2.4, IIT posits the identity between consciousness and integrated information; complexes in the system directly contribute to consciousness by integrating information. Technically, only com- plexes should be regarded as physical substrates of conscious experience, and they deserve to be called a 'locus' of consciousness. Despite a significant change concerning whether the overlapping or inclusion among complexes is possible, IIT maintains that a system can be condensed into multiple complexes. Finding such complexes in the system is the main focus of IIT on defining the local and temporal origin of consciousness. It should be emphasized that the notions of integrated information and complex provide possible explanations for some fundamental properties of experience. In calculating integrated information, nothing outside of the complex matters. How much information is generated by the complex, or how much information is lost by MIP, is purely intrinsic to the complex. This in- trinsicness of integrated information accounts for why experience is essentially intrinsic. In other words, experience is integrated information, and integrated information is intrinsic. Therefore, experience is intrinsic. This characteristic of experience can be also noted as privacy: "Since inte- grated information is generated within a complex and not outside its boundaries, the experience is necessarily private and related to a single point of view or perspective"(Tononi 2008, p.295). Appealing to the central concepts such as integrated information and complex, IIT appears to open up the prospects of making sense of essential features of experience. While integrated information is undeniably the key concept in the theory, it also creates a 7. This idea of integration might be closely related to the notion of synergy information proposed by Virgil and Koch(Virgil and Koch 2014). 8. Technically, how complexes are defined depends on which version of IIT is taken. As one reviewer noted, a complex is defined as a set of elements which produces a local maximum of integrated conceptual information on a system level, which quantified by Φmax in IIT 3.0. While this is true, strictly speaking, notions such as integrated conceptual information, local maxima, and mechanism-system distinction were explicitly introduced since IIT 3.0. One cannot find any of these before IIT 3.0. In the IIT 2.0, complexes are defined differently, as sets of elements that produce integrated information. As we have noted at the beginning of Section 2, our purpose is not briefly presenting the current version of IIT. (If it was, we would not cite any of the literature based on IIT 2.0 framework, including (Tononi 2008) or (Balduzzi and Tononi 2008), (Balduzzi and Tononi 2009). The focus is on the core concepts that play essential roles throughout all versions of IIT. Thus, until the end of Section 2, we temporarily choose to ignore conceptual differences among various versions and use some terms very loosely. In Section 2, for instance, "integrated information" covers integrated information in IIT 2.0 as well as integrated conceptual information and maximally integrated conceptual information in IIT 3.0. Accordingly, Φ can refer both Φ and Φmax. 7 problem which renders the application of IIT to real systems practically intractable. As specified above, calculating integrated information involves finding MIP, which requires creating all possi- ble partitions and measuring all the informational difference between the non-partitioned and the partitioned. With the growth of the number of the elements organizing the system, it becomes obvious that the amount of computation will dramatically increase. Consequently, one faces a serious combinatorial explosion in finding MIP.9 Due to this computational burden, applying IIT to neural substrates or artificial robots is currently infeasible. At the current stage of the theory, since direct empirical data supporting IIT are unavailable, researchers have tried to find efficient algorithms for finding MIP(Kitazono, Kanai, and Oizumi 2018; Hidaka and Oizumi 2018), or to develop approximations or proxy measures of Φ.10 The absence of experimental validity is a decisive disadvantage for IIT to become a solid theory claim to the science of consciousness. 2.4 Identity between consciousness and integrated information As mentioned in Section 2.3, IIT identifies consciousness with integrated information in the first place. Particularly in IIT, levels of consciousness are identified with quantities of integrated information, while qualities of consciousness are identified with informational structures derived from integrated information. From these identifications, IIT attempts to account for both how conscious a system is and how it feels. According to IIT, a level of consciousness is nothing but an amount of integrated information. Therefore, one can know how conscious the system is by calculating the amount of integrated information produced by that system.11 Consciousness is not all-or-nothing. Rather, as shown by the experience of falling asleep or that of anesthesia, consciousness is a matter of graded levels. IIT claims that a level of consciousness can be quantitatively measured by a value of integrated information, which is referred to Φ. Since this 'quantifying consciousness' has drawn considerable scholarly attention, many IIT studies thus far have been dedicated to finding correlations between 9. One reviewer mentioned that the levels of MIP need to be distinguished: the MIP on the level of small phi and the MIP on the level of conceptual information big Phi. It is of course true that these are two different forms of partitions that are respectively applied to the levels of mechanism and system. However, as clarified in footnote 8, such notions are restricted to the current version. They cannot be applied regardless of versions. 10. A large variety of modified Φ has been proposed as an estimation for Φ. ΦE is modulated by the Markovian discrete system and can be applied to continuous time series data(Barrett and Seth 2011), and Φ∗ is modu- lated by substituting the notion of decoding perspective of information that facilitates the overall computation procedure(Oizumi et al. 2016). However, they do not contain the main theoretical updates of the IIT, such as cause-effect information and the distinction between φ and Φ. For IIT 3.0, Marshall, Gomez-Raminez, and Tononi have proposed State Differentiation (SD) as a proxy measure of Φ, which is much easier to draw out from experimental data than the original Φ(Marshall, Gomez-Ramirez, and Tononi 2016). Still, it leaves the degree of integration being not properly measured. It is also insufficient to assume that SD functions are a complete form of measure in that it is applied to cellular animats; it is still analyzing the toy problem of the causal system. Recently, Tegmark has proposed several kinds of modified Φ through various definitions of informational distance and by normalizing integrated information using diverse techniques(Tegmark 2016). 11. Rich integration of neuronal connection is widely known as a major feature of the cerebral cortex, based on the anatomical structure of the brain. The fact that the brain cortex is constructed as a complicated neuronal network can explain how consciousness arises from such anatomical structure and why there exists such direct correlation between Φ and consciousness. 8 levels of consciousness and corresponding Φ values. Some of the results from analyzing EEG data and computer simulations suggest that Φ can be a reliable measure of consciousness(Tononi 2008). Indeed, the idea of the possibility to measure consciousness in a quantitative manner alludes to the science of conscious experience. It is the identification of levels of consciousness and Φ values that make this idea possible. On the other hand, IIT claims that a quality of experience is just an informational struc- ture assessed from integrated information. This informational structure can be represented as a geometrical shape in the multidimensional space that "completely and univocally specifies the quality of experience"(Tononi 2008, p.224). If so, how the system feels can be known by deriving what shape is represented by the integrated information it generates. Each version of IIT pro- vides sophisticated procedures for illustrating shapes like polytopes on multidimensional space from given integrated information. These shapes specify informational relationships generated by complexes. It appears to be obvious that, if it is successful, this geometrical representa- tion provides useful tools for analyzing qualities of experience. Qualities of experience have several fundamental aspects to be explained, such as similarities and differences, richness, het- erogeneities, as well as compositional structures. Once qualities of experience are identified with shapes assessed from integrated information, those fundamental aspects can be explained by analyzing geometrical characteristics of shapes in the multidimensional space. This explanatory potential of geometrical approach might be the most distinctive part of IIT; "what it is like to experience" could be explained by the "geometry of integrated information"(Balduzzi and Tononi 2009). The identity between consciousness and integrated information has further implications. First, in clinical contexts, quantifying consciousness by Φ value might play a significant role in treating pathological cases of patients in coma or vegetative state. As the locked-in syndrome case suggests, how to judge whether or not one is conscious has been an extremely controversial issue. However, if Φ is indeed a level of consciousness, we have a simple answer: a patient is conscious only when his or her brain generates non-zero Φ. This answer immediately leads to a liberal approach to the consciousness of non-humans. For animal consciousness, animals can be conscious if they generate integrated information at all. The same applies to artificial con- sciousness, as there is no reason to a priori exclude the possibility that artificial intelligence can be conscious. The only thing that matters upon consciousness is whether or not the candidate system produces non-zero Φ(Tononi and Koch 2015). Second, the idea of qualities of experience as geometrical shapes appears to entail that experience is substrate-independent. According to IIT's geometrical approach, systems with different states and connections can produce the same informational structures(Balduzzi and Tononi 2008, 2009). From the assumption that qualities of experience are nothing but informational structure geometrically depicted from integrated information, it follows that qualities of experience and their physical substrates can come apart. This substrate-independence might account for, at least partially, why consciousness seems non- physical(Tegmark 2017). In this respect, many theoretically and practically promising predictions and explanations come from the identification of consciousness and integrated information. However, the notion of consciousness as integrated information also raises some perplexing issues. Since the theory identifies Φ into a level of consciousness, IIT must ascribe experience 9 to seemingly unconscious systems. Surprisingly, according to IIT, a photodiode with the binary states on/off is minimally conscious, because it produces non-zero Φ.12 What is more, a lattice structure composed of a single kind of logic gates can be highly conscious, even more conscious than that of a human. In this way, IIT predicts that the functional zombie is empirically possible in principle(Oizumi, Albantakis, and Tononi 2014); even if the two systems are functionally equivalent, there can be a situation when one generates Φ, but the other does not(see Oizumi, Albantakis, and Tononi 2014, Figure 21). The theoretical development of IIT is not without a sense of irony; core concepts promising ample explanatory and predictive potentials are also bringing counterintuitive and problematic defects to the theory. In the remainder of this paper, these intriguing issues will be analyzed in further detail. 3 Major Transitions in IIT While the core ideas of IIT are more or less preserved, it has undergone several significant revi- sions through the updates from its prototype to the very latest version. These revisions made the framework of IIT more theoretically and technically articulated. Not only its theoretical structure but also the details of its mathematical model have been changed. Therefore, under- standing how IIT has acquired its current form requires a deeper analysis. Axioms, postulates, and the notion of purviews have been introduced, the concept of information has been revised, the distance matrix for repertoires has been changed, and levels of information integration have been divided. In what follows, we explain what these changes are about and how they affect IIT. 3.1 From thought experiments to systematic formulation: Phenomenological axioms and ontological postulates As occasionally mentioned in the literature, what consciousness is and what a physical system should be to generate consciousness have never been explicitly described until IIT has devel- oped into its latest version. The fundamental properties of experience were taken for granted by appealing to phenomenology, and the required properties for physical systems to produce experi- ence were merely motivated or suggested by a number of thought experiments. The photodiode and camera thought experiments were introduced from the early version of IIT(Tononi 2004, 2008), and the Internet thought experiment was added during the updates to IIT 3.0(Tononi 2012). Based on the idea that conscious experience is specific in a particular way, the photo- diode thought experiment motivates that physical systems must specify its possible causes or effects to generate consciousness. The digital camera thought experiment suggests that physical systems must be causally integrated since it appears that experience is unified and integrated. By contrasting information of the Internet and that of experience, the Internet thought experiment speculates that information produced by physical systems must be maximally integrated. While 12. Again, the reason why a photodiode should be treated as minimally conscious can be different according to versions of IIT. In IIT 2.0, the photodiode is minimally conscious, since it produces one bit of integrated information. However, in IIT 3.0, it is so because it generates non-zero integrated conceptual information. Although this distinction is important, for the purpose of Section 2, we do not deal with differences among versions. See footnote 8. 10 these thought experiments are interesting in themselves and might be helpful to understand the motivations behind the theory, they never clearly argued for or even specified the fundamental features of the essential properties of experience. In IIT 3.0, the situation has changed. Now, the fundamental properties of consciousness and requirement for physical systems are explicated and posited in the very beginning of the theoretical formulation. First and foremost, phenomenological axioms are introduced; these axioms are phenomenological in the sense that they are all concerned with the fundamental properties of experience. Each of five axioms corresponds to each of the essential properties of experience. The existence axiom states that consciousness exists. The composition axiom says that it is compositional. The information axiom states that it is informative. The integration axiom claims that it is integrated. Finally, the exclusion axiom says that one consciousness excludes another consciousness(Tononi 2012; Oizumi, Albantakis, and Tononi 2014). These properties mentioned in the axioms are supposed to be fundamental, as any experience must have them. Next, corresponding to the phenomenological axioms, ontological postulates are posited: these postulates are ontological in that they prescribe what mechanisms should generate conscious- ness. There are five postulates which lay parallel to each axiom. The existence postulate says mechanisms in a state must exist. The composition postulate says that mechanisms must be structured. The information postulate claims that mechanisms must produce information by specifying selective possible causes and effects within the system. The integration postulate states that mechanisms must integrate information. Finally, the exclusion postulate says that mechanisms must generate only the maximally integrated information(Tononi 2012; Oizumi, Al- bantakis, and Tononi 2014; Tononi and Koch 2015). As in phenomenological axioms, properties mentioned in ontological postulates are essential and necessary for every physical mechanism to generate consciousness. Moreover, the latter three postulates -- information, integration, and exclusion -- are applied to the two different levels of calculation; mechanisms and systems of mechanisms. Contents of postulates vary through the system depending on which level they are applied to.13 In virtue of these axioms and postulates, IIT becomes a top-down and theory-driven ap- proach, rather than as a bottom-up and experiment-driven approach to consciousness; the set of axioms and postulates comes first and later comes the mathematical model. Empirical exper- iments can be designed and conducted only under the models and theories. Consequently, by declaring its axioms and postulates, IIT can clarify both its theoretical framework on modeling the consciousness and the experiments. However, while clarification is one thing, justification is another thing. The introduction of the axioms and postulates raises a number of questions. First, on what ground must phenomenologi- cal axioms be accepted? That is, why should those axioms be considered axiomatic? The axioms themselves appear to be based on phenomenological intuitions or introspection. The axioms are "assumed to be self-evident from the intrinsic perspective of a conscious entity"(Oizumi, Alban- takis, and Tononi 2014, Supplementary 1, p.1). However, what if such intuitions or introspections are wrong? Moreover, how can each ontological postulate follow from each phenomenological ax- 13. For further details on ontological postulates, see (Oizumi, Albantakis, and Tononi 2014). 11 iom? Though the postulates are strictly parallel to axioms, there seems to be an unbridged gap between them. For instance, it is not clear how we can draw the information postulate; it is not clear if the mechanisms should specify selective causes and effects within the system from the information axiom, which states that consciousness is informative. Therefore, the rationales for positing phenomenological axioms and ontological postulates remain controversial. Recently, Bayne addresses precisely the axiomatic foundations of IIT(Bayne 2018). Bayne argues that some of the phenomenological axioms are not self-evident, and others seem to be self-evident but fail to practically or theoretically constrain the theory of consciousness at all. He suggests that IIT would be on firmer ground if it adopts what he calls 'natural kind approach'(Bayne 2018). While the verdict may still be out, it appears that the axiomatic approach and seemingly following postulates are not secured as they seem.14 3.2 From effective information to cause-effect information: New information and its metric The revision of the notion of information might be one of the most significant developments during the updates of IIT. In the early versions of the theory, information generated by a system was defined as effective information(ei)(Tononi 2008). There were two kinds of repertoires: a potential repertoire, which is a probability distribution of the past states when no current state of the system is known, and an actual repertoire, which is a probability distribution of the past states when a particular current state of the system is known. One can measure the distance between the potential and the actual repertoires by applying Kullback-Leibler Divergence(KLD), and such distance could be thought of a sort of relative entropy. This distance or relative entropy directly equals to the effective information. In IIT 3.0, however, a new form of information is introduced: cause-effect information(cei)(Tononi 2012; Oizumi, Albantakis, and Tononi 2014). The cause-effect information differs from effective information in many important aspects. First, unlike ei, cei involves the system's past and future(Tononi 2012; Oizumi, Albantakis, and Tononi 2014). In calculating ei, potential and actual repertoires are only of the past states of the system. In calculating cei, however, repertoires concern both the past and future states of the system. These repertoires can be thought of as probabilistic expressions of how the current state of the mechanism would be caused by the past states of the system and how it would cause the future states of the system. On one hand, there are unconstrained past repertoire and cause repertoire. The former is a probability distribution of the past states of a certain mechanism of the system when the current state of the given mechanism is not known. It always produces maximum entropy of past states. The latter is defined as a probability distribution of the past states of a certain mechanism of the system when the current state of the given mechanism is known. On the other hand, there are unconstrained future repertoire and effect repertoire. The former is a probability distribution of the future states of a certain mechanism when the current state of the given mechanism is perturbed in every possible way. The latter repertoire is a probability distribution of the future states of a certain mechanism of the system when the current state of the given mechanism is known. The current state of the given mechanism 14. We the authors gratefully thank the reviewer who recommended the recent literature. It was truly helpful to strengthen our argument. 12 specifies cause and effect repertoires of a certain mechanism of the system. It should be noted that unconstrained past and future repertoires and cause and effect repertoires are calculated independently of each other. Therefore, repertoires should be calculated twice in calculating cei; therefore, while calculating ei requires only two repertoires, four repertoires are required to calculate cei. Second, while ei concerns only the states of the given system itself, cei can involve the past/future states of mechanisms other than the given mechanism(Tononi 2012; Oizumi, Alban- takis, and Tononi 2014). For ei, only the past states of the same system should be taken into account, as potential and cause repertoires are defined as probability distributions of the past states of the system itself and nothing else. However, in calculating cei produced by the mecha- nism, not only the past/future states of the mechanism itself but also those of other mechanisms of the system can be considered. Said differently, the repertoires required to calculate cei of the given mechanism are not restricted to the same mechanism. For example, if the whole system is composed of elements A, B, and C with binary outputs 1 and 0, and the selected mechanism is A, A's current state 1 can specify the cause repertoire of the past states of any mechanism, including B, BC, AC, ABC, or even A itself. Similarly, it can also specify the past repertoire of the future state of any mechanism(see Oizumi, Albantakis, and Tononi 2014, Figure 4). Thus, to calculate cei generated by A in 1, one must first decide which mechanism to be paired with A. In principle, any mechanism of the system, which could be represented as the power set of the total elements of the system, can be paired with AB. In IIT 3.0, this idea of pairing is introduced as purview. When A in 1 is paired with ABC and the cause repertoire is calculated, the purview of A is represented as Ac/ABC p. If it is paired with ABC and the effect repertoire is calculated, the purview of A is represented as ABc/ABC f (see also Oizumi, Albantakis, and Tononi 2014, Figure 4). Once the purview is fixed, other elements outside the purview remain unconstrained and do not affect cause and effect repertoires. However, calculating unconstrained repertoire does not require a specific purview, because there is no difference on unconstrained repertoires among different current purviews. By discriminating the mechanism's purview, the causal analysis could be extended to every mechanism of the system. Third, whereas ei is measured by KLD, cei is measured by a different metric. As explained above, to calculate ei, we must measure the distance between the potential and the actual repertoires. In the earlier versions, it was KLD which was used to measure the distance. KLD is the most intuitive index for measuring the reduction of entropy, which directly relates to the quantity of information generated in terms of relative entropy. Since entropy and information were regarded as symmetrical, KLD was chosen for the scale of distance during the early versions of IIT. However, technically, KLD should not be considered as a proper metric, since it is not symmetric, does not obey triangular inequality, and is unbounded. In addition, non-compensated KLD measures only the reduction of uncertainty and does not account for the difference between states, which appears to be crucial in calculating information. For these reasons, another measure should be introduced as a new scale(see Oizumi, Albantakis, and Tononi 2014, Supplementary 2). Therefore, from IIT 3.0, Earth Mover's Distance(EMD) is used to measure the distance between repertoires. This is also known as Wasserstein distance, which is the distance function 13 defined by the minimum cost of redistributing the "dirt piles" to the location elsewhere(Oizumi, Albantakis, and Tononi 2014). Given that distributed probabilities can be thought of as "dirt piles", one can think of a distance between two repertoires as the minimum cost of distributing "dirt piles". In IIT, there are in fact two kinds of EMD. First, a general EMD is applied in calculating cause-effect and integrated information on the level of mechanisms. Second, an extended EMD is used to calculate that on the level of the systems of mechanisms. Nevertheless, in both cases, the point of using EMD remains the same. By using EMD, not only the reduction of entropy but also the difference between states is taken into account in calculating information. From IIT 3.0, the quantitative value of information is not represented by bit, since the unit of the distance measured by EMD is not a bit. In sum, EMD appears to be a more appropriate metric for IIT than KLD. Based on EMD, the distance between unconstrained past repertoire and cause repertoires is defined as the cause information(ci). The cause information illustrates possible causes of the mechanism's current state when a purview of the mechanism is fixed. Similarly, the distance between unconstrained future repertoire and effect repertoire is defined as the effect information(ei). This implies that effect information signifies possible effects of the mechanism's current state when a purview of the mechanism is fixed. In sum, effect information can be calculated from the distance between those repertoires and is quantified by EMD as same as cause repertoire.15 Finally, there is an informational principle that should be applied to cause-effect information. In the earlier versions of IIT, measuring the distance between repertoires was all that mattered. The measured distance was the amount of ei of the system. In IIT 3.0, however, there is more than just measuring the distance. As explained so far, there are two pairs of repertoires that lead to two kinds of information: cause and effect information. Then, which information should be accounted as the mechanism's information? At this point, the Information Bottleneck Principle(IBP) is introduced(Oizumi, Albantakis, and Tononi 2014). IBP forces one to choose the minimum of cause and effect information. The motivation behind IBP comes from intrinsic and causal notion of information: since information in IIT is supposed to be intrinsic to the system, the information that can be detected only by the external observer must be excluded. Suppose if the mechanism in a state only generates cause information, but no effect information. This implies that the mechanism being in such a state does not make any difference to the system. In that case, although the mechanism still belongs inside the system, it does not give any causal interaction among the system's other mechanisms. Hence, such cause information produced purely by the mechanism cannot be detected from the intrinsic perspective of the system. The same holds when the mechanism in the state produces only effect information, but no cause information. This observation enforces IBP so that the smaller one between cause and effect information is taken as cei. Since the concept of information is at the heart of the theory, the transition from ei to cei articulates the framework of IIT in a number of important aspects. By taking into account both the past and the future, the notion of information becomes more causal; it involves causes and 15. After IIT 3.0, a large variety of distance functions, such as Hilbert-space distance and Shannon-Jensen distance, have been newly proposed as metrics of informational difference(Tegmark 2016). It would be important to consider the characteristics of each measure in order to broaden the explanatory power of IIT. 14 effects. The application of IBP makes it more intrinsic. Specifically, as we will see in Sections 3.3- 3.4, when it comes to integrated information on the level of mechanisms, considering all possible purviews of the given mechanism plays a crucial role in enforcing the exclusion postulate. This involves the central notions of IIT 3.0, including concepts, conceptual structure, and other related ideas. However, the technical complexity is the other side of the theoretical articulation. As men- tioned above, the computational burden is doubled, since repertoires and information must be calculated twice. The multidimensional space for geometrical representation of concepts is also doubled(see Oizumi, Albantakis, and Tononi 2014, Figure 15). Moreover, calculating integrated information of mechanisms becomes more computationally complicated, because it should con- cern possible purviews of the mechanism. To calculate integrated information of the mechanism, MIP should be found in each possible purview(see Oizumi, Albantakis, and Tononi 2014, Figure 8). In this sense, it appears to be clear that the introduction of purview worsens the combinatorial explosion. In what follows, all these technical issues will be analyzed in further detail. 3.3 From one phi to two phis: the distinction between Φ and φ Before IIT 3.0, there was only one kind of integrated information; all integrated information calculated from the system was noted as Φ. Cause-effect repertoires were inferred from the mechanism, and it was all that mattered in calculating integrated information. However, from IIT 3.0, the distinction between the level of mechanism and that of systems of mechanisms has been introduced. According to this distinction of levels, a distinction between kinds of integrated information has been made(Tononi 2012; Oizumi, Albantakis, and Tononi 2014). On the one hand, there is integrated information generated from mechanisms, which is indicated as φ(small phi); on the other hand, there is integrated conceptual information produced by systems of mechanisms, which is represented as Φ(large phi). φ and Φ differ from each other both in their concepts and calculations. Integrated information φ succeeds the motivation "more than the sum of its parts" from older versions of IIT and is still analyzed on mechanisms. According to IIT, if there is a difference between the sum of the cause-effect information created by the partition of mechanism and the cause-effect information generated by the unpartitioned mechanism, and this difference directly refers to the information which mechanism forms as a whole entity. Any possible subset of a mechanism which can make difference on repertoire could be a candidate for partition.16 On the level of the mechanisms, φ can be measured by making a partition on a given purview; for example, the purview of ABC in 100 is defined over the past mechanisms in a state. As briefly explained in Section 2.3, among all possible partitions, MIP is selected for the calculation of integrated cause information, φcause. The purview of ABC also can be defined over the future mechanisms in a state. Applying the same procedure, integrated effect information, φeffect, can be calculated. By IBP, one can have integrated information φ. In this way, φ can be calculated 16. It is interesting that one of the variable subsets at a certain time might be empty as a result of a particular partition. Furthermore, partitioning can be thought of a method of making certain mechanisms causally inactive. This process is called 'virtualizing the element' or 'injecting noise to the mechanism'. For more detailed analysis, see (Krohn and Ostwald 2017). 15 respectively from every single purview available on a certain mechanism(see Oizumi, Albantakis, and Tononi 2014, Figure 8).17 However, since there can be many possible past/future purviews on a mechanism, one mech- anism in a state can have a multiple possible φs. Here, one of the ontological postulates comes in: the exclusion postulate states that, in order to contribute to experience, the mechanism must have only one set of possible causes and effects which is maximally irreducible, while all other sets should be excluded(Tononi 2012; Oizumi, Albantakis, and Tononi 2014). It means that only the cause-effect repertoire of the mechanism that provides the maximum value of φ, φmax, should be taken. As φ is defined as the minimum of φcause and φeffect, in order to find φmax, one must max and maximally irreducible effect find maximally irreducible cause repertoire that yields φcause max would be repertoire that provides φeffect φmax. The maximally irreducible cause repertoire is called core cause, the maximally irreducible effect repertoire core effect. The pair of core cause and effect is noted as Maximally Irreducible Cause and Effect repertoire(MICE). MICE or the mechanism which specifies MICE is called a core concept, or just concept.18 In short, by the exclusion postulate, the highest value of φ should be chosen among all possible φs produced by the mechanism and is defined as φmax. Here, the mechanism that produces φmax is should be regarded as the concept. max first. Then, the minimum of φcause max and φeffect After finding concepts, one can calculate the amount of integrated conceptual information at the level of systems of mechanisms. Concepts can be illustrated as points in the multidimensional space called concept space, and these points would make 'constellations' among the coordinate space. In IIT 3.0, the constellation of concepts is defined as conceptual structure(Oizumi, Alban- takis, and Tononi 2014). As each mechanism specifies its own MICE, the system of mechanisms specifies its own conceptual structure in the concept space. From this conceptual structure, one can calculate Conceptual Information(CI) generated by the system of mechanisms. As CI corresponds to the cause-effect information, it is quantified in a similar way; as there must be unconstrained past and future repertoires for calculating cause-effect information, there must be the "null" concepts for calculating CI. "Null" concepts are the unconstrained past and fu- ture repertoires in which the state of the system of mechanism is undecided.19 By applying 17. For the details of these computational steps, see (Oizumi, Albantakis, and Tononi 2014), Figure 6. 18. IIT 3.0 show a serious inconsistency in using the term concept: on one hand, 'concept' seems to refer MICE, a maximally irreducible cause-effect repertoire. In (Oizumi, Albantakis, and Tononi 2014), it is said: "the notion of a concept: the maximally irreducible cause-effect repertoire of a mechanism"(p.3). On the other hand, it is also used to indicate mechanism which specifies the MICE: "If the MICE exists, the mechanism constitutes a concept."(p.3), "concept(φmax): A mechanism that specifies a maximally irreducible cause-effect repertoire(MICE or quale "sensu stricto")"(p.5, Table 1), and "A mechanism that specifies a maximally irreducible cause and effect(MICE) constitutes a concept"(p.9). What is worse is that the term is described as denoting both: "Concept: A set of elements within a system and the maximally irreducible cause-effect repertoire it specifies, with its associated value of integrated information(φmax)"(p.5, Box 1). To avoid possible confusions, we choose the second use. In this article, the term concept will always refer to the mechanism specifying MICE. 19. The "null" concepts are named so because they specify unconstrained past and future repertoires if considered as mechanisms; in other words, it is the concept that specifies nothing. Although it is perceived as only a superficial notion on designating unconstrained repertoire, however, it also can be illustrated in conceptual space along with other concepts(see Oizumi, Albantakis, and Tononi 2014, Figure 11). By this way, "null" concept refers to the concept specifying its current purview of mechanism as an empty set. 16 the extended version of EMD, it can be quantified how much CI is produced by the system of mechanisms.20 The calculation of integrated conceptual information Φ is also analogical to that of φ. Once the purview of the system of mechanisms is given, MIP can be found by partitioning21 the purview. By measuring the difference of CI between the unpartitioned and the partitioned, it can be calculated how much the integrated conceptual information is generated by the system of mechanisms. Again, as for φ, there can be many possible Φs as all possible unidirectional partitions of the set of elements should be considered. Here, the exclusion postulate comes in again; it enforces only one complex among all other overlapping systems of mechanisms to contribute to consciousness. Thus, the one that generates the maximum of Φ should be chosen as Φmax. Finally, the conceptual structure that gives rise to Φmax is defined as Maximally Integrated Conceptual Structure(MICS). The system of mechanisms that produces Φmax and so specifies MICS is defined as complex.22 In IIT 3.0, such MICS generated by the complex is directly identified as the subjective experience. By introducing the distinction between φ and Φ, now it is much logical to explain the gener- ation of consciousness using integrated information in further detail. In IIT 3.0, MICE is called quale "sensu stricto", which means quale in the narrow sense. Since this sort of quale includes 'redness of red' or 'painfulness of pain', it can be considered to be quale in the philosophical debates. On the other hand, MICS is called quale "sensu lato", which means quale in a wide 20. Extended EMD differs from original EMD by its methods on calculation. The distance between cause-effect repertoire and unconstrained cause-effect repertoire is measured by EMD, and each EMD of cause and effect distributions are added up, then it is multiplied by φmax of the concept, which functions as the weight of each concept. In short, extended EMD is used on the level of systems of mechanisms by multiplying φmax as each distribution's weight. Even if the details on the calculation vary, the fundamentals on calculating the distance, which is redistributing the probability distribution, do not change. For further detail on applying extended EMD, see (Oizumi, Albantakis, and Tononi 2014), Supplementary 2. 21. The partitioning in the level of systems of mechanisms must be unidirectional(Oizumi, Albantakis, and Tononi 2014). Unidirectional partitioning is done by virtualizing elements; when a mechanism is injected with noise, information disappears, as the mechanism gets considered as external noise and loses its intrinsic causal power. Unidirectional partitioning could be thought of as injecting noise between subsets for only to a certain direction of the connection. Thus, partitioning the direction of connection between subsets on the system level is analogous to virtualizing the elements on the mechanism level. For further detail on virtualizing the elements, see also (Krohn and Ostwald 2017). 22. As a result, there comes an important change in defining the complex since IIT 3.0. In IIT 3.0, due to the exclusion postulate, complexes cannot be nested or overlap at all. In the earlier versions, however, since the exclusion axiom/postulate was not introduced yet, complexes could partially or wholly overlap. Meanwhile, there can be multiple complexes in one system. IIT predicts that one system can be condensed into several complexes. The complex that has Φmax is called major complex, and the complex which does not overlap, but has Φ smaller than Φmax is called minor complex. Since they have their own Φmax, they are considered as an individual complex. Minor complex can be thought of as a local maximum which implies 'locally condensed minimal consciousness'(see Oizumi, Albantakis, and Tononi 2014, Figure 16). There should be a single major complex in general situations, but there could be multiple major complexes according to circumstances. For example, split brain syndrome or dissociative disorders could be explained as clinical examples of the main complex being split into two or more. At the same time, minor complexes could be thought of as preconsciousness; the constituent of consciousness which can contribute to the reaction of extrinsic inputs. Continuous flash suppression could also be explained through the function of the minor complex(Oizumi, Albantakis, and Tononi 2014). 17 sense(Oizumi, Albantakis, and Tononi 2014). As mentioned above, IIT equates experience with MICS. As mechanisms in the complex maximally integrate cause-effect information, concepts are generated, and we have qualia. As the complex maximally integrates CI, MICS is produced and we have an experience. Based on the distinction and the exclusion postulate, the notions of concepts and MICS can be defined. These central notions of IIT 3.0 enable one to explain how experience arises from its physical substrates in a bottom-up manner. All these articulated explanations essentially start from the distinction between φ and Φ. However, this distinction between φ and Φ also raised several problems for IIT. First and fore- most, computing Φ and the application to real systems became computationally intractable(Oizumi, Albantakis, and Tononi 2014). Almost every aspect of calculation doubled: MIP had to be found twice; once at the level of mechanisms and twice at the level of systems of mechanisms. Owing to the distinction between φ and Φ, it appears that the combinatorial explosion in IIT extremely deteriorated. In turn, such computational infeasibility rendered the empirical prospect of IIT more pessimistic.23 At the cost of an articulated bottom-up explanation of experience, the theory had to face serious practical problems in retaining empirical validity. Another problem emerges from the exclusion postulate. As explained above, the exclusion postulate enforces that only the mechanisms which give rise to φmax or the systems of mechanisms which provide Φmax must be taken as the concept or complex. However, if there are several different MICEs that yield exactly the same Φmax, then which repertoire or conceptual structure should be taken? Clearly, being as biggest does not involve being as unique. Nevertheless, the exclusion postulate says nothing about this problematic underdetermination of quale(Krohn and Ostwald 2017). Moreover, as we have noted at the end of Section 2.1, what if exactly the same Φmax are produced at the different spatio-temporal grains? When two equivalent Φmaxs are detected both from the level of neuronal units and from the level of cerebral lobes, the exclusion postulate cannot tell which level should be taken as the 'locus' of conscious experience. At least at the current stage of the theory, IIT does not have any theoretical resource to deal with such issues. 3.4 From vector geometry to point geometry: the geometry of integrated information IIT has always assessed integrated information in a geometrical manner. Nonetheless, several updates from IIT 1.0 to 3.0 brought a number of changes in the geometry of integrated informa- tion. Since IIT's central notions, such as concepts and conceptual structure, are closely related to the geometry of integrated information, a deeper analysis of why and how the geometry has been revised would be needed. In earlier versions of IIT, the space for representing informational structures was dubbed qualia space, the multidimensional space which has its axes for each possible state of the sys- 23. One of the reviewers noted that the source of the combinatorial explosion is the combination principle. For example, as the calculation of φ needs be performed according to the combination principle, it must be carried out over all possible subsets of the candidate set and for each of those subsets over all possible purviews. Nevertheless, the main constraint is, that the number of unique bipartitions rises exponentially with the cardinality of the set(see Krohn and Ostwald 2017, Appendix). 18 tem(Tononi 2008; Balduzzi and Tononi 2008, 2009).24 The geometrical shape expressing the in- formational structure was called quale. The quale is constituted by q-arrows and points in qualia space: points represent actual repertoires specified by the system in a state when a certain con- nection -- a set of causal connections -- is added. Furthermore, q-arrows represent informational relationships between each actual repertoire specified by the added connection. Thus, the point "at the bottom" of the quale is a potential repertoire specified by the system in a state, when no connection is added (the "null set"). On the other hand, the point "at the top" of the quale is the actual repertoire, when all connections are added (the "full set"). By adding each connection from the potential repertoire, it is possible to analyze how much ei does the system gains by each connection. This can provide detail about which connection informationally contributes to the quale. All points connected by all q-arrows illustrate a geometrical figure like a "polytope"(see Balduzzi and Tononi 2009, Figure 3). The major point of the geometry of the earlier versions of IIT is that q-arrows are represented as vectors. Interpreting q-arrows as vectors, one can find many properties of the informational relationships constituting the quale: the length of the q-arrow represents how much ei is generated by adding a connection. For instance, the length of the q-arrow connecting "the bottom" and "the top" of the quale represents ei of the system in a state. In addition, the direction of the q-arrow expresses the particular way how adding a connection sharpens repertoires. One of the most interesting properties of q-arrows, however, would be entanglement(γ): when a q-arrow cannot be decomposed into an exact vector sum of its sub-q-arrow, then it is considered as tangled. When the q-arrow is tangled, it means that there is integrated information gained by adding up the corresponding connection. The way how the q-arrow is tangled can be measured by vector calculation. The difference between the length of the q-arrow and that of the vector sum of its sub-q-arrows is quantified by γ. In this sense, entanglement represents how much information a q-arrow generates above and beyond its components. In an earlier version of IIT, the q-arrow with γ>0 was defined as a concept. Moreover, complexes could be defined by comparing γ of each concept; a concept with relatively high γ was called as a mode. Before IIT 3.0, these articulated analyses were available from the vector analysis of q-arrows(Tononi 2008; Balduzzi and Tononi 2008, 2009). However, in IIT 3.0, such vector analysis is no longer available(Tononi 2012; Oizumi, Al- bantakis, and Tononi 2014). As explained in Section 3.3, the concept cannot be defined as an entangled q-arrow. Rather, it is defined as MICE plotted as a point in the concept space. In- stead of the null set, the current version posits the "null concept", which is the unconstrained repertoire specified by the system of mechanisms when no mechanism of the system is given. As explained in Section 3.3, one can calculate how much CI is generated by the mechanism by applying extended EMD. Thus, in concept space, the distance between two concepts does not capture how much ei is generated by adding a connection to the system in a state. Rather, it captures how much CI is generated by adding a mechanism to the system of mechanisms. Due to these differences, adding connections, specifying informational relationships between repertoires, and analyzing q-arrows cannot be found in the current version of concept space. In sum, all 24. For example, when n is the total number of the system's elements and each element can have only two possible outputs, dimension of 2n is required to constitute the qualia space which represents the system. 19 analyses and notions grounded by vector calculus of q-arrows are not available in IIT 3.0. Prima facie, the geometry of integrated information appears to be simplified. The transition from effective information to cause-effect information also affects the geometry of IIT. The transition in the notion of information doubles the concepts and space. Since the theory was based on ei, there were only the past repertoires. Therefore, just one space was required to represent concepts. In IIT 3.0, however, space must represent both the past and future states, because the theory is built on the notion of cei. As a result, there must be two repertoires that give φmax: core causes and effects. Since the concepts cover not only the past but also the future repertories, the space where the concepts are represented, and the geometrical structures made from concepts are also doubled. In other words, the multidimensional space and MICS must cover both of the past and future. As the definition of information changes, almost everything in the geometry of IIT appears to be doubled: points, space, and geometrical structures. In this sense, the geometry of IIT seems to be rather complicated. In a nutshell, the evolution of the geometry of integrated information has two sides. On one hand, it has been simplified in that all the articulated vector analyses for q-arrows are not used anymore. On the other hand, however, it has been complicated in that every aspect of the geometrical approach should be counted twice. We believe that these double aspects of the geometry of IIT are consequences of transitions to other central notions of the theory. 4 Theoretical Issues in IIT Despite its scientifically interesting prospects, IIT also faces several theoretical problems. These problems concern IIT's principles, core concepts, and their possible consequences. Despite a few exceptions, many recent literatures are almost focused on particular technical issues(Kitazono, Kanai, and Oizumi 2018; Hidaka and Oizumi 2018).25 This focus is fully understandable, since, without overcoming various technical barriers, there would be hardly empirical advances for IIT. However, theoretical problems deserve more attention, as it is theoretical considerations that enable us to judge whether or not the theory is worth pursuing in the first place. Despite such importance, theoretical issues have been largely overlooked in IIT debates, and relatively fewer studies have addressed this topic. Therefore, such theoretical problems IIT require a closer analysis. Of note, while there might be many issues concerning IIT's theoretical aspects, in the present paper, we focus on three major problems that appear to raise serious questions about the plausibility of the theory. 4.1 Sophisticated panpsychism: Unjustified scientific authority The first issue is that IIT embraces a form of panpsychism. Panpsychism has traditionally been ignored as full-fledged mysticism. The view that extremely simple organisms and even seemingly non-living things have 'a small piece of mind' sounds counterintuitive enough. IIT, however, admits a variety of examples that could support a sophisticated sort of panpsychism. A 25. Rare exceptions are (Bayne 2018) and (Krohn and Ostwald 2017). The former provides critical assessments of the axiomatic approach of IIT 3.0. The latter illustrates the important and disturbing conceptual issue of "magic cuts" which can violate IIT's fundament intuition: "the whole is more than the sum of its parts". 20 representative case is that of photodiodes(Tononi 2008; Oizumi, Albantakis, and Tononi 2014). According to IIT, a photodiode, which is designed to react to various external stimulations only by lighting on and off, is "minimally conscious." It means that the photodiode has a minimal level of consciousness and a certain quality of experience as well. Nonetheless, the photodiode might be the last one we ascribe experience to. It is difficult to believe that such a simple micro- mechanism could have a certain kind of consciousness. There is another example which appears to be the opposite of the photodiode case. Aaronson(Aaronson 2014b) has clearly shown that, if IIT is right, a lattice constituted by just connecting one kind of simple logic gates over and over could have a high value of Φ. According to Aaronson's description(Aaronson 2014b), XOR gates arranged in a 2D square grid would be conscious. Much incredible result is that such increasing of Φ is proportional to the length and breadth of the grid.26 Therefore, by a simple recursive procedure of connecting more XOR gates, there could always be a huge physical lattice which is more conscious than a normal human being! Though this is truly unbelievable, IIT clearly allows these examples. If the photodiode refers to the micro-case of panpsychism, the lattice could be its macro-case. The problem is that those simple and non-organic things' being conscious is so counterintuitive that it would rather be easier to take it as counterexample than as evidence. If IIT predicts that those simple systems which are apparently unconscious could be conscious, at least for many, such prediction itself would be enough to present reductio ad absurdum against IIT. Hence, the charge of panpsychism should be taken seriously in deciding whether or not IIT is theoretically plausible. If it is certain that no simple object such as a photodiode or a XOR grid is conscious and therefore panpsychism is wrong, IIT must be wrong too. What makes this issue more problematic is that the founder of IIT is seemingly undaunted by those critiques above. IIT does not just allow diverse panpsychistic cases. It actually argues for it, by demonstrating those counterintuitive cases in a detailed manner. Tononi's reply(Tononi 2014) shows his confidence that all those counterintuitive cases are actually the evidence for IIT. Tononi emphasizes that when science and popular intuitions or common-sense conflict with each other, it is always science that takes priority(Tononi 2014). According to Tononi, this is the primary reason why we should count those hard-to-swallow examples as evidence(Tononi 2014). The history of science is full of reversions of commonsense by innovative scientific discoveries. Since IIT is a scientific theory, the fact that IIT produces several counterintuitive predictions cannot be a strong reason to reject it. Rather, in Tononi's view(Tononi 2014), it is our widely entrenched intuition that must be corrected. Said differently, IIT might be on the edge of "scientific revolution," and Tononi might be, following Aaronson's witty phrase(Aaronson 2014a), "the Copernicus-of-consciousness." Nonetheless, Tononi's reply(Tononi 2014) could be objected in several ways. First, it is not obvious at all that IIT is really able to claim its priority over commonsense or intuition. Even if it is true that science tends to override culturally and historically widespread intuitions, the question remains whether IIT has any right to do so. Technically, not all hypotheses of science 26. To be fair, Aaronson's calculation was based on IIT 2.0 so that it cannot be directly applied to the current version. Unlike IIT 2.0, IIT 3.0 does not require procedure of normalization. Thanks for the reviewer who reminded this point. 21 can have the right to correct popular intuitions. In Kuhnian terms, only the so-called "normal science," which has successfully secured, well-established methodologies, exemplars, problem- solving procedures, basic beliefs and values shared by members of the scientist society, can argue for its right over commonsense and intuition(Kuhn 1962). However, it seems undeniably clear that, in the current stage, IIT cannot be such normal science of consciousness. For now, it is nothing more than an interesting working hypothesis that should wait for rigorous examination from the current scientist society. Moreover, as repeatedly pointed out in Section 3, IIT suffers from a number of technical issues preventing empirical experiments and practical applications. No direct evidence has been obtained by empirical studies conducted on real physical systems. Despite the growing body of empirical studies resting on the IIT framework, no IIT theorist has been able to apply the pure IIT 3.0 to neural data such as brain signals. Given its present status in the field, IIT appears to be unfit to serve as a hypothesis of normal science. Rather, IIT is more likely to be something in between "pre-science" and normal science, which might be one possible candidate of "paradigm shift" in the field of consciousness studies. Then, IIT's panpsychistic predictions cannot be prior to our general intuitions about consciousness. A heavy burden of proof is still on the side of IIT, and our anti-panpsychistic intuition should be taken as default. Aaronson's comment(Aaronson 2014a) reveals this situation: "The anti-common-sense view gets all its force by pretending that we're in a relatively late stage of research -- namely, the stage of taking an agreed-upon scientific definition of consciousness, and applying it to test our intuitions -- rather than in an extremely early stage, of agreeing on what the word "consciousness" is even supposed to mean(italics added)". The problematic implication of sophisticated panpsychism does not lie only in the conflict with the strong intuitions, which is external to IIT. It also lies in the logical development of the structure of the theory, which is internal to IIT. It is the most original and unique feature of IIT that the theory starts from a number of phenomenological axioms. Yet, the problem is that the axioms are taken for granted in IIT. They are assumed to be self-evident. However, taking some- thing for granted or assuming it to be self-evident is just another way of accepting it as intuitive. In this sense, it is IIT itself that strongly depends on a set of intuitions. IIT is fundamentally grounded on several phenomenological intuitions.27 Hence, if IIT allows panpsychistic cases and 27. To this matter of grounding IIT, one reviewer has raised an interesting point. The reviewer predicted that "defendants of IIT would argue that the set of axioms is qualitatively different from anti-panpsychist intuitions in that they are not only self-evident but also directly accessible from a first-person perspective". While this might be true, this reply seems to raise another issue about the direct accessibility of consciousness and its fundament properties, on which phenomenological axioms are about. This 'direct accessibility from the first-person point of view' has usually been discussed under the title of introspection. In order to claim that introspection lends further support to the phenomenological axioms, one must first prove that such introspection is significantly reliable enough to have some evidential force. However, it is controversial if introspection is significantly reliable; rather, a growing number of empirical studies suggest that introspection is not a reliable source of evidence. Once this point is taken, the alleged qualitative difference between anti-panpsychistic intuition supporting common sense and phenomenological axioms grounding IIT becomes doubtable. Though the reliability of introspection deserves deeper analysis, in the current context, raising doubt against introspection is enough to elaborate our argument by blurring the difference between anti-panpsychistic and phenomenological intuition. For a thorough critical assessment of the reliability of introspection, see (Schwitzgebel 2008) and (Schwitzgebel 2013). Smithies and Stoljar also present ample philosophical arguments for or against the special nature of introspection(Smithies and Stoljar 2012). 22 denies opposing intuitions, a charge of double standards could be raised. On one hand, IIT strongly holds some intuitions by calling them "phenomenological axioms". On the other hand, it easily dismisses other intuitions by treating them unscientific commonsense. Nonetheless, how can IIT justify this selective adoption of intuitions? Why does it adopt one group of intuitions but reject another? If axioms of IIT are considered as a significant type of phenomenological intuition concerning what consciousness is, anti-panpsychistic intuitions should also be taken to be equally important phenomenological insights about what consciousness is not. At least in the current version of IIT, we cannot find any principled reason to take axioms for granted and to reject other intuitions about consciousness. Once IIT wants to deny anti-panpsychistic intuitions as prejudices of scientifically unenlightened laymen, it should do the same thing with its own underlying intuitions. However, what such denial of its own axioms really amounts to is just a self-refutation. Therefore, without providing further reason to take its axioms and ignore anti-panpsychistic intuitions, IIT cannot be free of its charge of double standards of contrasting intuitions. To sum up, sophisticated panpsychism implied by IIT threatens IIT itself in two ways. First, considering IIT's premature status, the panpsychistic charge gives a very good reason to defy IIT. As long as no strong evidence is provided, panpsychism alone could suffice not to believe IIT. In addition, it raises the charge of double standards to seemingly equivalently respectable intuitions. Being fundamentally founded by "phenomenological axioms", it is difficult for IIT to dismiss opposing intuitions. 4.2 Fading and dancing qualia: Radical dissociation between experience and cognition The second issue with IIT is that as Cerullo has pointed out(Cerullo 2015), IIT faces the fading and dancing qualia arguments.28 Fading and dancing qualia are basically thought experiments designed by David Chalmers(Chalmers 1972). As suggested by their names, fading qualia de- scribe an imaginable situation where qualia become more and more eroded. Dancing qualia show another scenario that the whole qualia are replaced by totally different qualia. Their purpose is to show that, in our natural world, any attempt to detach experience from the functional orga- nization of a system would face extremely counterintuitive consequences. Despite the richness of detail, in the context of IIT, the relevant point is simple: IIT appears to entail anti-functionalism or anti-computationalism so that it commits to a possibility which fading and dancing qualia rule out. Fading qualia start with the assumption of the physical system and its functional organization in our world. Since functional organization is a matter of abstraction, it must be fixed how far the organization should be grained. In fading qualia, functional organizations are supposed to be sufficiently fine-grained to fix physical systems' behavioral capacities. Following this assumption, if two physical systems share their functional organization, all their behaviors must be identical. Another assumption is multiple realizations without experience. It is assumed that there are 28. Although Cerullo highlights the point(Cerullo 2015), he does not provide a specific description or analysis in his work. By contrast, Shanahan provides a more clear and comprehensive analysis(Shanahan 2015). Both of them concerns upon the problem of fading and dancing qualia anyhow. 23 multiple kinds of materials in implementing one organization, but only some of them support the phenomenal qualities of experience accompanied by the organization, while others do not. Now let us imagine that the functional organization of Mary's brain is realized by neurons. Then, Mary sees a ripe tomato and feels a visually red feeling. In her brain, maybe somewhere in her visual cortex, there is a neural correlate of that red quale. However, something strange happens. The neurons composing her neural correlate of phenomenal redness are now substituted by silicon chips one by one. Given multiple realizations, this replacement must be possible. The crucial point is that, although those chips are perfect functional equivalents of Mary's neurons, they do not support any quale at all. A natural consequence is that her vividly red experience becomes murkier, and eventually disappears. The problem is that Mary's functional organization never undergoes any change, despite the gradual qualitative change of her experience. She would still manifest exactly the same bodily and verbal behaviors as before. Moreover, considering that her brain function is perfectly the same, it is reasonable to think that her cognitive states are also the same as before. If cognitive states of Mary, such as her judgments or beliefs about experience, do not remain intact and change following the eroding visual experience, such cognitive states would radically come apart from the functional organization of Mary' brain. Nothing in the functional organization would correspond to the change of cognitive states. Chalmers argues that this kind of dissociation is highly unlikely, by saying "If such a major change in cognitive contents were not mirrored in a change in functional organization, cognition would float free of internal functioning like a disembodied Cartesian mind "(Chalmers 1972, p.258). This is why he claims that "There is simply no room in the system for any new beliefs to be formed", "[u]nless one is a dualist of a very strong variety"(Chalmers 1972, p.258). As free-floating, disembodied cognitive states are deeply problematic and counterintuitive, it is safe to assume that cognitive states do not suffer any change.29 As a result, Mary neither notices nor is aware of anything. This is fading qualia in a nutshell. It seems highly unlikely that such situation could really occur in our world. What is worse is that Mary is perfectly rational and functional in every other aspect, except for her beliefs about her visual experience. She is not pathological or deeply confused. Nevertheless, she suffers somewhat systematic errors concerning her experience. Whenever a substitution occurs, she forms a wrong belief that she is still seeing the red tomato. Clearly, this systematic error of rational subject is hardly acceptable in our natural world. Dancing qualia is another version of fading qualia. In fading qualia, the phenomenal aspect of the experience gets gradually eroded and ends up to none. In dancing qualia, however, the phenomenal aspect does not totally vanish. Instead, it keeps changing itself. Mary does not suffer the gradual neuron-silicon replacement. Nonetheless, she has a certain neuroprosthetic device, which functions identically to her natural neural correlates of the reddish quale. This time, despite its function, the device does not support the reddish quale. Suppose that it grounds a blue quale instead. And there is a switch that alters Mary's neural correlate to the device. Then, what would happen if someone turns the switch on? Ex hypothesi, Mary's visual experience will suddenly become blue-like. If the switch turns off, the opposite would happen. Hence, as someone turns the switch on and off, Mary's visual quale will dance back and forth! The trouble 29. One of the reviewers advised that there should be more rationales to claim that cognitive states are fixed under the gradual replacement. For more on the debate, see (Chalmers 1972), p.247-274. 24 is that Mary would not be able to notice any change in her visual field. Since the device is the perfect functional duplicate of Mary's neural correlates, the functional organization of her brain remains exactly the same. As in fading qualia, Mary's cognitive states would be intact, regardless of the change of the phenomenal aspect of her visual experience. If so, Mary would neither notice nor be aware of any change, even if visual qualia are dancing "in front of her eye"! For the same reason as in fading qualia, it appears that this consequence must be rejected. The relevant point in the context of IIT is that IIT essentially allows these implausible cases. Fading and dancing qualia are possible only on the assumption that there could be functionally identical, but phenomenally different systems. In the IIT framework, neurons and silicon chips, neural correlates and the neuroprosthetic device could be such systems. The only way to detour the unwelcomed consequences appears to be denying the possibility of the functionally identical, but phenomenally different systems. IIT, however, does not and even cannot deny that possibil- ity. According to IIT 3.0, even if two physical systems perfectly share their functions, they can be different in Φmax they produce. Considering that maximally integrated conceptual information is experience in IIT, the claim that function and Φmax can come apart implies anti-functionalism or anti-computationalism about consciousness. There could be zombie systems which perform exactly the same as conscious systems but do not have any experience at all. This is not just a speculation; indeed, Oizumi, Albantakis, and Tononi design such a zombie system and demon- strate how it works(see Oizumi, Albantakis, and Tononi 2014, Figure 21). If a zombie system is possible, there is no reason not to believe silicon chips in fading qualia or neuroprosthetics in dancing qualia. Then, IIT should accept those unacceptable consequences anyway. It is IIT's anti-functionalism that opens the door to fading and dancing qualia. In front of the implausible results of fading and dancing qualia, there are only two logical ways for IIT to reply: to dodge the bullet or to bite it. Nonetheless, none of the two appears to be available without significant revisions of the theory. On one hand, if IIT wants to dodge the bullet, it must show how Mary could notice the change in her visual experience, even if her brain functions remain exactly the same. It is highly likely that, if there is no difference in the brain functions, the same will apply to information processing. In IIT as a paradigm of cognitive science and artificial intelligence, it is widely accepted that, in order to notice or be aware of something, there should be corresponding activities of information processing. However, by the assumption of functional identity, Mary cannot have any new information processing corresponding to the change of quale. Then, how can Mary notice or be aware of the experiential change? On the other hand, if IIT tries to bite the bullet, all the debates concerning the charge of double standards resurface again. IIT cannot merely say "Though being counterintuitive, it's true nonetheless". IIT is scientifically so premature that it is not in a position to simply override strong intuitions in the name of science. Furthermore, since IIT itself takes some intuitions as primitive, it cannot easily dismiss other intuition as ungrounded. In one way or another, it seems difficult for IIT to defy the intuition that the radical dissociation between cognition and experience is impossible. In one way or another, IIT can neither dodge, nor bite the bullet of fading and dancing qualia. All in all, IIT cannot deal with fading and dancing qualia. Holding anti-functionalism about consciousness, IIT does not have theoretical resources to explain how the system which function- ally remains identical could notice its phenomenal changes. On the other hand, accepting the 25 possibility of unnoticeable phenomenal change is extremely counterintuitive to that, if IIT allows such notion, many would reject IIT. As in the panpsychism debate, due to its dependence on intuitions, IIT cannot merely dismiss the intuition that a rational and functioning system must be able to be aware of its own experiential changes. Anyway, IIT faces serious troubles. 4.3 The paradox of certainty: Loss of certainty undercuts existence In Section 4.2, we argued that, although the empirical possibility of radical experience-cognition dissociation causes a serious counterintuitive consequence, IIT cannot dodge this consequence. In this section, we attempt to show that such radical experience-cognition dissociation causes another problem: the loss of certainty about consciousness. We believe that this loss of certainty can undercut the very foundation of IIT: the existence of consciousness. We, or at least many of us, appear to be certain about our consciousness. Our own con- sciousness might be the only thing we can be certain about. However, the argument from fading and dancing qualia shows that our phenomenal beliefs or judgments can be detached from our consciousness even when we are fully alert and attended. If this is the case, we ourselves might be suffering fading and dancing qualia as well. That is, we might be like Mary who cannot be aware of the absence of her own visual consciousness. If so, even if we strongly believe or take for granted that we are conscious here and now, it is possible that we are not. As Descartes doubted, an omnipotent demon might manipulate our perceptual experience to make us believe the exis- tence of the external world, even if there is no such world. Similarly, something might control our cognitive system to make us believe the existence of our experience, even if there is no such thing as experience at all. Then, how can we be so sure about that we are conscious here and now? In other words, is there any guarantee that we are not deluded zombies who think that they are conscious if experience and cognition about the experience can come radically apart? It is clear that the radical experience-cognition dissociation deprives us of the certainty of consciousness. And if IIT allows the dissociation, it cannot secure the certainty of consciousness. Some might deny the certainty of consciousness. Although the certainty of our own experi- ence appears to be the last thing we can deny, whether or not we are really certain about our experience is surely debatable. Nevertheless, it appears that IIT cannot easily deny the certainty of consciousness, because the theory appears to be grounded in it: the first phenomenological axiom states that consciousness exists. Furthermore, this existence of conscious experience is supposed to be certain. Indeed, it is clearly argued that consciousness is certain when Tononi paraphrases(Tononi 2012, p.296) Descartes' cogito ergo sum: "I experience therefore I am". The very starting point of IIT, the existence axiom, necessarily requires the certainty of conscious- ness. If we are not certain about our own consciousness, why should we struggle for a scientific theory of consciousness? Therefore, the possibility of the radical experience-cognition dissociation provides a some- what delightful and disturbing paradox against IIT: If IIT is true, radical experience-cognition dissociation is actually possible. If so, we cannot be certain about our own consciousness. If we cannot be certain about our own consciousness, IIT cannot get off the ground. Therefore, if IIT is true, there is no reason to suppose that it is true. We call this argument the paradox of certainty. IIT appears to simultaneously require and reject the certainty of consciousness. 26 It seems that the only possible reply from IIT would be denying the empirical possibility of the radical experience-cognition dissociation. However, as we have seen in Section 4.2, the problem is that, at least in the current version of the theory, it is difficult to find any rationale for such denial. Considering the fact that IIT actually argues for the functional zombie system, it is doubtable that IIT can deny such possibility. In fact, we cannot find any consideration about how experience affects beliefs or judgments, and vice versa in IIT. While IIT appears to have a great deal with how experience is generated from its physical substrate, it does not provide much insight into how the subject can be aware of that generated experience. Said differently, IIT is blind to the question of how we can secure self-knowledge or metacognition about our own experience. This is the topic of the last section of this paper. 4.4 Metacognitive accessibility: Missing link in IIT What is the main source of the theoretical problems mentioned thus far? We think the culprit here is disregarding cognitive aspect of consciousness.30 In IIT, the explanation of how the experience could be cognitively accessed by a subject is totally absent. IIT never takes account of metacognition in explaining consciousness, and we believe that it is this neglect of metacognition that generates all theoretical problems IIT faces. Due to its ignorance of metacognition of consciousness, IIT can ascribe consciousness to simple systems lacking metacognitive mechanisms, such as photodiodes or logic grids. Though photodiodes and logic grids produce integrated information, it is highly unlikely that these simple physical systems are equipped with metacognitive mechanisms. Given that they lack metacog- nition, those systems do not, and even cannot, have cognitive access to integrated information of their own. There is no photodiodes and logic grids' metacognition of their integrated infor- mation. Under the IIT framework, this metacognitive inaccessibility implies that photodiodes and logic grids cannot know or be aware of their own consciousness. While they are conscious, they cannot know that they are conscious! However, this lack of metacognition and its strange consequence do not prevent IIT to ascribe consciousness to simple systems, as it does not concern metacognitive access to consciousness at all. Furthermore, since IIT appears to neglect how metacognition and experience could be associ- ated, it allows the radical dissociation between metacognition and experience, which is shown by fading and dancing qualia and ultimately results in the paradox of certainty. In fading and danc- ing qualia, unlike in the panpsychistic cases, the system has metacognitive access to integrated information it produces. That is, Mary has a metacognitive belief about her visual experience. The problem is that her metacognitive access systematically produces wrong beliefs about her own experience. In fading qualia, Mary is usually right about what she sees. However, as soon as the process of neuron-to-silicon replacement begins, Mary starts to have wrong beliefs about what she sees. In dancing qualia, whenever the switch turns on, Mary becomes wrong about her visual experience. In both cases, Mary's being wrong is very systematic in that it strongly correlates with the replacement. Mary's systematically being wrong indicates that her metacog- nitive access to her visual experience systemically results in wrong beliefs. However, since there 30. Cerullo makes a similar point(Cerullo 2015). After distinguishing incognitive and cognitive consciousness, he argues that IIT only deals with incognitive one, which is tantamount to consciousness without subject. 27 is no consideration about how the system metacognitively accesses its own experience in IIT, it cannot help but allow the absurdities of fading and dancing qualia. In addition, once the radical experience-cognition dissociation is admitted as possible, there appears to be no way to eschew the paradox of certainty. Given the tight relationship between experience and cognitive access, IIT's neglect of metacog- nition is somewhat surprising. Phenomenologically, there appears to be a close, even constitutive relation between metacognition and experience. Despite philosophical debates surrounding the distinction between phenomenal vs. access consciousness(Block 1995, 2007), we believe that there could be experience without actual metacognitive access. Nevertheless, this does not mean that there could be an experience that cannot be metacognitively accessible. It sounds absurd and even unintelligible that a conscious experience is absolutely out of our range of metacognition. Such experience must be a conscious experience we cannot be conscious of, which is unconscious by its nature. Hence, it appears that metacognitive accessibility, not actual metacognitive access, is necessarily involved in having consciousness. That is, metacognitive accessibility is a necessary condition for something to be a conscious experience.31 Therefore, we argue that any scientific theory of consciousness must take account of the metacognitive accessibility of consciousness. However, no matter which version it may take, IIT does not seem to consider why and how metacognitive accessibility must be taken into account when it comes to explaining conscious experience. Accordingly, we strongly suggest that the first step to deal with the theoretical problems mentioned so far is introducing metacognitive accessibility in the IIT framework. Phenomenological axioms, ontological postulates, and math- ematical models of IIT should be revised in order to reflect the necessary connection between metacognitive accessibility and consciousness. Once we can successfully assimilate metacognition into IIT, we could have a better version of the theory, which would deserve to be called 'IIT 4.0.' 5 Conclusion IIT has been a center of the debate surrounding the science of consciousness. Many of those who are engaged in the field displayed interest in the theory, and some raised serious doubts and criticisms. It is worth to assess what IIT is about and why it is controversial. In this paper, we have critically examined the theoretical evolution and related issues of IIT. We have introduced basic concepts, which might be considered as the core of IIT. Both IIT's explanatory power and limits appear to be already embedded in its core concepts. We have also described how the 31. For a similar point, see Chalmers(Chalmers 1997) who argues against Block(Block 1995) that, even when there is phenomenal consciousness(P-con) without access consciousness(A-con), it does not mean that there is not accessible consciousness. According to Chalmers, once A-con is defined in terms of availability for global control, P-con always goes along with A-con(Chalmers 1997). Since global availability requires only accessibility, the original notion of A-con should be modified from access consciousness to accessible consciousness. Our suggestion here could be taken as claiming that, if an experience is phenomenally conscious, it must be accessibly conscious. It is worth noting that this transition from access to accessibility is what distinguishes Chalmers(Chalmers 1997) and us from those who follow Higher Order Theory of consciousness(HOT)(Rosenthal 1986, 2005). In HOT, for a mental state to be conscious, it must be actually accessed by a higher order state. Our suggestion, however, does not demand actual higher order, metacognitive access. All that required is that the state must be metacognitively accessible. No actual higher order state needs to be there. 28 theory has been updated throughout the last decade. In some aspects, those major transitions can be thought as a progress. However, in other aspects, some of the issues were worsened, and even new problems emerged. Specifically, the principled part of the framework of IIT, its phenomenological axioms, and ontological postulates raise serious questions about the scientific status of the theory, the possibility of radical dissociation between experience and cognition, and the logical structure of the theory. We have suggested that focusing on our ability to access our own experience through metacognition might be one way to deal with these theoretical issues. The cognitive relationship between metacognition and consciousness might push IIT one step forward in becoming the science of consciousness. Author Contribution HP wrote Section 2 and 3; KM wrote Section 1, 4 and 5; All authors reviewed the manuscript; HP documented the manuscript in LATEX. References Aaronson, Scott. 2014a. Giulio tononi and me: a phi-nal exchange. https://www.scottaarons on.com/blog/?p=1823. . 2014b. Why i am not an integrated information theorist (or, the unconscious expander). https://www.scottaaronson.com/blog/?p=1799. Balduzzi, David, and Giulio Tononi. 2008. Integrated information in discrete dynamical systems: motivation and theoretical framework. PLOS Computational Biology 4 (6): e1000091. http s://doi.org/10.1371/journal.pcbi.1000091. . 2009. Qualia: the geometry of integrated information. PLOS Computational Biology 5 (8): e1000462. https://doi.org/10.1371/journal.pcbi.1000462. Barrett, Adam B., and Anil K. Seth. 2011. Practical measures of integrated information for time- series data. PLOS Computational Biology 7 (1): e1001052. https://doi.org/10.1371/jou rnal.pcbi.1001052. Bateson, Gregory. 1972. Step to ecology of mind. University of Chicago Press. Bayne, Tim. 2018. On the axiomatic foundations of the integrated information theory of con- sciousness. Neuroscience of Consciousness 2018 (1). https://doi.org/10.1093/nc/niy 007. Block, Ned J. 1995. On a confusion about the function of consciousness. Behavioral and Brain Sciences 18:227 -- 247. . 2007. Consciousness, accessibility, and the mesh between psychology and neuroscience. Behavioral and Brain Sciences 30 (5-6): 499 -- 548. 29 Cerullo, Michael A. 2015. The problem with phi: a critique of integrated information theory. PLOS Computational Biology 11 (9): e1004286. https://doi.org/10.1371/journal.pcbi .1004286. Chalmers, David J. 1972. The conscious mind. Philosophy of Mind Series. Oxford University Press. . 1997. Availability: the cognitive basis of consciousness? Behavioral and Brain Sciences 30:148 -- 149. Hidaka, Shohei, and Masafumi Oizumi. 2018. Fast and exact search for the partition with minimal information loss. PLOS ONE 13 (9): e0201126. https://doi.org/10.1371/journal.pone .0201126. Horgan, John. 2015. Can integrated information theory explains consciousness? https://blogs .scientificamerican.com/cross-check/can-integrated-information-theory-explai n-consciousness. Kitazono, Jun, Ryota Kanai, and Masafumi Oizumi. 2018. Efficient algorithms for searching the minimum information partition in integrated information theory. Entropy 20 (3): 173. https://doi.org/10.3390/e20030173. Krohn, Stephan, and Dirk Ostwald. 2017. Computing integrated information. Neuroscience of Consciousness 2017 (1): nix017. https://doi.org/10.1093/nc/nix017. Kuhn, Thomas S. 1962. The structure of scientific revolutions. University of Chicago Press. Marshall, William, Jaime Gomez-Ramirez, and Giulio Tononi. 2016. Integrated information and state differentiation. Frontiers in Psychology 7:926. https : / / doi . org / 10 . 3389 / fpsyg .2016.00926. Oizumi, Masafumi, Larissa Albantakis, and Giulio Tononi. 2014. From the phenomenology to the mechanisms of consciousness: integrated information theory 3.0. PLOS Computational Biology 10 (5): e1003588. https://doi.org/10.1371/journal.pcbi.1003588. Oizumi, Masafumi, Shun-ichi Amari, Toru Yanagawa, Naotaka Fujii, and Naotsugu Tsuchiya. 2016. Measuring integrated information from the decoding perspective. PLOS Computa- tional Biology 12 (1): e1004654. https://doi.org/10.1371/journal.pcbi.1004654. Rosenthal, David M. 1986. Two concepts of consciousness. Philosophical Studies 49:329 -- 359. . 2005. Consciousness and mind. Oxford University Press. Schwitzgebel, Eric. 2008. The unreliability of naive introspection. Philosophical Review 117:245 -- 273. . 2012. Why tononi should think that the united states is conscious. https://schwitzspl inters.blogspot.com/2012/03/why-tononi-should-think-that-united.html. . 2013. Perplexities of consciousness. Life and Mind: Philosophical Issues in Biology and Psychology. The MIT Press. 30 Schwitzgebel, Eric. 2014. Tononi's exclusion postulate would make consciousness (nearly) irrel- evant. https://schwitzsplinters.blogspot.com/2014/07/tononis-exclusion-postul ate-would-make.html. Shanahan, Murray. 2015. Ascribing consciousness to artificial intelligence. https://arxiv.org /abs/1504.05696. Smithies, Declan, and Daniel Stoljar. 2012. Introspection and consciousness. Oxford University Press. Tegmark, Max. 2016. Improved measures of integrated information. PLOS Computational Biology 12 (11): e1005123. https://doi.org/10.1371/journal.pcbi.1005123. . 2017. Life 3.0: being human in the age of artificial intelligence. Penguin UK. Tononi, Giulio. 2001. Information measures for conscious experience. Archives Italiennes de Bi- ologie 139 (4): 367 -- 371. . 2004. An information integration theory of consciousness. BMC Neuroscience 5:42. htt ps://doi.org/10.1186/1471-2202-5-42. . 2008. Consciousness as integrated information: a provisional manifesto. The Biological Bulletin 215 (3): 216 -- 242. https://doi.org/10.2307/25470707. . 2012. Integrated information theory of consciousness: an updated account. Archives Italiennes de Biologie 150 (2-3): 290 -- 326. . 2014. Why scott should stare at a blank wall and reconsider (or, the conscious grid). https://www.scottaaronson.com/blog/?p=1799. Tononi, Giulio, Melanie Boly, Marcello Massimini, and Christoph Koch. 2016. Integrated infor- mation theory: from consciousness to its physical substrate. Nature Reviews Neuroscience 17:450 -- 461. https://doi.org/10.1038/nrn.2016.44. Tononi, Giulio, and Christoph Koch. 2015. Consciousness: here, there, and everywhere? The Royal Society 370 (1668). https://doi.org/10.1098/rstb.2014.0167. Virgil, Griffith, and Christoph Koch. 2014. Prokopenko m. (eds) guided self-organization: incep- tion. emergence, complexity and computation. Chap. 6. Springer. 31
1609.08855
1
1609
2016-09-28T10:43:41
Low-dimensional firing rate dynamics of spiking neuron networks
[ "q-bio.NC", "cond-mat.dis-nn" ]
Starting from a spectral expansion of the Fokker-Plank equation for the membrane potential density in a network of spiking neurons, a low-dimensional dynamics of the collective firing rate is derived. As a result a $n$-order ordinary differential equation for the network activity can be worked out by taking into account the slowest $n$ modes of the expansion. The resulting low-dimensional dynamics naturally takes into account the strength of the synaptic couplings under the hypothesis of a not too fast changing membrane potential density. By considering only the two slowest modes, the firing rate dynamics is equivalent to the one of a damped oscillator in which the angular frequency and the relaxation time are state-dependent. The presented results apply to a wide class of networks of one-compartment neuron models.
q-bio.NC
q-bio
Low-dimensional firing rate dynamics of spiking neuron networks Maurizio Mattia Istituto Superiore di Sanit`a viale Regina Elena 299, 00161 Rome, Italy [email protected] May 8, 2018 Abstract Starting from a spectral expansion of the Fokker-Plank equation for the membrane potential density in a network of spiking neurons, a low-dimensional dynamics of the collective firing rate is derived. As a result a n-order ordinary differential equation for the network activity can be worked out by taking into account the slowest n modes of the expansion. The resulting low-dimensional dynamics naturally takes into account the strength of the synaptic couplings under the hypothesis of a not too fast changing membrane potential density. By considering only the two slowest modes, the firing rate dynamics is equivalent to the one of a damped oscillator in which the angular frequency and the relaxation time are state-dependent. The presented results apply to a wide class of networks of one-compartment neuron models. The collective dynamics of neuronal networks can be complex even when simplified one-compartment models of neurons are considered for modeling. Complexity arises from the spiking sparse-in-time nature of the inter-neuronal communication, the high-dimensionality of the system and the quenched randomness in the synaptic couplings. Reducing such complexity relying on mean-field approaches has a long history in theoretical neuroscience [15, 2, 5, 13, 12, 11], although adopted approximations often limit the general applicability of the resulting simplified dynamics. Here, with the aim to widen such effectiveness by relying on a population density approach [14, 1, 8, 3, 7, 10], a low-dimensional dynamic mean-field description is provided for the instantaneous emission/firing rate ν(t) of a network of spiking neurons. In this framework, the network dynamics is described by the Fokker-Planck (FP) equation for the membrane potential density. A suited spectral expansion of the FP operator [7, 10] is the approach followed focusing on a homogeneous pool of interacting integrate-and-fire (IF) neurons. Notation summary To start, some of the results in [10] and a brief description of the adopted notation are provided. Under mean-field approximation the density p(v, t) of neurons with membrane potential v at time t follows a FP equation with suited boundary conditions: (1) In general, the FP operator L ≡ L(p) is nonlinear because it depends on the instantaneous firing rate ν(t) given by the flux of realizations crossing the emission threshold θ: ∂t p(v, t) = L p(v, t) ≡ −∂v Sp(v, t) . ν(t) = Sp(θ, t) . 1 The spectrum {λn} of the FP operator, Lφn(cid:105) = λn φn(cid:105) , provides a moving basis {φn(cid:105)} driven by the first two instantaneous moments of incoming currents, which are time-varying and state-dependent. Equation (1) can be rewritten as the dynamics of the expansion coefficients an(t) of the density p(v, t) on such basis: (cid:88) n p(cid:105) = anφn(cid:105) . An emission rate equation results:(cid:40) (cid:126)a = (Λ + C ν) (cid:126)a + (cid:126)c ν ν = Φ + (cid:126)f · (cid:126)a , (2) where (cid:126)a = {an}n(cid:54)=0. (cid:126)f = {Sφn(θ, t)}n(cid:54)=0 are the fluxes in θ for non-stationary modes (n (cid:54)= 0). Synaptic coupling in stationary (n = 0) and non-stationary modes is expressed in the vector (cid:126)c, cn = (cid:104)∂ν ψnφ0(cid:105) ∀n (cid:54)= 0, and the matrix C, Cnm = (cid:104)∂ν ψnφm(cid:105) ∀n, m (cid:54)= 0 , where (cid:104)ψn are the eigenfunctions of the adjoint operator L+. Λ is the diagonal matrix of the eigenvalues of L Λnm = λn δnm ∀n, m (cid:54)= 0. Low-dimensional ordinary differential equation for ν A set of uncoupled IF neurons In absence of synaptic coupling, incoming currents to the neurons do not depend on ν(t). Hence, both (cid:126)c = 0 and C = 0, and eigenfunctions of L are independent from ν(t). Under these conditions Eq. (2) simplifies as , (3) (cid:40) (cid:126)a = Λ (cid:126)a ν = Φ + (cid:126)f · (cid:126)a with constant coefficients Φ, (cid:126)f and Λ. Recursively deriving with respect to time both equations in (3), one obtains: ν = (cid:126)f · (cid:126)a ν = (cid:126)f · (cid:126)a t ν = (cid:126)f · ∂n ∂n t (cid:126)a ··· and such that t (cid:126)a = Λ ∂n−1 ∂n t (cid:126)a = Λ2 ∂n−2 (cid:126)a = ··· = Λn (cid:126)a , t t ν = (cid:126)f · Λn (cid:126)a . ∂n 2 Stopping derivations to the n-th order, an approximated expression for the emission rate equation results: and  .  ν ν ... ∂n t ν ν = Φ + f1 a1 + f2 a2 + ··· + fn an  = (cid:88) n f2 a2 ... λ2 1 λ2 2 ... ... λn 1 λn 2 ··· λn ··· λ2 ... . . . ··· λn λ1 λ2  f1 a1 ν + ··· + (−1)n (cid:89) fn an n 1 λj 0<j≤n The coefficient matrix of this linear system is the Vandermonde's matrix (in a non-classical form), for which is known an explicit expression of its inverse. Hence, the solutions for the unknown {fi ai}n≥i>0 can be worked out allowing to rewrite the approximated emission rate equation as: t ν = Φ − ν ∂n − (cid:88) ν + (4) 1 1 λj 0<j≤n λj λk 0<k≤n 0<j<k Note that, this ordinary differential equation (ODE) depends only on the eigenvalues {λn} and the gain function Φ. Assuming an ordered spectrum of L, such that Re(λn) ≥ Re(λm) if n > m, this n-th order approximation of the emission rate equation is neglecting the dynamics at timescales smaller than 1/ Re(λ1)n. As for IF neurons the eigenvalues are hierarchically distributed in couples of complex conjugates or in couples with similar real values, the second order (n = 2) approximation can be effective enough: (5) Here, n = +1,−1 has been used for convenience instead of n = 1, 2. An example subset of eigenvalues is shown in Fig. 1 for VIF neurons, a simplified IF neuron with constant leakage and a reflecting barrier in v = 0 [4, 10]. λ+1 λ−1 λ+1 ν + + 1 ν = Φ − ν . (cid:18) 1 − (cid:19) 1 λ−1 Figure 1. Distribution of eigenvalues for VIF neurons. First 3 couples of modes are considered. Mean µ and standard deviation σ of the input current are changed keeping constant the emission rate at ν = 5 Hz. Darker and lighter colors are for larger and smaller Re(λn), respectively. Red (blue) curves are for λn with n > 0 (n < 0). A, λn distribution on the complex plane (Re(λ), Im(λ)). Eigenvalues are complex conjugates at positive drift (µ > 0). Angular frequency at positive drift is expected to be ω0 = 2π n ν. B-C, real and imaginary parts of the eigenvalues. 3 -4-224-1200-800-4000m (q/s)-1Re l (s)nB-1200-800-400-100-50501000-1Re l (s)n-6pn-4pn-1Im l (s)n-2pn6pn4pn2pnA50100-100-50-4-224m (q/s)-1Im l (s)nl+1l+2l+3l-1l-2l-3C For a set of uncoupled IF neurons, Eq. (5) extends the Wilson-Cowan equation [15] in two ways: i) it is a second-order ODE such that damped oscillations can be described even when afferent currents are constant; ii) equilibrium is approached with decay time and oscillation period (if any) which depend not only on the single-neuron parameters but also on the state of the input current. This can be well appreciated rewriting Eq. (5) as the dynamics of a damped oscillator in the force field established by Φ − ν: (5(cid:48)) τ 2 ν + 2τ ν = (1 + τ 2 ω2 0)(Φ − ν) , where τ is the input-dependent decay time, 1 τ = − λ+1 + λ−1 2 , and ω0 is the angular frequency of the damped oscillations (if real) 0 = − (λ+1 − λ−1)2 ω2 4 . Note that this approximated dynamics is the same as the one in which only the first two non-stationary modes in the spectral expansion are taken into account, and it is equivalent to the complex-valued firing rate dynamics introduced in [12]. An example of τ and ω0 is shown in Fig. 2 for VIF neurons. Figure 2. Decay time τ (A) and oscillation frequency Re(ω0)/2π (B) in Eq. (5(cid:48)) for VIF neurons at different µ and σ. Light and dark red lines, set of (µ, σ) at constant firing rate ν (see colored labels). In B, iso-frequency curves overlap contour lines of ω0 at positive drift µ and vanishing σ, as expected. A network of synaptically coupled IF neurons A similar approach can be applied to the case of a homogeneous pool of synaptically coupled neurons. Now all functions like Φ, (cid:126)f and Λ, depend implicitly on the emission rate ν(t). This because all these functions depend on the moments of incoming currents, which in turn are modulated by the recurrent activity of the network. First time derivatives of the expression in Eq. (2) for the stationary mode are ν(1 − Φ(cid:48)) = (cid:126)f(cid:48) · (cid:126)a ν + (cid:126)f · (cid:126)a and ν(1 − Φ(cid:48)) − Φ(cid:48)(cid:48) ν2 = (cid:126)f(cid:48)(cid:48) · (cid:126)a ν2 + 2 (cid:126)f(cid:48) · (cid:126)a ν + (cid:126)f(cid:48) · (cid:126)a ν + (cid:126)f · (cid:126)a , 4 0.51.01.52.02.500.050.100.150.20t (s)1/2s (q/s)-4-2024m (q/s)n = 2 Hzn = 4 Hzn = 6 Hzn = 8 Hz0.51.01.52.02.51/2s (q/s)-4-2024m (q/s)n = 2 Hzn = 4 Hzn = 6 Hzn = 8 Hz02468Re(w)/2p (Hz)0AB where Φ(cid:48) = ∂ν Φ, Φ(cid:48)(cid:48) = ∂2 ν Φ, and the same holds for (cid:126)f(cid:48) and (cid:126)f(cid:48)(cid:48). From the expression for non-stationary modes in Eq. (2), a bit lengthy formula for (cid:126)a can be worked (cid:126)a =(cid:2)Λ(cid:48) ν + C(cid:48) ν2 + C ν + (Λ + C ν)2(cid:3) (cid:126)a + (Λ + C ν) (cid:126)c ν + (cid:126)c(cid:48) ν2 + (cid:126)c ν . out: Putting together all these expressions a not so useful approximation for the emission rate equation results. Things change if a not too fast dynamics is considered. Because the adopted moving basis {φn(cid:105)} rapidly adapts to the continuously varying moments of incoming currents, the hypothesis of a slow enough ν(t) dynamics allows to consider the coefficients an of the same order of the time derivative of the input: an = O( ν). So far, for small enough ν a good approximation is to neglect all the nonlinear terms (cid:126)a ν, (cid:126)a ν2, (cid:126)a ν, ν2 and (cid:126)a ν. Above expressions are then simplified as follows: and (cid:126)a = Λ2 (cid:126)a + Λ (cid:126)c (cid:126)ν + (cid:126)c ν ν(1 − Φ(cid:48) − (cid:126)f · (cid:126)c) = (cid:126)f · Λ (cid:126)a ν(1 − Φ(cid:48) − (cid:126)f · (cid:126)c) − ν (cid:126)f · Λ (cid:126)c = (cid:126)f · Λ2 (cid:126)a . This set of equations together with the one for the stationary mode are a linear system equivalent to the one shown in the derivation of Eq. (4), and it can be solved in the same way: (cid:18) 1 − + 1 λ−1 λ+1 (cid:19)(cid:32) 1 − Φ(cid:48) − (cid:126)f · Λ−1 (cid:126)c Tr(Λ−1) (cid:33) ν + 1 λ+1 λ−1 (1 − Φ(cid:48) − (cid:126)f · (cid:126)c) ν = Φ − ν . (6) Here a second order ODE for ν(t) is recovered as for the uncoupled case in Eq. (5). It can be usefully rewritten as: (6(cid:48)) α2(ν) ν + α1(ν) ν = Φ − ν , where (cid:126)f · Λ−1 (cid:126)c = f+1 c+1/λ+1 + f−1 c−1/λ−1 and Tr(Λ−1) = 1/λ+1 + 1/λ−1. Of course, in the limit of synaptically uncoupled neurons (Φ(cid:48) = 0 and (cid:126)c = 0) these coefficients are the same as in Eq. (5). State-dependent decay time and angular frequency are now τ (ν) = 2α2(ν) α1(ν) ω0(ν)2 = 1 α2(ν) − 1 τ (ν)2 , respectively. In Fig. 3 both τ and ω0 are shown for different synaptic efficacies J. Infinitesimal moments µ and σ of the input current to the VIF neurons are varied keeping constant the output firing at ν = 5 Hz. Two remarks: i) τ is independent from the coupling intensity; ii) oscillatory behavior appears even at negative drift (µ < 0) when J > 0, and it can disappear at positive drift for suited inhibitory feedbacks (J < 0). 5 with α1(ν) = − (cid:19)(cid:32) + 1 λ−1 1 − Φ(cid:48) − (cid:126)f · Λ−1 (cid:126)c Tr(Λ−1) (cid:33) , (cid:18) 1 λ+1 1 α2(ν) = λ+1 λ−1 (1 − Φ(cid:48) − (cid:126)f · (cid:126)c) Figure 3. Decay time τ (A) and oscillation frequency Re(ω0)/2π (B) versus mean afferent current µ for different synaptic efficacies J. As in the other figures, VIF neurons are considered. Total infinitesimal mean is assumed to be µ = µ0 +  J ν, where µ0 is the mean of external currents and  = 0.2 is the probability to have two neurons synaptically coupled. The mean-field infinitesimal variance is σ2 = σ2 variance of external currents. Both in panels A and B, J is varied from 0 to 0.5θ at steps of 0.1θ (curve colors from indigo to red, respectively). Firing rate is kept constant (ν = 5 Hz) by varying µ0 and σ2 0 accordingly. C, contour lines of angular frequencies shown for both excitatory (J > 0) and inhibitory (J < 0) synaptic couplings. Damped oscillations (ω0 > 0) appear at both positive and negative drifts (black border). Contour lines correspond to constant Re(ω0)/2π from 0 Hz to 5.5 Hz at steps of 0.5 Hz (from dark to light blue, respectively). 0 +  J 2 ν, where σ2 0 is the For VIF neurons is also interesting to see how τ varies at different firing rates, as shown in Fig. 4. Time scales are increasingly shorter for large firing rates, as expected. Furthermore, 1/τ is well approximated by a linear combination of µ and ν. This should not surprise because for VIF neurons Re(λ±1) (cid:39) −2π2 σ2/θ2, and in the plane (µ, σ) iso-frequency curves are rotated parabolas: σ(µ)2 − σ(0)2 ∝ µ (see Fig. 2 and [10]). Such relationship for other kinds of IF neurons is expected to depend on the particular shape of their current-to-rate gain functions Φ. 0 → ext J 2 The case of non-stationary external currents The introduced approach can be used also in presence of non-stationary external currents ( µ0 (cid:54)= 0 and σ0 (cid:54)= 0). Considering synaptic current also due to the spiking activity of neurons outside the network, and firing at rate νext(t), the first two moments of this additional current become µ0 → ext Jext νext +µ0 and σ2 0. What shown in the previous Section applies here provided that also the derivatives with respect to νext are considered. An alternative way to reduce the dimensionality of the emission rate dynamics in this case, is to rely on a more straightforward adiabatic approximation [6, 9]. For very slow external currents ( νext (cid:28) Re(λ±1)), the moving basis instantaneously adapts to the time varying input, and its movement is only mildly driven by the intrinsic timescales determined by the eigenvalues of L. Under this hypothesis, (cid:126)a (cid:39) 0 and Eq. (2) reduces to ext νext + σ2 (cid:26) 0 = Λ (cid:126)a + (cid:126)c ν + (cid:126)cext νext ν = Φ + (cid:126)f · (cid:126)a , where the elements of (cid:126)cext are obtained deriving (cid:104)ψn with respect to νext and the nonlinear terms ν (cid:126)a are neglected. Finally, the following first order dynamics results τ ν = Φ − ν + F , (7) where τ (ν) = (cid:126)f · Λ−1 (cid:126)c is the activity dependent timescale and F (t) = − (cid:126)f · Λ−1 (cid:126)cext νext is an additional time-dependent forcing term. Although, all the modes of the expansion here can be considered, this approximation cannot express damped oscillations. In order words this approximation is neglecting the intrinsic diffusive dynamics related to the λn. 6 -4-224510152025300m (q/s)t (ms)A123456-4-2240m (q/s)Re(w)/2p (Hz)00J = 0.5 qJ = 0 qBn = 5 Hz-0.4-0.20.00.20.4-4-2240m (q/s)J (q)5.5 Hz5 Hz4.5 Hz0 HzC Figure 4. Decay time τ (A) versus mean current µ at different constant firing rates ν. ν varies from 5 Hz to 25 Hz at steps of 4 Hz (colors from light to dark red, respectively). B, 1/τ versus µ, colors as in panel A. Acknowledgments I thank E. S. Schaffer and L. F. Abbott for a stimulating discussion from which originated my interest in the presented simplification of the FP equation. I also thank M. Augustin for a careful reading of the manuscript and the many discussions about it, and E. Hugues for an interesting feedback on the uncoupled neuron case. References [1] L. F. Abbott and Carl van Vreeswijk. Asynchronous states in networks of pulse-coupled oscillators. Phys. Rev. E, 48(2):1483–1490, 1993. [2] Daniel J. Amit and M Tsodyks. Quantitative study of attractor neural network retrieving at low spike rates: I. substrate–spikes, rates and neuronal gain. Network, 2(3):259–273, 1991. [3] Nicolas Brunel and Vincent Hakim. Fast global oscillations in networks of integrate-and-fire neurons with low firing rates. Neural Comput., 11(7):1621–71, 1999. [4] Stefano Fusi and Maurizio Mattia. Collective behavior of networks with linear (VLSI) integrate- and-fire neurons. Neural Comput., 11(3):633–52, 1999. [5] Wulfram Gerstner. Time structure of the activity in neural network models. Phys. Rev. E, 51(1):738–758, 1995. [6] Guido Gigante, Maurizio Mattia, and Paolo Del Giudice. Diverse population-bursting modes of adapting spiking neurons. Phys. Rev. Lett., 98(14):148101, 2007. [7] Bruce W Knight. Dynamics of encoding in neuron populations: some general mathematical fea- tures. Neural Comput., 12(3):473–518, 2000. [8] Bruce W Knight, D Manin, and Lawrence Sirovich. Dynamical models of interacting neuron popula- tions in visual cortex. In E.C. Gerf, editor, Symposium on Robotics and Cybernetics: Computational Engineering in Systems Applications, Cite Scientifique, Lille, France, 1996. [9] Daniele Linaro, Marco Storace, and Maurizio Mattia. Inferring network dynamics and neuron properties from population recordings. Front. Comput. Neurosci., 5(00043), 2011. [10] Maurizio Mattia and Paolo Del Giudice. Population dynamics of interacting spiking neurons. Phys. Rev. E, 66:051917, 2002. 7 -4-224510152025300m (q/s)t (ms)n = 5 Hzn = 10 HzA100200300400500-4-2240m (q/s)0-11/t (s)n = 25 Hzn = 5 HzB [11] Ernest Montbri´o, Diego Paz´o, and Alex Roxin. Macroscopic description for networks of spiking neurons. Phys. Rev. X, 5(2):021028, 2015. [12] Evan S Schaffer, Srdjan Ostojic, and Larry F Abbott. A Complex-Valued Firing-Rate Model That Approximates the Dynamics of Spiking Networks. PLoS Comput. Biol., 9(10):e1003301, 2013. [13] O Shriki, David Hansel, and Haim Sompolinsky. Rate models for conductance-based cortical neuronal networks. Neural Comput., 15(8):1809–41, 2003. [14] Alessandro Treves. Mean-field analysis of neuronal spike dynamics. Network, 4(3):259–84, 1993. [15] Hugh R. Wilson and Jack D. Cowan. Excitatory and inhibitory interactions in localized populations of model neurons. Biophys. J., 12(1):1–24, 1972. 8
1706.04192
4
1706
2017-10-31T19:05:38
A Theoretical Solution of the Mind-Body Problem: An Operationalized Proof that no Purely Physical System Can Exhibit all the Properties of Human Consciousness
[ "q-bio.NC" ]
This article presents an operationalized solution to the mind-body problem which relies on rigorously defined theoretical reasoning rather than philosophical argument. We identify a specific operation which is a necessary property of all healthy human conscious individuals -- specifically the operation of self-certainty, or the capacity of healthy conscious humans to "know" with certainty that they are conscious. This operation is shown to be inconsistent with the properties possible in any meaningful definition of a physical system. This inconsistency is demonstrated by proving a "no-go" theorem for any physical system capable of human logical reasoning, if this reasoning is required to be both sound and consistent. The proof of this theorem is both general -- it applies to any function whereby evidence affects the state of some physical system -- and recursive, since any physical process subserving a function of this type is shown to imply another such function. Thus for at least one aspect of human consciousness, the mind-body problem is now conclusively resolved.
q-bio.NC
q-bio
A Theoretical Solution of the Mind-Body Problem: An Operationalized Proof that no Purely Physical System Can Exhibit all the Properties of Human ConsciousnessCatherine M Reason1London, United KingdomThis article presents an operationalized solution to the mind-body problem which relies on rigorously defined theoretical reasoning rather than philosophical argument. We identify a specific operation which is a necessary property of all healthy human conscious individuals -- specifically the operation of self-certainty, or the capacity of healthy conscious humans to "know" with certainty that they are conscious. This operation is shown to be inconsistent with the properties possible in any meaningful definition of a physical system. This inconsistency is demonstrated by proving a "no-go" theorem for any physical system capable of human logical reasoning, if this reasoning is required to be both sound and consistent. The proof of this theorem is both general -- it applies to any function whereby evidence affects the state of some physical system -- and recursive, since any physical process subserving a function of this type is shown to imply another such function. Thus forat least one aspect of human consciousness, the mind-body problem is now conclusively resolved.IntroductionOf all the problems in the study of human consciousness, the mind-body problem has undoubtedly proven the most intractable to date. Chalmers (1996) has referred to this as the Hard Problem, of consciousness, and a study of the literature on this topic finds little agreement as to its likely solution. Some approaches (see for example Churchland, 1996; Dennett 1996; Hacker 2010) deny the problem even exists at all. So far, however, all approaches to the mind-body problem have been philosophical in character, and scientific analysis has tended to avoid the question. Chalmers' own approach to the problem relies on the notion of "philosophical zombies" -- beings physically identical to conscious human beings but without subjective experience. Chalmers' arguments are typically philosophical in nature, relying on concepts which are expressed in ambiguous verbal forms and which therefore provide opportunities for considerable misinterpretation. The beginnings of a way out of this impasse came in 1995 when Gilbert Caplain published an outline for a proof that consciousness could not be a computational property using arguments drawn from computer science and logic, rather than philosophy. Caplain's proof was expressed only in outline, and necessarily relied on 1 To whom correspondence should be addressed at [email protected] much philosophical terminology such as knowledge, justification and belief without giving precise theoretical definitions of these terms. However Caplain's proof did incorporate one highly innovative development -- it did not rely on some general, poorly defined notion of consciousness as an entity, but instead focused on one particular property associated with human consciousness. Furthermore the property Caplain chose, the capacity of human beings to know with certainty that they are conscious, is a property that can be represented as a specific operation. Thus Caplain pointed the way to an operationalized approach to the mind-body problem, in which undefined or poorly defined notions of consciousness or subjectivity could be replaced by rigorously defined operations. Caplain's proof therefore offered the promise of being expressible not as a philosophical argument, but as a logical theoremderiving from rigorously defined premises.In a recent paper the present author (Reason 2016) attempted to expand Caplain's outline, by showing how it could be generalized to all physical systems and by attempting to remove all ambiguity arising from the use of undefined philosophical terms. Informal communications subsequently revealed this attempt to be only partially successful, and considerable ambiguity remained over the use of such terms as certainty and reliability. Two issues quickly became apparent from these informal communications. Firstly, that most readers were of the view that there must be something wrong with the proof; that it was in some way artificial, or contrived, or "playing with words". Secondly, that it was extremely difficult for some readers to disentangle their thinking from their intuitions and introspections about their own conscious experience. The only way of dealing with these difficulties is to present theproof in as rigorous a manner as I am able, even if the level of detail required for this sometimes looks like overkill.The purpose of this current paper is threefold: Firstly, to explain the logic of Caplain'sproof in an intuitive, easily accessible form; secondly, to illustrate how an operationalapproach can convert a verbal philosophical argument into a rigorous theoretical proof; and thirdly, to express Caplain's proof in just such a rigorous, theoretical form. Philosophers who read this paper are asked to bear in mind that there are important differences between a theoretical proof and a philosophical argument; many of the misconceptions which philosophers have had about this proof are the result of not fully appreciating these differences. Some of these misconceptions will be listed toward the end of this article.The rationale behind Caplain's proofThe theoretical argument underlying Caplain's proof is extremely simple, and is in fact simply an extension of Descartes' evil demon thought experiment -- which tends to be presented in modern formulations as a "brain in a vat". For those unfamiliar with this thought experiment, here is a brief summary. We are asked to consider the possibility that we are in fact disembodied brains, kept alive in tanks or vats of nutrients, and that all our experiences of the outside world are in fact illusions created through direct stimulation of our sensory receptors by electrical or other means. The situation is in fact exactly that described in the well-known "Matrix" series of movies.The upshot of this is that our sensory experience is identical regardless of whether it is2 ultimately generated by the external world we think we perceive, or whether it is generated directly by means of some sort of deceptive apparatus. So we can never know for certain, on the basis of sensory experience, whether the external world reallyexists.Caplain's proof simply extends this thought experiment from sensory perception to introspection. In a computing machine, the output of every act of either perception or introspection must be represented by some sort of symbolic state. This state must be the result of some computational process, and the only way of establishing the accuracy of that process is by the application of some other process. This leads to an infinite regress along the lines of Munchhausen's Trilemma.The version of Caplain's proof which I am presenting here is constructed according to the following rationale. First, we operationalize consciousness in terms of an operation I shall call self-certainty, which will be described more fully in the next section. This enables us to represent consciousness as a simple function (the function of answering a YES/NO question). Self-certainty is thus a specific example of a general class of functions, consisting of all functions which answer YES/NO questions.If we can then show that, for a given class of physical systems, no YES/NO question can be answered with certainty, we will have shown that self-certainty is impossible inthat class of physical systems. The class of physical systems we choose is that class of physical systems which can exhibit humanlike logical reasoning, since human beings must by definition exhibit humanlike reasoning. Any system capable of humanlike logical reasoning we shall call H.The proof which is to be presented here is the sequence of inferences in H which establish that no physical system reasoning in H can answer any YES/NO question with certainty. Therefore, if this proof A is sound, no physical system capable of reasoning in H is capable of self-certainty.It follows that, if we assume that all healthy conscious human beings are capable of self-certainty, we have shown that no healthy human being can be an exclusively physical system.Explaining the Operational ApproachConsider the situation in this somewhat contrived little thought experiment:Two philosophers are arguing about whether buses run on time in northwest London. Philosopher A wants to know whether the H12 bus runs on time. He decides to choose the Pinner High Street bus stop and counts how often the bus is on time at that particular stop. He finds the bus is late more often than not at that particular stop, andso concludes that the H12 bus does not run on time. So here we have:Premise 1: If the H12 bus runs on time then it will stop at Pinner High Street on time;3 Premise 2: The H12 bus does not stop at Pinner High Street on time;Conclusion: The H12 does not run on time.Clearly one can argue about the sampling difficulties associated with this little experiment, but let us leave that aside. Instead we introduce Philosopher B, who examines the experiment and declares the first premise to be unsound. According to Philosopher B, buses stop for all sorts of reasons; they may break down, or they may get stuck in traffic. Or it may be a public holiday and the buses are not running at all. It does not follow, says Philosopher B, that because buses do not stop at some particular place, and at some particular time, they must be running late. Philosopher A, according to Philosopher B, has just made an unsound assumption about the meaning of the term "bus stop".Of course it should be simple to work this out. After all, everyone knows what a bus stop is and everyone should know what Philosopher A means by a bus stop. But philosophy is prone to a rather peculiar disorder in situations like these, which I shall term a "sigh of relief and stop thinking" reaction. By which I mean that Philosopher B has found an ambiguous meaning which he can misinterpret to his benefit, and so immediately sighs with relief and stops thinking.Philosopher A now returns to his study and publishes an explanation of exactly what he means by a bus stop, and why his experiment depends only on that meaning of the term "bus stop", and not on any other meaning. But by this time Philosopher B has moved on. He has completely forgotten the point of the original argument and is now making general assertions such as "There is no reason to suppose any relationship between the lateness of buses and the timing of their stops. In fact the whole notion of a 'bus stop' probably lacks any coherent meaning to begin with."The purpose of this admittedly rather absurd scenario is to illustrate a situation which will be familiar to anyone who has ever followed a philosophical debate; a situation which might be called "motivated misunderstanding". Philosophical arguments are phrased in natural language terms, but their referents often involve metaphysical issues, or other issues which are hard to define unambiguously. Chalmers (2011) refers to this type of problem as a "verbal dispute", and suggests various ways of resolving the difficulty. However his methods of resolution depend on the assumptionthat both parties to a dispute are motivated to come to a common understanding. One might suspect, given human nature, that some parties to a dispute might not have an incentive to arrive at a common understanding if the result is that their particular views are disproved or discredited. Such individuals would therefore be motivated to prevent the ambiguity being resolved. It would be difficult to arrive at a common understanding in such a case.In science this situation is avoided, because scientific investigations try to avoid metaphysical questions and to concentrate on problems where the relevant factors can be strictly defined, observed and measured. The operational approach is one way of doing this (Bridgeman 1927). In the operational approach to measurement, for example, one avoids the temptation to regard measurements of physical quantities as 4 "proxies" for some sort of metaphysical property. Consider the operation of measuring the length of some physical object by placing a measuring rod alongside it.One might be tempted to regard the reading from the measuring rod as a proxy of some sort for some metaphysical property called "length" which cannot be directly observed. However this raises inevitable questions about how the relationship between the proxy and the underlying metaphysical property can be established, giventhat the latter cannot be directly observed. These difficulties are circumvented if one regards the physical quantity "length" to be just the result obtained by placing a measuring rod alongside something. The output of the measuring operation is therefore to be taken as synonymous with the physical property "length". One can thus carry out measurements of physical quantities without worrying about what metaphysical properties, if any, may be associated with those quantities.Of course there are limits to how far one can apply this approach to questions of interest to philosophers, since issues of metaphysics are fundamental to many of thesequestions. The nature of consciousness -- in particular the so-called "Hard Problem" of consciousness -- is a case in point. Caplain's work, however, shows that even in thestudy of consciousness, there are problems which can be treated by the operational approach and so lifted from the realm of philosophical speculation into the domain of theoretical and scientific analysis.To see how this is so, we must note that Caplain's proof applies not to some indefinable metaphysical quality of consciousness, but to a very specific and well-defined operation -- that is, the operation whereby a conscious human being comes to "know" (in Caplain's words) with certainty that they are conscious. This operation can be expressed as the process of answering a binary question -- a question for whichthe answer is either YES or NO. If a conscious human subject asks the question "Am I certainly conscious?" , then there should be some operation associated with that subject's consciousness which is capable of giving the answer YES to that question. One can then ask whether that operation is one that can be performed by some physical system (although Caplain's original outline proof applied only to computing machines). If it turns out that the answer to this question is no, then the mind-body problem is at least partially resolved without any consideration having to be given to the metaphysical nature of consciousness itself. That is to say, one could prove that consciousness cannot wholly be a physical process without worrying about what consciousness itself actually is. This advantage of the operational approach saves us a lot of work. For example, one currently fashionable explanation for the Hard Problemis Illusionism (see for example Frankish 2016, Blackmore 2016). Illusionism is basedon the premise that the properties of consciousness are not what they subjectively appear to be. But the operational approach enables us the sidestep the question of whether this doctrine is true, or even whether it is coherent. Of course in order to do this one needs to do far more than just operationalize consciousness. One also needs to operationalize the concept of a physical system itself, and also define rigorously any other assumptions which are necessary to demonstrate an inconsistency between the operation of ascertaining that one is certainly conscious and the properties of physical systems. How might this be done?If our intention is to demonstrate an inconsistency between an operationalized 5 definition of consciousness and an operationalized definition of a physical system, then it is essential that any definitions and terms employed be unambiguous. Any terms associated with metaphysics, such as belief, or knowledge, or justification, should be avoided since these are likely to be interpreted in subtly different ways by different readers, and these different ways are likely to have hidden assumptions built into them. Likewise it is important not to get bogged down in the quagmire of humanepistemology -- we are not interested in the question of how human beings "know" things, or "understand" things, or how they become certain of things. All such questions are irrelevant to our task and any terminology relating to such questions is to be avoided. To the extent we are interested in epistemology at all, we are interestedonly in the epistemology of physical systems.Assumptions necessary for an operationalized formulation of the proofIn order to express the proof in a form that can reasonably be described as rigorous, it is necessary to list the assumptions and premises on which his proof depends, in a clear and unambiguous form. This we shall now proceed to do2. We first need some sort of definition of a physical system. It should be borne in mind here that the purpose of our definitions is to eliminate ambiguity, not to provide a comprehensive description. We can therefore operationalize the definition of a physical system in terms of some general property which a physical system must possess, without attempting to capture the full meaning of the term "physical system".We therefore define a physical system as some set of physical processes, where a physical process is defined as follows:Definition: A physical process is some entity which has an objective existence and is capable of evolving in time.Philosophers use the notion of supervenience to refer to entities which are not themselves physical systems but whose existence is nonetheless consistent with the doctrine of physicalism. An entity S is said to be supervenient on some basis B if two things which have the same B properties must have the same S properties (Kim 1984). Mental processes are thus assumed to supervene on physical processes. Hence the following expression:Principle of Physicalism3: All mental processes supervene on some physical process or set of processes.(The notion of a mental process will be expressed more formally in the next section.) Our operationalized notion of consciousness also needs to be expressed a little more specifically. We shall call this notion self-certainty:Definition: Self-certainty is the capacity of at least some conscious beings to verify 2The premises listed here were originally given in Reason (2016). They are presented here in an expanded form, together with explanations as to why they have been chosen.3 Referred to as Principle F' in Reason (2016)6 with absolute certainty that they are conscious -- that is, to give the answer YES to thequestion "Am I certainly conscious?"It is important to emphasize that it is only absolute certainty which is at issue here. Caplain's proof does not imply that physical systems cannot be reasonably certain of being conscious, or that they might have a certainty of being conscious which is contingent on additional assumptions. For example Chalmers (2012) postulates an analogous situation involving mathematical reasoning; in this example an expert mathematician is given cause to doubt the accuracy of his own mathematical musings by being offered the possibility that he has been administered some sort of drug whichdestroys the capacity for mathematical reasoning. Chalmers suggests that the mathematician could regain confidence in the accuracy of his own cogitations by insulating himself from all such questions (that is to say, by ignoring them). Certainty of this sort, which depends on additional assumptions, is not covered by Caplain's proof and does not constitute self-certainty. We can express this as:Condition 1: An entity is self-certain only if its certainty is absolute (beyond any possibility of error) and does not depend on additional assumptions.It is also necessary to emphasize that self-certainty does not require assumptions to bemade about the nature of consciousness. Any such assumptions -- for example, that consciousness equates to wakefulness, or that consciousness requires some specific sort of self-awareness -- should therefore not be made. A conscious being might, for example, be experiencing a lucid dream or some other altered state of consciousness. So long as the conscious being can verify that it is in some conscious state or other, the requirement for self-certainty is satisfied. We express this as:Condition 2: Self-certainty does not require that the conscious state which is found to exist has any particular property or set of properties.4 Our next two premises refer to the rules of logical inference by which Caplain's proof can be established. Our objective is to prove Caplain's result as a rigorous theorem, but it must be admitted that this cannot be done in the context of any specified formal system. Even if such a formal system should exist, we do not know what it is. Fortunately, however, these details can be overlooked. If we start with the assumption that our own ability to reason logically, as human beings, is in principle sound5 (which is a necessary assumption for any sort of theoretical work whatsoever) 4The significance of this condition cannot be overstated. Philosophers who find the operational approach objectionable (and this includes just about every philosopher who has ever contacted me on the subject) argue that nothing can be inferred about the nature of consciousness from such operational definitions. Condition 2 illustrates that this situation is intentional. The no-go theorem to be proven in the following section depends on this operational definition alone, and not on any philosophical notionsabout what consciousness is or should be. To underline this, one can express self-certainty as illusory self-certainty, as in the sentence "It is absolutely certain that I have at least the illusion of being conscious" without affecting the validity of the proof.5 At first glance this may seem to be inconsistent with Condition 1 -- however this is not the case. Condition 1 requires only that self-certainty does not require additional assumptions ; however it is necessary to make additional assumptions in order to go beyond self-certainty and to prove Caplain's result.7 then we can conclude that the necessary rules of inference are precisely those which enable us, as human beings, to reason logically. Any system which incorporates theseparticular rules of inference, we shall call H. This leads to the following two assumptions:Assumption 1: There exists a physical system M which supports the logical system H;Assumption 2: H is in principle both sound and consistent."Sound in principle" simply means that if a system reasoning with H makes a mistake,there is no reason in principle why it should eventually not correct that mistake.6 H does not therefore condemn any system using it to fallacious logical reasoning which can never be corrected. By "correct", we mean of course that H incorporates a capacity for classical inference. There may well be many different formal systems, with different sets of axioms and different rules of inference, which exhibit the necessary properties of H, or it may be that there is no such formal system. By defining H in the way we do, we manage to avoid this question entirely. However since H is defined only implicitly, derivations in H cannot be given in the way they would be for formally specified systems. Such derivations must be presented instead in terms of the intuitive logic which supervenes on H. For this reason the following proof must be presented in somewhat tedious detail, in order that each inference may be checked individually by the reader.An operational formulation of Caplain's proofWe are now in a position to present Caplain's proof in a rigorous form. The proof willnow be presented in a number of stages, and each milestone will be clearly identified and numbered. The reader is invited to consider the reasonableness of each inference as it is presented. Should any reader find the accompanying profusion of Greek letters and other algebraic symbols somewhat bewildering, they may find it useful to consult the simpler proof in Reason (2016), in which they do not appear.It will be convenient to restrict our considerations to mental processes which can be represented as functions. We will define M to be some physical system, and j to be some function which yields the answer to a YES/NO question. We shall begin with the simplest case in which all questions can be considered to refer to propositions which are either TRUE or UNTRUE. The function j therefore represents the mappingTRUE ® YESUNTRUE® NO6 Another way of looking at this, which is perhaps strictly more accurate, is to assert that H is sound butthat individual intelligent beings are noisy theorem-provers in H. That is to say, intelligent beings can prove theorems in H but are subject to random errors. However they are not subject to systematic errors.8 where TRUE and UNTRUE are the possible truth-values of some proposition p. We shall call the range of this function the evidence state of p. So if T represents the truth-value of p, and K represents the evidence state of p, the mapping can be represented asK = j(T)We now present this as our first milestone:Milestone 1: Any mental process which is equivalent to the process of answering a YES/NO question can be expressed as a function of the form j.The equivalence covers mental processes which may not themselves be answers to YES/NO questions, but can be represented as such: For example, the mental process of thinking "I exist" is equivalent to the process of answering the YES/NO question "Do I exist?"In a physical system M, an evidence state of some proposition p can be any state of Mcorrelated with the correct truth-value for p, and the function j will be performed by some physical process P, as is required by the Principle of Physicalism. This gives us a general format for mental processes in a physical system. If T is the truth value for some proposition p, and K is the corresponding evidence state for T, thenK = O(P)where O is the output of some physical process P. This is to say that if TRUE is the correct truth value for p, then M will evolve to state KT, and if TRUE is not the correcttruth value for p, then M will evolve to some state nonKT. Thus P executes the mapping:TRUE ® KTnot TRUE ® nonKTwhere KT and nonKT are separate states of M. We shall call this mapping p.It is possible that T is a physical state of M, and furthermore that T is the same physical state as K; KT however cannot be the same physical state as nonKT. Some decision is therefore always required as to whether M evolves to KT or to nonKT. Thisdecision , and the subsequent evolution of M, constitute the process P. This leads us to our second milestone:Milestone 2: Any physical instantiation of a function in the form of j can be expressed as a mapping in the form of p, which will be performed by some physical process P.Now consider the mapping which does the reverse of p, that is the mapping:9 TRUE ® nonKTnot TRUE ® KTa mapping which we shall call r. It is nonetheless clear that r has the same form as p,and by Milestone 2 there will be some physical process which performs it, which we shall call R. We shall express this as our next milestone:Milestone 3: For every mapping of the form p performed by a physical process P, there will be some contradictory mapping of the form r and some physical process R which performs this7.But if K represents the evidence state of some true proposition, how is M to ascertain whether the state K has been produced by P or by R? It clearly matters, because if K has been produced by P then KT will represent the answer YES to some question, but if K has been produced by R, then KT will represent the answer NO to that same question. To resolve this difficulty, it is clearly necessary for M to ascertain whether K has been produced by P or by R!Another way of looking at this is to say that P is an accurate instantiation of j but thatR is an inaccurate instantiation of j. In this case KT will always represent the evidence state YES, but if K has been produced by P then K will be a correct evidencestate, whereas if K has been produced by R then K will be an incorrect evidence state.This brings us to the notion of failure, which is always a necessary consideration when dealing with processes in the real world, as opposed to the idealized mathematical functions they perform. We can express this notion in terms of the following axiom:8Axiom of Fallibility: Given any arbitrarily selected physical process, whose properties have not been ascertained, it is impossible to be certain that this process will or will not perform some given function.In other words, if we have a physical system M in which there exists some evidence state K, produced by some process which is supposed to instantiate a function of the form j, it is impossible to be sure whether that state has been produced by the processP or by the process R without ascertaining any of the properties of the process. We might be able to deduce through logical deduction what that process should be, but wecannot through logical deduction alone deduce what that process actually is in any given case. How could we ascertain the properties of this mysterious process? We could either examine the process directly, or we could examine the mapping performed by the process, which entails examining the states between which the mapping occurs. In 7 This can always be done by applying a logical NOT operator to the truth value before performing j ☻. Therefore if P exists, so does R.8 It is convenient at this point to express the notion of failure as an axiom of H, but it is not strictly necessary. Later we shall show how it is possible to express the notion of failure without assuming the Axiom of Fallibility.10 the former case we could ask, for example, "Does this process have the properties of Por of R?" This is clearly equivalent to asking either the questions "Does this process have the properties of P?" and "Does this process have the properties of R?" These are clearly YES/NO questions and so by Milestone 1 can be expressed as functions of the form j -- let us call this particular function j1. In the second case we could ask, for example, "Does this process map T on to K?" This is also clearly a YES/NO question and so is also expressible as a function of the form j. Let us call this particular function j2. By Milestone 2, any physical instantiation of these functions can be expressed as a mapping of the form p, which will be performed by some physical process. In either case call this process P'.By Milestone 3 there will exist for this mapping a corresponding mapping of the formr, which will be performed by some physical process we shall call R'.From the Axiom of Fallibility it follows that one cannot determine if the actual physical process which instantiates j1 or j2 in the physical system M, is P' or R' (or indeed some completely different third process as yet unidentified) without ascertaining the properties of this actual process. But ascertaining the properties of anactual physical process (as opposed to an idealized one) is equivalent to a function of the form j. And by the Principle of Physicalism, M can only perform functions of theform j by means of some physical process. This leads to the following rather awkward state of affairs:A physical system M can only perform some function of the form j by means of somephysical process;M can only ascertain the correctness of some physical process by performing some function of the form j.This is either circular or leads to an infinite regress. The nature of the infinite regress is easily demonstrated by examining the functions which need to be performed. If thefirst function can be represented as K0 = j0(T0)where K0 is the evidence state of some proposition p0 and T0 is the truth value of p0; then subsequent functions can be represented asKi = ji(Ti)where Ki and Ti are the evidence state and truth value respectively of some proposition pi where pi is equivalent to answering the YES/NO question:"Is the function which answers the question equivalent to pi - 1 correctly performed?"11 Clearly the number of questions, which would need to be answered in order to establish that any given j-type function has been performed correctly, is at least countably infinite.The upshot of all this is that M can never establish with provable accuracy whether any of its processes is accurate or inaccurate9, since establishing this requires a non-terminating sequence of processes. It does not yet follow, however, that M cannot arrive at the conclusion that all its processes are accurate. M might, indeed, be correctto conclude that all of its processes are accurate. However such a conclusion cannot be a logically valid theorem in H; there is no logically sound chain of reasoning in H by which M can show that any of its processes is accurate. To put this in slightly less technical language, it is impossible to prove in H that any of M's processes is accurate.If this is not yet obvious, consider the following: We have established above a chain of reasoning which shows that any attempt by M to establish the accuracy of any of itsprocesses logically implies an infinite regress. We have established this conclusion by means of our own human capacity for logical reasoning; a capacity which we have labeled H. By Assumption 2, H is in principle sound and consistent. If H is sound and consistent, then it cannot support a stable10 chain of reasoning which would lead to a conclusion inconsistent with other conclusions already established in H; that is to say, with other theorems in H. M might be able to derive a logically valid proof in some other system, say nonH, that some of its processes were accurate; but then by definition M could not be a human being.11 One can therefore say that the statement "M cannot prove the accuracy of any of its processes" is equivalent to a theorem in H.If at this point we invoke Assumption 1, then M itself can establish that none of its processes is provably accurate. Since M can reason according to the system H, M canprove any theorem that can be proven by a human being; this includes the proposition "M cannot prove the accuracy of any of its processes". To express this in slightly less formal terms, we can say that if a human being, reasoning according to the principles of H, can show that M cannot establish the accuracy of any of its own processes, then M can do the same.This is an extremely raw and primitive version of the proof we need, which applies only to propositions which are either categorically true or categorically untrue. In real9It would strictly be more accurate to say that the accuracy of M cannot be proven in the logical systemH.10Since H is only sound and consistent in principle, it could temporarily support an invalid chain of reasoning inconsistent with other theorems in H. However such a chain of reasoning would have to be unstable, by which I mean that H would also have to support a chain of reasoning showing this fallacious chain to be invalid. See also Note 5.11 Such a system might, for example, contain axioms which simply assert that certain physical processes are accurate. It would be necessary for such a system to be so constituted that these additional axioms did not lead to contradictions with other properties of the system; this would entail nonH abandoning certain axioms, such as the Axiom of Fallibility, and having different rules of inference from H. The rules of H, for example, allow M to ask the question "Is there any reason why these additional axioms should not be dropped?"12 life, we are often concerned with degrees of probability as much as simple truth or untruth. Often we will have to deal with questions where we simply do not know the answer; where the possible answers to a YES/NO question will be YES, NO and MAYBE. This is easily remedied, however, by expressing such questions in the form:"Is it the case that some proposition p is certainly true?"To avoid ambiguity we shall define any statement of the form "p is certainly true" as meaning that there is zero probability of p being untrue, or that there exists no possible world in which p is untrue. Statements of the form "M is certain of p" will be defined as meaning "M is in the evidence state YES for some particular propositionp".All questions of this type can be expressed as YES/NO questions. We can now encapsulate the reasoning in this section as a specific proof, which we shall call Proof A. This proof establishes that no physical system which possesses reasoning capacities equivalent to those of a human being can correctly answer YES to the following question:"Am I certain that I am answering some given YES/NO question correctly?"We can express this as our next milestone:Milestone 4: For any suitably qualified physical system M (that is, one which is capable of human-level reasoning), and for any proposition equivalent to the answer to a YES/NO question, there can exist some physical process of type R which will answer that question inaccurately. Therefore any suitably qualified physical system can deduce that it can never be certain that its answer to any YES/NO question is correct.This is an extremely general result and it is thus worthwhile checking for any possibleloopholes. The reader might, for example, consider that since Proof A applies only to physical systems, it might not apply to M if M is not certain that it is a physical system. We can deal with this by expressing Milestone 4 in a conditional fashion:"If M is a physical system there can always exist some physical process of type R which will answer any given YES/NO question inaccurately."Clearly this statement is provable in H. In order to establish that Proof A certainly does not apply to M, M must be able to answer "YES to the question:"Is it certainly true that M is not a physical system?"We will call the proposition equivalent to this question p0. M can now ask the question p1:"Is there some process R0 in M that would answer p0 incorrectly?"Whatever the answer to p1, M can now ask the question p2:13 "Is there some process R1 in M that would answer p1 incorrectly?"Clearly we find ourselves in another infinite regress. In fact generally for any proposition pi which can be expressed as a YES/NO question, there will be some proposition pi + 1 of the form:"Is there some process Ri in M that would answer pi incorrectly?"Thus the sequence of questions which need to be answered in order to establish the truth of the proposition"M is certainly not a physical system"is non-terminating. Since M is in fact a physical system, each such question is equivalent to a j-type function which must be performed by some physical process, which means the sequence of such processes is also non-terminating. So long as M isin fact a physical system, it does not matter if M itself is not certain that it is a physical system.Having established this, our first application of Proof A is to the question"Is it certain that I am conscious?"where this question is subject to Conditions 1 and 2 as described in then previous section. Any healthy, conscious human being should be able to answer YES to this question. Proof A, however, shows that no suitably qualified physical system can do so. Conscious human beings, therefore, cannot be exclusively physical systems.Before proceeding to the next section, we will now deal with the outstanding matter of the Axiom of Fallibility. The purpose of this axiom is to express the concept of failure, by asserting that there is no way of establishing through purely logical means whether some particular physical process is working correctly or incorrectly. Some readers, however, and most especially philosophers, may be attracted to the idea of simply dropping this axiom. One can however deal with the notion of failure without expressing it explicitly as an axiom. Consider some function of the form j which is performed by some physical process. M can now ask the question"Is it certain that the process which was supposed to perform the function j did not in fact perform a mapping of the form r?"to which the answer must be YES. The process of answering this question is itself a function of the form j -- call this j'. By the Principle of Physicalism, any physical system which has performed this function must have done so via some physical process. M can now ask the question"Is it certain that the process which was supposed to perform the function j' did not infact perform a mapping of the form r?"14 to which the answer must be YES. Since the process of answering this question is also a function of the form j, which we can call j'', it must by the Principle of Physicalism have been performed by some physical process.. In fact in general any function ji which is assumed to have been performed correctly implies the existence of another function ji + 1 which answers the question equivalent to that assumption. This implies a non-terminating sequence of functions and a corresponding non-terminating sequence of physical processes to perform them. There is, in other words,no escape from the infinite regress by abandoning the Axiom of Fallibility.Verbal summary of the proofProof A has two important properties. Firstly, it is generalizable; it applies to any mapping of the form j. Secondly, it is recursive; it shows that the question of whether any given j-type function is performed correctly can itself be represented as a j-type function, to which Proof A applies. We shall now summarize the proof given in the last section in a somewhat more accessible, verbal form. In doing so we shall illustrate the properties of generality andrecursiveness, and show how these can be used to deal with the apparent objections tothe proof which have been raised informally by some philosophers. It will be useful to base our verbal summary on the notion of evidence; however we shall ignore all considerations about the nature of evidence and instead operationalize the function of evidence as a sequence of YES/NO questions. That is, if M has evidence that some proposition is true, then it is the function of that evidence to answer the question "Is this proposition certainly true?" with either a YES or a NO.This function we shall refer to as the evidence function, which relates the presence or absence of some evidence to the evidence state K of some proposition. If we define a state E so that E is 1 when some evidence exists, and E is 0 when that evidence does not exist, then the evidence function relates E to K as follows:If E = 1 then K = YESIf E = 0 then K = NOBy the Principle of Physicalism, any evidence function in M must be performed by a process which supervenes on some physical process or other. Since we have so far ascertained nothing about this process, we shall call it X.Philosophers, and most particularly epistemologists, traditionally use possible worlds semantics to describe conditions of possibility and necessity. For convenience we shall adhere to this convention henceforth; if some proposition is possibly true, we shall express this by saying there exists some possible world in which that propositionis true. We shall refer to a possible world as epistemically available to M if M cannot rule out the possibility that M exists in such a world, given the evidence currently available to M. We shall say that M is certain of some proposition if M can correctly decide that there is no possible world epistemically available to M, in which the evidence state which corresponds to the truth of that proposition is itself incorrect.15 We shall express the first part of our proof as the following logical argument, which we shall call Castor:It follows from the Principle of Physicalism that any evidence function must supervene on some objectively real process X. Since X is an objectively real process it cannot be guaranteed a priori to be a correct instantiation of any function. We shall say that X fails on any occasion on which it does not correctly map some element from the domain of the function to its range. If X fails than its output will be an incorrect evidence state.How can M be certain that the evidence state K is correct, ie that the mapping from E to K has been correctly performed? Only by ruling out the possibility that X has failed. In other worlds, there must be no possible world epistemically available to M in which X has failed. The question of whether X has failed can be represented as the question "Is it certainly the case that X has not failed?" The process of answering thisquestion is itself equivalent to a new evidence function.This argument is general -- it applies to any evidence function. We can express this asthe second part of our proof, which we shall call Pollux:Castor has generality, ie it applies to any evidence function. Therefore Castor can be applied recursively to the new evidence function generated by Castor. This will generate a third evidence function, to which Castor can be applied recursively yet again. Indeed each time Castor is applied to any evidence function, it will generate a another evidence function to which Castor can be applied. Therefore the correctness of any objectively real process X cannot be ascertained without performing a non-terminating sequence of functions, each of which (by the Principle of Physicalism) must be performed by some objectively real process.Castor and Pollux together are equivalent to the Proof A given in the previous section.By splitting the proof into two sections in this way, the reader can clearly see how Proof A exhibits the two useful properties of generality and recursiveness. Castor establishes generality; Pollux shows that Castor can be applied recursively.The foregoing tells us that M can never guarantee its own correctness when evaluating the truth of any proposition.. However this result has not been proven by M, but by us, the readers. It has been proven in whatever logical system or systems underly our reasoning processes, which we must assume are sound in principle. As inthe previous section, let us call this system H. However Assumption 1 allows M to "inherit" H, as it were, and so to prove Proof A for itself. This reasoning also allows us to disregard the detailed character of H itself, since these details effectively cancel out. The reader's understanding of Proof A can effectively be regarded as a derivationof Proof A in the logical system H.We shall now express the outcome of Proof A as a theorem, which we can call Gemini:16 No suitably qualified physical system can exhibit the property of self-certainty, and in any system which exhibits self-certainty, the process which subserves self-certainty cannot depend on any objectively real process.Readers should note that Proof A is a proof of exactly this theorem, and not any other. A former reviewer of this paper attempted to paraphrase this theorem into a generic and rather ambiguous statement about the epistemology of consciousness, and then expressed objections based on the resulting ambiguities. Such a practice is strongly tobe discouraged. A theoretical proof is a proof of a specific, well-defined proposition, derived from a specific set of assumptions according to clear rules of inference. In this respect a theoretical proof is quite different from a philosophical argument.Another common objection from philosophers is that conscious entities do not need tocalculate or prove that they are conscious, but can in some way simply "ascertain" it. The mistake here is to assume that physical systems must necessarily have the same properties as conscious human beings. Any such "ascertaining" can still be represented as an evidence function, and if such an evidence function is performed by some physical process X (as is required by the Principle of Physicalism) then Proof A will apply to it. Readers should be careful not to attribute unconditionally to physical systems the properties of their own consciousness. Proof A applies generally to any system capable of reasoning in H, if both the Principle of Physicalism and the Axiom of Fallibility are assumed.Another objection which has been made frequently to me in private communications is that it would be a trivial matter to program a machine to answer "Yes" whenever thequestion "Are you certainly conscious?" was asked. Of course this is true, but such a machine would, by definition, not be reasoning according to the rules of H. Thereforesuch an objection has no relevance to the Gemini theorem.It has also been suggested that the correctness of X does not have to be established forM to be certain that the evidence state K is correct,12 since in the possible world whereX is correct, M may have access to other, different evidence from the possible world in which X fails. However this argument implicitly assumes another evidence function, and Proof A will apply recursively to any such function. One can express this evidence function in terms of the question "Does this other, different evidence demonstrate the truth of some proposition?" Since Proof A is generalizable to all evidence functions, it will apply to this one This is true regardless of what evidence M may have access to, since evaluating any such evidence is equivalent to the question "Does this evidence demonstrate the truth of some proposition?" and the function of answering such a question will always be subject to Proof A. No matter what evidence M may have access to, therefore, M can never be certain that the evidence function implied by that evidence has been correctly performed.Some philosophers insist that any notion of certainty entailing an infinite sequence of propositions is epistemologically unacceptable and unnecessary, and that certainty requires only that M is certain of some proposition p. This would be a reasonable remark if the notion of certainty employed here were axiomatic, but for physical 12 I am indebted to David Chalmers for this objection, and also for the one in the subsequent paragraph.17 systems reasoning in H this is not the case. The requirement for an infinite sequence of functions is in fact a therorem in H, the proof of which is implicit in Proof A. This proof can be expressed explicitly as follows: We start by assuming that M is certain of some proposition p. By the Principle of Physicalism, the process of determining p must supervene on some physical process, say X. From the Axiom of Fallibility, M can deduce that X may be fallible. By Assumption 1, M can deduce that if X is fallible, then the evidence state for p may be incorrect, in which case M cannot be certain of p. Therefore to be certain of p, M must have some way of determining that X is operating correctly. By the Principle of Physicalism, this must supervene on some physical process, say X*. But what process is X*? It cannot be X, since M relies on the process X* to ensure that X is operating correctly. X therefore depends on X*, not the other way round.Therefore M must assume that X* is some new process X', which is different from X. But from the Axiom of Fallibility, M can deduce that X' may be fallible. Since M's certainty of p depends on X, and the correctness of X depends on X', M must have some means of determining that X' is operating correctly. Let us call this process X**. But what process is X**? It cannot be X', since the correctness of X' depends on the correctness of X**. Neither can it be X, since establishing the correctness of Xrequires the correctness of X'. Therefore we require some new process X'' ... and so on ad infinitum. This leaves us with no choice but to discard the assumption which led to the regress -- that is, the Principle of Physicalism.A rather more serious objection to the Gemini proof is that the rules of H by which theproof is derived are never explicitly defined. The proof is derived using the rules of classical logic and we must therefore assume, firstly, that our ability as human beings to use these rules is sound, and secondly, that our judgement that these rules are the correct ones to apply in this situation is also sound. It is this second judgement which is impossible to formalize. Somehow, H must allow us (and any system which reasons like us) to decide that it is classical logic, rather than, say, some paraconsistent or quantum logic, which is the correct logic to use in this case. I do notbelieve that this is a serious problem, but it does underline how the proof depends on accepting Assumption 2. A brief mention must be made of two alternative philosophical approaches which some people appear to believe constitute loopholes in Proof A. The first is coherentism -- the doctrine that although individual processes may not be reliable, a large ensemble of processes together might be. This is obviously vulnerable to applying Proof A simultaneously to every process in the ensemble. Alternatively, the ensemble can simply be treated as a single process. More fundamentally, however, coherentism cannot be proven to be true; it must therefore be assumed to be true, which entails adding another axiom to H. Since the rules of H must allow M to ask the question "Is there any reason why this axiom should not be dropped?" and since the answer to this question is by definition NO, coherentism cannot establish absolute certainty.The second philosophical approach which requires some mention is constitutivism (Shoemaker, 1990). The key principle here is that the evidence state for the question "Am I certainly conscious?" is a physical state which overlaps wholly or partly with 18 the physical basis of consciousness itself. This idea was dealt with briefly in Reason (2016); the problem here is that, while self-certainty might be allowable if constitutivism is true, the process of determining if constitutivism is true can itself be represented as an evidence function, to which Proof A will of course apply. More specifically, for constitutivism to work, M must (by the rules of H) be able to answer YES to the question "Is the evidence state for consciousness physically identical with consciousness itself?" This question clearly implies an evidence function, to which Proof A applies. Proof A can be expressed in as a couplet of statements, as follows. For any suitably qualified physical system:Any evidence function must, by the Principle of Physicalism, be performed by some physical process.The correctness of any physical process must, by the Axiom of Fallibility, be determined by some evidence function.All philosophical objections to Proof A can be dealt with by applying this Gemini couplet. What kind of process is not a physical process?The reader may well now be wondering, if Proof A is as generalizable as it appears to be, why it does not simply apply to all processes, whether physical or not. What sort of characteristic could possibly differentiate physical processes, which are subject to Proof A, from non-physical processes, which are not subject to Proof A?The answer is that physical processes are defined as being objectively real -- that is, they are assumed to have properties which exist independently of the subjective stateswhich record the values of those properties. Therefore if we assume a j-type functionperformed by some objectively real process X, there will be an objectively real fact about whether X is a process of type P, or a process of type R. The infinite regress, in the form of the non-terminating sequence of processes generated by Proof A, arises from the need to identify exactly what this objective property is. No such difficulty arises if no objectively real process exists, since there is then no objective property to be identified. The consequence of this however is that whatever process performs the j-type function equivalent to self-certainty, this process cannot depend on any objectively real process -- it cannot, in other words, depend on any process which is external to human consciousness. Another way of looking at this is to say that any process which performs self-certainty must be subjectively real but not objectively real.DiscussionThe Gemini theorem itself shows only that self-certainty is impossible in any physicalsystem. In order to apply this to the mind-body problem we need to make a further 19 assumption, which is that self-certainty is in fact possible for human beings. This is by its nature an empirical question. Some will regard it as intuitively obvious that human beings are capable of self-certainty, given that self-certainty does not require us to make any assumptions about the nature of consciousness or the properties of conscious states13. For those who do not, however, there is a clear empirical prediction which can be made.Gemini implies that self-certainty will entail a violation of the conservation of energy (as was explained in Reason, 2016). We can express this violation as an inequality between the energy produced in a physical system and the energy consumed by that system. The difference between energy produced and energy consumed is denoted bythe symbol c (Reason 2016). It is this inequality which we must somehow detect experimentally.One should bear in mind here that there is no a priori reason why c should be a single event; it is quite possible the c effect in the brain is made up of many separate, localized effects which all contribute toward a change in the brain's physical evolution. If one denotes these local effects by cn, then the total effect c will be givenas follows: c = c1 + c2 + …….. + cnSince some of these local effects will be negative, it is theoretically possible for the c effect over the whole brain to sum to zero. This would imply an ad hoc degree of symmetry in the brain's functional architecture which seems to me unlikely. Nonetheless we are more likely to be able to detect non-zero c effects in highly localized regions of the brain than in larger, more general regions. In order to do this we need to rely on proxy measures of energy production such as cerebral blood flow, and proxy measures of energy consumption such as changes in the brain's electromagnetic field. If Ep represents some proxy measure of energy production, and Ec some proxy for energy consumption, then the absolute c value in some small three-dimensional compartment of the brain will be:13A special problem exists in the case of self-referential statements such as the question "Is it certain that I exist?" In such a case a system reasoning in H could get as far as Milestone 2 and then argue, along the lines of Descartes' Cogito ergo sum, that it does not really matter if the physical process which subserves the answering of that question is accurate or not -- the mere fact that such a physical process exists is enough to answer the question in the affirmative. One can express this in the form of aCartesian Principle: If P exists then M exists (since P is a process in M). Descartes' Cogito can thus be represented in H as a syllogism of the form:If P exists then M exists;P exists;Therefore M exists.However this syllogism requires M to establish that it is the case that P exists. Since this clearly entailsa j-type function, Proof A will apply to it, as before.20 cxyz = ((f(Ep) + energyin) - (g(Ec) + energyout))where energyin and energyout are terms representing the energy flowing into and out of the compartment from other compartments. The Gemini theorem predicts that this quantity, summed over three dimensions, will be non-zero. In practice, measuring theenergy transfer between compartments is not practical. The result of this is that only strongly asymmetric total c effects (those that do not sum to zero) are likely to be detectable. Such asymmetric effects will be indicated by a residual effect cres given by:cres = f(Ep) - g(Ec)when this is summed over three dimensions. The detection of symmetric c effects, however, relies on assuming that the energyin and energyout terms can both be set to zero. Whether this is possible is unclear at the present time..What sort of proxy measures of energy production and energy consumption are possible? The first problem to be confronted is that in vivo neuroimaging techniques assume the conservation of energy, and therefore do not explicitly distinguish betweenenergy production and consumption, regarding both as proxy measures for neural activity. Obviously we cannot make this assumption. We therefore need to eaxmine possible neuroimaging techniques a little more closely.The most widespread imaging technology used in neuroscientific research is Functional Magnetic Resonance Imaging (fMRI). This relies on the differing magnetic properties of oxygenated versus deoxygenated blood (Huettel, Song and McCarthy, 2014). Increased neural activity is thought to trigger a release in blood oxygenation levels, which is detectable in the form of the Blood Oxygenation level Dependent (BOLD) signal. Since there is a timelag between the increase in neural activity and the resulting increase in blood oxygenation, the temporal resolution of this technique is limited to about five seconds. Spatial resolution for this technique is less than five millimeters. This may be sufficient for detecting symmetric c but it is not clear if intercompartmental energy transfer can be ignored using this method.A more direct measure of energy production is Positron Emission Tomography. This can be used to measure glucose uptake by use of the radioactive tracer fluorodeoxyglucose (FDG). (See Vanitha 2011 for a brief description of the technique.) However the requirement for a radioactive tracer, together with the complexity and expense of the equipment, make this method problematic for pure research purposes. A possibly more promising approach is Phosphorus Magnetic Resonance Spectroscopy (P MRS) which measures the energy generated from the hydrolysis of ATP (Zhu, Qiao, Du, Xiong, Liu, Zhang, Ugurbil and Chen, 2012). Measuring the other side of the problem, energy consumption, turns out to be a little more constricted. The most obvious technology is Magnetoencephalography (MEG), a description of which can be found in Hamalainen, Hari, Ilmoniem, Knuutila and Lounasmaa (1993). MEG measures neural activity in the form of magnetic fields generated by dendritic currents. Since it requires at least 50,000 neurons to generate afield strong enough to be detectable, this technique has limited spatial resolution (> 21 3mm) which depends on making assumptions about the nature of the neural activity which has generated the signal; however it has excellent temporal resolution. A less well-known technique is Event-Related Optical Signal detection (Gratton and Fabiani 2001). This method detects neuronal activity by via selective absorption of infra-red light. It has excellent temporal resolution but its spatial resolution is limited to about a centimeter. It also has the disadvantage that the technique only works to a depth of up to 5cm into the head, and it is not clear if c can be assumed to occur within the cortex. However a potential advantage of EROS is that it can be carried out simultaneously with a measurement of hemodynamic response in the form of NIRS, or Near Infra-Red Signal detection (Gratton, Goodman-Wood and Fabiani 2001). This property could be especially useful in detecting c.It is the my hope that some or all of these techniques may offer the basis for a practical experiment to detect the c effect. Anyone who believes they can assist in carrying out such an experiment is urged to contact the author.ReferencesBlackmore, S. (2016) Delusions of Consciousness, Journal of Consciousness Studies, 23 (11-12), 52-64.Bridgeman, P. W. (1927) The Logic of Modern Physics. New York: McMillan.Caplain, G. (1995) Is consciousness a computational property? Informatica, 19, 615-619.Chalmers, D. J. (1996) The Conscious Mind: In Search of a Fundamental Theory. New York: Oxford University Press.Chalmers, D. J. (2011) Verbal Disputes, Philosophical Review, 120, 515-566.Chalmers, D. J. (2012) Constructing the World, Oxford: Oxford University Press.Churchland, P. S, (1996) The Hornswoggle Problem, Journal of Consciousness Studies, 3 (5-6), 401-408.Dennett, D. C. (1996) Facing backwards on the problem of consciousness, Journal of Consiousness Studies, 3 (1), 4-6.Frankish, K. (2016) Illusionism as a Theory of Consciousness, Journal of Consciousness Studies, 23 (11-12), 11-39.Gratton, G., and Fabiani, M. (2001) The event-related optical signal: A new tool for studying brain function. International Journal of Psychophysiology, 42, 109-121.Gratton, G., Goodman-Wood, M. R., and Fabiani, M. (2001) Comparison of neuronal22 and hemodynamic measures of the brain response to visual stimulation: An optical imaging study. Human Brain Mapping, 13, 13-25.Hacker, P. (2010) Hacker's Challenge, The Philosopher's Magazine, 51 (51), 23-32.Hamalainen. M., Hari, R., Ilmoniem, R. J., Knuutila, J., and Lounasmaa, O. V. (1993) Magnetoencephalography -- theory, instrumentation, and applications to noninvasive studies of the working human brain. Review of Modern Physics, 65 (2), 413-497.Huettel, S.A., Song, A. W., and McCarthy, G. (2014) Functional magnetic resonance imaging: Basic science (3rd edition) Sunderland, Mass: Sinauer.Kim, J. (1984) Concepts of Supervenience, Philosophy and Phenomenological Research, 45, pp. 153-176.Reason, C. M. (2016) Consciousness is Not a Physically Provable Property, Journal of Mind and Behavior, 37 (1), 31-46.Shoemaker, S. (1990) First-Person Access, Philosophical Perspectives, 4, 187-214.Vanitha, A. (2011) Positron emission tomography in neuroscience research. Annals ofNeurosciences, 18 (2) 36.Zhu, X. H., Qiao. H., Du, F., Xiong. Q., Liu, X., Zhang, X., Ugurbil, K., and Chen, W.(2012) Quantitative imaging of energy expenditure in human brain. Neuroimage, 60 (4), 2107-2117.23
1203.3596
3
1203
2012-06-08T01:26:57
Formation of antiwaves in gap-junction-coupled chains of neurons
[ "q-bio.NC", "nlin.PS", "physics.bio-ph" ]
Using network models consisting of gap junction coupled Wang-Buszaki neurons, we demonstrate that it is possible to obtain not only synchronous activity between neurons but also a variety of constant phase shifts between 0 and \pi. We call these phase shifts intermediate stable phaselocked states. These phase shifts can produce a large variety of wave-like activity patterns in one-dimensional chains and two-dimensional arrays of neurons, which can be studied by reducing the system of equations to a phase model. The 2\pi periodic coupling functions of these models are characterized by prominent higher order terms in their Fourier expansion, which can be varied by changing model parameters. We study how the relative contribution of the odd and even terms affect what solutions are possible, the basin of attraction of those solutions and their stability. These models may be applicable to the spinal central pattern generators of the dogfish and also to the developing neocortex of the neonatal rat.
q-bio.NC
q-bio
Formation of Anti-Waves in Gap Junction Coupled Chains of Neurons Alexander Urban Department of Physics, University of Pittsburgh, 100 Allen Hall 3941 O'Hara Street, Pittsburgh, PA 15260 Bard Ermentrout Department of Mathematics, University of Pittsburgh, 139 University Place, Pittsburgh, PA 15260 (Dated: November 6, 2018) Abstract Using network models consisting of gap junction coupled Wang-Buszaki neurons, we demonstrate that it is possible to obtain not only synchronous activity between neurons but also a variety of constant phase shifts between 0 and π. We call these phase shifts intermediate stable phase- locked states. These phase shifts can produce a large variety of wave-like activity patterns in one-dimensional chains and two-dimensional arrays of neurons, which can be studied by reducing the system of equations to a phase model. The 2π periodic coupling functions of these models are characterized by prominent higher order terms in their Fourier expansion, which can be varied by changing model parameters. We study how the relative contribution of the odd and even terms affect what solutions are possible, the basin of attraction of those solutions and their stability. These models may be applicable to the spinal central pattern generators of the dogfish and also to the developing neocortex of the neonatal rat. 2 1 0 2 n u J 8 ] . C N o i b - q [ 3 v 6 9 5 3 . 3 0 2 1 : v i X r a 1 I. INTRODUCTION In the past few decades, neuroscientists have discovered the importance of oscillations in the brain. Oscillations occur in many neural networks and play key roles in sensory as well as motor systems [1][2]. On the smallest scale of the nervous system, individual neurons can behave as complex nonlinear oscillators.[3] On a larger scale, oscillatory patterns in EEG (electroencephalographic) recordings have been shown to correspond to different cognitive states[3]. Phase oscillator models (obtained from systems of weakly coupled oscillators) have proven extremely useful in the analysis of such systems [4][5]. A pair of symmetric weakly coupled oscillators always has at least two possible periodic (phase-locked) states: the synchronous state and the anti-phase atate (where they are a half a cycle out of phase). Suppose that for some range of parameters the synchronous solution is stable and as one of these parameters is varied, the synchronous solution loses stability. In a symmetrically coupled system the synchronous solution generically loses stability in a pitchfork bifurcation. When this pitchfork bifurcation is supercritical it gives rise to two new stable phase-locked solutions in which the phase difference between oscillators is some constant between 0 and π. We refer to such solutions as Intermediate stable phase locked states. Intermediate stable phase locked states between pairs of neurons can be produced in a variety of neuron models by using a variety of coupling schemes as well as adjusting the parameters which affect the excitability of individual neurons [6][7][8]. The first work examining this behavior in conductance-based models was by Vanvreeswijk et al [8]. In this paper, the authors studied integrate-and-fire as well as Hodgkin-Huxley neurons. Cymbalyuk has both modeled and experimentally demonstrated an intermediate stable phase-locked state in a system of two "silicon neurons"[9][10]. These intermediate phase-locked states can lead to wave-like behavior in large networks. Wave activity is ubiquitous in the brain. Traveling waves of electrical activity in the brains of animals have been observed in a variety of species and occur in a diverse set of structures. These structures include the retina, olfactory cortex [11], peri-geniculate nucleus[12], neocortex [13] and spinal cord [1], among others [14],[12],[13],[11],[15][12]. In phase models, waves correspond to constant nonzero phase-differences between successive pairs of oscillators. There are many ways to generate such phase differences (see [16], section 8.3.5) such as a gradient in natural frequencies [4], pacemakers, or manipulation of the 2 boundary conditions [17]. In some swimming organisms, it is very important that a fixed phase-lag be maintained over a wide range of frequencies. For example, in the lamprey, the lag is about 2π/100 [? ], while for the crayfish it is 2π/4 [19]. As we will show in this paper, intermediate stable phase-locked states provide a simple way to produce waves with a stable fixed inter-oscillator phase-difference. The paper is organized as follows. We first show that by varying the excitability of gap- junction coupled Wang-Buszaki neurons, we can produce intermediate stable phase-locked states. We then explore the consequences of these states in one-dimensional nearest-neighbor coupled chains. We prove the existence and stability of a wide variety of complex waves including simple waves and zig-zag waves. We then study two-dimensional arrays and find two-dimensional analogues of zig-zag and regular waves. II. GAP JUNCTION COUPLING BETWEEN PAIRS OF WANG-BUSZAKI NEU- RONS A. Measuring the Phase Difference Between Spiking Neurons The Wang-Buszaki model is a conductance based neuron model derived from the Hodgkin Huxley model. It was originally used to describe fast spiking interneurons in the hippocam- pus [20]. In this section we examine pairs of Wang-Buszaki neurons reciprocally coupled with gap junctions and ask what happens as we vary parameters which affect the excitability of the individual cells. Gap junctions are specialized ion channels connecting the cytoplasm of the pre and post-synaptic cell. A depolarizing ionic current is driven by the potential difference between the cells. In order to understand the role that gap junctions play in rhythmically oscillating networks, consider just two neurons coupled with gap junctions. The coupled neuron equations will have the following form: C V1 = −IN a − Ik − IL − I0 + g(V2 − V1) C V2 = −IN a − Ik − IL − I0 + g(V1 − V2) 3 (1) These currents are given by the equations: ∞h(V − VN a) IN a = gN am3 Ik = gkn4(V − Vk) IL = gL(V − VL) (2) C is the the membrane capacitance measured in units of µF/cm2. The parameters: VN a,Vk and VL are the reversal potentials for the ion channels. The parameters g, gN a, gk and gL are constants describing the conductances of their respectively labeled ion channels. Typically, they are measured in units of mS/cm2. The m and the h are the time-dependent activation and decativation variables which are described by the equations: dh dt dm dt = η(αh(V )(1 − h) − βh(V )h) = η(αm(V )(1 − m) − βm(V )m) (3) In these equations, η is an overall channel switching rate determined by the temperature of the system[21]. The nonlinearities αi(V ), βi(V ), as well as the parameters used in the simulations are referenced in the Appendix A. Previously, Pfeuty et al. have studied pairs of Wang-Buszaki neurons with gap junction coupling and shown that by varying the potassium and sodium conductances, one can vary the stability of the synchronous in-phase and anti-phase states [22]. In order to find pa- rameter regimes in which intermediate phase-locked states exist, we varied the potassium conductance, gk and also the temperature dependent rate constant, η. Parameters such as these play key roles in the behavior of both central pattern generators and large scale oscilla- tory networks [23][24]. The most important consequence of varying these parameters is that the absolute refractory period decreases. The neuron is less hyperpolarized after the action potential for larger η or smaller gK [22]. In other words, this smaller absolute refractory region allows for the neuron to fire more quickly [22]. Consider the upper half of Figure 1: Panels A and B are calculations of the phase difference between two Wang-Buszaki neurons computed after 1500 ms of integration time for different values of gk (the maximal potassium conductance) and η (the temperature dependent time constant). The phase difference, φ was measured as the difference between the times at which V1 and V2 cross zero with a posi- tive slope divided by the period of the oscillation. Holding the stimulus current constant at 4 Phase difference vs. η of two Wang−Buszaki Neurons 1 Phase difference vs. gk of two Wang−Buszaki Neurons 1 B. e c n e r e f f i d e s a h P 0.8 0.6 0.4 0.2 0 4 4.5 5 5.5 6 6.5 gk η=5 7 7.5 8 8.5 9 H functions obtained by varying gk A. e c n e r e f f i d e s a h P 0.8 0.6 0.4 0.2 η=5 η=5.5 η=6 η=6.5 η=7 ) φ ( H 0 3 2 1 0 −1 −2 0 3.5 4 4.5 5 5.5 η 6 6.5 7 7.5 8 H functions obtained by varying η gk=9 C. 0.1 0.2 0.3 0.4 0.5 φ 0.6 0.7 0.8 0.9 1 D. ) φ ( H 2 1 0 −1 −2 0 gk=5 gk−6 gk=6.5 gk=7 gk=9 0.1 0.2 0.3 0.4 0.5 φ 0.6 0.7 0.8 0.9 1 FIG. 1. A. The phase difference between two Wang Buszaki neurons (Full model) as a function of the parameter η. The diagram clearly illustrates a pitchfork bifurcation connecting the synchronous and anti-phase solutions. B. Plot of the phase difference for varying values of gk. C. Plot of the odd part of the interaction function (cf equations (8,9)) for different values of η (the dimensionless temperature dependent time constant). D. Calculation of the odd part of the interaction function for different values of gk. The zeros of these functions correspond to the stable phase locked solutions. i0 = 0.63nA and varying either the parameter η or gk, we are able to demonstrate a super- critical pitchfork bifurcation in the phase difference between neurons. Figure 1A illustrates the bifurcation of the system of two neurons from synchrony to anti-phase behavior as η is increased. For values of η between 5.0 and 7.0 the system passes through all possible stable relative phase-locked solutions varying between 0.0 and π. Figure 1B is a plot of the phase difference between neurons as we decrease gk. As we decrease gk between 8.0 and 5.0 we also obtain intermediate stable phase-locked solutions. 5 B. Calculation Of The Interaction Function In order to simplify the analysis, we reduce our system of neurons to a phase model. This can be accomplished by applying Malkins Theorem [2]. Consider two weakly coupled systems (By weak, we mean that the coupling acts to only affect the phase and not the amplitude of the oscillation [2]) of coupled differential equations describing tonically firing neurons: 1 = F (V1) + ǫC1(V1, V2) + O(ǫ2) V ′ 2 = F (V2) + ǫC2(V2, V1) + O(ǫ2). V ′ Then Malkin's theorem states: V1(t) = V0(θ1) + O(ǫ), V2(t) = V0(θ2) + O(ǫ) where V0 is the solution to the homogenous equation. Furthermore: dθ1 dt dθ2 dt = 1 + ǫH1(θ2 − θ1) + O(ǫ2) = 1 + ǫH2(θ1 − θ2) + O(ǫ2) where: Hj(φ) := 1 T Z P 0 V ∗(t) · Cj[V0(t), V0(t + φ)] dt. Here V ∗ is the adjoint of the linearized equation (4). Let φ := θ2 − θ1. Then: dφ dt = −2ǫHodd(φ) := ǫ[H(−φ) − H(φ)]. (4) (5) (6) (7) (8) (9) The right hand sides of these phase model equations are described by interaction functions which are derived from the full model. Zeros of Hodd(φ) determine the possible phase-locked states and those for which H ′ odd(φ) > 0 are stable. We calculated the interaction functions for several values of gk and η. Figure 1C and 1D shows the odd portions of the interaction functions calculated with various parameter values. We write these interaction functions in terms of their Fourier series: H(x) = a0 2 + ∞ Xn=1 (cid:18)an cos(nx) + bn sin(nx)(cid:19) (10) For example, Table I shows the first few Fourier modes of the interaction function computed from the full model with η = 6. This table demonstrates the substantial contribution of higher order Fourier terms to the interaction function near the bifurcation point. 6 a0 = 5.1974931 b1 = 0.47408548 a1 = −2.9970722 a2 = −0.92187762 b2 = −0.36833799 a3 = −0.44113794 b3 = −0.2577318 a4 = −0.25482759 b4 = −0.15762125 a5 = −0.16416954 b5 = −0.09083201 a6 = −0.11295291 b6 = −0.048487604 TABLE I. Fourier Coefficients of the Interaction function, equation (8) (computed from the full model with η = 6). Near the bifurcation point from synchrony there are substantial higher order even and odd Fourier terms. Observe that the first two odd Fourier terms in Table I are the dominant ones (b1, b2). As the shape of Hodd(φ) depends only on the odd terms, aj have no effect on the existence and stability of the locked solutions for a pair of symmetrically coupled oscillators. However, once there are more than just two oscillators, the even terms play an important role, particularly as fasr as stability is concerned. We find that as η changes between 5 and 7, the Fourier coefficient, b2 remains negative and b1 goes to zero. For this reason, in the rest of this paper, we will use a truncated version of the H function that contains only three terms: H(x) = b1 sin(x) + b2 sin(2x) + a1 cos(x) (11) Unless otherwise stated, we typically take b1 = 1 and b2 = −.75, giving zeros of Hodd at φ = 0, π,± cos−1(2/3) ≈ ±0.847. We remark that this particular form of Hodd(φ) arises in the bead on a hoop instability [25](Chapt 3.5) and in a model for the coordination of finger tapping ([26]). If b2 < 0 is fixed and negative, then as b1 decreases from a large positive value, the synchronous solution loses stability (at b1 = −2b2) and a branch of intermediate stable phase-locked solutions bifurcates. This branch remains stable until b1 becomes sufficiently negative b1 < 2b2 whereupon the anti-phase solution is stable. III. WAVES IN LARGE NETWORKS This section is a study of wave behavior in both chains and two dimensional arrays of neurons with nearest neighbor coupling in regimes where there is an intermediate stable 7 phase-locked state for pair-wise coupling. Primarily, we study phase models which use the interaction functions (or approximations of them) derived from the Wang-Buszaki model. Our models may be relevant to patterns of wave activity in the neonatal rat. Peinado et al. was able to observe wave activity in gap-junction coupled interneurons in rat neocortex prior to day twelve of development[13]. Furthermore, he was able to enhance these waves by applying halothane and picrotoxin. Picrotoxin blocks inhibitory synapses while halothane reduces the potassium conductance. In general, he observed that the reduction in potassium directly led to the formation of waves. Since our intermediate stable phase-locked states can occur in gap junction coupled neurons by reducing the potassium conductance, this may be experimental evidence that this effect plays a role in wave formation in a two-dimensional network. We focus on a specific type of solution to the phase equations known as anti-waves. Anti- waves were first studied in two papers by Ermentrout and Kopell[17][27]. Similar phenomena have been examined by Strogatz et al., (uniformly twisted waves) and Blasius et al., (1D quasiregular concentric waves)[28][29] [30]. Anti-waves consist of waves either initiated at the ends and colliding in the middle or waves initiating from the middle and terminating at the ends. The latter are, in a sense, equivalent to one-dimensional target patterns. The wavenumber for these waves (the spatial gradient of the phase) shows an abrupt change of sign which we will call a kink. Similar phase waves have been demonstrated in mechanical systems[31]. Anti-waves have been experimentally observed in the spinal cords of dogfish [32] and may well be present in other biological tissue. For instance, similar patterns of electrical activity have been observed in the muscle of the colon of a cat [33]. Central pattern generators in the fins of electric fish have also been known to produce anti-waves[27]. These animals are able to produce a variety of complex waves in which the "kink" or lead oscillator in the wave is able to shift[27]. We want to stress that our mechanism for generating anti-waves is fundamentally different than previous papers. Furthermore, our anti-waves can form anywhere along the chain and are much more robust to perturbations than in previous models[17]. The rest of this paper is organized into several sections. We begin in section III A by discussing the basic phase models and boundary conditions. This is followed by an analysis of both an ordinary traveling wave (III B) and anti-wave solutions (III C) . In section III D, we demonstrate that the probability of obtaining a particular solution depends on the relative contribution of 8 the even component of the interaction function. We show that starting from the anti-wave solution, if the even component is sufficiently large, perturbations initiated at one end of the chain can propagate down the chain and shift the position of the kink. Finally, in V we demonstrate that this analysis can be extended to higher dimensions and that a variety of anti-wave patterns are possible in a two dimensional oscillator arrays. A. Models And Boundary Conditions The models we consider were introduced by Kopell and Ermentrout in a 1986 paper[17]. These models primarily describe networks of neurons with nearest neighbor coupling. In analyzing these equations we apply two types of boundary conditions: periodic boundary conditions and non-reflecting boundary conditions. For a system of N + 1 neurons with periodic boundary conditions, the system of phase equations may be written: θ1 = ω1 + HL(θN +1 − θ1) + HR(θ2 − θ1) θ2 = ω2 + HL(θ1 − θ2) + HR(θ3 − θ2) ... θN +1 = ωN +1 + HL(θN − θN +1) + HR(θ1 − θN +1). (12) In these equations, the ωi represents the natural frequencies of the oscillators. We denote the coupling in the two possible directions as HL(φ) and HR(φ). In general, we assume that the coupling is isotropic, so that: HR(φ) = HL(φ) = H(φ). Ultimately, we are interested in phase-locking behavior, thus we make the change of variables: φj = φj+1 − φj. This results in a system of N phase equations: 9 N Xj=1 φ1 = ∆ω1 + H(φ2) + H(−φ1) − H( φj) − H(φ1) φ2 = ∆ω2 + H(φ3) + H(−φ2) − H(−φ1) − H(φ2) φj = ∆ωj + H(φj+1) + H(−φj) − H(−φj−1) − H(φj) ... ... φN = ∆ω1 + H(−φN ) + H(− N Xj=1 φj) − H(−φN −1) − H(φN ). (13) In these equations ∆ωi is the frequency gradient between oscillators. In most of our simu- lations, we assume identical frequencies so ∆ωi = 0 The second type of boundary condition that we use is a variation of what are known as non- reflecting boundary conditions [34]. Non-reflecting boundary conditions are implemented in order to attempt to eliminate reflections and to "trick" the neurons at the ends of the chains, neuron 1 and neuron N + 1, into behaving as though the chain is infinite. If one were to think of the chain as being a continuous system, then the boundary conditions are simply a statement that dφ(x,t) dx = 0 when evaluated at the ends of the chain. There is a precedent for using such boundary conditions in nonlinear oscillator problems, for an example, see [31] . Applying these boundary conditions to our phase equations, we have[35]: θ0 = θ2 θN +1 = θN −1, and our phase equations become: φ1 = ∆ω1 + H(φ2) + H(−φ1) − 2H(φ1) φ2 = ∆ω2 + H(φ3) + H(−φ2) − H(−φ1) − H(φ2) φj = ∆ωj + H(φj+1) + H(−φj) − H(−φj−1) − H(φj) ... ... φN = ∆ωN + 2H(−φN ) − H(−φN −1) − H(φN ). 10 (14) (15) full model: g=.01 η=6.0 full model: g=.01 η=6.0 phase model: η=6.0 phase model: η=6.0 5438 5456 5476 5496 ) s m ( e m i t 5872 5890 5908 ) s m ( e m i t ) s m ( e m i t 5817 5835 5853 5817 5835 5853 5817 5835 8 2 2 4 6 index η=6.2 g=.019 4 6 index η=6.2 g=.019 5,855 5872 5890 5908 ) s m ( e m i t 40 20 0 −20 −40 −60 ) s m ( e m i t 5 10 15 20 5 10 15 8 10 20 2 8 10 4 6 index η=6.2 ) s m ( e m i t 5 10 15 20 5 10 15 20 6 5 4 3 2 1 2 4 6 index 8 10 η=6.2 ) s m ( e m i t ) s m ( e m i t 5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20 2 4 6 index 8 10 2 4 6 index 8 10 2 4 6 index 8 10 2 4 6 index 8 10 FIG. 2. Examples of traveling waves in both the Wang-Buszaki model (left) and the phase model (right) corresponding to two different values of the temperature dependent time constant η. The phase model reproduces the dynamics of the full model. Furthermore, it is clear from the phase model, that as one increases the constant η, the wavelength of the traveling waves decreases. Coupling strength in the full model is small: It must be on the order of g = 0.01 µS/cm2 to ensure phase-locking. B. Traveling Waves In Chains Of Coupled Oscillators The wave solution for the traveling wave can be written: φj = jφ∗ + ∆ωt, (16) where j is the index of the oscillator and φ∗ is the phase shift between adjacent oscillators. Substituting this solution into the equations with non-reflecting boundary conditions, we see that 16 is a solution provided that H(φ∗) = H(−φ∗). Since our interaction function has both odd and even components, this statement is true only if φ∗ is the root of H(φ)odd. Thus, the stable phase-locked states for pair-wise symmetrically coupled oscillators determine the wavenumber, φ∗. In a continuum limit, we can consider θ(x, t) to be a function of both 11 position and time. If we take the derivative of θ(x, t) with respect to t we obtain the expression[17]: which is equivalent to: dθ dt + ∂θ ∂x ∂x ∂t = 0, ω + φ∗vθ = 0 ω φ∗ . vθ = (17) (18) This is an expression for the phase velocity of the wave. Therefore, we see that we may identify the wavenumber of the system as: k = φ∗. The stable phase-locked state between pairs of oscillators defines the wavenumber of a traveling wave. In the last section it was demonstrated that as we vary constants η and gk in the Wang-Buszaki model, the stable fixed point changes. This translates to a change in wavelength in a chain of neurons. Figure 2 shows a comparison between the phase model and full model for a variety of gk and η. The four panels on the left correspond to the full model. The panels on the right correspond to the phase model. The phase model quantitatively reproduces the dynamics of the full model. The phase model clearly demonstrates that the wavenumber increases for increasing η. Coupling strength in the full model is small: it must be on the order of 0.01 to get agreement with the phase models. Stronger coupling results in only synchronous dynamics. C. Anti-Waves In Chains of Coupled Oscillators Anti-waves have been studied by Ermentrout and Kopell in two separate publications [17][27]. The previous mechanisms for generating anti-waves rely on extremely long chains (essentially infinite) or chains with distal connections. Assuming an isotropic chain with no gradient in the natural frequencies, the intermediate stable phase-locked state defined by H(φ∗)odd = 0 will generate traveling waves. If the fixed point φ∗ is identified as the wavenumber k, then the one kink anti-wave solution can be written: φj = kj + ωt φj = −kj + ωt 12 j < j∗ j ≥ j∗. (19) A. Full model: Periodic Boundary conditions e m i t s m C. e m i t s m 300 305 310 315 320 325 330 335 300 305 310 315 320 325 330 335 5 10 Neuron index 15 20 Full Model Non−Reflecting Boundary conditions 5 10 Neuron index 15 20 B. e m i t n o i t a r g e t n I D. e m i t n o i t a r g e t n I 0 0.5 1 1.5 2 0 0.5 1 1.5 2 10 0 −10 −20 −30 −40 −50 −60 10 0 −10 −20 −30 −40 −50 −60 Phase Model: Periodic Boundary conditions 10 15 20 5 Neuron index Phase model: Non−reflecting Boundary Conditions 5 10 Neuron index 15 20 6 5 4 3 2 1 6 5 4 3 2 1 FIG. 3. Four examples of anti-waves in both rings and chains of oscillators computed with η = 6.0. A. Is a wave in a chain of Wang-Buszaki neurons with periodic boundary conditions. B. Is a wave in a chain of phase oscillators with periodic boundary conditions. C. Is a wave in a ring of Wang- Buszaki neurons with non-reflecting boundary conditions (full model). D. Is the phase model reduction of panel C: it is a wave in a chain of phase oscillators with non-reflecting boundary conditions. In these equations j∗ represents the position of the kink. By substituting the above expres- sion into 15 we see again that the anti-wave is a solution provided that H(k) = H(−k) (k is the root of Hodd(k)) . Figure 3 demonstrates examples of anti-waves obtained in the Wang- Buszaki model compared with a phase model. Figures 3A and 3C are waves generated in the full model with non-periodic and non-reflecting boundary conditions, respectively. Figures 3B and 3D are the equivalent phase models. D. Obtaining Different Wave solutions from Random Initial Conditions In order to see the variety of anti-waves, we will start chains of oscillators with random initial phases to estimate the basins of attraction [28]. The interaction function we use (11) allows any number of shocks or kinks in the anti-wave, constrained only by the length of the chain. However, if we start the equations from random initial conditions, the probability of getting a certain number of shocks varies as we change the size of the Fourier components. The main point is that even if the odd portion of the interaction function allows for a multiple 13 A. −1 −0.5 1 a 0 0.5 0 C. 1 b −0.5 0 0.5 1 0 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0.35 0.3 0.25 0.2 0.15 0.1 0.05 B. −1 −0.5 1 b 0 0.5 0 D. 1 b −0.5 0 0.5 1 0 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0.6 0.5 0.4 0.3 0.2 0.1 0 5 5 15 10 N 15 10 N 5 5 15 10 N 15 10 N FIG. 4. Probability of obtaining different solutions of equation (15) for twenty oscillators as a function of the Fourier coefficients of the interaction function. N designates the number of shocks in an anti-wave. N = 0 corresponds to a traveling wave. A. The probability distribution with H(φ) = a1 cos(φ) + b1 sin(φ) + 0.75 sin(2φ) as a function of a1 with b1 = 1. B. The probability distribution with a1 = 1 and varying b1. C. The probability distribution calculated with an interaction function: H(φ) = −3 cos(φ)− .92 cos(2φ) + b1 sin(φ)− 0.75 sin(2φ). D. The probability distribution calculated using the interaction function: H(φ) = −3 cos(φ) + b1 sin(φ) − 0.75 sin(2φ). For each parameter value the equations were solved from 10000 random initial conditions. shock solution, the probability of the system converging to this solution from random initial conditions may be extremely low and is determined by the magnitude of both the even and first odd Fourier modes. Figure 4 shows the probability distributions of obtaining various anti-wave and traveling wave solutions as a function of the magnitudes of both the even and odd Fourier terms of the interaction function. Panel 4A is a plot of the probability of obtaining either an N-shock anti-wave solution or a traveling wave (0−shock) solution as a function of a1. From the plot, we see that the probability of obtaining a traveling wave solution approaches zero as a1 → 0 and the probability distribution shifts towards N = 6. 14 A. 0 50 100 e m i t n o i t a r g e t n I 150 200 250 300 350 400 0 50 100 Equation number 150 200 6 5 4 3 2 1 0 B. n o i t i s o p k c o h s l a n F i 108 107 106 105 104 103 102 101 100 99 0 0.5 Initial amplitude (A) 1 1.5 FIG. 5. A. Compactons initiated from either side of the chain can shift the phase shock in the anti-wave back and forth. The interaction function used was: H(φ) = cos(φ) + sin(φ)− 0.75 sin(2φ) B. The shift in the kink for varying initial compacton amplitudes. As a1 increases towards 1, the probability of obtaining the traveling wave solution increases until it is the most probable state. Panel 4B is a plot of the probability distribution as a function of b1 with a1 = 1. Panel 4B shows a trend similar to 4A. For b1 = 0 the probability of obtaining a traveling wave solution from random initial conditions is close to 0. The solution with the maximal probability corresponds to an anti-wave with N = 9 shocks or kinks. As b1 is increased towards 1 or decreased towards −1 the probability distribution shifts towards N = 0. That is, solutions with fewer kinks become more probable. Panels 4C and 4D demonstrate the effect of higher order even terms. Panel 4D demonstrates that using the interaction function with two even terms (H(φ) = cos(φ) + b1 sin(φ) − 3 4 sin(2φ)) results in a probability distribution in which the traveling wave solution is the most probable solution for b1 → 1 and the 1 kink anti-wave solution is the most probable for b1 → −1. If we do not include this second order term, as demonstrated in panel 4C the most probable solution corresponds to an (N = 4 shocks) anti-wave. The main point of this is that, even though only the odd Fourier modes determine the solutions to equations 13 and 15, the even Fourier modes can drastically affect the basin of attraction of those solutions. 15 E. Moving The Shock Position With Impulses Electric fish have been observed to produce complex anti-wave type patterns[27]. What is more, the lead oscillator or (kink) in these anti-waves has been observed to be able to shift position. We present a mechanism which shifts the position of the kink/shock in the anti-wave without changing its shape. Over the last fifteen years, Pikovsky and Rosenau have written several papers studying waves in oscillator lattices with purely even coupling. They showed that such interaction functions can be derived from networks of Josephson junctions and Van-der Pol oscillators[36][37][38]. Waves in these networks take the form of solitary pulses which retain their shape. Pikovsky and Rosenau have coined these waves "compactons"[36][37][38]. Collisions between different compactons have been studied in detail. Nothing, however, has written about compactons colliding with a phase boundary (like the phase shock in the anti-wave). Compacton-like pulses are possible not only in systems with a purely even interaction function, but they can be observed in systems with odd terms, provided the even component is large enough[38]. The odd portion of the interaction function acts to dissipate the initial pulse, but a compacton-like wave may still travel large distances even with a substantial odd term. Compactons are possible in systems in which the interaction function generates anti-waves. For instance, we consider a system of 200 phase-difference equations with non-reflecting boundary conditions. The interaction function used is H(φ) = cos(φ) + sin(φ)− .75 sin(2φ). The initial conditions are the one kink anti-wave solution with an extra pulse stimulus applied at the end. Thus we have the one kink/shock solution: φj(0) = k φj(0) = −k N 2 N 2 , j < j ≥ with the addition of a pulse: φj(0) = k + A 2 (1 + cos((j − x0)π/σ)) j − x0 < σ. (20) (21) Here A is the amplitude, x0 is the position of the pulse and σ is the width of the pulse. In Figure 5 the pulse width used is σ = 10 and k = .84106. This pulse is the form used by Pikovsky et al. to generate compactons, but other initial conditions may work as well[38]. 16 Figure 5C. shows multiple compactons traveling on top of an anti-wave and colliding with the shock located at N = 100. The anti-wave is composed of the stationary white and grey regions. The grey region of the plot corresponds to the solution φ = .841 whereas the white region corresponds to: φ = 5.44. Upon collision, the shock shifts but retains its shape. In this manner, multiple pulses initiated at the ends of the chain may be used to shift the shock back and forth. Figure 5B. is a plot of the shift in the kink as a function of the initial compacton amplitude. The maximum shift obtained is approximately 9 sites. If A > 1.5 radians, larger amplitude pulses do not necessarily provide larger position shifts. If the pulse is too large, it will destroy the "perfect kink" solution. Thus, near the supercritical pitchfork bifurcation, not only can the shock of an anti-wave form anywhere along the the chain but precisely because of this property, it can be shifted around by an impulse (compacton). In this way, the additional even and odd Fourier terms produce a central pattern generator which is malleable. Perhaps this is a mechanism by which the hindbrain of a fish could send impulses to the rest of its spinal cord to modify the fish's swimming pattern. IV. STABILITY ANALYSIS In analyzing the stability of anti-wave solutions, one of the main questions we want to address is how the relative contributions of each Fourier mode contributes to the stability of the solution. Additionally, we want to examine the importance of other parameters in the model, such as the length of the chain and the position of the kink. In these chain models there are four basic wave solutions which are of interest. Two of the solutions correspond to a traveling wave in either direction (left to right or right to left) and two correspond to anti-waves. The first anti-wave solution describes a wave emanating from the center of the chain and propagating in both directions outward. The second anti-wave solution corresponds to two waves emanating from the edges and colliding in the center of the chain. Using our simplified interaction function we begin by analyzing the simplest (shortest) chain possible, the three neuron system. The three phase equations (22) describing the neurons can be condensed to two by a change of variables. Linearizing these equations (23) about the anti-wave solution results in a 2x2 Jacobian. Thus, the problem is simple enough so that we can solve for the eigenvalues as a function of a1 and subsequently show where and how the anti-wave solution loses stability. Once we have proved stability for this simple case, we 17 discuss longer chains and specifically, we analyze the effects of the magnitude of the even component on the stability of various anti-wave solutions. A. Stability Analysis Of The Three Oscillator System Our equations for the system with non-reflecting boundary conditions are: θ1 = 2H(θ2 − θ1) θ2 = H(θ3 − θ2) + H(θ1 − θ2) θ3 = 2H(θ2 − θ3). We then write them in terms of their phase differences: φ1 = 2H(−φ2) − H(−φ1) − H(φ2) φ2 = H(−φ1) + H(φ2) − 2H(φ1). (22) (23) Note that these equations are invariant under a reflection: φ1 → −φ2 φ2 → −φ1 Figure 6 is a plot of the nullclines of the system 23 for various values of a1 using b1 = 1 and b2 = −.75. There are two anti-wave solutions indicated by the boxes and two traveling wave solutions which are circled. One can see that when there is no even component the solutions to the system possesses perfect reflection symmetry. As soon as a1 is nonzero, the system loses this symmetry. We want to analyze the stability of the anti-wave analytically. Choosing φ1 = k and φ2 = −k, we linearize our equations about this system and write down the Jacobian as follows: M0 =   H ′(−k) −H ′(−k) − 2H ′(k) −2H ′(k) − H ′(−k) H ′(−k)   , Solving for the eigenvalues of this expression, we have λ1,2 = −2H ′(k),−2H ′(−k) − 2H ′(k). (24) (25) As long as the derivative of these H functions evaluated at this solution is positive, then the solution will be stable. Substituting the simplified interaction function (26) and its 18 a1=0 a1=1.2 A. 6 B. 6 1 φ 5 4 3 2 1 0 0 1 2 3 φ 2 1 φ 5 4 3 2 1 0 4 5 6 0 1 2 3 φ 2 1.2 C. 1.1 1 φ 1 0.9 0.8 0.7 saddle point Branch point a1=1.118 φ 1=.8411 stable node 4 5 6 0.6 0 0.5 1.5 2 1 a1 FIG. 6. Phase portrait of the two-dimensional system, (23) where H(φ) = a1 cos(φ) + sin φ − 0.75 sin 2φ. The φ1 (φ2) nullcline is in red(green). A. The nullclines and the flow for a1 = 0. The fixed points corresponding to anti-waves are enclosed with boxes. The fixed points corresponding to traveling waves are circled. B. The nullclines and the flow for a1 = 1.2. C. Bifurcation diagram computed with AUTO: The anti-wave solution loses stability near a1 = 1.118. In this plot, the stable solution is represented by a solid line whereas the unstable solution is represented by the dotted lines. derivative into equation 25. results in the eigenvalue expressions: equation 27. H(φ) = b1 sin(φ) + b2 sin(2φ) + a1 cos(φ) H ′(φ) = b1 cos(φ) + 2b2 cos(2φ) − a1 sin(φ) λ1 = −2(b1 cos(k) + 2b2 cos(2k) − a1 sin(x)) λ2 = −4(b1 cos(k) + 2b2 cos(2k)) (26) (27) At the critical value of the parameter acritical = 1.1, the first eigenvalue vanishes. For larger values of a1, the fixed point becomes unstable. The mirror symmetry of the equations, along with the bifurcation diagram in Figure 6, suggest that this is a subcritcal pitchfork bifurcation. This is verified in [39] where we explicity calculate the normal form equations. 19 B. Stability Analysis For The Traveling Wave Proving the stability of the traveling wave (with no kinks) is relatively straightforward. We begin by writing down the equations for the "jth" oscillator. Since the system is nearest neighbor coupling, the general expression is: φj = H(φj+1) + H(−φj) − H(−φj−1) − H(φj). (28) Defining a = H ′(φ∗), b = H ′(−φ∗) and linearizing the equations about the wave solution, the equations for the phase are simply a discretized version of Laplace's equation: aφj+1 − (a + b)φj + bφj−1 = λφj. We may solve for the equations by assuming a general solution: φj = Axj 1 + Bxj 2, (29) (30) and invoking the boundary conditions: φ0 = φ1 and φN −1 = φN .Plugging our solution 30 into 29, we may solve for the eigenvalues of the system: ℜλ = 2√ba cos πm N − 1 − (a + b). Therefore the wave solution will always be stable provided that Or alternately that a, b > 0. Stability is lost for a, b < 0. 2√ba ≤ a + b (31) (32) C. Stability for The Anti-Wave under More General Conditions Stability analysis of the anti-wave solutions may be proven using a combination of the Gershgorin Circle Theorem and numerically computing the eigenvalues of the linearized equations. In these examples, we use non-reflecting boundary conditions. The argument will be identical for periodic boundary conditions. One starts by assuming a 1 shock solution and linearizing the phase the equations about this solution. The Jacobian will have matrix elements of the form: ai,i = −(H ′(φj) + H ′(φj)) ai,i+1 = H ′(φj+1) ai,i−1 = H ′(φj−1) (33) 20 All other matrix elements are zero. Once more, define: a = H ′(φ∗), b = H ′(−φ∗). denotes the location of the kink, then the matrix elements corresponding to the kink are: If l or, al,l = −(a + b) al,l+1 = b al,l−1 = b al,l = −(a + b) al,l+1 = a al,l−1 = a (34) (35) depending on the orientation of the shock. The Gershgorin Circle Theorem [40] tells us that all of the eigenvalues of the matrix lay in the union of disks centered at the diagonal elements of the matrix with radii less than the absolute value of the sum of the row entries not including the diagonal terms. If the Jacobian of the anti-wave is described by equation 34 all eigenvalues will lie in the union of three disks: one centered at −(a + b) with radius a + b, one centered at −2a + b (corresponding to the ends of the chain) with radius a and one centered at −(a + b) with radius 2b. If we assume that the even term a1 is positive and increasing, then H ′(−φ) ≥ H ′(φ) b ≥ a, 2b > a + b. (36) Thus the disks will extend beyond the origin, and we will not be able to say anything about stability using this theorem. On the other hand, if the solution is a shock oriented in the opposite direction, such as in equation 35, the disk corresponding to the equation at the shock is centered at −(a+b) and extends out with radius 2a. The disk will always lie in the left half of the imaginary plane for a1 positive. Therefore, this solution will always be stable. In the former case, in which the fractured wave is not necessarily stable, one must apply numerical methods to explicitly calculate the eigenvalues of the Jacobian. We are interested how variables such as the position of the shock and the length of the chain effect the stability of the solution as we vary the magnitude of the first even Fourier mode. Again, we use the interaction function: H(φ) = b1 sin(φ) + b2 sin(2φ) + a1 cos(φ), where b1 = 1 and b2 = −.75. We start by examining a fifty-one oscillator chain (described by fifty equations) in which we move the position of the shock. Figure 7A shows the critical value of a1 at which the shock solution loses stability as a function of shock position where the position varies between site 4 and site 46. The parameter acritical is determined by calculating the eigenvalues of the 21 A n o i t i s o p k n i k ) r e b m u n e t i s ( 50 40 30 20 10 0 1 Hopf bifurcation 2 3 4 5 acritical 6 7 8 9 0.4 0.3 0.2 B Real part of the eigenvalue with maximal real part as a function of a1 N=22 N=23 N=24 N=25 N=26 N=27 N=28 N=29 λ ℜ 0.1 0 −0.1 −0.2 0 acrit vs. b2 for N=50 2 4 6 a1 8 10 12 C −0.5 −1 2 b −1.5 −2 −2.5 0 2 4 6 acrit 8 10 12 FIG. 7. A. Critical value of a1 as a function of shock position. B. Eigenvalue with the maximal real part as a function of a1 for different shock positions. N denotes the location of the shock. C. An interaction function with a more negative b2 can also possesses a larger a1 before the solution becomes unstable. Jacobian for various values of a1 and determining when the eigenvalue with the maximal real part becomes positive. Panel 7B plots the real part of the eigenvalue with maximal real part as a function of a1: as the position shifts, the eigenvalues lose stability at different values of a1. The type of bifurcation by which they lose stability changes as well. For shocks located at even numbered sites, the system loses stability in a Hopf bifurcation as a complex conjugate pair of eigenvalues crosses the origin simultaneously. For shocks located at odd sites the system apparently loses stability in a sub-critical pitchfork bifurcation (This has not been rigorously proven). Figure 7C is a plot of acritical for the eigenvalues of the fifty-one oscillator Jacobian linearized about solutions corresponding to varying values of b2. The plot clearly demonstrates that shock solutions corresponding to a larger value of b2 can support a larger even component before becoming unstable. Figure 8 is a plot of acritical vs. the number of phase difference equations (number of oscillators in a chain). The shock is always located centrally: e.g. for 101 oscillators, the shock is located at N = 50. For an odd number of phase difference equations, the shock is position is obtained by dividing the number of equations in half and rounding up. Figure 8 shows phenomena similar to Figure 10A. That is, a solution will lose stability for different a1 dependent on the position of the 22 l a c i t i r c a 14 12 10 8 6 4 2 0 0 Hopf Bifurcation 5 10 15 Subcritical Pitchfork Bifurcation 35 40 45 50 20 phase difference equation number 25 30 FIG. 8. Critical value of a1 as a function of chain length where N represents the number of phase difference equations φN . Depending on the length of the chain and position of the shock, the solution may lose stability in either a Hopf bifurcation or what is believed to be a subcritical pitchfork bifurcation. The shock is always located centrally: (e.g. for 101 oscillators, the shock is located at N=50. For an odd number of phase difference equations, the shock is position is obtained by dividing the number of equations by half and rounding up.) shock in the solution as well as the number of oscillators in the chain. In the case that the shock is perfectly centered, we expect the solution to lose stability in a sub-critical pitchfork. Regardless of the position of the shock, even for relatively short chains the anti-wave solution will be stable for a relatively large even component, whose magnitude is at least as large as the first odd Fourier mode. The manner in which the solution loses stability and the size of the even component it can support depend on both the position of the kink/shock and the length of the chain. The Gershgorin circle theorem tells us that one of the anti-wave solutions will always be stable, no matter how long the chain. V. PATTERN FORMATION IN TWO-DIMENSIONAL ARRAYS Anti-wave patterns can be observed in two-dimensional networks as well. Using nearest neighbor coupling, the differential equation for a single oscillator is: dθx,y dt = H(θx+1,y − θx,y) + H(θx,y+1 − θx,y) + H(θx,y−1 − θx,y) + H(θx−1,y − θx,y), (37) where x, y are discrete indices describing the location of an oscillator. These indices run from 1 to N where N 2 is the number of differential equations in the array. As in the one dimensional case, we use non-reflecting boundary conditions. Examples of these patterns 23 A. C. x e d n i r o t a l l i c s O 20 40 60 80 100 x e d n i r o t a l l i c s O 20 40 60 80 100 10 20 30 10 20 30 40 50 60 Oscillator index 40 50 60 Oscillator index B. D. x e d n i r o t a l l i c s O 20 40 60 80 100 x e d n i r o t a l l i c s O 20 40 60 80 100 10 20 30 10 20 30 70 80 90 100 70 80 90 100 40 50 60 Oscillator index 40 50 60 Oscillator index 70 80 90 100 70 80 90 100 FIG. 9. Examples of two-dimensional patterns. The horizontal and vertical axis are the oscil- lator indices. A. Snapshot of a slowly varying wave pattern obtained from compacton-like initial conditions: a 2-dimensional pulse is initiated in the upper left-hand corner of the array on a synchronous background. The interaction function used is H(φ) = sin(φ) − cos(φ) − .75 sin(2φ) B. Stable stationary fractured pattern obtained from random initial conditions and interaction function:−2 cos(φ) − 0.518 sin(φ) − 1.31 sin(2φ) − .933 sin(3φ) C. Stationary anti-wave generated with H(φ) = cos(φ) + sin(φ) − .75 sin(2φ). D. Stationary traveling wave generated with the same interaction function as C. and the interaction functions used to generate them are are illustrated in Figure 9. Patterns 9B,9C and 9D are stationary and numerically stable. Pattern 9B is generated from random initial conditions. Patterns 9C and 9D are stable but have a very small basin of attraction. The initial conditions must be almost identical to the actual anti-wave or traveling wave solution. Analogous to the one-dimensional case, a plane wave may be represented by the solution: Whereas the shock solution may be written: θxy = kxx + kyy + ωt. θxy = kxx + kyy θxy = −kxx + kyy x ≤ x∗ x > x∗. (38) (39) In these equations x∗ is the location of the shock. A shock may be thought of as a discon- tinuous boundary between the two traveling waves. The fractured pattern in Figure 9B is composed of many small waves of varying kx and ky which form shocks. Fractured patterns 24 are generated with random initial conditions where the phases are chosen between 0 and 2π[41] [42]. VI. CONCLUSION Intermediate stable phase-locked states occur in a variety of neuron models and are capable of producing a pattern of wave-like activity known as an anti-wave. This type of activity was first observed in the dogfish spinal cord by Grillner [32] and it may occur in other networks as well[33][12]. The mechanism we use for generating these waves may have an experimental basis: Peinado et al. have shown that modulating the potassium conductance of gap junction coupled neurons in the neonatal mouse neocortex leads to wave behavior [13]. The mechanism for generating traveling waves and anti-waves is more flexible than past models, because it allows one to modulate the wavelength of the wave by adjusting properties such as the potassium conductance or temperature dependent time constant. It allows for a huge variety of phase-locked patterns in a chain or array of neurons. In the case of anti- wave generation, it does not require distal connections and the shock can form anywhere along the chain. The shock position can also be moved by colliding phase compactons. This may provide a mechanism by which an animal such as the dogfish could switch between different swimming patterns. More generally, near the bifurcation from in-phase synchrony, the interaction functions of the phase model have large higher order Fourier modes. This is in contrast to many oscillator models that were used in the past, which only include the lowest odd Fourier mode. The odd modes determine the value of the stable-phase locked solution but the even modes affect the stability of the anti-wave solution. Varying the relative even component affects the basin of attraction of a particular solution, or the probability that the phases will converge to a particular anti-wave solution from random initial conditions. Finally, we note that this behavior is not only relevant to chains but to two dimensional arrays as well. Appendix A: The Wang-Buszaki Model The parameters used in the Wang-Buszaki model are: vsyn = −60.5mV, gL = 0.1µS, vL = −65mV, gN a = 35 µS/cm2, VN a = 55mV, gK = 9 µS/cm2, VK = −90mV, ai0 = 4 25 τi = 15ms and i0 = .63 nA/cm2 . The nonlinearites mentioned in Section II A are given by the following expressions: αm(V ) = 0.1(V + 35.0)/(1.0 − exp(−(V + 35.0)/10.0)) βm(V ) = 4.0 exp(−(V + 60.0)/18.0) M∞(V ) = αm(V )/(αm(V ) + βm(V )) αh(V ) = 0.07 exp(−(V + 58.0)/20.0) βh(V ) = 1.0/(1.0 + exp(−(V + 28.0)/10.0)) H∞(V ) = αh(V )/(αh(V ) + βh(V )) αn(V ) = 0.01(V + 34.0)/(1.0 − exp(−(V + 34.0)/10.00)) βn(V ) = 0.125 exp(−(V + 44.0)/80.0) N∞(V ) = αn(V )/(αn(V ) + βn(V )) (A1) [1] S. Grillner, Science 228, 143 (1985), http://www.sciencemag.org/content/228/4696/143.full.pdf, http://www.sciencemag.org/content/228/4696/143.abstract [2] E. Izhikevich, Dynamical Systems in Neuroscience: The Geometry of Excitabilty and Bursting (MIT press: Cambridge, Massachusetts, 2007) [3] G. Buzski and A. Draguhn, Science 304, 1926 (2004), http://www.sciencemag.org/content/304/5679/1926.full.pdf, http://www.sciencemag.org/content/304/5679/1926.abstract [4] A. H. Cohen, P. J. Holmes, and R. H. Rand, Journal of Mathematical Biology 13, 345 (1982), ISSN 0303-6812, 10.1007/BF00276069, http://dx.doi.org/10.1007/BF00276069 [5] B. Ermentrout, J. Flores, and A. Gelperin, Journal of Neurophysiology 79, 2677 (1998), http://jn.physiology.org/content/79/5/2677.full.pdf+html, http://jn.physiology.org/content/79/5/2677.abstract [6] B. Ermentrout, M. Pascal, and B. Gutkin, Neural Computation 13, 1285 (2001), http://www.mitpressjournals.org/doi/pdf/10.1162/08997660152002861, http://www.mitpressjournals.org/doi/abs/10.1162/08997660152002861 26 [7] J. G. Mancilla, T. J. Lewis, D. J. Pinto, and B. W. Connors, The Journal of Neuroscience J. 27, Rinzel, 2058 (2007), http://www.jneurosci.org/content/27/8/2058.full.pdf+html, http://www.jneurosci.org/content/27/8/2058.abstract [8] C. Vreeswijk, L. F. Abbott, and G. Bard Ermentrout, Journal of Compu- tational Neuroscience 1, 313 (1994), ISSN 0929-5313, 10.1007/BF00961879, http://dx.doi.org/10.1007/BF00961879 [9] A silicon neuron refers to an electronic circuit designed to mimic a biological neuron [10] G. S. Cymbalyuk, G. N. Patel, R. L. Calabrese, S. P. De- Weerth, and A. H. Cohen, Neural Computation 12, 2259 (2000), http://www.mitpressjournals.org/doi/pdf/10.1162/089976600300014926, http://www.mitpressjournals.org/doi/abs/10.1162/089976600300014926 [11] A. Gelperin, Journal of Experimental Biology 202, 1855 (1999), http://jeb.biologists.org/content/202/14/1855.full.pdf+html, http://jeb.biologists.org/content/202/14/1855.abstract [12] U. Kim, T. Bal, and D. A. McCormick, Journal of Neurophysiology 74, 1301 (1995), http://jn.physiology.org/content/74/3/1301.full.pdf+html, http://jn.physiology.org/content/74/3/1301.abstract [13] A. Peinado, Journal of Neurophysiology 85, 620 (2001), http://jn.physiology.org/content/85/2/620.full.pdf+html, http://jn.physiology.org/content/85/2/620.abstract [14] G. Ermentrout and D. Kleinfeld, Neuron 29, 33 (2001), ISSN 0896-6273, http://www.sciencedirect.com/science/article/B6WSS-42D33KV-7/2/9571b90cc1aea3c7aa28c6ff54dea0c8 [15] Y. Momose-Sato, K. Sato, and M. Kinoshita, European Journal of Neuroscience 25, 929 (2007), ISSN 1460-9568, http://dx.doi.org/10.1111/j.1460-9568.2007.05352.x [16] G. B. Ermentrout and D. Terman, Mathematical Foundations of Neuroscience (Springer, 2012) [17] N. Kopell and G. B. Ermentrout, Communications on Pure and Applied Mathematics 39, 623 (1986), ISSN 1097-0312, http://dx.doi.org/10.1002/cpa.3160390504 [18] P. C. Bressloff, S. Coombes, and B. de Souza, Phys. Rev. Lett. 79, 2791 (Oct 1997), http://link.aps.org/doi/10.1103/PhysRevLett.79.2791 [19] S. Jones, B. Mulloney, T. Kaper, and N. Kopell, The Journal of neuroscience 23, 3457 (2003) [20] G. B. Xiao-Jing Wang, The Journal of Neuroscience 16, 6402 (October 1996) 27 [21] η = QT −Tbase 10 : in this equations, Q0 is the "ratio of the rates for an increase in temperature of 100C" [16]. [22] B. Pfeuty, G. Mato, D. Golomb, and D. Hansel, The Journal of Neuroscience 23, 6280 (2003), http://www.jneurosci.org/content/23/15/6280.full.pdf+html, http://www.jneurosci.org/content/23/15/6280.abstract [23] M. P. Nusbaum and M. P. Beenhakker, Nature (London) 417, 343 (May 2002) [24] S. P. Javedan, R. S. Fisher, H. G. Eder, K. Smith, and J. Wu, Epilepsia 46, 574 (2002) [25] S. H. and Strogatz, Physica D: Nonlinear Phenomena 143, 1 (2000), ISSN 0167-2789, http://www.sciencedirect.com/science/article/pii/S0167278900000944 [26] H. Haken, J. Kelso, and H. Bunz, Biological cybernetics 51, 347 (1985) [27] G. B. Ermentrout and N. Kopell, SIAM Journal on Applied Mathematics 54, pp. 478 (1994), ISSN 00361399, https://sremote.pitt.edu:11017/stable/2102230 [28] D. A. Wiley, S. H. Strogatz, and M. Girvan, CHAOS 16, 015103 (2006), ISSN 10541500, http://dx.doi.org/10.1063/1.2165594 [29] B. Blasius and R. Tonjes, Phys. Rev. Lett. 95, 084101 (Aug 2005), http://link.aps.org/doi/10.1103/PhysRevLett.95.084101 [30] These are not to be confused with fractured waves studied by Kopell (also known as s-waves) [43]. [31] B. Denardo, B. Galvin, A. Greenfield, A. Larraza, S. Put- terman, and W. Wright, Phys. Rev. Lett. 68, 1730 (Mar 1992), http://link.aps.org/doi/10.1103/PhysRevLett.68.1730 [32] S. Grillner, Experimental Brain Research 20, 459 (1974) [33] J. Christensen and R. Hauser, American Journal of Physiology -- Legacy Con- tent 221, 1033 (1971), http://ajplegacy.physiology.org/content/221/4/1033.full.pdf+html, http://ajplegacy.physiology.org/content/221/4/1033.short [34] K. W. Thompson, Journal of Computational Physics 68, 1 (1987), ISSN 0021-9991, http://www.sciencedirect.com/science/article/pii/0021999187900416 [35] To see that this is consistent with the preceding statement, start with the H(−φj−1) = H(φj), using a Taylor series to expand out each side and evaluating at the wave solution, we have 28 the expression: Since H ′(φ) 6= 0, then it must be that ( dφ dφ dx H ′(−φ) dxx=0) = 0. = 0 (2) [36] A. Pikovsky and P. Rosenau, Physica D: Nonlinear Phenomena 218, 56 (2006), ISSN 0167- 2789, http://www.sciencedirect.com/science/article/pii/S0167278906001382 [37] K. Ahnert and A. Pikovsky, Phys. Rev. E 79, 026209 (Feb 2009), http://link.aps.org/doi/10.1103/PhysRevE.79.026209 [38] K. Ahnert and A. Pikovsky, Chaos 18, 037118 (Sep. 2008), arXiv:0806.1833 [nlin.PS] [39] A. Urban, Intermediate Stable Phase Locked States in Oscillator Networks, Ph.D. thesis, Uni- versity of Pittsburgh (2011) [40] Gershgorin Circle Theorem: Let:A = [aij] be an arbitrary n x n matrix with elements that may be complex and let: Λi = n Xj=1,i6=j aij f or i = 1, 2, ...n Then all of the eigenvalues λi ofA lie in the union of n disks Γi where: This wording of the Gershgorin Circle theorem was taken from: Γi : λ − aii ≤ λi f or i = 1, 2, ...n (3) (4) Tables of Integrals, Series, and Products by I.S.Gradshteyn and I.M. Ryzhik[44]. [41] All equations in this section were integrated with Euler's method. The step size used is .01. [42] The Fourier terms used may or may not be calculated from the full model. For instance, an interaction function with an artificially inflated second odd mode will produce an interesting fractured pattern, whereas many interaction functions computed from the full model will not. [43] D. Somers and N. Kopell, Physica D: Nonlinear Phenomena 89, 169 (1995), ISSN 0167-2789, http://www.sciencedirect.com/science/article/pii/0167278995001980 [44] I. Gradshteyn and I. Ryzhik, Tables of Integrals, Series and Products (Academic Press, 2000) 29
1610.05982
3
1610
2017-06-21T07:58:06
The temporal evolution of the central fixation bias in scene viewing
[ "q-bio.NC" ]
When watching the image of a natural scene on a computer screen, observers initially move their eyes towards the center of the image --- a reliable experimental finding termed central fixation bias. This systematic tendency in eye guidance likely masks attentional selection driven by image properties and top-down cognitive processes. Here we show that the central fixation bias can be reduced by delaying the initial saccade relative to image onset. In four scene-viewing experiments we manipulated observers' initial gaze position and delayed their first saccade by a specific time interval relative to the onset of an image. We analyzed the distance to image center over time and show that the central fixation bias of initial fixations was significantly reduced after delayed saccade onsets. We additionally show that selection of the initial saccade target strongly depended on the first saccade latency. Processes influencing the time course of the central fixation bias were investigated by comparing simulations of several dynamic and statistical models. Model comparisons suggest that the central fixation bias is generated by a default activation as a response to the sudden image onset and that this default activation pattern decreases over time. Our results suggest that it may often be preferable to use a modified version of the scene viewing paradigm that decouples image onset from the start signal for scene exploration and explicitly controls the central fixation bias. In general, the initial fixation location and the latency of the first saccade need to be taken into consideration when investigating eye movements during scene viewing.
q-bio.NC
q-bio
The temporal evolution of the central fixation bias in scene viewing Lars O. M. Rothkegel1∗, Hans A. Trukenbrod1, Heiko H. Schutt1,2, Felix A. Wichmann2−4, and Ralf Engbert1 1University of Potsdam, Germany 2Eberhard Karls University Tubingen, Germany 3Bernstein Center for Computational Neuroscience Tubingen, Germany 4Max Planck Institute for Intelligent Systems, Tubingen, Germany June 22, 2017 ∗To whom correspondence should be addressed: Lars Rothkegel Department of Psychology & Cognitive Sciene Program University of Potsdam Am Neuen Palais 10 14469 Potsdam Germany E-mail: [email protected] Phone: +49 331 9772370, Fax: +49 331 9772794 Temporal evolution of the central fixation bias 1 Abstract When watching the image of a natural scene on a computer screen, observers initially move their eyes towards the center of the image - a reliable experimental finding termed central fixation bias. This systematic tendency in eye guidance likely masks attentional selection driven by image properties and top-down cognitive processes. Here we show that the central fixation bias can be reduced by delaying the initial saccade relative to image onset. In four scene-viewing experiments we manipulated observers' initial gaze position and delayed their first saccade by a specific time interval relative to the onset of an image. We analyzed the distance to image center over time and show that the central fixation bias of initial fixations was significantly reduced after delayed saccade onsets. We additionally show that selection of the initial saccade target strongly depended on the first saccade latency. Processes influencing the time course of the central fixation bias were investigated by comparing simulations of several dynamic and statistical models. Model comparisons suggest that the central fixation bias is generated by a default activation as a response to the sudden image onset and that this default activation pattern decreases over time. Our results suggest that it may often be preferable to use a modified version of the scene viewing paradigm that decouples image onset from the start signal for scene exploration and explicitly controls the central fixation bias. In general, the initial fixation location and the latency of the first saccade need to be taken into consideration when investigating eye movements during scene viewing. Introduction How humans visually explore natural scenes depends on multiple factors. Eye movements are influenced by low level image properties (e.g., chromaticity, orientation, luminance, 2 Rothkegel et al. and color contrast; Itti, Koch, & Niebur, 1998; Torralba, 2003; Le Meur, Le Callet, Barba, & Thoreau, 2006) as well as higher level cognitive processes like the observers' scene understanding (Loftus & Mackworth, 1978; Henderson, Weeks Jr, & Hollingworth, 1999), task (Yarbus, Haigh, & Rigss, 1967; Castelhano & Henderson, 2008), or probability of reward (Hayhoe & Ballard, 2005; Tatler, Hayhoe, Land, & Ballard, 2011). Besides low-level image features and high-level cognition, systematic tendencies have a strong impact on how humans look at pictures (Tatler & Vincent, 2009; Le Meur & Liu, 2015). A dominant systematic tendency in natural scene viewing is the central fixation bias (CFB; Buswell, 1935; Tatler, 2007; Tseng, Carmi, Cameron, Munoz, & Itti, 2009). Regardless of stimulus material (Tatler, 2007; Tseng et al., 2009), head position (Vitu, Kapoula, Lancelin, & Lavigne, 2004), initial fixation position (Tatler, 2007; Bindemann, Scheepers, Ferguson, & Burton, 2010), or image position (Bindemann, 2010), the eyes tend to initially fixate close to the center of an image when presented to a human observer on a computer screen. After several explanations of the CFB had been ruled out, two hypotheses remained. First, the image center might be the best location to maximize information extraction from scenes (Najemnik & Geisler, 2005; Tatler, 2007) – at least for typical photographs found in image databases and on the internet (c.f; Wichmann, Drewes, Rosas, & Gegenfurtner, 2010). Second, the center of an image provides a strategic advantage to start inspection of a suddenly appearing stimulus (Tatler, 2007). Since realworld visual input doesn't suddenly appear and peripheral information of an upcoming stimulus is usually available, the CFB might be a laboratory artifact to some degree. Also, natural visual stimuli do not have rigid boundaries like a computer screen. A reduction of the CFB in mobile eye tracking data ('t Hart et al., 2009; Ioannidou, Hermens, Hodgson, et al., 2016) supports this idea. Temporal evolution of the central fixation bias 3 A previous study from our lab resulted in a strong reduction of the CFB on initial fixations compared to similar experiments. In this study we manipulated the initial fixation by requiring participants to maintain fixation on a starting position close to the border of the screen for 1 s (Rothkegel, Trukenbrod, Schutt, Wichmann, & Engbert, 2016). In addition, some images in this study had asymmetric conspicuity distributions, with interesting or salient image parts on either side of the image, but less so in the center. Thus, the reduction of the CFB in our scene viewing experiment could have been generated by three aspects: extreme initial starting positions, delayed initial saccades, and the saliency bias of the images we used. To investigate the principles underlying the reduced CFB, we designed and analyzed four experiments, in which observers started exploration from different positions within an image and were required to maintain fixation for various time intervals after image onset (pre-trial fixation time). Our study used the images investigated in the most frequently cited paper on the central fixation bias (Tatler, 2007), in order to exclude any influence of the images on the reduction of the CFB. We hypothesized that (i) a forced prolonged initial fixation decouples image onset from the signal to start exploration and leads to a reduced CFB on the second fixation which in turn reduces the bias on subsequent fixations (due to the short saccade amplitudes of humans during scene perception; Tatler & Vincent, 2008) and that (ii) the magnitude of the reduction varies with the duration of the prolonged initial fixation. Here we show that the CFB of early eye movements can be reduced by dissociating initial eye movements from a sudden image onset by 125 ms and more. Increasing the delay of the initial response by more than 125 ms produced only marginal differences. In addition, we show that the initial saccade latency predicts the strength of the CFB on a trial-by- 4 Rothkegel et al. trial basis. The pre-trial fixation time primarily assures that the initial fixation is long enough to avoid a strong orienting response to the center of an image. Finally we compare our data to five different models of saccade generation in scene view- ing. An extended version of a recently published model of saccade generation (Engbert, Trukenbrod, Barthelm´e, & Wichmann, 2015) with an additional initial central activation that decreases over time best explains the data in terms of maximum likelihood (Schutt et al., 2017) and qualitative progression of the CFB over time. In terms of maximum likelihood a model that imitates the visual attention span of humans combined with the empirical fixation map explains the data equally well but fails to reproduce the qualitative progression of the CFB over time. General Methods Stimuli A set of 120 images was presented on a 20-inch CRT monitor (Mitsubishi Diamond Pro 2070: frame rate 120 Hz, resolution 1280×1024 pixels; Mitsubishi Electric Corporation, Tokyo, Japan) in Experiment 1, 2 and 4 and on a different 20-inch CRT monitor (Iiyama Vision Master Pro 514: frame rate 100 Hz, resolution 1280×1024 pixels; Iiyama, Nagano, Japan) in Experiment 3. The images were the same as in Tatler's (2007) original study on the central fixation bias. Images were indoor scenes (40 images), outdoor scenes with man-made structures present (e.g., urban scenes; 40 images), and outdoor scenes with no man-made structures present (40 images). Images were taken using a Nikon D2 digital SLR using its highest resolution (4 megapixel). All pictures had a size of 1600×1200 pixels. For the presentation during the experiment, images were converted to a size of 1200×900 pixels and centered on a screen with gray borders extending 64 pixels to the top/bottom Temporal evolution of the central fixation bias 5 and 40 pixels to the left/right of the image. In Experiments 1, 2 and 4 the images covered 31.1◦ of visual angle in the horizontal and 23.3◦ in the vertical dimension. In Experiment 3 images covered a larger proportion of the visual field with 36.25◦ of visual angle in the horizontal and 27.20◦ in the vertical dimension due to a reduced viewing distance. Participants Participants were students of the University of Potsdam and of nearby high schools. Num- ber of participants for each experiment is indicated below. They received credit points or a monetary compensation of 8 Euro for their participation in any of the four experiments. The average duration of one experimental session was 40-45 minutes. All participants had normal or corrected-to-normal vision. The work was carried out in accordance with the Declaration of Helsinki. Informed consent was obtained for experimentation by all participants. General Procedure Participants were instructed to position their heads on a chin rest in front of a computer screen at a viewing distance of 70 cm (60 cm in Exp. 3). Eye movements were recorded binocularly (monocularly in Exp. 3) using an Eyelink 1000 video-based-eyetracker (SR- Research, Osgoode/ON, Canada) with a sampling rate of 500 Hz (1000 Hz in Exp. 3 and downsampled to 500 Hz for our analysis). Trials began with a black fixation cross presented on gray background. After successful fixation, an image was presented. After onset of the image, the fixation cross remained visible on top of the image for a variable duration. We refer to this duration as the pre-trial fixation time. Participants were instructed to keep their eyes on the fixation cross until it disappeared. If participants moved their eyes before the pre-trial fixation time elapsed, a mask of Gaussian white noise 6 Rothkegel et al. was displayed and the trial started anew with the initial fixation check. After successful initial fixation, participants were instructed to explore the scene freely for five seconds in all experiments. Experiments were run with the Matlab software (MATLAB, 2015) using the Psychophysics (Brainard, 1997; Pelli, 1997; Kleiner et al., 2007) and Eyelink (Cornelissen, Peters, & Palmer, 2002) toolboxes. Data Analysis Data preprocessing and saccade detection. For saccade detection we applied a velocity- based algorithm (Engbert & Kliegl, 2003; Engbert & Mergenthaler, 2006). Saccades had a minimum amplitude of 0.5◦ and exceeded an average velocity during a trial by 6 (median- based) standard deviations for at least 6 data samples (12 ms). The epoch between two subsequent saccades was defined as a fixation. Distance to Center over Time We computed the mean distance of the eye position to the image center DT C as a function of pre-trial fixation time (T ). This was computed as follows n(cid:88) m(cid:88) xjk(t) − xcenter , (1) DT C T = 1 m · n j=1 k=1 where xjk(t) indicates gaze position of participant j on image k at time t and xcenter indicates the image center. As a continuous-time measure, we computed the DTC of each sample of the eye-position time series. In this representation, a larger DTC indicates a less pronounced CFB and vice versa. For all experiments we visualized the mean DT C(t) to the image center for the entire 5 s observation window for each pre-trial fixation time. The observation window started at t = 0 with the disappearance of the fixation marker. Temporal evolution of the central fixation bias 7 All figures were created with the ggplot2 package (Wickham, 2009) of the R-Language of Statistical Computing. Influence of the initial fixation on the second fixation. The pre-trial fixation time influ- enced the DTC on early fixations. To further investigate this influence, we plot the DTC of the second fixation as a function of overall saccade latency from image onset. We com- puted linear mixed models (Bates, Machler, Bolker, & Walker, 2015) with initial saccade latency and pre-trial fixation time as fixed effects, the DTC of the second fixation as the dependent variable and an intercept for subjects and images as random factors. To com- pute the models, we transformed DTC with the boxcox function of the R package MASS (Venables & Ripley, 2002) to follow a normal distribution. We obtained significance levels with the lmerTest package (Kuznetsova, Brockhoff, & Christensen, 2013). Contrasts were defined as sum contrasts. This means that each pre-trial fixation time is compared to the overall mean of distance to center. To be able to compare the different factor levels to the overall mean, the highest pre-trial fixation time in each experiment was left out. In all experiments we excluded saccades with a latency smaller than or equal to 80 ms as anticipatory. Density Maps of eye positions over time. To visualize the temporal evolution of eye positions in our experiments, we computed movies of 2D density maps for the different pre- trial fixation times and each eye position of the time-series recorded for each experiment. Based on a kernel density estimation via diffusion (Botev, Grotowski, & Kroese, 2010), we estimated density maps for the first two seconds (after removal of the fixation cross) in each experiment. These movies are available as supplementary material. 8 Methods Rothkegel et al. Experiment 1 Participants. We recorded eye movements from 40 participants in Experiment 1 (34 female, 14–39 years old; 2 from a nearby high school) Procedure. In Experiment 1 the fixation cross was presented at the horizontal meridian 5.6◦ (256 Pixels) away from the left or right border of the monitor. This position was chosen to reproduce the findings of a strongly reduced central fixation bias observed in an earlier study (Rothkegel et al., 2016), where participants experienced a pre-trial fixation time of 1 s. A proportion of 20% of participants explored the image immediately after successful fixation without an additional pre-trial fixation time (0 ms). This corresponds to the standard scene viewing paradigm. For all other participants the fixation cross remained on top of the image for a duration of 125 ms, 250 ms, 500 ms, or 1000 ms. Pre-trial fixation time was used as a between-subject factor, i.e., each participant was tested with one of five pre-trial fixation times. Figure 1 illustrates a representative trial with the starting position on the left side of the screen. Fixation Check 2 was nonexistent for participants with a 0 ms pre-trial fixation time. Results Distance to Center over Time. In Experiment 1, the DT C initially decreased for all conditions (i.e., the CFB increased). There was a pronounced effect that mean fixation positions tended to be closer to the image center when participants were allowed to explore an image immediately after image onset, i.e., with a pre-trial fixation time of 0 ms (black curves in Fig. 2a). Surprisingly, for the first 4 participants (Block 1) of this group the effect Temporal evolution of the central fixation bias 9 Figure 1. Schematic illustration of the experimental procedure of Experiment 1 with a starting position close to the left border of the screen. After a short fixation check of 200 ms (Fixation Check 1) the image is presented. A second fixation check between 0 and 1000 ms controls if par- ticipants move their eyes after image onset. After a succesful second fixation check, participants are allowed to freely move their eyes. was visible throughout the whole observation time of 5 s. A second group of participants in the 0 ms condition (Block 2) did not replicate the stronger CFB through the whole observation time. In addition, there was a gradual reduction of the CFB for pre-trial fixation times from 125 ms to 250 ms (red and green curve). DT C for pre-trial fixation times of 250 and 500 ms hardly differed (green vs. blue curve). The minimum for the pre-trial fixation time of 1000 ms occured later in time because of disproportionately long saccade latencies of the first saccade after a forced fixation on the fixation cross of 1000 ms (cyan curve). Distance to Center on the second fixation. Figure 2b shows the influence of initial saccade latency on the mean DTC of the second fixation for the five pre-trial fixation times. Each bin represents a quintile of the distribution of saccade latencies in each condition. A clear relation between DTC of the second fixation and latency of the initial saccade is visible Fixation Check 1 5 s of freeobservationFixation Check 2 (0;125;250;500;1000ms) 10 a Rothkegel et al. b Figure 2. Experiment 1. a) Mean distance to center over time (DT C(t)) for the five different pre-trial fixation times with starting positions close to the border of the screen. Confidence intervals indicate standard errors as described by Cousineau (2005). Block 1 represents partici- pants 1-20, Block 2 participants 21-40 who were originally tested as a follow up experiment to consolidate the results. b) Mean distance to center of the second fixation as a result of initial fix- ation duration. Error bars are 95% confidence intervals of 1000 bootstrap samples as suggested by Efron and Tibshirani (1994). for the pre-trial fixation times of 0 ms and 125 ms. Overall, short saccade latencies led to a small average DTC (i.e., a strong initial CFB) whereas long latencies led to a larger average DTC (i.e., a less pronounced initial CFB). Table 1 shows the output of the LMM for Experiment 1. The DTC for a pre-trial fixation time of 0 ms is significantly lower than the average DTC and for a pre-trial fixation time of 500 ms it is significantly higher. This means that a saccade immediately after a sudden image onset led to the strongest overall CFB in this experiment. The initial saccade latency is highly significant regardless of the pre-trial fixation time. The model also shows that an interaction between saccade latency and pre-trial fixation time exists. If the image appears directly after a succesfull fixation check (pre-trial fixation of 0 ms) the 46810020004000time [ms]distance to image center [°]block11+22pre−trial fixation time [s]00.1250.250.51lllllllllllllllllllllllll34567500100015002000initial saccade latency [ms]distance to center on second fixation [°]pre−trial fixation time [s]lllll00.1250.250.51 Temporal evolution of the central fixation bias 11 Table 1: Output of LMM for Experiment 1 Fixed Effect Estimate 1.856 −0.925 −0.148 0.185 0.496 0.751 1.685 0.245 −0.210 −0.886 t 23.546∗∗∗ −6.106∗∗∗ −1.053 1.320 3.660∗∗∗ 6.951∗∗∗ 4.976∗∗∗ 1.178 −1.199 −5.726∗∗∗ (Intercept) 0 ms 125 ms 250 ms 500 ms saccade latency saccade latency x 0 ms saccade latency x 125 ms saccade latency x 250 ms saccade latency x 500 ms Random effects variance: Subjects Random effects variance: Images Log-Likelihood Deviance AIC BIC N SE 0.079 0.151 0.140 0.140 0.136 0.108 0.339 0.208 0.175 0.155 0.1498 0.1477 -7135.53 14271.07 14297.07 14380.64 4575 ∗p < .05, ∗∗p < .01, ∗∗∗p < .001 influence of saccade latency is significantly higher than on average (see row saccade latency × 0 ms). If pre-trial fixation time is as long as 500 ms, the influence of saccade latency is significantly weaker than on average (see line saccade latency × 500 ms). This interaction suggests that after a certain threshold time is reached, the influence of increasing saccade latency disappears. Discussion Experiment 1 led to a reduction of the CFB on the initial saccade target for all pre-trial fixation times of 125 ms and more during scene perception from extreme starting positions (Fig. 2a). A pre-trial fixation time of 125 ms produced an intermediate CFB, whereas longer pre-trial fixation times produced asymptotic behavior. With a pre-trial fixation time of 0 s the DTC was smaller throughout almost the whole observation time of 5 s 12 Rothkegel et al. for the first group of participants. However, this effect was not replicated in a retest with 20 new participants. In this experiment the long influence of the pre-trial fixation time indeed vanished. The early effect of the CFB did not differ in the two groups of participants. The CFB of the second fixation did strongly depend on the latency of the initial saccade (Fig. 2b). Thus, the early differences between pre-trial fixation times in Figure 2a are driven by differences in the distribution of intial saccade latencies. These results replicated our earlier findings of a reduced CFB during scene perception by introducing a non-zero pre-trial fixation time (Rothkegel et al., 2016). A delay of 125 ms seemed sufficient to achieve a considerable reduction. In addition, our results suggested that the most important mediating factor of the CFB was the latency of the first saccadic response. Saccades with brief saccade latencies were on average directed more strongly towards the center than saccades with long saccade latencies. Experiment 2 To assure that our results from Experiment 1 were not mainly induced by the extreme starting positions we conducted another experiment with starting positions closer to the image center. Methods Participants. We recorded eye movements from 20 participants for Experiment 2 (17 female; 14-28 years old; 1 from a nearby high school). Procedure. Experiment 2 was similar to Experiment 1 except that the fixation cross was presented on a donut shaped ring with a distance of 2.6◦ to 7.8◦ (100 to 300 pixels) to the center. We used this donut shaped ring to obtain intermediate starting positions neither Temporal evolution of the central fixation bias 13 a b Figure 3. Experiment 2. a) Mean distance to center over time (DT C(t)) for the five different pre-trial fixation times with starting positions on a donut shaped ring around the image center. Confidence intervals indicate standard errors as described by Cousineau (2005). b) Mean dis- tance to center of the second fixation as a result of initial saccade latency and pre-trial fixation time. Bins represent quintiles of the saccade latency distribution. Errorbars are the standard error of the mean. too close nor too far away from the center so that fixations could be directed both towards and away from the center. In addition, the donut shaped ring of starting positions made the initial starting position less predictable. This setup was thus slightly different to the experiment conducted by Tatler (2007) where the initial starting position was randomly chosen from a circle (fixed radius) around the image center. Results Distance to Center over Time. In Experiment 2, where the starting positions were located on a ring around the image center, the eyes moved initially even further towards the image center in the 0 ms pre-trial condition (black curve in Fig. 3a was the only curve with a 46810020004000time [ms]distance to image center [°]pre−trial fixation time [s]00.1250.250.51lllllllllllllllllllllllll3456500100015002000initial saccade latency [ms]distance to center on second fixation [°]pre−trial fixation time [s]lllll00.1250.250.51 14 Rothkegel et al. pronounced negative slope in the beginning). A difference in DT C was visible until about 600 ms after offset of the fixation marker. Later during the trial, the curves converged for all pre-trial conditions and reached a stable DTC for the rest of the trial. Qualitatively, we also observed a small initial difference in DTC between short pre-trial fixation times of 125 ms and 250 ms and pre-trial fixation times of 500 ms and 1000 ms. Distance to Center of the second fixation. As in Experiment 1, we found a strong influence of the latency of the first saccade on the DTC of the second fixation for small pre-trial fixation times (Fig. 3b). The results of the linear mixed model in Experiment 2 (Tab. 2) were similar to Experiment 1. The most important results are the significantly lower DTC of the 0 ms pre-trial fixation time compared to the average and the significant increase in DTC for higher saccade latencies. As in Experiment 1 an interaction between saccade latency and pre-trial fixation time is visible. This is especially true for the 0 ms condition, where the influence of saccade latency significantly increases compared to the average influence. In Experiment 2 the only significant decrease in saccade latency influence is visible for a pre-trial fixation time of 250 ms. The overall direction of the influence (increasing influence of saccade latency for pre-trial fixation times of 0 & 125 ms vs. decreasing influence for pre-trial fixation times of 250 & 500 ms) remains as in Experiment 1. Discussion If the starting position was close to the image center all pre-trial fixation times of 125 ms or longer (Fig. 3a) led to a reduction of the CFB on early fixations. After around 600 ms this influence disappeared. Furthermore, a clear relation between latency of the first saccade and the CFB of the second fixation was visible (Fig. 3b). Thus, the results Temporal evolution of the central fixation bias 15 Table 2: Output of LMM for Experiment 2 Fixed Effect Estimate 2.167 −0.899 −0.010 0.228 0.255 0.468 1.126 0.327 −0.541 −0.235 SE 0.094 0.186 0.177 0.174 0.177 0.110 0.291 0.229 0.204 0.208 0.1112 0.1333 (Intercept) 0 ms 125 ms 250 ms 500 ms saccade latency saccade latency x 0 ms saccade latency x 125 ms saccade latency x 250 ms saccade latency x 500 ms Random effects variance: Subjects Random effects variance: Images Log-Likelihood Deviance AIC BIC N −3599.60 7199.20 7225.20 7299.56 2253 t 22.965∗∗∗ −4.829∗∗∗ −0.054 1.308 1.439 4.258∗∗∗ 3.870∗∗∗ 1.430 −2.656∗∗ −1.132 ∗p < .05, ∗∗p < .01, ∗∗∗p < .001 replicated our observations from Experiment 1 and demonstrated that a reduced CFB was not exclusively generated by the extreme starting positions used in Experiment 1. Experiment 3 The results from Experiment 1 and 2 showed that a pre-trial fixation time of 125 ms was enough to reduce the central fixation bias on early fixations. The difference of the CFB between pre-trial fixation times larger than 125 ms was relatively small. To investigate the minimum pre-trial fixation time for a substantial CFB reduction, we conducted a third experiment with pre-trial fixation times ranging from 0 to 125 ms in six equidistant steps. We changed the between-subject design of pre-trial fixation time to a within- subject design to reduce the influence of individual participants (cf., Exp. 1). Hence, every participant was tested with all pre-trial fixation times. Since effects were maximal 16 Rothkegel et al. in the first experiment we used the same extreme starting positions as in Experiment 1. Methods Participants. We recorded eye movements from 24 participants for Experiment 3 (20 female; 20–29 years old) Procedure. In Experiment 3, participants experienced pre-trial fixation times between 0 and 125 ms in steps of 25 ms (0, 25, 50, 75, 100, 125 ms). Each of the six pre- trial fixation times was presented in a block of 20 images, pseudo-randomized across participants. In total participants viewed 120 images. Note, the experiment was tested with a different setup (monitor, eye-tracker, etc.; see general methods section for details). Thus, the absolute value of DTC is not directly comparable between Experiment 3 and the remaining experiments. Results Distance to Center over Time. As in Experiment 1 and 2 the eyes initially moved towards the center for all pre-trial fixation times (Fig. 4a). The difference between pre-trial con- ditions was not as clearly visible as in previous experiments. Even the difference between the 0 and the 125 ms condition was relatively small. The smaller difference was probably due to the blocked design where pre-trial fixation times changed after 20 trials during the experiment for each participant. Nonetheless, curves with a pre-trial fixation time smaller than or equal to 50 ms had smaller minima than the ones with pre-trial fixation times larger than 50 ms (see inset in Fig. 4a). Distance to Center of the second fixation. The influence of the first saccade latency on the distance to center of the second fixation is clearly visible in Figure 4b. The influence Temporal evolution of the central fixation bias 17 a b Figure 4. Experiment 3. a) Mean distance to center over time (DT C(t)) for the six different pre-trial fixation times with starting positions close to the left and right border. Confidence intervals indicate standard errors as described by Cousineau (2005). b) Mean distance to center of the second fixation as a result of initial saccade latency and pre-trial fixation time. Bins represent quintiles of the saccade latency distribution. Errorbars are the standard error of the mean. seemed even clearer than in previous experiments. However, the range of the distance to center values was larger in this experiment as a result of the increased magnitude of the image in visual degree. Saccade latencies were more homogeneous in Experiment 3. The difference of mean saccade latencies (pre-trial fixation time + saccade latency after removal of the fixation marker) between the 0 and 125 ms condition was much smaller (57 ms) than in Experiments 1 (154 ms) and 2 (138 ms). A linear mixed model for Experiment 3 showed that DTC of the second fixation did not show an independent influence of pre-trial fixation time (Tab. 3). However, we replicated a significant influence of the first saccade latency on DTC of the second fixation. Shorter saccade latencies led to fixations closer to the center of an image. An interaction between pre-trial fixation time and saccade latency was not observed. 681012020004000time [ms]distance to image center [°]pre−trial fixation time [s]00.0250.050.0750.10.125●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●45678200400600initial saccade latency [ms]distance to center on second fixation [°]pre−trial fixation time [s]●●●●●●00.0250.050.0750.10.125llllllllllllllllllllllllllllll45678200400600initial saccade latency [ms]distance to center on second fixation [°]pre−trial fixation time [s]llllll00.0250.050.0750.10.125 18 Rothkegel et al. Table 3: Output of LMM for Experiment 3 Fixed Effect Estimate 2.187 0.056 −0.096 −0.069 0.055 −0.015 1.056 −0.334 0.161 0.071 −0.093 0.153 SE 0.110 0.076 0.086 0.072 0.081 0.078 0.110 0.201 0.249 0.208 0.227 0.226 0.2308 0.1359 (Intercept) 0 ms 25 ms 50 ms 75 ms 100 ms saccade latency saccade latency x 0 ms saccade latency x 25 ms saccade latency x 50 ms saccade latency x 75 ms saccade latency x 100 ms Random effects variance: Subjects Random effects variance: Images Log-Likelihood Deviance AIC BIC N −3990.68 7981.35 8011.35 8099.75 2679 t 19.832∗∗∗ 0.731 −1.118 −0.949 0.673 −0.190 9.564∗∗∗ −1.662 0.645 0.340 −0.409 0.678 ∗p < .05, ∗∗p < .01, ∗∗∗p < .001 The saccade latencies of Experiment 3 were rather similar between the different pre-trial fixation times. There was a clear difference between the three lowest pre-trial fixation times (mean saccade-latencies of 315, 320 & 321 ms) compared to the three longer pre- trial fixation times (mean saccade-latencies of 365, 352 & 371 ms). Thus somewhere around 75 ms seems to be the lowest pre-trial fixation time to produce an influence in viewing behaviour. Discussion Experiment 3 was conducted to investigate the minimum pre-trial fixation time necessary for a reduction of the early central fixation bias. All pre-trial conditions showed a similar behavior with a tendency of an early CFB as measured by the DTC. We observed the Temporal evolution of the central fixation bias 19 weakest DTC effect for pre-trial fixation times of 125 ms (inset in Fig. 4a). Pre-trial fixation times equal to or smaller than 50 ms generated fixation positions closest to the image center. Differences in DTC could be explained by the influence of the first saccade latency on the selection of the second fixation location (Fig. 4b). Thus, saccade latencies are the most important factor modulating the CFB. A post-hoc analysis revealed that saccade latencies were only affected in conditions with pre-trial fixation times larger than 50 ms. We conclude that a minimum pre-trial fixation time of around 75 ms is needed to prolong saccade latencies in order to reduce the CFB in scene viewing. Experiment 4 In Experiment 4, participants started exploration at the center of the screen. This starting position was chosen to quantify the influence of pre-trial fixation times in a standard scene viewing paradigm. Methods Participants. In this experiment we recorded eye movements from 10 participants (3 male; 18–36 years old) Procedure. Experiment 4 followed the same procedure as the preceding experiments but participants started observation in the center of the screen. We tested pre-trial fixation times of 0, 125 and 250 ms since we observed only subtle changes of results for longer pre-trial fixation times in Experiment 1 and 2. As in Experiment 3, we used a within subject design for the three different pre-trial fixation times such that participants viewed blocks of 40 images for each pre-trial fixation time. 20 a Rothkegel et al. b Figure 5. Experiment 4. a) Mean distance to center over time (DT C(t)) for the three different pre-trial fixation times with starting positions in the center of the image. Confidence intervals indicate standard errors as described by Cousineau (2005). b) Mean distance to center of the second fixation as a result of initial saccade latency and pre-trial fixation time. Bins represent quintiles of the saccade latency distribution. Errorbars are the standard error of the mean. Results Distance to Center over Time. Contrary to the first experiments initial gaze positions could only move away from the image center with central starting positions in Experiment 4 (Fig. 5a). Therefore, DTC gradually increased until it reached an asymptote. Between pre-trial conditions, DTC differed with respect to the point in time, when curves started to monotonically increase (pre-trial fixation times: 250 ms < 125 ms < 0 ms). Although pre-trial fixation times were chosen to be equidistant, curves for 125 ms and 0 ms pre-trial conditions (red & black curve) converged later than curves for the 250 ms and 125 ms pre-trial conditions (green & red curve). This demonstrated that pre-trial fixation times of 125 ms or more reduces the CFB of early fixations even during scene viewing with central starting positions. 0246810020004000time [ms]distance to image center [°]pre−trial fixation time [s]00.1250.25lllllllllllllll45675001000initial saccade latency [ms]distance to center on second fixation [°]pre−trial fixation time [s]lll00.1250.25 Temporal evolution of the central fixation bias 21 Distance to Center of the second fixation. Latencies of the first saccade were longer in this experiment than in any of the other experiments. This observation is in line with results from face perception, where the initial fixation is longer when participants start exploring a face in the center (Arizpe, Kravitz, Yovel, & Baker, 2012). Due to the increased number of long initial saccade latencies, an influence of saccade latency on the second fixation location was not as clearly visible as in the previous experiments (Fig. 5b). Results of a linear mixed model for Experiment 4 partially replicated the main results from Experiment 1–3. The DTC of the second fixation was significantly smaller for a pre- trial fixaiton time of 0 ms. Ths influence of saccade latency on distance to center of the second fixation did not reach a level of significance of 95% in Experiment 4. The direction of the influence did stay positive though and nearly reached the level of significance. The fact that saccade latency was not a significant predictor is a result of the rather long latencies and a small number of participants. By removing initial saccade latencies of higher than 1 s (which normally are very rare) saccade latency becomes a significant predictor (p < 0.03). The interaction between saccade latency and pre-trial times showed that the influence of saccade latency on DTC was, as observed in Experiments 1 and 2, significantly larger for a pre-trial fixation time of 0 ms. Discussion In our last experiment we investigated the effect of pre-trial fixation times on the CFB in a standard scene viewing experiment where participants start exploration from the image center. As expected DTC increased in all conditions continuously until it reached an asymptote. The point in time when DTC started to increase varied for different pre-trial fixation times. We measured the earliest response for pre-trial fixation times of 250 ms and the slowest response after no pre-trial fixation times (0 ms). If we remove latencies of 22 Rothkegel et al. Table 4: Output of LMM for Experiment 4 Fixed Effect Estimate 2.858 −0.309 0.155 0.224 0.454 −0.189 t 18.194∗∗∗ −3.312∗∗∗ 1.877 1.837 2.611∗∗ −1.222 (Intercept) 0 ms 125 ms saccade latency saccade latency x 0 ms saccade latency x 125 ms Random effects variance: Subjects Random effects variance: Images Log-Likelihood Deviance AIC BIC N SE 0.157 0.093 0.083 0.122 0.174 0.154 0.2607 0.1845 −1877.00 3754.01 3772.01 3817.26 1128 ∗p < .05, ∗∗p < .01, ∗∗∗p < .001 higher than 1 s we can replicate an influence of saccade latencies on DTC of the second fixation. In general saccade latency seems to be a strong mediating factor of the CFB. In addition, we observed long initial saccade latencies when participants started at the image center. This is particularly worrying, because the first fixation is usually omitted from analyses in scene viewing experiments. When comparing Experiment 4 to the remaining experiments, the CFB was strongest when participants started at the image center without pre-trial fixation time (0 ms). Only after about 1 s DTC (& CFB) was comparable between experiments and pre-trial conditions. Since most scene viewing experiments last five seconds or less (c.f., data sets in MIT saliency benchmark; Bylinskii et al., 2015) a substantial proportion of fixations is biased towards the center during a standard scene viewing experiment. A combination of a non-zero pre-trial fixation time and adjustments of the starting position will reduce the CFB and may help to better understand target selection during scene viewing. We will further comment on this issue in the general discussion. Temporal evolution of the central fixation bias 23 Figure shows the influence of initial saccade latency on distance to image center for all 4 experiments combined. A clear increase of DTC is visible between 150 and 400 ms. Because initial saccade latencies above 400 ms do not show an influence, pre-trial fixation times above 250 ms did not produce effects noteworthy. This also explains why in Ex- periment 4 the rather long saccade latencies were not a significant predictor for the CFB. We conclude our experiments by stating that the initial saccade latency is the dominant factor influencing the early central fixation bias in scene viewing and that by delaying the initial saccade we ensured that the CFB was not exhibited as strongly as in the standard scene viewing paradigm. Figure 6. fixation for all 4 experiments combined. Influence of initial saccade latency on the distance to image center of the second Model simulations Method Results from our experiments suggest that the early CFB evolves systematically over time with initial saccade latency as a major determinant. To further investigate the time lllllllllllllllllllllllll3456200400600800initial saccade latency [ms]distance to image center [°] 24 Rothkegel et al. course of the early CFB, we simulated scanpaths generated by different computational models for all experiments and compared simulated scanpaths to our empirical data. Additionally we computed the likelihood of the empirical data under each model (Schutt et al., 2017). Overall we compared five models. First, we simulated three control models: (i) a model that selects fixations proportional to the densities of the empirically observed fixations, (ii) a model which multiplies the fixation density with a Gaussian around the current fixation before selecting the next fixation, (iii) a model that selects fixations from a central activation map to produce a pure CFB. In a second step, we simulated two versions of the dynamical SceneWalk model (Engbert et al., 2015): (iv) the original model and (v) a version of the original model with a modulation of the activation map by a sudden image onset. For the sampling from a central activation map model (iii) and the extended SceneWalk model (v) parameters had to be estimated. We used a standard optimization algorithm (fminsearch) implemented in MATLAB (MATLAB, 2015) to obtain the parameters with maximum likelihood (Bickel & Doksum, 2015; Schutt et al., 2017) of fixations 2–4 of half of the participants (Exp. 1–4: N = 20/10/12/5) and a quarter of the images (N = 30). We estimated parameters from the second to fourth fixation only for efficiency reasons and since DT C curves reached a stable value for later fixations. All models were implemented on a grid of 128×128 cells. For all models we normalized activation maps for target selection to a sum of one to obtain a probability value for each grid cell. For each empirical scanpath, we computed a simulated scanpath with the same number of fixations and fixation durations. Density Sampling. As the most straightforward statistical control, we simulated scanpaths by randomly sampling from a 2D density map of all fixations on a given image, i.e. Temporal evolution of the central fixation bias 25 the empirical fixation map generated by all participants (N = 94). First, we applied kernel density estimation using the SpatStat package (Baddeley & Turner, 2005) of the R Language for Statistical Computing (R Core Team, 2014). Based on a Gaussian kernel function with a bandwidth parameter of λ = 1.05 (smoothing parameter of the original publication of our SceneWalk model) we computed the empirical fixation density map for each image. Second, to simulate a scanpath (i.e., a fixation sequence), we sampled randomly from this map proportional to the local density so that the normalized density at a particular location translated into the probability to generate a fixation at that position. Gaussian Model. Next, we implemented a statistical model that sequentially sampled from the empirical fixation map via a Gaussian-shaped aperture around fixation to mimic the limits of visual acuity and the attentional span. For a given fixation position x, the empirical fixation map was weighted by a two dimensional Gaussian centered around x, with a standard deviation of 4.88◦ visual angle. The same standard deviation was used for the size of the attention map of the SceneWalk model by Engbert et al. (see after next section). Fixation locations were sampled from this combined map. To generate a scanpath in this model, the map was recomputed after each fixation. Center Bias Model. Sampling of scanpaths of the pure CFB model was similar to the sampling of the empirical density map model. Contrary to the empirical density, we estimated an elliptical Gaussian around the image center to account for the CFB. Standard deviations for the horizontal and vertical elongation of the Gaussian were estimated as described in the Methods section. The estimated values for σx for the four experiments were 6.9◦, 6.6◦, 7.4◦ and 6.9◦ and for σy 4.3◦, 4.4◦, 4.6◦ and 4.4◦ indicating that the center bias was elongated in the horizontal direction. This is in agreement with the finding described by Clarke and Tatler (2014) for an image independent baseline. The largest 26 Rothkegel et al. values were estimated in Experiment 3, where the size of the image in degree of visual angle was larger than the other experiments (see General Methods). Scene Walk Model. This recently published saccade generation model (Engbert et al., 2015) proposes that eye movements are driven by two different time-dependent neural activation maps. A fixation map memorizes previous fixations and tags visited fixations locations, making them less probable to be fixated again shortly afterwards. Thus, this map serves as an inhibition of return mechanism (Itti & Koch, 2001; Klein, 2000). A second map, the attention map, reflects the attentional allocation on the given scene for a speficic fixation position. To compute the attention map, first an intermediate map is computed by multiplying a two dimensional Gaussian distribution centered at the current fixation position with the empirical saliency map of the image to reflect the reduced processing in the periphery. The influence of attention maps from previous fixations declines over time and thus the previous attention map is increasingly replaced by the map of the new fixation. An equivalent mechanism is used to control the dynamics of inhibition, i.e., the fixation map. The attention and inhibition maps prior to the first fixation are set to uniform distributions. After computation of the two maps for the current fixation position and duration, they are combined by subtracting the fixation map from the attention map to a target map. After the maps are combined, a target is chosen proportional to the relative activations (Luce, 1959) of the target map. Thus, positions where the fixation map is high whereas the attention map is low are rarely fixated and vice versa. Model parameters where chosen as in the published version of the SceneWalk model (Engbert et al., 2015). Two differences to the original publication were that i) we only computed new maps for each new fixation instead of computing new maps every ∆t = 10 ms. This was done because we chose fixation durations from the Temporal evolution of the central fixation bias 27 experimental data for our simulations and did thus not need to compute the activation maps for other possible fixation durations (see Rothkegel et al., 2016; Schutt et al., 2017) and ii) negative values of the grid and all values smaller than a constant η were transformed by an exponential function such that each value on the grid was larger than zero. This was done because the likelihood of a scanpath containing any impossible fixation would be zero. Thus we could not differentiate between models in terms of likelihood otherwise. After transforming these small or negative values the combined target map was normalized again to obtain a sum of 1. SceneWalk StartMap. Since the original SceneWalk model was not intended to reproduce the early CFB (see results of model simulations), we developed a modified version of the original model that takes the sudden image onset during scene perception into account. We made two changes. First, contrary to the original SceneWalk model with a constant activation across the entire attention map at the beginning of a trial, we decided to use an attention map with higher activations near the center of an image than at the periphery (see Fig. 7b). This was motivated by the sudden image onset that may lead to an initial prioritization of central locations. Second, we realized that the decay of the attention map was too fast during the initial fixation. Therefore, we added a parameter that specified the rate of decay during the initial fixation. For all other fixations we used the same decay parameter as during the original simulations. The initial center map was an elliptical Gaussian (cf., Center Bias model). We estimated two parameters for the standard deviations of the center map. The horizontal standard deviation σx was estimated at values of 3.5◦, 1.8◦ and 3.9◦ for Experiments 1–3. The vertical standard deviation σy for Experiment 1–3 was estimated at 2.3◦, 2.3◦ and 2.4◦. The parameters estimated for Experiment 4 were very large with σx = 136.0◦ and σy = 28 Rothkegel et al. Figure 7. Fixations 1–4 of a randomly chosen trial with a pre-trial fixation time of 0 ms. In this example the initial fixation duration is very short and thus the initial central activation map of the SceneWalk StartMap has a strong influence on the first saccade target. 4.2◦. This resulted in small initial differences in activations between center and periphery for simulations of Experiment 4 and was similar to the constant activations in the original model. The reason for this behaviour arises from the architecture of the model. Since activations in the attention map rise near fixation, central activations are prioritized initially when participants start to explore a scene near the image center. Figure 7 shows the simulated fixations 1–4 of a randomly chosen trial for the two different versions of the SceneWalk model. The initial fixation duration in this trial was very short (t = 184 ms) and thus the initial attention map of the SceneWalk StartMap model is biased stongly towards the center whereas the original SceneWalk model does not show an increased central activation during initial fixations. SceneWalkSceneWalkStartMapFixations 1-2Fixations 2-3Fixations 3-4 Temporal evolution of the central fixation bias 29 Results We simulated saccadic sequences with the same starting positions, number of fixations, and fixation durations as observed empirically. We computed the DT C of saccadic se- quences simulated for each experiment (Fig. 8 & 9a), compared the effect of saccade latencies on DTC of the second fixation (Fig. 9b), and computed the information gain for Fixation 2–10 for the different models (Fig. 10). The initial fixation was excluded from the likelihood estimation since the starting position was determined by the experiment. Distance to Center over Time. The average DTC over time is plotted in Figure 8. We averaged across pre-trial fixation times in a first analysis since DTC curves did not differ between pre-trial fixation times for all models but the SceneWalk StartMap model. In all experiments the SceneWalk StartMap model (green curve) produced a pattern most similar to the empirical data (black curve) but underestimated the empirical early CFB in Experiments 1–3. However, it was the only model to produce the initial characteristic dip of the DTC curves. While the original SceneWalk model (pink curve) was not able to generate an early CFB, it converged with the empirical data after 1-2 s and reached a stable DTC. The Density model (yellow curve) sampled fixations proportional to the empirical fixation density and was the first model to reach a stable DTC. Since DTC of the Density model and the empirical data converged, we can conclude that fixations across participants were distributed randomly and proportional to the empirical density of all fixation locations on an image after about 2 s. While participants initially explored the center, later fixations were not directed further into the periphery than expected by the overall distribution of fixations. Both the Gaussian model (cyan curve) and Center Bias model (blue curve) moved initially too slowly towards the center and reached a stable DTC after about 1 s that 30 Rothkegel et al. Figure 8. Distance to image center over time for the empirical data and the 5 saccade generation models. We did not split the data into groups of pre-trial fixation time for visibility purposes. The SceneWalk StartMap model (green curve) is the only one that progresses similarly to the empirical data (black curve). The Gauss model (cyan curve) and the center bias sampling (blue curve) do not converge with the empirical data in the end. Inlays represent the summed deviation of model curves to the empirical data. was too close to the center. The bad performance of the Gaussian model demonstrated that a second mechanism is needed to move fixations away from the current fixation location. In the SceneWalk model this is implemented by inhibitory tagging via the fixation map (c.f., Rothkegel et al., 2016). The initial decline of DTC of the Center Bias model was weaker and slower than observed in our experiments (Exp. 1–3). This indicated that the CFB on the second fixation was stronger than estimated from Fixation 2–4. The inlays in Figure 8 show the overall deviation of model curves from the empirical data. The SceneWalk StartMap outperformed other models and deviated the least in all four Experiment 1Experiment 2Experiment 3Experiment 4020004000020004000020004000020004000024681012time [ms]distance to center [°]ModelDensity ModelGauss ModelCenter BiasSceneWalkSceneWalkStartMapData Temporal evolution of the central fixation bias 31 a b a) Distance to image center over time for the empirical data and the SceneWalk Figure 9. StartMap model for the different pre-trial fixation times in Experiment 1. b) Influence of the initial saccade latency on the distance to image center on the second fixation for the empirical data and the SceneWalk StartMap model in Experiment 1. experiments. Next, we investigated the temporal evolution of the DTC for different pre-trial fixation times for the SceneWalk StartMap model (Fig. 9a). The SceneWalk StartMap model only took the initial saccade latency after image onset into account. This produced a qualitatively similar pattern for the different pre-trial fixation times as seen in the data. The qualitative progression for most pre-trial fixation times and experiments was similar to what was observed empirically. It is eminent though that the central fixation tendency produced by the model was too weak when compared to the data. This was probably a result of the method and the fixations used for the parameter estimation (see discussion). Finally, we evaluated the relation between latencies of the first saccade and DTC of the second fixation (Fig. 9b). Again this influence was only visible in the SceneWalk StartMap model as the other models do not incorporate a mechanism that depends on the initial saccade latency. The SceneWalk StartMap model produced a result pattern similar to DataSceneWalkStartMap01000200068106810time [ms]distance to image center [°]llllllllllllllllllllllllllllllllllllllllllllllllllDataSceneWalkStartMap50010001500200034567893456789initial saccade latency [ms]distance to center on second fixation [°]pre−trial fixation time [s]lllll00.1250.250.51 32 Rothkegel et al. the empirical data with a similar progresson of lines and a differentiation between pre- trial fixation times. However, the early CFB on the second fixation was too small in all experiments, i.e. the distance to center in all simulations was too large. Likelihood. We computed the likelihood for each model given the empirical data (Schutt et al., 2017). As explained above all model parameters were estimated on Fixation 2 to 4 of half of the participants (N = 5–20) and a quarter of the images (N = 30) for each experiment. The likelihood was evaluated on the other half of participants and on the rest of the images (N = 90) for Fixation 2 to 10. All models were computed on the same grid of 128×128 grid cells. The grid cells were normalized to obtain probabilities. This translates to a likelihood for a random fixation on a uniform map of 1/(128× 128) = 2−14. Thus the information gain for a fixation i at point (x,y) of a model map u can be computed as gaini = log2(u[xi, yi]) + 14. (2) The computed value represents the information gained relative to a random process that generates a uniform distribution of fixations. Figure 10 shows the information gain of all models for Fixation 2 to 10. The Density model (yellow curve) predicts fixation locations well, but the information gained from this model decreases over time. Thus saliency models could explain less information for later fixations than for the initial ones. The CenterBias explained initial fixations similarly well as the density model, but the gained information declined towards zero for later fixations. Both stationary models were outperformed by all dynamic models on all fixations except for the second. For this second fixation only the SceneWalk StartMap model (green curve) generated higher information gain than the static models in experiment 1 and 3. The original SceneWalk model (pink curve) performed less good on the second fixation, but Temporal evolution of the central fixation bias 33 Figure 10. Mean information gain in bit per fixation for the different models for the Fixation 2 to 10. Fixation 1 is not simulated by the models and set to a likelihood of 1. Higher values of information gain mean that empirically measured fixations are more likely in the corresponding model to occur. makes the same predictions from the third fixation on. Finally, the Gaussian model performed equally well as the SceneWalk models in terms of information gain. Thus we confirm earlier findings that the restriction to nearby saccade targets is the most important influence to include into a model. However, as seen in the previous section, this is not sufficient to explain the DTC over time. Across experiments and models, the likelihood declined over time. This was due to the larger variability of later fixation locations between participants. Discussion Model simulations showed that adjusting the SceneWalk model with an initial map can reasonably explain the central fixation bias. While the information gain of the Gaussian model was similar, the SceneWalk StartMap model was the only model that generated an Experiment 1Experiment 2Experiment 3Experiment 4246810246810246810246810012ordinal fixation number within trialinformation gain [bit/fix]ModelDensity ModelGauss ModelCenter BiasSceneWalkSceneWalkStartMap 34 Rothkegel et al. early dip in DTC curves, qualitatively replicated differences in DTC curves between pre- trial fixation times, and replicated saccade latency effects on DTC of the second fixation. In particular for starting positions near the border of the screen, a central activation in the attention map was needed to qualitatively reproduce the initial CFB. However, the initial CFB generated by the SceneWalk StartMap model was too weak. The same was true for the Center Bias model. This suggests that the procedure to estimate the central bias from the second to fourth fixation underestimated the initial CFB. Incorporating a stronger bias might further improve predictions of the SceneWalk StartMap model. Fixations after about 2 s of exploration were distributed as if drawn from the distribution randomly. Later fixations were not directed towards the periphery more due to the initial CFB. Both SceneWalk models approached a similar DTC during the final viewing period, while the simpler Gaussian model stayed too near to the center. General Discussion During scene viewing the eyes have a strong tendency to fixate near the center of an image which potentially masks other bottom-up and top-down effects of saccadic target selection. In a previous study (Rothkegel et al., 2016) with starting positions near the image border and an experimentally delayed first saccade after the onset of an image we observed a considerable reduction of the central fixation bias (CFB; Tatler, 2007). Here, we investigated this reduction in four scene viewing experiments. We manipulated starting positions and the latency of the initial saccade. Contrary to the original scene viewing paradigm, where participants started exploration immediately after image onset, we delayed the initial saccadic response by instructing participants to start exploration Temporal evolution of the central fixation bias 35 only after disappearance of a fixation marker. As a measure of the central fixation bias we computed the distance to center (DTC) of the eyes over time. In all experiments the disappearance of a fixation marker 125 ms after image onset led to an early reduction of the DTC in comparison to trials where the fixation marker disappeared simultaneously to the image onset (original scene viewing paradigm). The reduction of the CFB was particularly pronounced in experiments with pre-trial fixation time as a between subject factor (Exp. 1 & 2). But the reduction of the CFB was even visible when participants started observation at the center of an image (Exp. 4). In a within-subject design with very short pre-trial fixation times (Exp. 3) the manipulation did not significantly reduce the early CFB. A systematic investigation of the disappearance delays of the fixation marker demonstrated that a delay between 125 ms and 250 ms is sufficient for a strong and reliable reduction of the CFB. Shorter delays produced stronger CFBs, whereas longer delays only modestly reduced the CFB but severely changed the time course of the scene viewing paradigm. With long delays the preview time before exploration of an image increased substantially in contrast to the standard scene viewing paradigm. The distance to center of the second fixation was well predicted by the latency of the initial saccade (time from image onset) across experiments. Short saccade latencies led to a strong bias towards the center whereas longer saccade latencies were less systematically directed towards the center. Hence, the latency of the initial response seemed to primarily account for the observed differences of the CFB. Our findings are in agreement with the note communicated earlier that a sudden image onset during scene viewing represents an artificial laboratory situation and may cause unnatural saccadic behavior ('t Hart et al., 2009; Tatler et al., 2011). However, the sudden image onset seems to primarily affect the tendency of the first saccade to move the eyes towards the center of an image. Due to the dependence of fixation locations (Engbert 36 Rothkegel et al. et al., 2015), subsequent fixations are more likely located near the center. Hence, only after about 2 s the initial CFB disappeared and fixations were randomly distributed on an image in our experiments. The reason for this reduction, however, remains unknown. Since our manipulation changed the initial saccade latency and reduced the early CFB, the CFB could be a result of an orienting response due to a luminance change or a result of an initial fixation near the center for fast extraction of the gist (Tatler, 2007). Computational models that aim at predicting the allocation of visual attention on an image are based on the extraction of image features (Itti et al., 1998; Borji & Itti, 2013) and top-down cognitive processes (Navalpakkam, Arbib, & Itti, 2005; Cerf, Harel, Einhauser, & Koch, 2008). These models are evaluated by comparing human fixations to a weighted distribution of different influences (Bylinskii et al., 2015; Borji & Itti, 2013; Le Meur & Baccino, 2013; Borji, Cheng, Jiang, & Li, 2015). Although bottom-up and top-down influences as well as a combination of the two can predict human fixations (Bylinskii et al., 2015), the CFB is a strong predictor that improves goodness-of-fit more than any other single feature (Judd, Ehinger, Durand, & Torralba, 2009; Bylinskii et al., 2015). Therefore, most static visual attention models rely heavily on the implementation of a CFB (Kummerer, Wallis, & Bethge, 2015). The early CFB during scene viewing seems to be an automated, stereotyped response of the saccadic system to a sudden image onset. It masks other bottom-up and top-down factors of saccade target selection and its strength critically depends on the duration of a trial since it primarily affects early fixations. Therefore, a reduction of the CFB during scene viewing, as generated by our paradigm, provides a better understanding of target selection and a more rigorous test of visual attention models than the original scene viewing paradigm. At the minimum, the latency of the first saccade needs to be taken into account, since it strongly influences subsequent viewing behavior. Temporal evolution of the central fixation bias 37 By extending a recently published model of saccade generation (SceneWalk model; En- gbert et al., 2015; Schutt et al., 2017) we were able to account for the empirical data. The model is based on two competing pathways that provide potential saccade targets (attention map) and keep track of recently fixated locations (fixation map). To generate a strong early CFB, we needed to assume that the sudden image onset led to a strong central activation in the attention map. The dynamic model was the only model to qualitatively reproduce the CFB and the relation between saccade latency and the distance to center of the second fixation. However, in its current form the model underestimated both effects. A control model that randomly selected saccade targets from the distribution of empir- ically observed fixation locations (Density model) performed worse than the SceneWalk model but demonstrated that fixations are distributed randomly and proportional to the empirical distribution on the image after about 2 s. Note, a very similar target selection mechanism is often assumed in saliency models where targets are sampled randomly from a saliency map. By incorporating systematic eye movement tendencies these models im- prove (c.f., Le Meur & Liu, 2015). A model similar to the SceneWalk model but without a fixation map (Gauss Model) performed well in terms of likelihood but did not repro- duce the temporal evolution of the CFB or influences of saccade latencies. Finally, a pure central fixation bias model (CenterBias model) performed poorly on all measures. Our results imply to use a modified version of the scene viewing paradigm to study bottom- up and top-down processes of target selection beyond the CFB. To minimize the influence of the sudden image onset, we suggest to use a fixation marker that disappears about 125 ms after image onset. In addition, due to the dependence of successive fixations, scene exploration should not exclusively start near the image center. Instead initial fixations (fixation markers) should be evenly distributed across the entire image or even with a preference towards the periphery. Central parts of the image will be fixated when the 38 Rothkegel et al. eyes move towards the other side of an image. Finally, sudden onsets of stimuli are often used in other laboratory tasks as well (e.g., visual search, face perception). To what extend our results generalize to other domains remains an open question but an early initial CFB might also bias initial fixations in these tasks. Conclusion Delaying the first saccadic response by 125 ms or more relative to image onset reduced the central fixation bias, which is most pronounced during early fixations. The latency of the first saccade after image onset was the main predictor for the distance to image center of the second fixation in all four experiments relatively independent of the time we enforced. Thus our results suggest to use a modified version of the scene viewing paradigm to better understand target selection beyond the central fixation bias. Acknowledgements This work was supported by Deutsche Forschungsgemeinschaft (grants EN 471/13-1 and WI 2103/4-1 to R. E. and F. A. W., resp.). We thank Benjamin Tatler for providing us with the images of natural scenes from his seminal paper about the central fixation bias. References Arizpe, J., Kravitz, D. J., Yovel, G., & Baker, C. I. (2012). Start position strongly influences fixation patterns during face processing: Difficulties with eye movements as a measure of information use. PloS one, 7 (2), e31106. doi: http://dx. doi.org/10.1371/journal.pone.0031106. Temporal evolution of the central fixation bias 39 Baddeley, A., & Turner, R. (2005). SPATSTAT: An R package for analyzing spa- tial point patterns. Journal of Statistical Software, 12 (6), 1–42. Retrieved from http://www.jstatsoft.org/v12/i06/ Bates, D., Machler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67 (1), 1–48. doi: 10.18637/jss.v067.i01 Bickel, P. J., & Doksum, K. A. (2015). Mathematical statistics: basic ideas and selected topics (Vol. 1). Boca Raton: CRC Press. Bindemann, M. (2010). Scene and screen center bias early eye movements in scene viewing. Vision Research, 50 (23), 2577–2587. Bindemann, M., Scheepers, C., Ferguson, H. J., & Burton, A. M. (2010). Face, body, and center of gravity mediate person detection in natural scenes. Journal of Experimental Psychology: Human Perception and Performance, 36 (6), 1477–1485. Borji, A., Cheng, M.-M., Jiang, H., & Li, J. (2015). Salient object detection: A benchmark. IEEE Transactions on Image Processing, 24 (12), 5706–5722. Borji, A., & Itti, L. (2013). State-of-the-art in visual attention modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 (1), 185–207. Botev, Z. I., Grotowski, J. F., & Kroese, D. P. (2010). Kernel density estimation via diffusion. Annals of Statistics, 38 (5), 2916-2957. Brainard, D. H. (1997). The psychophysics toolbox. Spatial Vision, 10 , 433–436. Buswell, G. T. (1935). How people look at pictures. Chicago: University of Chicago Press. Bylinskii, Z., Judd, T., Borji, A., Itti, L., Durand, F., Oliva, A., & Torralba, A. (2015). MIT saliency benchmark. http://saliency.mit.edu/. Castelhano, M. S., & Henderson, J. M. (2008). The influence of color on the perception of scene gist. Journal of Experimental Psychology: Human Perception and Performance, 34 (3), 660–675. Cerf, M., Harel, J., Einhauser, W., & Koch, C. (2008). Predicting human gaze using low- level saliency combined with face detection. In Advances in neural information processing 40 Rothkegel et al. systems (Vol. 20, pp. 241–248). Clarke, A. D., & Tatler, B. W. (2014). Deriving an appropriate baseline for describing fixation behaviour. Vision Research, 102 , 41–51. Cornelissen, F. W., Peters, E. M., & Palmer, J. (2002). The eyelink toolbox: eye tracking with matlab and the psychophysics toolbox. Behavior Research Methods, 34 (4), 613–617. Cousineau, D. (2005). Confidence intervals in within-subject designs: A simpler solution to loftus and masson's method. Tutorials in Quantitative Methods for Psychology, 1 (1), 42–45. Efron, B., & Tibshirani, R. J. (1994). An introduction to the bootstrap. New York: Chapman and Hall. Engbert, R., & Kliegl, R. (2003). Microsaccades uncover the orientation of covert attention. Vision Research, 43 (9), 1035–1045. Engbert, R., & Mergenthaler, K. (2006). Microsaccades are triggered by low retinal image slip. Proceedings of the National Academy of Sciences, 103 (18), 7192–7197. Engbert, R., Trukenbrod, H. A., Barthelm´e, S., & Wichmann, F. A. (2015). Spatial statistics and attentional dynamics in scene viewing. Journal of Vision, 15 (1), 14. doi: 10.1167/15.1.14 Hayhoe, M., & Ballard, D. (2005). Eye movements in natural behavior. Trends in Cognitive Sciences, 9 (4), 188–194. Henderson, J. M., Weeks Jr, P. A., & Hollingworth, A. (1999). The effects of semantic consis- tency on eye movements during complex scene viewing. Journal of Experimental Psychol- ogy: Human Perception and Performance, 25 (1), 210. Ioannidou, F., Hermens, F., Hodgson, T., et al. (2016). The centrial bias in day-to-day viewing. Journal of Eye Movement Research, 9 (6), 1–13. Itti, L., & Koch, C. (2001). Computational modelling of visual attention. Nature Reviews Neuroscience, 2 (3), 194–203. Itti, L., Koch, C., & Niebur, E. (1998). A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20 (11), Temporal evolution of the central fixation bias 41 1254–1259. Judd, T., Ehinger, K., Durand, F., & Torralba, A. (2009). Learning to predict where humans look. In IEEE 12th International Conference on Computer Vision (pp. 2106–2113). Klein, R. (2000). Inhibition of return. Trends in Cognitive Sciences, 4 (4), 138–147. Kleiner, M., Brainard, D., Pelli, D., Ingling, A., Murray, R., Broussard, C., et al. (2007). What's new in psychtoolbox-3. Perception, 36 (14), 1–16. Kummerer, M., Wallis, T. S., & Bethge, M. (2015). Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences, 112 (52), 16054– 16059. Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2013). lmertest: Tests for random and fixed effects for linear mixed effect models (lmer objects of lme4 package). R package version, 2 (6). Le Meur, O., & Baccino, T. (2013). Methods for comparing scanpaths and saliency maps: strengths and weaknesses. Behavior Research Methods, 45 (1), 251–266. Le Meur, O., Le Callet, P., Barba, D., & Thoreau, D. (2006). A coherent computational approach to model bottom-up visual attention. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28 (5), 802–817. Le Meur, O., & Liu, Z. (2015). Saccadic model of eye movements for free-viewing condition. Vision Research, 116 , 152-164. doi: 10.1016/j. visres.2014.12.026 Loftus, G. R., & Mackworth, N. H. (1978). Cognitive determinants of fixation location during picture viewing. Journal of Experimental Psychology: Human Perception and Perfor- mance, 4 (4), 565–572. Luce, R. D. (1959). Individual Choice Behavior: A Theoretical Analysis. New York: Wiley. MATLAB. (2015). version 8.6.0 (r2015b). Natick, Massachusetts: The MathWorks Inc. Najemnik, J., & Geisler, W. S. (2005). Optimal eye movement strategies in visual search. Nature, 434 (7031), 387–391. Navalpakkam, V., Arbib, M., & Itti, L. (2005). Attention and scene understanding. In Neuro- 42 Rothkegel et al. biology of attention (pp. 197–203). San Diego, CA: Elsevier. Pelli, D. G. (1997). The videotoolbox software for visual psychophysics: Transforming numbers into movies. Spatial Vision, 10 (4), 437–442. R Core Team. (2014). R: A language and environment for statistical computing [Computer software manual]. Vienna, Austria. Retrieved from http://www.R-project.org/ Rothkegel, L. O. M., Trukenbrod, H. A., Schutt, H. H., Wichmann, F. A., & Engbert, R. (2016). Influence of initial fixation position in scene viewing. Vision Research, 129 , 33–49. Schutt, H. H., Rothkegel, L. O. M., Trukenbrod, H. A., Reich, S., Wichmann, F. A., & Engbert, R. (2017). Likelihood-based parameter estimation and comparison of dynamical cognitive models. Psychological Review , Advance online publication. doi: 10.1037/rev0000068 Tatler, B. W. (2007). The central fixation bias in scene viewing: Selecting an optimal viewing position independently of motor biases and image feature distributions. Journal of Vision, 7 (14), 1–17. doi: 10.1167/7.14.4 Tatler, B. W., Hayhoe, M. M., Land, M. F., & Ballard, D. H. (2011). Eye guidance in natural vision: Reinterpreting salience. Journal of Vision, 11 (5), 1–23. Tatler, B. W., & Vincent, B. T. (2008). Systematic tendencies in scene viewing. Journal of Eye Movement Research, 2 (2), 1–18. Tatler, B. W., & Vincent, B. T. (2009). The prominence of behavioural biases in eye guidance. Visual Cognition, 17 (6-7), 1029–1054. 't Hart, B. M., Vockeroth, J., Schumann, F., Bartl, K., Schneider, E., Koenig, P., & Einhauser, W. (2009). Gaze allocation in natural stimuli: comparing free exploration to head-fixed viewing conditions. Visual Cognition, 17 , 1132–1158. Torralba, A. (2003). Modeling global scene factors in attention. Journal of the Optical Society of America, 20 (7), 1407–1418. Tseng, P.-H., Carmi, R., Cameron, I. G., Munoz, D. P., & Itti, L. (2009). Quantifying center bias of observers in free viewing of dynamic natural scenes. Journal of Vision, 9 (7), 1–16. Venables, W. N., & Ripley, B. D. (2002). Modern Applied Statistics with S (Fourth ed.). New Temporal evolution of the central fixation bias 43 York: Springer. Vitu, F., Kapoula, Z., Lancelin, D., & Lavigne, F. (2004). Eye movements in reading isolated words: Evidence for strong biases towards the center of the screen. Vision Research, 44 (3), 321–338. Wichmann, F. A., Drewes, J., Rosas, P., & Gegenfurtner, K. R. (2010). Animal detection in natural scenes: critical features revisited. Journal of Vision, 10 (4), 1–27. Wickham, H. (2009). ggplot2: Elegant graphics for data analysis. New York: Springer. Retrieved from http://ggplot2.org Yarbus, A. L., Haigh, B., & Rigss, L. A. (1967). Eye Movements and Vision (Vol. 2). New York: Plenum Press.
1605.04033
2
1605
2016-10-24T14:15:20
Structural Pathways Supporting Swift Acquisition of New Visuo-Motor Skills
[ "q-bio.NC" ]
Human skill learning requires fine-scale coordination of distributed networks of brain regions that are directly linked to one another by white matter tracts to allow for effective information transmission. Yet how individual differences in these anatomical pathways may impact individual differences in learning remains far from understood. Here, we test the hypothesis that individual differences in the organization of structural networks supporting task performance predict individual differences in the rate at which humans learn a visuo-motor skill. Over the course of 6 weeks, twenty-two healthy adult subjects practiced a discrete sequence production task, where they learned a sequence of finger movements based on discrete visual cues. We collected structural imaging data during four MRI scanning sessions spaced approximately two weeks apart, and using deterministic tractography, structural networks were generated for each participant to identify streamlines that connect cortical and sub-cortical brain regions. We observed that increased white matter connectivity linking early visual (but not motor) regions was associated with a faster learning rate. Moreover, we observed that the strength of multi-edge paths between motor and visual modules was also correlated with learning rate, supporting the role of polysynaptic connections in successful skill acquisition. Our results demonstrate that the combination of diffusion imaging and tractography-based connectivity can be used to predict future individual differences in learning capacity, particularly when combined with methods from network science and graph theory.
q-bio.NC
q-bio
Structural Pathways Supporting Swift Acquisition of New Visuo-Motor Skills 6 1 0 2 t c O 4 2 ] . C N o i b - q [ 2 v 3 3 0 4 0 . 5 0 6 1 : v i X r a Ari E. Kahn1,2,3, Marcelo G. Mattar4, Jean M. Vettel3,5,2, Nicholas F. Wymbs6, Scott T. Grafton5, and Danielle S. Bassett2,7,8 1Department of Neuroscience, University of Pennsylvania, Philadelphia, PA 19104 USA 2Department of Bioengineering, University of Pennsylvania, Philadelphia, PA 19104 USA 3Human Research and Engineering Directorate, U.S. Army Research Laboratory, Aberdeen, MD 21001 USA 4Department of Psychology, University of Pennsylvania, Philadelphia, PA 19104 USA 6Department of Physical Medicine and Rehabilitation, Johns Hopkins University, Baltimore, MD 21218 USA 5Department of Psychological and Brain Sciences, University of California, Santa Barbara, CA 93106 USA 7Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104 USA 8To whom correspondence should be addressed: [email protected]. ABSTRACT 1 Human skill learning requires fine-scale coordination of distributed networks of brain regions linked by white matter tracts to allow for effective information transmission. Yet how individual differences in these anatomical pathways may impact individual differences in learning remains far from understood. Here, we test the hypothesis that individual differences in structural organization of networks supporting task performance predict individual differences in the rate at which humans learn a visuo-motor skill. Over the course of 6 weeks, twenty healthy adult subjects practiced a discrete sequence production task, learning a sequence of finger movements based on discrete visual cues. We collected structural imaging data, and using deterministic tractography generated structural networks for each participant to identify streamlines connecting cortical and sub-cortical brain regions. We observed that increased white matter connectivity linking early visual regions was associated with a faster learning rate. Moreover, the strength of multi-edge paths between motor and visual modules was also correlated with learning rate, supporting the potential role of extended sets of polysynaptic connections in successful skill acquisition. Our results demonstrate that estimates of anatomical connectivity from white matter microstructure can be used to predict future individual differences in the capacity to learn a new motor-visual skill, and that these predictions are supported both by direct connectivity in visual cortex and indirect connectivity between visual cortex and motor cortex. Keywords: diffusion MRI, motor sequence learning, graph theory, polysynaptic networks, discrete sequence production Introduction Human skill learning is a complex phenomenon that involves the fine-scale coordination of disparate cortical and subcortical regions (Dayan and Cohen, 2011). This coordination critically depends on the effective transmission of information across white matter tracts, which link distant brain regions in cortico-cortical networks and cortico-subcortical loops (Lynch and Tian, 2006). Lesions or injuries to these interconnected tracts – particularly in motor and visual systems – can directly cause deficits in skill learning (Ding et al., 2001). The exact extent of these deficits is difficult to predict, largely due to the fact that white matter tracts form a complex interconnected network (Sporns et al., 2005). Damage to this network can have broadly distributed repercussions on processing, causing loss of information transmission (Scantlebury et al., 2014), or detrimental alterations in transmission patterns (Crofts et al., 2011). The interconnected nature of white matter tracts not only complicates response to injury, but it also forms a fundamental substrate for individual differences in brain anatomy that may have non-trivial effects on cognition and behavior. White matter connectivity displays large-scale differences across individuals (Bassett et al., 2011), being modulated by age (Betzel et al., 2014), gender (Tunc et al., 2016; Ingalhalikar et al., 2014), genetics (Hong et al., 2015), and prior experience (Sampaio-Baptista et al., 2013; Scholz et al., 2009). How these individual differences may account for individual differences in skill learning is not fully understood. Gaining such an understanding could directly inform therapeutic interventions to enhance recovery of motor skills after brain injury (Tomassini et al., 2011), and furthermore could potentially inform training paradigms to enhance motor-visual expertise in healthy individuals (Neumann et al., 2016). Here, we examine if connectivity networks defined by diffusion MRI are predictive of individual differences in the rate at which subjects acquire a simple visuo-motor task (Wymbs and Grafton, 2015). In a discrete sequence production (DSP) task, subjects perform a sequence of finger movements based on visual cues (Rhodes et al., 2004). Once the correct key for each movement is pressed, the visual cue for the next sequence element is presented without delay. Consequently, a DSP task allows subjects to develop exceptionally fast, contiguous movements, much like an expert pianist performing a keyboard arpeggio. Efficient acquisition of this specific visuo-motor skill requires a gradual autonomy of visual and motor functional subnetworks (Bassett et al., 2015, 2013) (Fig. 1A–B). Initially, a person relies on the visual cue to perform a finger movement, an action that requires integration between motor and visual cortices; however, once a sequence becomes overlearned, a subject has mastered direct motor-motor associations where a given finger movement is the cue for the next finger movement. These functional network changes may depend on underlying structure, shown to be a fundamental driver of brain dynamics 3/32 at rest (Becker et al., 2015; Goni et al., 2014; Honey et al., 2009) and during task performance (Hermundstad et al., 2013, 2014; Jarbo and Verstynen, 2015; Smith et al., 2009; Osher et al., 2016). Furthermore, individual variability in behavior has been linked to differences in structural networks (Johansen-Berg, 2010), and IQ and motor speed have been associated with greater white matter connectivity (Li et al., 2009) and fractional anisotropy (FA) (Hirsiger et al., 2016). Prior work in word learning tasks also suggests that increased myelination, axonal diameter, and fractional anisotropy in tracts implicated in task processing are associated with better performance (Lopez-Barroso et al., 2013; Wong et al., 2011). Building on these prior studies, we hypothesized that individuals with greater structural connectivity in motor and visual cortices (and particularly in primary motor and visual cortices) would show faster learning rates than individuals with less connectivity. We also set out to test whether these structural differences remained constant over the 6 weeks of practice (Le Bihan and Johansen-Berg, 2012) or changed appreciably with training (Scholz et al., 2009; Blumenfeld-Katzir et al., 2011; Taubert et al., 2012). Finally, due to the prevalence of physically extended sets of polysynaptic connections in the visual-motor system, we hypothesized that individual differences in long distance walks on the graph of structural connections between visual and motor cortex would correspond to individual differences in learning rate. To address these hypotheses, we examined diffusion tensor imaging data acquired from 20 healthy young adult subjects over the course of 6 weeks of training on the DSP task (Bassett et al., 2013, 2014, 2015; Wymbs and Grafton, 2015). Subjects were scanned in 4 separate sessions, including a scan on day 1 before training began and then a scan approximately every 2 weeks. Between scanning sessions, subjects practiced a set of ten-element sequences at home using a program installed on their laptop computers, and behavioral performance was assessed by calculating the movement time (MT) for each sequence defined as the duration between the first button press and the last button press in the sequence. The learning rate for each participant was computed as the first exponential drop-off parameter in a double-exponential fit of the MT as a function of trials practiced across the entire 6 weeks of training. To compare individual differences in learning rate to the organization of white matter connectivity, we generated structural networks from the 4 diffusion tensor imaging scans using a deterministic fiber tracking algorithm (Fig. 1C), which provided estimates of the number of streamlines connecting pairs of cortical and subcortical regions derived from brain atlases (Fig. 1D). We observe three main results: individual differences in learning rate are significantly correlated with white matter connectivity in visual (but not motor) cortex, these relationships are consistent across the 6 weeks of task practice, and individuals with faster learning rates also show greater walk strength linking motor and visual cortices, a measure suggesting increased strength of polysynaptic pathways. Figure 1 4/32 Materials and Methods Participants and Experimental Design: Participants: Twenty-two right-handed participants (13 females and 9 males; mean age, 24 years) volunteered and provided informed consent in writing in accordance with the guidelines of the Institutional Review Board of the University of California, Santa Barbara. All had normal or corrected vision and no history of neurological disease or psychiatric disorders. We excluded two participants because one participant failed to complete the experiment and the other had excessive head motion (persistent head motion greater than 5mm during the MRI scanning). We also had technical problems for two participants and were unable to collect DTI data during the pretraining session for Scan 1. Finally, for one additional subject, Scan 1 was removed due to the total of estimated streamlines differing by more than 3 standard deviations from the subject mean. Therefore, the structural analysis includes 17 participants for Scan 1 and 20 participants for Scans 2–4. Experimental setup and procedure: The DSP training protocol occurred over a six week period with four MRI scanning sessions spaced two weeks apart on Day 1, Day 14, Day 28, and Day 42 (Fig. 2A). On Day 1 of the experimental protocol, the participants completed their first MRI session, Scan 1, and the experimenter (N.F.W.) installed the training module on the participant's personal laptop and taught them how to use it for at-home training sessions. Participants were required to do the training for a minimum of 10 out of the 14 days in each 2-week period between the subsequent scanning sessions for Scans 2-4. All participants completed the full expected training regimen; none completed less than 10 full training sessions. In their at-home training sessions, participants practiced a set of 10-element sequences using their right hand in a discrete sequence production (DSP) task (Mattar et al., 2015; Bassett et al., 2015, 2013, 2014; Wymbs and Grafton, 2015). Sequences were presented using a horizontal array of five square stimuli, and the key responses were mapped from left to right, such that the thumb corresponded to the leftmost stimulus and the pinky finger corresponded to the rightmost stimulus (Fig. 2B). A square highlighted in red served as the imperative stimulus, and the next square in the sequence was highlighted immediately after each correct key press. If an incorrect key was pressed, the sequence was paused at the error and restarted upon the appropriate key press. Participants had an unlimited amount of time to respond and complete each trial. Figure 2 Each practice trial began with the presentation of a sequence-identity cue that identified one of six sequences. These six sequences were presented with three different levels of exposure, in order to acquire data over a larger range of learning stages while controlling for the effect of scanning day (Table 1). The two extensively trained (EXT) sequences were identified with 5/32 a colored circle (cyan for sequence A and magenta for B), and they were each practiced for 64 trials during every at-home training session. The two moderately trained (MOD) sequences were identified by triangles (red for sequence C and green for D) and each practiced for 10 trials in every session. The two minimally trained (MIN) sequences were identified by black outlined stars (filled with orange for sequence E and white for F) and only practiced for 1 trial each during the at-home training sessions. Participants were given feedback every ten trials that reported the number of error-free sequences and the mean time required to complete them. During each of the four MRI scanning sessions, we collected functional Echo Planar Imaging EPI data and structural imaging data from MPRAGE and DTI scans. In the functional runs, participants performed 300 trials of the self-paced DSP task using the same block structure with feedback as the at-home practice sessions, but the sequences were presented equally for a total of 50 trials for each of the six trained sequences. We have previously reported results from functional analyses (Mattar et al., 2015; Bassett et al., 2015, 2013, 2014; Wymbs and Grafton, 2015). In this paper, we analyze the structural data and examine individual variability in structural connections among the distributed motor and visual regions of interest that were derived directly from the functional neuroimaging studies of this same dataset (Bassett et al., 2015). In this previous work, a set of motor and visual regions that formed functional modules were identified in a data-driven fashion whose task-based modulation tracked the effects of training. Here we build on the identification of these regions of interest by studying their structural connectivity derived from diffusion imaging. Table 1 Estimating Learning Rates for Individual Participants For each sequence, we defined the MT as the duration between the time of the first button press and the time of the last button press. For the set of sequences of a single type (i.e., sequence A,B,C,D,E, and F), we estimated the learning rate by fitting a double exponential function to the MT data (Schmidt and Lee, 2005),(Rosenbaum, 2014) using a robust outlier correction in MATLAB (using the function "fit.m" in the Curve Fitting Toolbox with option "Robust" and type "Lar"): MT = D1e−tκ + D2e−tλ , where t is time, κ is the exponential drop-off parameter (which we called the learning rate) used to describe the fast rate of improvement, λ is the exponential drop-off parameter used to describe the slow, sustained rate of improvement, and D1 and D2 are real and positive constants. The magnitude of κ indicates the steepness of the learning slope: individuals with larger κ values have a steeper drop-off in MT, suggesting that they are faster learners (Dayan and Cohen, 2011; Yarrow et al., 2009). The decrease in MT has been used to quantify learning for several decades (Snoddy, 1926; Crossman, 2010). Several functional forms have been suggested for the fit of MT (Newell and Rosenbloom, 1993; Heathcote et al., 2000), 6/32 and variants of an exponential are viewed as the most statistically robust choices (Heathcote et al., 2000). In addition, the fitting approach that we used has the advantage of estimating the rate of learning independent of initial performance or performance ceiling. For the purpose of measuring effects on learning rate, we used the average value of κ for the two extensively-trained sequences, for which we had the greatest number of trials practiced (Table 1). While we do not have explicit information on the computing power of each subject's laptop, the learning rate that we study is independent of the starting MT, the ending MT, and the mean MT. Instead, it is a measure of the rate of change in MT. Thus, any differences in computing power cannot be used to explain the results. Moreover, we should mention that error rates on this task are on the order of 1× 10−3 (Bassett et al., 2015), and error rates are not significantly correlated with learning rates (r = 0.34, p = 0.13) (Bassett et al., 2015). Neuroanatomical Data and Associated Methods In this section, we briefly describe the neuroanatomical data acquired from participants, as well as computational methods associated with data preprocessing, structural network construction, and statistical analyses. Data Acquisition All scans were acquired on a 3T Siemens TIM Trio scanner with a 12-channel phased-array head coil at the University of California, Santa Barbara. Each data acquisition session included both a diffusion tensor imaging (DTI) scan as well as a high-resolution T1-weighted anatomical scan. The structural scan was conducted with an echo planar diffusion weighted technique acquired with iPAT using an acceleration factor of 2. The diffusion scan was 30-directional with a b-value of 1000s/mm2 and TE/TR = 94/8400 ms, in addition to two b0 images. Matrix size was 128×128 with a slice number of 60. Field of view was 230×230mm2 and slice thickness was 2mm. Acquisition time per DTI scan was 9:09min. The anatomical scan was a high-resolution three-dimensional T1-weighted sagittal whole-brain image using a magnetization prepared rapid acquisition gradient-echo (MPRAGE) sequence. It was acquired with TR = 2300 ms; TE=2.98 ms; flip angle = 9 degrees; 160 slices; 1.10mm thickness. DTI Preprocessing DTI is both highly sensitive to subject movement (Yendiki et al., 2013) and susceptible to directional eddy currents, which can cause distortions in the brain volume (Jezzard et al., 1998). To address these issues, we performed the following data preprocessing using the FMRIB Software Library (FSL v5.0.8) (Smith et al., 2004; Jenkinson et al., 2012). First, individual subject masks of the brain were created with the Brain Extraction Tool (BET) (Smith, 2002) for use in later registration and 7/32 correction tools, which require an accurate estimation of the spatial extent of the brain. We applied the EDDY correction tool (Andersson and Sotiropoulos, 2016) which simultaneously models both motion effects and eddy current distortions, and corrects them relative to a b = 0 image collected at the beginning of the scan. Next, subject scans were transformed into a common space to compare regional connectivity between subjects. Using FMRIB's Linear Image Registration Tool (FLIRT) (Jenkinson and Smith, 2001; Jenkinson et al., 2002), scans were registered to the anatomical T1 image, and then the anatomical scan was in turn registered to the Montreal Neurological Institute (MNI) space MNI152 template using FMRIB's Nonlinear Image Registration Tool (FNIRT). Motion correction also impacts the effective b-matrix directions since the rotated images are no longer aligned with the scanner; therefore, we used the output of EDDY to rotate the b-vectors to match the changes induced by the motion correction procedure (Leemans and Jones, 2009). Using DSI-Studio (http://dsi-studio.labsolver.org), orientation density functions (ODFs) within each voxel were recon- structed from the corrected scans in native diffusion space in order to minimize sampling distortions (Cieslak and Grafton, 2014). We then used the reconstructed ODFs to perform a whole-brain deterministic tractography using DSI-Studio (Yeh et al., 2013). We generated 1,000,000 streamlines per subject, with a maximum turning angle of 35 degrees(Bassett et al., 2011) and a maximum length of 500mm(Cieslak and Grafton, 2014). By holding the number of streamlines between participants constant, we use the number of streamlines that connect brain region pairs as an estimate of the strength of the connection and examine individual variability in structural connectivity(Griffa et al., 2013). Table 2 Network Construction To examine the relationship between structural connectivity and individual differences in learning rate, we constructed networks for each subject where nodes are atlas regions and edges are the measured connection strength between region pairs (Hagmann et al., 2008). The nodes of the network were derived from spatially-defined regions of a brain atlas, and we utilized two complementary atlas parcellations to confirm that our results are not specific to the particular regional boundaries chosen by one atlas. First, we used the Harvard-Oxford atlas to allow for direct comparison to functional network studies of this same task (Bassett et al., 2013, 2014, 2015), and we combined the Harvard-Oxford cortical and subcortical atlases into a single 111-region atlas by giving cortical labels precedence whenever a single voxel was assigned to both a cortical and subcortical region. In our intra- vs interhemisphere analysis, we exclude the brainstem region from this atlas since it crosses the midline. As a complementary parcellation, we chose the anatomically-defined AAL atlas, originally developed in Statistical Parametric Mapping (SPM) 8/32 (Tzourio-Mazoyer et al., 2002), which divides each brain hemisphere into 45 regions. For both atlases, we used a version in MNI-space that was then warped into subject-specific native space using FNIRT. Across both atlases, the edges of the network were derived from streamlines that started and ended between the region pair and excluded streamlines that passed through one or both of the regions. Weighted connectivity matrices were then generated from the atlases and DTI reconstructions such that the matrix W contained elements Wi j whose values were equal to the number of streamlines with end-to-end connectivity between regions i and j. All diagonal elements in the matrix were set to 0 to eliminate self-connections. To correct for differing region sizes, each matrix element was divided by the sum of the volumes of regions i and j (Hagmann et al., 2008). That is, Bi j = Wi j vis+v js where vis is the number of voxels in region i for subject s. The resultant connectivity matrix for each subject and scan was then normalized to give a connection strength A such that Ai j = Bi j ∑i, j Bi j , ensuring that all scans had identical total connection strength. Network Statistics Based on the functional analysis of this dataset (Bassett et al., 2015), we examined whether individual variability in structural connectivity among distributed regions of the motor and visual systems was correlated with learning rate (Mattar et al., 2015). For both of these systems, we calculated the mean connection strength within the system by averaging the weights of all edges connecting pairs of nodes within the system (see Table 2). We report our results both at the single-scan level as well as an average over the 4 scans of each subject. Results are consistent in the two cases. To analyze the impact of indirect connectivity between motor and visual regions, we computed walk strength, a measure of the connection strength between two regions that accounts for indirect paths of varying walk lengths. Here, a walk is defined as a path from one point in the graph to another that may pass along the same edge more than once (Fig. 6A). Given a graph G and its adjacency matrix A, An provides the connection strength between all pairs of nodes when examining walks of length n (Estrada and Hatano, 2007). For instance, streamlines directly connecting primary visual cortex to primary motor cortex would be a walk of length 1, whereas the combination of streamlines connecting primary motor cortex first to thalamus and then to primary visual cortex would be a walk of length 2. Note that the term "length" is used in a topological sense, where walks with more steps are considered to have longer length (Crofts and Higham, 2009). We base our analysis on a similar metric, communicability, which is defined such that walks of all lengths contribute to network communication, but longer walks increasingly contribute less. For an unweighted graph, the network communicability is simply given as ∑∞ n=1 An n! (Estrada and Hatano, 2007). In a weighted graph, an additional normalization is needed to prevent highly connected nodes 9/32 from unduly dominating the estimate (Crofts and Higham, 2009). A typical solution is to divide all weights Ai j by(cid:112)did j where di is the degree of node i, given by di = ∑∞ k=1 Aik(Higham et al., 2007). While communicability provides a single metric of communication between nodes, it does not provide information on the contributions of specific walk lengths. To address this limitation, we define the walk strength as the normalized strength of walks of length n, which is given as Sn where S = D− 1 2 and D = diag(di), or the matrix whose diagonal is given by the values di.(Crofts and Higham, 2008). 2 AD− 1 Statistical Testing Analysis was performed in Python using a collection of freely available packages: Numpy/Scipy, Pandas, stastmodels and Jupyter. Correlations reported throughout the paper are Pearson correlations at an α level of 0.05. Data was corrected for multiple comparisons using Bonferroni, False Discovery Rate (Benjamini and Hochberg, 1995), and the form p < 0.05 n , where n is the number of comparisons. Results: Visual Streamline Connectivity Correlates with Learning Figure 3 Our general aim was to uncover the structural network correlates of individual differences in learning rate for a common visuo-motor task (Wymbs and Grafton, 2015). Because direct connections between motor and visual cortices are not present at this large scale, we separately consider connectivity within motor areas and within visual areas previously identified in a functional analysis of this dataset (Mattar et al., 2015; Bassett et al., 2015) (see Table 2 and Fig. 1A–B). We explicitly test the fundamental hypothesis that individuals with greater mean structural connectivity in motor and visual cortices would show faster learning rates (κ; see Methods) than individuals with less connectivity. We observed a highly significant correlation between visual-visual streamlines and learning rate across all subjects (Pearson correlation coefficient r = 0.50, with corresponding one-tailed p-value of p = 0.0125, significant after Bonferroni correction; Fig. 3A). In contrast, we observed no significant correlation between motor-motor streamlines and learning rate (r = 0.07, p = 0.389; Fig. 3B). Within the subset of connections linking visual regions with one another, we expected that connection strength within a given hemisphere would be particularly relevant given that interhemispheric transfer of information is not as relevant in this task as it is in other tasks manipulating perceptual reference frames (Bernier and Grafton, 2010) or narrow visual fields (Doron et al., 2012). Consistent with our hypothesis, we found that the observed correlation between the visual-to-visual connection strength and learning rate was largely driven by intrahemispheric streamlines (Pearson correlation coefficient r = 0.68, p = 0.0005; 10/32 Fig. 3C), while no significant correlation was observed among interhemispheric connections (r = 0.01, p = 0.512; Fig. 3D). We verified these same relationships in the AAL atlas (Fig. 3E–H). The structural connection strength among visual-visual region pairs accounts for individual variability in learning rate. It is again more pronounced in both overall visual (r = 0.44, p = 0.027) and intrahemispheric visual-visual connectivity (r = 0.62, p = 0.0002), while no significant correlation is observed either in motor-motor connectivity (r = 0.03, p = 0.449) or in interhemispheric visual connections (r = 0.10, p = 0.666). Reliability of Connectivity-Based Predictors of Learning: Figure 4 Next, we asked whether individual differences in the white matter connections that predicted learning rate would remain constant (Le Bihan and Johansen-Berg, 2012) or change appreciably (Scholz et al., 2009; Blumenfeld-Katzir et al., 2011; Taubert et al., 2012) over 6 weeks of practice. We performed the same analysis as before but individually applied to each scan, restricting ourselves to the set of visual intrahemispheric connections (Fig. 4A–D). Across the four scan sessions, we observed a positive relationship between the visual-to-visual connection strength and learning rate: the p-values for scans 2, 3, and 4 all pass a Bonferroni correction for n = 4 tests and scan 1 was close to significant at p = 0.05. Pearson correlation coefficients and corresponding p-values for scan 1 were r = 0.41, p = 0.050, for scan 2 were r = 0.72, p = 0.0002, for scan 3 were r = 0.61, p = 0.002 , and for scan 4 were r = 0.71, p = 0.0003. These results suggest that the connectivity-learning relationship remained constant over 6 weeks of practice. In addition to being robust across scanning sessions, the connectivity-learning relationship is also robustly observed when we segregated the brain into 90 (rather than 111) regions using a separate atlas. Specifically, using the automated anatomical labeling (AAL) atlas, we observed a significant correlation between learning rate and intrahemispheric visual connection strength across all four scan sessions after Bonferroni correction for n = 4 tests (Fig. 4E–H). Pearson correlation coefficients and corresponding p-values for scan 1 were r = 0.51, p = 0.011, for scan 2 were r = 0.62, p = 0.002, for scan 3 were r = 0.61, p = 0.002, and for scan 4 were r = 0.54, p = 0.003. The scan-independent relationship between learning rate and visual-to-visual connectivity suggests the possibility that visual-to-visual connectivity itself is consistent across the 6 weeks of training, consistent with previous reports in other learning contexts (Le Bihan and Johansen-Berg, 2012). To directly assess the reliability of visual-to-visual connectivity, we performed two separate analyses: one at the level of white matter streamlines and the second at the level of fractional anisotropy across voxels. First, we computed the intraclass correlation coefficient (ICC)(Shrout and Fleiss, 1979) to assess the reliability of visual 11/32 connectivity across scanning sessions across the subset of subjects present for all four scans (n = 17). Using a two-way ANOVA on visual-to-visual connection strength, we found no main effect of scanning session (F(3,48) = 1.35, p = 0.27). Furthermore, the ICC is extremely high (ICC(1,1) = 0.83), which indicates the high reliability of visual-to-visual connection strength across scanning sessions. Additionally, we performed voxel level univariate analyses to test for reliability of fractional anisotropy across the whole brain over the 6 weeks of learning. A repeated measures ANOVA was calculated across the four DTI scan sessions. An f -omnibus test demonstrated no significant effects (p > 0.05, FDR corrected). In addition, a paired t-test between scans 1 and 4 was performed. There were no significant differences of FA values (p > 0.05, FDR corrected). These results support the conclusion that white matter microstructure remains consistent over the 4 scans, supporting the observed inter-scan reliability of our results. Anatomical Specificity of Connectivity-Learning Relationship Figure 5 To better understand the relationship between intrahemispheric visual connectivity and variability in learning rate κ, we examined which visual region pairs were driving this effect. This examination had the added benefit of assessing whether different connections predicted behavior differently: although a positive trend was expected given the results in Fig. 3, it is possible that a few smaller regions might show the opposite relationship. To address these questions, for each visual region pair we calculated the Pearson correlation coefficient between the subject learning rate and the mean connection strength between the two regions across scans (see Fig. 5A). We found significant correlations (uncorrected) between learning rate and individual differences in the connections between five pairs of visual regions: right intracalcarine and right cuneal cortex (r = 0.64, p = 0.0012), right cuneal cortex and right occipital pole (r = 0.42, p = 0.032), left intracalcarine and left cuneal cortex (r = 0.56, p = 0.005), left supracalcarine and left occipital cortex (r = 0.38, p = 0.049), and left supracalcarine and left lingual gyrus (r = 0.4, p = 0.039). Only the first of these relationships passed FDR correction for multiple comparisons (Fig. 5B). Role of Indirect Connectivity in Learning Prediction Our results have revealed structural correlates in direct connections within visual and motor regions; however, prior fMRI studies have linked changes in learning rate to functional connectivity between motor and visual areas (Mattar et al., 2015; Bassett et al., 2015). To examine structural predictors between regions, we turn to recently developed mathematical techniques in the domain of network science that allow us to directly examine the effects of indirect connectivity (an estimate of polysynaptic 12/32 transmission potential across extended physical distances) in brain networks. Specifically, we compute variable walk lengths between any two nodes in a network. Direct connections are a walk length of 1, while connections that pass through one intermediary region have a walk length of 2; connections that pass through two intermediary regions have a walk length of 3, and so on (Fig. 6A). We hypothesized that as we examined sufficiently long walk lengths, the connectivity between motor and visual regions would become increasingly correlated with individual differences in learning rate. Figure 6 Our results confirm this hypothesis, demonstrating that the length-specific connectivity between motor and visual regions was increasingly correlated with individual differences in learning rate as walk length increased. As shown in Fig. 6, at walks of length 15, individual differences in walk strength between motor and visual regions were significantly correlated with individual differences in learning rate (Pearson correlation coefficient r = 0.39, p = 0.004). As walk length continued to increase, the correlation approached an asymptote which can be observed at n = 40 with r = 0.56, p = 0.005. We confirmed these assessments of statistical significance using a non-parametric null model wherein we shuffled node assignments to "visual" or "motor" sets, thereby choosing a random set of pseudo visual-motor edges. We examined the correlation at walks of n = 40 on repeated null model samples, and constructed a 95% threshold for the correlation coefficient from the null distribution. We observed that walks of length n = 18 and beyond all exceeded this threshold, and at n = 40 our data was significant compared to the null model at p = 0.008. These results indicate the importance of indirect connections between motor and visual cortices in facilitating the learning of a visuo-motor task. Discussion In this study, we assess whether individual differences in structural connectivity can account for individual differences in learning a visuo-motor task. Participants practiced a set of ten-element sequences over a six-week period, and we collected structural imaging data during four MRI scanning sessions spaced two weeks apart. We mapped structural connectivity between brain regions in large networks of interest in motor and visual systems, identified by prior assessments of functional neuroimaging data during task performance (Bassett et al., 2015). We observed a significant correlation between visual (but not motor) structural connectivity and learning rate across participants, and this relationship was consistent across the 4 scanning sessions. Interestingly, this correlation was strongest in direct connections among visual regions within the same hemisphere. However, an assessment of network walk strength also revealed a significant correlation between the strength of indirect connections between motor and visual cortices and individual differences in learning rate, suggesting the potential importance 13/32 of physically extended polysynaptic information transmission for skill acquisition. The Relationship Between White Matter Microstructure and Human Behavior. Our primary hypothesis posited that individual variability in white matter microstructure connecting task-relevant regions would account for individual differences in skill acquisition. Consistent with our hypothesis, we observed a significant relationship between intrahemispheric connections among visual region pairs and variability in learning rate on a discrete sequence production task practiced over the course of 6 weeks. Previous studies have offered preliminary evidence to suggest that structural differences in specific brain regions (although not networks) correlate with individual differences in skill learning (Tomassini et al., 2011; Tuch et al., 2005). Our work extends these previous studies by demonstrating that the degree of connectivity within visual regions is correlated with individual differences in learning rate on a simple motor-visual task. While previous studies have focused on regional or tract-specific changes in fractional anisotropy (FA) in white matter, we demonstrate that tractography-based approaches capture individual differences in white matter that directly support skill acquisition. Of note, we do not observe longitudinal changes of fractional anisotropy in our study population over the course of training, suggesting that our diffusion measures of connectivity are remarkably stationary. Not surprisingly then, we found a remarkably consistent relationship between individual differences of connectivity and learning rate across all four DTI scanning sessions. While both tractography and FA-based approaches can reveal important structural differences, a tractography-based approach allows us to leverage network-based tools to understand brain- and system-wide dynamics. Although we expected structural variability in both visual and motor systems would correlate with individual variability in learning rate, we only found a significant relationship with intrahemispheric connections among visual regions. We speculate that this may be due to the nature of the motor task itself. The participants learned to quickly press one of five buttons following a visual cue. This specific action (a button press) is not a particularly novel movement for these participants, all of whom have already developed a wide variety of dexterous skills such as typing. Due to the ubiquity of this action over the course of development, the structural connectivity within motor cortex may already be at a ceiling, obscuring any correlation with learning. Alternatively, it is possible that changes in motor connectivity with learning may only be measured at smaller spatial scales. In contrast to the simple button press, the more challenging skill that the subjects mastered was the spatial mapping between visual stimuli and motor commands. It is intuitively plausible that the ability to learn this mapping efficiently is fundamentally dependent on visual resources for detailed encoding of spatial information. Indeed, a wide range of visuo-motor tasks have demonstrated strong reliance on occipital areas in mapping arbitrary stimuli with specific motor responses as well as sequences of responses (Grafton et al., 1994, 1995; Diedrichsen and Kornysheva, 2015; Wiestler et al., 2014). 14/32 Structure is Consistent Across Scanning Sessions. Our results demonstrated a relationship between individual variability in learning rate and connection strength between visual regions that was consistent across the four scanning sessions. This consistency is particularly interesting in light of prior work showing changes in brain network connectivity as a function of experience-dependent plasticity (Lindenberger et al., 2006; Pascual-Leone et al., 2005). Indeed, researchers actively debate the time scales at which these structural changes occur (May, 2011; Holtmaat and Svoboda, 2009; Keller and Just, 2015) and whether these changes can be detected using current diffusion weighted imaging techniques (Thomas and Baker, 2013; Lovd´en et al., 2013). Some of the most well-known experience-dependent plasticity changes have been reported from motor learning tasks (Zatorre et al., 2012). Using multi-week training paradigms in juggling, some of these studies have identified both volumetric changes in visual and parietal cortices (Draganski et al., 2004; Scholz et al., 2009) and fractional anisotropy changes in the posterior intraparietal sulcus (Scholz et al., 2009). Complementary work has examined structural correlates for professional piano players, identifying volumetric differences in motor and parietal regions (Gaser and Schlaug, 2003) as well as structural connectivity differences in DTI data within the corticospinal tracts that connect motor cortex with the brainstem and spinal cord (Bengtsson et al., 2005). These training induced changes may arise from activity-dependent myelination (Fields, 2015), which in turn may contribute to the observed changes in functional connectivity during long-term motor learning (Sampaio-Baptista et al., 2015). However, unlike juggling or extensive piano practice, our participants did not train on a complex visuo-motor task, but instead, they learned a pairing between a visual cue and a required finger movement for a set of six sequences. In the context of this fine-motor training, we observed a stable relationship between visual connectivity and subject learning rate across all four scans, independent of the number of trials practiced. A Putative Role for Physically Extended Polysynaptic Connections. Because prior work in functional neuroimaging has linked changes in learning rate to functional connectivity between motor and visual areas (Mattar et al., 2015; Bassett et al., 2015), we directly assessed indirect connectivity defined as a variant of network communicability that we called walk strength. This metric computes variable walk lengths where paths between two nodes can have increasing numbers of intermediary steps. For example, a two-step walk could be taken from visual cortex through thalamus to motor cortex. We found that as walk length increased, individual differences in motor-visual connectivity were increasingly correlated with learning rate. These results suggest a role for physically extended sets of polysynaptic connections between motor and visual cortices that support the acquisition of this visuo-motor skill. Such a role is consistent with previous work in computational neuroscience highlighting the role of highly structured circuits in sequence generation and memory (Rajan et al., 2016; Hermundstad et al., 2011). Indeed, in computational models at the neuron level, architectures reminiscent of chains (Levy et al., 2001; Fiete et al., 15/32 2010) and rings are particularly conducive to the generation of sequences. Our results complement these insights at small spatial scales to suggest that long-distance (chain-like) paths at the large scale of white matter tractography are supportive of sequence production. In future, it may be interesting to assess the generalizability of these results across other sequential learning tasks, and to determine the degree to which additional measurements of indirect connectivity (Goni et al., 2014) may differentially relate to learning rate, performance accuracy, and reaction time (Tuch et al., 2005). Methodological Considerations First, it is important to note that in this study, we rely on DTI and white matter tractography to estimate subject-specific and whole-brain structural connectivity. However, it is important to note that DTI-based tractographic reconstructions present a number of limitations. Among these is the tendency of current methods to present false positives and false negatives when compared to histological studies (Thomas et al., 2014; Reveley et al., 2015). However, diffusion imaging remains the only reliable method for studying human white matter structure noninvasively. Moreover, we expect potential tractography biases to be consistent across subjects, allowing us to accurately access individual differences in white matter architecture and its relationship to behavior. Second, it is also important to note that it is not possible using these techniques to decipher the number of synapses present along the tracts between two regions, nor is it possible to decipher the number of synapses present along long-distance paths in the network. Thus, while the data supports a role for physically extended polysynaptic pathways, it does not directly speak to their microstructure. Third, it is important to note that while we hypothesized that interhemispheric connections would be less important for this task than for other tasks that require the manipulation of perceptual reference frames (Bernier and Grafton, 2010) or that utilized a single visual hemifield (Doron et al., 2012), it is nevertheless possible that interhemispheric connections also play a role. It will be important in the future to implement higher resolution diffusion imaging to clarify the potential role of interhemispheric connections in the learning of this novel visuo-motor skill. Fourth, it is interesting to ask whether the structural drivers of individual differences in learning rate are anatomically co-located with observed changes in functional connectivity during task performance. In fact, evidence suggests that this is not the case, and that instead regions that show individual differences in structural connectivity that are predictive of individual differences in learning rate are not the same as the regions that display changes in functional connectivity with training (Bassett et al., 2015). Together, these data suggest that further study is needed to understand the relationships between individual differences in structural connectivity and functional connectivity, and how they relate to gross changes in behavior or to individual differences in learning rate. Finally, we note that the lack of longitudinal changes in the strength of connectivity (measured both with 16/32 fractional anisotropy and with the number of reconstructed streamlines between pairs of large-scale brain regions) could be explained either by neuroscientific or methodological factors. It is important to note that with this particular data set, we are unable to determine the origin of this consistency with complete confidence. Conclusion We identified variability in structural connectivity that accounts for individual differences in learning rate over six weeks of training on a visuo-motor skill. Our analysis revealed direct connections among intrahemispheric visual regions as well as indirect connections between visual and motor cortices that suggests an underlying mechanism for differences in behavior. Clinically, these results offer novel biomarkers that may prove useful in predicting the time scales of motor rehabilitation following stroke and brain injury. In particular, because individuals with greater visual connectivity show swifter learning rates, a clinician may be able to predict the rate at which a patient will re-learn a motor skill after a stroke based on the degree to which their visual system (and its indirect connections with the motor system) remain intact. More generally, our results may inform personalized training paradigms for healthy individuals; individuals with greater visual connectivity – and greater strength of indirect connectivity between motor and visual systems – may require less training to obtain the same proficiency as an individual with lesser connectivity and greater training. While speculative at this point, these possibilities motivate future work in clarifying the utility of white matter architecture in optimizing visuo-motor training across healthy and injured populations. Acknowledgments D.S.B. and A.K. acknowledge support from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army Research Laboratory and the Army Research Office through contract numbers W911NF-10-2-0022 and W911NF-14- 1-0679, the National Institute of Mental Health (2-R01-DC-009209-11), the National Institute of Child Health and Human Development (1R01HD086888-01), the Office of Naval Research, and the National Science Foundation (BCS-1441502, BCS-1430087, and PHY-1554488). S.T.G. and N.F.W. acknowledge support from the National Institute of Neurologic Disease and Stroke (1-P01-NS44393). The content is solely the responsibility of the authors and does not necessarily represent the official views of any of the funding agencies. References Andersson, J. L. R. and Sotiropoulos, S. N. (2016). An integrated approach to correction for off-resonance effects and subject movement in diffusion MR imaging. NeuroImage, 125:1063–1078. 17/32 Bassett, D. S., Brown, J. A., Deshpande, V., Carlson, J. M., and Grafton, S. T. (2011). Conserved and variable architecture of human white matter connectivity. Neuroimage, 54(2):1262–1279. Bassett, D. S., Wymbs, N. F., Porter, M. A., Mucha, P. J., and Grafton, S. T. (2014). Cross-linked structure of network evolution. Chaos, 24(1):013112. Bassett, D. S., Wymbs, N. F., Rombach, M. P., Porter, M. A., Mucha, P. J., and Grafton, S. T. (2013). Task-based core-periphery organization of human brain dynamics. PLoS computational biology, 9(9):e1003171. Bassett, D. S., Yang, M., Wymbs, N. F., and Grafton, S. T. (2015). Learning-induced autonomy of sensorimotor systems. Nat Neurosci, 18(5):744–751. Becker, C., Pequito, S., Pappas, G. J., Miller, M. B., Grafton, S. T., Bassett, D. S., and Preciado, V. (2015). Accurately predicting functional connectivity from diffusion imaging. arXiv, 1512:02602. Bengtsson, S. L., Nagy, Z., Skare, S., Forsman, L., Forssberg, H., and Ull´en, F. (2005). Extensive piano practicing has regionally specific effects on white matter development. Nature Neuroscience, 8(9):1148–1150. Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society Series B . . . . Bernier, P. M. and Grafton, S. T. (2010). Human posterior parietal cortex flexibly determines reference frames for reaching based on sensory context. Neuron, 68(4):776–788. Betzel, R. F., Byrge, L., He, Y., Goni, J., Zuo, X. N., and Sporns, O. (2014). Changes in structural and functional connectivity among resting-state networks across the human lifespan. Neuroimage, 102(Pt 2):345–357. Blumenfeld-Katzir, T., Pasternak, O., Dagan, M., and Assaf, Y. (2011). Diffusion MRI of structural brain plasticity induced by a learning and memory task. PLoS One, 6(6):e20678. Cieslak, M. and Grafton, S. T. (2014). Local termination pattern analysis: a tool for comparing white matter morphology. Brain Imaging Behav, 8(2):292–299. Crofts, J. J. and Higham, D. J. (2008). Communicability in complex brain networks. arXiv.org. Crofts, J. J. and Higham, D. J. (2009). A weighted communicability measure applied to complex brain networks. J R Soc Interface, 6(33):411–414. Crofts, J. J., Higham, D. J., Bosnell, R., Jbabdi, S., Matthews, P. M., Behrens, T. E., and Johansen-Berg, H. (2011). Network analysis detects changes in the contralesional hemisphere following stroke. Neuroimage, 54(1):161–169. Crossman, E. R. F. W. (2010). A theory of the acquisition of speed-skill. Ergonomics, 2(2):153–166. 18/32 Dayan, E. and Cohen, L. G. (2011). Neuroplasticity subserving motor skill learning. Neuron, 72(3):443–454. Diedrichsen, J. and Kornysheva, K. (2015). Motor skill learning between selection and execution. Trends in Cognitive Sciences, 19(4):227–233. Ding, Y., Yao, B., Lai, Q., and McAllister, J. P. (2001). Impaired motor learning and diffuse axonal damage in motor and visual systems of the rat following traumatic brain injury. Neurol Res, 23(2–3):193–202. Doron, K. W., Bassett, D. S., and Gazzaniga, M. S. (2012). Dynamic network structure of interhemispheric coordination. Proc Natl Acad Sci U S A, 109(46):18661–18668. Draganski, B., Gaser, C., Busch, V., Schuierer, G., Bogdahn, U., and May, A. (2004). Neuroplasticity: changes in grey matter induced by training. Nature, 427(6972):311–312. Estrada, E. and Hatano, N. (2007). Communicability in complex networks. arXiv.org, (3):036111. Fields, R. D. (2015). A new mechanism of nervous system plasticity: activity-dependent myelination. Nat Rev Neurosci, 16(12):756–767. Fiete, I. R., Senn, W., Wang, C. Z., and Hahnloser, R. H. R. (2010). Spike-time-dependent plasticity and heterosynaptic competition organize networks to produce long scale-free sequences of neural activity. Neuron, 65:563–576. Gaser, C. and Schlaug, G. (2003). Brain structures differ between musicians and non-musicians. The Journal of neuroscience : the official journal of the Society for Neuroscience, 23(27):9240–9245. Goni, J., van den Heuvel, M. P., Avena-Koenigsberger, A., Velez de Mendizabal, N., Betzel, R. F., Griffa, A., Hagmann, P., Corominas-Murtra, B., Thiran, J. P., and Sporns, O. (2014). Resting-brain functional connectivity predicted by analytic measures of network communication. Proc Natl Acad Sci U S A, 111(2):833–838. Grafton, S., Hazeltine, E., and Ivry, R. (1995). Functional mapping of sequence learning in normal humans. Journal of Cognitive Neuroscience, 7(4):497–510. Grafton, S. T., Woods, R. P., and Tyszka, M. (1994). Functional imaging of procedural motor learning: relating cerebral blood flow with individual subject performance. Human Brain Mapping, 1(3):221–234. Griffa, A., Baumann, P. S., Thiran, J.-P., and Hagmann, P. (2013). Structural connectomics in brain diseases. NeuroImage, 80:515–526. Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Wedeen, V. J., and Sporns, O. (2008). Mapping the structural core of human cerebral cortex. PLoS biology, 6(7):e159. Heathcote, A., Brown, S., and Mewhort, D. J. (2000). The power law repealed: the case for an exponential law of practice. 19/32 Psychonomic bulletin & review, 7(2):185–207. Hermundstad, A. M., Bassett, D. S., Brown, K. S., Aminoff, E. M., Clewett, D., Freeman, S., Frithsen, A., Johnson, A., Tipper, C. M., Miller, M. B., Grafton, S. T., and Carlson, J. M. (2013). Structural foundations of resting-state and task-based functional connectivity in the human brain. Proc Natl Acad Sci U S A, 110(15):6169–6174. Hermundstad, A. M., Brown, K. S., Bassett, D. S., Aminoff, E. M., Frithsen, A., Johnson, A., Tipper, C. M., Miller, M. B., Grafton, S. T., and Carlson, J. M. (2014). Structurally-constrained relationships between cognitive states in the human brain. PLoS Comput Biol, 10(5):e1003591. Hermundstad, A. M., Brown, K. S., Bassett, D. S., and Carlson, J. M. (2011). Learning, memory, and the role of neural network architecture. PLoS Comput Biol, 7(6):e1002063. Higham, D. J., Kalna, G., and Kibble, M. (2007). Spectral clustering and its use in bioinformatics. Journal of Computational and Applied Mathematics, 204(1):25–37. Hirsiger, S., Koppelmans, V., Merillat, S., Liem, F., Erdeniz, B., Seidler, R. D., and Jancke, L. (2016). Structural and functional connectivity in healthy aging: Associations for cognition and motor behavior. Hum Brain Mapp, 37(3):855–867. Holtmaat, A. and Svoboda, K. (2009). Experience-dependent structural synaptic plasticity in the mammalian brain. Nature reviews. Neuroscience, 10(9):647–658. Honey, C. J., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J. P., Meuli, R., and Hagmann, P. (2009). Predicting human resting-state functional connectivity from structural connectivity. Proc Natl Acad Sci U S A, 106(6):2035–2040. Hong, S. B., Zalesky, A., Park, S., Yang, Y. H., Park, M. H., Kim, B., Song, I. C., Sohn, C. H., Shin, M. S., Kim, B. N., Cho, S. C., and Kim, J. W. (2015). COMT genotype affects brain white matter pathways in attention-deficit/hyperactivity disorder. Hum Brain Mapp, 36(1):367–377. Ingalhalikar, M., Smith, A., Parker, D., Satterthwaite, T. D., Elliott, M. A., Ruparel, K., Hakonarson, H., Gur, R. E., Gur, R. C., and Verma, R. (2014). Sex differences in the structural connectome of the human brain. Proc Natl Acad Sci U S A, 111(2):823–828. Jarbo, K. and Verstynen, T. D. (2015). Converging structural and functional connectivity of orbitofrontal, dorsolateral prefrontal, and posterior parietal cortex in the human striatum. The Journal of neuroscience : the official journal of the Society for Neuroscience, 35(9):3865–3878. Jenkinson, M., Bannister, P., Brady, M., and Smith, S. (2002). Improved Optimization for the Robust and Accurate Linear Registration and Motion Correction of Brain Images. NeuroImage, 17(2):825–841. 20/32 Jenkinson, M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M. W., and Smith, S. M. (2012). FSL. NeuroImage, 62(2):782–790. Jenkinson, M. and Smith, S. (2001). A global optimisation method for robust affine registration of brain images. Medical image analysis, 5(2):143–156. Jezzard, P., Barnett, A. S., and Pierpaoli, C. (1998). Characterization of and correction for eddy current artifacts in echo planar diffusion imaging. Magnetic resonance in medicine, 39(5):801–812. Johansen-Berg, H. (2010). Behavioural relevance of variation in white matter microstructure. Current opinion in neurology, 23(4):351–358. Keller, T. A. and Just, M. A. (2015). Structural and functional neuroplasticity in human learning of spatial routes. NeuroImage, 125:256–266. Le Bihan, D. and Johansen-Berg, H. (2012). Diffusion MRI at 25: exploring brain tissue structure and function. NeuroImage, 61(2):324–341. Leemans, A. and Jones, D. K. (2009). The B-matrix must be rotated when correcting for subject motion in DTI data. Magnetic resonance in medicine, 61(6):1336–1349. Levy, N., Horn, D., Meilijson, I., and Ruppin, E. (2001). Distributed synchrony in a cell assembly of spiking neurons. Neural Netw, 14:815–824. Li, Y., Liu, Y., Li, J., Qin, W., Li, K., Yu, C., and Jiang, T. (2009). Brain anatomical network and intelligence. PLoS Comput Biol, 5(5):e1000395. Lindenberger, U., Li, S.-C., and Backman, L. (2006). Delineating brain-behavior mappings across the lifespan: substantive and methodological advances in developmental neuroscience. Neuroscience and biobehavioral reviews, 30(6):713–717. Lopez-Barroso, D., Catani, M., Ripolles, P., Dell'Acqua, F., Rodriguez-Fornells, A., and de Diego-Balaguer, R. (2013). Word learning is mediated by the left arcuate fasciculus. Proc Natl Acad Sci U S A, 110(32):13168–13173. Lovd´en, M., Wenger, E., Martensson, J., Lindenberger, U., and Backman, L. (2013). Structural brain plasticity in adult learning and development. Neuroscience and biobehavioral reviews, 37(9 Pt B):2296–2310. Lynch, J. C. and Tian, J. R. (2006). Cortico-cortical networks and cortico-subcortical loops for the higher control of eye movements. Prog Brain Res, 151:461–501. Mattar, M., Wymbs, N. F., Bock, A., Aguirre, G., Grafton, S. T., and Bassett, D. S. (2015). Predicting future learning from baseline network architecture. Submitted. May, A. (2011). Experience-dependent structural plasticity in the adult human brain. Trends in cognitive sciences, 15(10):475– 21/32 482. Neumann, N., Lotze, M., and Eickhoff, S. B. (2016). Cognitive expertise: An ale meta-analysis. Hum Brain Mapp, 37(1):262– 272. Newell, A. and Rosenbloom, P. S. (1993). Mechanisms of skill acquisition and the law of practice. MIT Press. Osher, D. E., Saxe, R. R., Koldewyn, K., Gabrieli, J. D. E., Kanwisher, N., and Saygin, Z. M. (2016). Structural Connectivity Fingerprints Predict Cortical Selectivity for Multiple Visual Categories across Cortex. Cerebral cortex (New York, N.Y. : 1991), 26(4):1668–1683. Pascual-Leone, A., Amedi, A., Fregni, F., and Merabet, L. B. (2005). The plastic human brain cortex. Annual review of neuroscience, 28(1):377–401. Rajan, K., Harvey, C. D., and Tank, D. W. (2016). Recurrent network models of sequence generation and memory. Neuron, 90(1):128–142. Reveley, C., Seth, A. K., Pierpaoli, C., Silva, A. C., Yu, D., Saunders, R. C., Leopold, D. A., and Ye, F. Q. (2015). Superficial white matter fiber systems impede detection of long-range cortical connections in diffusion MR tractography. Proceedings of the National Academy of Sciences of the United States of America, 112(21):E2820–8. Rhodes, B. J., Bullock, D., Verwey, W. B., Averbeck, B. B., and Page, M. P. (2004). Learning and production of movement sequences: behavioral, neurophysiological, and modeling perspectives. Hum Mov Sci, 23(5):699–746. Rosenbaum, D. A. (2014). Human Motor Control. Elsevier. Sampaio-Baptista, C., Filippini, N., Stagg, C. J., Near, J., Scholz, J., and Johansen-Berg, H. (2015). Changes in functional connectivity and GABA levels with long-term motor learning. Neuroimage, 106:15–20. Sampaio-Baptista, C., Khrapitchev, A. A., Foxley, S., Schlagheck, T., Scholz, J., Jbabdi, S., DeLuca, G. C., Miller, K. L., Taylor, A., Thomas, N., Kleim, J., Sibson, N. R., Bannerman, D., and Johansen-Berg, H. (2013). Motor skill learning induces changes in white matter microstructure and myelination. J Neurosci, 33(50):19499–19503. Scantlebury, N., Cunningham, T., Dockstader, C., Laughlin, S., Gaetz, W., Rockel, C., Dickson, J., and Mabbott, D. (2014). Relations between white matter maturation and reaction time in childhood. J Int Neuropsychol Soc, 20(1):99–112. Schmidt, R. A. and Lee, T. D. (2005). Motor Control and Learning. A Behavioral Emphasis. Human Kinetics Publishers. Scholz, J., Klein, M. C., Behrens, T. E., and Johansen-Berg, H. (2009). Training induces changes in white-matter architecture. Nat Neurosci, 12(11):1370–1371. Shrout, P. E. and Fleiss, J. L. (1979). Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin. 22/32 Smith, S. M. (2002). Fast robust automated brain extraction. Human brain mapping, 17(3):143–155. Smith, S. M., Fox, P. T., Miller, K. L., Glahn, D. C., Fox, P. M., Mackay, C. E., Filippini, N., Watkins, K. E., Toro, R., Laird, A. R., and Beckmann, C. F. (2009). Correspondence of the brain's functional architecture during activation and rest. Proceedings of the National Academy of Sciences of the United States of America, 106(31):13040–13045. Smith, S. M., Jenkinson, M., Woolrich, M. W., Beckmann, C. F., Behrens, T. E. J., Johansen-Berg, H., Bannister, P. R., De Luca, M., Drobnjak, I., Flitney, D. E., Niazy, R. K., Saunders, J., Vickers, J., Zhang, Y., De Stefano, N., Brady, J. M., and Matthews, P. M. (2004). Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage, 23:S208–S219. Snoddy, G. S. (1926). Learning and stability: a psychophysiological analysis of a case of motor learning with clinical applications. Journal of Applied Psychology, 10(1):1–36. Sporns, O., Tononi, G., and Kotter, R. (2005). The human connectome: A structural description of the human brain. PLoS Comput Biol, 1(4):e42. Taubert, M., Villringer, A., and Ragert, P. (2012). Learning-related gray and white matter changes in humans: an update. Neuroscientist, 18(4):320–325. Thomas, C. and Baker, C. I. (2013). Teaching an adult brain new tricks: a critical review of evidence for training-dependent structural plasticity in humans. NeuroImage, 73:225–236. Thomas, C., Ye, F. Q., Irfanoglu, M. O., Modi, P., Saleem, K. S., Leopold, D. A., and Pierpaoli, C. (2014). Anatomical accuracy of brain connections derived from diffusion MRI tractography is inherently limited. Proceedings of the National Academy of Sciences of the United States of America, 111(46):16574–16579. Tomassini, V., Jbabdi, S., Kincses, Z. T., Bosnell, R., Douaud, G., Pozzilli, C., Matthews, P. M., and Johansen-Berg, H. (2011). Structural and functional bases for individual differences in motor learning. Hum Brain Mapp, 32(3):494–508. Tuch, D. S., Salat, D. H., Wisco, J. J., Zaleta, A. K., Hevelone, N. D., and Rosas, H. D. (2005). Choice reaction time performance correlates with diffusion anisotropy in white matter pathways supporting visuospatial attention. Proceedings of the National Academy of Sciences, 102(34):12212–12217. Tunc, B., Solmaz, B., Parker, D., Satterthwaite, T. D., Elliott, M. A., Calkins, M. E., Ruparel, K., Gur, R. E., Gur, R. C., and Verma, R. (2016). Establishing a link between sex-related differences in the structural connectome and behaviour. Philos Trans R Soc Lond B Biol Sci, 371(1688). Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., Mazoyer, B., and Joliot, M. 23/32 (2002). Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. NeuroImage, 15(1):273–289. Wiestler, T., Waters-Metenier, S., and Diedrichsen, J. (2014). Effector-independent motor sequence representations exist in extrinsic and intrinsic reference frames. Journal of Neuroscience, 34(14):5054–5064. Wong, F. C., Chandrasekaran, B., Garibaldi, K., and Wong, P. C. (2011). White matter anisotropy in the ventral language pathway predicts sound-to-word learning success. J Neurosci, 31(24):8780–8785. Wymbs, N. F. and Grafton, S. T. (2015). The Human Motor System Supports Sequence-Specific Representations over Multiple Training-Dependent Timescales. Cerebral cortex (New York, N.Y. : 1991), 25(11):4213–4225. Yarrow, K., Brown, P., and Krakauer, J. W. (2009). Inside the brain of an elite athlete: the neural processes that support high achievement in sports. Nature reviews. Neuroscience, 10(8):585–596. Yeh, F.-C., Verstynen, T. D., Wang, Y., Fern´andez-Miranda, J. C., and Tseng, W.-Y. I. (2013). Deterministic diffusion fiber tracking improved by quantitative anisotropy. PloS one, 8(11):e80713. Yendiki, A., Koldewyn, K., Kakunoori, S., Kanwisher, N., and Fischl, B. (2013). Spurious group differences due to head motion in a diffusion MRI study. NeuroImage, 88C:79–90. Zatorre, R. J., Fields, R. D., and Johansen-Berg, H. (2012). Plasticity in gray and white: neuroimaging changes in brain structure during learning. Nature Publishing Group, 15(4):528–536. 24/32 Scan 1 Scan 2 Scan 3 Scan 4 MIN sequences MOD sequences EXT sequences 50 50 50 110 200 740 170 350 230 500 1,430 2,120 Table 1. Number of trials practiced of each sequence type at the start of each scanning session. 25/32 Motor Visual L,R Precentral gyrus L,R Intracalcarine cortex L,R Postcentral gyrus L,R Cuneus cortex L,R Superior parietal lobule L,R Lingual gyrus L,R Supramarginal gyrus, anterior L,R Supercalcarine cortex L,R Supplemental motor area L,R Occipital Pole L Parietal operculum cortex R Supramarginal gyrus, posterior Table 2. Brain areas in motor and visual systems derived directly from functional neuroimaging studies of the same task (Bassett et al., 2015). 26/32 Figure 1. Structural Connectivity in Motor and Visual Networks of Interest. (A–B) Previous research suggests that increased skill on the discrete sequence production (DSP) task requires concerted functional network changes in distributed regions of motor (A) and visual (B) systems (Bassett et al., 2015, 2013); see Table 2 for region names. (C) To assess structural correlates of individual differences in learning rate on the DSP task, we performed deterministic diffusion imaging tractography on 4 scans dispersed evenly throughout the 6 weeks of training. (D) We constructed structural networks using diffusion imaging tractography and the 111 cortical and subcortical regions in the Harvard-Oxford atlas to examine individual variability in connectivity strength. We also show that our results are robust across atlases, replicating our findings in the 90 cortical and subcortical region parcellation of the automated anatomical labeling (AAL) atlas. 27/32 0.00.5Connection StrengthRegionRegionACBD Figure 2. Overview of training, task paradigm, and MT estimation. (A) Training Schedule. Subjects underwent four scans, each approximately two weeks apart. Subjects practiced once a day for at least 10 days between each scanning session. (B) Subjects viewed a screen on which stimuli were displayed. Each sequence was preceded with the display of a sequence identity cue, which informed the subject which of six sequences would follow. During the sequence, subjects saw five horizontally arranged squares. For each element of the sequence, one box was highlighted for the subject, providing information on the key to press. Upon completion of the task, a fixation cross was displayed for a short Inter-Trial Interval (ITI), and every 10 trials performance feedback was provided. The squares were spatially mapped onto a key-pad, one corresponding to each finger in addition to the thumb (see insert). (C) Double exponential fit of MT to the number of trials practiced. The fit is shown for the fastest learner, the slowest learner, and the mean across all subjects. 28/32 EXT (64 trials/sequence)MOD (10 trials/sequence)MIN (1 trial/sequence)home training exampleSCAN1.5 hrSCAN1.5 hrSCAN1.5 hrSCAN1.5 hr"baseline training"(50 trials/sequence)trainingsessions: 6 weeks of trainingA14 days14 days14 days#1-10#11-20#21-30(50 trials/sequence)(50 trials/sequence)C0500100015002000Trial123456Movement Time (s)Mean FitFastest LearnerSlowest LearnerSubject Movement TimeOver Trials sequence identity cue+ Feedback2000 ms2000 ms + variable ITIBathumbindexaebcdS-R mappingbcdemiddleringpinky DSP task production Figure 3. Correlations between mean connection strength and learning rate across all scanning sessions. (A) We observe a significant correlation between the average strength of connections linking visual regionss and the learning rate. (B) No such relationship is observed for connections linking motor regions. The correlation between learning rate and visual-visual connectivity is largely driven by intrahemispheric (C) rather than interhemispheric (D) connections. (E-H): We observe the same relationships in the AAL atlas as in the Harvard Oxford atlas for each subset of connections. 29/32 Connection StrengthLearning Rate, κVisual-Visual ConnectivityConnection StrengthLearning Rate, κMotor-Motor ConnectivityConnection StrengthLearning Rate, κIntrahemispheric Visual ConnectivityConnection StrengthLearning Rate, κInterhemispheric Visual ConnectivityABCD−0.0020.0000.0020.0040.0060.0080.0100.0120.000450.00060.000750.00020.00030.00040.00050.0020.0030.0040.0050.0060.0070.0080.0090.0020.0030.0040.0050.0060.0070.0080.0090.000240.000300.000360.0000.0020.0040.0060.0080.000350.000450.00055r=0.499p=0.0125Connection StrengthLearning Rate, κConnection StrengthLearning Rate, κ0.0020.0030.0040.0050.0060.0070.0080.0090.00040.00060.00080.00100.0010.0020.0030.0040.0050.0060.0070.0080.0090.0100.00060.00070.0008Connection StrengthLearning Rate, κConnection StrengthLearning Rate, κ0.000150.00030.000450.00060.0020.0030.0040.0050.0060.0070.0080.0090.00090.00100.00110.00120.00080.0010.0020.0030.0040.0050.0060.0070.0080.009Harvard-OxfordVisual-Visual ConnectivityMotor-Motor ConnectivityIntrahemispheric Visual ConnectivityInterhemispheric Visual ConnectivityEFGHAutomated Anatomical Labelingr=0.067p=0.389r=0.681p=0.0005r=0.007p=0.512r=0.031p=0.449r=0.436p=0.027r=0.102p=0.666r=0.624p=0.002 Figure 4. Connectivity-learning relationship by scan (A–D): The structural connection strength between intrahemispheric visual-visual region pairs accounts for individual variability in learning rate, and this relationship is stable across the four scanning sessions. This relationship is significant after Bonferroni correction for Scans 2, 3, and 4 in the Harvard-Oxford atlas, with near significance in Scan 1. (E-H): We replicated these results within the AAL atlas, showing significance in all four scan sessions. 30/32 Connection StrengthConnection StrengthConnection StrengthConnection StrengthABCDConnection StrengthLearning Rate, κConnection StrengthConnection StrengthConnection StrengthScan 10.0010.0020.0030.0040.0050.0060.0070.0080.0090.00040.00050.00060.0007Scan 20.0000.0020.0040.0060.0080.0100.0120.00050.000650.0008Scan 3−0.0020.0000.0020.0040.0060.0080.0100.0120.000450.000600.00075Scan 4−0.0020.0000.0020.0040.0060.0080.0100.000450.000600.00075r=0.411p=0.050r=0.706p=0.0003r=0.613p=0.002r=0.715p=0.0002Learning Rate, κLearning Rate, κLearning Rate, κLearning Rate, κ0.0010.0020.0030.0040.0050.0060.0070.0080.0090.00080.000950.00110.001250.0020.0030.0040.0050.0060.0070.0080.0090.000900.001050.00120−0.0020.0000.0020.0040.0060.0080.0100.0120.00080.00100.00120.0010.0020.0030.0040.0050.0060.0070.0080.0090.0100.000850.001000.00115r=0.551p=0.011r=0.616p=0.002r=0.608p=0.002r=0.593p=0.003EFGHScan 1Scan 2Scan 3Scan 4Learning Rate, κLearning Rate, κLearning Rate, κHarvard-OxfordAutomated Anatomical Labeling Figure 5. Anatomical specificity of visual-to-visual connection strength correlation with learning rate. (A) Significant correlation coefficients (uncorrected) between connection strength and learning rate κ for intrahemispheric visual regions are shown in colored boxes. We note relationships that pass a correction for multiple comparisons of the form p < 0.05 n , where n is the number of comparisons. Gray boxes were not included in the analysis, representing either duplicate entries or interhemisphere connections. (B) The reconstructed streamlines are shown for the only region pair that survives FDR correction: the connection between right intracalcarine cortex (green region) and right cuneal cortex (purple region). 31/32 0.420.640.380.40.56Left Cuneal CortexLeft Intracalcarine CortexLeft Lingual GyrusLeft Occipital PoleLeft Supracalcarine CortexRight Cuneal CortexRight Intracalcarine CortexRight Lingual GyrusRight Occipital PoleRight Supracalcarine CortexRight Supracalcarine CortexRight Occipital PoleRight Lingual GyrusRight Intracalcarine CortexRight Cuneal CortexLeft Supracalcarine CortexLeft Occipital PoleLeft Lingual GyrusLeft Intracalcarine CortexLeft Cuneal CortexRegional Specificity ofConnectivity-Learning Relationship−0.650.65Streamlines Connecting HighlyCorrelated Regional PairsPearson Correlation CoefficientABRight Intracalcarine CortexRight Cuneal Cortex Figure 6. Indirect connections between visual and motor cortex facilitate learning. (A) Indirect connections between regions of interest can be quantified by walks on a structural network. Consider a toy graph in which paths exist from the source (S) to the target (T ). The eventual flow of information between S and T will not only be influenced by direct connections, but also by indirect walks of length greater than one. (B) Correlation between individual differences in motor-visual connection strength at increasing walk lengths (measured by (D− 1 2 )n) and individual differences in learning rate κ. The correlation 2 AD− 1 becomes significant at a walk length of n = 15. The red line indicates the p = 0.05 significance level calculated from the expectation of Pearson correlation coefficients in normal data; the black line indicates the p = 0.05 significance level calculated from a non-parametric permutation based null model in which node labels have been shuffled uniformly at random. 32/32 Correlation between Visual-Motor Connectivity and Learning RateSTABDirect ConnectivitySTABSTABTwo-stepSTABThree-stepABWalk Length−0.2−0.10.00.10.20.30.40.50.610010-110-210-305101520253035400510152025303540Walk LengthPearson Correlationp-value
1605.04463
1
1605
2016-05-14T20:10:00
Diversity of emergent dynamics in competitive threshold-linear networks: a preliminary report
[ "q-bio.NC", "nlin.AO" ]
Threshold-linear networks consist of simple units interacting in the presence of a threshold nonlinearity. Competitive threshold-linear networks have long been known to exhibit multistability, where the activity of the network settles into one of potentially many steady states. In this work, we find conditions that guarantee the absence of steady states, while maintaining bounded activity. These conditions lead us to define a combinatorial family of competitive threshold-linear networks, parametrized by a simple directed graph. By exploring this family, we discover that threshold-linear networks are capable of displaying a surprisingly rich variety of nonlinear dynamics, including limit cycles, quasiperiodic attractors, and chaos. In particular, several types of nonlinear behaviors can co-exist in the same network. Our mathematical results also enable us to engineer networks with multiple dynamic patterns. Taken together, these theoretical and computational findings suggest that threshold-linear networks may be a valuable tool for understanding the relationship between network connectivity and emergent dynamics.
q-bio.NC
q-bio
Diversity of emergent dynamics in competitive threshold-linear networks: a preliminary report Katherine Morrison1,2, Anda Degeratu3, Vladimir Itskov1 & Carina Curto1 May 13, 2016 1 Department of Mathematics, The Pennsylvania State University, University Park, PA 16802 2 School of Mathematical Sciences, University of Northern Colorado, Greeley, CO 80639 3 Mathematisches Institut, Albert-Ludwig-Universitat Freiburg, Freiburg, Germany 79104 Abstract. Threshold-linear networks consist of simple units interacting in the presence of a threshold nonlinearity. Competitive threshold-linear networks have long been known to exhibit multistability, where the activity of the network settles into one of potentially many steady states. In this work, we find condi- tions that guarantee the absence of steady states, while maintaining bounded activity. These conditions lead us to define a combinatorial family of competitive threshold-linear networks, parametrized by a sim- ple directed graph. By exploring this family, we discover that threshold-linear networks are capable of displaying a surprisingly rich variety of nonlinear dynamics, including limit cycles, quasiperiodic attractors, and chaos. In particular, several types of nonlinear behaviors can co-exist in the same network. Our mathematical results also enable us to engineer networks with multiple dynamic patterns. Taken together, these theoretical and computational findings suggest that threshold-linear networks may be a valuable tool for understanding the relationship between network connectivity and emergent dynamics. Contents 1 Introduction 2 Mathematical results 3 Simulations 3.1 Dynamic diversity from network connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Network engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 5 5 7 4 Proofs 8 8 4.1 Permitted sets background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Proof of Theorem 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 4.3 Proof of Theorem 2.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 5 Acknowledgments 1. Introduction 12 Dynamic networks consist of nodes and their interactions. They are commonly-used models in fields as disparate as ecology, economics, and neuroscience. Even when the building blocks are simple, these networks can display rich emergent dynamics, whose complexity cannot be reduced to a sum of con- stituent parts. For this reason, dynamic networks are exemplars of complex systems. Moreover, the most interesting dynamic phenomena that arise are fundamentally nonlinear behaviors, such as multistability, periodic attractors, and chaos. 1 Despite this, dynamic networks are often approximated using linear models – namely, linear systems of ordinary differential equations (ODEs). This is because the accompanying mathematical theory is extremely well-developed. Indeed, we may say that networks with linear interactions are the complex systems we already understand. While reducing more complicated models to linear approximations can be a useful approach, this strategy also poses severe limitations. Phenomena such as multistability, chaos, and robust periodic attractors (limit cycles) simply do not occur in linear models. Can we replace linear systems of ODEs with something "almost" linear – simple enough that a useful mathematical theory can be developed, yet capable of capturing the full variety of nonlinear behavior? Motivated by this question, we study the dynamics of threshold-linear networks. Emergent dynam- ics in these networks are not inherited from intrinsically oscillating nodes or a fluctuating external drive – instead, they can be attributed solely to the structure of connectivity, given by a matrix W . The non- linear behavior stems entirely from a simple threshold at each node, which guarantees that the activity of individual units cannot go negative. This is natural in any system where the dynamic variables repre- sent fundamentally non-negative quantities, such as the size of a population, a chemical concentration, or the firing rate of a neuron. Though these systems look essentially linear, the presence of the thresh- old changes everything. With it, the entire repertoire of complex nonlinear behavior comes into play: multistability, limit cycles, quasi-periodic behavior, and even deterministic chaos emerges. Many things were previously known about the stable fixed points of threshold-linear networks, espe- cially in the case of symmetric W [1, 2, 3, 4]. In this work, we were motivated by trying to find conditions that guarantee the absence of stable fixed points, so that network activity is forced into regular (or irreg- ular) oscillations. Our initial results led us to define a combinatorial family of competitive threshold-linear networks, which we call the CTLN model, that is parametrized by a simple directed graph. By exploring this family, we have discovered that threshold-linear networks are capable of displaying a surprisingly rich variety of nonlinear dynamics. Our mathematical results have also enabled us to engineer networks with multiple dynamic patterns, with distinct attractors corresponding to different initial conditions. Taken to- gether, our theoretical and computational results suggest that threshold-linear networks may be a valuable tool for studying the relationship between emergent dynamics and network connectivity. 2. Mathematical results Competitive threshold-linear networks. A threshold-linear network is a rate model consisting of n nodes, with dynamics governed by the system of ordinary differential equations:  n(cid:88) = −xi + dxi dt  Wijxj + θ , i = 1, . . . , n. (2.1) j=1 + The dynamic variables x1, . . . , xn give the activity levels1 of nodes 1, . . . , n. The matrix entries Wij are directed connection strengths between pairs of nodes, the parameter θ ∈ R is an external drive to each node, and the threshold-nonlinearity [·]+ is given by [y]+ = max{y, 0}. Although these networks have been around for decades in the neural networks community, the math- ematical theory is still a work in progress. It began in earnest about 15 years ago, with work by Hahn- loser, Seung, and others [1, 5]. Theirs was the first serious attempt to develop a mathematical theory of threshold-linear networks to rival that of Hopfield networks [6]. Not surprisingly, the initial results were confined to the case where W is a symmetric matrix. In [1], precise conditions were found to guaran- tee that network activity always converges to a stable fixed point, and a characterization was given of symmetric threshold-linear networks exhibiting multistability. 1If the nodes are neurons, the activity level is typically called a 'firing rate.' 2 A competitive threshold-linear network is a special case of (2.1) where we impose the additional restrictions: Wij ≤ 0 and Wii = 0 for all i, j = 1, . . . , n, and θ > 0. It is easy to see that such networks always have bounded activity. In particular, for each i we have ≤ −xi + θ, −xi ≤ dxi dt and thus if x(0) ∈ [0, θ]n, then x(t) ∈ [0, θ]n for all t > 0. What can we say about the dynamics inside this box? In the simplest competitive networks, where all interactions are equal so that Wij = w for all i (cid:54)= j, we have two extreme cases. First, note that the self-inhibition (or decay rate) of a single node, which has been normalized to −1, provides a natural scale for the strength of inhibition. If inhibition is strong, so that w < −1, we obtain a classical winner-take-all (WTA) network. Such a network has a stable fixed point corresponding to each single node, and no other attractors. The activity always converges to the fixed point of the "winning" node. If, on the other hand, inhibition is weak, so that w > −1, then the network synchronizes and activity always converges to a single stable fixed point in which all nodes have equal activity. Neither case is particularly exciting. When a competitive network has both strong and weak inhibition, things get more interesting – especially for asymmetric networks with no stable fixed points. Competitive networks with no stable fixed points. Our first new result establishes sufficient condi- tions such that a competitive network has no stable fixed points (or steady states). The statement of the theorem makes use of a simple2 directed graph GW on n vertices, that is defined from the n × n connectivity matrix W as follows: GW has an edge from j → i ⇔ Wij > −1 (for i (cid:54)= j). Note that Wij represents the influence of node j on node i. The edges of GW correspond to inhibitory interactions that are weaker than the self-inhibition of each node. Recall that a graph is oriented if there are no bi-directional connections, and a sink is a vertex with no outgoing edges. We can now state our first theorem, whose proof is given in Section 4. Theorem 2.2. Consider a competitive threshold-linear network with connectivity matrix W and associated graph GW . Suppose that: (i) GW is an oriented graph with no sinks, and (ii) whenever j → i in GW , Wij < . 1 Wji Then the network (2.1) has bounded activity and no stable fixed points. The CTLN model. Because the main relevant feature of the networks in Theorem 2.2 is the structure of the graph GW , we now specialize to competitive networks with only two values for the connections: one value for Wij < −1, and another for Wij > −1. To any simple directed graph G on n vertices and for any 0 < ε < 1 and δ > 0, we can associate a corresponding n × n connectivity matrix W = W (G, ε, δ) as follows:  0 −1 + ε −1 − δ Wij = if i = j, if i ← j in G, if i (cid:54)← j in G. (2.3) 2A graph is simple if it has no multiple edges and no self-loops. Simple graphs have binary adjacency matrices with zeros on the diagonal. 3 Figure 1: Competitive threshold-linear networks with strong and weak inhibition. (A) (Left) All interactions between nodes in the network are either strongly inhibitory (red arrows) or weakly inhibitory (black arrows). (Right) The graph of the network retains only the arrows corresponding to weak inhibition. (B) The equations for a CTLN network. (C) (Left) An oriented graph on 3 nodes, with the corresponding adjacency matrix below. (Right) Network activity follows the arrows in the graph, with peak activity occurring sequentially in the cyclic order 123. (D) Cliques correspond to stable fixed points, but only if they are target-free cliques. The clique 12 is target-free, but 23 is not. Simulations in (C-D) were run with θ = 1, ε = 0.25, and δ = 0.5. Clearly, GW = G, and W satisfies the conditions of a competitive network. We refer to threshold-linear networks of the form (2.1), with θ > 0 and W = W (G, ε, δ) as in (2.3), as the Combinatorial Threshold-Linear Network (CTLN) model. Note that this model is completely specified by the choice of directed graph, G, along with three positive real parameters: ε, δ, and θ (see Figure 1A- C). Moreover, condition (ii) of Theorem 2.2 is always satisfied, provided ε < δ/(1 + δ). We thus have the following result, obtained by specializing Theorem 2.2 to this case. Theorem 2.4. Let G be an oriented graph with no sinks, and consider the associated CTLN model (2.1) with W = W (G, ε, δ), as in (2.3). If ε < , then the network has bounded activity and no stable fixed points. 1 + δ δ Figure 1C displays the smallest oriented graph with no sinks, together with the activity of the corre- sponding CTLN model. The solutions to (2.1) for this W always converge to the same perfectly periodic trajectory, irrespective of the initial conditions. Stable fixed points in the CTLN model. To state the next theorem, we need a few graph-theoretic definitions. A subset of vertices σ is a clique of G if i → j and j → i for all pairs i, j ∈ σ. In other words, a clique is a subset of nodes that is all-to-all bidirectionally connected. We say that a vertex k is a target of σ if k /∈ σ and i → k for each i ∈ σ. If the clique σ has no targets, we say that it is target-free. A clique is maximal if it is not contained in any larger clique of G. Note that all target-free cliques are necessarily maximal, but maximal cliques need not be target-free. For example, the graph in Figure 1D has two maximal cliques, but only one of them is target-free. It turns out that only the target-free cliques 4 ABCactivitytime123D123051015202530051015202530timeactivitydynamics1234512345networkgraph can support3 stable fixed points. Theorem 2.5. Let G be a simple directed graph, and consider the associated CTLN model with W = W (G, ε, δ) for any choice of the parameters ε, δ, θ > 0 with ε < 1. If σ is a clique of G, then there exists a stable fixed point x∗ of (2.1) with support σ if and only if σ is target-free. The proof is given in Section 4. While Theorem 2.5 identifies precisely which cliques support stable fixed points, there may still exist additional stable fixed points that do not correspond to cliques. We have not, however, been able to find any such example. This leads us to the following conjecture: Conjecture 2.6. Consider the CTLN model W = W (G, ε, δ), where G is a simple directed graph. There exists a stable fixed point of (2.1) with support σ if and only if σ is a target-free clique of G. 3. Simulations 3.1. Dynamic diversity from network connectivity The CTLN model captures a surprisingly rich diversity of nonlinear dynamics. Figure 2 displays adjacency matrices for three different graphs on n = 25 nodes, along with two-dimensional projections of solutions that are periodic (A), chaotic (B), and quasi-periodic (C). These behaviors are examples of emergent dynamics, because they depend on the structure of the network and are not predictable from the local properties of individual nodes. Figure 2: Nonlinear dynamics of the CTLN model. Top panels show adjacency matrices for three oriented graphs on n = 25 nodes with no sinks, satisfying the conditions of Theorem 2.4. (Black = 1, white = 0, and gray = diagonal elements, which are ignored.) The corresponding CTLN model networks produce (A) a limit cycle, (B) a chaotic attractor, and (C) quasi-periodic behavior. Bottom panels show random two-dimensional projections of the 25-dimensional trajectories. Note that in Figure 2, and in all other simulations throughout this paper, we have used exactly the same CTLN model parameters as in Figure 1: θ = 1, ε = 0.25, and δ = 0.5, which satisfy the condition ε < δ/(1+δ) from Theorem 2.4. Differences in network dynamics thus arise solely as a result of differences in the underlying graph G. Even within the same network, different initial conditions may lead to different patterns of activity. The network in Figure 3 possesses several emergent nonlinear phenomena: multistability, a limit cycle, and chaos. The selected attractor depends only on the choice of initial conditions. Note that the graph in Figure 3 is not oriented, and the fixed points could be predicted using Theorem 2.5. 3We say that a subset of vertices supports a stable fixed point if there exists a fixed point attractor x∗ of the dynamics such k = 0 for all k /∈ σ. In other words, the nodes that are active at x∗ are precisely the ones in the i > 0 for all i ∈ σ, and x∗ that x∗ subset σ. 5 chaoslimit cycleABCquasi-periodic Figure 3: Variety of emergent dynamics in a single network with n = 8 nodes. Solutions corresponding to different initial conditions for the network on the left are shown, with the color of each activity trace matching the color of the corresponding node. The network has four attractors: two stable fixed points, one limit cycle, and one chaotic attractor. The equations for the dynamics are identical in each case; only the initial conditions differ. The plots on the far right are random two-dimensional projections of the 8-dimensional trajectories for the limit cycle and chaotic attractor, respectively. In Figure 1C, we saw that the activity in the limit cycle consisted of nodes being activated in a regular sequence. Sequential patterns of activation are common in the CTLN model, and are often irregular and surprising. In Figure 4, a network with 7 nodes has a single emergent sequence (irrespective of initial conditions), following only one of many possible cycles in the graph. In particular, it is surprising that the activity of node 2 decays to zero, while there are other nodes whose activity persists despite having a smaller in-degree. This shows that local properties of the graph are not sufficient to predict the emergent dynamics; the resulting sequence is somehow shaped by the structure of the graph as a whole. Figure 4: Emergence of an irregular sequence. (A) A graph on 7 nodes. The CTLN model makes no difference between black and gray edges, but black edges are highlighted here because they correspond to the emergent sequence of activation. (B) Node 2 decays to zero after a short period of transient activity, while the remaining nodes settle into a limit cycle of sequential activation with ordering 634517. The same sequence emerges irrespective of initial conditions, and is robust to small perturbations in the matrix W . Limit cycles need not be sequential – they can also display complex rhythms, including synchronous or quasi-synchronous activity for a subset of the nodes. The network in Figure 5 has two high-activity nodes (3 and 6) that peak at different times, while nodes 2, 4 and 7 are approximately synchronous. Note that node 6 has the highest peak activity, even though it has the lowest in-degree among all nodes in the network. Using Theorem 2.4, we can also generate large random networks that are guaranteed never to settle into a steady state. Such networks can exhibit spontaneous transitions between distinct patterns of net- work activity. In Figure 6, the total population activity trace has a sharp qualitative change around t = 80; this is reminiscent of state transitions in cortical networks observed during light anesthesia and sleep [7]. 6 16234587activitytimelimit cycleprojection015020025030050100chaotic attractoractivitytime010203040506070809010001020304050timefixed point 201020304050timefixed point 1activityactivity1234567timefiring rate01020405060301673452total pop activityAB Figure 5: Emergence of a complex rhythm. A graph on 7 nodes (left) yields a CTLN model whose activity always settles into the same limit cycle (right). The activity in this limit cycle is a rhythmic, with some nodes that are quasi-synchronous rather than sequential in their activation. Figure 6: Spontaneous state transition. A random network of 50 nodes satisfying the conditions of Theorem 2.4 exhibits a spontaneous state transition from irregular to periodic behavior. Total population activity (black trace) is the sum of the activity for all 50 nodes. 3.2. Network engineering The power of Theorems 2.4 and 2.5 is that they enable us to reason about the graph G in order to make surprisingly strong yet accurate predictions about the resulting network dynamics. Such mathematical results are thus valuable tools for designing networks with prescribed dynamic properties, and underscore the potential of the CTLN model for understanding the relationship between connectivity and emergent dynamics in networks with simple nonlinearities. Rhythmic patterns of activity, supporting locomotion and other functions, arise in Central Pattern Gen- erator circuits (CPGs) throughout the nervous system [8, 9]. The CTLN model provides a natural frame- work for CPGs. For example, Figure 7 shows that limit cycles corresponding to two different quadruped gaits (bound and trot) can coexist in the same network, with the network selecting one pattern over the other based solely on initial conditions. Note that the network in Figure 7A has four different cliques, but they all have targets and hence none of them can support stable fixed points (see Theorem 2.5). More generally, Theorem 2.5 enables us to construct networks that continually transition between cliques without getting "stuck," since we know precisely which cliques in a graph correspond to steady states. Figure 8A depicts a network with a series of overlapping cliques, only the last of which is target- free. We can also engineer networks by patching together modules that individually yield limit cycles, rather than cliques. Figure 8B depicts a network that has 6 overlapping limit cycles, corresponding to subsets of nodes 1-5, 4-8, 7-11, 10-14, 13-17, and 16-20. The network activity will stay in a single limit cycle indefinitely, unless it receives an external "kick" helping it to transition to an adjacent limit cycle. 7 1234567timefiring rate0510152025303540timeactivity (a.u.)total pop activity50 nodesspontaneous state transition04080120160200 Figure 7: A Central Pattern Generator circuit for quadruped motion. (A) The graph of a CTLN model with 8 nodes that produces two distinct quadruped gaits: 'bound' and 'trot.' Note that the orientation of the two hind legs (LH, RH) is flipped for ease of drawing the graph. Arrows into or out of a clique of two nodes (grouped via gray arcs) represent connections to all nodes in the clique. (B,C) The network activity produces two distinct gaits, based only on differences in initial conditions. Our final example shows that a single network, with very simple architecture, can have multiple quasi- periodic attractors. The network in Figure 9A has n nodes and n− 2 quasi-periodic attractors. Each quasi- periodic orbit selects a single node from the inner (shaded) region, which forms a dynamic sequence with the blue and gray outer nodes (Figure 9B). 4. Proofs In this section we prove Theorems 2.2 and 2.5 (and, as a consequence, Theorem 2.4 – an immediate corollary of Theorem 2.2). Our proofs build on our previous work in [2, 3, 4]. First, we need some background on permitted sets in threshold-linear networks. 4.1. Permitted sets background An important concept that has been studied in connection to threshold-linear networks is that of a permit- ted set. Permitted sets were defined in [1] for generalized networks of the form (2.1), having the added flexibility that the external input θ can vary between nodes. Definition 4.1. Consider the threshold-linear network x = −x + [W x + b]+, where b ∈ Rn is a vector of external inputs (constant in time). A subset of nodes σ ⊂ [n] is a permitted set if there exists a b such that the network possesses an asymptotically stable fixed point x∗ with support σ. Note that permitted sets of a network depend only on W , so the definition still makes sense for our networks (2.1). For b = θ1, however, we do not in general expect a permitted set to have a corresponding fixed point. On the other hand, if a given σ is not a permitted set, there can be no stable fixed point having support σ. The requirement that σ is a permitted set is thus a necessary, but not sufficient, condition for the existence of a stable fixed point of (2.1) having support σ. Permitted sets are straightforward to detect from W . For any n × n matrix A, and index subset σ = {(cid:96)1, . . . , (cid:96)k} ⊂ [n], the principal submatrix Aσ is the k× k matrix whose entries are given by (Aσ)ij = A(cid:96)i,(cid:96)j . Theorem 4.2 ([2, 3]). Consider the threshold-linear network (2.1) on n nodes. A subset of nodes σ ⊂ [n] is a permitted set if and only the principal submatrix (−I + W )σ is stable – i.e., if all its eigenvalues have negative real part. 8 boundtrotboundtrottimefiring rate0102030405060timeAfiring rate0102030405060LFRFRHLHLH LF RF RHLH LF RF RHtimeBC (A) (Top) A chain of six overlapping 5-clique Figure 8: Regular and irregular sequential activity from modular architectures. modules, each in a different color. Nodes belonging to two adjacent 5-cliques are double-colored. Black nodes receive output edges from all nodes in one 5-clique, and feed forward onto all nodes in the next 5-clique. Note that only the final 5-clique (purple) is target-free. (Middle) The solution of the network when the nodes in the first 5-clique are initialized to 0.1, and all other nodes are initialized to 0. Darker regions correspond to higher firing rates; note that the outer target nodes are omitted. (Bottom) Same solution as in middle plot. The firing rate curves are colored by the nodes in the graph, with overlap nodes receiving both colors. The activity moves slowly from one clique to the next until it stabilizes on the last clique, which is target-free. The total population activity is given by the black trace above. (B) (Top) A chain of six overlapping modules, where each module is a cyclic tournament on five nodes. Nodes highlighted in red receive small kicks during the simulation. (Middle) The solution of the network when nodes 1 and 2 are initialized to 0.1, and all other nodes are initialized to 0. Initially the network activity is confined to the first module, and cycles among those nodes. Every 15 time units, a small kick is given to the middle node in the next module. The timing of these kicks and the label of the affected node are shown as red pulses along the bottom of the plot. (Bottom) Same solution as in the middle plot, with individual firing rates shown in color. All simulations were run with θ = 1, ε = 0.25, and δ = 0.5. This result was first obtained for symmetric W in [1], but later generalized to arbitrary W in [2]. 4.2. Proof of Theorem 2.2 In [3, Lemma 1] it was shown that if all 2 × 2 principal submatrices of an n × n matrix have negative trace and are unstable, then all larger principal submatrices are also unstable. Putting this together with Theorem 4.2 immediately implies the following lemma: Lemma 4.3. Let W be the connectivity matrix of a threshold-linear network. If all 2× 2 principal submatri- ces of −I + W have negative trace and are unstable, then the network has no permitted sets containing more than one node. Recalling that there can be no stable fixed points supported on subsets of nodes that are not permitted sets, we obtain: If all 2 × 2 Corollary 4.4. Consider the threshold-linear network (2.1), with diagonal entries Wii = 0. principal submatrices of −I + W are unstable, then the network has no stable fixed points supported on more than one node. 9 time010203040506012345612189158101416207111317190102030405060708090timefiring rate2468101214161820neuron2468101214161820firing rate369121518neuronAB Figure 9: Multiple quasi-periodic attractors in the same network. (A) A graph on n nodes, where there are n − 2 nodes in the middle (inner) layer. The top (blue) outer node feeds onto all nodes in the middle layer, while the bottom (gray) outer node receives connections from all nodes in the middle layer and feeds back onto the top node. This architecture produces n − 2 different quasi-periodic attractors that each involve the top node, one of the middle nodes, and the bottom node. (B) Two distinct quasi-periodic attractors, one involving the green middle node (top), and the other involving the red middle node (bottom). In each solution, the activity of all other middle nodes is small and synchronous, and is depicted in black. Random projections of the activity (right) indicate that they are quasi-periodic trajectories and not perfect limit cycles. To create the top plot, the green node was initialized at 0.1 and all others at 0; for the bottom plot, the red node was initialized at 0.1 and all others at 0. All simulations were run with θ = 1, ε = 0.25, and δ = 0.5. We are now ready to prove Theorem 2.2. Proof of Theorem 2.2. We have already seen that the activity is bounded. To see that there are no stable fixed points, we first show that there can be no stable fixed points supported on two or more nodes because the network has no permitted sets σ of size σ ≥ 2. By Corollary 4.4, it suffices to show that all 2× 2 principal submatrices of −I + W are unstable. Each of these matrices, , has negative trace, and is thus stable if and only if its determinant ∆ is positive. It is unstable if ∆ ≤ 0. We compute: (cid:18) −1 Wij Wji −1 (cid:19) (cid:18) −1 Wij Wji −1 (cid:19) ∆ = det = 1 − WijWji. Since the network is competitive and GW is an oriented graph, there are two cases: (a) −1 < Wij ≤ 0 and Wji ≤ −1 (or vice versa), or (b) both Wij ≤ −1 and Wji ≤ −1. In case (a), the graph GW must have an edge j → i, so by hypothesis (ii) in the theorem we know that WijWji > 1 (note that the inequality reverses, since Wji is negative). We thus have ∆ < 0. In case (b), we immediately see that WijWji ≥ 1 and thus ∆ ≤ 0. We conclude that all 2 × 2 principal submatrices of −I + W are unstable, and thus the network has no stable fixed points with 2 or more active nodes. Next, we show by contradiction that the network has no fixed points supported on a single node (i.e., i > 0 and there is no winner-take-all behavior). Suppose x∗ is a fixed point supported on node i, with x∗ j = 0 for all j (cid:54)= i, and recall that Wii = 0. It follows that x∗ n(cid:88) j=1 x∗ i = Wijx∗ j + θ = Wiix∗ i + θ = θ. On the other hand, for any k (cid:54)= i we must have x∗ k = 0, where x∗ k = Wkjx∗ j + θ = [Wkix∗ i + θ]+ = [Wkiθ + θ]+.  n(cid:88)  j=1 + 10 0102030405060708090100timefiring rate0102030405060708090100timefiring rateprojectionAB Now recall that GW has no sinks (by hypothesis (i) of the theorem), and so there exists at least one vertex (cid:96) such that i → (cid:96). This means W(cid:96)i > −1, and thus x∗ (cid:96) > 0, contradicting the assumption that the fixed point was supported only on node i. 4.3. Proof of Theorem 2.5 To prove Theorem 2.5, we make use of the fixed point conditions that were derived in [4]. Lemma 4.5. Consider the CTLN model with W = W (G, ε, δ), and suppose x∗ is a fixed point of (2.1) supported on a clique σ of G. Then x∗ is a stable fixed point, and x∗ σ = θ ε + (1 − ε)σ 1σ. Proof. If σ is a clique of G, then (−I + W )σ = (−1 + ε)11T − εIσ, with eigenvalues σ(−1 + ε) − ε and −ε. Clearly, these are negative for 0 < ε < 1, so we can conclude that (−I + W )σ is stable and thus It follows that any fixed point x∗ with support σ is stable and unique [3, Corollary σ is a permitted set. 9] (see also [4, Section 1.1]). To verify the formula for x∗ σ, we simply check that it satisfies the fixed point equation x∗ σ > 0, we can drop the threshold nonlinearity to obtain the equivalent constraint, (I − Wσ)x∗ σ = θ1σ. Now plugging in the desired expression for x∗ σ + θ1σ]+. Since x∗ σ = [Wσx∗ σ yields: σ =(cid:0)(1 − ε)11T + εIσ (cid:1) (I − Wσ)x∗ θ ε + (1 − ε)σ 1σ = θ ε + (1 − ε)σ ((1 − ε)σ1σ + ε1σ) = θ1σ. We can now prove Theorem 2.5. Proof of Theorem 2.5. (⇒) Suppose x∗ is a stable fixed point with support σ, where σ is a clique of G. Then Wij = −1 + ε for all pairs i, j ∈ σ. To see that σ must be a target-free clique, suppose that σ has a target k /∈ σ. This implies that Wki = −1 + ε for each i ∈ σ. It follows that (cid:34)(cid:88) i∈σ x∗ k = Wkix∗ i + θ (cid:35) (cid:34) = + (cid:88) (cid:35) (cid:20) (cid:21) (−1 + ε) x∗ i + θ i∈σ σ from Lemma 4.5 to obtain (cid:80) + εθ = ε + (1 − ε)σ > 0, + θσ where we have used the expression for x∗ contradicts the fact that x∗ k = 0, since k /∈ σ. We thus conclude that σ must be a target-free clique. ε + (1 − ε)σ . This (⇐) Suppose σ is a target-free clique. Since σ is a clique, it follows from Lemma 4.5 that if a fixed point x∗ with support σ exists, then it must be unique and stable, with x∗ ε + (1 − ε)σ 1σ. Clearly, x∗ σ > 0. To guarantee that the fixed point with support σ exists, however, we must also check that for each k /∈ σ we have dxk/dtx=x∗ = 0, so that i∈σ x∗ σ = i = θ (cid:35) (cid:34)(cid:88) i∈σ 0 = −x∗ k + Wkix∗ i + θ + (4.6) holds with x∗ there exists ik ∈ σ such that Wkik = −1 − δ. We thus have i = θ k = 0 and x∗ (cid:88) It follows that(cid:2)(cid:80) i∈σ i∈σ Wkix∗ Wkix∗ ε + (1 − ε)σ for each i ∈ σ. Since σ is a target-free clique, for any k /∈ σ (cid:18)−1 − δ + (σ − 1)(−1 + ε) + = 0, and equation (4.6) holds for each k /∈ σ since x∗ ε + (1 − ε)σ < 0. ε + (1 − ε)σ k = 0. −θδ (cid:19) + 1 = i + θ ≤ θ i + θ(cid:3) 11 5. Acknowledgments CC was supported by NSF DMS-1225666/1537228, NSF DMS-1516881, and an Alfred P. Sloan Research Fellowship. VI was supported by Joint NSF DMS/NIGMS grant R01GM117592, NSF DMS-1122519, NSF IOS-155925, and a DARPA Young Faculty Award. CC, KM, and VI also gratefully acknowledge the support of the Statistical and Applied Mathematical Sciences Institute, under grant NSF DMS-1127914. References [1] R. H. Hahnloser, H.S. Seung, and J.J. Slotine. Permitted and forbidden sets in symmetric threshold- linear networks. Neural Comput., 15(3):621–638, 2003. [2] C. Curto, A. Degeratu, and V. Itskov. Flexible memory networks. Bull. Math. Biol., 74(3):590–614, 2012. [3] C. Curto, A. Degeratu, and V. Itskov. Encoding binary neural codes in networks of threshold-linear neurons. Neural Comput., 25:2858–2903, 2013. [4] C. Curto and K. Morrison. Pattern completion in threshold-linear networks. Accepted to Neural Com- putation. Available at http://arxiv.org/abs/1512.00897, 2016. [5] X. Xie, R. H. Hahnloser, and H.S. Seung. Selectively grouping neurons in recurrent networks of lateral inhibition. Neural Comput., 14:2627–2646, 2002. [6] J.J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci., 79(8):2554–2558, 1982. [7] C. Curto, S. Sakata, S. Marguet, V. Itskov, and K.D. Harris. A simple model of cortical dynamics explains variability and state dependence of sensory responses in urethane-anesthetized auditory cortex. J. Neurosci., 29(34):10600–10612, 2009. [8] E. Marder and D. Bucher. Central pattern generators and the control of rhythmic movements. Curr. Bio., 11(23):R986–996, 2001. [9] R. Yuste, J.N. MacLean, J. Smith, and A. Lansner. The cortex as a central pattern generator. Nat. Rev. Neurosci., 6:477–483, 2005. 12
1703.05414
1
1703
2017-03-15T22:54:15
A Multitaper, Causal Decomposition for Stochastic, Multivariate Time Series: Application to High-Frequency Calcium Imaging Data
[ "q-bio.NC", "physics.data-an" ]
Neural data analysis has increasingly incorporated causal information to study circuit connectivity. Dimensional reduction forms the basis of most analyses of large multivariate time series. Here, we present a new, multitaper-based decomposition for stochastic, multivariate time series that acts on the covariance of the time series at all lags, $C(\tau)$, as opposed to standard methods that decompose the time series, $\mathbf{X}(t)$, using only information at zero-lag. In both simulated and neural imaging examples, we demonstrate that methods that neglect the full causal structure may be discarding important dynamical information in a time series.
q-bio.NC
q-bio
A Multitaper, Causal Decomposition for Stochastic, Multivariate Time Series: Application to High-Frequency Calcium Imaging Data Andrew T. Sornborger James D. Lauderdale Department of Mathematics, University of California, Davis, CA Department of Cellular Biology, University of Georgia, Athens, GA Email: [email protected] Email: [email protected] Abstract -- Neural data analysis has increasingly incorporated causal information to study circuit connectivity. Dimensional reduction forms the basis of most analyses of large multivariate time series. Here, we present a new, multitaper-based decom- position for stochastic, multivariate time series that acts on the covariance of the time series at all lags, C(τ ), as opposed to standard methods that decompose the time series, X(t), using only information at zero-lag. In both simulated and neural imaging examples, we demonstrate that methods that neglect the full causal structure may be discarding important dynamical information in a time series. Index Terms -- Neural Imaging, Multivariate Time Series, Ma- trix Decomposition, Causality, Multitaper Methods, Spectral Analysis, Dimension Reduction. I. INTRODUCTION The technology for imaging neural systems has improved to the point where we can rapidly be overcome by the sheer size of our datasets. For this reason, a significant amount of effort has been expended to develop dimensional reduction methods. Independent component analysis [1], [2], non-negative ma- trix factorization [3], generalized linear models [4], principal component analysis [5], wavelet analyses [6], [7], and other matrix factorizations [8] have all been used to advantage in the analysis of neural imaging data. The lagged-covariance (cid:27) C(τ ) = E dtX(t)XT (t − τ ) and its Fourier transform, the cross-spectrum, C(τ )e−2πif τ dτ S(f ) = (1) (2) (cid:26)(cid:90) (cid:90) are fundamental to the description of a stationary random process, X(t). For a stationary process, the cross-spectrum is uncorrelated as a function of frequency, f. Therefore, statisti- cal estimates of stationary processes are often performed in the frequency domain. Since estimates of the cross-spectrum are approximately independent and identically distributed (i.i.d.), confidence intervals can typically be calculated more easily in the frequency domain. Despite the fact that the Fourier transform can decorrelate a stationary process, uncertainty principles are also fundamental to the analysis of time series. They are a statement that a function that has well-localized support in one set of coor- dinates, say time, becomes delocalized when transformed to another set of coordinates, i.e. frequency. Therefore, although the statistics of a stationary process are more tractable in frequency coordinates, if the covariance of the process, C(τ ), has more localized temporal support than the cross-spectrum, S(f ), then frequency-by-frequency estimation of covariance matrices can result in a loss of statistical power, since the power can be spread across many frequencies. This statement also holds in the opposite direction if the cross-spectrum has more localized frequency support than the covariance. The goal of this paper is to improve the detection and estimation of lagged-covariance or cross-spectral information using multivariate, multitaper methods [9], [10]. We show that these methods can capture information in the data that goes unseen with standard dimensional reduction methods based only on zero-lag information. II. METHODS To attack this problem, we set up a framework for the detection and estimation of statistically significant covariance or cross-spectral information in multivariate data. The dimen- sional reduction of covariance or cross-spectral tensors is a technique that we have found to be very useful in revealing the structure latent in time series data. We use multitaper methods to construct our estimates. Multitaper methods provide a set of approximately independent estimates of the covariance or cross-spectrum allowing us to form consistent estimates with low broad-band frequency bias. A. The Probabilistic Model We will assume that the time series of interest may be least approximately or over a given time represented, at window, as a band-limited, weakly stationary process X(t) = dZ(f )e2πif t . (3) Here, dZ(f ) represents a vector of length P of orthogonal increment Cram´er processes (c.f. [9], [11], [12], [13]) for which E{dZ(f )} = 0 (4) (cid:90) fN −fN and E{dZ(f )dZ(f(cid:48))†} = . (5) With these assumptions, the covariance at lag τ may be written C(τ ) = E dt X(t)XT (t − τ ) (cid:26) S(f )df, 0, f = f(cid:48) f (cid:54)= f(cid:48) (cid:27) dt dZ(f ) dZ†(f(cid:48))e2πi(f−f(cid:48))te2πif τ (cid:90) fN −fN (cid:40)(cid:90) = E giving (cid:26)(cid:90) (cid:90) fN (cid:90) fN −fN −fN (cid:41) , (6) (7) C(τ ) = df S(f )e2πif τ . Here and below, bolded letters denote vectors; capital, un- bolded letters denote matrices; small, unbolded letters denote scalars; T denotes the matrix transpose, † denotes the Hermi- tian transpose (complex conjugated and transposed) and E{} denotes an expectation value. B. Causal Decomposition: Decomposing Covariance and Cross-Spectral Tensors the Lagged- The most commonly used method for decomposing a data matrix is the singular value decomposition (SVD). This de- composition is used to estimate the principal components of stochastic processes and is so useful in the study of system dynamics (among other things) that it was reinvented many times during the 20th century and is also called the proper orthogonal decomposition (POD), the Karhunen-Lo`eve (KL) transform and the method of empirical eigenfunctions. The SVD of a vector time series is a decomposition of the X(t) = undnvn(t) , (8) form where (cid:90) rankX(cid:88) n=1 (9) dtX(t)X(t)T un = λnun , the {ui} are mutually orthogonal, the {di = λi} are listed in descending order and {vi(t)} are the normalized projections i X(t)/di, and are mutually orthogonal. Note that vi(t) = uT the operator (cid:82) dtX(t)X(t)T on the left of (9) is just the √ zero-lag covariance matrix, C(0). Other decompositions, such as those mentioned in the Introduction are also based on reductions that use zero-lag information. Typically, the SVD is first used to find left, un, and right, vn(t), eigenvector bases for X(t). Then, for dimensional reduction, the standard procedure is to truncate the decompo- sition by setting a set of small singular values, {di}, to zero and reconstructing X(t) in the subspace of the remaining left and right eigenvectors. As we saw in the Introduction and above, for stationary processes, the fundamental quantity of interest for modeling a stochastic system is the set of lagged covariance matrices, C(τ ), or the cross-spectral matrices at different frequencies, Fig. 1. The two covariance matrices generated by autoregressive process (16). A) C(0), the covariance matrix with support at zero lag; B) C(20), the covariance matrix at lag τ = 20. This matrix is identical to that at τ = 40 and τ = 60. (cid:88) n (cid:88) S(f ). These translationally-invariant quantities measure how the stochastic system evolves over time depending on its previous history. Therefore, instead of applying the SVD to the raw data, X(t), we apply it to estimates of the tensor C(τ ) or S(f ) (note that once τ or f is discretized, these quantities are indeed rank-three tensors). The SVD is not unique for tensors of rank higher than two. However, since τ and f are clearly special since they are ordered, we will decompose C(τ ) or S(f ) along these indices. Thus, the decomposition of the covariance becomes C(τ ) = Knanpn(τ ) , (10) where the {Kn} are eigen-covariance matrices, the {an} are singular values and the {pn(τ )} are time- or lag-like eigenvectors. Here, n = 1, . . . , N, where N = rank(X)2. An analogous treatment results in the decomposition of the cross-spectrum S(f ) = Hnbnqn(f ) , (11) n where the {Hn} are eigen-cross spectral matrices, the {bn} are singular values and the {qn(f )} are frequency-like eigen- vectors. Since these are decompositions of quantities that are funda- mentally related to the causal structure of stationary processes, we will refer to Eqs. (10) and (11) as Causal Decompositions (CDs). C. Multitaper Estimation To obtain consistent statistical estimates of C(τ ) and S(f ), we resort to multitaper spectral analysis [9]. Multitaper meth- ods have been used successfully in many applications in the analysis of neural time series [14]. Multitaper analysis is based on the projection of a time series onto a set of m = 1, ..., M orthogonal 'tapers' called Slepians (also called discrete prolate spheroidal sequences or DPSSs). Slepians are a complete set of functions that are ordered in terms of their frequency concentration and serve as optimal band-pass filters provided the maximum index M < 2T W − 3. Here, T is the number of points in the time series and W is a user-defined frequency bandwidth. The data is first tapered, then Fourier eigen-cross spectral matrices, singular values and frequency- like eigenvectors of the SVD of S(f ) (and the corresponding objects for the decomposition of C(τ )). Jackknife (leave-one- out) estimates are well-suited for application in these circum- stances. First, we outline the jackknife estimation procedure for S(f ). We define Sj(f ) = 1 M − 1 Jm(f )Jm(f )† . (14) (cid:88) (j) where(cid:80) (j) indicates the sum over all M tapers except the j'th taper. We now perform an SVD on Sj(f ) resulting in sets of eigen-covariance matrices, {{Hi}j}, singular values, {{bi}j}, and frequency-like eigenvectors, {{qi}j}, where i = 1, . . . , N and j = 1, . . . , M. From these sets, we calculate estimated means and variances for {Hi}, {bi}, and {qi}. With estimates of the mean and variance, it is straightforward to calculate confidence intervals for the eigen-covariance matrices, eigenvectors and singular values. This procedure for calculating mean and variance estimates of the objects resulting from the decomposition of S(f ) may also be used to calculate mean and variance estimates of the objects resulting from the decomposition of C(τ ). In this case, we perform SVDs on C j(τ ), where (cid:90) C j(τ ) = df Sj(f )e2πif τ , (15) and calculate the jackknife estimates as we did for Sj(f ). Notice that even though C j(τ1) is not independent of C j(τ2), the sample C j1 (τ ) is independent of C j2 (τ ) because it is calculated with Slepian tapers whose Fourier transform is orthogonal to the other Slepian tapers in the frequency domain. Thus the samples are approximately i.i.d.. This is an example of a transformation-based-bootstrap method [15], which leaves out independent samples in the frequency domain. For practical purposes, the CDs in (10) and (11) may be computed as 'raw' estimates, i.e. without an attempt to esti- mate statistics of the process. The use of multitaper methods does increase the time required to calculate a CD. Therefore, for cursory or preliminary analyses, it should be remembered that the number of tapers may be decreased or multitaper methods can be done away with entirely. The CD may be computed by simply forming a 'raw' estimate of C(τ ) or S(f ), then performing the SVD on the raw estimate. In this case, we have no confidence intervals, but we can get a sense of which eigen-matrices, {Hn} or {Kn}, represent the dominant contributions to the covariance or cross-spectrum, the form of the singular value spectrum, and the form of the eigenvectors {pn(τ )} or {qn(f )}. III. RESULTS We tested our method on both simulated and actual neural imaging data. The simulated data analysis is meant to demon- strate a situation in which a standard data reduction would not detect any statistically significant structure, but a CD should find statistically significant correlations in the data. The Fig. 2. A) The exact simulated AR process covariance at zero-lag; B) The exact AR process eigen-covariance at lags 20, 40, and 60. Note that this is equal to A−AT (see text); C,D) The second estimated eigen-covariances from CDs computed using multitaper methods in the frequency domain with time- bandwidth products TW = 5 and 40, respectively. Note that the first eigen- covariance was to very close approximation the identity matrix; E) Estimated second eigen-covariance from a non-multitaper CD computed in the time (lag) domain; F) Fidelity (correlation of estimated with exact eigen-covariance) of second eigen-covariance as a function of TW. 'CV' denotes the fidelity of the non-multitaper estimate taken in the time (lag) domain. Here, we see that T W ∼ 15 is sufficient for good esimation. For these simulations, T = 10, 000. transformed, resulting in a set of M frequency estimates, so- called eigenestimates, with low broad-band bias (i.e. estimates of amplitudes at a given frequency are relatively uninfluenced by frequencies outside of the bandwidth W ). The data is averaged across the eigenestimates, resulting in smoothed spectral estimates with reduced variance (relative to a single, tapered periodogram). Because multitaper techniques rely on a pre-defined band- width W , the CDs described here may be smoothed by the user to different degrees. More smoothing gives rise to smaller confidence intervals, but estimates that have lower resolution in frequency, whereas less smoothing results in larger confidence intervals, but higher frequency resolution. Define the tapered eigenestimate hm(t)X(t)e−2πif t , (12) Jm(f ) ≡ T(cid:88) t=1 M(cid:88) m=1 where hm(t) is a Slepian function. A multitaper estimate of S(f ) is given by S(f ) = 1 M − 1 Jm(f )Jm(f )† (13) C(τ ) may then be estimated via (7). D. Statistical Significance The estimate S(f ) is Wishart distributed with 2(M − P ) degrees of freedom [10]. However, the distribution of C(τ ) is more complicated, due to the Fourier transform in (7). Furthermore, we wish to obtain confidence intervals on the neural imaging data that we analyze are 1000Hz, fluorescence line-scan measurements of seizure-related calcium activity using the ratiometric, genetically-encoded calcium indicator cameleon YC2.1 in a larval zebrafish. We demonstrate that with this data set our method detects structure that would not have been detected with a standard analysis. A. Simulated Data We used an autoregressive (AR) process to generate a stationary multivariate time series, X(t). The process was X(t) = AX(t−20)+AX(t−40)+AX(t−60)+N(t) , (16) where the matrix A is depicted in Fig. 1B and N(t) is an i.i.d. Gaussian random vector. Note that with this matrix, there was no feedback in the time series. Time series 6 through 9 were simply added to time series 3 at the fixed lags of 20, 40 and 60 time units. Therefore the covariance tensor of this time series was diagonal at τ = 0 (Fig. 1A) (capturing the variance of each time series), proportional to A at lags 20, 40, and 60 and proportional to AT at lags −20, −40, and −60. Because all variables in the process have the same variance, C(0) is proportional to the identity matrix and an SVD of this data, whose left singular vectors are eigenvectors of C(0), outputs random right/left singular vectors with ordered, normally distributed singular values. In Fig. 2, we present results from a CD of the simulated AR process. The first eigen-cross-spectral estimates of the multita- per CD captured the diagonal C(0) cross-spectral matrix (see Fig. 1A for the theoretical matrix, the first eigen-cross-spectral matrix estimate is not shown). The second eigen-cross-spectral matrix of the multitaper CD captured the covariance due to A (Fig. 2B, this matrix is H2 ∼ A − AT ), with estimation improving as T W increased (Fig. 2C,D). Note also that, since the covariance tensor of this time series only has narrow (δ- function) support at the discrete lags 0, ±20, ±40, ±60, the uncertainty principle predicts that the cross-spectral tensor will have broad support across many frequencies, but the covariance tensor will have sharp (δ-function) support as a function of time (lag). For this reason, the 'raw' CD in the time (lag) domain is more statistically efficient than low T W multitaper estimators (in the sense of smallest mean squared error, Fig. 2F) due to the few, high signal-to-noise peaks in covariance support (Fig. 1E, red point in F). However, we have no confidence intervals for such raw estimators. We would expect processes with narrow frequency support to be more efficiently computed in the frequency domain. B. High-frequency Calcium Imaging Data Calcium imaging data taken at high frequencies are capable of detecting individual action potentials, but at the cost of a reduced signal-to-noise due to the reduced number of photons per image in an acquisition. Thus, dimensional reduction methods are important for the analysis of such data. Action potentials in vivo are typically modeled as a stochastic point process. Because individual neurons communicate with each other with synaptic propagation times on the order of tens of Fig. 3. A) Location of line-scan in 9 dpf larval zebrafish; B) Raw imaging data segment (one of 60 taken at 1000Hz); C) Normalized singular values from multitaper CD (see text); D, E, F) Eigen-cross-spectral matrices and eigen-cross-spectra from the three events labelled with asterisks in C). The eigen-cross-spectra are plotted with 2-σ confidence intervals.) milliseconds, one expects that a data-reduction using an SVD would not give useful results since only zero-lag correlations are being taken into account. This is indeed the case with the data set that we analyzed here. However, using a multitaper CD, we found a number of statistically significant events that went undetected with the SVD analysis. The line-scan calcium imaging data that we analyzed was taken from a line through the dorsal central nervous system of a 9 days-post-fertilization (dpf) larval zebrafish (Fig. 3A). The data was taken in a set of 60 contiguous 8.192 second segments (Fig. 3B) for a total of 491.52 seconds (∼ 8 minutes). A multitaper CD was performed on each segment. In order to visualize statistically significant events, the spectrum of a data matrix with normally distributed noise of equal power to the data in a given segment was subtracted from the spectrum of the CD (Fig. 3C). A series of statistically significant events were detected, three of which are shown in Fig. 3D-F in more detail. Each of these panels shows the two most significant eigen-cross-spectral matrices and the logarithm of the absolute value of the eigen-cross-spectrum. Note that the first eigen-cross-spectral matrix is, to a very good approximation, the identity. This is the matrix that would be diagonalized by the right eigenvectors of an SVD and would fail to identify the cross-spectral information contained in the second (and subsequent) eigen-cross-spectral matrices and eigen-cross-spectra. The eigen-matrices and eigen-spectra show different struc- ture from event to event. We will not interpret this here because it is impossible to interpret causal relationships in line scan data; the whole neuronal circuit is not sampled. However, due to the variety of relationships between line-scan pixels in the data, it is clear that a rich variety of neural mechanisms could be at play. Given a complete sample of neurons, using, for instance light-sheet microscopy, one would be able to make better guesses at the neural circuits that underlie these structures. IV. DISCUSSION AND CONCLUSIONS We have shown that the causal decomposition presented here can detect information that would normally go unnoticed in the standard multivariate dimensional reduction methods that are used to analyze imaging data. The CDs described here are a useful way of summariz- ing the information contained in the covariance and cross- spectrum. Like the standard SVD, which may be truncated to denoise a dataset X(t), the CD may be used to denoise C(τ ) or S(f ). The results of a CD may be readily visualized to give the user a clearer understanding of the underlying covariance or cross-spectral structure of a time series. They may also be used to obtain improved estimates of multivariate auto- regressive processes and, hence, improved predictions from measurements of the process. If the user wants to use the CD to dimensionally reduce the dataset (not the covariance or cross- spectrum), since the eigen-cross-spectra are hermitian, S(f ) = S†(−f ), S(f ) + S†(−f ) (alternatively, C(τ ) = C(−τ )T , thus C(τ ) + C(−τ )T ) will have an orthogonal, set of real eigenvectors. These may be listed in order of their covariance, thresholded, and the dataset may be projected on them, giving a reconstruction of the data retaining causal information. We have made no attempt to optimize the decompositions. Improvements could be made for instance because the ma- trices S(f ) are hermitian and S(f ) = S†(−f ). Similarly, C(τ ) = C T (−τ ). We note that a CD may be also be useful when performed on the coherency, amplitude spectrum, or other objects often investigated in the analysis of time series, instead of the cross-spectrum. By using multitaper techniques, our approach gives us statistics that may be used to obtain confidence intervals on the various objects (eigen-covariance, singular values and lag- or frequency-like eigenvectors) that result from the decom- position. They may also be used to test whether the objects resulting from the decomposition, for instance the singular value spectrum, of a given covariance or cross-spectrum are statistically different from others. As an example, one can form a two-sample t-statistic to test for non-stationarity of a process by calculating CDs of two different time-windows of the process and testing whether the distribution of singular values differs between them. ACKNOWLEDGMENT This work was funded by the National Institutes of Health, CRCNS program NS090645 (ATS and JDL). REFERENCES [1] M. J. McKeown, S. Makeig, G. G. Brown, T. P. Jung, S. S. Kindermann, A. J. Bell, and T. J. Sejnowski, "Analysis of fMRI data by blind separation into independent spatial components," Hum Brain Mapp, vol. 6, no. 3, pp. 160 -- 188, 1998. [2] A. Hyvarinen and E. Oja, "Independent component analysis: algorithms and applications," Neural Netw, vol. 13, no. 4-5, pp. 411 -- 430, 2000. [3] L. Tao, J. D. Lauderdale, and A. T. Sornborger, "Mapping Functional Connectivity between Neuronal Ensembles with Larval Zebrafish Trans- genic for a Ratiometric Calcium Indicator," Front Neural Circuits, vol. 5, p. 2, 2011. [4] S. J. Kiebel, J. B. Poline, K. J. Friston, A. P. Holmes, and K. J. Worsley, "Robust smoothness estimation in statistical parametric maps using linear model," Neuroimage, standardized residuals from the general vol. 10, no. 6, pp. 756 -- 766, Dec 1999. [5] R. M. Everson, A. K. Prashanth, M. Gabbay, B. W. Knight, L. Sirovich, and E. Kaplan, "Representation of spatial frequency and orientation in the visual cortex," Proc. Natl. Acad. Sci. U.S.A., vol. 95, no. 14, pp. 8334 -- 8338, Jul 1998. [6] T. P. Patel, K. Man, B. L. Firestein, and D. F. Meaney, "Automated quantification of neuronal networks and single-cell calcium dynamics using calcium imaging," J. Neurosci. Methods, vol. 243, pp. 26 -- 38, Mar 2015. [7] C. Park, N. A. Lazar, J. Ahn, and A. Sornborger, "A multiscale analysis of the temporal characteristics of resting-state fMRI data," J. Neurosci. Methods, vol. 193, no. 2, pp. 334 -- 342, Nov 2010. [8] A. Sornborger, J. Broder, A. Majumder, G. Srinivasamoorthy, E. Porter, S. S. Reagin, C. Keith, and J. D. Lauderdale, "Estimating weak ratio- metric signals in imaging data. II. Meta-analysis with multiple, dual- channel datasets," J Opt Soc Am A Opt Image Sci Vis, vol. 25, no. 9, pp. 2185 -- 2194, Sep 2008. [9] D. Thomson, "Spectrum estimation and harmonic analysis," Proc. IEEE, vol. 70, pp. 1055 -- 1096, 1982. [10] A. T. Sornborger and T. Yokoo, "A multivariate, multitaper approach to detecting and estimating harmonic response in cortical optical imaging data," J. Neurosci. Methods, vol. 203, no. 1, pp. 254 -- 263, Jan 2012. [11] D. Brillinger, Ed., Time Series: data analysis and theory. Philadelphia, USA: Society for Industrial and Applied Mathematics, 2001. [12] A. Walden, "A unified view of multitaper multivariate spectral estima- tion," Biometrika, vol. 87, pp. 767 -- 788, 2000. [13] D. Percival and A. Walden, Eds., Spectral analysis for physical applica- tions: Multitaper and conventional univariate techniques. Cambridge, UK: Cambridge University Press, 1993. [14] Mitra, P.P. and Bokil, H., Observed Brain Dynamics. Oxford, UK: Oxford University Press, 2008. [15] Lahiri, S.N., Resampling Methods for Dependent Data. New York, NY: Springer, 2012.
1802.02523
1
1802
2018-01-16T01:51:18
Plasma Brain Dynamics (PBD): A Mechanism for EEG Waves Under Human Consciousness
[ "q-bio.NC", "physics.med-ph" ]
EEG signals are records of nonlinear solitary waves in human brains. The waves have several types (e.g., a, b, g, q, d) in response to different levels of consciousness. They are classified into two groups: Group-1 consists of complex storm-like waves (a, b, and g); Group-2 is composed of simple quasilinear waves (q and d). In order to elucidate the mechanism of EEG wave formation and propagation, this paper extends the Vlasov-Maxwell equations of Plasma Brain Dynamics (PBD) to a set of two-fluid, self-similar, nonlinear solitary wave equations. Numerical simulations are performed for different EEG signals. Main results include: (1) The excitation and propagation of the EEG wave packets are dependent of electric and magnetic fields, brain aqua-ions, electron and ion temperatures, masses, and their initial fluid speeds; (2) Group-1 complex waves contain three ingredients: the high-frequency ion-acoustic (IA) mode, the intermediate-frequency lower-hybrid (LH) mode, and the low-frequency ion-cyclotron (IC) mode; (3) Group-2 simple waves fall within the IA band, featured by one or a combination of the three envelopes: sinusoidal, sawtooth, and spiky/bipolar. The study proposes an alternative model to Quantum Brain Dynamics (QBD) by suggesting that the formation and propagation of the nonlinear solitary EEG waves in the brain have the same mechanism as that of the waves in space plasmas
q-bio.NC
q-bio
Cosmos and History: The Journal of Natural and Social Philosophy, vol. 13, no. 2, 2017 PLASMA BRAIN DYNAMICS (PBD): A MECHANISM FOR EEG WAVES UNDER HUMAN CONSCIOUSNESS John Z. G. Ma ABSTRACT: EEG signals are records of nonlinear solitary waves in human brains. The waves have several types (e.g., α, β, γ, θ, δ) in response to different levels of consciousness. They are classified into two groups: Group-1 consists of complex storm-like waves (α, β, and γ); Group-2 is composed of simple quasilinear waves (θ and δ). In order to elucidate the mechanism of EEG wave formation and propagation, this paper extends the Vlasov-Maxwell equations of Plasma Brain Dynamics (PBD) to a set of two-fluid, self-similar, nonlinear solitary wave equations. Numerical simulations are performed for different EEG signals. Main results include: (1) The excitation and propagation of the EEG wave packets are dependent of electric and magnetic fields, brain aqua-ions, electron and ion temperatures, masses, and their initial fluid speeds; (2) Group-1 complex waves contain three ingredients: the high-frequency ion-acoustic (IA) mode, the intermediate-frequency lower-hybrid (LH) mode, and, the low-frequency ion-cyclotron (IC) mode; (3) Group-2 simple waves fall within the IA band, featured by one or a combination of the three envelopes: sinusoidal, sawtooth, and spiky/bipolar. The study proposes an alternative model to Quantum Brain Dynamics (QBD) by suggesting that the formation and propagation of the nonlinear solitary EEG waves in the brain have the same mechanism as that of the waves in space plasmas. KEYWORDS: Plasma brain dynamics (PBD); Quantum brain dynamics (QBD); Consciousness; EEG; nonlinear solitary wave 1. INTRODUCTION Consciousness resides mainly in the outer layer of the cerebrum, cerebral cortex, with a www.cosmosandhistory.org 185 COSMOS AND HISTORY 186 interconnected with each other with each neuron to link with up to 104 other neurons, forming a highly intricate system to pass signals via as many as 1000 trillion synaptic Consciousness is dominated by the prefrontal cortex in the brain neuronal system to advance has been reached toward understanding the measured low-frequency, nonlinear, electromagnetic brain waves in diagnosis of EEG (electroencephalography) thickness of (2~5)×10-3 m and a surface area of 0.16~0.4 m2,1 giving a volume of (3.2~20)×10-4 m3. The adult male human brain of an average of 1.5 kg has 86 billion neurons (nerve cells) and 85 billion non-neuronal cells.2 The average volume density of neurons turns out to be 4.3×1013 ~2.7×1014 neurons /m3.3 Cortical neurons are connections.4 express the brain cognitive ability.5 After the mathematical brain model was suggested in the early 1940s,6 Quantum Brain Dynamics (QBD) was developed from Ricciardi and Urnezawa's pioneer work7 to account for the neuro-and-cognitive mechanism of human consciousness.8 QBD deals with a couple of basic fields: (1) the water (H2O) rotational field, and (2) the electromagnetic field.9 However, in spite of great progress having been achieved hitherto with sophisticated theories, models, or numerical simulations,10 little or MEG (magnetoencephalography),11 not mentioning the qualitative or quantitative 1 Nunez PL, Srinivasan R. 2006. Electric fields of the brain: The neurophysics of EEG, 2nd ed. Oxford: Oxford University Press, p.6. 2 Herculano-Houzel S. (1) 2009. The human brain in numbers: A linearly scaled-up primate brain. Front. Human Neurosci. 3, 31, pp.1-11; (2) 2016. The human advantage: A new understanding of how our brain became remarkable. Cambridge, MA: MIT Press, p.79. 3 Teplan M 2002. Fundamentals of EEG measurement. Measurement Sci. Rev. 2, 2, pp.1-11. 4 Mastin L 2010. Neurons & synapses. In: The human memory. http://www.human- memory.net/brain_ neurons.html 5 Gabi M, Neves K, Masseron C, et al 2016. No relative expansion of the number of prefrontal neurons in primate and human evolution. PNAS, 113, 34, 9617-9622. 6 McCulloch WS, Pitts W 1943. A logical calculus of the ideas immanent in nervous activity. 5, pp.115-133; Reprint: 1990. Bull. Math. Biol. 52, 1/2, pp.99-115. 7 Ricciardi LM, Umezawa H 1967. Brain and physics of many-body problems. Kybernetik, 4, 2, pp.44-48. 8 Vitiello G 2011. Hiroomi Umezawa and quantum field theory. NeuroQuantology, 9, 3, pp.402- 412. 9 Jibu M, Yasue K. 1995. Quantum brain dynamics. In: Quantum brain dynamics and consciousness: An introduction. Amsterdam: John Benjamins Publishing, pp.163-166. 10 E.g., (1) Başar E 2010. From quantum mechanics to the quantum brain. NeuroQuantology, 8, 3, pp.319-321. (2) Hameroff S 2012. How quantum brain biology can rescue conscious free will. Front Integr Neurosci. 6, 93, pp.1-17. (3) Sakane S, Hiramatsu T, Matsui T 2016. Neural network for quantum brain dynamics: 4D CP1+U(1) gauge theory on lattice and its phase structure. arXiv:1610.05443v1 [cond-mat.dis-nn]. 11 (1) For a detailed review, see: van Vugt M 2002. Oscillations in the brain: A dynamic memory model. Thesis, University College Utrecht. http://www.ai.rug.nl/~mkvanvugt/UCUthesis.pdf. (2) JOHN Z.G. MA 187 data-fit visualizations of the holistic behavior of neuronal functions. To be more intriguingly, different from the pioneer recognitions in QBD, e.g., Penrose's neural "firing and not firing,"12 and Penrose-Hameroff's neural microtubules,13 new regular neuron firing and for kink-like polarization excitations in microtubules.14 calculations of the neural de-coherence rates indicated that consciousness of the human brain should be thought of as a "classical rather than quantum" neural process, both for Therefore, it is necessary to review the physics in consciousness. Fortunately, an alternative model, Plasma Brain Dynamics (PBD), was proposed in the early 1970s by Hokkyo to deal with the collective features of the cortical phase transitions in neural populations of the living brain.15 The preliminary study was based on plasma kinetics and made use of a quasi-linear approximation to the collisional Liouville equation to derive the distribution function, F (r,v), of brain activities in the neural phase space (i.e., 3D position space r and 3D velocity space v). Different from Ricciardi and Urnezawa's result that the stable memory of the brain contributed by the Bose condensation into the ground state of the quantum many-body system of Goldstone particles, Hokkyo concluded that a brain is (1) awaken but inattentive for a single-humped F peaked at v=0; (2) awaken but attentive for F to possess a small second hump on the tail; and, (3) experiencing a memorizing process when the second hump is flattening to form a plateau which persists for indefinitely long time. The brain plasma system can be treated as a collision-free media where neuronal activities could be mediated by long-range extracellular flows16 and the collective behavior of the neuronal network could be described by equations of collision-free Musso F, Brinkmeyer J, Mobascher A, et al 2010. Spontaneous brain activity and EEG microstates: A novel EEG/fMRI analysis approach to explore resting-state networks. Neuroimage, 52, pp.1149- 1161. (3) Atasoy S, Donnelly I, Pearson J 2016. Human brain networks function in connectome- specific harmonic waves. Nature Communications, 7, 10340. 12 Penrose R 1989. The emperor's new mind: concerning computers, minds and the laws of physics. Oxford: Oxford University Press. 13 Hameroff S, Penrose R 2003. Conscious events as orchestrated space-time selections. NeuroQuant. 1, pp.10-35. 14 Tegmark M 2000. Importance of quantum decoherence in brain processes. Phys Rev E 61, 4 Pt B, pp.4194-4206. 15 Hokkyo N 1972. A plasma model of brain dynamics. Prog. Theoret. Phys., 48, 4, pp.1191-1195. 16 (1) Linkenkaer-Hansen K, Nikouline VV, Palva JM, Ilmoniemi RJ 2001. Long-range temporal correlations and scaling behavior in human brain oscillations. J Neurosci, 21, 4, pp.1370-1377. (2) Vuksanovic V, Hövel P 2014. Functional connectivity of distant cortical regions: Role of remote synchronization and symmetry in interactions. NeuroImage, 97, pp.1-8. COSMOS AND HISTORY 188 set of Vlasov-Maxwell equations to evaluate brain functions by merely taking ion movements on the basis of stochastic analyses.17 Such a system can be described by a dynamics into consideration18 in the nervous extracellular space where the interstitial occupies about 85% of the brain volume.19 Because the extracellular electric signals, like fluid is in contact with the cerebrospinal fluid from the ventricular surfaces; surrounded by the extracellular space there exists larger intracellular neuronal compartment which microscopic-level spiking activity of neuronal assemblies, mesoscopic-level local field potential (LFP, also known as 'micro-EEG'), and macroscopic-level EEG, provide insights into the cooperative behavior of neurons, their average synaptic input and their spiking output,20 we formulate collision-free plasma dynamics from Vlasov-Maxwell equations in this paper to focus on the mechanism of the EEG waves via data-fit modeling for the neuronal activity of the brain consciousness. It deserves a special mention here that any results of EEG analyses can be used for MEG due to the fact that the two types of signals have the same source of excitation, i.e., the ionic currents generated by biochemical processes at the cellular level, but recorded respectively in response to electric and magnetic effects of the brain activities.21 17 Touboul J 2012. Mean-field equations for stochastic firing-rate neural fields with delays: Derivation and noise-induced transitions. Phys D: Nonlin Phenomena, 241, 15, pp.1223-1244. 18 Tozzi A, Peters JF 2016. Towards plasma-like collisionless trajectories in the brain. viXra, 1610.0014v1, pp.1-10. http://vixra.org/pdf/1610.0014 v1.pdf 19 Syková E, Nicholson C 2008. Diffusion in brain extracellular space. Physiol Rev, 88, 4, pp.1277- 1340. 20 Buzsáki G, Anastassiou CA, Koch C 2012. The origin of extracellular fields and currents -- EEG, ECoG, LFP and spikes. Nature Rev, Neurosci, 13, pp.407-420. 21 da Silva FL 2013. EEG and MEG: Relevance to Neuroscience. Neuron, 80, 5, pp.1112-1128. JOHN Z.G. MA 189 Figure 1. EEG signals in a rest state with closed eyes (adapted from Figure 2 in ref.22) Figure 1 gives an example of the propagating EEG wave packages in a typical brain.22 It exposes following features of consciousness: (1) There exist five levels of consciousness (c.f., Thompson23): Semiconscious (bottom panel, δ: 0.5-4 Hz; coma, dreamless-sleeping); Subconscious (lower middle panel, θ: 4-8 Hz; drowsy, idling, dreaming, deep-meditation); Conscious (middle panel, α: 8-14 Hz; relaxed, reflecting, light-meditation, visualization); Ultraconscious (upper middle panel, β: 14-30 Hz; perception, alerting, concentration); and, Superconscious (top panel, γ: 30-42 Hz; focus, religious ecstasy). itself extrinsically with nonlinear, solitary (2) Consciousness manifests electromagnetic waves which are divided into a couple of groups: Group-1 consists of complex amplitude-modulated stormy waves24 (as given in the upper three panels: α, β, and γ) which are characterized by the superimposition of multi-frequency waveforms; and, Group-2 is of simple 22 Campisi P, La Rocca D, Scarano G 2012. EEG for automatic person recognition. Computer, 45, pp.87-89. 23 Thompson E 2015. Dreamless Sleep, the embodied mind, and consciousness. In: T Metzinger & JM Windt am Main: MIND Group. doi: 10.15502/9783958570351, pp.1-19. 24 Ma JZG, Hirose A 2010. Lower-hybrid (LH) oscillitons evolved from ion-acoustic (IA)/ion- cyclotron (IC) solitary waves: Effect of electron inertia. Nonlin. Proc. Geophys. 17, pp.245-268. (Eds). Open MIND, 37(T), Frankfurt COSMOS AND HISTORY 190 quasilinear waves25 (as given in the lower two panels: θ and δ) which are characterized by envelopes between sinusoidal and saw-tooth waveforms of one-or-two dominant frequencies. (3) The fluctuation of waveforms is approximately symmetric to the resting potential (0 µV), indicating that the brain functioning is bounded to, or, at least can be reduced to, a kind of two-species system of oppositely charged particles which are maintained by billions of neurons.26 The dynamical interaction or competition of the two ingredients in the system pumps the excitation, development, and propagation of electric (or magnetic) signals which are able to be measured as EEG (or MEG) waveforms. As the first study on the EEG mechanism through formulating the new PBD paradigm, the purpose of this paper lies in gaining substantial insights into the complicated features of the real-time EEG oscillations in realistic situations. The work will provide a reference to illustrate the synaptic, coherent propagating flows in the nervous extracellular space of the brain. No similar studies have yet been reported so far in literature. To reduce the complexity of the problem while still being able to develop a tenable approach toward the distinctive nonlinear characteristics of EEG signals of the brain, we assume a basic two-component plasma system consisting of positively and negatively charged ions. That is to say, all the ions with the same polarity in the brain are reduced to one-type, singly- charged species with a reduced mass, mi, which is around tens of proton mass, mp. In addition, we suggest that all of the test particles of the brain plasma under modeling are well inside the extracellular space thereby being able to neglect all the edge effects in a 3D Cartesian frame of reference. Furthermore, we neglect the relativistic effects of all the test particles considering the fact that the globally coherent speeds of the brain activity occurred in the five levels of consciousness are no more than Vc ~ 5 m/s.27 specifically, the travelling speeds of the EEG signals were estimated with values, on average, of α: 6.5±0.9 m/s; θ: 4.0±0.9 m/s; as well as all of the signals fall within 3~11 m/s.28 Previous studies show that not only the axonal actions of the neuronal system are similar to the scaled equivalents of plasma lightning,29 but also the 25 Ma JZG, Hirose A 2009. Parallel propagation of ion solitons in magnetic flux tubes. Phys. Scr., 79: 045502. 26 Herculano-Houzel S 2009. The human brain in numbers: a linearly scaled-up primate brain. Front. Hum. Neurosci., 3, 31, pp.1-11. 27 Alexander DM, Nikolaev AR, Jurica P, et al 2016. Global Neuromagnetic Cortical Fields Have Non-Zero Velocity. PLoS ONE, 11, 3, e0148413. 28 Patten TM, Rennie CJ, Robinson PA, Gong P 2012. Human cortical traveling waves: Dynamical properties and correlations with responses. PLoS ONE. 7, 6, e38392, pp.1-10. 29 Persinger MA 2012. Brain electromagnetic activity and lightning: potentially congruent scale- invariant quantitative properties. Front Integr Neurosci, 6, 19, pp.1-7. JOHN Z.G. MA 191 2. BRAIN PLASMA MODEL AND TWO-SPECIES VLASOV-MAXWELL EQUATIONS cerebral cortex and its white matter system of corticocortical fibers turns out to be a system somewhat analogous to the earth's ionospheric shell.30 We are thus inspired to extend our work on space plasma dynamics applicable to brain plasma dynamics, with the first step to elucidate the mechanism of EEG signals. The layout of the paper is as follows: Section 2 estimates brain plasma parameters and set up a set of two-fluid Vlasov-Maxwell equations; Section 3 derives a set of nonlinear, self-similar differential equations for the excitation and propagation of EEG waves; Section 4 exposes the data-fit modeling of the solitary EEG waves under different conditions. The last section gives conclusions along with some concise discussions. SI units are used throughout the paper except wherever the conventional units are more conveniently used. Among the neurons of the cerebral cortex, there always exist transmissions of impulses which induce dendritic synapses to drive excitatory and inhibitory postsynaptic potentials. The triggered currents move through dendrites and cell body to the axon base, and pass through the membrane to the extracellular space. It was suggested that the superimposition of the potentials derived from the mixture of the extracellular currents generated by those neurons with uniformly oriented dendrites gives rise to EEG signals,31 which has a characteristic magnitude of the electric field, Ec = 2 mV/m for an order of ~200 µV with a typical distance of 10 cm.32 The involved ions consist mostly of small ions like H+, Na+, K+, Ca2+, and Cl-.33 In both the intracellular and extracellular spaces, the concentration of negative ions (124.0 mM) is far less than that of positive ones (317.5 mM), giving the charge number densities of n+ ≈ 1.9×1026 m-3, and n- ≈39% n+,34 with n+ ~ 1/1000 of the molecular number density of water or the free electron density in copper. Because the brain is electrically neutral (i.e., as many positive charges as negative charges), the excess positive charges in the brain should be balanced by 30 Kozlowski M, Marciak-Kozlowska J 2012. On the Temperature and Energy of the Brain Waves: Is there Any Connection with Early Universe? NeuroQuantology, 10, 3, pp.443-452. 31 Freeman WJ 1992. Tutorial on neurobiology: From single neurons to brain chaos. Int. J. Bifurcation Chaos, 02, pp.451-482, 32 Nunez PL, Srinivasan R 2006. Electric fields of the brain: The neurophysics of EEG. 2nd Ed. Oxford: Oxford University Press. p.12,23. 33 Jibu M, Yasue K 1995. Quantum Brain Dynamics and Consciousness: An Introduction. Amsterdam: John Benjamins Publishing. p.99. 34 Phillips R, Kondev J, Theriot J 2013. Physical biology of the cell. 2nd Ed. Chapter 17: Biological electricity and the Hodgkin-Huxley model. New York: Garland Science. Table 17.1. p.685. COSMOS AND HISTORY 192 the abundant electrons which come from macromolecules such as nucleic acids and proteins in the brain (c.f. p.685 in ref.33). in this case, this paper assumes a model to the first order which treats the negative ions as a kind of "dust" ingredient in the brain, and neglect their contributions to reduce the complexity of the study, while still being able to obtain significant solutions toward elucidating the dominant EEG features as shown in Figure 1. With this simplification, we treat the brain plasma consists of negatively charged electrons only and singly charged one- species ion with a reduced mass mi from all of the ions. On the one hand, the E×B drift of brain plasma particles in the typical external geomagnetic field, Bext ~ 5×10-5 T, has a characteristic speed, Vd = Ec / Bext ~ 40 m/s, which applies for all particles including electrons, positive and negative ions. On the other hand, the brain aqua ions, [M(H2O)n]z+ (where parameters M, n, and z are a metal atom, solvation number, and charge number, respectively), has a typical thermal speed, 𝑉𝑉𝑇𝑇𝑇𝑇=𝜂𝜂�8𝑘𝑘𝑏𝑏𝑇𝑇𝑇𝑇0/𝜋𝜋𝜋𝜋𝑇𝑇 =105 m/s (where kb =1.38 ×10-23 J/K, Ti0 = 310 K, mi =148mp=2.48×10-25 kg (mp is the proton mass) for M=Ca and n=6; with η~0.5, a relaxation coefficient due to de-coherence), while brain electrons are of 𝑉𝑉𝑇𝑇𝑇𝑇=�8𝑘𝑘𝑏𝑏𝑇𝑇𝑇𝑇0/𝜋𝜋𝜋𝜋𝑇𝑇=109 km/s for the same temperature Ti0. Thus, Vd ~ VTi ≪ VTe, suggesting that the electric field is more heavily correlated with the dynamics of ions rather than that of electrons. If Bext decreases and Vd is competitive to VTe, electron dynamics will certainly be different. This paper does not deal with large E×B drift case which is used for astronauts or those under clean magnetic environment. The two types of particles are therefore described by different collision- free Vlasov-Maxwell equations. Reduced from the collisional Boltzmann equation, the mandatory set of the Vlasov-Maxwell equations for ions and electrons are expressed jointly as follows:15,35 (1) in which t is time, subscript α = (i,e) denote ion and electron species, respectively; and 35 Boyd TJM, Sanderson JJ 2003. The physics of plasmas. Cambridge: Cambridge University Press. pp.253-255. ⎩⎪⎪⎨⎪⎪⎧�𝜕𝜕𝜕𝜕𝜕𝜕+𝐯𝐯∙∇+𝐚𝐚𝛼𝛼∙∇𝐯𝐯�𝑓𝑓𝛼𝛼=0 ∇×𝐄𝐄=−𝜕𝜕𝐁𝐁𝜕𝜕𝜕𝜕 ∇×𝐁𝐁=µ𝐣𝐣 ∇∙𝐄𝐄=𝑞𝑞𝜀𝜀 ∇∙𝐁𝐁=0 JOHN Z.G. MA 193 𝛼𝛼 , 𝐣𝐣=�𝑒𝑒𝛼𝛼�𝐯𝐯𝑓𝑓𝛼𝛼d𝐯𝐯 3. TWO-FLUID, SELF-SIMILAR NONLINEAR SOLITARY WAVE EQUATIONS IN BRAIN PLASMA 𝐚𝐚𝛼𝛼=𝑒𝑒𝛼𝛼𝜋𝜋𝛼𝛼(𝐄𝐄+𝐯𝐯×𝐁𝐁) , 𝑞𝑞=�𝑒𝑒𝛼𝛼�𝑓𝑓𝛼𝛼d𝐯𝐯 𝛼𝛼 where 𝑒𝑒𝑇𝑇=−𝑒𝑒 and 𝑒𝑒𝑇𝑇=+𝑒𝑒. In addition, 𝜀𝜀=𝜀𝜀𝑟𝑟𝜀𝜀0~80𝜀𝜀0 (p.118: ref.31) and 𝜇𝜇= 𝜇𝜇𝑟𝑟𝜇𝜇0~10𝜇𝜇0,36 in which 𝜀𝜀0 and 𝜇𝜇0 are the electric permittivity and magnetic permeability of free space, respectively. EEG waves are the collective manifestation of the brain plasma particles in the presence of both the internal brain electric & magnetic fields and the external geomagnetic field. The elucidation of the collective features depends on the fluid formulations of the brain plasma system. By introducing density n, velocity u, and pressure p (the diagonal value of tensor 𝐏𝐏��⃗) as follows:37 𝑛𝑛𝛼𝛼=�𝑓𝑓𝛼𝛼d𝐯𝐯 , 𝐮𝐮𝛼𝛼=1𝑛𝑛𝛼𝛼�𝐯𝐯𝑓𝑓𝛼𝛼d𝐯𝐯 , 𝐏𝐏��⃗𝛼𝛼=𝜋𝜋𝛼𝛼�(𝐯𝐯−𝐮𝐮)(𝐯𝐯−𝐮𝐮)𝑓𝑓𝛼𝛼d𝐯𝐯 Equation (1) provides the following self-similar set of two-fluid equations for the brain plasma system by assuming a slab model where the propagation of EEG signals is along the x-coordinate in the Cartesian frame (x,y,z): For the electron fluid: ⎩⎪⎪⎨⎪⎪⎧�𝑢𝑢𝑇𝑇𝑒𝑒−𝜉𝜉𝑚𝑚𝑢𝑢𝑇𝑇𝑒𝑒�𝑑𝑑𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑑𝑑 =−�𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑧𝑧−𝑢𝑢𝑇𝑇𝑧𝑧𝐵𝐵𝑒𝑒� 𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑑𝑑 =−�𝑆𝑆𝑒𝑒−𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑧𝑧+𝑢𝑢𝑇𝑇𝑧𝑧𝐵𝐵𝑒𝑒0� 𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑢𝑢𝑇𝑇𝑧𝑧𝑑𝑑𝑑𝑑 =−�𝑆𝑆𝑧𝑧+𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑒𝑒−𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑒𝑒0� (2) For the ion fluid: 36 Georgiev DD 2003. Electric and magnetic fields inside neurons and their impact upon the cytoskeletal microtubules. Tech. Rep. http://cogprints.org/3190/ 37 Schunk RW, Nagy FA 2000. Ionospheres: physics, plasma physics, and chemistry. Cambridge: Cambridge University Press. pp.50-51. Cogprints Report, U. Southampton, UK, COSMOS AND HISTORY 194 ⎩⎪⎪⎨⎪⎪⎧𝜉𝜉𝑚𝑚�𝑢𝑢𝑇𝑇𝑒𝑒−3𝑆𝑆𝑛𝑛2 𝜉𝜉𝑇𝑇𝑢𝑢𝑇𝑇𝑒𝑒3�𝑑𝑑𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑑𝑑 =𝐸𝐸𝑒𝑒+𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑧𝑧−𝑢𝑢𝑇𝑇𝑧𝑧𝐵𝐵𝑒𝑒 𝜉𝜉𝑚𝑚𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑑𝑑 =𝑆𝑆𝑒𝑒−𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑧𝑧+𝑢𝑢𝑇𝑇𝑧𝑧𝐵𝐵𝑒𝑒0 𝜉𝜉𝑚𝑚𝑢𝑢𝑇𝑇𝑒𝑒𝑑𝑑𝑢𝑢𝑇𝑇𝑧𝑧𝑑𝑑𝑑𝑑 =𝑆𝑆𝑧𝑧+𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑒𝑒−𝑢𝑢𝑇𝑇𝑒𝑒𝐵𝐵𝑒𝑒0 (3) For electric and magnetic fields: 𝑑𝑑𝐸𝐸𝑒𝑒𝑑𝑑𝑑𝑑=−𝜉𝜉𝑚𝑚𝑛𝑛sc ⎩⎪⎪⎨⎪⎪⎧ �1−𝑀𝑀2𝜉𝜉V2𝜉𝜉𝑚𝑚 �𝑑𝑑𝐵𝐵𝑒𝑒𝑑𝑑𝑑𝑑=𝑆𝑆𝑛𝑛𝜉𝜉V2�𝑢𝑢𝑇𝑇𝑧𝑧𝑢𝑢𝑇𝑇𝑒𝑒−𝑢𝑢𝑇𝑇𝑧𝑧𝑢𝑢𝑇𝑇𝑒𝑒� �1−𝑀𝑀2𝜉𝜉V2𝜉𝜉𝑚𝑚 �𝑑𝑑𝐵𝐵𝑧𝑧𝑑𝑑𝑑𝑑=−𝑆𝑆𝑛𝑛𝜉𝜉V2�𝑢𝑢𝑇𝑇𝑒𝑒𝑢𝑢𝑇𝑇𝑒𝑒−𝑢𝑢𝑇𝑇𝑒𝑒𝑢𝑢𝑇𝑇𝑒𝑒� (4) in which ⎩⎨⎧𝑆𝑆𝑛𝑛=𝑢𝑢𝑒𝑒0−𝑀𝑀, 𝑆𝑆𝑒𝑒= 𝐸𝐸𝑒𝑒0−𝑀𝑀𝐵𝐵𝑧𝑧0 , 𝑆𝑆𝑧𝑧= 𝐸𝐸𝑧𝑧0+𝑀𝑀𝐵𝐵𝑒𝑒0 𝑛𝑛sc=𝑛𝑛𝑇𝑇−𝑛𝑛𝑇𝑇 , 𝑛𝑛𝑇𝑇=𝑆𝑆𝑛𝑛𝑢𝑢𝑇𝑇𝑒𝑒 , 𝑛𝑛𝑇𝑇=𝑆𝑆𝑛𝑛𝑢𝑢𝑇𝑇𝑒𝑒 𝐵𝐵𝑒𝑒=𝐵𝐵𝑒𝑒0 𝐸𝐸𝑒𝑒=𝑆𝑆𝑒𝑒+𝑀𝑀𝐵𝐵𝑧𝑧 , 𝐸𝐸𝑧𝑧=𝑆𝑆𝑧𝑧−𝑀𝑀𝐵𝐵𝑒𝑒 , (5) In Equations (2)~(5), subscript "0" refers to the equilibrium state, and X is the self-similar coordinate after the transformation of X=x−Mt,38 where 𝑀𝑀=𝑉𝑉ph/𝑐𝑐𝑠𝑠= �1+𝛾𝛾/𝜉𝜉𝑇𝑇2>1 is the Mach number which is independent of X (γ is the adiabatic index) and determined by Vph and 𝑐𝑐𝑠𝑠=𝑉𝑉𝑇𝑇𝑇𝑇�𝛾𝛾𝜋𝜋/8/𝜂𝜂, the phase speed and the local sound speed of the ion acoustic (IA) waves [Eq.(10) in ref.39]. All of the parameters are dimensionless with normalizations of n by n0; (x,y,z) by electron Debye length λDe; u by cs; p by p0; B by a pseudo-magnetic field B0=meωpi/e (ωpi is the ion plasma frequency, ); E by E0=csB0. In addition, 𝜉𝜉𝑚𝑚=𝜋𝜋𝑇𝑇/𝜋𝜋𝑇𝑇 is ratio between ion and electron masses, 𝜉𝜉𝑇𝑇=𝑇𝑇𝑇𝑇0/𝑇𝑇𝑇𝑇0=(𝑝𝑝𝑇𝑇0𝑛𝑛𝑇𝑇0)/(𝑝𝑝𝑇𝑇0𝑛𝑛𝑇𝑇0) is the ratio between electron and ion temperatures, and 𝜉𝜉V=𝜐𝜐𝑇𝑇𝑇𝑇√𝜀𝜀𝑟𝑟𝜇𝜇𝑟𝑟/𝑐𝑐~28𝜐𝜐𝑇𝑇𝑇𝑇/𝑐𝑐 is the ratio between electron thermal speed and the speed of light in the brain, where c is the speed of light in free space. Note that 𝜀𝜀0𝐸𝐸02/𝑝𝑝𝑇𝑇0=(𝜋𝜋𝑇𝑇/𝜋𝜋𝑇𝑇)2. This set of nonlinear equations was derived previously to describe the satellite-measured coherent solitary waves excited in the two-fluid system in 38 Lee LC, Kan JR 1981. Nonlinear ion-acoustic waves and solitons in a magnetized plasma, Phys. Fluids 24, pp.430-436. JOHN Z.G. MA 195 4. MECHANISM OF EEG WAVES space plasmas.39 The only difference of the two-fluid formulations between the space plasma and brain plasma situations arise from the electric permittivity and magnetic permeability: in the former, the two parameters are those of the free space, 𝜀𝜀0 and 𝜇𝜇0; by contrast, in the latter, they are updated by 𝜀𝜀 and 𝜇𝜇. This fact brings us to make use of the data-fit modeling results in the classical ionospheric and magnetospheric physics to reexamine the physics of consciousness via EEG signals. The approach will thus provide an alternative paradigm to QBD in the study of the neural processes. Under different electron conditions, Equations (2)~(5) describe both stormy EEG waves (Group 1: α, β, and γ) and simple EEG waves (Group 2: θ and δ) in Figure 1. The boundary conditions of the simulations at X=0 satisfy: 𝑛𝑛𝑇𝑇≈𝑛𝑛𝑇𝑇= 𝑛𝑛0, 𝑢𝑢𝑇𝑇𝑒𝑒=𝑢𝑢𝑇𝑇𝑒𝑒=𝑈𝑈0−𝑀𝑀, 𝐸𝐸𝑒𝑒=𝐸𝐸𝑒𝑒0, 𝐸𝐸𝑧𝑧=𝐸𝐸𝑧𝑧0, 𝐵𝐵𝑒𝑒=𝐵𝐵𝑒𝑒0, 𝐵𝐵𝑒𝑒=𝐵𝐵𝑒𝑒0, and 𝐵𝐵𝑧𝑧=𝐵𝐵𝑧𝑧0. Note that in linear IA regime, once 𝜉𝜉𝑇𝑇 is given, 𝑀𝑀 can be easily obtained; by contrast, in the nonlinear regime, the decoupled two parameters contribute jointly to the formation of solitary waves. Because the mass of the negatively charged electrons, 𝜋𝜋𝑇𝑇=9.1094×10−31 kg, is far smaller than that of the positively charged ions, 𝜋𝜋𝑇𝑇, which is tens of the mass of protons, 𝜋𝜋𝑝𝑝=1836𝜋𝜋𝑇𝑇=1.6726×10−27kg, it is reasonable to neglect the electron inertia in the complicated set of the nonlinear Equations (2)~(5) by assuming a zero electron mass. In this case, the dynamics of electrons are neglected and the simplest nonlinear solitary waves are obtained in the IA mode propagating in the direction parallel to the field lines of the external magnetic field, Bext. A detailed analysis of the IA solitary waves was given in a little more complicated case,40 where two types of isothermal Boltzmann electrons are included, one type is background electrons and the other is energetic ones. For the simpler case with only one type of background electrons, the reduced set of Equations (2)~(5) is as follows [Equation (9) in ref.39]: 39 Ma JZG, Hirose A 2010. Lower-hybrid (LH) oscillitons evolved from ion-acoustic (IA) / ion- cyclotron (IC) solitary waves: effect of electron inertia. Nonlin. Processes Geophys. 17, pp.245–268. 40 Ma JZG, Hirose A 2009. Parallel propagation of ion solitons in magnetic flux tubes, Phys. Scripta. 79, 045502, pp. 1-13. 4.1 Simple EEG waves (Group 2) COSMOS AND HISTORY 196 𝑛𝑛 2 𝑑𝑑2𝜑𝜑𝑑𝑑𝑑𝑑2=𝑒𝑒𝜑𝜑−� �1+ 1𝑀𝑀∗2�3𝜉𝜉𝑇𝑇−2𝜑𝜑��+��1+ 1𝑀𝑀∗2�3𝜉𝜉𝑇𝑇−2𝜑𝜑��2− 12𝜉𝜉𝑇𝑇𝑀𝑀∗2 (6) in which ϕ(X) is the normalized electrostatic potential in unit of kBTe0/e to simulate the output of the EEG measurements, and 𝑀𝑀∗=𝑈𝑈0−𝑀𝑀. Notice that the output of EEG signals is the potential drop, 𝜙𝜙(𝐸𝐸)=∑(𝑑𝑑𝜑𝜑/𝑑𝑑𝑑𝑑)⌋𝑛𝑛Δ𝑑𝑑𝑛𝑛 ≈ )𝑑𝑑=𝐸𝐸𝑑𝑑, rather than the potential ϕ(X), where d is the full distance between (∑ 𝐸𝐸𝑛𝑛𝑛𝑛 the two electrodes to measure the drop. Figure 2 presents an example of the 𝜙𝜙(𝐸𝐸) waveforms simulated under different sets of input parameters in the data-fit modeling. First of all, there exist three types of nonlinear solitary waves in view of the envelopes: sinusoidal (lower left panel), sawtooth (upper right panel), and spiky/bipolar (lower right panel). These nonlinear waveforms are developed from the non-wave structure (upper left panel) which is superimposed upon the background propagating linear IA waves and non-propagating IC oscillations (see Appendix B in ref.39 for the dispersion relations). Though the envelopes appear distinct among the three waves, previous FFT analyses have exposed that the dominant frequencies of the three solitary waves are all in the IA band, however, the number of harmonic ingredients is different from one type to the other. Take a look at Group 2 in Figure 1 and compare it with the three solitary waves in Figure 2. Not surprisingly, either θ or δ signals of the measured EEG waves turns out to be either one of the three types, or the composition of two or three of them. No doubt, it is evident that the simple EEG waves in Group 2 are the nonlinear IA solitary waves propagating in the brain. Besides, there exist only two specific components of ions which are responsible for the excitation and propagation of the brain IA solitary waves. One component is at a lower initial subsonic speed, 𝑈𝑈0𝑙𝑙, which moves opposite to X, and the other one is at a higher initial supersonic speed, 𝑈𝑈0ℎ, which moves along X. In either cases, the difference between the two speeds and the transonic Mach number, M=1.04 (for γ=3 and 𝜉𝜉𝑇𝑇=5.8), must be over-transonic with a value of 𝑀𝑀∗ to drive solitary waves. For 𝑀𝑀∗=1.3, 𝑈𝑈0𝑙𝑙=−0.26 and 𝑈𝑈0ℎ=2.34. Interestingly, as displayed in the two left panels of Figure 2, a small increase in 𝑀𝑀∗ from 1.3 to 1.4 propels the formation of the sinusoidal waves. In order words, there exists a minimum 𝑈𝑈0 for either the subsonic or supersonic ions lower than which no solitary waves are able to be excited to produce EEG signals. JOHN Z.G. MA 197 Figure 2. Simulations of simple EEG signals, 𝜙𝜙(𝐸𝐸) Furthermore, similar to the existence of the minimum 𝑈𝑈0, there exist a maximum initial value in potential ϕ(X), as shown in the upper two panels of the figure. At ϕ0=0.1 no solitary waves appear for 𝑀𝑀∗=1.3 and 𝜉𝜉𝑇𝑇=5.8; when ϕ0 decreases, coherent waveforms come into being. At ϕ0=0.01 a train of sawtooth envelopes is developed to exhibit the solitary structures of propagating waves. In comparison, the ratio between electron and ion temperatures, 𝜉𝜉𝑇𝑇, also plays a role in the modulation of solitary waves. For example, a minimum ratio exists only above which can waves be driven; in addition, as given in the two right panels, the waves become spiky with the increase of the ratio. It deserves mentioning that the direction of the wave propagation in simulations is opposite to that of the measured signals due to the different frames of reference. However, there do remain oppositely propagating solitary waves simultaneously in the curvilinear coordinates owing to the presence of centrifugal and Coriolis factors.41 41 Ma JZG 2010. Nonlinear ion-acoustic (IA) waves driven in a cylindrically symmetric flow. Astrophys Space Sci. 330, pp.87–94. COSMOS AND HISTORY 198 4.2 Complex EEG waves (Group 1) In the simple EEG waves, electrons are assumed to response to the external electric and magnetic fields instantly to keep the plasma neutrality at any time. However, this assumption so idealizes the real situations as not to offer a validated mechanism to account for the complex EEG signals as illustrated in Group 1 of Figure 1. We have to relax the constraint on the electron inertia and take into account the role played by the electron mass. In this case, the dynamics of both ion and electron particles are involved in the nonlinear set of the highly coupled Eq. (2)~(5). No analytical solutions can be derived as that given in Eq. (6) for the simple EEG waves. Numerical simulations are required to reveal the solitary structures which should certainly be modulated by the newly embedded ingredient to display unknown solitary wave packets. These packets should be different from the three simple modes (sinusoidal, sawtooth, and spiky/bipolar) in the IA regime. Figure 3. Simulations of complex EEG α-signals, 𝜙𝜙(𝐸𝐸) A typical set of input parameters used for simulations is taken as follows: mi/mp =148, giving 𝜉𝜉𝑚𝑚=271728; 𝜉𝜉𝑇𝑇=𝑇𝑇𝑇𝑇0/𝑇𝑇𝑇𝑇0=10, giving 𝜉𝜉V=0.03; and JOHN Z.G. MA 199 U0=(0.3,0,0), M=1, E0=(0,0,1), B0=(0.2,0,1). The results are given in Figures 3, 4, and 5 to expose the mechanism of how the EEG signals in Group-1, α, β, and γ, come into existence. Firstly, Figure 3 illustrates the general modulations of the input parameters on the EEG envelopes with an emphasis on the EEG α-signals. These parameters under consideration in the present study include the ion & electron mass ratio, mi/mp, their temperature ratio, 𝑇𝑇𝑇𝑇0/𝑇𝑇𝑇𝑇0, electron relative speed 𝜉𝜉V, as well as the initial speeds of plasma fluids along x, Ux0. The top panel presents the wave packets under the condition restricted by the introduced typical set of input parameters. The FFT power spectra in space plasmas39 shows that the oscillating frequencies contain following three bands: spiky high-frequency IA band, oscillating intermediate-frequency lower-hybrid (LH) band, and fluctuating low- frequency ion-cyclotron (IC) band. Though space and brain plasmas own different electric permittivity and magnetic permeability, the similar waveforms of the nonlinear solitary waves propagating in the two regions remind us at least qualitatively to assume that EEG signals have the same mechanism in these three wave bands. Quantitative studies are needed for the data-fit modeling in clinic applications. This is beyond the scope of the present study. COSMOS AND HISTORY 200 Figure 4. Simulations of complex EEG β-signals, 𝜙𝜙(𝐸𝐸) The top panel includes roughly two wave packets and 180 oscillations. The ratio of ~90 falls just in the range of that in IC and IA frequencies in ionospheric auroral F-layer.42 With a reduced mi/mp =40 in the upper middle panel, there is a little unnoticeable 6-period decrease in the oscillations, however, with an obvious increase in the magnitude of the wave packets from below 0.2 in the top panel to above 0.3 in this panel. When 𝑇𝑇𝑇𝑇0/𝑇𝑇𝑇𝑇0 enhances from 10 to 100, however, the lower middle panel exposes a totally different train which contains 26 solitary packets and every packet owns 6 IA oscillations. These packets should be related to LH frequencies to which electrons are involved in the modulations, while the IC oscillations are undiscernible due to the long wave lengths. Notice that 𝜉𝜉V is dependent of Te0. By contrast, a decrease in Ux0 from 0.3 in the top panel to 0.1 in the bottom panel reduces the IC wavelength and an additional packet appears. 42 Burchill JK, Knudsen DJ, Bock BJJ, et al 2004. Core ion interactions with BBELF, lower hybrid, and Alfven waves in the high-latitude topside ionosphere, J. Geophys. Res. 109, A01219. JOHN Z.G. MA 201 Secondly, since a brain contains ions with respective masses, Figure 4 particularly displays the impact of the mass on the solitary wave packets of 𝑇𝑇𝑇𝑇0/𝑇𝑇𝑇𝑇0=100 and Ux0=0.1, with an emphasis on the EEG β-signals. when mi/mp decreases from 148 in the top panel to 100 in the upper middle panel, the IA-LH train of the 21 wave packages 21 in the top panel does not have obvious changes. By contrast, after the ratio reduces to 50 in the lower middle panel, the low- frequency IC feature emerges, along with approximately doubled high-frequency IA oscillations. Strikingly, with the ratio goes down to 40, meaning the dominance of the atomic K+ and/or Ca2+, over the aqua-ions in the brain plasma dynamics, a kind of new train of the oscillating solitary packets comes to birth, named as nonlinear "oscillitons" in space physics.43 Oscillitons are embodied with a full spectra of oscillations from the lower IC end to the higher IA end. The frequency ratio in the case of the bottom panel is IC:LH:IA~1:10:60. Finally, a direct comparison between Figure 1 and Figures 3 & 4 let us aware of the fact that the α-type EEG signal has similar appearances presented jointly by all of the panels in Figure 3; and the β-type signal does with all of the panels in Figure 4. Therefore, it is feasible to have a parameterized study to quantitatively determine realistic dimensional parameters through data-fit modeling. For the highly stormy γ-type EEG signal, the production of the highly nonlinear oscillitons in the bottom panel of Figure 4 guides us to believe that only low mass ions are responsible for the origin of this special type of entities in the brain plasma. By simply adjusting Ux0 from 0.1 to 0.3, Figure 5 reconstructs the measured γ-envelopes given in the top panel of Figure 1. Though a preliminary result it is, we see that the simulations are capable of signifying evidently the irregular occurrence of the measured stormy train of the amplitude-modulated wave packets. Figure 5. Simulations of complex EEG γ-signals, 𝜙𝜙(𝐸𝐸) 43 Sauer K, Dubinin E, McKenzie JF 2001. New type of soliton in bi-ion plasmas and possible implications. Geophys. Res. Lett. 28, pp.3589–3592. COSMOS AND HISTORY 202 5. SUMMARY AND CONCLUSION According to the recorded wave forms, the EEG signals can be divided into two distinct groups. Group-1 owns highly nonlinear structures with a train of storm-like wave packets the amplitude of which are modulated violently to display complex envelopes. This group contains α, β, and γ types. By contrast, Group-2 is composed of quasilinear waves with deformed amplitudes deviated more or less from linear waves. This group includes θ and δ types. Since the plasma model of brain dynamics was formulated in the early 1970s,15 new advance has been reported toward the collision-free processes happening in the brain plasma system (e.g., ref.18). More importantly, QBD was suggested to explain the neuro- and-cognitive mechanism of human consciousness.8 However, little work is known to elucidate the mechanism of the electric EEG signals measured in the human brain under different situations of the mental consciousness. We derived a set of two-fluid, self-similar, nonlinear solitary wave equations from PBD's Vlasov-Maxwell equations. This model treats brain aqua-ions and electrons as two different fluids which are coupled with each other in the presence of the internal electric and magnetic fields and the external geomagnetic field. The E×B draft becomes a criterion to differentiate the effect of the plasma particles. By making use of the dimention-free formalism, we perform numerical simulations to fit with the five types of the EEG signals. Following results are obtained: (1) In the external geomagnetic field, Bext, the E×B speed in the brain is in the same order of the ion thermal speed, but much smaller than the electron thermal speed. Thus, brain electric field has little correlation with the dynamics of electrons. (2) The formation of the EEG waves is dependent of not only electric and magnetic fields, but also brain aqua-ions, electron and ion temperatures, masses, and their initial fluid speeds. Different sets of these input parameters contribute to different solitary wave packets. (3) When electron inertia is neglected, Group-2 simple waves come into being within the IA band. The waves are featured by one or a combination of the three envelopes: sinusoidal, sawtooth, and spiky/bipolar, among which the number of harmonic ingredients is different. (4) When electron inertia is considered, Group-1 complex waves emerge as the superimpositions of following two or three components: the high-frequency spiky IA mode, the intermediate-frequency oscillating LH mode, and, the low-frequency fluctuating IC mode. JOHN Z.G. MA 203 By introducing the different plasma brain model from QBD, the present paper proposes a new qualitative overview on the mechanism of the brain EEG waves. Though the waves in the brain are shown to have similar dynamical processes to those in space plasmas, quantitative studies are necessary in clinic diagnoses of measured EEG signals. [email protected] Philosophy, Cosmology and Consciousness California Institute of Integral Studies 1453 Mission St., San Francisco, CA 94103
1506.03157
2
1506
2015-11-05T14:15:51
A new framework for Euclidean summary statistics in the neural spike train space
[ "q-bio.NC", "stat.AP" ]
Statistical analysis and inference on spike trains is one of the central topics in the neural coding. It is of great interest to understand the underlying structure of given neural data. Based on the metric distances between spike trains, recent investigations have introduced the notion of an average or prototype spike train to characterize the template pattern in the neural activity. However, as those metrics lack certain Euclidean properties, the defined averages are nonunique, and do not share the conventional properties of a mean. In this article, we propose a new framework to define the mean spike train where we adopt a Euclidean-like metric from an $L^p$ family. We demonstrate that this new mean spike train properly represents the average pattern in the conventional fashion, and can be effectively computed using a theoretically-proven convergent procedure. We compare this mean with other spike train averages and demonstrate its superiority. Furthermore, we apply the new framework in a recording from rodent geniculate ganglion, where background firing activity is a common issue for neural coding. We show that the proposed mean spike train can be utilized to remove the background noise and improve decoding performance.
q-bio.NC
q-bio
The Annals of Applied Statistics 2015, Vol. 9, No. 3, 1278–1297 DOI: 10.1214/15-AOAS847 c(cid:13) Institute of Mathematical Statistics, 2015 A NEW FRAMEWORK FOR EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE By Sergiusz Wesolowski, Robert J. Contreras and Wei Wu Florida State University Statistical analysis and inference on spike trains is one of the cen- tral topics in the neural coding. It is of great interest to understand the underlying structure of given neural data. Based on the metric distances between spike trains, recent investigations have introduced the notion of an average or prototype spike train to characterize the template pattern in the neural activity. However, as those metrics lack certain Euclidean properties, the defined averages are nonunique, and do not share the conventional properties of a mean. In this article, we propose a new framework to define the mean spike train where we adopt a Euclidean-like metric from an Lp family. We demonstrate that this new mean spike train properly represents the average pat- tern in the conventional fashion, and can be effectively computed using a theoretically-proven convergent procedure. We compare this mean with other spike train averages and demonstrate its superiority. Furthermore, we apply the new framework in a recording from rodent geniculate ganglion, where background firing activity is a common is- sue for neural coding. We show that the proposed mean spike train can be utilized to remove the background noise and improve decoding performance. 1. Introduction. Neural spike trains are often called the language of the brain and are the focus of many investigations in computational neuro- science. Due to the stochastic nature of the spike discharge record, proba- bilistic and statistical methods have been extensively investigated to exam- ine the underlying firing patterns [Rieke et al. (1999), Brown et al. (2002), Kass, Ventura and Brown (2005), Box, Hunter and Hunter (1978), Kass and Ventura (2001)]. However, these methods focus only on parametric rep- resentations at each given time and therefore can prove to be limited in data-driven problems in the space of spike trains. Received June 2014; revised May 2015. Key words and phrases. Spike train metrics, Euclidean summary statistics, mean spike train, neural coding, geniculate ganglion, background noise removal. This is an electronic reprint of the original article published by the Institute of Mathematical Statistics in The Annals of Applied Statistics, 2015, Vol. 9, No. 3, 1278–1297. This reprint differs from the original in pagination and typographic detail. 1 2 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU Alternative approaches for analyzing spike train data are based on metri- cizing the spike train space. Over the past two decades, various methods have been developed to measure distances or dissimilarities between spike trains, for example, the distances in discrete state space, discrete time mod- els [Lim and Capranica (1994), MacLeod, Backer and Laurent (1998), Rieke et al. (1999)], those in discrete state space, continuous time models [Victor and Purpura (1996), Aronov et al. (2003), Aronov and Victor (2004), Victor, Goldberg and Gardner (2007), Wu and Srivastava (2011)], those in contin- uous state space, continuous time models [van Rossum (2001), Houghton and Sen (2008), Houghton (2009)], and a number of others [Schreiber et al. (2003), Kreuz et al. (2007), Quian Quiroga, Kreuz and Grassberger (2002), Hunter and Milton (2003), Paiva, Park and Pr´ıncipe (2009)]. An ongoing pursuit of great interest in computational neuroscience is defining an average that can represent tendency of a set of spike trains. What follows is the problem of defining basic summary statistics reflecting the intuitive properties of the mean and the variance, which are crucial for further statistical inference methods. In particular, to make the first-order statistic, mean, convenient for constructing new framework and inference methods, it should satisfy the following properties: 1. The mean is uniquely defined in a certain framework. 2. The mean is still a spike train. 3. The mean represents the conventional intuition of average like in the Euclidean space. 4. The mean depends on exact spike times only, and is independent of the recording time period. 5. The mean can be computed efficiently. N PN Property 3 can be described as follows: given a set of N spike trains with each having K spikes, we denote these trains using vectors {(xi1, . . . , xiK)}N i=1. Then the mean spike train is expected to resemble 1 i=1(xi1, . . . , xiK ). In Victor and Purpura (1997), the authors considered a "consensus" spike train, which is the centroid of a spike train set (under the Victor–Purpura metric). This idea was further generalized in Diez, Schoenberg and Woody (2012) to a "prototype" spike train which does not have to belong to the given set of spike trains, but its spike times are chosen from the set of all recorded spike times. Recently, a notion of an "average" based on ker- nel smoothing methods was introduced in Julienne and Houghton (2013). In Wu and Srivastava (2011, 2013), the authors proposed an elastic metric on inter-spike intervals to define a mean directly in the spike train space. However, none of these approaches satisfies the 5 desirable properties listed above, and therefore may result in limited use in practical applications. In this article, we propose a new framework for defining the mean spike train. We adopt a recently developed metric related to an Lp family with EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE 3 p ≥ 1, which inherits desirable properties in the special case of p = 2 [Dubbs, Seiler and Magnasco (2010)]. This metric is a direct generalization of the Victor–Purpura metric, and we refer to it as a GVP (Generalized Victor– Purpura) metric. We will demonstrate that this new mean spike train satis- fies all aforementioned 5 properties. In particular, the new framework is the only one (over all available methods) that has desirable Euclidean proper- ties on the given spike times. Such properties are crucial for the definition of summary statistics such as the mean, variance, and covariance in the spike train space. In general, these 5 properties assure intuitiveness of the sum- mary statistics, as well as efficiency in their estimation. In contrast, previous methods have issues such as nonuniqueness, dependence on model assump- tions, or more complicated computations, and therefore do not result in the same level of performance (see the detailed comparison in Section 2.5). One direct application of the mean spike train is in neural decoding in the rodent peripheral gustatory system [Wu et al. (2013)]. The neural data was recorded from single cells in geniculate ganglion, as the spiking activity in these neurons modulated with respect to different taste stimuli on the tongue. It is commonly known that spontaneous spiking activity can be observed even if only artificial saliva is applied. Thus, the neural response is actually a mixture of a background activity and a taste-stimulus activity. In this article we demonstrate using simulation as well as real data that the proposed new framework can be used to remove the background activity, which leads to improvement in decoding performance. In Section 2 we define the new framework by introducing the mean spike train under the GVP metric, and provide an efficient algorithm to estimate it. In Section 3 we extend this framework by developing a statistical ap- proach for noise removal and apply the method to the experimental data. We then discuss the merits of the new framework in Section 4. Finally, in the Appendix, we provide mathematical details on the convergence of the mean estimation algorithm. 2. New framework. Before we turn to describing the methods, it is nec- essary to set up a formal notation in the spike train space. 2.1. Notation. Assume S is a spike train with spike times 0 < s1 < s2 < · · · < sM < T , where [0, T ] denotes a recording time domain. We denote this spike train as S = (sj)M j=1 = (s1, s2, . . . , sM ). We define the space of all spike trains containing M spikes to be SM and M =0 SM . This can be equivalently the space of all spike trains to be S =S∞ described as a space of all bounded, finite, increasing sequences. 4 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU A time warping on the spike times (or inter-spike intervals) has been commonly used to measure distance between two spike trains [Victor and Purpura (1996), Dubbs, Seiler and Magnasco (2010), Wu and Srivastava (2011)]. Let Γ be the set of all time-warping functions, where a time warping is defined as an orientation-preserving diffeomorphism of the domain [0, T ]. That; that is, Γ = {γ : [0, T ] → [0, T ]γ(0) = 0, γ(T ) = T, 0 < γ(t) < ∞}. It is easy to verify that Γ is a group with the operation being the composition of functions. By applying γ ∈ Γ on a spike train S = (sj)M j=1, one obtains a warped spike train γ(S) = (γ(sj))M j=1. 2.2. GVP metric. In Dubbs, Seiler and Magnasco (2010), a new spike train metric was introduced with parameter p ∈ [1, ∞). This metric is a direct generalization of the classical Victor–Purpura (VP) metric (VP is a special case when p = 1), and we refer to it as the Generalized Victor– Purpura (GVP) metric. In particular, when p = 2, this metric resembles a Euclidean L2 distance. Assume that X = (xi)M j=1 are two spike trains in [0, T ]. For λ(> 0), the GVP metric between X and Y is given in the following form: i=1 and Y = (yj)N (1) dGVP[λ](X, Y ) = min γ∈Γ(cid:18)EOR(X, γ(Y )) + λ2 X{i,j:xi=γ(yj )} (xi − yj)2(cid:19)1/2 , where EOR(·, ·) denotes the cardinality of the Exclusive OR (i.e., union minus intersection) of two sets. That is, EOR(X, γ(Y )) measures the number of unmatched spike times between X and γ(Y ) and can be computed as EOR(X, γ(Y )) = M + N − 2 M N Xi=1 Xj=1 1{γ(yj )=xi}, where 1{·} is an indicator function. The constant λ(> 0) is the penalty co- efficient. We emphasize that dGVP is a proper metric, that is, it satisfies positive definiteness, symmetry, and the triangle inequality. It shares a lot of similarities with the classical L2 norm. Similarly to the result in Wu and Srivastava (2013), one can show that the optimal time warping between two spike trains X = (xi)M i=1 and Y = (yj)N j=1 must be a strictly increasing, piece-wise linear function, with nodes mapping from points in Y to points in X. Based on this fact, a dynamic programming algorithm was developed to compute the distance dGVP with the computational cost of the order of O(M N ). Using the bipartite graph matching theory, another efficient algorithm was also developed to compute dGVP in the cost of O(M N ) [Dubbs, Seiler and Magnasco (2010)]. EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE 5 2.3. Definition of the summary statistics and their properties. Conven- tional statistical inferences in the Euclidean space are based on basic quan- tities such as mean and variance. For statistical inferences in the spike train space, we can analogously use a Euclidean spike train metric to define these summary statistics as follows. For a set of spike trains S1, S2, . . . , SK ∈ S where the corresponding num- bers of spikes are n1, n2, . . . , nK (arbitrary nonnegative integers), respec- tively, we define their sample mean using the classical Karcher mean [Karcher (1977)] as follows: K (2) S∗ = argmin S∈S dGVP[λ](Sk, S)2. Xk=1 When the mean spike train S∗ is known, the associated (scalar) sample variance, σ2, can be defined in the following form: (3) σ2 = 1 K − 1 K Xk=1 dGVP[λ](Sk, S∗)2. The computation of this variance is straightforward, and the main challenge is in computing the mean spike train for any λ(> 0). Before we move on to the computational methods for the summary statis- tics, we list several basic theoretical properties of the mean spike trains using the dGVP metric. The proofs are omitted here to save space: 1. The optimal time warping between two spike trains must be a con- tinuous, increasing, and piece-wise linear function between subsets of spike times in these two trains. 2. Let spike trains X = (xi)M i=1, Y = (yi)M i=1 ∈ SM be defined on [0, T ]. If λ2 < 1/(M T 2), then dGVP[λ](X, Y ) = λ2 M (xi − yi)2!1/2 Xi=1 . 3. Assume a set of spike trains S1, S2, . . . , SK ∈ S with n1, n2, . . . , nK If λ2 < 1/ spikes, (KNmaxT 2), then the number of spikes in the mean train is the median of {nk}K respectively, and let Nmax = max(n1, n2, . . . , nK). 4. Let spike trains S1, . . . , SK ∈ SM . If λ2 < 1/(KM T 2), then the mean k=1. spike train has a conventional closed-form: 1 K K Xk=1 Sk. 6 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU 2.4. Computation of the mean spike train. To compute the mean spike train S∗ under the GVP metric, we need to estimate two unknowns: (1) the number of spikes n, and (2) the placements of these spikes in [0, T ]. For a general value of λ > 0, neither the matching term nor the penalty term is dominant, and therefore we cannot identify the number of spikes in the mean before estimating their placements [Wu and Srivastava (2011)]. A key challenge is that we need to update the number of spikes in the algorithm. In this article, we propose a general algorithm to estimate the mean spike train. We initialize the number of spikes in the mean spike train equal to the maximum of {n1, n2, . . . , nK}, and then adjust this number during the iterations. We present, here, how the Karcher mean in equation (2) can be efficiently computed using a convergent procedure. 2.4.1. Algorithm. Assume that we have a set of K spike trains, S1, . . . , SK i=1. with n1, n2, . . . , nK spikes, respectively. Denote Sk = (sk Then the sum of squared distances in equation (2) is i=1 and S = (si)n i )nk K Xk=1 (4) dGVP[λ](Sk, S)2 = K Xk=1 min γ∈Γ(cid:18)EOR[Sk, γ(S)] + λ2 X{i,j:sk i =γ(sj )} (sk i − sj)2(cid:19). We develop here an iterative procedure to minimize PK k=1 dGVP[λ](Sk, S)2 (as a function of S) and estimate the optimal S∗. This new algorithm has four main steps in each iteration: Matching, Adjusting, Pruning, and Checking, and we refer to it as the MAPC algorithm. In particular, the Adjusting step corresponds to the Centering step in the MCP-algorithm in Wu and Srivastava (2013); in contrast to the nonlinear warping-based Centering- step, the Adjusting step utilizes the Euclidean property and updates the mean spike train in an efficient linear fashion. The Checking step is mainly used to avoid underestimating the number of spikes in the mean. This step adds one spike into the current mean and checks if such an addition results in a better mean (i.e., smaller mean squared distances). In contrast, such a problem is not addressed in the MCP algorithm. Matching–Adjusting–Pruning–Checking (MAPC) Algorithm: 1. Let n = max{n1, n2, . . . , nK}. (Randomly) set initial times for the n spikes in [0, T ] to form an initial S. 2. Matching step: Use the dynamic programming procedure [Wu and Srivas- tava (2011)] to find the optimal matching γk from S to Sk, k = 1, . . . , K. EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE 7 That is, (5) γk = argmin γ∈Γ (cid:18)EOR[Sk, γ(S)] + λ2 X{i,j:sk i =γ(sj)} (sk i − sj)2(cid:19). 3. Adjusting step: (a) For k = 1, . . . , K, j = 1, . . . , n, define rk j =(cid:26) sk i , sj, if ∃i ∈ {1, . . . , nk}, s.t. γk(sj) = sk i , otherwise. n), k = 1, . . . , K. Then we update the mean 4. Pruning step: Remove spikes from the proposed mean ¯S that are matched (b) Denote Rk = (rk spike train to be S = 1 1 , . . . , rk i=1 Ri. K PK less than K/2 times. (a) For each j = 1, . . . , n, count the number of times sj appears in {γk(Sk)}K k=1. That is, hj =PN dated mean spike train as S∗. Then k=1 1sj ∈γk(Sk). (b) Remove sj from S if hj ≤ K/2, j = 1, . . . , n, and denote the up- S∗ = {sj ∈ Shj > K/2}. 5. Checking step: To avoid being stuck in a local minimum, we check if an insertion or/and deletion of a specific spike can improve the mean estimation: (a) Let S∗ be S∗ except one spike with the minimal number of ap- pearances (randomly chosen if there are multiple spikes at the minimum) in the Pruning step. Then, update the mean as if dGVP[λ](Sk, S∗)2 < otherwise. dGVP[λ](Sk, S∗)2, K Xk=1 (b) Let S∗∗ be the current mean S∗∗ with one spike inserted at random within [0, T ]. Then update the mean as S∗∗ =  S∗, S∗, S∗∗∗ =  S∗∗, S∗∗, K Xk=1 K Xk=1 if dGVP[λ](Sk, S∗∗)2 < otherwise. dGVP[λ](Sk, S∗∗)2, K Xk=1 6. Mean update: Let S = S∗∗∗ and n be the number of spikes in S. 8 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU 7. If the sum of squared distances stabilizes over steps 2–6, then the mean spike train is the current estimate and we can stop the procedure. Oth- erwise, go back to step 2. that the sum of squared distances (SSD), PK Denote the estimated mean in the mth iteration as S(m). One can show k=1 dGVP[λ](Sk, S(m))2, de- creases iteratively as a function of m. As 0 is a natural lower bound, the SSD will always converge when m gets large. The detailed proof is given in the Appendix. Note that this MAPC algorithm takes only linear computational order on the number of spike trains in the set. In practical applications, we find that this algorithm has great efficiency in reaching a reasonable convergence to a mean spike train. In general, when the penalty coefficient λ gets large, the optimal time warping chooses to have fewer matchings between spikes to lower the warping cost. Some of the spikes in the mean will be removed during the iterations to minimize the SSD. In the extreme case, when λ is sufficiently large, any time warping would be discouraged (as that will result in a larger distance than simply the Exclusive OR operation). In this case, the mean spike train will be an empty train. This result indicates that in order to get a meaningful estimate of the mean spike train, the penalty coefficient λ should not take a very large value. In practical use, one may use a cross-validation to decide the optimal value of λ. 2.4.2. Illustration of the MAPC algorithm. To test the performance of the MAPC algorithm, we illustrate the mean estimation using 30 spike trains randomly generated from a homogeneous Poisson point process. Let the total time T = 1 (sec), the Poisson rate ρ = 8 (spikes/sec). The individual spike trains are shown in Figure 1A. The number of spikes in these trains varies Fig. 1. A: 30 spike trains generated from a homogeneous Poisson process. Each vertical line indicates a spike. B: Estimation results when λ2 = 6. Upper panel: The sum of squared distances (SSD) over all iterations. Lower panel: The estimated mean spike train over all iterations. The initial is the spike train on the top row (0), and the final estimate is the spike train on the bottom row (12th). C: Estimation result when λ2 = 60. EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE 9 from 3 to 13, with the median value of 9. Therefore, n, the number of spikes in the mean, is initialized to be 13 and we adopt randomly distributed 13 spikes in [0, T ] as the initial for the mean in each case. We let λ2 vary between 6 and 60 to show the behavior for a small and a large warping penalty. The result for the MAPC algorithm for λ2 = 6 is shown in Figure 1B. The upper panel shows the evolution of the SSD in equation (2) versus the iteration index. We see that it takes only a few iterations for the SSD to decreasingly converge to a minimum. The estimated mean spike trains over iteration are shown in the lower panel. Apparent changes are observed in the first few (1 to 5) iterations, and then the process stabilizes. Note that the spikes in the mean train are approximately evenly spaced, which properly captures the homogeneous nature of the underlying process. We also note that the number of spikes in this mean spike train is 9, which equals the median of the numbers of spikes in the set. The result for λ2 = 60 is shown in Figure 1C. With a larger penalty, the optimal time warping between spike trains chooses to have fewer matchings between spikes to lower the warping cost. Some of the spikes in the mean are removed during the iteration. In this case, the convergent SSD is about 150, which is greater than the SSD when λ2 = 6 (at about 80). Note that when λ is even larger, we expect fewer or even no spikes to appear in the estimated mean. 2.5. Advantages over previous methods of averaging spike trains. There have been multiple ideas of capturing the general trend in a set of spike trains, which include the "consensus" spike train [Victor and Purpura (1997)], the "prototype" spike train [Diez, Schoenberg and Woody (2012)], the "aver- age" spike train [Julienne and Houghton (2013)], and the "mean" spike train [Wu and Srivastava (2011)]. However, none of those concepts satisfies all de- sirable 5 properties of a mean spike train listed in the Introduction section. We have summarized the most relevant differentiating features in Table 1. In the case of the "consensus" and "prototype" spike trains, one main problem lies in the nonuniqueness of the results in the spike train space, which arises directly from the underlying metric used (resembles Manhattan distance). If the estimated spiking times of those averages are restricted to spiking times in the sample sets, then the estimates can be unreliable, particularly when sample sizes are relatively small. The "average" design uses the van Rossum metric, which relies on kernel-smoothing of the spike trains. The estimation of the "average" is based on a greedy algorithm, but the accuracy and the kernel dependence of the method have not been carefully examined. In this article, we propose a new notion of a "mean" spike train based on the kernel- free GVP metric. The key advantages behind our design are the Euclidean properties of GVP distance and the subsequent Karcher mean definition [in 10 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU Comparison of average of spike trains for different methods Table 1 "mean" (proposed framework) "consensus" Victor and Purpura (1997) Method "average" "prototype" "mean" Diez, Schoen- Julienne and Wu and Srivastava Houghton berg and Woody (2012) (2013) (2011) Metric used GVP metric Victor– Victor– Purpura metricPurpura metric metric van Rossum Elastic metric Properties (in 1, 2, 3, 4, 5 2, 4, 5 2, 4, 5 1, 2, 4 1, 2, 5 Introduction) satisfied Domain Full spike train space Number of spikes, Median of λ2 ≪ 1 {n1, . . . , nN } Given sample set NA Spike times if n1 = · · · = nN , λ2 ≪ 1 cj = 1 N PN k=1 skj Restricted to sample set Given spike times Full spike Full spike train space train space set NA Restricted to sampled spike times NA NA median of {n1, . . . , nN } ISI-based nonlinear form Uniqueness in the Almost surely Nonunique full space Nonunique Not known Almost surely equation (2)]. The new framework satisfies all 5 desirable properties, which distinguishes it from others. It is worth noting that, to the best of our knowledge, the GVP metric is one of only two spike train metrics that have Euclidean properties [the other one is the Elastic metric proposed in Wu and Srivastava (2011)]. However, the Elastic metric satisfies only 3 out of the 5 properties, and the GVP metric has two apparent advantages over it: first, the Elastic metric estimates the mean spike train explicitly depending on recording intervals. Such depen- dence may introduce an additional level of noise arising from experimental parameters, making the inference less reliable. In contrast, the mean spike train under the GVP metric relies only on exact spike times in the given data and is independent of recording intervals (Property 4). Second, the fact that the Elastic mean is estimated through the inter-spike intervals (ISI) makes it difficult to capture the intuition behind the result, whereas the GVP mean is estimated directly through spike times and resembles the intuition of the mean estimation (Property 3). For illustrative purposes, we compare spike train "averages" of all methods using the 30 spike trains in Section 2.4.2, where the data is simulated under a homogeneous Poisson process. A natural expectation is that these averages should be equi-distantly spaced across the time domain. We adopt a simple EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE11 Fig. 2. Averaged spike trains according to four different methods. measure for the equi-distant spacing-compute the standard deviation of the ISI in each train, denoted by SDISI. Basically, smaller SDISI values indicate more even spacing. In Figure 2, we show the averages estimated using the GVP mean, Elastic mean, "Prototype," and "Consensus" methods. (In the GVP and Elastic methods, we let penalty coefficients be sufficiently small.) It is found that SDISI in the GVP mean is 0.019, the smallest over all four methods. 3. Application in noise removal. The notion of the mean spike train has a direct application in neural decoding. In this article we examine how the mean can be used to remove spontaneous activity in geniculate ganglion neurons and improve decoding performance. 3.1. Noise-removal method. Geniculate ganglion neurons exhibit spiking response to the chemical stimulus applied on the taste buds on the tongue. Such neuronal activity is commonly used in the neural coding in the periph- eral gustatory system [Di Lorenzo, Chen and Victor (2009), Breza, Nikonov and Contreras (2010), Lawhern et al. (2011)]. We note that these neurons exhibit responses even if there is no stimulus applied or the stimulation is a control solution-artificial saliva. That is, the observed spike trains under the stimulation are likely to be a mixture of the spontaneous activity and responses to the taste stimuli. In the context of neural decoding, such spon- taneous activity can be viewed as "background noise," and a "de-noised" spiking activity is expected to better characterize the neural response with respect to the taste stimulus and result in better decoding performance. Previous approaches to the noise-removal focus mainly on spike count across a time domain and do not have a temporal matching between spikes. In this paper, we propose a novel noise-removal procedure based on our new framework. In Figure 3 we describe the schematic idea of incorporating the noise removal in statistical inference. The procedure assumes that the ob- served data is a "sum" of isolated neuron responses and their spontaneous activity. To improve the neural decoding, we at first use the stimulus-free spike recordings to estimate the mean background noise with the MAPC 12 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU Fig. 3. Scheme differentiating the noise-removal approach from standard inference on spike train data. Dashed boxes indicate the components of standard inference framework, the solid lines indicate where the noise-removal framework is introduced. algorithm. Then, we "subtract" the mean out from the observed stimulus- dependent data. Obviously, for random variables X, Y in vector spaces one cannot assume that X = X + Y − ¯Y ( ¯Y denotes the mean of Y ) is a noise re- duced "version" of X + Y . However, in the space of spike trains we managed to establish this procedure, utilizing the warping matchings on the GVP mean. The procedure of obtaining X = X ⊕ Y ⊖ ¯Y indeed gives a noise- reduced version X of a point pattern X ⊕ Y . This approach is possible with definitions of the addition ⊕ and the subtraction ⊖ in the spike train space as follows: Adding spike trains. We assume that the noise is additive and adding spike trains is achieved by union set operation. That is, let X = (x1, . . . , xN ) and Y = (y1, . . . , yM ) be two spike trains of length N and M , respectively. We define a spike train Z = X ⊕ Y as a spike train of length N + M such that Z = X ⊕ Y = ({x1, . . . , xN } ∪ {y1, . . . , yM }). Subtracting spike trains. Defining the subtraction is more challenging, as it cannot follow directly from the set operations. This is due to the fact that it is unlikely to have coinciding spike times in two different spike trains. To perform the subtraction, we turn to the definition of the GVP metric and optimal warping between two spike trains [equation (1)]. We define the subtraction of a spike train Y from a spike train X as removing all spike positions from X that are matched with spikes in Y under the optimal warping γ. We say that a spike pair (xi, yj) is matched if xi = γ(yj) for some pair (i, j). Formally, let X = (x1, . . . , xN ), Y = (y1, . . . , yM ) be two spike trains and γ be the optimal warping between them according to the dGVP metric. We define the subtraction of Y from X, noted by Z = X ⊖ Y , as follows: Z = X ⊖ Y = ({x1, . . . , xN } \ {xi : xi = γ(yj) for some pair (i, j)}). Once the ⊕ and ⊖ are established, we can describe the noise-removal method as follows: we "subtract" the mean(Y ) from the observed X ⊕ Y EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE13 using the matching of the GVP metric and obtain a spike train X ⊕ Y ⊖ mean(Y ) by removing matched spikes. We at first use a simulation for illus- tration of both ⊕ and ⊖ operations in Section 3.2. The new method is then applied to a real experimental data set in Section 3.3. 3.2. Result for simulated data. To illustrate the noise-removal frame- work, we first generate 40 independent realizations {µi}40 i=1 of a homogeneous Poisson process on [0, 2] with constant intensity α = 10. These simulations represent the noise and are used to estimate the mean background noise µ with the MAPC algorithm. The results are shown in Figure 4A. Next we generate two sets of 20 independent spike trains, {Xi}20 i=1 and {Yi}20 i=1, as realizations of an inhomogeneous Poisson processes (IPP) with intensity functions ρX(t) = exp(−(t − 1.5)2) and ρY (t) = exp(−(t − 0.5)2), respectively. The generated spike trains are shown in Figure 4B. In our framework they correspond to the underlying true neuronal signals. In the third step we obtain the equivalent of the "observed" data, by adding the previously generated noise for each generated µi to the corre- sponding spike trains Xi and Yi. The combined results are shown in Fig- ure 4C. In this case adding spike trains is understood in the set opera- tions terms. We obtain spike trains following Poisson processes Xi ⊕ µi ∼ IPP(ρX + α), Yi ⊕ µi+20 ∼ IPP(ρY + α). The mean background noise spike train µ is then subtracted out from each realization of the noised data set, according to the procedure described in Section 3.1. For each i = 1, . . . , 20, we obtain the noise removed spike trains: Xi ⊕ µi ⊖ µ, Yi ⊕ µi+20 ⊖ µ, shown in Figure 4D. Illustration of the noise addition and the noise removal with the use of the ⊕, ⊖ Fig. 4. operations. A: Background noise-40 spike trains generated from HPP(10); the mean back- ground noise is presented with dashed lines in the bottom row. B: 2 × 20 spike trains from IPP(ρX) (asterisks) and IPP(ρY ) (circles), respectively. C: Sum of spike trains from A and B. D: Spike trains after the background noise removed. 14 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU Fig. 5. The noise-removal influence on classification performance with respect to in- creasing noise level α. A: The classification performance for the noisy data. The bold lines represent the average classification score among 50 simulations, the dotted lines in- dicate the standard deviation from the average classification score. B Same as A, but for the noise-removed data. C: Mean classification score curves from A (dashed line) and B (dotted line). We repeat this simulation procedure 50 times for each level of α ∈ [2, 20] and perform classification on the noisy data Xi ⊕ µi, Yi ⊕ µi+20 as well as on the noise-removed: Xi ⊕ µi ⊖ µ, Yi ⊕ µi+20 ⊖ µ. The classification score is obtained by a standard leave-one-out cross-validation. We record the aver- age score (the classification accuracy) with the standard deviation for each α level; the result is shown in Figure 5A, B. As anticipated with the increas- ing noise intensity α, the classification performance on each of the noisy and noise-reduced data sets declines. However, if we compare the two average classification scores presented in Figure 5C, we see that the noise-removal framework outperforms the classification on the noisy data once the noise intensity level α becomes not negligible. This result indicates that the pro- posed noise-removal procedure can help increase the contrast between dif- ferent classes and result in an improvement in classification analysis. Next, we will examine this procedure on a real experimental data set. 3.3. Result in real data in gustatory system. Here we apply the noise- removal procedure to neural response in the gustatory system and test if the decoding (i.e., classification with respect to taste stimuli) can improve after the spontaneous activity is removed. The data consists of spike train recordings of rat geniculate ganglion neurons and was previously used in Wu et al. (2013). Briefly, adult male Sprague–Dawley rat's geniculate ganglion tongue neurons were stimulated with 10 different solutions over a time period of 5 seconds: 0.1 M NaCl, 0.01 M citric acid (CA), 0.003, 0.01, 0.03, and 0.1 M acetic acid (AA), and each AA mixed with 0.1 M NaCl. Each stimulus was presented 2–4 times. Stimulus trials were divided into three time regions: a 5-second pre-stimulus period, a 5-second stimulus application period, and a EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE15 Fig. 6. An example of spike trains from Cell 10. Each group of 3 or 4 rows corresponds to a different type of stimuli applied. A: The 5-second pre-stimulus spike trains, whose mean spike train, calculated by the MAPC algorithm, is shown by the thick vertical bars on the top of the panel. B: The 5-second stimulus period. C: The same 5-second period of spike trains as in B, but with spontaneous activity subtracted out. 5-second post-stimulus period. During the first and third regions, a control solution of artificial saliva was applied. During the stimulus period, one of the 10 aforementioned solutions was applied. In this study we focus on classifying the given spike trains according to the 10 stimuli presented in each of 21 observed neurons. In Figure 6A, B we present the real data recordings in the first and second time regions from one example neuron. The spike trains in the pre-stimulus 5-second period reflect spontaneous activity with artificial saliva applied. They are treated as "noise" data, in contrast to stimulus-dependent responses. We compute their mean spike train with the parameter λ2 = 0.001 (a small value to get more spikes in the mean). The result is shown in the top row in Figure 6A. This mean properly summarizes spiking activity during the pre-stimulus period. In the next step, we subtract out this mean noise from the data during the stimulus period (spike trains between the 5th and 10th second). The noise-removed spike trains are shown in Figure 6C. We can now compare the decoding performance using the observed stimulus-response data and the "noise-removed" data. To reliably evaluate classification scores, we take the approach of leave-one-out cross-validation. In both cases of the observed data and the noise-removed data, the class 16 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU Fig. 7. The result of the noise-removal procedure applied to each of the recorded 21 cells. The marker coding is the same for both panels and indicates the influence of the noise-removal approach on the classification score: black circles-increase, grey diamond- s-decrease, black asterisks-unchanged. A. Raw classification scores for each cell in each condition. B. Same result as in A, but in terms of classification score increase with respect to mean noise size. The vertical black line corresponds to the noise size cutoff of 10 spikes. is assigned according to the nearest neighbor's class under the dGVP met- ric. In this classification analysis, we use λ2 = 225, a relatively larger value to emphasize the importance of both matching term and penalty term in equation (1). The comparison on the classification accuracy is shown in Figure 7A. In 10 out of 21 cells the classification was improved after the noise-removal procedure and in only 4 cells the classification was hindered. Classification in 7 cells remained unchanged after noise removal, which seems to be quite significant. To explain this issue, we have investigated the size of the mean background noise and its influence on increase in classification performance. It turned out, as seen in Figure 7B, that in 5 out of these 7 unchanged cells, the pre-stimulus spiking is negligible (the estimated mean spike train has 0 or 1 spike). In those cases, obviously, subtracting out the mean noise spike train will bare minimum influence. The remaining two cases are associated with the opposite problem of the noise size-the number of spikes in the pre-stimulus period is comparable or greater than the number of spikes in the stimulation period. When such mean noise is subtracted out, it also can take away relevant information, thus it may not improve the decoding. Those noise size issues are consistent with common intuition behind the noise removal. However, it is worth noting that the size of the estimated mean background noise can be controlled in our new framework, by adjusting the penalty parameter λ (Section 2.4). More investigation will be conducted on the selection of λ in our future work. EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE17 When focusing on cells that have significant noise influence in their spik- ing pattern (at least 10 spikes in the estimated mean noise spike train), we see that in the majority of cases (8 out of 13) the noise removal has improved the classification score (Figure 7B). In extreme cases we obtain up to 20% improvement in the decoding performance. (Note that with 10 different stimuli, a random guess results in 10% average accuracy.) Only 3 out of the 13 cells indicate loss of information and 2 are not influenced by noise removal. In summary, we find that our notion of mean background noise for spike train data is in agreement with the common understanding of the additive noise for the majority of recorded neurons. Moreover, the proposed noise- removal framework effectively improves neural decoding, provided that the pre-stimulus spiking has a high enough intensity. 4. Discussion. In this article we propose a new framework for defining the mean of a set of spike trains and the deviation from the mean. We pro- vide an efficient algorithm for computing the mean spike train and prove the convergence of the method. The framework is based on the dGVP metric [Dubbs, Seiler and Magnasco (2010)] which resembles the Euclidean dis- tance. This concept gives an intuitive sample mean point pattern, in which the spike positions in the mean are averaged among matched spikes in a set of all spike trains. Our summary statistics provide the basis for inference on point pattern data and in this article we utilize it to develop a mean-based noise-removal approach. We show that our procedure improves the classification score for simulated inhomogeneous Poisson point process data with various nonneg- ligible noise levels. We have also applied the new tools to a neural decoding problem in a rat's gustatory system. It is found that the mean point pattern approach and the noise-removal framework significantly improved the neural decoding among the set of 21 neurons. In the noise-removal framework, we have defined the operations of ad- dition and subtraction between spike trains with the use of the matching component of the GVP metric. We, however, note that those operations do not satisfy the law of associativity. For more advanced analysis, it is desir- able to establish an algebraic structure on the space of point patterns. Thus, refining those approaches will be pursued in the further work. Once the algebraic structure is established, statistical models can be built and regression analysis can be performed. With this setting and the already developed mean and the deviation from the mean approaches, we expect to develop classical statistical inferences such as hypothesis tests, confidence intervals/regions, FANOVA (functional ANOVA), FPCA (functional PCA), and regressions on functional data [Ramsay and Silverman (2005), Valder- rama (2007)]. All these tools are expected to provide a new methodology 18 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU for more effective analysis and modeling of neural spike trains or any point pattern data in general. APPENDIX Theorem 1 (Convergence of the MAPC algorithm). Denote the esti- mated mean in the mth iteration of the MAPC-algorithm as S(m). Then the k=1 dGVP(Sk, S(m))2 decreases iteratively. That sum of squared distances PK is, dGVP(Sk, S(m+1))2 ≤ K Xk=1 dGVP(Sk, S(m))2. K Xk=1 Proof. The proof will go through steps 2–6 of the algorithm-in each step we will show that the overall distance to the proposed mean S(m) = (s(m) 1 1. Matching: In the mth iteration, we find the optimal matching γk from n ) is nonincreasing: , . . . , s(m) S(m) to Sk for each k ∈ {1, . . . , K}. Having those, we can write dGVP(Sk, S(m))2 K Xk=1 = K Xk=1 EOR(Sk, γk(S(m))) + λ2 K Xk=1 X{i,j:sk i =γk(s(m) j (sk i − s(m) j )2. )} 2. Adjusting: By definition, we update S(m) to S(m) = (s(m) , . . . , s(m) n ) = 1 k=1 Rk, where Rk = (rk 1 , . . . , rk n) with rk j =( sk i , s(m) j if ∃i ∈ {1, . . . , nk}, s.t. γk(s(m) otherwise j ) = sk i , , k = 1, . . . , K, j = 1, . . . , n. 1 K PK i =γk(s(m) j (sk i −s(m) j )} )2 =PK k=1Pn j=1(rk j −s(m) j k=1 × )2 ≥PN j=1(rk Let γ be the piecewise linear warping function from S(m) to S(m), that )2. j k=1P{i,j:sk j − s(m) Hence,PK Pn is, S(m) = γ(S(m)). Then dGVP(Sk, S(m))2 K Xk=1 EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE19 K K n EOR(Sk, γk(S(m))) + λ2 EOR(Sk, γk(S(m))) + λ2 j − s(m) j )2 K (rk Xj=1 Xk=1 Xk=1 X{i,j:sk Xk=1 K (sk i − s(m) j )2 )} i =γk(s(m) j ≥ ≥ = ≥ K Xk=1 Xk=1 Xk=1 Xk=1 K K EOR(Sk, γk ◦ γ−1( S(m))) + λ2 X{i,j:sk i =γk◦γ−1(s(m) j (sk i − s(m) j )2 )} dGVP(Sk, S(m))2. S(m) in which all spikes appear in {γk(Sk)}K the result 3. Prunning: S∗(m) = {sj ∈ S(m)PK PK k=1 EOR(Sk, γk ◦ γ−1( S(m))) + λ2PK in the Adjusting step, we have PK )} Using the basic rule of the Exclusive OR, it is easy to find that k=1 1sj ∈γk(Sk) ≥ K/2} is a subset of k=1 at least K/2 times. Based on k=1 dGVP(Sk, S(m))2 ≥ )2. i − s(m) k=1P{i,j:sk i =γk◦γ−1(s(m) (sk j j EOR(Sk, γk ◦ γ−1( S∗(m))) ≤ K Xk=1 EOR(Sk, γk ◦ γ−1( S(m))). K Xk=1 Let S∗(m) = (s∗(m) S∗(m). Then, 1 , . . . , s∗(m) n∗ ), where n∗ denotes the number of spikes in dGVP(Sk, S(m))2 K Xk=1 EOR(Sk, γk ◦ γ−1( S∗(m))) + λ2 K Xk=1 dGVP(Sk, S∗(m))2. ≥ ≥ K Xk=1 Xk=1 K X {i,j:sk i =γk◦γ−1(s∗(m) j (sk i − s∗(m) j )2 )} 4. Checking: Finally, we perform the checking step to avoid the possible local minima in the pruning process. In the test if a spike can be removed from S∗(m), we let S∗(m) be S∗(m) except one spike with a minimal number 20 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU of appearances. Then update the mean spike train as dGVP(Sk, S∗(m))2, K K if otherwise. dGVP(Sk, S∗(m))2 < Xk=1 Xk=1 k=1 dGVP(Sk, S(m))2 ≥PN k=1 dGVP(Sk, S∗∗(m))2. In the test if a spike can be added to S∗∗(m), we let S∗∗(m) be S∗∗(m) with one spike inserted at random within [0, T ]. Then update the mean spike train as S∗(m), S∗(m), S∗∗(m) =  It is easy to verify thatPN S∗∗∗(m) =  It is easy to see that PK S∗∗(m), S∗∗(m), if dGVP(Sk, S∗∗(m))2 < K Xk=1 otherwise. dGVP(Sk, S∗∗(m))2, K Xk=1 k=1 dGVP(Sk, S∗∗∗(m))2. Us- ing step 6, the mean at the (m + 1)th iteration is S(m+1) = S∗∗∗(m). Hence, k=1 dGVP(Sk, S(m))2 ≥PK dGVP(Sk, S(m+1))2 ≤ K Xk=1 K Xk=1 dGVP(Sk, S(m))2. (cid:3) REFERENCES Aronov, D. and Victor, J. D. (2004). Non-Euclidean properties of spike train metric spaces. Phys. Rev. E (3) 69 061905, 9. MR2096492 Aronov, D., Reich, D. S., Mechler, F. and Victor, J. D. (2003). Neural coding of spatial phase in V1 of the macaque monkey. J. Neurophysiol. 89 3304–3327. Box, G. E. P., Hunter, W. G. and Hunter, J. S. (1978). Statistics for Experimenters: An Introduction to Design, Data Analysis, and Model Building. Wiley, New York. MR0483116 Breza, J. M., Nikonov, A. A. and Contreras, R. J. (2010). Response latency to lingual taste stimulation distinguishes neuron types within the geniculate ganglion. J. Neurophysiol. 103 1771–1784. Brown, E. N., Barbieri, R., Ventura, V., Kass, R. E. and Frank, L. M. (2002). The time-rescaling theorem and its application to neural spike train data analysis. Neural Comput. 14 325–346. Diez, D. M., Schoenberg, F. P. and Woody, C. D. (2012). Algorithms for computing spike time distance and point process prototypes with application to feline neuronal responses to acoustic stimuli. J. Neurosci. Methods 203 186–192. Di Lorenzo, P. M., Chen, J. and Victor, J. D. (2009). Quality time: Representation of a multidimensional sensory domain through temporal decoding. J. Neurosci. 29 9227– 9238. Dubbs, A. J., Seiler, B. A. and Magnasco, M. O. (2010). A fast Lp spike alignment metric. Neural Comput. 22 2785–2808. MR2760538 Houghton, C. (2009). Studying spike trains using a van Rossum metric with a synapse- like filter. J. Comput. Neurosci. 26 149–155. MR2481024 EUCLIDEAN SUMMARY STATISTICS IN THE NEURAL SPIKE TRAIN SPACE21 Houghton, C. and Sen, K. (2008). A new multineuron spike train metric. Neural Com- put. 20 1495–1511. MR2410365 Hunter, J. D. and Milton, J. G. (2003). Amplitude and frequency dependence of spike timing: Implications for dynamic regulation. J. Neurophysiol. 90 387–394. Julienne, H. and Houghton, C. (2013). A simple algorithm for averaging spike trains. J. Math. Neurosci. (JMN) 3 1–14. Karcher, H. (1977). Riemannian center of mass and mollifier smoothing. Comm. Pure Appl. Math. 30 509–541. MR0442975 Kass, R. E. and Ventura, V. (2001). A spike-train probability model. Neural Comput. 13 1713–1720. Kass, R. E., Ventura, V. and Brown, E. N. (2005). Statistical issues in the analysis of neuronal data. J. Neurophysiol. 94 8–25. Kreuz, T., Haas, J. S., Morelli, A., Abarbanel, H. D. I. and Politi, A. (2007). Measuring spike train synchrony. J. Neurosci. Methods 165 151–161. Lawhern, V., Nikonov, A. A., Wu, W. and Contreras, R. J. (2011). Spike rate and spike timing contributions to coding taste quality information in rat periphery. Frontiers in Integrative Neuroscience 5 1–14. Lim, D. and Capranica, R. R. (1994). Measurement of temporal regularity of spike train responses in auditory nerve fibers of the green treefrog. J. Neurosci. Methods 52 203–213. MacLeod, K., Backer, A. and Laurent, G. (1998). Who reads temporal information contained across synchronized and oscillatory spike trains? Nature 395 693–698. Paiva, A. R. C., Park, I. and Pr´ıncipe, J. C. (2009). A reproducing kernel Hilbert space framework for spike train signal processing. Neural Comput. 21 424–449. MR2477866 Quian Quiroga, R., Kreuz, T. and Grassberger, P. (2002). Event synchronization: A simple and fast method to measure synchronicity and time delay patterns. Phys. Rev. E (3) 66 041904, 9. MR1935193 Ramsay, J. O. and Silverman, B. W. (2005). Functional Data Analysis, 2nd ed. Springer, New York. MR2168993 Rieke, F., Warland, D., de Ruyter van Steveninck, R. and Bialek, W. (1999). Spikes: Exploring the Neural Code. MIT Press, Cambridge, MA. MR1983010 Schreiber, S., Fellousb, J. M., Whitmerc, D., Tiesingaa, P. and Sejnowskib, T. J. (2003). A new correlation-based measure of spike timing reliability. Neurocomputing 52- 54 925–931. Valderrama, M. J. (2007). An overview to modelling functional data. Comput. Statist. 22 331–334. MR2336339 van Rossum, M. C. W. (2001). A novel spike distance. Neural Comput. 13 751–763. Victor, J. D., Goldberg, D. H. and Gardner, D. (2007). Dynamic programming algo- rithms for comparing multineuronal spike trains via cost-based metrics and alignments. J. Neurosci. Methods 161 351–360. Victor, J. D. and Purpura, K. P. (1996). Nature and precision of temporal coding in visual cortex: A metric-space analysis. J. Neurophysiol. 76 1310–1326. Victor, J. D. and Purpura, K. P. (1997). Sensory coding in cortical neurons. Ann. N.Y. Acad. Sci. 835 330–352. Wu, W. and Srivastava, A. (2011). An information-geometric framework for statistical inferences in the neural spike train space. J. Comput. Neurosci. 31 725–748. MR2864743 Wu, W. and Srivastava, A. (2013). Estimating summary statistics in the spike-train space. J. Comput. Neurosci. 34 391–410. MR3061973 22 S. WESOLOWSKI, R. J. CONTRERAS AND W. WU Wu, W., Mast, T. G., Ziembko, C., Breza, J. M. and Contreras, R. J. (2013). Statistical analysis and decoding of neural activity in the rodent geniculate ganglion using a metric-based inference system. PLoS ONE 8 e65439. S. Wesolowski Department of Mathematics Florida State University Tallahassee, Florida 32306-4510 USA E-mail: [email protected] R. J. Contreras Department of Psychology Florida State University Tallahassee, Florida 32306-4301 USA E-mail: [email protected] W. Wu Department of Statistics Florida State University Tallahassee, Florida 32306-4330 USA E-mail: [email protected]
1807.11036
1
1807
2018-07-29T10:21:59
Modeling Human Decision-making: An Overview of the Brussels Quantum Approach
[ "q-bio.NC", "quant-ph" ]
We present the fundamentals of the quantum theoretical approach we have developed in the last decade to model cognitive phenomena that resisted modeling by means of classical logical and probabilistic structures, like Boolean, Kolmogorovian and, more generally, set theoretical structures. We firstly sketch the operational-realistic foundations of conceptual entities, i.e. concepts, conceptual combinations, propositions, decision-making entities, etc. Then, we briefly illustrate the application of the quantum formalism in Hilbert space to represent combinations of natural concepts, discussing its success in modeling a wide range of empirical data on concepts and their conjunction, disjunction and negation. Next, we naturally extend the quantum theoretical approach to model some long-standing `fallacies of human reasoning', namely, the `conjunction fallacy' and the `disjunction effect'. Finally, we put forward an explanatory hypothesis according to which human reasoning is a defined superposition of `emergent reasoning' and `logical reasoning', where the former generally prevails over the latter. The quantum theoretical approach explains human fallacies as the consequence of genuine quantum structures in human reasoning, i.e. `contextuality', `emergence', `entanglement', `interference' and `superposition'. As such, it is alternative to the Kahneman-Tversky research programme, which instead aims to explain human fallacies in terms of `individual heuristics and biases'.
q-bio.NC
q-bio
Modeling Human Decision-making: An Overview of the Brussels Quantum Approach Diederik Aerts1, Massimiliano Sassoli de Bianchi3, Sandro Sozzo3 and Tomas Veloz4 1 Center Leo Apostel for Interdisciplinary Studies and Department of Mathematics Brussels Free University, Krijgskundestraat 33, 1160 Brussels (Belgium) E-Mail: [email protected] 2 Center Leo Apostel for Interdisciplinary Studies Brussels Free University, Krijgskundestraat 33, 1160 Brussels (Belgium) and Laboratorio di Autoricerca di Base, via Cadepiano 18, 6917 Barbengo (Switzerland) E-Mail: [email protected] 3 School of Business and Centre IQSCS, University of Leicester University Road, LE1 7RH Leicester (United Kingdom) E-Mail: [email protected] 4 Universidad Andres Bello, Departamento Ciencias Biol´ogicas Facultad Ciencias de la Vida, 8370146 Santiago (Chile) and Instituto de Filosof´ıa y Ciencias de la Complejidad Los Alerces 3024, Nunoa, Santiago (Chile) E-Mail: [email protected] Abstract We present the fundamentals of the quantum theoretical approach we have developed in the last decade to model cognitive phenomena that resisted modeling by means of classical logical and probabilistic struc- tures, like Boolean, Kolmogorovian and, more generally, set theoretical structures. We firstly sketch the operational-realistic foundations of conceptual entities, i.e. concepts, conceptual combinations, proposi- tions, decision-making entities, etc. Then, we briefly illustrate the application of the quantum formalism in Hilbert space to represent combinations of natural concepts, discussing its success in modeling a wide range of empirical data on concepts and their conjunction, disjunction and negation. Next, we naturally extend the quantum theoretical approach to model some long-standing 'fallacies of human reasoning', namely, the 'conjunction fallacy' and the 'disjunction effect'. Finally, we put forward an explanatory hypothesis according to which human reasoning is a defined superposition of 'emergent reasoning' and 'logical reasoning', where the former generally prevails over the latter. The quantum theoretical ap- proach explains human fallacies as the consequence of genuine quantum structures in human reasoning, i.e. 'contextuality', 'emergence', 'entanglement', 'interference' and 'superposition'. As such, it is alter- native to the Kahneman-Tversky research programme, which instead aims to explain human fallacies in terms of 'individual heuristics and biases'. Keywords: Quantum structures; cognition; concept theory; decision theory; human reasoning. 1 1 The quantum cognition research programme Researchers in cognitive and social science have assumed for many years, often implicitly, that human judgment and decision-making under uncertainty can be modeled by means of set theoretical structures, i.e. sets and operations between sets. These structures are very similar to those used in distributive logic, formalized by George Boole ('Boolean logic'), and probability theory, axiomatized by Kolmogorov ('Kolmogorovian probability') [1]. There is agreement, in both the psychology and physics communities, to refer to these structures as 'classical structures', as they were originally used in classical physics and later extended to psychology, computer science, statistics, economics, finance, etc. In decision theory, classical structures have been, again implicitly, incorporated into the so-called 'Bayesian paradigm': any source of uncertainty can be formalized probabilistically in a Kolmogorovian sense. Expected utility theory (EUT), which provides the normative foundations of rational behavior under uncertainty, then guarantees that decision-makers should behave as if they maximized EU with respect to a Kolmogorovian probability measure formalizing their subjective uncertainty about the world [2]. More recently, this classical paradigm has been seriously challenged by a number of paradoxical findings in cognitive psychology, revealing that classical structures are generally unable to model concrete human decisions. This has made highly problematical the interpretation of a whole set of cognitive phenomena in terms of Boolean logic and Kolmogorovian probability. In regard to that, empirical literature has identified the following two major types of 'deviations from classicality'. (i) 'Probability judgment errors': people judge the probability of the conjunction 'A and B' (disjunction 'A or B') of two events A and B as higher (lower) than the probability of one of them, violating in this way the 'monotonicity law of Kolmogorovian probability'. (ii) 'Decision-making errors': people prefer action A over action B if they know that an event Z occurs, and also if they know that event Z does not occur, but they prefer B over A if they do not know whether Z occurs or not, violating in this way the 'law of total Kolmogorovian probability'. Errors of type (i) manifest, for example, in terms of 'over/under-extension effects' in typicality and membership judgments on conceptual categorization [3, 4, 5, 6]. But, the most famous evidence of prob- ability judgment error was discovered by Kahneman (Nobel prize in economics 2002) and Tversky (one of the most cited psychologists) by means of their 'Linda story' and is known as the 'conjunction fallacy' [7]. The conjunction fallacy has then been systematically confirmed and found generating further effects, fallacies, paradoxes and contradictions (see, e.g., [8, 9] for a review of cognitive fallacies). In 1992, Tversky and Shafir detected a significant example of decision-making error in the 'disjunction effect' [10], which led to the conclusion that people do not generally apply rational reasoning when taking concrete decisions but, in certain situations, they are instead strongly influenced by contextual aspects of a cognitive nature, like 'uncertainty aversion'. The disjunction effect violates the law of total Kolmogorovian probability, and also the 'sure thing principle', one of the fundamental axioms of EUT. A similar violation was observed in 1961 by Daniel Ellsberg, who noticed that people tend to avoid maximizing EU if they have to take decisions under an uncertainty aversion scenario ('Ellsberg paradox') [11]. A quite influential solution of the above difficulties came from Kahneman and Tversky themselves, who put forward that the above deviations are 'genuine fallacies of human reasoning', elaborating a theory based on 'judgment heuristics and individual biases' to explain them [12]. The Kahneman-Tversky solution, though valid at an intuitive level, does not seem to provide sufficient clarifications with respect to the deep mechanisms underlying human reasoning, hence it is generally accepted that efforts should be made to move beyond the fragmentation of existing approaches to develop a more coherent, comprehensive and deductive theoretical hypothesis [13]. An important alternative considers the traditional logical and probabilistic modelling schemes as too restrictive to capture the way in which the human mind works in a concrete judgment or decision. This 2 approach therefore adopts, with great success, the more general mathematical formalism of quantum prob- ability, which is known to be non-Kolmogorovian in a very precise sense. Impressive results have already been obtained by this 'quantum cognition research programme', in the modelling of conjunction and dis- junction fallacies, human similarity judgments, disjunction effects and Ellsberg-type paradoxes, knowledge and meaning representation, question order effects, etc. [8, 9, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25]. We illustrate here the Brussels contribution to this novel, exciting and potentially cutting-edge research programme. In Section 2, we present the foundations of the operational-realistic approach to human cog- nition we have recently developed relying on a two-decade research on the axiomatic approaches to quan- tum physics, the origins of quantum probability and the appearance of quantum structures outside the micro-world. We show that any conceptual entity, that is, a single concept, a conceptual combination, a proposition, or a more complex decision-making entity, can be abstractly described in terms of the oper- ationally defined notions of 'states', 'contexts', 'properties', 'measurements' and 'outcome probabilities', and that impressive analogies recur between the operational description of measurement processes and the effects of context on quantum and conceptual entities. This indicates that the quantum formalism and its quantum probability structure are natural candidates to model cognitive phenomena. Then, we formulate in Section 3 an explicit quantum theoretic approach to model combinations of two concepts, overcoming the classical modeling difficulties that arise in over- and under-extension effects on conceptual categorization. Next, we apply in Sections 4 and 3 the quantum theoretical approach to model the conjunction fallacy and the disjunction effect, respectively, explaining the empirical deviations from classical modeling in terms of genuine quantum structures. Finally, we formulate in Section 6 a theoretical hypothesis which, on the one side explains emergence of quantum structures in cognition and, one the other side, provides a unifying and coherent explanation across the cognitive effects mentioned above. This explanatory hypothesis inter- prets human reasoning as a superposition of 'logical reasoning' and 'emergent reasoning', and research on cognitive fallacies allows to demonstrate that the former generally prevails over the latter. The Brussels quantum modeling approach to conceptual entities was the starting point for the devel- opment of a quantum theoretical approach to 'meaning entities' that can be associated with documents and texts, like those of the World Wide Web, which is presented in [26]. 2 The Brussels operational-realistic approach to cognition It is important to note that the modeling effectiveness of quantum theory regarding the situations mentioned in Section 1 is not necessarily due to the existence of microscopic quantum processes in the human brain, at least at the present stage of research. So, why should cognition go into quantum? From our point of view, the reason for such a bold change of methodological perspective can be traced back to the studies on the axiomatic foundations of quantum physics, the origins of quantum probability and the differences between classical and quantum theories. More precisely, we were mainly guided by: (i) the deep analogies between quantum particles and conceptual entities with respect to 'potentiality' and 'contextuality'; (ii) the role played by quantum probability in formalizing experimental situations where potentiality and contextuality are present. It is indeed well known in quantum physics that, in a quantum measurement process, the measurement context physically influences the quantum particle that is measured in a 'non-deterministic way', actualizing one outcome within a set of possible measurement outcomes, as a consequence of the interaction between the quantum particle itself and the measurement context. Suppose now that a statistics of measurement outcomes is collected after a sequence of many repeated measurement processes on an arbitrary quantum entity, and such that (i) the measurement actualises properties of the entity that were not actual before 3 the measurement started; (ii) different outcomes and actualisations are obtained probabilistically. What type of probability can formalize such experimental situation? It cannot be a Kolmogorovian probability, because Kolmogorovian probability formalizes a situation of lack of knowledge about properties of the entity that were already actual before the measurement started. On the other hand, one can demonstrate that a situation where a measurement context actualizes properties that were only potential before the measurement started can be represented in a generalized quantum probabilistic framework [27, 28]. How about a human decision process? We realized that a decision process is generally made in a state of genuine potentiality, which is not of the type of a lack of knowledge of an actuality. Consider, for example, a survey including the question "are you a smoker or not?", and suppose that 21 participants over a whole sample of 100 participants answer 'yes' to this question [29]. In the large number limit, 0.21 is interpreted as the probability of finding a smoker in this sample of participants. This probability clearly formalizes a 'lack of knowledge about an actuality', because each participant 'is' a smoker or 'is not' a smoker before the property was tested, hence before the experiment to test it -- the survey -- started. Consider now the question "are you for or against the use of nuclear energy?", and suppose that 31 participants answer 'yes' to this question. In this case, the resulting probability 0.31 cannot formalize a lack of knowledge about an actuality. Indeed, due to the nature of the question, it is very plausible that some of the participants had no opinion about it before the survey, hence for these participants the outcome was brought into existence and influenced by the context at the time the question was asked, including the specific conceptual structure of how the question was formulated. This is how context plays an essential role whenever the human mind is concerned with outcomes of cognitive tests. Like in physics, the probability that represents a 'lack of knowledge about an actual cognitive property' is Kolmogorovian, while the probability that represents a 'context-driven actualization of a cognitive property' is non-Kolmogorovian and, possibly, quantum [27]. The effect context plays on a conceptual entity in a typicality judgment process is equally fundamental. Consider, for example, the concept Pet, and suppose that a sample of participants are asked to rank a list of items, like Dog, Cat, Squirrel, Goldfish, Parrot, etc., as typical examples of Pet. Most of the respondents will judge items like Snake and Spider as not very typical examples of Pet, preferring to them items like Dog and Cat. But, consider now the context expressed by Did you see the type of pet he has? This explains that he is a weird person. It is very plausible that people will judge Snake and Spider as very typical examples of Pet in this context, while Dog and Cat will score a lower typicality than in the case in which Pet is considered in the absence of any context. By using a terminology that will become familiar at the end of this section, we say that, "in a typicality measurement, the probability that the conceptual entity Pet changes its state to Snake is low in the absence of any context, while it is high in the presence of the context Weird" [30, 31]. The impressive analogies above, between quantum and conceptual entities, led us to look into common theoretical foundations for quantum physics and cognition, which could justify the use of a shared mathe- matical formalism to model microscopic and cognitive phenomena. In this respect, long-standing research exists in mathematical physics that attempts to recover the Hilbert space formalism of quantum theory from physically justified axioms, relying on well defined empirical notions, directly connected with the op- erations that are usually performed in a laboratory. One of the well-known approaches to the foundations of quantum physics and quantum probability is the 'Geneva-Brussels approach', initiated by Jauch [32] and Piron [33] in Geneva, and further developed by the Brussels research team (see, e.g., [34]). This research produced a formal approach, the 'State Context Property' (SCoP) formalism', where any physical entity is described in terms of the basic notions of 'state', 'context' and 'property', which arise as a consequence of concrete physical operations on macroscopic apparatuses, such as preparation and registration devices, performed in 'spatiotemporal' domains, such as physical laboratories. Measurements, outcomes, state transformations and probabilities can then be expressed in terms of these more fundamental notions. Also, if suitable axioms are imposed on the mathematical structure underlying the SCoP formalism, the Hilbert 4 space structure of quantum theory emerges as a unique (up to isomorphisms) mathematical representation. When the SCoP formalism was considered, it was clear from the beginning that its validity was very general, that is, it was able to describe not only physical entities but also entities of a more abstract nature, like conceptual entities, cultural artifacts, the mind of a person, etc. [35]. The possibility of obtaining an 'operational and realistic description' of conceptual entities, in the sense of using the SCoP formalism in cognitive domains to formalize conceptual entities in terms of states, contexts, properties, measurements and outcome probabilities, was recently re-emphasized and the essence of the approach explained in a more systematic way [36]. We briefly illustrate this operational-realistic foundation of human cognition in the following. Let us firstly consider the empirical phenomenology of cognitive psychology. Like in physics, where laboratories define precise spatiotemporal domains, we introduce 'psychological laboratories' where cog- nitive tests, or measurements, are performed. These measurements are performed in situations that are specifically 'prepared' for the tests, which includes experimental devices (if any), structured questionnaires, human participants interacting with the questionnaires in written answers, or each other, e.g., an inter- viewer and an interviewed. Whenever empirical data are collected from the responses of a sample of participants, a statistics of the obtained outcomes arises. These empirical facts allow identifying entities, states, contexts, measurements, outcomes, and probabilities, as follows. The complex of experimental procedures conceived by the experimenter, the experimental design and the cognitive effect under investigation, are typically associated with a process of preparation of the state of a conceptual entity A, defined by these same procedures. Hence, like in physics, the preparation procedure sets the initial state pA of the conceptual entity A under study. Let us consider, for example, a questionnaire where a participant is asked to rank on a 7-point Likert scale the membership of a list of items with respect to the concept Fruits. The questionnaire defines the state pF ruits of the conceptual entity Fruits. It is true that cognitive situations exist where the preparation of the state of a conceptual entity is hardly controllable. Nevertheless, the state of the conceptual entity, defined by means of such a process is a 'state of affairs', i.e. expresses a 'reality of the conceptual entity'. In other words, once the entity is prepared in a given state, the latter is independent of any measurement, and can be confronted with the different participants in an empirical test, leading to outcome data and their statistics, exactly like in physics. A context e is an element that can provoke a change of state of the conceptual entity. For example, the concept Juicy can function as a context for the conceptual entity Fruits leading to Juicy Fruits, which can then be considered as a state of the conceptual entity Fruits. A special context is the one introduced by the measurement itself. Indeed, when the cognitive test starts, an interaction of a cognitive nature occurs between the conceptual entity A under study and a participant in the experiment, during which the state pA of the conceptual entity A generally changes, being transformed into another state p. Also this cognitive interaction is formalized by means of a context e. For example, if the participant is asked to choose the most typical example of Fruits in a list of items, containing Olive, Almond, Apple, etc., and the answer is Apple, then the initial state pF ruits of the conceptual entity Fruits changes to pApple, i.e. to the state describing the situation "the fruit is an apple", as a consequence of the contextual interaction with the participant. The state change of a conceptual entity due to a context may be either 'deterministic', hence in prin- ciple predictable under the assumption that the state before the context acts is known, or 'intrinsically probabilistic', in the sense that only the probability µ(p, e, pA) that the state pA of A changes to the state p is is given and is different from 1. In the example above on typicality judgments, the typicality of the item Apple for the concept Fruits is formalized by means of the probability µ(pApple, e, pF ruits) of the state transition pF ruits → pApple, where the context e is the context of the typicality measurement. An interesting aspect concerns the final state of a conceptual entity after a judgment or decision. As in physics, we assume the existence of a nonempty class of cognitive tests that correspond to 'ideal first 5 kind measurements', in the sense that the measurement outcome xk uniquely determines the final state pk of the conceptual entity A after the measurement described by the context e. The operational-realistic approach above naturally suggests the canonical quantum representation in Hilbert space, where a conceptual entity is associated with a Hilbert space, states are represented by Hilbert space unit vectors, measurement outcomes are represented by orthonormal (ON) basis vectors, measurement contexts are represented by spectral families of orthogonal projection operators, probabilities and state transformations are represented by the Born and Luders rule of quantum probability, respectively. It is worth emphasizing that in our approach states of conceptual entities describe 'modes of being' of these entities, while participants act in cognitive tests as contexts for conceptual entities, changing their states. This means that states do not describe subjective beliefs of a person, or collection of persons. Subjective beliefs are instead incorporated in the cognitive interaction between the conceptual situation and the human participants deciding on that situation. In this respect, our operational-realistic approach to human cognition departs from other quantum cognition approaches (see, e.g., [9, 17, 18, 19, 36]). We conclude this section with an epistemological consideration. Our research investigates the validity of quantum theory as a general, unitary and coherent theory for human cognition. Our quantum theoretical models, elaborated for specific conceptual situations and data, derive from quantum theory as a consequence of the assumptions about this general validity. As such, these models are subject to the technical and epistemological constraints of quantum theory. In other terms, our quantum modeling rests on a 'theory based approach', and should be distinguished from an 'ad hoc modeling based approach', only devised to fit data. While one should be generally suspicious of models in which free parameters are added on an 'ad hoc' basis to fit data, we believe that the success of our 'theory based models' to reproduce different data sets constitutes in itself a convincing argument to support its advantage over traditional modeling approaches and to extend its use to more complex cognitive situations. 3 Modeling combinations of natural concepts We present in this section a quantum theoretical framework in complex Hilbert space to represent concepts and their combinations, showing that it explains the empirical deviations from classicality observed in conceptual categorization as genuine expressions of quantum structures [14, 15, 16, 30, 31, 37]. How concepts combine to form complex conceptual structures, like sentences and texts, how combina- tions carry new meaning, and how human reasoning processes are activated in these combinations, have always fascinated psychologists, logicians, philosophers and linguists, and have deep implications on ap- plied disciplines, such as computer science and artificial intelligence. Many efforts were devoted to these challenges, with very few groundbreaking results. Already at an intuitive level, defining and formalizing concepts is a complex task, which has to account for the dimension, from the 'more abstract' to the 'more concrete', of a concept. Saying Thing is different from expressing the more concrete concepts of Fruit, Apple, or even, more specifically, This Apple, which is finally associated with a tangible object, for instance that I would be holding in this moment in my hand. Different approaches were presented to tackle the problem of defining what a concept is and, more important, to represent concepts and their combinations, like simple conjunctions and disjunctions. The first attempt of clarifying 'what a concept is' can be traced back to Aristotle and is known as the 'classical', or 'rule-based' view of concepts. According to the classical view, all instances of a concept share a common set of necessary and sufficient defining properties. In this view, a concept can be identified with a 'set of instantiations'. However, it was already known in the 1950s that (i) in some cases, one cannot give a set of characteristics or rules defining a concept; (ii) it is often unclear whether an object is a member of a particular category; (iii) conceptual membership of an instance strongly depends on the context [38]. A major blow to the classical view came from Rosch's work on color [39, 40]. This work showed 6 that colors do not have any particular criterial attributes or definite boundaries, and instances differ with respect to how typical they are of a concept. This led to formulation of 'prototype theory', according to which concepts are organized around family resemblances, and consist of characteristic, rather than defining, features [41]. These features are weighted in the definition of the 'prototype'. Rosch showed that people rate conceptual membership as 'graded', with degree of membership of an instance corresponding to conceptual distance from the prototype. Another major approach to concept theory comes from 'exemplar theory', according to which a concept is represented by a set of 'salient instances' of it stored in memory, rather than a set of defining or characteristic features [42, 43]. A third recognized approach is 'theory theory', according to which concepts take the form of 'mini-theories' [44] or schemata [45], in which the causal relationships among properties are identified. More recently, other theories of concepts have also been proposed, e.g., constraint theory [46], 'connectionist CONCAT model' [47] and 'emergent binding model [48]. Coming to concept representation, for years there has been an implicit assumption that classical set theoretical structures could be used to represent concepts and their combinations. Obviously, taking into account the above aspects of concepts, like 'context-dependence', 'vagueness' and 'graded typicality', re- quired introducing probabilistic structures and fuzzy notions, like the 'fuzzy set representation of concepts' [49]: for every item X, a concept A is associated with a 'membership function' µ(A) in such a way that the conjunction 'A and B' and the disjunction 'A or B' of two concepts A and B respectively satisfy the 'minimum rule of fuzzy set conjunction' and the 'maximum rule of fuzzy set disjunction', i.e. (cid:104) (cid:104) (cid:105) (cid:105) µ(A and B) = min µ(A), µ(B) µ(A or B) = max µ(A), µ(B) (1) (2) This would still allow one to associate classical (fuzzy) set theoretical connectives to conceptual conjunction and disjunction. However, a whole set of experimental findings revealed that these these fuzzy set theoretical structures are still insufficient to model even simple combinations of two concepts, as follows. (i) 'Guppy effect', or 'Pet-Fish problem': people judge an item like Guppy to be a very typical example of the conjunction Pet-Fish, without judging Guppy to be a typical example of either Pet or Fish [3]. (ii) 'Over-extension and under-extension of membership weights': people judge the membership weight of the conjunction (disjunction) to be higher (lower) than the membership weight of one or both the component concepts [4, 5]. (iii) 'Borderline contradictions': people judge a sentence like "John is tall and John is not tall" to be true for some borderline case John [6]. Findings (i) -- (iii) raise the so-called 'combination problem', i.e. how consistently representing the com- bination of two (or more) concepts in terms of the representation of the component concepts. Recent studies reveal that the deviation from classicality in conceptual categorization is even deeper. To this end let us analyze Hampton's experiments on membership weights. schematized in Figure 1. Consider a pair of natural concepts, say Fruits, Vegetables, their conjunction Fruits and Vegetables, and their disjunction Fruits or Vegetables. Then, consider various items of these concepts, say Apple, Tomato, Broccoli, Olive, Coconut, Mushroom, Almond, Raisin, Acorn, etc. Next, perform a cognitive test in the form of a questionnaire submitted to a sample of participants in which they have to rank on a '7-point Likert scale' {+3, +2, +1, 0,−1,−2,−3}, membership of these items with respect to Fruits, Vegetables and their conjunction Fruits and Vegetables and disjunction Fruits or Vegetables. In this case, +3 corresponds to 'strong membership', while −3 corresponds to 'strong non-membership'. Thus, a person judging Apple as a very strong member of Fruits will write +3 on the questionnaire. Finally, collect the relative frequency of positive answers. In the large number limit, we call 'membership weight' the obtained number. 7 Figure 1. A schematic representation of Hampton's experiments on membership weights, where partici- pants were asked to rank the membership of certain items, like Olive, with respect to Fruits, Vegetables and their disjunction Fruits or Vegetables, on a '7-point Likert scale'. Hampton collected membership weights of several items with respect to pairs of natural concepts and their conjunction [4] and/or disjunction [5]. Most of Hampton's data cannot be represented by a single probability function over a σ-algebra of sets satisfying the axioms of Kolmogorov [1]. We illustrate this important result in the following by means of general results and concrete examples. Let us start by considering the conceptual conjunction. Let X denote an item and let µ(A), µ(B) and µ(A and B) denote the membership weights of X with respect to the concepts A, B and their conjunction 'A and B', respectively. We say that µ(A), µ(B) and µ(A and B) are 'classical conjunction data' if a non-empty set Ω, a σ-algebra A ⊆ P(Ω) (P(Ω) denotes the power set of Ω), a Kolmogorovian probability function1 p : E ∈ A −→ [0, 1], and events EA, EB ∈ A exist such that: µ(A) = p(EA) µ(B) = p(EB) µ(A and B) = p(EA ∩ EB) (3) (4) (5) One can then show that, given an item X, the membership weights µ(A), µ(B) and µ(A and B) are 'classical conjunction data' if and only if the following inequalities are simultaneously satisfied: (cid:104) (cid:105) ≤ 0 µ(A and B) − min µ(A), µ(B) µ(A) + µ(B) − µ(A and B) ≤ 1 (6) (7) Equation (6) expresses monotonicity of probability with respect to the conjunction, while (7) is a more technical result within Kolmogorovian probability [14]. Most of Hampton's data on the conjunction of two concepts violate at least one of the inequalities (6) and (7). In particular, Hampton called 'overextension' a violation of (6) [4]. Consider, for example, the item Razor. Hampton measured membership of Razor with respect to Weapons, Tools and their conjunction Weapons and Tools, finding µ(A) = 0.63, µ(B) = 0.78 and µ(A and B) = 0.83, respectively. Since 0.83 is 1A normalized function p : E ∈ A −→ [0, 1] is said 'Kolmogorovian' if it satisfies the following axioms: (i) p(Ω) = 1 and i p(Ei), for every sequence {Ei}i of pairwise disjoint events Ei [1]. (ii) p(∪iEi) =(cid:80) 8 greater than both 0.63 and 0.78, we get that (6) is violated, and Razor is said to be 'double overextended' with respect to the conjunction Weapons and Tools. Let us now come to conceptual disjunction. Let X denote an item and let µ(A), µ(B) and µ(A or B) denote the membership weights of X with respect to the concepts A, B and their disjunction 'A or B', respectively. We say that µ(A), µ(B) and µ(A or B) are 'classical disjunction data' if a non-empty set Ω, a σ-algebra A ⊆ P(Ω), a Kolmogorovian probability function p : E ∈ A −→ [0, 1], and events EA, EB ∈ A exist such that: µ(A) = p(EA) µ(B) = p(EB) µ(A or B) = p(EA ∪ EB) (8) (9) (10) One can then show that, given an item X, the membership weights µ(A), µ(B) and µ(A or B) are 'classical disjunction data' if and only if the following inequalities are simultaneously satisfied: (cid:104) (cid:105) ≥ 0 µ(A or B) − max µ(A), µ(B) µ(A) + µ(B) − µ(A or B) ≥ 0 (11) (12) Equation (11) expresses monotonicity of probability2 with respect to the disjunction, while (12) is again a more technical result within Kolmogorovian probability [14]. Most of Hampton's data on the disjunction of two concepts violate at least one of the inequalities (11) and (12). In particular, Hampton called 'underextension' a violation of (11) [5]. Consider, for example, the item Ashtray. Hampton measured membership of Ashtray with respect to Home Furnishing, Furniture and their disjunction Home Furnishing or Furniture, finding µ(A) = 0.70, µ(B) = 0.30 and µ(A or B) = 0.25, respectively. Since 0.25 is lower than both 0.70 and 0.30, we get that (11) is violated and Ashtray is said to be 'double underextended' with respect to the disjunction Home Furnishing or Furniture. The conclusion one draws from the above results is immediate. Classical (fuzzy) set theoretical logical and probabilistic structures are too limited to represent even simple combinations of two concepts. On the other hand, a concept is not any more a set of instantiations according to the operational-realistic approach to cognition, as we have explained in Section 2. A concept is rather an entity in a specific state changing under the influence of a context [36]. We elaborated on this, developing a quantum theoretical approach to model concepts and their combinations, showing that it faithfully represents various sets of data on the Guppy effect on typicality [30, 31], over- and under-extensions of membership weights [14, 37] and borderline contradictions [50]. We present here the quantum theoretical model in Hilbert space for the conjunction and the disjunction of two concepts that allows one to represent most of Hampton's data, explaining at the same time the observed deviations from classicality as consequences of genuine quantum effects, namely, 'contextuality', 'emergence', 'interference', and 'superposition'. We use the mathematical formulation and notation of quantum theory that was introduced by Paul Maurice Dirac, one of the founding fathers of quantum theory [52]. Let us start by the quantum model for the conjunction of two concepts. As above, let X denote an item and let µ(A), µ(B) and µ(A and B) denote the membership weights of X with respect to the concepts A, B and their conjunction 'A and B', respectively. We associate this conceptual situation with a Hilbert space H over the field C of complex numbers. Then, we represent A by the unit vector A(cid:105) ∈ H , i.e. (cid:104)AA(cid:105) = 1, and we represent B by the unit vector B(cid:105) ∈ H, i.e. (cid:104)BB(cid:105) = 1. For the sake of simplicity, 2The monotonicty law of Kolmogorovian probability is globally expressed by the inequalities p(EA∩ EB) ≤ p(EA), p(EB) ≤ p(EA ∪ EB). 9 we assume that A(cid:105) and B(cid:105) are orthogonal, i.e. the scalar product (cid:104)AB(cid:105) = 0. Next, we represent the conjunction 'A and B' by the linear combination A and B(cid:105) = (A(cid:105) + B(cid:105)) 1√ 2 (13) Finally, we represent the 'decision measurement' of a person judging membership of X with respect to A, B and 'A and B' by an orthogonal projection operator M over H , i.e. M†M = M .3 We interpret the membership weight as a probability of membership and represent the latter by the Born rule of quantum probability [52]. We have: µ(A) = (cid:104)AMA(cid:105) µ(B) = (cid:104)BMB(cid:105) µ(A and B) = (cid:104)A and BMA and B(cid:105) (14) (15) (16) (cid:104)BMA(cid:105)∗ = By using (14) -- (16) and applying linearity of Hilbert space and hermiticity of M , (cid:104)AM†B(cid:105) = (cid:104)AMB(cid:105) (where (cid:104)BMA(cid:105)∗ is the complex conjugate of (cid:104)BMA(cid:105)), we then get: i.e. µ(A and B) = = ((cid:104)A + (cid:104)B)M (A(cid:105) + B(cid:105)) (cid:16)(cid:104)AMA(cid:105) + (cid:104)AMB(cid:105) + (cid:104)BMA(cid:105) + (cid:104)BMB(cid:105)(cid:17) 1 2 1 2 µ(A) + µ(B) 2 + Re(cid:104)AMB(cid:105) = (17) where Re(cid:104)AMB(cid:105) is the real part of the complex number (cid:104)AMB(cid:105). We recognize in Re(cid:104)AMB(cid:105) the typ- ical interference term of the quantum double-slit experiment. Its presence expresses quantum interference, as this term produces a fluctuation around the average value 1 2 (µ(A) + µ(B)), which is what one would classically expect in the double-slit experiment. two cases, as follows. We now manipulate (17) to get the 'quantum probability formula for the conjunction', splitting it into (I) If µ(A) + µ(B) ≤ 1, we get: µ(A and B) = µ(A) + µ(B) 2 +(cid:112)µ(A)µ(B) cos θc (18) (19) (II) If µ(A) + µ(B) > 1, we get: µ(A and B) = µ(A) + µ(B) 2 +(cid:112)1 − µ(A)(cid:112)1 − µ(B) cos θc where θc is the 'interference angle for the conjunction'. The quantum model above can be effectively realized in the Hilbert space C3, i.e. the set of all ordered triples of complex numbers.4 Then, let {(1, 0, 0), (0, 1, 0), (0, 0, 1)} be the canonical orthonormal (ON) basis of C3. We again distinguish two cases, as follows. idempotency, i.e. M 2 = M · M = M . 3We remind that an orthogonal projection operator is a liner operator which satisfies hermiticity, i.e. M† = M , and 4Indeed, A(cid:105) and B(cid:105) are orthogonal vectors, and also MA(cid:105) and (1 − M )A(cid:105) and MB(cid:105) and (1 − M )B(cid:105) are, representing three data points µ(A), µ(B) and µ(A and B) requires a Hilbert space of at least dimension 3. 10 (I) If µ(A) + µ(B) ≤ 1, µ(A), µ(B) (cid:54)= 0, 1, then (14) -- (18) are satisfied by the choice: (cid:17) (cid:16)(cid:112)1 − µ(A), 0,(cid:112)µ(A) (cid:115) (cid:16) − (cid:115) µ(A)µ(B) 1 − µ(A) ,− A(cid:105) = B(cid:105) = eiθc (cid:17) ,(cid:112)µ(B) 1 − µ(A) − µ(B) 1 − µ(A) (20) (21) while the orthogonal projection operator M projects on the one-dimensional subspace generated by (0, 0, 1). (II) If µ(A) + µ(B) > 1, µ(A), µ(B) (cid:54)= 0, 1, then (14) -- (17) and (19) are satisfied by the choice: (cid:16)(cid:112)µ(A), 0,(cid:112)1 − µ(A) (cid:17) (cid:16)(cid:115) (1 − µ(A))(1 − µ(B)) A(cid:105) = B(cid:105) = eiθc µ(A) (cid:115) , (cid:17) ,−(cid:112)1 − µ(B) µ(A) + µ(B) − 1 µ(A) (22) (23) while the orthogonal projection operator M projects on the two-dimensional subspace generated by (1, 0, 0) and (0, 1, 0).5 Let us consider again the item Razor with respect to Weapons, Tools and Weapons and Tools. Since µ(A) + µ(B) = 0.63 + 0.78 = 1.41 > 1, we use the quantum formulas in (19), (22) and (23) to model the data. In particular, we get from (19): (cid:16) 2µ(A and B) − µ(A) − µ(B) (cid:112)1 − µ(A)(cid:112)1 − µ(B) (cid:17) θc = arccos = 64.02◦ (24) Thus, the membership weights µ(A) = 0.63, µ(B) = 0.78 and µ(A and B) = 0.83 are represented by θc = 64.02◦, A(cid:105) = (0.79, 0, 0.61) and B(cid:105) = ei64.02◦ (0.36, 0.81,−0.47) in the quantum model. In this case, double overextension of Razor can be explained as an effect of 'constructive interference' (θc < 90◦) between the concepts Weapons and Tools in the superposition Weapons and Tools. Let us now come to the quantum model for the disjunction. Again, let X denote an item and let µ(A), µ(B) and µ(A or B) denote the membership weights of X with respect to the concepts A, B and their disjunction 'A or B', respectively. We associate this conceptual situation with a complex Hilbert space H , represent A and B by the orthogonal unit vectors A(cid:105) and B(cid:105), respectively, and the disjunction 'A or B' by the normalized superposition with equal weights A or B(cid:105) = 1√ (A(cid:105) + B(cid:105)). Finally, we represent the 'decision measurement' of a person judging membership of X with respect to A, B and 'A or B' by the orthogonal projection M . 2 By using again the Born rule of quantum probability for the membership weights µ(A) and µ(B), we find: µ(A or B) = (cid:104)A or BMA or B(cid:105) = µ(A) + µ(B) + Re(cid:104)AMB(cid:105) 2 As previously, (25) mathematically takes two forms, as follows. (I) If µ(A) + µ(B) ≤ 1, we have: µ(A or B) = (II) If µ(A) + µ(B) > 1, we have: µ(A or B) = µ(A) + µ(B) 2 +(cid:112)µ(A)µ(B) cos θd µ(A) + µ(B) 2 +(cid:112)1 − µ(A)(cid:112)1 − µ(B) cos θd (25) (26) (27) 5The situation in which µ(A) = 0 or µ(B) = 0 requires some further technicalities and a more complex Hilbert space structure, the 'Fock space', which will be introduced later. We do not dwell on this aspect here, for the sake of brevity. 11 Experiment by Hampton on conceptual conjunction [4] µ(A) µ(B) µ(A and B) 0.725 0.825 0.825 θc 76.83◦ A(cid:105) (0.85, 0, 0.52) e−iθcB(cid:105) (0.26, 0.87,−0.42) X=Desk Lamp A =Furniture B =Household Appliances 0.87 0.81 67.56◦ 0.9 (0.93, 0, 0.36) X=Mint A =Food B =Plant 0.5 0.9 0.95 77.02◦ (0.88, 0, 0.48) (0.17, 0.88,−0.44) (0.21, 0.89,−0.39) X=Tree House A =Building B =Dwelling Experiment by Hampton on conceptual disjunction [5] µ(A) µ(B) µ(A or B) 0.5 0.7 0.4 θd 121.09◦ A(cid:105) (0.71, 0, 0.71) e−iθdB(cid:105) (0.55, 0.63,−0.55) X=Rat A =Pets B =Farmyard Animals 0.7 0.7 1 121.90◦ (0.73, 0, 0.68) (0.61, 0.45,−0.66) X=Tomato A =Fruits B =Vegetables 0.4 0.7 0.95 53.13◦ (0.71, 0, 0.71) (0.71, 0,−0.71) X=Cake Tin A =Household Appliances B =Kitchen Utensils Table 1. Quantum modeling of some of Hampton's data on the conjunction and the disjunction of two concepts. where θd denotes the 'interference angle for the disjunction'. Also here, the quantum model can be realized in the Hilbert space C3, and formulas analogous to (20) -- (23) hold with θd in place of θc [14]. Let us consider again the item Ashtray with respect to Home Furnishing, Furniture and Home Furnish- ing or Furniture. Since µ(A) + µ(B) = 0.70 + 0.30 = 1, we use the quantum formulas (26), (20) and (21), with θd in place of θc, to model the data. In particular, we get from (26): = 123.06◦ (28) (cid:16) 2µ(A or B) − µ(A) − µ(B) (cid:112)µ(A)µ(B) (cid:17) θd = arccos Thus, the membership weights µ(A) = 0.70, µ(B) = 0.30 and µ(A or B) = 0.25 are represented by θd = 123.06◦, A(cid:105) = (0.55, 0, 0.84) and B(cid:105) = ei123.06◦ (−0.84, 0,−0.55) in the quantum model. Hence, double underextension of Ashtray can be explained as an effect of 'destructive interference' (90◦ < θd < 180◦) between the concepts Home Furnishing and Furniture in the superposition Home Furnishing or Furniture. The quantum probability formulas (17) and (25) enable representation of most of Hampton's data on conjunctions/disjunctions of 2 concepts. We report some relevant cases in Table 1. It is worth observing, at this stage, that the quantum theoretical model in the Hilbert space H is not able to fully represent conjunctions and disjunctions of two concepts in all their generality. Indeed, a more general structure is needed in this case, because of reasons that will become clear in Section 6. This more general algebraic structure is a 'two-sector Fock space' F , i.e. F = H ⊕ (H ⊗ H ) (⊕ denotes the direct sum of linear vector spaces), where the individual Hilbert space H is called the 'sector 1 of F ', while the tensor product Hilbert space H ⊗ H is called the 'sector 2 of F ' [14, 37, 50, 51]. The Fock space representation explains conceptual phenomena that are much deeper than the ones observed above. To understand what we mean by this, let us study a relevant example, as follows. 12 Hampton measured membership weights of the item Olive with respect to the concepts Fruits, Vegetables and their disjunction Fruits or Vegetables, finding 0.5, 0.1 and 0.8, respectively [5]. In this case, (12) is violated, hence these data are non-classical. Now, consider again Olive and its membership weights with respect to the concepts Fruits, Vegetables and their conjunction Fruits and Vegetables. We performed the test, finding 0.56, 0.63 and 0.65, respectively [51]. In this case, (6) is violated, hence these data are again non-classical. However, we notice that in both the disjunction and the conjunction, Olive is double overextended, with similar weights for the combined concept. This clearly indicates that situations exist where people do not really take into account whether the connective 'or' or the connective 'and' is considered, but they actually estimate whether the given item is a member of the new emergent concept, obtained by combining two concepts, but in a way that is independent of the fact that the combination is realized through a disjunction or a conjunction. An intuitive and general notion of 'conceptual emergence' thus arises, which will be precisely defined in Section 6. The quantum theoretical approach in Fock space enables discovery of new non-classical effects. In this regards, we recently generalized Hampton's experiments, extending them to conjunctions and negations of two natural concepts. We tested in an experiment [51] membership weights of items with respect to: (i) natural concepts A and B; (ii) conceptual negations 'not A', 'not B'; (iii) conjunctions 'A and B', 'A and not B', 'not A and B', and 'not A and not B'. As expected, we identified systematic (double) overextension in all the conjunctions above, as well as deviations from classicality in conceptual negations. But, we also found a new non-classical effect [53, 54]. Indeed we identified unexpected, systematic and significant deviations from the marginal law. Consider, for example, the quantity µ(A and B) + µ(A and not B) − µ(A) (29) In a Kolmogorovian probability framework, (29) should be identically zero, for every item and pair of concepts. Since we know that membership weight data violate Kolmogorovian modeling, one would expect that the left side of (29) is different from 0 and generally depends on the item and the pair of concepts that are tested. We instead found that the following equality is approximately satisfied µ(A and B) + µ(A and not B) − µ(A) ≈ 1 2 (30) In addition, the numerical value on the right side of (30) does not depend on either the item or the pair of concepts that are tested. This non-classical pattern is predicted by our quantum theoretical model in Fock space for the conjunction and negation of two natural concepts. More precisely, it can be shown that the left hand side of (30) is given by interference terms oscillating around the value 0.5, which will generally compensate each other. Thus, observing that emergent reasoning (represented by sector 1 of Fock space) typically prevails over logical reasoning (represented by sector 2 of Fock space), as confirmed by the experimental data, it follows that the approximate equality (30) will hold in general [53, 54]. This means that our quantum model is able to capture a robust non-classical effect which goes much deeper than just over- and under-extension, but is part of an intrinsic mechanism of concept formation [37, 54, 55]. Hence, the quantum theoretical approach for conceptual entities and their combinations produces highly predictive models, which have allowed us to identify further genuine quantum effects in the combination of concepts. For example, we identified an experimental violation of classical Bell's inequalities for the concept combination The Animal Acts, and elaborated a quantum representation for it, thus proving that the violation can be explained in terms of 'quantum entanglement' [56, 57]. We also detected 'quantum indistinguishability of Bose-Einstein type' in specific combinations of identical concepts, e.g., Eleven An- imals [58]. We will not go into detail here about the Fock space modeling of conceptual combinations, nor analyse how entanglement and indistinguishability were identified. We instead limit ourselves to the 13 Figure 2. Quantum interference effects in the disjunction Fruits or Vegetables of the natural concepts Fruits and Vegetables. The typical interference fringes of the double-slit experiment are recognizable. For more details about how this figure was obtained, see for instance [59]. 'one sector' Hilbert space model, and this mainly for three reasons. Firstly, because of its mathemati- cal simplicity and tractability. Secondly, because Hilbert space representations already enable to grasp essential aspects of quantum structures in conceptual combinations (see, for example, Figure 2, where we illustrate an impressive effect of quantum interference we identified in the disjunction of Fruits and Vegetables [59].). Thirdly, because non-classical phenomena identified in other areas of cognitive science can be equally realized in a Hilbert space C3, as we will see in Sections 4 and 5. 4 Application to conjunctive fallacies Non-classical effects have been identified in different domains of cognitive science, as mentioned in Section 1. In this section, we apply the quantum conceptual approach presented in Section 3 to model conjunctive and disjunctive fallacies, which are examples of probability judgment errors. In 1983, Kahneman and Tversky identified an important deviation of classicality in decision science, which is known as the 'conjunction fallacy' [7, 8, 9]. More precisely, Kahneman and Tversky performed a cognitive test on a sample of participants, presenting them a questionnaire containing a story about a liberal woman named 'Linda', as follows. "Linda is 31 years old, single, outspoken and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations." Participants were then interrogated about the likelihood of the following events: (i) Linda is a bank teller; (ii) Linda is a bank teller and is active in the feminist movement. Kahneman and Tversky found that 85% of the respondents judged option (ii) to be more likely than option (i). Obviously, the test violates monotonicity of classical probability, and such violation exactly expresses the conjunction fallacy [7]. Several empirical studies have confirmed the conjunction fallacy, and various competing proposals have 14 been formulated at a theoretical level (for a literarary review, see, for example [60]). As mentioned in Section 1, an influential proposal of solution came from Kahneman and Tversky themselves within their theory of 'individual heuristics and biases' [12]. They formulated the 'judgment heuristics hypothesis', according to which people prefer the judgment heuristics of 'representativeness' to standard logical rea- soning, and option (ii) is more representative than option (i) of Linda's activities. An alternative proposal is the 'misunderstanding hypothesis': people misunderstand some terms of the problem, like 'and' and 'probability'. More recently, some authors have put forward that the conjunction fallacy is a consequence of 'question order effects', that is, people first judge the most likely event (i) and then the event (ii), where (i) and (ii) are incompatible in the standard quantum sense [8, 9]. Morier and Borgida performed a more complete test of the fallacy in 1984 [61]. They asked a sample of participants to rank the likelihood of the following events: (i) Linda is a feminist; (ii) Linda is a bank teller; (iii) Linda is a feminist and a bank teller; (iv) Linda is a feminist or a bank teller. Let us denote by µ(A), µ(B), µ(A and B) and µ(A or B) the mean probability that Linda is judged as a feminist, bank teller, feminist and bank teller, and feminist or bank teller, respectively. Morier and Borgida found the following mean probability judgments' µ(A) = 0.83 > 0.60 = µ(A or B) µ(A and B) = 0.36 > 0.26 = µ(B) (31) (32) Equation (32) expresses a conjunction fallacy, because the probability of the conjunction is judged as higher than the probability of one of the two events. Equation (31) expresses instead a 'disjunction fallacy', because the probability of the disjunction is judged as lower than the probability of one of the two events [61]. By referring to the terminology introduced in Section 3, we recognize strong similarities between con- junction fallacy and conceptual overextension on the one side, and between disjunction fallacy and con- ceptual underextension on the other side. And, indeed, the quantum theoretical model of conceptual conjunction and disjunction can be successfully applied to also model conjunctive and disjunctive fallacies, respectively. Let us start by the conjunction fallacy and let us refer to the quantum representation of the conjunction of two concepts in Section 3. We denote the concept Feminist by A and the concept Bank Teller by B. The conjunction 'A and B' then corresponds to the conceptual conjunction Feminist and Bank Teller. As above, we denote by µ(A), µ(B) and µ(A and B) the mean probability that the item Linda is judged to be a Feminist, a Bank Teller and a Feminist and Bank Teller, respectively. These probabilities can be interpreted as membership weights of the item Linda with respect to the concepts Feminist, Bank Teller and Feminist and Bank Teller, respectively, in accordance with the analysis in Section 3. We associate the overall conceptual situation above with a Hilbert space H over complex numbers. Then, as in Section 3, we represent A and B by the unit vectors A(cid:105) and B(cid:105), respectively, of H . We again assume that A(cid:105) and B(cid:105) are orthogonal, and represent the conjunction 'A and B' by the normalized (A(cid:105) + B(cid:105)), as in(13). Finally, the decision measurement of a person judging superposition state vector 1√ 2 whether Linda is 'a feminist', 'a bank teller', and 'a feminist and a bank teller' is represented by the orthogonal projection operator M over H . By using the Born rule of quantum probability, we can now write (14), (15), (16), where the latter, as we observed already, can be written in the form: (cid:40) µ(A)+µ(B) +(cid:112)µ(A)µ(B) cos θc +(cid:112)1 − µ(A)(cid:112)1 − µ(B) cos θc µ(A and B) = 2 µ(A)+µ(B) 2 if µ(A) + µ(B) ≤ 1 if µ(A) + µ(B) > 1 (33) 15 A(cid:105) = (0.91, 0, 0.41) B(cid:105) = ei121.44◦ (0.39, 0.33,−0.86) (36) in the canonical basis of C3. The explanation we give to the modeling above is that, when the item Linda is considered, together with her story, the concepts Feminist and Bank Teller 'destructively interfere' in the conjunction Feminist and Bank Teller, meant as a newly emerging conceptual entity. We can also provide a quantum representation for the disjunction fallacy in [61]. We now let the (A(cid:105) + B(cid:105)) represent the conceptual disjunction Feminist or Bank Teller. Then, superposition vector (20) (or (22)), (21) (or (23)) and (33) hold with θd in place of θc. In [61], the mean judgment probabilities are µ(A) = 0.83, µ(B) = 0.26 and µ(A or B) = 0.60. Since µ(A) + µ(B) = 1.09 > 1, we get: 1√ 2 (cid:16) 2µ(A or B) − µ(A) − µ(B) (cid:112)1 − µ(A)(cid:112)1 − µ(B) (cid:17) = 81.08◦ θd = arccos (34) (35) (37) (38) This completes the construction of a quantum theoretic model for the conjunction fallacy. One easily shows that the quantum model can be realized in the Hilbert space C3, and (20) and (21) hold in case (I), while (22) and (23) hold in case (II), Section 3. Let us now show that the quantum model faithfully represents conjunctive data in [61]. In this case, the mean judgment probabilities are µ(A) = 0.83, µ(B) = 0.26 and µ(A and B) = 0.36. Since µ(A) + µ(B) = 1.09 > 1, we get from (22), (23) and (33): (cid:16) 2µ(A and B) − µ(A) − µ(B) (cid:112)1 − µ(A)(cid:112)1 − µ(B) (cid:17) = 121.44◦ θc = arccos A(cid:105) = (0.91, 0, 0.41) B(cid:105) = ei81.08◦ (0.39, 0.33. − 0.86) (39) in the canonical basis of C3. As we can see, the quantum model of conceptual disjunction faithfully represents data on the disjunction fallacy. The explanation we give to the modeling above is that, when the item Linda is considered, together with her story, the concepts Feminist and Bank Teller 'constructively interfere' in the disjunction Feminist or Bank Teller, meant as a newly emerging conceptual entity. Several tests on Linda-like stories have been performed in the last thirty years, finding systematic devi- ations from classicality. We report in Table 2 some of these data, together with the corresponding values from the quantum model. For each experimental test, we take the average of the judgment probabilities across the various Linda-like stories. To conclude the section, conjunction and disjunction fallacies can be respectively interpreted as the decision theory counterparts of overextension and underextension of concept theory. The quantum effects of contextuality, interference and superposition can again naturally account for the observed deviations from classicality in these human reasoning fallacies. In analogy with the case of double overextension in conceptual conjunctions, the quantum model also predicts the presence of 'double conjunction fallacies', though it predicts that double conjunction fallacies will be less likely than single conjunction fallacies. Similarly, the quantum model predicts the presence of 'double disjunction fallacies', in analogy with the case of double underextension in conceptual disjunctions. 5 Application to disjunctive effects We have seen in Section 4 that violations of classical logical reasoning occur in concrete human decisions in the form of probability judgement errors. In this section, we want to show that pitfalls of classical decision theory are also observed in the form of decision-making errors. 16 Tests on the conjunction fallacy Experiment µ(A) µ(B) µ(A and B) T & K (1983) F & P (1996) Lu (2015) 0.61 0.61 0.59 0.38 0.37 0.32 0.51 0.39 0.37 θc 88.21◦ 102.15◦ 101.28◦ A(cid:105) (0.62, 0, 0.78) (0.62, 0, 0.78) (0.64, 0, 0.77) e−iθcB(cid:105) (−0.77,−0.16, 0.62) (−0.76,−0.23, 0.61) (−0.68,−0.47, 0.57) Tests on the disjunction fallacy Experiment µ(A) µ(B) µ(A or B) Fisk (2002) Lu (2015) 0.63 0.57 0.33 0.26 0.54 0.54 θd 82.44◦ 71.05◦ A(cid:105) (0.61, 0, 0.79) (0.66, 0, 0.75) e−iθdB(cid:105) (−0.75,−0.33, 0.57) (−0.59,−0.63, 0.51) Table 2. Quantum modeling of mean probability data averaged across different Linda-like tests in Tversky and Kahneman (T & K) [7], Fisk and Pidgeon (F & P) [62] and Lu [63] for the conjunction fallacy, and Fisk [64] and Lu [63] for the disjunction fallacy. As mentioned in Section 1, EUT is the predominant model of decision-making under uncertainty: it prescribes that in the presence of uncertainty, decision makers should choose in such a way to maximize their utility, or degree of satisfaction [2]. However, the 'Ellsberg' [11] and 'Machina paradoxes' [65] reveal that cognitive factors play a fundamental role in concrete decision-making processes which are not accounted for by EUT. One of the building axioms of EUT, Savage's 'sure thing principle', is specifically violated in the Ellsberg and Machina paradoxes. We extensively studied in [24, 66, 67] how the quantum theoretical approach solves these paradoxical situations. For sake of brevity, we do not deal with them here, but we instead discuss the violation of the sure thing principle that is observed in other decision-making experiments. Savage stated the sure thing principle being inspired by the following story [2]. "A businessman contemplates buying a certain piece of property. He considers the outcome of the next presidential election relevant. So, to clarify the matter to himself, he asks whether he would buy if he knew that the Democratic candidate were going to win, and decides that he would. Similarly, he considers whether he would buy if he knew that the Republican candidate were going to win, and again finds that he would. Seeing that he would buy in either event, he decides that he should buy, even though he does not know which event obtains, or will obtain, as we would ordinarily say." Tversky and Shafir tested the sure thing principle in an experiment where they presented a group of students with a 'two-stage gamble', that is, a gamble which can be played two sequential times [9, 10]. At each stage the decision consisted in whether or not playing a gamble that has an equal chance of winning, say $200, or losing, say $100. The key result is based on the decision for the second bet, after finishing the first bet. The experiment included three situations: (i) the students were informed that they had already won the first gamble; (ii) the students were informed that they had lost the first gamble; (iii) the students did not know the outcome of the first gamble. Tversky and Shafir found that 69%, that is, the majority of the students who knew they had won the first gamble, decided to play again; 59%, again the majority of the students who knew they had lost the first gamble, decided to play again; but only 36% of the students who did not know whether they had won or lost chose to play again (equivalently, 64%, that is, the majority, decided not to play again) [10]. The two-stage gamble experiment violates Savage's sure thing principle. Indeed, students generally 17 decide to play again if they know they won, and they also decide to play again if they know they lost, but they generally decide not to play again when they do not know whether they won or lost. More generally, the experiment performed by Tversky and Shafir violates the total law of Kolmogorovian probability.6 Indeed, if we denote by µ(P ) the total probability that a student decides to play again without knowing whether s/he has won or lost the first gamble, by µ(W ) and µ(L) the probability that the student wins and loses the first gamble, respectively, by µ(PW ) the conditional probability that the student decides to play again knowing s/he has won, and by µ(PL) the conditional probability that the student decides to play again knowing s/he has lost, then it is not possible to find any value of µ(W ) and µ(L) = 1− µ(W ) such that µ(PW ) = 0.69 and µ(PL) = 0.59, p(P ) = 0.36 and the law of total probability µ(P ) = µ(W )µ(PW ) + µ(L)µ(PL) (40) is satisfied. This violation of the laws of Kolmogorovian probability is called the 'disjunction effect'. An equivalent formulation of the disjunction effect is known as the 'Hawaii problem', and it is again due to Tversky and Shafir [10], while a disjunction effect also occurs in the 'prisoner's dilemma' [9]. A seemingly plausible explanation is that the origin of the violation of the sure thing principle in the disjunction effect is 'uncertainty aversion', that is, people prefer certain over uncertain events. Let us now work out a quantum theoretical model for the disjunction effect. Like in the disjunction fallacy, we need to suitably apply the quantum modelof Section 3 about the disjunction of two concepts. To this end, let us denote by A the conceptual situation of 'having won the first gamble', by B the conceptual situation of 'having lost the first gamble', and by 'A or B' the conceptual situation of 'having won or lost the first gamble', and we denote by µ(A), µ(B) and µ(A or B), respectively, the corresponding probabilities. The participant (student) has to make a decision whether to play again -- positive outcome, or not to play again -- negative outcome. We associate the overall conceptual situation above with a Hilbert space H over complex numbers. Then, as in Sections 3 and 4, we represent A by the unit vector A(cid:105) and B by the unit vector B(cid:105) of H . We again assume that A(cid:105) and B(cid:105) are orthogonal and represent the disjunction 'A or B' by means of their normalized superposition vector A or B(cid:105) = 1√ (A(cid:105) + B(cid:105)). The decision measurement has two outcomes, say 'yes' and 'no', hence it is represented by the orthogonal projection operator M over H . In the two-stage gamble experiment by Tversky and Shafir, we have that the probability of the outcome 'yes', that is, 'play again', in the 'win' situation, that is, state vector A(cid:105), is 0.69, hence we write µ(A) = 0.69. The probability of the outcome 'yes', that is, 'play again', in the 'lost' situation, that is, the state vector B(cid:105), is 0.59, hence we write µ(B) = 0.59. The probability of the outcome 'yes', that is, 'play again', in the 'win or lost' situation, that is, the superposition state vector A or B(cid:105) = 1√ (A(cid:105) + B(cid:105)), is 0.36, hence we write µ(A or B) = 0.36. 2 2 In accordance with the Born rule of quantum probability, formulae (25)-(27) apply again, so we can write: µ(A or B) = (cid:104)A or BMA or B(cid:105) = µ(A) = (cid:104)AMA(cid:105) µ(B) = (cid:104)BMB(cid:105) (cid:40) µ(A)+µ(B) = 2 µ(A)+µ(B) 2 µ(A) + µ(B) +(cid:112)µ(A)µ(B) cos θd +(cid:112)1 − µ(A)(cid:112)1 − µ(B) cos θd 2 + Re(cid:104)AMB(cid:105) if µ(A) + µ(B) ≤ 1 if µ(A) + µ(B) > 1 (41) (42) (43) 6In Kolmogorovian probability, one proves the law of total probability, namely, p(EA) = p(EB)p(EAEB)+p(E(cid:48) where E(cid:48) B = Ω \ EB denotes the 'complement event' with respect to EB. B)p(EAE(cid:48) B), 18 Experimental tests on the two-stage gamble version of the disjunction effect e−iθdB(cid:105) µ(A) µ(B) µ(A or B) A(cid:105) Experiment T & S (1992) K et al. (2001) L & B (2007) 0.69 0.72 0.63 0.58 0.47 0.45 0.37 0.48 0.41 θd 137.26◦ 107.37◦ 106.75◦ (0.73, 0, 0.68) (0.85, 0, 0.53) (0.79, 0, 0.61) (0.61, 0.45,−0.66) (0.45, 0.51,−0.73) (0.57, 0.36,−0.74) Table 3. Quantum modeling of average data on the-stage gamble tests Tversky and Shafir (T & S) [10], Kuhberger (K) et al. [68] and Lambdin and Burdsal (L & B) [69]. Equation (43) includes both cases (I) and (II) in Section 3 for the disjunction of two concepts. In particular, since µ(A) + µ(B) = 0.69 + 0.59 = 1.28 > 1 in the Tversky-Shafir's experiment we are considering [10], coming to the C3 realization, we get: Since in the Tversky-Shafir's experiment we are considering [10], we have µ(A) + µ(B) = 0.69 + 0.59 = 1.28 > 1, it is (27) that we have to use, so coming to the C3 realization, we get: (cid:16) 2µ(A or B) − µ(A) − µ(B) (cid:112)1 − µ(A)(cid:112)1 − µ(B) (cid:17) = 141.76◦ with the coordinates of two vectors A(cid:105) and B(cid:105), in the canonical basis of C3, being given by: θd = arccos (cid:16)(cid:112)µ(A), 0,(cid:112)1 − µ(A) (cid:17) (cid:16)(cid:115) (1 − µ(A))(1 − µ(B)) (cid:115) , A(cid:105) = B(cid:105) = eiθd = ei141.76◦ (0.43, 0.64,−0.64) µ(A) = (0.83, 0, 0.56) µ(A) + µ(B) − 1 µ(A) (cid:17) ,−(cid:112)1 − µ(B) (44) (45) (46) and the orthogonal projection operator M projecting on the subspace generated by (1, 0, 0) and (0, 1, 0). It is worth reminding that, also in this disjunction effect situation, we have applied two key quan- (A(cid:105) + B(cid:105)) to represent 'A or B', and tum features, namely, superposition, in the linear combination 1√ 2 interference, as the effect appearing in (43). It then follows from the value of the interference angle θd = 141.76◦ that the disjunction effect can be explained as an effect of 'destructive interference' created by the conceptual situations A and B in the combination 'A or B'. It is also important to notice that uncertainty aversion can be interpreted as the effect of the overall cognitive landscape influencing the decision. Hence, effects of context play a crucial role in the disjunction effect too. The two-stage gamble was repeated by other scholars, finding systematic deviations from classicality. We report in Table 3 some of these data, together with the corresponding values from the quantum model. In addition, the quantum theoretical model above also works with the disjunction effect detected in the Hawaii problem and the prisoner's dilemma [24]. In the previous sections, we have briefly illustrated the modeling effectiveness of the quantum theoretical approach to human cognition we have developed in Brussels in the last decade. In the section that follows, we intend to explain where the quantum cognitive effectiveness comes from. 6 A unifying explanatory hypothesis We present in this section an hypothesis we have recently elaborated to explain the appearance of genuine quantum structures in cognitive phenomena. The explanatory hypothesis reveals very stable patterns of 19 human reasoning, enlightening at the same time some fundamental traits of its deepest nature [55]. We have extensively discussed here the fundamental limitations of classical set theoretical formalisms, like those employed to formalize Boolean logic and Kolmogorovian probability, in the modeling of several cognitive phenomena. These deviations from classicality put at stake the descriptive but, to some extent also the normative, foundations of rational decision theory, according to which any source of uncertainty can be formalized probabilistically in a Kolmogorovian sense [9]. The Kahneman-Tversky research programme interprets the deviations above as true fallacies of human reasoning and already provides a step forward, also because it introduces, albeit implicitly and only partially, non-Kolmogorovian models of probability [12]. However, the explanation in terms of individual biases and judgment heuristics, though valid at an intuitive level, cannot be the definitive answer to the problem, as it does not provide any explanation in terms of human reasoning processes. The quantum cognition research programme has the significant advantage of exploiting the modeling flexibility offered by the quantum mechanical formalism with respect to classical set theoretical formalisms. We have however seen in Section 2 that there is no evidence that human reasoning processes can be explained in terms of microscopic quantum processes. Thus, one is naturally led to wonder how and why the quantum formalism is so effective in representing complex cognitive phenomena, like categorization, judgment, decision-making and perception? The quantum theoretical approach to cognition illustrated here suggests a simple explanatory hypothe- sis on the mechanisms that underlie human reasoning. According to our explanatory hypothesis, reasoning is a specifically structured superposition of two processes: 'logical reasoning' and 'emergent reasoning'. Logical reasoning combines conceptual entities, that is, concepts, combinations of concepts, propositions and also more complex decision entities, by applying the rules of Boolean logic, though in some cases enriched by a Kolmogorovian probabilistic semantics (see, e.g., graded membership in conceptual combina- tions, Section 3). Emergent reasoning enables instead formation of combined conceptual entities as newly emerging entities, in the case of concepts, new concepts, in the case of propositions, new propositions, etc., carrying new meaning, linked to the meaning of the constituent conceptual entities, but with a connection that cannot be formalized by the algebra of Boolean logic. These two mechanisms act in superposition in human thought during a reasoning process, the first one being guided by an algebra of logic, the second one following a mechanism of emergence. In the above perspective, human reasoning can be mathematically formalized in a Fock space, where only the first two sectors are active, and the states of conceptual entities are represented by unit vectors of this 'two-sector Fock space'. More specifically, 'sector 1 of Fock space', formed by a single Hilbert space, models 'conceptual emergence', hence the combination of two concepts is represented by a superposition vector of the vectors representing the component concepts in this Hilbert space, allowing quantum interference between conceptual entities to play a role in the process of emergence. 'Sector 2 of Fock space', i.e. a tensor product of two identical copies of this Hilbert space, models a conceptual combination from the combining concepts by requiring the rules of logic for the logical connective used for the combining, i.e. conjunction or disjunction, to be satisfied in a probabilistic setting. This quantum-theoretic modeling suggested us to call 'quantum conceptual thought' the process occurring in sector 1 of Fock space, 'quantum logical thought' the process occurring in sector 2. The relative importance of emergence or logic in a specific cognitive process is measured by the 'degree of participation' of sectors 1 and 2. The abundance of evidence of deviations from classical logical reasoning in concrete human decisions, together with our results in Section 3 on the mechanisms of concept formation identified in the conjunction and negation of natural concepts, induce to draw the conclusion that emergence constitutes the dominant dynamics of reasoning, but a dynamics of logic is systematically present too, though only at a lower level. Coming now to the non-classical effects detected in cognition and analyzed in Sections 3, 4 and 3, 20 we believe that over- and under-extension effects in typicality and membership judgments, conjunctive and disjunctive fallacies, and disjunction effects are a consequence of the dominant dynamics and their nature is emergence. This explains why the quantum formalism in Hilbert space (or sector 1 of Fock space) is so effective to model empirical data about these effects. It follows that paradoxes, fallacies, effects and contradictions are not 'errors from a default logical reasoning' but, rather, natural expressions of the dominant emergent reasoning. On the other side, the results presented in Sections 3, 4 and 3 are so significant that one is naturally led to wonder whether logical aspects do play any role in the reasoning processes generating the different effects. This is particularly evident in conjunctive fallacies and disjunctive effects, where all data can be represented in sector 1 of Fock space and a mechanism of conceptual emergence is sufficient to explain the observed deviations from classicality. We believe that this is due to the fact that original experiments on human judgment and decision-making, such as those by Hampton [4, 5], Tversky-Kahneman [7] and Tversky-Shafir [10], were explicitly designed to identify deviations from classical logical reasoning in concrete decisions, hence they on purpose 'triggered' emergent aspects of human reasoning. The situation is more complex in conceptual categorization, where 'pure classical data' cannot be modeled in sector 1 of Fock space alone, but they already require sector 2, as we have briefly sketched in Section 3. Hence, there are reasons to believe that decision-making experiments can be designed where both Fock space sectors are needed to model the data, because the cognitive effect under study would trigger both aspects of human reasoning, emergent and logical, with a similar level of participation to the overall cognitive process. References [1] Kolmogorov, A. N. (1933). Grundbegriffe der Wahrscheinlichkeitrechnung, Ergebnisse Der Mathe- matik; translated as Foundations of Probability. New York: Chelsea Publishing Company, 1950. [2] Savage, L. (1954). The Foundations of Statistics. New York: John Wiley & Sons. [3] Osherson, D., & Smith, E. (1981). On the adequacy of prototype theory as a theory of concepts. Cognition 9, 35 -- 58. [4] Hampton, J. A. (1988a). Overextension of conjunctive concepts: Evidence for a unitary model for concept typicality and class inclusion. Journal of Experimental Psychology: Learning, Memory, and Cognition 14, 12 -- 32. [5] Hampton, J. A. (1988b). Disjunction of natural concepts. Memory & Cognition 16, 579 -- 591. [6] Alxatib, S., & Pelletier, J. (2011). On the psychology of truth gaps. In Nouwen, R., van Rooij, R., Sauerland, U., & Schmitz, H.-C. (Eds.), Vagueness in Communication (pp. 13 -- 36). Berlin, Heidelberg: Springer-Verlag. [7] Tversky, A. & Kahneman, D. (1983). Extension versus intuitive reasoning: The conjunction fallacy in probability judgment. Psychological Review 90, 293 -- 315. [8] Busemeyer, J. R., Pothos, E. M., Franco, R., & Trueblood, J. S. (2011). A quantum theoretical explanation for probability judgment errors. Psychological Review 118, 193 -- 218. [9] Busemeyer, J. R., & Bruza, P. D. (2012). Quantum Models of Cognition and Decision. Cambridge: Cambridge University Press. [10] Tversky, A., & Shafir, E. (1992). The disjunction effect in choice under uncertainty. Psychological Science 3, 305 -- 309. 21 [11] Ellsberg, D. (1961). Risk, ambiguity, and the Savage axioms. Quarterly Journal of Economic 75, 643 -- 669. [12] Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science 185, 1124 -- 1131. [13] Shah, A. K., & Oppenheimer, D. M. (2008). Heuristics made easy: An effort-reduction framework. Psychological Bulletin 134, 207 -- 222. [14] Aerts, D. (2009). Quantum structure in cognition. Journal of Mathematical Psychology 53, 314 -- 348. [15] Aerts, D., Broekaert, J., Gabora, L., & Sozzo, S. (2013). Quantum structure and human thought. Behavioral and Brain Sciences 36, 274 -- 276. [16] Aerts, D., Gabora, L., & S. Sozzo, S. (2013). Concepts and their dynamics: A quantum -- theoretic modeling of human thought. Topics in Cognitive Science 5, 737 -- 772. [17] Haven, E., & Khrennikov, A. Y. (2013). Quantum Social Science. Cambridge: Cambridge University Press. [18] Pothos, E. M., & Busemeyer, J. R. (2013). Can quantum probability provide a new direction for cognitive modeling? Behavioral and Brain Sciences 36, 255 -- 274. [19] Wang, Z., Solloway, T., Shiffrin, R. M., & Busemeyer, J. R. (2014). Context effects produced by question orders reveal quantum nature of human judgments. Proceedings of the National Academy of Sciences 111, 9431 -- 9436. [20] Bruza, P. D., Wang, Z., & Busemeyer, J. R. (2015). Quantum cognition: A new theoretical approach to psychology. Trends in cognitive sciences 19, 383 -- 393. [21] Melucci, M. (2015). Introduction to Information Retrieval and Quantum Mechanics. Berlin: Springer. [22] Haven, E., & Khrennikov, A. (2016). Statistical and subjective interpretations of probability in quantum-like models of cognition and decision making. Journal of Mathematical Psychology 74, 82 -- 91. [23] Kvam, P. D., Pleskac, T. J., Yu, S., & Busemeyer, J. R. (2016). Interference effects of choice on confidence: Quantum characteristics of evidence accumulation. Proceedings of the National Academy of Sciences 112, 10645 -- 10650. [24] Aerts, D., Haven, E., & Sozzo, S. (2018). A proposal to extend expected utility in a quantum proba- bilistic framework. Economic Theory 65, 1079 -- 1109. [25] Pothos, E. M., Busemeyer, J. R., Shiffrin, R. M., & Yearsley, J. M. (2017). The rational status of quantum cognition. Journal of Experimental Psychology: General 146, 968 -- 987. [26] Aerts, D., Sassoli de Bianchi, M., Sozzo, S., & Veloz, T. (2018, forthcoming). Modeling Meaning Associated with Documental Entities: Introducing the Brussels Quantum Approach. [27] Aerts, D. (1986). A possible explanation for the probabilities of quantum mechanics. Journal of Math- ematical Physics 27, 202 -- 210. [28] Pitowsky, I. (1989). Quantum Probability, Quantum Logic. Lecture Notes in Physics vol. 321. Berlin: Springer. 22 [29] Aerts, D., & Aerts, S. (1995). Applications of quantum statistics in psychological studies of decision processes. Foundations of Science 1, 85 -- 97. [30] 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. [31] Aerts, D., & Gabora, L. (2005). A theory of concepts and their combinations II: A Hilbert space representation. Kybernetes 34, 192 -- 221. [32] Jauch, J. M. (1968). Foundations of Quantum Mechanics. Reading, MA: Addison Wesley. [33] Piron, C. Foundations of Quantum Physics. Reading, MA: Reading. [34] Aerts, D. (1999). Foundations of quantum physics: A general realistic and operational approach. International Journal of Theoretical Physics 38, 289 -- 358. [35] Aerts, D. (2002). Being and change: foundations of a realistic operational formalism. In: D. Aerts, M. Czachor and T. Durt (Eds.), Probing the Structure of Quantum Mechanics: Nonlinearity, Nonlocality, Probability and Axiomatics, pp. 71 -- 110. Singapore: World Scientific. [36] Aerts, D., Sassoli de Bianchi, M., & Sozzo, S. (2016). On the foundations of the Brussels operational- realistic approach to cognition. Frontiers in Physics doi: 10.3389/fphy.2016.00017 [37] Aerts, D., & Sozzo, S. (2016). Quantum structure in cognition: Origins, developments, successes and expectations, in The Palgrave Handbook of Quantum Models in Social Science: Applications and Grand Challenges, Haven, E., & Khrennikov, A. Eds., 157-193 (London: Palgrave & Macmillan). [38] Wittgenstein, L. (1953/2001). Philosophical Investigations. Blackwell Publishing. [39] Rosch, E. (1973). Natural categories, Cognitive Psychology 4, 328 -- 350. [40] Rosch, E. (1978). Principles of categorization. In Rosch, E. & Lloyd, B. (Eds.), Cognition and catego- rization. Hillsdale, NJ: Lawrence Erlbaum, pp. 133 -- 179. [41] Rosch, E. (1983). Prototype classification and logical classification: The two systems. In Scholnick, E. K. (Ed.), New trends in conceptual representation: Challenges to Piaget theory?. New Jersey: Lawrence Erlbaum, pp. 133 -- 159. [42] Nosofsky, R. (1988). Exemplar-based accounts of relations between classification, recognition, and typicality. Journal of Experimental Psychology: Learning, Memory, and Cognition 14, 700 -- 708. [43] Nosofsky, R. (1992). Exemplars, prototypes, and similarity rules. In Healy, A., Kosslyn, S., & Shiffrin, R. (Eds.), From learning theory to connectionist theory: Essays in honor of William K. Estes. Hillsdale NJ: Erlbaum. [44] Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual coherence. Psychological Review 92, 289 -- 316. [45] Rumelhart, D. E., & Norman, D. A. (1988). Representation in memory. In Atkinson, R. C., Hernsein, R. J., Lindzey, G., & Duncan, R. L. (Eds.), Stevens handbook of experimental psychology. New Jersey: John Wiley & Sons. [46] Costello, J., & Keane, M. T. (2000). Efficient creativity: Constraint-guided conceptual combination. Cognitive Science 24, 299 -- 349. 23 [47] Van Dantzig, S., Raffone, A., & Hommel, B. (2011). Acquiring contextualized concepts: A connec- tionist approach. Cognitive Science 35, 1162 -- 1189. [48] Thagard, P., & Stewart, T. C. (2011). The AHA! experience: Creativity through emergent binding in neural networks. Cognitive Science 35, 1 -- 33. [49] Zadeh, L. (1982). A note on prototype theory and fuzzy sets. Cognition 12, 291 -- 297. [50] Sozzo, S. (2014). A quantum probability explanation in Fock space for borderline contradictions. Journal of Mathematical Psychology 58, 1 -- 12. [51] Sozzo, S. (2015). Conjunction and negation of natural concepts: A quantum-theoretic modeling. Jour- nal of Mathematical Psychology 66. 83 -- 102. [52] Dirac, P. A. M. (1958). Quantum mechanics, 4th Ed. Oxford: Oxford University Press. [53] Aerts, D., Sozzo, S., & Veloz, T. (2015). Quantum structure of negation and conjunction in human thought. Frontiers in Psychology doi: 10.3389/fpsyg.2015.01447. [54] Aerts, D., Sozzo, S., & Veloz, T. (2015). New fundamental evidence of non-classical structure in the combination of natural concepts. Philosophical Transactions of the Royal Society A 374, 20150095. [55] Aerts, D., Sozzo, S., & Veloz, T. (2015). Quantum structure in cognition and the foundations of human reasoning. International Journal of Theoretical Physics 54, 4557 -- 4569. [56] Aerts, D., & Sozzo, S. (2011). Quantum structure in cognition. Why and how concepts are entangled. Quantum Interaction. Lecture Notes in Computer Science 7052, 116 -- 127. Berlin: Springer. [57] Aerts, D., & Sozzo, S. (2014). Quantum entanglement in conceptual combinations. International Journal of Theoretical Physics 53, 3587 -- 3603. [58] Aerts, D., Sozzo, S., & Veloz, T. (2015). Quantum nature of identity in human concepts: Bose- Einstein statistics for conceptual indistinguishability. International Journal of Theoretical Physics 54, 4430 -- 4443. [59] Aerts, D. (2009a). Quantum particles as conceptual entities: A possible explanatory framework for quantum theory. Foundations of Science 14, pp. 361 -- 411. [60] Moro, R. (2009). On the nature of the conjunction fallacy. Synthese 171, 1 -- 24. [61] Morier, D., & Borgida, E. (1984). The conjunction fallacy: A task specific phenomenon? Personality and Social Psychology Bulletin 10, 243 -- 252. [62] Fisk, J. E., & Pidgeon, N. (1996). Component probabilities and the conjunction fallacy: Resolving signed summation and the low component model in a contingent approach. Acta Psychologica 94, 1 -- 20. [63] Lu, Y. (2015). The conjunction and disjunction fallacies: Explanations of the Linda problem by the equate-to-differentiate model. Integrative Psychological and Behavioral Science, 1 -- 25. [64] Fisk, J. E. (2002). Judgments under uncertainty: Representativeness or potential surprise? British Journal of Psychology 93, 431 -- 449. 24 [65] Machina, M. J. (2009). Risk, ambiguity, and the dark -- dependence axioms. American Economic Review 99, 385 -- 392. [66] Aerts, D., & Sozzo, S. (2016). From ambiguity aversion to a generalized expected utility. Modeling preferences in a quantum probabilistic framework. Journal of Mathematical Psychology 74, 117 -- 127. [67] Aerts, D., Geriente, S., Moreira, C., & Sozzo (2018). Testing ambiguity and Machina preferences within a quantum-theoretic framework for decision-making. Journal of Mathematical Economics doi 10.1016/j.jmateco.2017.12.002. [68] Kuhberger, A., Kamunska, D., & Perner, J. (2001). The disjunction effect: Does it exist for two-step gambles? Organization Behavior and Human Decision Processes 85, 250 -- 264. [69] Lambdin, C., & Burdsal, C. (2007). The disjunction effect reexamined: Relevant methodological issues and the fallacy of unspecified percentage comparisons Organization Behavior and Human Decision Processes 103, 268 -- 276. 25
1501.01854
1
1501
2015-01-08T14:09:01
Summary of Information Theoretic Quantities
[ "q-bio.NC" ]
Information theory is a practical and theoretical framework developed for the study of communication over noisy channels. Its probabilistic basis and capacity to relate statistical structure to function make it ideally suited for studying information flow in the nervous system. As a framework it has a number of useful properties: it provides a general measure sensitive to any relationship, not only linear effects; its quantities have meaningful units which in many cases allow direct comparison between different experiments; and it can be used to study how much information can be gained by observing neural responses in single experimental trials, rather than in averages over multiple trials. A variety of information theoretic quantities are in common use in neuroscience - including the Shannon entropy, Kullback-Leibler divergence, and mutual information. In this entry, we introduce and define these quantities. Further details on how these quantities can be estimated in practice are provided in the entry "Estimation of Information-Theoretic Quantities" and examples of application of these techniques in neuroscience can be found in the entry "Applications of Information-Theoretic Quantities in Neuroscience".
q-bio.NC
q-bio
Summary of Information Theoretic Quantities   Robin  A.A.  Ince1,  Stefano  Panzeri1,2    and  Simon  R.  Schultz3   1  Institute  of  Neuroscience  and  Psychology,  58  Hillhead  Street,  University  of  Glasgow,  Glasgow  G12  8QB,  UK   2  Center  For  Neuroscience  and  Cognitive  Systems,  Italian  Institute  of  Technology,  Corso  Bettini  31  –  38068  Rovereto   (Tn)  Italy   3  Department  of  Bioengineering,  Imperial  College  London,  South  Kensington,  London  SW7  2AZ,  UK     7  pages  3443  words.   Definition Information  theory  is  a  practical  and  theoretical  framework  developed  for  the  study  of  communication  over   noisy  channels.  Its  probabilistic  basis  and  capacity  to  relate  statistical  structure  to  function  make  it  ideally   suited   for   studying   information   flow   in   the   nervous   system.   As   a   framework   it   has   a   number   of   useful   properties:  it  provides  a  general  measure  sensitive  to  any  relationship,  not  only  linear  effects;  its  quantities   have  meaningful  units  which  in  many  cases  allow  direct  comparison  between  different  experiments;  and  it   can   be   used   to   study   how   much   information   can   be   gained   by   observing   neural   responses   in   single   experimental   trials,   rather   than   in   averages   over   multiple   trials.   A   variety   of   information   theoretic   quantities   are   in   common   use   in   neuroscience   –   including   the   Shannon   entropy,   Kullback-­‐Leibler   divergence,  and  mutual  information.  In  this  entry,  we  introduce  and  define  these  quantities.  Further  details   on  how  these  quantities  can  be  estimated  in  practice  are  provided  in  the  entry  “Estimation  of  Information-­‐ Theoretic  Quantities”  and  examples  of  application  of  these  techniques  in  neuroscience  can  be  found  in  the   entry  “Applications  of  Information-­‐Theoretic  Quantities  in  Neuroscience”.   Detailed Description Information theoretic quantities Entropy  as  a  measure  of  uncertainty   Information   theory   derives   from   Shannon’s   theory   of   communication   (Shannon,   1948;   Shannon   and   Weaver,   1949).   Information,   as   we   use   the   word   technically,   is   associated   with   the   resolution   of   uncertainty.   The   underpinning   theoretical   concept   in   information   theory   is   thus   the   measurement   of   uncertainty,   for   which   Shannon   derived   a   quantity   called   entropy,   by   analogy   to   statistical   mechanics.   Shannon  proved  that  the  only  quantity  suitable  for  measuring  the  uncertainty  of  a  discrete  random  variable   X  is:     H X ( ) =− ∑ K P x ( )log x P x ( )     2 (1)   where  X  can  take  a  number  of  values  x  according  to  the  probability  distribution  P(x).  The  constant  K  we   take  to  be  one.  When  the  logarithm  is  taken  to  base  2,  the  resulting  units  of  the  entropy  H(x)  are  called  bits   (when  the  natural  logarithm  is  used  the  term  is  nats).   Entropy  can  be  thought  of  as  a  non-­‐parametric  way  to  measure  the  variability  of  a  distribution.  Spread  out   distributions   (with   high   variability)   will   have   high   entropy   since   all   potential   outcomes   have   similar   probabilities  and  so  the  outcome  of  any  particular  draw  is  very  uncertain.  On  the  other  hand,  concentrated   distributions  (with  low  variability)  will  have  lower  entropy,  since  some  outcomes  will  have  high  probability,   allowing  a  reasonable  guess  to  be  made  about  the  outcome  of  any  particular  draw.  In  this  respect,  entropy   can  be  thought  of  as  a  generalised  form  of  variance;  unlike  variance,  which  is  a  2nd  order  statistic  and  only   meaningful  for  uni-­‐modal  distributions,  entropy  can  give  meaningful  values  for  any  form  of  distribution1.   2 As   an   example,   consider   the   roll   of   an   unbiased   6   sided   die   performed   under   a   cup.   With   no   external   knowledge   about   the   roll,   an   observer   would   believe   any   of   the   numbers   are   equally   likely   –   a   uniform   log 6 .  Now  a   distribution  over  the  6  possible  faces.  From  equation  (1),  the  entropy  of  this  distribution  is   third   party   peeks   under   the   cup   and   tells   our   observer   that   the   die   is   showing   an   even   number.   This   knowledge  reduces  the  uncertainty  about  the  result  of  the  roll,  but  by  how  much?  After  being  told  the  die   is   showing   an   even   number,   the   number   of   possibilities   is   reduced   from   6   to   3,   but   each   of   the   even   log 3.  So  the  reduction  in  the  observers   numbers  remain  equally  likely.  The  entropy  of  this  distribution  is   −= uncertainty,  measured  as  the  difference  in  entropy,  is   ,  or  1  bit.  We  can  quantify   the  knowledge  imparted  by  the  statement  “the  result  is  even”  as  1  bit  of  information.  This  corresponds  to  a   reduction  in  uncertainty  of  a  factor  of  two  (from  6  to  3  possible  outcomes).  For  this  example,  in  both  the   before   and   after   situations   all   the   possibilities   were   equally   likely   (uniform   distributions)   but   the   methodology   can   be   applied   to   any   possible   distribution.   Note   that   the   uniform   distribution   is   the   maximum  entropy  distribution  over  a  finite  set.  Any  other  distribution  would  have  lower  entropy  since  it  is   not  possible  to  be  less  uncertain  about  a  possible  outcome  than  when  all  possibilities  are  equally  likely  –                                                                                                                             log 6 log 3 log 2 2 2 2 2 1  The   differential   entropy   (the   continuous   analogue   of   the   discrete   entropy   discussed   here)   of   a   Normal   distribution  is  proportion  to  the  logarithm  of  the  variance.   there   is   no   structure   to   allow   any   sort   of   informed   guess.   For   further   applications   of   the   concept   of   maximum  entropy  see  the  entry  on  “Estimating  information-­‐theoretic  quantities”.     The  fact  that  H(X)  should  be  maximised  by  a  uniform  distribution  is  one  of  three  axioms  Shannon  started   from  in  order  to  derive  the  form  of  Eq.  (1).  The  others  are  that  impossible  events  do  not  contribute  to  the   uncertainty,  and  that  the  uncertainty  from  a  combination  of  independent  events  should  be  the  sum  of  the   uncertainty  of  the  constituent  events.  These  three  conditions  necessarily  lead  to  Eq.  (1),  although  it  can  be   reached  via  many  other  routes  as  well  (Chakrabarti  and  Chakrabarty,  2005).     Mutual  Information   The   example   of   a   die   roll   motivates   how   a   difference   between   entropies   can   quantify   the   information   conveyed   about   a   set   of   possible   outcomes.   In   the   case   of   two   discrete   random   variables   –   here   we   consider  S,  representing  a  set  of  stimuli  which  are  presented  during  an  experiment,  and  R,  a  set  of  recorded   responses  –  this  is  formalised  in  a  quantity  called  the  mutual  information.  This  is  a  quantity  measuring  the   dependence   between   the   two   random   variables   and   which   can   be   defined   in   terms   of   entropies   in   the   following  three  equivalent  forms:   I R S ( ; )   =− =− = H S H S R H R H R S H R H S H R S , ( ( ( ) ( ) ) ( ( ) ) ) +− ( )     (2)   H(R),H(S)   are   the   individual   entropies   of   each   random   variable   as   discussed   above,   H(R,S)   is   the   joint   entropy   of   the   two   variables   and   H(RS),   H(SR)   are   conditional   entropies.     The   conditional   entropy   is   defined  as     H X Y ( ) = ∑ y P y H X Y ( ) ( = y )     (3)   Where  H(XY=y)  is  defined  as  in  Eq.  (1),  but  with  P(x)  replaced  by  the  conditional  probability  P(xy).  The   conditional  entropy  H(XY)  represents  the  average  entropy  of  X  when  the  value  of  Y  is  known.     The  three  forms  of  Eq.  (2)  above  each  illustrate  a  particular  interpretation  of  the  mutual  information.  From   the   first   form,   we   can   see   it   quantifies   the   average   reduction   in   uncertainty   about   which   stimulus   was   presented  based  on  observation  of  a  response  R  answering  the  question  “if  we  observe  R,  by  how  much   does  that  reduce  our  uncertainty  about  the  value  of  S?”.  Equivalently,  the  second  form  shows  us  that  it   similarly  answers  the  question  “if  we  observe  S,  by  how  much  does  that  reduce  our  uncertainty  about  the   value   of   R?”.     As   the   third   form   shows   explicitly,   mutual   information   is   symmetric   in   its   arguments.   The   third   form   also   shows   that   information   is   the   difference   in   entropy   between   a   model   in   which   the   two   variables  are  independent  (given  by  P(r)P(s)  with  entropy  equal  to  H(R)  +  H(S))  and  the  true  observed  joint   distribution,   P(r,s)   (with   entropy   H(R,S)).   This   shows   that   for   two   independent   variables   the   mutual   information  between  them  is  equal  to  zero  and  illustrates  that  information  measures  how  far  responses   and  stimuli  are  from  independence.     A  common  measure  of  difference  between  probability  distributions  is  the  Kullback-­‐Leibler  (KL)  divergence   (Kullback  and  Leibler,  1951).     D P Q KL ( ) = ∑ x P x ( )log 2 P x ( ) Q x ( )     (4)   Note  that  this  is  not  a  true  “distance”  metric  since  it  is  not  symmetric  –  the  KL  divergence  between  p  and  q   is   different   to   that   between   q   and   p.   As   can   be   seen   from   the   third   expression   of   Eq.   (2),   the   mutual   information  is  just  the  KL  divergence  between  the  joint  distribution  and  the  independent  model  formed  as   the  product  of  the  marginal  distributions:   ) KL ( , ) ) I R S D P r s P r P s ( ) ( ; ( ) P r s ( , ) P r P s ( ) ( ) ( P r s ( , )log = = ∑ 2 rs       (5)   In  the  above  S  and  R,  represent  discrete  random  variables  where  one  is  viewed  as  a  stimulus,  and  the  other   is   a   recorded   normal   response   –   but   these   could   of   course   be   any   two   discrete   random   variables   (for   example,   a   variable   representing   wildtype   vs   a   genetic   manipulation,   behavioural   responses,   or   other   intrinsic   signals),   and   we   will   see   shortly   that   information   and   entropy   can   be   easily   generalised   to   continuous  signals.   A   natural   question   is   “does   mutual   information   correspond   to   a   measure   of   discriminability?”   If   by   discriminability  one  means  the  measure  d-­‐prime  (Green  and  Swets,  1966),  the  answer  is  in  general  “no”.   However,  in  the  specific  case  where  we  are  measuring  the  transmission  of  information  about  one  of  two   equi-­‐probable   stimuli   (or   equivalently,   presence/absence   of   a   stimulus),   mutual   information   has   such   an   interpretation.  This  can  be  seen  by  reaching  for  another  “distance-­‐like  measure”  –  this  time,  one  that  is   symmetric,   the   Jensen-­‐Shannon   (JS)   Divergence,   a   symmetrized   version   of   the   KL   Divergence   (Lin,   1991;   Fuglede  and  Topsoe,  2004):     D P Q JS ( ) = 1 2 ⎛+⎞ D P ⎜ ⎝ KL P Q ⎛+⎞ ⎟ ⎠ 2 + 1 2 D Q KL ⎜ ⎝ P Q 2 ⎟ ⎠     (6)   It   can   be   fairly   easily   seen   that   in   the   case   where   we   have   only   two   stimuli,   s1   and   s2,   the   mutual   information  I(R;S)  can  be  written     I R S ( ; ) = D P r s P r s ( ) ( JS ( 1 ) 2     (7)   Thus   the   mutual   information   can   be   considered   to   measure   how   far   apart   (how   discriminable)   the   distributions   of   responses   are,   given   the   two   stimuli.   Note   that   while   mutual   information   generalises   to   multiple  stimuli,  it  is  not  entirely  clear  that  the  concept  of  discriminability  does,  in  a  useful  way.   In  summary,  the  mutual  information  is  a  measure  of  how  strongly  two  variables  are  related,  similar  in  spirit   to  correlation  but  with  some  specific  advantages.  Firstly,  it  is  a  completely  general  measure;  it  places  no   assumptions  or  models  on  the  variables  under  consideration  and  is  sensitive  to  any  possible  relationships,   including   non-­‐linear   effects   and   effects   in   high   order   statistics   of   the   distributions.   Second,   it   has   meaningful   units   allowing   direct   comparison   across   different   experiments   and   even   with   behavioural   performance.   Finally,   it   permits   several   nice   interpretations   related   to   its   calculation   as   a   single   trial   property  and  involving  its  relationship  to  decoding  performance.   Other  information  theoretic  quantities     Similar  procedures  to  those  used  to  estimate  entropy  and  mutual  information  can  be  used  to  estimate  a   number   of   other   information   theoretic   quantities.   We   mention   several   such   quantities   here   for   completeness.   Multi-­‐information   Note  that  the  mutual  information  between  a  pair  of  random  variables  naturally  generalises  to  the  concept   of  the  multivariate  mutual  information,  or  multi-­‐information:  the  mutual  information  between  n  variables,   I(X1;...;Xn) =   P(x1,...,xn)log2 ∑ x1...xn = I(X1;...;Xn−1) − I(X1;...;Xn−1 Xn) P(x1,...,xn) P(xi) i∏  .   (8)   Note   that   multi-­‐information,   unlike   standard   mutual   information,   can   take   negative   values.   Multi-­‐ information   has   found   use   in   neuroscience   in   the   study   of   patterns   of   activity   in   neural   ensembles   (Schneidman  et  al.,  2006).  Note  that  this  is  not  the  only  way  to  generalize  mutual  information  beyond  two   variables,  and  a  related  quantity,  the  interaction  information,  can  also  be  defined  (McGill,  1954;  Bell,  2003;   Jakulin  and  Bratko,  2003).   Conditional  mutual  information   The  conditional  mutual  information  is  the  expected  value  of  the  mutual  information  between  two  variables   given  a  third  (Cover  and  Thomas,  1991),     I(X;Y Z) = H(X Z)− H(X Y,Z)     (9)   This   quantifies   the   relationship   between   variables   X   and   Y,   while   controlling   for   the   influence   of   Z.   In   neuroscience,  this  can  be  useful  for  example  to  investigate  the  coding  of  correlated  stimulus  features  (Ince   et  al.,  2012).  Consider  two  correlated  stimulus  features  S1  and  S2.  If  it  is  found  that  I(R;S1)  >  0,  this  could  be   because   the   response   is   modulated   by   feature   S1,   but   it   might   be   that   the   response   is   modulated   by   feature   S2   and   the   relationship   between   response   and   S1   follows   from   the   correlation   between   the   features.  Considering  I(R;  S1  S2)  can  resolve  this  situation  and  reveal  if  the  feature  S1  is  truly  represented.     Entropy  and  information  rates   Most  biological  systems  function  not  as  discrete  realisations  from  a  static  process  (like  the  roll  of  a  die),  but   rather  operate  continuously  as  time-­‐varying  dynamic  processes.  This  brings  to  mind  the  notion  of  the  rate   at   which   a   source   generates   information   –   e.g.   if   we   were   tossing   a   coin   once   per   second,   and   the   outcomes  of  the  coin  tosses  were  independent,  then  we  would  be  generating  information  at  a  rate  of  1   bit/sec.  In  general,  the  entropy  rate  of  a  stochastic  process  is  defined  as     h(X) = lim n→∞ 1 n H(X1, X2,..., Xn)  .   By  extension,  the  mutual  information  rate  is     i(R;S) = lim n→∞ 1 n I(r1,r2,...,rn;S)     (10)   (11)   with  units  of  bits/sec.  Asymptotic  entropy  and  information  rates  can  therefore  be  estimated  indirectly  by   calculating  information  for  sufficiently  long  sequences  of  time  bins.  If  the  calculation  is  repeated  for  smaller   and   smaller   time   bins   it   is   possible   to   extrapolate   the   resulting   discrete   information   value   to   the   instantaneous   limit,   making   the   rate   calculation   independent   of   both   sequence   length   and   bin   width   (Strong   et   al.,   1998).   It   is   also   possible   to   calculate   them   directly   using   a   Bayesian   probabilistic   model   (Kennel  et  al.,  2005;  Shlens  et  al.,  2007).   Information  theoretic  quantities  for  continuous  variables   The   quantities   above   are   all   defined   on   random   variables   taking   discrete   values.   However,   entropy   and   other  information  quantities  can  also  be  defined  on  continuous  spaces.  In  the  case  of  entropy,  replacing   summation  by  integration  yields  the  differential  entropy:     hdiff ( X ) = p(x)log2 p(x) X ∫     (12)   Other  information  theoretic  quantities  can  be  defined  analogously,  in  general  by  replacing  sums  over  the   discrete  spaces  with  integrals  over  the  continuous  spaces.  Note  that  differential  entropy  does  not  entirely   generalize  the  properties  of  the  discrete  Shannon  entropy,  and  thus  this  quantity  has  not  been  widely  used   in   neuroscience,   and   is   therefore   beyond   the   scope   of   this   article.   In   contrast,   the   Kullback-­‐Leibler   divergence  and  mutual  information  both  generalize  in  a  straightforward  manner  to  continuous  spaces.  For   instance,  the  mutual  information  between  two  random  variables  with  joint  density  p(x,y)  is     I(X;Y ) = p(x,y)log2 ∫ p(x,y) p(x)p(y) dxdy  .   (13)   Acknowledgements   Research   supported    by   the   SI-­‐CODE   (FET-­‐Open,   FP7-­‐284533)   project   and   by   the    ABC   and   NETT   (People   Programme   and   PITN-­‐GA-­‐2011-­‐289146)     projects   of    the    European    Union's  Seventh  Framework    Programme    FP7  2007-­‐2013.    Marie     Curie     Actions  PITN-­‐GA-­‐2011-­‐290011   References   Bell  AJ  (2003)  The  co-­‐information  lattice.  4th  Int  Sympo  Sium  Indep  Compon  Anal  Blind  Signal  Sep  ICA2003   Nara  Jpn:921g926.   Chakrabarti   CG,   Chakrabarty   I   (2005)   Shannon   entropy:   axiomatic   characterization   and   application.   Int   J   Math  Math  Sci  2005:2847–2854.   Cover  TM,  Thomas  JA  (1991)  Elements  of  information  theory.  Wiley  New  York.   Fuglede   B,   Topsoe   F   (2004)   Jensen-­‐Shannon   divergence   and   Hilbert   space   embedding.   In:   International   Symposium  on  Information  Theory,  2004.  ISIT  2004.  Proceedings,  pp  31–.   Green   DM,   Swets   JA   (1966)   Signal   detection   theory   and   psychophysics.   Wiley   New   York.   Available   at:   http://andrei.gorea.free.fr/Teaching_fichiers/SDT%20and%20Psytchophysics.pdf  [Accessed  January   17,  2014].   Ince  RAA,  Mazzoni  A,  Bartels  A,  Logothetis  NK,  Panzeri  S  (2012)  A  novel  test  to  determine  the  significance  of   neural   selectivity   to   single   and   multiple   potentially   correlated   stimulus   features.   J   Neurosci   Methods  210:49–65.   Jakulin  A,  Bratko  I  (2003)  Quantifying  and  Visualizing  Attribute  Interactions.  arXiv:cs/0308002  Available  at:   http://arxiv.org/abs/cs/0308002  [Accessed  February  13,  2012].   Kennel   MB,   Shlens   J,   Abarbanel   HDI,   Chichilnisky   E   (2005)   Estimating   entropy   rates   with   Bayesian   confidence  intervals.  Neural  Comput  17:1531–1576.   Kullback  S,  Leibler  RA  (1951)  On  Information  and  Sufficiency.  Ann  Math  Stat  22:79–86.   Lin  J  (1991)  Divergence  measures  based  on  the  Shannon  entropy.  IEEE  Trans  Inf  Theory  37:145–151.   McGill  WJ  (1954)  Multivariate  information  transmission.  Psychometrika  19:97–116.   Schneidman  E,  Berry  II  MJ,  Segev  R,  Bialek  W  (2006)  Weak  pairwise  correlations  imply  strongly  correlated   network  states  in  a  neural  population.  Nature  440:1007–1012.   Shannon  CE  (1948)  A  mathematical  theory  of  communication,  Bell  Syst.  Bell  Syst  Tech  J  27:379–423.   Shannon  CE,  Weaver  W  (1949)  The  mathematical  theory  of  communication  (Urbana,  IL.  Univ  Ill  Press  19:1.   Shlens   J,   Kennel   MB,   Abarbanel   HDI,   Chichilnisky   E   (2007)   Estimating   information   rates   with   confidence   intervals  in  neural  spike  trains.  Neural  Comput  19:1683–1719.   Strong  SP,  Koberle  R,  de  Ruyter  van  Steveninck  RR,  Bialek  W  (1998)  Entropy  and  Information  in  Neural  Spike   Trains.  Phys  Rev  Lett  80:197–200.      
1507.08747
2
1507
2015-11-25T08:45:46
Does normal pupil diameter differences in population underlie the color selection of the #dress?
[ "q-bio.NC" ]
The fundamental question that arises from the color composition of the #dress is: 'What are the phenomena that underlie the individual differences in colors reported given all other conditions like light and device for display being identical?'. The main color camps are blue/black (b/b) and white/gold (w/g) and a survey of 384 participants showed near equal distribution. We looked at pupil size differences in the sample population of 53 from the two groups plus a group who switched (w/g to b/b). Our results show that w/g and switch population had significantly ( w/g <b/b, p-value = 0.0086) lower pupil size than b/b camp. A standard infinity focus experiment was then conducted on 18 participants from each group to check if there is bimodality in the population and we again found statistically significant difference (w/g < b/b , p-value = 0.0132). Six participants, half from the w/g camp, were administered dilation drops that increased the pupil size by 3-4mm to check if increase in retinal illuminance will trigger a change in color in the w/g group, but the participants did not report a switch. The results suggest a population difference in normal pupil-size in the three groups.
q-bio.NC
q-bio
Does normal pupil diameter differences in population underlie the color selection of the #dress? Kavita Vemuri*,1 Kulvinder Bisla,1 SaiKrishna Mulpuru,1 Srinivasa Varadarajan2 1. International Institute of Information Technology, Hyderabad, India 2. L V Prasad Eye Institute, Hyderabad (*[email protected]) The fundamental question that arises from the color composition of the #dress is: 'What are the phenomena that underlie the individual differences in colors reported given all other conditions like light and device for display being identical?'. The main color camps are blue/black (b/b) and white/gold (w/g) and a survey of 384 participants showed near equal distribution. We looked at pupil size differences in the sample population of 53 from the two groups plus a group who switched (w/g to b/b). Our results show that w/g and switch population had significantly ( w/g <b/b, p-value = 0.0086) lower pupil size than b/b camp. A standard infinity focus experiment was then conducted on 18 participants from each group to check if there is bimodality in the population and we again found statistically significant difference (w/g < b/b , p-value = 0.0132). Six participants, half from the w/g camp, were administered dilation drops that increased the pupil size by 3-4mm to check if increase in retinal illuminance will trigger a change in color in the w/g group, but the participants did not report a switch. The results suggest a population difference in normal pupil-size in the three groups. 1.INTRODUCTION The debate and initial experiments to understand the reasons for individual variation in the colors people see in an over-exposed blurry photograph of a #dress has bought into spotlight the deliberation on color as perception or physiological. The #dress is stated to be originally blue/black by the designer but when viewed under identical conditions one set of the viewers reported it was white/gold (w/g), while others saw it as blue/black (b/b) and some have switched mostly from w/g to b/b. A task which required participants to color match patches from the dress showed differences between the two color camps due to lightness and not in chromaticity [1]. An online survey with a sample size of 1401 [2] indicated gender differences, with more females and older population choosing w/g; they also found that people who spent longer periods in artificial light reported seeing the color as b/b. A bias towards bluish illumination in interpretation of illuminance and color constancy was reported in another study [3]. Review [4] of the initial findings from [1,2,3] suggests that results need to be analyzed with a volume of literature on color perception from sensory inputs to the underlying neural mechanisms. Testing for variation due to perceptions of light, hue and chromaticity a large hue-angle range with points being in both gold and black was reported [5]. The studies to date confirm that there are three main population groups – the w/g’ers’, b/b’ers and the switchers. The question that still remains is: what could be the reason for individuals to see 2 different colors given all other viewing conditions the same?. To explore the intrigue, we investigated the pupil size variations different colors including the colors of the dress – white, gold, blue (light to dark shades), black, brown and grey. Our initial premise was based on spectral sensitivity of the eye and the role of the pupil in retinal illuminance, and hence an experiment was design to measure pupil response to color from a display to check if the RGB and the colors from the dress would show difference in either the size or response latency as could be seen when shifting from mesopic to photopic level. The human eye's pupil diameter falls between 2 and 8 mm and determines the optical transfer function, and the amount of light that falls on the retina. The fovea in the retina is approximately 1.4 mm in diameter with rod-free areas of 0.5 mm with 0.3 mm being populated by purely by cones (L and M cones). Various factors are known to affect pupil size and pupillary light reflex, like retinal illuminance [6, 7,8], accommodative state [9], and iris color and age [10]. The pupillary light reflex is in turn controlled by photosensitive retinal ganglion cells which contain the photopigment melanopsin [11,12] and also by external inputs channeled by the rods and cones [13]. The cones are responsible for our color sense. The retinal illuminance mediating the pupil light reflex has been studied as a function of luminance, stimulus size and the wavelength (color) and estimated as a product of the corneal field density and the stimulus illuminance [14,15]. The wavelength dependency of corneal field density [8,16] studied by varying the retinal illumninace with flashing colors indicated that pupil light reflex was stimulus illuminance and size dependent under certain incidence conditions retinal Extending illumninace and pupil light reflex, the present study was designed to examine the color (wavelength) modulated pupil diameter changes by presented different color slides in natural viewing conditions – like on a laptop LCD screen. The pupil size variation was studied using an infrared camera based eye tracker at 30 Hz for 23 colors including RGB, black/white. Using a range of colors, we could get the variation within each color the understanding light, of stimuli and the change over the 130 second duration of the experiment. The large data set gives a time- averaged stable pupil size for each participant. 2. METHODS A. Participants: A survey was completed by 384 individuals (male: 290 and female:94, mean age: 22) all of Asian-Indian ethnicity with iris color in shades of brown. Fifty three participants (b/b: 21, w/g: 16 and switch: 16, mean age: 27 years) took part in the pupil size measurement experiment using a Tobii X30 eye tracker. Twenty from this subset were naïve, that is, they had not seen the image before. The artificial dilation using prescription drops was administered by an ophthalmologist on 6 participants from the above set (b/b: 3, w/g: 3, though one ‘switched’ just prior to the experiment). Upon comprehensive eye examination by the ophthalmologist, all subjects were found to have best corrected visual (Snellen) acuity of 20/20 or better and no visual abnormality. Informed consent was obtained from all participants and the eye tracking studies are approved by the human ethics committee of the International Institute of Information Technology, Hyderabad. B. Apparatus The Tobii X30 and X120 eye trackers were used to collect pupil size data for the color slides and the infinity focus experiments respectively. The light flux emitted by the LCD screen of the HP laptop and the background light in the room were measured by a digital light meter (Lutron LX-102). . C. Stimuli and experimental procedure Color Slide experiment: 22 different color shades between blue and red were displayed in no particular color order for 3 seconds on a HP Pavilion laptop screen (15inch). Included were the hues of the dress colors (shades of white, blue, black, brown and gold) obtained by reading the pixel values from a digitized photograph of the #dress. The pixel value at different points on the dress digital photograph was acquired and these values were input into the paintbrush tool to make the color slides. To reduce color/luminance adaptation effects, a filler slide in a shade of gray adjusted using the RGB values was introduced between successive color slides for 3 seconds. The participants were seated at 60 cms from the screen and were instructed to look at the screen as normally as possible. The Tobii X30 eye tracker continually recorded the pupil size at the rate of 30 Hz while the participants looked at the color slides. A total of 43 slides were shown in a lighted room (background illumination: 168 lux) and with the lights in the room turned off (room illumination: 7 lux) and the only light being form the laptop screen. At the beginning and the end of the color slide show the #dress image was shown and the participants were asked to indicate the color composition they saw as b/b or w/g in 2 choice forced alternative manner. In addition, participants were asked if they had at any time ‘switched’; those who answered 'Yes' formed the 3rd group. The subset of participants who switched, mentioned seeing a w/g the first time and then continued to see b/b, except for one who could switch at ‘will’; pupil data was collected from this participant, but not presented in this report. All the participants were given 5 minutes of adaptation time before the experiment. Infinity focus:A black-colored ‘X’ sign was projected on a grayish wall 4 meters away from a participant for 3 seconds in a well-lit room with both artificial and daylight mixed (230lux). An instruction slide at the beginning and the #dress were again shown and the choice noted down at the beginning and end while pupil size was collected using a Tobii X120. Out of the 36 participants, 24 were new and 6 took part in the previous experiment. All other procedures were as before. Artificial Pupil dilation:Six participants from among the 53 went through a comprehensive eye test in an eye clinic and were also tested for color vision. A day after the eye test, the participants took part in the pupil size measurement where slides with the red, green, blue colors as shown in the first experiment plus black was preceded and followed by the dress photograph with the choice question. An ophthalmologist reviewed the reports and administered dilation drop (1 drop) in a light controlled room (background luminance: 168 lux) and the experimental color slides was repeated twice with a gap of 30 minutes between the two presentations. 3. RESULTS AND DISCUSSION Survey :To check for age and gender dependency, a survey with 384 participants (male: 290 and female: 94) all of Asian-Indian origin with iris color in shades of brown and across age groups (10 to 60 years) was conducted. The survey participants viewed the image on their personal electronic devices like laptops, smart phones, desktop monitors and filled an online question with choices: b/b, white/gold or seen both and we found no dependency on age or gender with 48.5% reported seeing the dress in blue/black and 51.5% selecting the white/gold-brown option. With regards to gender the break up was 50% of males and 51% of female participants seeing the dress as blue and black while 49% male and 48% female report the color as white and gold/brown. Around 10% of the survey participants mentioned the ability to switch from white/gold-brown initially to blue/black but not being able to switch back Color slide experiment: Pupil size measurements while viewing colors displayed on a light-emitting LCD screen were collected from 53 respondents with the following break-up: blue/black (21), white/gold (16) and switch (16). For the analysis, 2 participants' data were discarded due to insufficient eye tracking data. To look at group wise pupil size response differences for each color the average pupil size across all subjects in each group for individual colors was plotted for both light conditions (Fig 1a). As average values across the 3 second color stimulus period was considered, pupillary unrest or hippus which is said to induce a low frequency random fluctuation (0.02 to 2Hz) with amplitude of approximately 0.25 [18] effects were not regarded. The average pupil diameter for the b/b group is higher across all the colors compared to the w/g or switch groups. Interestingly, the w/g and switch group had nearly equal pupil size across most of the colors. The adjacent color variation in pupil size is similar for all three groups, with small differences (0.1 to 0.2mm) between the b/b and w/g/switch for grey-red or red-grey and grey-pink-grey-orange sequences. The data collected when the room lights were turned off and the only light emitting device in the room was the laptop screen, the b/b group’s pupil size shows a dc shift downwards with a maximum of 0.49 mm for ‘dark_blue’ and a minimum for the ‘purple’; also the inter-color pupil size change for green-grey-red-grey-orange sequence was 0.2mm on an average in the dark room. Similar and significant changes were not observed for the w/g or switch groups except for the ‘grey_blue’ (0.2mm). The pupil size changes for the RGB and white/black colors adhere to expected gradients with ‘black’ screen inducing a pupil size increase and bright ‘white’ forced a reflexive size decrease. The change noticed in the pupil size to color is in accordance to the way the brain compares the responses from cones types. For example, in the log luminance versus pupil diameter plot (Figure 1b,c) we notice constriction of the pupil for ‘green’ and ‘yellow’ and dilation for ‘black’. The theoretical pupil diameter calculated using the equation: pupil size = 4.9-3tanh (0.4 logL), where L is luminance measured by the light meter (Lx102) at 2inch from the screen for each color [19] is also included in the plots for lighted (Figure 1b) and dark room (Figure 1c) conditions. The statistical significance was estimated by a Wilcoxon-Mann- Whitney Test, with the condition: w/g <b/b , with inputs as the average pupil size across all colors for each member of the group and the p-value was 0.0086 (for lighted room) and 0.0391 (for dark room) indicative of significant difference especially for the lighted room condition between the two groups in pupil size. Considering that the switch and w/g group show similar values we plot the average pupil size for left/right eye estimated from 90 data points from seven colors: black, blue, green, red, grey, yellow, white individual participant for the b/b and w/g groups in the lighted room condition (the trend is similar for dark room data too but for shifted down by 2-3mm for the b/b group). A 2-tailed unequal variance T-Test was also committed for each color across the 2 groups, and the p-values for each of the 7 main colors are listed in the table 1. The within-subject size variation for the colors is substantial as expected from color perception even at constant light intensity of the emitting surface (laptop). Table 1: The p-values for 7 colors from a 2-tailed unequal variance T-Test on the pupil size for the 2 groups (wg, bb). Color GREY YELLOW BLUE GREEN BLACK RED WHITE p-value(b/b, w/g) 0.000710167 0.000407803 0.00004.12045 0.000192334 0.000330135 0.000957234 0.002005474 Infinity focus The difference in pupil size between the two groups evidenced lead us to check if there was a population bimodality and the #dress has exposed it and hence an ‘infinity’ focus experiment as done in regular eye tests was set up. The repeat participants from the first experiment were 6 for each group, while 12 were new with a total of 18 in each group. The average pupil size from 5 seconds of data while focused on a ‘X’ sign 4 meters from the participant is co-plotted in Figure 3 for both groups in same lighting conditions as the first experiment. . The 2- sample T-test at unequal variance between the two groups for the condition wg< bb, estimated a T-value = 2.343 and p-value of 0.0127. Artificial Pupil dilation Would increase in pupil dilation artificially of the w/g especially bring about a change in color choice of the #dress? The hypothesis was prompted by the difference we noticed in pupil size between the b/b and w/g group. Six participants, 3 from each group, were administered dilation drops under the supervision of an ophthalmologist after all the required eye tests. One subject switched just before the pre-dilation test. A smaller slide set with RGB and black interspersed with ‘gray’ slides were shown for 3s and the #dress was shown once before and after the color slide show. The pupil size data collected from pre-dilation drops administration, 20 minutes into dilation state and 35 minutes later show the expected inter-subject variation in the dilation size (pre-dilation:~3.3mm to dilated state: 7.9mm) and except for one participant from the w/g group, larger pupil size was seen in b/b group in dilated state. The pupil size change as a function of the luminance (Figure 4) for the pre-post dilation for the six subjects show that in the dilated state the color response is almost flat for 5 participants expect for the ‘switch’ participant. But the color of the dress reported by the participants did not change from their initial choices, that is, the w/g group did not switch to b/b when dilated. from the Our initial premise of a possible group-wise differential response to colors from the lower or higher wavelengths (Blue or red) was not observed but we found new and interesting results summarized as : a) The average pupil size across all the 20 participants who report black/blue is significant higher compared to the other two groups, which alludes to a possible bimodality in participant set, which has been highlighted by the current #dress experiment setup . The statistically analysis to check for significance also confirms the results. b) No significant difference is noticed for the participants who either see the dress as white/gold or are able to switch. c) As expected the pupil size varies as a function of the color with relative difference between the base gray color and the consecutive color ranging from 0.2 to 0.6 mm and this was found to be approximately uniform across the groups. d) the statistical significance test infinity_focus experiment confirms the findings from the color_slide experiment of a probable bimodal population and that normal pupil size is plausibly associated with the perceived color. The artificial dilation of the pupil did not bring the switch in color choice which suggests that light flux is not the main physiological parameter driving color identification differences. A possibility that needs to be explored is the S,L and M cone density differences in the two groups. Another factor that requires further experiments is to check is by artificially constricting the pupil size the retinal illuminance in the b/b group is reduced and a switch to w/g is realized. It was reported few participants could see b/b and in our study too a few of the w/g participants reported that they could see b/b as the brightness of the screen was reduced by 60% and the b/b group could see the dress as ‘w/g’ at maximum brightness. The b/b group having consistently higher normal pupil dilation also indicates that macular pigment differences in these two groups need to be studied and also if subtle contrast sensitivity difference exists between these two groups [5] that by constricting the pupil a 4. Summary We have presented original results from three experiments that indicate a possible bimodality in normal pupil size within people, which could be a one of the factors that leads to the difference in the colors seen in the #dress. The statistically significant estimate of difference in normal pupil size for the two groups throws up more questions that require systematic experiments by increasing the sample size and by measuring cone density, astigmatism and macular pigmentation differences in the two groups. Acknowledgements The authors sincerely thank the effort put in by Elvis Singhal, summer intern at the institute in collecting the preliminary data. Also thank all the participants and the ophthalmologist, Dr. Preetha, for the enthusiasm and interest in the experiment. Links: #dress: (http://swiked.tumblr.com/post/112073818575/ guys-please-help-me-is-this-dresswhite-and) Tobii: http://www.tobii.com/. References 1. K. R. Gegenfurtner, M. Bloj, M. Toscani. The many colours of ‘the dress’. Current Biology, Volume 25, Issue 13, Pages R543-R544(29 June 2015). 2. R. Lafer-Sousa, K. L. Hermann, Bl R. Conway. Striking individual differences in color perception uncovered by ‘the dress’ photograph. Current Biology, Volume 25, Issue 13, Pages R545-R546 (29 June 2015 3. A.Winkler, and M.A.Webster. Asymmetries in blue–yellow color perception and the color of ‘the dress’. Curr. Biol. 25, R547–R548 (2015). L.Spillmann, J.S.Werner, 4. D. H. Brainard, A. C. Hurlbert. . Colour Vision: Understanding Current Biology.Volume 25, Issue 13, Pages R551-R554 (29 June 2015). #TheDress. 5. M. Melgosa, L. Gómez-Robledo, I.M.Suero, M.D Fairchild . What can we learn from a dress with ambiguous colors?. Color Res. Appl., 40: 525–529 (2015). Heller, F.PerryD.L.Jewett, J.D.LevineAutonomic components of the human pupillary light reflex.InvestOphthalmol Vis Sci. 31:156-16 (1990). 7. H.J.Wyatt, J.F.Musselman. Evidence for an 6. P.H. unbalanced pathway from nasal retina, and for signal cancellation brain-stem. Vision Res.21:513-525(1981). in size influence 8. D.H.McDougal, P.D.Gamlin.The of intrinsically-photosensitive retinal ganglion cells on the spectral sensitivity and response dynamics of the human pupillary light reflex. Vision Res.50 (2010). 9. G.A. Fry. The to accommodation and convergence.Am J Optom. 22: 451–465 (1945). relation of pupil 10. B. Winn, D. Whitaker, D.B. Elliott, N. J. Phillips. Factors Affecting Light-Adapted Pupil Size in Normal Human Subjects. Invest Ophthalmol Vis Sci. 35:1132-1137 (1994). 11. P.D.Gamlin, D.H.McDougal,J. , V.C.Smith, K.Yau, D.M.Dacey. Human and macaque pupil responses driven by melanopsin- containing retinal ganglion cells. Vision Research, 47 (7), 946–954(2007). 12. A.D.Guler, J.L.Ecker , C.M.Altimus, H.W.Liao,S. Hattar. Melanopsin cells are the principal conduits for rod-cone input to non-image-forming vision Nature, 453 (7191), 102– 105 (2008). , G.S.Lall , S.Haq Pokorny Dacey, H.W.Liao, Peterson, F.R.Robinson, V.C. Smith, J.Pokorny, P.D. Gamlin. Melanopsin-expressing ganglion cells in primate retina signal colour and irradiance and project to the LGN. Nature, 433 (7027), 749–754(2005). B.B 13. D.M. 14. B.H.Crawford.The Dependence of Pupil Size upon External Light Stimulus under Static and Variable Conditions.Proceedings of the Royal Society of London. Series B - Biological Sciences, 121(823), 376–395(1936). 15. P.A.Stanley, A.K.Davies. The effect of field of view size on steady state pupil diameter.Ophthalmic & Physiological Optics, 15(6), 601–603 (1995). 16. J.C.Park, J.J. McAnany. Effect of stimulus size and luminance on the rod-, cone-, and melanopsin- mediated pupillary light reflex Journal of Vision, 15(3):13, 1–13 (2015). 17. L.Stark, F.W. Campbell, J. Atwood. Nature, 182(4639), 857–858 (1958). 18. P. Moon, D.E. Spencer. (1944). On the Stiles- Crawford effect. Journal of the Optical Society of America, 34(6), 319–329 (1944).
1609.08980
1
1609
2016-09-28T15:59:12
Assessment of corticospinal tract dysfunction and disease severity in amyotrophic lateral sclerosis
[ "q-bio.NC", "q-bio.TO" ]
The upper motor neuron dysfunction in amyotrophic lateral sclerosis was quantified using triple stimulation and more focal transcranial magnetic stimulation techniques that were developed to reduce recording variability. These measurements were combined with clinical and neurophysiological data to develop a novel random forest based supervised machine learning prediction model. This model was capable of predicting cross-sectional ALS disease severity as measured by the ALSFRSr scale with 97% overall accuracy and 99% precision. The machine learning model developed in this research provides a new, unique and objective diagnostic method for quantifying disease severity and identifying subtle changes in disease progression in ALS.
q-bio.NC
q-bio
in amyotrophic Assessment of corticospinal tract dysfunction and disease severity lateral sclerosis. Rahul Remanan, M.B.B.S.*, Viktor Sukhotskiy, Ph.D. graduate student, Mona Shahbazi, N.P., Edward P. Furlani, Ph.D., Dale J. Lange, M.D. * Corresponding author Abstract: The upper motor neuron dysfunction in amyotrophic lateral sclerosis was quantified using triple stimulation and more focal transcranial magnetic stimulation techniques that were developed to reduce recording variability. These measurements were combined with clinical and neurophysiological data to develop a novel random forest based supervised machine learning prediction model. This model was capable of predicting cross-sectional ALS disease severity as measured by the ALSFRSr scale with 97% overall accuracy and 99% precision. The machine learning model developed in this research provides a new, unique and objective diagnostic method for quantifying disease severity and identifying subtle changes in disease progression in ALS. Amyotrophic lateral sclerosis (ALS) is a progressively fatal neuro- degenerative condition. It is associated with a differential involvement of upper motor neuron (UMN) and lower motor neuron (LMN) pathways. To better understand the differential impact of UMN and LMN components on disease related disability and mortality, an objective quantification of different components in the ALS disease process is needed. The UMN involvement in ALS is difficult to measure objectively. Its contribution to overall disease severity is important, yet not well quantified. The most widely used clinical measures for UMN involvement are: identification of overactive reflexes, detection and grading spasticity, eliciting a plantar response and timed repetitive functional tests. Such findings are not specific and difficult to precisely quantitate. Important components of the UMN system are the motor cortex and the corticospinal tract. The corticospinal tract is the pathway responsible for conveying kinematic (1, 2) and dynamic (3) state information from motor networks responsible for the control of voluntary muscle activity. Any functional loss in the corticomotoneuronal pathway can manifest as a loss of motor function (1) and impact disease severity in ALS. Hence, it is important to understand the disease severity changes in relation to corticospinal tract functional loss. Transcranial magnetic stimulation (TMS) allows non-invasive brain stimulation of the motor cortex and is helpful to quantitate UMN involvement (4). Though TMS measurements are useful for measuring UMN involvement, there is an inherent variability associated with the most commonly used TMS parameters. Therefore, the results are difficult to clinically interpret. A simplified set of TMS parameters that are easier to interpret would have better utility in a clinical setting. Furthermore, these measures could provide us with meaningful insights into the extent of UMN involvement contributing to the overall disease severity. The identification of a set of TMS parameters that enable ALS disease severity prediction was the goal of this study. Methods: Subjects and ethical review of procedures: An approval to perform TMS and clinical assessments of disease severity was obtained from an institutional review board. Thirty eight subjects with clinical features and diagnosis consistent with motor neuron disease (MND) were included in this study. There were no re-imbursements for participating in this study (see supplementary methods). Disease severity measures: The disease severity of subjects with ALS was measured using ALSFRSr(5) and AALS(6). The ALSFRSr is an easy to administer scale and a strong indicator of survival and global function (5). It is well-suited for identifying any possible relationship between disease severity and TMS parameters. Since, all motor evoked potentials (MEP) following TMS were recorded from abductor digiti minimi (ADM), these MEP recordings directly represent the degree of upper extremity UMN dysfunction. However, ALSFRSr lacks an objective measurement of upper extremity functional loss, a requirement for direct comparison of TMS results. The AALS on the other hand is based on a weighted score approach using a variety of subcomponent measures. Therefore, the AALS was included here as a secondary disease severity measure. The subcomponents of AALS include: FVC, timed 20ft walking test, timed standing up from a chair, timed standing up from lying down, timed climbing up and down 4 steps of stairs, timed propelling of wheelchair 20ft, grip force testing using dynamometer, pinch grip testing using pinch gauge, Purdue peg board test, block-and-board test and manual muscle testing for both upper and lower extremities(5). Each of these variables has an independent predictive value for ALS disease severity. The individual measures used in this scale allowed us to dissect specific upper extremity dysfunction and a subcomponents analysis of AALS was therefore important. Also, ALSFRSr and AALS have unique characteristics as disease severity measures in ALS (7), hence the relevance of comparison of both measures. Another key disease severity measure was the impact of ALS disease process on quality of life(8). The McGill quality of life single item scale (MQoLSiS) is an easy to administer scale that can measure ALS related quality of life changes (9). We independently recorded ALSFRSr, total AALS scores and MQoLSiS for disease severity, along with raw measurement values for all the individual clinical measures that constituted AALS. TMS and peripheral nerve stimulation: During each visit, the following TMS parameters were recorded: MEP amplitude, central motor conduction time (CMCT), TST amplitude ratio (rAmp) and TST area ratio (rArea). Motor threshold estimation using MEP/CMAP with a likelihood of >50% was used as a measure for cortical excitability(10). Latency measurements from F-wave responses were used to calculate CMCT. The motor cortex stimulation was performed using a MagPro R-30 (MagVenture Inc., Alpharetta, GA, USA) magnetic stimulator with a figure-of-eight (butterfly) geometry MC-B70 coil (MagVenture Inc., Alpharetta, GA, USA). A figure-of-eight coil has a better ability to localize the area of stimulation than a traditional circular coil as demonstrated below. The electrical stimulations of Wrist (Wr) and erb's point (Ep) along with the corresponding electromyography (EMG) recordings were performed using an EMG console (Natus Neurology Incorporated, Middleton, WI, USA). For the Wr and the Ep electrical stimulation, adhesive 2cm diameter Ag/AgCl disposable electrodes (Natus Neurology Incorporated, Middleton, WI, USA) were used. Similar adhesive electrodes were used to record the responses from ADM muscle (see supplementary methods). We developed three-dimensional (3D) computational models For maximal activation of the corticomotoneuronal pathway using the figure-of-eight coil; a biphasic, counterclockwise current direction pulse was used(11). The stimulation intervals for performing the study were calculated by the TST software. The three sites of stimulation for TSTtest were the motor cortex – Wr – Ep. The TSTtest responses represent the fraction of surviving motor neurons in the corticospinal tract and were the result of timed collisions dictated by the software. The TSTcontrol recordings were made to rule out any peripheral nerve conduction abnormalities by applying Ep – Wr – Ep sequential stimulations. The amplitude and area under the curve of the TSTtest and TSTcontrol responses were used to calculate the TST amplitude and area ratios (12). TMS computational model: to investigate and compare the induced electric (E) field distributions generated by two different TMS coils: 1) a traditional circular coil (MagVenture C-100) and 2) the MagPro R-30 stimulator with MC-B70 figure-of-eight coil. The fields were computed using the Comsol AC/DC module with the Magnetic Fields (mef) physics interface, version 5.2 (see supplementary methods) (13). The coils were energized using a biphasic sinusoidal current. The resultant E fields for the two coil geometries subjected to this excitation are shown in figures 1a,b (see also, supplementary methods). This analysis shows clear differences between the field distributions of the two coil configurations. Specifically, the figure-of-eight coil produces a much more spatially focused E field than the circular coil, which is consistent with previous studies (14). Analysis and prediction model: Statistical analyses were performed using R 64 bit for Microsoft Windows, version 3.3.1(15). A supervised machine learning (ML) model using random decision trees(16, 17) was built using various parameters collected during this study. The goal of this ML model was to develop an accurate predictor for ALSFRSr scores. The ML model was tested against a double-blinded ALS disease severity dataset that was collected after its development. The ML model was also compared against a multiple linear regression model (MLR) for predicting ALSFRSr scores (see supplementary methods). Results: Clinical population: Following screening, a review of patient charts and the administration of informed consent, 38 subjects were included in the motor neuron disease (MND) category and underwent TMS. Applying Shapiro-Wilk test of normality on the age distribution of the subjects included in the study, the normality (p=0.60) of the sample population was established. All 38 subjects had either suspected, possible, probable, laboratory supported definite, or definite ALS according to El- Escorial criteria. The group comprised of 23 men with an average age of 56.26 (SD=11.76) years and 15 women of average age 57.87 (SD=7.69) years. The average duration of the disease for men was 2.52 (SD=2.66) years and for women, 3.80 (SD=5.03) years with no significant difference in disease duration between the two gender groups. Among the 38 subjects, 12 had findings limited to UMN dysfunction lasting for more than 4 years. Significant bulbar dysfunction in the form of dysarthria, swallowing difficulty or respiratory problems was identified in 8 subjects. A bulbar predominant disease process, with minimal or no involvement of upper or lower extremities were observed in 4 subjects. The average ALSFRSr score for men was 41.18 (SD=5.38) and 37.33 (SD=6.72) for women. A statistically significant difference was observed in ALSFRSr scores (p=0.01, 95% CI: 0.90, 6.80) between the two genders. The average AALS score for men was 49.86 (SD=14.94) and 55.27 (SD=19.72) for women. The mean quality of life as measured using McGill quality of life single item (MQoLSiS) questionnaire for men was 7.27 (SD=1.93) and 6.87 (SD=2.60) for women. No significant differences in AALS and MQoLSIS scores were observed between the two genders. The average pain or discomfort associated with the neurophysiology procedures were measured using visual analog scale. The scale reported an average discomfort of 4.00 (SD=2.83) for men and 3.80 (SD=3.02) for women with no significant difference in tolerating the procedure between the two genders. Triple stimulation test ratio: A simple linear regression was calculated to predict the dexterity task performances based on the triple stimulation test. A significant linear relationship was observed between triple stimulation test results and both the dexterity tasks used in this study. For block-and-board test based on the TST amplitude ratio (rAmp) resulted in a significant regression equation (F(1,72)=62.82, R2=0.46, p<0.01). Another simple linear regression using rAmp significantly predicted the Purdue peg board results (F(1,72)=37.16, R2=0.33, p<0.01). A significant linear relationship was also observed between triple stimulation test results and both of the objective measures of upper extremity strength. The resultant simple linear regression equations based on rAmp significantly predicted both the grip force (F(1,72)=8.57, R2=0.09, p<0.01) and lateral pinch grip strength (F(1,72)=4.93, R2=0.05, p=0.03). Other TMS and neurophysiology findings: An MEP response was observed in 29 subjects (76.32%) upon stimulation of both hemispheres. Absent MEP responses occurred in 7 subjects (18.42%). Two subjects (5.26%) had MEP responses from only one cortical hemisphere. Among those with absent MEPs 5 subjects (71.43%) had associated bulbar symptoms. Bulbar symptoms accompanied a significantly lower ALSFRSr score (p<0.01, 95% CI: 1.74, 8.24), MEP amplitude (p<0.01, 95% CI: 0.56, 2.37) and rAmp (p=0.04, 95% CI: 1.14, 42.02), compared to subjects without bulbar symptoms. Bulbar symptoms were also accompanied by a significantly higher motor threshold (p<0.01, 95% CI: -22.93, -3.54). No significant differences in CMAP amplitudes were observed between subjects with prominent bulbar symptoms and those without. A significant linear relationship was observed between CMCT, motor threshold and rAmp. The resultant simple linear regression equations using rAmp significantly predicted the CMCT (F(1,74)=113.70, R2=0.60, p<0.01) and threshold (F(1,74)=145.3, R2=0.66, p<0.01). Both CMCT and motor threshold showed inverse linear relationship with rAmp. Predicting disease severity using neurophysiological measures: the motor Using simple linear regression, we found that changes in rAmp was a significant predictor of disease severity measurements using ALSFRSr (F(1,72)=19.35, R2=0.20, p<0.01) and AALS (F(1,72)=11.03, R2=0.12, p<0.01). Simple linear regression models using the motor threshold and CMCT also (F(1,72)=13.5, R2=0.15, p<0.01), significantly predicted both ALSFRSr (F(1,72)=19.49, R2=0.20, p<0.01) and AALS (F(1,72)=6.70, R2=0.07, p=0.01), (F(1,72)=14.46, R2=0.16, p<0.01)) disease severity scores respectively. A summary of all the linear correlations between ALSFRSr scores and TMS parameters are shown in figure 2a. Combining the TMS parameters together resulted in a significant multiple linear regression equation (F(6,67)=4.35, R2=0.22, p<0.01). But, this MLR model constrained to the TMS data, needed improvement due to: 1) only accounting for a small proportion of the variability in ALSFRSr disease severity scores and 2) the floor effect of the prediction model. Therefore, to improve the prediction model, we combined both clinical and neurophysiological data. A correlation heat map (figure 2b) and the islands of statistically significant correlations plot (figure 2c), helped estimate the relative importance of each variable and build the MLR model for ALSFRSr prediction (see supplementary methods). The resultant MLR model accounted for a very high variance and significantly predicted the ALSFRSr scores (F(21,46)=21.37, R2=0.86, p<0.01). But, this prediction model was less reliable due to: 1) the failure to identify an accurate relationship between ALSFRSr score and several of the predictor variables and 2) poor precision for the predicted values. Therefore, the pooled dataset collected in this study was used to build a random forest based supervised ML model. Each variable used in this ML model had a unique significance in helping predict the disease severity (see supplementary methods). Since the variance accounted for by the MLR and ML prediction models were nearly 90%, a sample size of 38 subjects gave a statistical power of 99.96% for our conclusion that the MLR and ML based prediction models described here were significantly useful for disease prediction in ALS. Also, a subsequent double-blinded testing showed no significant differences between predicted and observed values for both MLR and ML models, with the ML model registering increased accuracy and precision, over the MLR model. Significantly, the random forest based supervised ML model was capable of predicting disease severity with an overall accuracy of >97% and a precision of >99% (table 2, figure 3a,b). Other notable findings were: 1) the local weighted smoothening for the ALSFRSr - triple stimulation test data demonstrated an initial flattened disease progression slope; 2) for the corresponding degree of UMN dysfunction, both the regression lines and the local weighted smoothening lines for the ML prediction values closely followed those for the observed ALSFRSr values (figure 3c). Discussion: Transcranial magnetic stimulation (TMS) follows Faraday-Maxwell's laws of electrodynamics (18). A brief duration electrical current due to a rapidly varying magnetic field is induced in the cortical interneurons. The TMS recordings are often highly variable, due to the nature of how MEPs are generated. The triple stimulation technique (TST) was developed to reduce some of the inherent variability with TMS. The amplitude ratio of TST is representative of the fraction of the functional network of neurons in the corticospinal tract(19). The central motor conduction time (CMCT) quantifies the loss of fastest conducting neurons in the corticospinal tract network(20) and despite its test-retest variability, the longitudinal changes in CMCT has been shown to be a predictor of disease severity(21). Here, we demonstrate for the first time the role of cross-sectional measures of TST amplitude and area ratios, CMCT and motor threshold TMS parameters to be independently useful in predicting the degree of clinical impairment in ALS. We also successfully tested the ML prediction model for ALS disease severity. Our modeling of the decaying magnetic field strength and the primary electric field induced on the surface of the brain, showed a significantly focal electric E field while using a figure-of-eight coil (figure 1a,b). A more focal field increased the accuracy and precision of the TMS recordings by selective activation of ADM. Another important requirement for accurate TST recordings was the nature of the TMS pulse. Usually single pulse protocols rely on a monophasic stimulation pulse. Since TST is dependent on maximal activation of the motor cortex neurons, our use of a biphasic pulse was important (supplementary methods). Improvement of TMS parameters based on physical properties of the brain stimulation coil, produced accurate TMS and TST recordings. An important clinical finding is the strong concordance between measures that measure of dexterity and TMS parameters. Clinical corticomotoneuronal change usually include dexterity measures that involve timed repetitive fine-motor tasks. The degree of corticomotoneuronal dysfunction studies as measured by TST scores significantly predicted timed fine-motor repetitive tasks like block-and-board test and the Purdue peg board tests. We also found that hand weakness was also significantly influenced by corticomotoneuronal dysfunction. We also observed a significant correlation between CMAP amplitude and all four of the TMS parameters. Similar to previous studies, we noticed an inverse relation between distal CMAP amplitude and motor threshold (22). These findings together are suggestive of a possible influence of corticomotoneuronal dysfunction on motor units. An exception to this inverse relation between the CMAP amplitude and the motor threshold was observed among subjects with bulbar symptoms. Bulbar symptoms usually accompany in-excitable motor cortex(23). This finding was consistent with our observation of significantly increased motor threshold, lower MEP amplitudes and lower TST amplitude ratios, despite robust CMAPs that accompanied bulbar symptoms. Since the neurophysiological effect of TMS is mediated predominantly through recruitment of the population of interneurons, these findings suggest an underlying upper motor neuron pathophysiology causing bulbar symptoms. These findings also suggest a differential involvement of motor cortex neuronal population in different phenotypes of ALS and are worth further exploration. We also observed that an initial flattening of the ALSFRSr slope occurred among subjects with early corticomotoneuronal involvement. This flattening of the ALSFRSr slope could be explained in the context of 'e pluribus unum' model analysis of ensembles of accumulators(24), which found that firing rates are largely invariant with respect to ensemble size if the accumulators share at least modestly correlated accumulation rates. The subset of neurons with properties that can be considered outliers compared to the rest did not significantly alter the network properties. This model along with the neurophysiological findings of our study can explain some of the underlying motor cortical changes associated with ALS. loss to exponential The initial flat slope of ALSFRSr in relation to the TMS parameters can be viewed as a state where the overall functional motor network characteristics undergo slower alterations compared in ensemble accumulators, an outcome due to the compensatory nature of the motor networks. Only when the neuronal loss exceeds a very large value, is a significantly observable influence in disease severity and neurophysiological changes recorded. These findings of an initial slow progression of the disease are also consistent with recent observations of cortical changes as the primary event in the disease process (25). The presence of a relatively stable curve of early disease progression compared loss of corticomotoneurons, followed by a rapid deterioration in clinical functions helps explain the non-linear disease progression pattern in ALS(26) and needs further investigation. to a greater functional As an isolated clinical tool, TST is extremely useful due to its small range of normal values. For simple stratification as normal or abnormal findings, the TST recordings are easier to interpret than other TMS parameters. But this simple stratification alone had very little significance in predicting disease severity. Despite its small range for normal values and a significant correlation with disease severity measures, the TST scores were only moderately strong predictors of disease severity. This was applicable for all other individual TMS measures used in this study. However, both clinical and neurophysiological measures, when pooled together using random forest based supervised ML, the ALS disease severity prediction was improved with over 97% overall accuracy, 99% precision and accounting for nearly 90% of the variance in ALSFRSr scores. Both the MLR and ML models were significant predictors of ALSFRSr scores. This indicates the robustness of the predictor variables used in both MLR and ML models. But, there is scope for improvement in accuracy, precision and ability to estimate the variations in disease progression. There have been previous proposals for combining ML and cortical excitability parameters (27). Combining additional TMS measures like the cortical silent period, input-output curve and paired pulse stimulation protocols might help add greater predictive power to the ML model. Also, neurophysiology measures and imaging measures have shown definite promise in identifying ALS disease severity (28). Therefore the usefulness of measures like motor unit number estimation and neuroimaging data including magnetic resonance (MR) spectroscopy, susceptibility weighted imaging and diffusion tensor imaging for improvement of the ML model, also need to be evaluated. In previous studies, the utility of TST as a measure to predict structural abnormalities in ALS (29) and improve diagnostic accuracy of ALS has been demonstrated (30). In our study, we demonstrate for the first time that cross sectional TST measures are moderately strong, significant predictors of disease severity on their own. Using ML, this predictive power could be significantly boosted. The TST amplitude and area ratios were identified as important variables for disease severity prediction. Combining various clinical and neurophysiological measures with TST scores significantly improved the ability of the ML model to identify small changes in ALS disease severity. As a neurophysiology procedure, TST is no more time consuming or painful than a proximal erb's point stimulation and information from this procedure is easy to interpret. In conclusion, quantifying corticomotoneuronal dysfunction using upper extremity TST is extremely useful and should be included in a TMS study of ALS subjects. The random forest based cross-sectional ALSFRSr score prediction technique utilizing TST is a potential future biomarker for ALS disease severity. References: 1. C. E. Lang, M. H. Schieber, Reduced muscle selectivity during individuated finger movements in humans after damage to the motor cortex or corticospinal tract. J Neurophysiol 91, 1722-1733 (2004). 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. H. Ackermann, I. Hertrich, I. Daum, G. Scharf, S. Spieker, Kinematic analysis of articulatory movements in central motor disorders. Mov Disord 12, 1019-1027 (1997). A. Andrykiewicz et al., Corticomuscular synchronization with small and large dynamic force output. BMC Neurosci 8, 101 (2007). W. J. Triggs et al., Transcranial magnetic stimulation identifies upper motor neuron involvement in motor neuron disease. Neurology 53, 605- 611 (1999). P. H. Gordon, R. G. Miller, D. H. Moore, Alsfrs-R. Amyotroph Lateral Scler Other Motor Neuron Disord 5 Suppl 1, 90-93 (2004). V. Appel, S. S. Stewart, G. Smith, S. H. Appel, A rating scale for amyotrophic lateral sclerosis: description and preliminary experience. Ann Neurol 22, 328-333 (1987). A. Voustianiouk et al., ALSFRS and appel ALS scores: discordance with disease progression. Muscle Nerve 37, 668-672 (2008). M. B. Bromberg, Assessing quality of life in ALS. J Clin Neuromuscul Dis 9, 318-325 (2007). J. S. Lou, D. Moore, P. H. Gordon, R. Miller, Correlates of quality of life in ALS: Lessons from the minocycline study. Amyotroph Lateral Scler 11, 116-121 (2010). J. C. Rothwell et al., Magnetic stimulation: motor evoked potentials. The International Federation of Clinical Neurophysiology. Electroencephalogr Clin Neurophysiol Suppl 52, 97-103 (1999). T. Kammer, S. Beck, A. Thielscher, U. Laubis-Herrmann, H. Topka, Motor thresholds in humans: a transcranial magnetic stimulation study comparing different pulse waveforms, current directions and stimulator types. Clin Neurophysiol 112, 250-258 (2001). 12. K. M. Rosler, O. Scheidegger, M. R. Magistris, Corticospinal output and loss of force during motor fatigue. Exp Brain Res 197, 111-123 (2009). 13. www.comsol.com. 14. F. S. Salinas, J. L. Lancaster, P. T. Fox, Detailed 3D models of the induced electric field of transcranial magnetic stimulation coils. Phys Med Biol 52, 2879-2892 (2007). 15. R. C. Team. (2014). 16. 17. L. Breiman, Random forests. Machine learning 45, 5-32 (2001). L. Breiman. (URL http://stat-www. berkeley. edu/users/breiman/RandomForests, R package version, 2006). J. C. Maxwell, A treatise on electricity and magnetism. (Clarendon press, 1881), vol. 1. 18. 19. M. R. Magistris, K. M. Rosler, A. Truffert, J. P. Myers, Transcranial stimulation excites virtually all motor neurons supplying the target muscle. A demonstration and a method improving the study of motor evoked potentials. Brain : a journal of neurology 121 ( Pt 3), 437-450 (1998). 20. D. Claus et al., Central motor conduction in degenerative ataxic disorders: a magnetic stimulation study. J Neurol Neurosurg Psychiatry 51, 790-795 (1988). 21. A. G. Floyd et al., Transcranial magnetic stimulation in ALS: utility of central motor conduction tests. Neurology 72, 498-504 (2009). 22. A. Eisen, M. Swash, Clinical neurophysiology of ALS. Clin Neurophysiol 112, 2190-2201 (2001). 23. A. Eisen, W. Shytbel, K. Murphy, M. Hoirch, Cortical magnetic stimulation in amyotrophic lateral sclerosis. Muscle Nerve 13, 146-151 (1990). 24. B. Zandbelt, B. A. Purcell, T. J. Palmeri, G. D. Logan, J. D. Schall, Response times from ensembles of accumulators. Proc Natl Acad Sci U S A 111, 2848-2853 (2014). 25. P. Menon, M. C. Kiernan, S. Vucic, Cortical hyperexcitability precedes lower motor neuron dysfunction in ALS. Clin Neurophysiol 126, 803-809 (2015). 26. P. H. Gordon et al., Progression in ALS is not linear but is curvilinear. J Neurol 257, 1713-1717 (2010). 27. D. Giordano et al., An integrated computer-controlled system for assisting researchers in cortical excitability studies by using transcranial magnetic stimulation. Computer methods and programs in biomedicine 107, 4-15 (2012). 28. H. Mitsumoto et al., Quantitative objective markers for upper and lower motor neuron dysfunction in ALS. Neurology 68, 1402-1410 (2007). 29. A. M. Grapperon et al., Association between structural and functional corticospinal involvement in amyotrophic lateral sclerosis assessed by diffusion tensor MRI and triple stimulation technique. Muscle Nerve 49, 551-557 (2014). 30. B. U. Kleine et al., Prospective, blind study of the triple stimulation technique in the diagnosis of ALS. Amyotroph Lateral Scler 11, 67-75 (2010). Status 2000 7 4.28 88.47% Figures and tables; Description Random forest regression Number of trees No. of variables tried at each split Mean of squared residuals Variance in ALSFRr scores explained Multiple linear regression Number of variables in the model Mean of squared residuals Variance in ALSFRr scores explained (adjusted R2) Significance level (p) Table 1: Summary of the characteristics of supervised random forest based machine learning and multiple linear regression models for ALSFRSr score prediction. 21 5.23 86.1% <0.01 Type of the model Random forest regression Precision SD Overall 99.42% Blinded data 99.22% Overall 96.95% Blinded data 97.53% 0.65% 97.56% 0.43% 92.06% 3.76% 95.39% 3.67% 86.62% Accuracy SD Multiple linear regression Table 2: Comparison of mean absolute error (MAE) and root mean squared error (RMSE) for machine learning and the multiple linear regression models. 3.70% 6.52% 6.38% 10.96% (a) (b) Figure 1: Computational modeling of E fields: (a) predicted E field norm 40 mm from the center of a circular coil, (b) predicted E field norm 40 mm from the center of a figure-of-eight coil. (a) (b) (c) Figure 2: Summary of the correlations observed between ALS disease severity and different clinical/neurophysiological measures: (a) matrix plot summarizing the significant (p<0.05) linear correlations between various TMS parameters, and ALS disease severities as measured using ALSFRSr, (b) heat-map plot summarizes the strength of correlations between all the modeling variables, (c) the islands of statistical significance plot highlights the significant (p<0.05) linear correlations between all the modeling variables. (a) (b) (c) Figure 3: Summary plots of the accuracy and precision of the machine learning (ML) disease severity for ALS prediction model: (a) histogram density plot of the accuracy and precision errors for the machine learning (ML) model and their corresponding normal distributions, (b) overall accuracy and precision errors for the multiple linear regression (MLR) and ML models compared against the blinded-test accuracy and precision errors for both models, (c) scatter plot comparison of TST amplitude ratio versus the observed (blue) and the ML predicted (red) ALSFRSr scores, with an overlay of simple linear regression (dotted) and local weighted smoothening (solid) lines. Competing financial interests: None of the authors or immediate families have any competing financial relationships. Author contributions: Rahul Remanan, M.B.B.S. The author made substantial contributions to the development of this research project, including the transcranial magnetic stimulation procedures, neurophysiology assessments, clinical assessments, statistical analyses and development of machine learning prediction toolkit. He is the corresponding author of this article, and led the development of both the main text and the supplementary methods sections of this article. Viktor Sukhotskiy, Ph.D. graduate student. The author made substantial contributions to this research project, specifically in the development of computational modeling of the transcranial magnetic stimulations (TMS). He also made significant contributions to the development of the main text and supplementary methods sections of this article, highlighting the details of this computational simulation model of TMS. Mona Shahbazi, N.P. The author made substantial contributions to the development of the research project, especially those sections involving clinical assessments. Her contributions to this article include development of sections detailing the clinical assessments and descriptions of the nature of the disease process in the main text and in the supplementary methods sections. Edward P. Furlani, Ph.D. The author made substantial contributions to the development of the research project, specifically in the development of computational modeling of the transcranial magnetic stimulations (TMS). His role was pivotal in overseeing the development of this precise computational model. He also made significant contributions to the development of the main text and the supplementary methods sections of this article, highlighting the details of computational simulation model of TMS. Dale J. Lange, M.D. The author's contributions to this research project include the development of transcranial magnetic stimulation procedures, clinical assessments and statistical analyses. He had a pivotal role in overseeing the development of the neurophysiology and clinical assessment toolkits used in this study. He also made significant contributions to the development of the main text and supplementary methods of this article, especially those highlighting the clinical assessments and the nature of the disease process. He had a significant role in conceptual development of both the main text and the supplementary methods sections of this research article. Materials & Correspondence: All correspondences and material requests should be addressed to Rahul Remanan. The author can be reached at: [email protected]. Supplementary methods: Magnetic field strength modeling: An understanding of physical mechanisms and properties of TMS is important in reducing measurement variability (1). Since the E field produced by the TMS coil(s) plays a critical role in this process, it is instructive to examine its spatial distribution for traditional coil configurations. To this end, we developed 3D computational models to investigate and compare the E field distributions generated by two different TMS coils, a circular coil (MagVenture C-100) and a butterfly, or "figure-of-eight" coil: the MagVenture MC-B70 model. The B and E fields generated by the coils were computed using the Comsol Multiphysics AC/DC module with the Magnetic Fields (mef) physics interface on Microsoft Windows, version 5.2(2). This program solves the Maxwell system of equations [1]-[8] in the frequency-domain: ∇ × 𝐇 = 𝐉 𝐁 = 𝛁 × 𝐀 𝐄 = − 𝜕𝐀 𝜕𝑡 − ∇ ∙ V 𝐉 = 𝛔𝐄 + 𝝏𝐃 𝝏𝒕 + 𝐉𝐞 𝛁 ∙ 𝐉 = 𝟎 𝐃 = 𝜀0𝜀𝑟𝐄 𝐁 = 𝜇0𝜇𝑟𝐇 𝐧𝟐 ∙ (𝐉𝟏 − 𝐉𝟐) = 0 [1] [2] [3] [4] [5] [6] [7] [8] where, E is the electric field intensity, D is the electric displacement, H is the magnetic field intensity, B is the magnetic flux density, J is the total current density, A is the magnetic vector potential, ε0 is the permittivity of vacuum, εr is the relative permittivity of a medium, μ0 is the permeability of vacuum, μr is the relative permeability, σ is the electrical conductivity, Je is the applied current density, i.e. the current density imposed in the coils, V is the electric scalar potential and n2 is the outward normal from medium 2 at interfaces between two media (J1 and J2 are respectively the current densities of medium 1 and 2). The Comsol Multiphysics software solves eqs. [1]-[8] numerically using the finite element method. The inputs to the computational models include the coil energization Je, the material domains with corresponding boundaries, the material properties (magnetic, electric and dielectric) and appropriate physical boundary conditions. The coils' energization parameters were: a biphasic sinusoidal current with a peak magnitude of 690.7 A per turn and a frequency of 3571 Hz, which was modeled by extending a 280 μs sinusoidal biphasic pulse into a steady-state signal. The biphasic (full-sine) pulse is detailed in the MagPro user guide (Pg. 18). The coil current peak magnitude was derived from the peak di/dt value (155 A/µs) displayed by the MagPro R-30 during actual usage of the coils. The models were implemented using rectangular prism computational domains (CDs), wherein the TMS coil was surrounded by air (figures S1a,b). Three-dimensional meshes of tetrahedral linear elements of second order were used to tessellate the entire CD. The dimensions of the CD were increased and the meshes were refined until a stable solution and convergence errors lower than 1% were achieved. The details of the computational models are summarized in table S1. The physical dimension data for the C-100 and MC-B70 coil material domains were taken from the coil vendor (3, 4). Simulations were carried out using the linear Flexible Generalized Minimum Residual (FGMRES) iterative solver with 10,000 maximum number of iterations. This solver avoided out of memory problems, provided a fast convergence and computing robustness. A fully coupled approach using an automatic Newton method was chosen with a minimum damping factor of 1x10-4 since it can ensure convergence effectiveness. These parameters were chosen because the initial conditions are well-known wherein all electromagnetic quantities are set to zero. The models were computed on a Precision® T7610 workstation (Dell Technologies Inc., Round Rock, TX, USA) with a 24 core Xeon® E5-2697 v2 processor (Intel Corp., Santa Clara, CA, USA) and using 64 GB DDR3 memory. The analysis for each coil took approximately 20 minutes to complete. Screening and subject selection: A total of 57 subjects were screened from a database within a neuromuscular clinic population. From this screening pool of 57 subjects, 38 were included in the motor neuron disease (MND) category and underwent TMS, registering a screen failure rate of 33.33%. Among those 19 subjects who were screen failures 9 (47.37%) did not meet the diagnostic criteria, 7 (36.84%) were cancellations or voluntary withdrawal from the research study, 2 (10.53%) were due to the severity of the disease preventing them to undergo the procedure and 1 (5.26%) due to the discomfort associated with the procedure. Recordings of the CMAP, erb's point, F-wave and MEP responses: The supramaximal distal compound muscle action potential (CMAP) and erb's point responses were required to perform TST. These responses were determined using an incremental stimulation intensity technique. The threshold surface electrode current required to elicit a ≥75μV amplitude response at 100μV sensitivity window from the ADM, was determined. The threshold response intensity was increased to four fold and a second set of responses were recorded. The corresponding responses were considered as the supramaximal distal CMAP and erb's point recordings. The stimulation intensity was increased by 5% to identify the plateauing of supramaximal response amplitudes. Among subjects in whom plateauing of the responses could not be elicited, stimulation intensity was increased by 5% intervals until a plateauing was observed. The corresponding responses were identified as the supramaximal distal CMAP and erb's recordings, among those subjects. A band-pass setting of 2 Hz - 10 kHz without 60Hz filtering was used for all the TST acquisitions(5). To measure the F-wave latency, a set of 10 supramaximal stimulation responses were recorded. The late responses were recorded at sensitivity windows of 100 μV, with a bandpass setting of 2Hz – 3kHz, and no 60Hz filtering. The late response waveforms were averaged and the earliest late response was used to calculate the F-wave latency. The CMCT was calculated using the equation [9], 𝐂𝐌𝐂𝐓 = 𝑀𝐸𝑃𝑙𝑎𝑡𝑒𝑐𝑛𝑦 − 𝐹−𝑤𝑎𝑣𝑒𝑙𝑎𝑡𝑒𝑛𝑐𝑦+𝐷𝑖𝑠𝑡𝑎𝑙 𝐶𝑀𝐴𝑃𝑙𝑎𝑡𝑒𝑛𝑐𝑦−1 2 [9] where, F-wave latency was used to compute the peripheral nerve conduction delay that occurred between the anterior horn of the spinal cord and the ADM(6). All study related recordings were performed in a specially designed room to minimize electrical interference. All recordings were made while the subject was sitting upright. The side from where the responses were recorded was suitably stabilized. The elbow joint was flexed at ~90° angle. The hands were in a supinated position. For MEP recordings, the initial placement of the stimulating coil was in such a way that the point of maximal magnetic field strength was along an imaginary point on the scalp, 4-5cm lateral to the vertex. The coil's longitudinal axis was in the anterior-posterior direction. The horizontal axis was along 45 degrees to an imaginary line connecting the vertex and tragus. The point of optimal 𝑛 𝑖 [10] [11] + 20.82 𝐀𝐋𝐒𝐅𝐑𝐒𝐫 = ∑ 𝛽𝑖𝑥𝑖 𝑛 𝑖 + 40.27 𝐀𝐋𝐒𝐅𝐑𝐒𝐫 = ∑ 𝛼𝑖𝑥𝑖 stimulation for the region of the motor cortex that controls the fifth digit was identified by measuring the ADM responses and adjusting the positioning of the coil accordingly. Contralateral facilitation of the same muscle groups was performed in some subjects to optimize the MEP recordings. During facilitation, an auditory feedback was used to maintain a muscle contraction of ~20-30% of the maximal effort. Multiple linear regression model: Due to the inherent challenges associated with neuroscience data, both MLR models and supervised ML, have their own benefits and pitfalls in pattern predictions (7, 8). Therefore, testing both sets of models was important. First MLR model used only the TMS parameters. The resultant linear eq. [10], where, αi is the coefficient for the variable xi and n is the total number of variables in the predictor model (table S3). Due to the limitations in the TMS model a more detailed MLR model using a combination of both clinical and neurophysiological data was developed for ALSFRSr score prediction. The resultant linear eq. [11], where, βi is the coefficient for the variable xi and n is the total number of variables in the predictor model (table S4). This MLR model was compared against the supervised ML based model that used the same predictor variables. Machine learning model for ALSFRSr prediction: The random forest based supervised machine learning prediction model has been successfully tested for ALSFRSr score predictions from longitudinal data (9, 10). Using this supervised machine learning technique, a cross sectional ALSFRSr score predictor was developed as part of this study. A total of 21 variables were used in the ensemble training by developing predictor vectors for the random forest regression (table S2a,b). The relative importance of each of the predictors was estimated using mean increase in error and mean increase in node purity (figure S2a,b). Using a total of 1280 random objects generated with Marsenne twister pseudorandom generator that can be reproduced (11), a forest of 2000 trees was built and tested (figure S2c). Random number generators are implemented at the architectural level for newer generation microprocessors (12, 13). But, to ensure a wider compatibility and better reproducibility of the prediction code, a pseudorandom number generator implementation was used here. The machine learning computations can be performed with minimal computational overhead. The prediction models were tested in OptiPlex® 9010 workstation (Dell Technologies Inc., Round Rock, TX, USA) with a 4 core Core® i7-3770 processor (Intel Corp., Santa Clara, CA, USA), 4GB DDR3 memory. The necessary code for running the prediction model and de-identified datasets are available on GitHub (14). Statistical tests and testing of the prediction models: The descriptive statistics and within group analysis were done using the code published in the GitHub repository. The outputs from the statistical testing of the dataset are also included here (14). The two-sample Welch's unequal variance t-test was used to identify significant differences within groups. Using the initial recruitment of 35 subjects, the MLR and ML The accuracy, precision and corresponding standard deviations (SD) for both the final models were developed, followed by an addition of three subjects to the study pool; two females and one male. Two had features of bulbar predominant motor neuron disease. These three subjects were used to build a blinded-test data by blinding the observed ALSFRSr scores from the prediction scores. These de-identified blinded datasets are also included in the GitHub repository (14). One of the subjects in the blinded-test dataset underwent the same list of procedures four years prior. Since the goal of this study was to identify cross- sectional neurophysiology and clinical data for ALSFRSr score prediction, this particular dataset was treated as a unique cross-sectional data-point. Addition of this subject also helped in identifying any possible pitfalls in ML model while handling longitudinal data instead of cross-sectional data. Various prediction models were tested against this blinded-test data. prediction models using MLR and ML were calculated using eqs. [12, 13, 14], ∆𝒂𝒄𝒄𝒖𝒓𝒂𝒄𝒚= ∑ (𝟏 − ∆𝒑𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏= ∑ (𝟏 − ∆𝑺𝑫= √𝟏 where, n is the total number of samples used in the prediction model. Both the MLR and ML prediction models computed unique ALSFRSr scores that corresponded to the right and left sides of each subject. These values were computed using a combination of both side dependent and side independent measures. The resultant values were used to compute the precision for both MLR and ML models. Since the ALSFRSr scale values ranges from 0 to 48, the maximum possible value for the scale was the denominator for computing percentage precision. × 𝟏𝟎𝟎 [14] 𝑨𝑳𝑺𝑭𝑹𝑺𝒓𝒑𝒓𝒆𝒅𝒊𝒄𝒕𝒆𝒅.𝒓𝒊𝒈𝒉𝒕−𝑨𝑳𝑺𝑭𝑹𝑺𝒓𝒑𝒓𝒆𝒅𝒊𝒄𝒕𝒆𝒅.𝒍𝒆𝒇𝒕 𝑨𝑳𝑺𝑭𝑹𝑺𝒓𝑶𝒃𝒔𝒆𝒓𝒗𝒆𝒅−𝑨𝑳𝑺𝑭𝑹𝑺𝒓𝒑𝒓𝒆𝒅𝒊𝒄𝒕𝒆𝒅 ) × 𝟏𝟎𝟎 [13] ∑ (∆𝒊) 𝒏 𝒊=𝟏 𝒏 ) × 𝟏𝟎𝟎 [12] 𝒏 𝒊=𝟏 𝒏 𝒊=𝟏 𝑨𝑳𝑺𝑭𝑹𝑺𝒓𝑶𝒃𝒔𝒆𝒓𝒗𝒆𝒅 𝟒𝟖 The random forest regression is based on random sampling of a vector, such that the corresponding tree predictor assumes a numerical value. Therefore the regression results are state dependent on the random attributes of the computational environment running the ML model. This results in slight differences in outputs for each separate run. The error analyses reported here are based on a randomly chosen single run of the prediction model based on the parameters described above. The outputs from the successful execution of the code that were used for the analyses, are also included in the GitHub repository(14). References: 1. D. C. Klooster et al., Technical aspects of neurostimulation: Focus on equipment, electric field modeling, and stimulation protocols. Neurosci Biobehav Rev 65, 113-141 (2016). www.comsol.com. www.magventure.com/en-gb/Products/MagPro-coils/Circular-Coils/C-100-Circular-Coil. www.magventure.com/en-gb/Products/MagPro-coils/Butterfly-Coils/MC-B70-Butterfly-Coil. J. Chu, R. C. Chen, Changes in motor unit action potential parameters in monopolar recordings related to filter settings of the EMG amplifier. Arch Phys Med Rehabil 66, 601-604 (1985). A. Samii, C. A. Luciano, J. M. Dambrosia, M. Hallett, Central motor conduction time: reproducibility and discomfort of different methods. Muscle Nerve 21, 1445-1450 (1998). 2. 3. 4. 5. 6. 7. 8. 9. 10. M. Helmstaedter, The mutual inspirations of machine learning and neuroscience. Neuron 86, 25-28 (2015). P. F. Smith, S. Ganesh, P. Liu, A comparison of random forest regression and multiple linear regression for prediction in neuroscience. J Neurosci Methods 220, 85-91 (2013). R. Kuffner et al., Crowdsourced analysis of clinical trial data to predict amyotrophic lateral sclerosis progression. Nat Biotechnol 33, 51-57 (2015). T. Hothorn, H. H. Jung, RandomForest4Life: a Random Forest for predicting ALS disease progression. Amyotroph Lateral Scler Frontotemporal Degener 15, 444-452 (2014). 11. M. Matsumoto, T. Nishimura, Mersenne twister: a 623-dimensionally equidistributed uniform pseudo-random number generator. ACM Trans. Model. Comput. Simul. 8, 3-30 (1998). https://software.intel.com/en-us/articles/intel-digital-random-number-generator-drng-software- implementation-guide. https://support.amd.com/TechDocs/24594.pdf. R. Remanan. ( https://github.com/rahulremanan/TST.git). 12. 13. 14. Supplementary figures and tables: (a) (b) Figure S1: Computational domain for numerical analysis of B and E fields showing TMS coil(s) in air: (a) single coil, (b) figure-of-eight coil. (a) (b) (c) Figure S2: Relative importance of each predictor vectors used in the random forest based machine learning model for ALSFRSr score prediction: (a) mean decrease in error plot, (b) the expected error rate using Gini impurity index during the random splits training of the prediction model, (c) relative error rate in predicting ALSFRSr compared against the number of prediction trees used to build the ML model. Simulation type Relative electric permittivity Electric conductivity (S/m) Relative magnetic permeability Dimensions C-100 1 6.0 x 107 Coil 1 Air block MC-B70 1 6.0 x 107 1 Coil Air block Coil height: 6 mm Outside diameter: 110 mm Inside diameter: 20 mm Cross sectional area (wire): 19.29 mm2 Length (wire): 3 m Width (x): 500 mm Depth (y): 500 mm Height (z): 750 mm Coil height: 6 mm Outside diameter: 97 mm Inside diameter: 27 mm Cross sectional area of wire: 21 mm2 Length (wire): 3 m Width (x): 500 mm Depth (y): 500 mm Height (z): 750 mm Total number of tetrahedral and boundary mesh elements Degrees- of- freedom 455733 4135240 505332 4612009 Table S1: Details of dimensions of computational domain, dielectric and magnetic properties of materials and media used for simulation of the C-100 and MC-B70 TMS coils. Vector Description of the vector V1 Age in years V2 Body mass index V3 McGill single item quality of life score V4 TST amplitude ratio as percentage V5 TST area ratio as percentage V6 MEP:CMAP amplitude ratio V7 MEP: proximal stimulation amplitude ratio V8 Inverse of central motor conduction time measured in milliseconds V9 Motor threshold as a percentage of the maximal output of the TMS stimulator V10 Forced vital capacity as percentage of the predicted V11 Inverse of the time to complete walking 20ft measured in seconds V12 Inverse of the time to standup from a chair measured in seconds V13 Inverse of the time to climb up and down 4 steps of stairs measured in seconds V14 Inverse of the time to propel a wheelchair 20ft measured in seconds V15 Grip force testing using Jamar dynamometer in kg V16 Lateral pinch grip in kg V17 Timed Purdue peg board test V18 Timed block and board test V19 Manual muscle strength testing of the upper extremity using MRC scale V20 Manual muscle strength testing of the lower extremity using MRC scale V21 Duration of the disease in years V22 Gender (0=male, 1=female) V23 Visual analogue scale V24 Presence or absence of bulbar symptoms (0=no, 1=yes) V25 Race (1=Caucasian, 2=Black/African American, 3=Asian, 4=Hispanic) V26 Handedness (0=left,1=right) V27 Smoking status (0=no, 1=yes) V28 Distal compound muscle action potential (CMAP) amplitude (μV) V29 Proximal stimulation response amplitude (μV) V30 Motor evoked potential amplitude (μV) Mean decrease in error Increase in node purity 14.49 9.87 12.40 0.77 3.37 6.84 5.26 7.82 2.79 15.00 35.56 26.17 26.31 28.89 21.34 13.22 12.12 7.15 5.73 4.21 8.44 54.60 38.64 39.48 17.04 26.37 35.13 27.80 33.82 9.76 47.31 672.24 414.26 231.18 434.36 142.17 55.89 62.14 57.52 9.58 12.94 29.20 Unused vectors for the ML model Table S2: Description of vectors and relative importance of each one when used for the development of random forest based supervised machine learning prediction of ALSFRSr scores (see figure S2a,b). Variable Coefficients t value Significance level B Standard error 1.464 0.148 1.628 0.108 -1.305 0.196 1.398 0.167 -0.034 0.973 1.402 0.165 0.336 0.97 (Constant) 20.8195 14.2171 V4 0.14831 0.09111 V5 -0.1174 0.08995 V6 11.937 8.5384 V7 -0.5043 14.9013 V8 28.6969 20.4667 V9 0.13729 0.14158 Significance codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Table S3: Summary of the multiple linear regression model using TMS measures as independent variables and the ALSFRSr score as the dependent variable. Variable Coefficients t value Significance level . B Standard error 0.08 0.02 0.01 2.93 0.24 1.18 0.01 2.74 -1.61 0.12 -0.26 0.80 -0.67 0.51 -1.19 0.24 0.08 1.81 0.11 1.63 0.80 0.43 0.91 0.11 0.00 3.41 0.72 0.48 3.27 0.00 -1.11 0.27 0.00 4.66 0.25 0.80 0.27 1.11 0.59 0.56 -3.21 0.00 -2.03 0.05 -1.82 0.08 0.04 0.05 0.11 0.29 -0.24 0.15 -0.01 0.04 -0.03 0.04 -5.77 4.85 13.15 7.27 18.87 11.56 0.07 0.00 45.31 13.30 1.86 2.59 25.03 7.65 -13.87 12.48 0.02 0.09 0.04 0.01 0.08 0.09 0.02 0.04 -0.79 0.24 -0.22 0.11 -0.18 0.10 (Constant) 40.27 13.74 V1 V2 V3 V4 V5 V6 V7 V8 V9 V10 V11 V12 V13 V14 V15 V16 V17 V18 V19 V20 V21 Significance codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Table S4: Summary of the multiple linear regression model using a combination of clinical and neurophysiology data as independent variables and the ALSFRSr score as the dependent variable. ** ** . ** ** *** ** * .
1902.00337
1
1902
2019-02-01T14:03:28
Neural dynamics of emotion and cognition: from trajectories to underlying neural geometry
[ "q-bio.NC" ]
This paper describes the outlines of a research program for understanding the cognitive-emotional brain, with an emphasis on the issue of dynamics: How can we study, characterize, and understand the neural underpinnings of cognitive-emotional behaviors as inherently dynamic processes? The framework embraces many of the central themes developed by Steve Grossberg in his extensive body of work in the past 50 years. By embracing head on the leitmotifs of dynamics, decentralized computation, emergence, selection and competition, and autonomy, it is proposed that a science of the mind-brain can be developed that is built upon a solid foundation of understanding behavior while employing computational and mathematical tools in an integral manner. A key implication of the framework is that standard ways of thinking about causation are inadequate when unravelling the workings of a complex system such as the brain. Instead, it is proposed that researchers should focus on determining the dynamic multivariate structure of brain data. Accordingly, central problems become to characterize the dimensionality of neural trajectories, and the geometry of the underlying neural space. At a time when the development of neurotechniques has reached a fever pitch, neuroscience needs to redirect its focus and invest comparable energy in the conceptual and theoretical dimensions of its research endeavor. Otherwise we run the risk of being able to measure 'every atom' in the brain in a theoretical vacuum.
q-bio.NC
q-bio
Neural dynamics of emotion and cognition: from trajectories to underlying neural geometry Luiz Pessoa [email protected] Department of Psychology Department of Electrical and Computer Engineering Maryland Neuroimaging Center University of Maryland, College Park, USA Abstract This paper describes the outlines of a research program for understanding the cognitive- emotional brain, with an emphasis on the issue of dynamics: How can we study, characterize, and understand the neural underpinnings of cognitive-emotional behaviors as inherently dynamic processes? The framework embraces many of the central themes developed by Steve Grossberg in his extensive body of work in the past 50 years. By embracing head on the leitmotifs of dynamics, decentralized computation, emergence, selection and competition, and autonomy, it is proposed that a science of the mind-brain can be developed that is built upon a solid foundation of understanding behavior while employing computational and mathematical tools in an integral manner. A key implication of the framework is that standard ways of thinking about causation are inadequate when unravelling the workings of a complex system such as the brain. Instead, it is proposed that researchers should focus on determining the dynamic multivariate structure of brain data. Accordingly, central problems become to characterize the dimensionality of neural trajectories, and the geometry of the underlying neural space. At a time when the development of neurotechniques has reached a fever pitch, neuroscience needs to redirect its focus and invest comparable energy in the conceptual and theoretical dimensions of its research endeavor. Otherwise we run the risk of being able to measure "every atom" in the brain in a theoretical vacuum. 1 1. The problem of emotion, cognition, and behavior This paper describes the outlines of a research program for understanding the cognitive- emotional brain, with an emphasis on dynamics: How can we study, characterize, and understand the neural underpinnings of cognitive-emotional behaviors as inherently dynamic processes? At the outset, I propose eliminating the distinction between emotion and cognition (Pessoa, 2018a). What is emotion? What are its defining characteristics? Are emotions distinct from feelings? Researchers have debated, and in fact agonized over, such questions for a very long time. And the debate continues. For example, nine essays are dedicated to the topic in the latest edition of The Nature of Emotion (Fox, Lapate, Davidson, & Shackman, 2018); and additional suggestions continue appearing (see, Fox, 2018). Such pursuit of the "essence of emotion" appears misguided. What researchers of the mind and brain are interested in, it could be argued, is understanding behaviors. Mind scientists seek to understand the structure of behaviors, their inherent logic. Brain scientists strive to unravel how the two domains, mental and neural, map to one another during behaviors. The framework described here is strongly influenced by many lines of research and thinking (as any intellectual endeavor, of course), and most of all by the research by Steve Grossberg1. 2. Grossbergian themes Grossberg has developed his theoretical framework for over 50 years. The breadth of his thinking is so enormous as to defy understanding. In this section, I will describe a series of themes that permeate his work, sometimes very explicitly, at times less so. Although the remainder of the paper will build upon more directly on only a few of the themes -- and centrally on dynamics -- all of them are viewed as essential to building an understanding of the cognitive-emotional brain. 1 I was a graduate student at the Department of Cognitive and Neural Systems from 1990 to 1995, and worked closely with Steve during the last year of my PhD. This work was continued after I returned to Brazil until 1998. Steve also taught an enormously inspiring informal seminar during my second or third year in which he outlined his research program. His infinite energy and untiring guidance have been constant sources of inspiration in my career. 2 2.1 Dynamics This theme is so central to Grossberg's work that it is fair to say that without it the work would not exist. In Grossberg's very first publication2, he states: Fundamental to the motivation of the new theory is the realization that the dynamics of many psychological problems may be viewed from a unified point of view once the geometrical substrates that characterize each separate problem are elaborated and distinguished (Grossberg, 1964; italics added). The very first equation of his opus (Grossberg, 1964) reads as follows: 𝑑𝑠𝑘 𝑑𝑡 = 𝛼(𝑀 − 𝑠𝑘)𝑇𝑘 − 𝐷𝑘, where that the "activation" 𝑠𝑘 was defined via a grow process whereby 𝑠𝑘 increased toward 𝑀 at rate 𝛼 and its total input 𝑇𝑘 (itself dependent on other activations), while also subject to a simple exponential decay, 𝐷𝑘. At first, it would appear that one would hardly have to emphasize dynamics as an important principle. Yet, experimental brain research is frequently, and even preponderantly, quasi-static. Data from almost any measurement modality (physiology, functional MRI, etc.) are epoched in terms of trials or segments that largely discard most temporal information. 2.2 Behavior Consideration of a very extensive body of behavioral data is essential. Contrast this to a sort of "tunnel vision" that is unfortunately widespread, as all too often researchers break into cliques that focus on apparently distinct sets of phenomena. For example, research addressing "appetitive" and "aversive" processing has been carried out by largely separate communities. More generally, researchers focus on "motivation" or "emotion," on "cognition" or "emotion," and so on. But behaviors do not obey boundaries, and thinking about diverse sources of data is necessary for deeper understanding. 2 A monograph published while in graduate school with over 400 pages. 3 Behavior is the founding pillar for explaining the brain. This reads like a truism. Unfortunately, it is not, as suggested for example by the popularity of the recent paper by Krakauer and colleagues (2017) in the powerful journal Neuron (see also Gomez-Marin, 2016; Gomez-Marin et al., 2014). Their general call to arms to embrace behavior and eschew a neuronal reductionistic bias has resonated with those who believe that ever more sophisticated measurement techniques are not enough to dissect the brain. What if the activity of every neuron could recorded in the brain of an animal during a certain behavior (see Ahrens et al., 2012; Lovett-Barron et al., 2017). What would be gained by doing so? Consider a device that can measure the exact state of a modern Airbus 380 aircraft (which weighs more than a million pounds), say an image of every atom (aircraft are mostly made of aluminum) at millisecond resolution. An adequate level of description of the aircraft and its parts is in terms of fluid dynamics and related aerodynamics, where issues related to compressible flow, turbulence, and boundary layers are important. Therefore, although it is conceivable that this future device could provide some useful information, the point made here is that additional data are only minimally useful without more advanced theoretical understanding -- of both mind (that is behavior) and brain. Returning to the theme of dynamics, in parallel with the way data are analyzed, behavior is frequently conceptualized in terms of discrete trials of relatively short duration. This approach is understandable from the perspective of experimental scientists who need trial averaging to handle noise. But behavior itself is inherently temporal, and neglecting that aspect seriously limits research progress. 2.3 Decentralization, heterarchy Understanding systems in terms of the interactions between their parts fosters a way of thinking that favors decentralized organization. It is the coordination between the multiple parts that leads to the behaviors of interest, not a "controller" that dictates the function of the system. In many sophisticated systems, and the brain is no exception, it is natural to think that many of its chief functions depend on centralized processes. For example, the prefrontal cortex may be viewed as a uniquely positioned brain sector where multiple types of information converge, 4 allowing it to then guide behavior (Fuster, 2000; Miller and Cohen, 2001). A contrasting view is one in which processing takes place in a distributed fashion via the interactions of constituent parts. Accordingly, instead of information flowing hierarchically to an "apex region" where all the pieces are combined, information flows in multiple directions without a strict hierarchy. An organization of this sort is termed a heterarchy to emphasize the notion that the flow of information is multidirectional (McCulloch, 1945). A vivid illustration of the problem of centralization involves "executive control" processes. Early models of executive function were built around the notion of a "controller" -- essentially a homunculus -- that regulates lower-level systems when needed (e.g., Baddeley, 1996; see Shallice, 1988). The inherent problems with such an approach were eventually recognized by several investigators, who called for a "fractionation" of the executive in more manageable (that is, less intelligent) units (Monsell and Driver, 2000). Functions such as "shifting," "updating," and "inhibition" (Miyake et al., 2000) became more prevalent when describing the executive. However, time and again the use of such constructs has amounted to a way of redescribing the object of study rather than actual explanation. To this day, the goal of "banishing the homunculus" remains a formidable challenge (Verbruggen et al., 2014). The historical conceptualization of the hypothalamus provides another useful example (for an excellent discussion, see Morgane, 1979). This structure is generally referred to as the "head ganglion of the autonomic nervous system." This rubric encapsulates a hierarchical theoretical view based on the idea of "descending" control: the area functions as a central controller of structures along the extent of the brainstem. Indeed, the hypothalamus has robust projections to multiple brainstem sites. However, no area is simply an outflow region (and thus a "head"); all areas receive multiple inputs. In the case of the hypothalamus, multiple brainstem sites that receive projections from the hypothalamus project back to it, following the general tendency of connections to be bidirectional. More critically, the hypothalamus is extensively and bidirectionally connected with most sectors of the cortex (Nieuwenhuys, Voogd, and Huijzen, 2008; Pessoa, 2017a). Far from a master controller, the hypothalamus is an integral node of cortical-subcortical communication. 2.4 Emergence 5 This is a central concept in all of Grossberg's work, as captured by this recent statement: Brain circuits give rise to these distinct psychological functions as emergent properties that arise from interactions among brain regions that work together as functional systems. (Grossberg, 2018, p. 2; italics in the original). Von Bertalanffy (1950, p. 135), one of the chief early proponents of complex systems theory, famously asserted "the necessity of investigating not only parts but also relations of organization resulting from a dynamic interaction and manifesting themselves by the difference in behavior of parts in isolation and in the whole organism". But what does it mean to say "difference in behavior of parts in isolation and in the whole organism"? Hence emergence3, a term originally coined in the 1870s to describe instances in chemistry and physiology where new and unpredictable properties appear that are not clearly ascribable to the elements from which they arise. But what does emergence mean? At the most basic level it reflects the notion that "something new appears." While fascinating, this proposition sits uncomfortably with experimental scientists. As presciently stated by Von Bertalanffy (1950, p. 142) himself, the "exact scientist therefore is inclined to look at these conceptions with justified mistrust." Unfortunately, the picture has not appreciably changed, despite stunning developments in mathematics and physics in understanding nonlinear dynamical systems in the last 50 years. Today, emergence can be defined precisely, and in ways that leave no room for vague allusions to "wholeness" or "system properties." In the present context, the body of work by Grossberg provides a clear demonstration of how "emergent properties" can be precisely defined. 2.5 Selection and competition Selection of information for further analysis is a key problem that needs to be solved for effective behavior. Indeed: 3 The term emergence appears to have been first proposed in the 1870s when used by George Henry Lewes in his book Problems of Life and Mind and taken up by Wilhelm Wundt in his Introduction to Psychology. 6 How can a limited-capacity information processing system that receives a constant stream of diverse inputs be designed to selectively process those inputs that are most significant to the objectives of the system? (Grossberg and Levine, 1987, p. 5015) If selection is the ubiquitous problem that must be effectively solved, competition is the mechanism by which stimuli, objects, actions, and so forth are selected. 2.6 Autonomy Central to Grossberg's theoretical framework is the notion of autonomy: Brains look the way that they do because they embody computational designs whereby individuals autonomously adapt to changing environments in real time. Grossberg (2018, p. 4; italics in the original). To understand the cognitive-emotional brain, it is necessary to consider that all animals need to function independently in diverse and challenging conditions and environments -- they need to be autonomous. All vertebrates have a brain architecture that allows a considerable amount of communication and integration of signals (Pessoa, 2018b; Pessoa et al., in preparation). Why this kind of architecture? One possibility is that it confers a high degree of flexibility that allows animals to cope with the complex interactions in their changing habitats, involving predators, prey, potential mates, and so on. Survival may benefit from circuits that can form in a combinatorial fashion, as the number of conditions related to the internal and externals worlds of the animal are exceedingly high. Consider a key system for both appetitive and defensive behaviors, the superior colliculus in the midbrain (Dean et al., 1989; Peek & Card, 2016; Pereira & Moita, 2016). It receives retinal inputs and has outputs that give it access to movements of head and neck, for example. In rodents, the superior colliculus could be involved in implementing the following rule: If unexpected movement is overhead, flee; otherwise, if movement is in the lower field, consider further exploration. However, simple rules based on stimulus features do not capture the flexibility of rodent behavior (think how hard it is to catch a rat! see Dean et al., 1989). In particular, rats freeze more frequently to novel stimuli in unfamiliar environments, such as an 7 open field. Clearly, the context in which a stimulus occurs is essential (Peek & Card, 2016; Pereira & Moita, 2016). More generally, one way to view the more elaborate architecture of birds and mammals (Striedter, 2005) is in terms of the enhanced potential for combinatorial interactions that they afford, such that the manner different signals can influence each other is considerably expanded -- and accordingly expand the range of behaviors. This overall type of architecture may produce circuits with local specificity but relatively large-scale sensitivity, a type of global-within-local design, which likely contributes to more plastic and sophisticated behaviors. Yet, integration is evolutionarily ancient -- it is a hallmark of the vertebrate brain -- and could explain the existence of complex behaviors now recognized in all vertebrate taxa. 2.7 Computational theory Brain research is a strongly empirical scientific enterprise. To be sure, research is inspired and guided by conceptual/theoretical thinking, although mostly in a qualitative fashion. But as Rabinovich and colleagues (2006, p. 48) state: "Neural networks [both natural and artificial] are complicated dynamical entities, whose properties are understood only in the simplest cases." Can the complex architecture that supports the cognitive-emotional brain be investigated without formal/mathematical tools? Given the richness of the multi-level interactions, does neuroscience need to migrate to a model that is closer to that of physics? Experimental physicists are not lacking in mathematical sophistication. Neuroscience, in contrast, has evolved into extremely sophisticated "laboratory techniques" that are often divorced from formal approaches. How should we train future generations of brain scientists? Grossberg's position on these questions is easy to predict, as he and colleagues created the Department of Cognitive and Neural Systems at Boston University in 1989 exactly to address this issue. 3. Decentralized computing: Top-down control versus circuit interactions The central claim is that, for interesting behaviors, most of the required explaining is not present at the level of isolated systems (perception, action, etc.) but at the level of the interactions between them (Pessoa, 2018a,b). Strictly speaking, however, the concept of an 8 interaction is not opposed to the notion of separable entities, as interactions can also refer to distinct variables, processes, or systems that, themselves, produce effects in a non-additive manner. Accordingly, a better term is integration, which implies sufficient intertwining between the putatively separate systems that their individuation becomes a linguistic short-cut. Let's consider the mechanisms of fear extinction (Figure 1A). When a conditioned stimulus (CS) no longer predicts the unconditioned stimulus (UCS) to which it was paired at some point in the past, a new relationship needs to be learned, namely the CS is no longer associated with the UCS -- this type of learning is called "extinction." The medial prefrontal cortex (PFC) plays an important role during extinction, as initially revealed via lesioning (Morgan et al., 1993) and subsequently by chemical manipulation of this area (for review, see Dunsmoor et al., 2015). As the medial PFC is extensively interconnected with the amygdala, an early idea was that the former would exert an inhibitory influence on the latter, thereby enabling the extinction of the conditioned response. At this level of description, fear extinction fits the scheme of separate entities interacting to generate a new behavior: cognition (tied to the medial PFC) controlling emotion (tied to the amygdala) in a top-down fashion. Yet, considering the PFC as "top" and the amygdala as "down" does not take into account the richness of the existing neuronal interactions. It is well known that the amygdala plays a critical role in aversive learning, that is, the initial CS-UCS learning. The amygdala plays a critical role in the acquisition and consolidation of fear extinction, too. Chemical blockage of amygdala mechanisms (in the basolateral amygdala) either impair or entirely Figure 1. Fear extinction and structure-function mapping. (A) Fear extinction. (B) Conceptualization of fear extinction in terms of the top-down regulation of the amygdala by the medial prefrontal cortex, with additional variables influencing the process. (C) Schematic representation of the connections between some of the brain regions involved, emphasizing a non-hierarchical view of the processes leading to fear extinction. The descriptors "valence," "regulation," and so on, are not tied to brain areas in any straightforward one-to-one fashion. Abbreviations: CS, conditioned stimulus; MPFC, medial prefrontal cortex; OFC, orbitofrontal cortex. prevent the acquisition of extinction (Herry et al., 2006). In addition, consolidation of extinction is supported by morphological changes in amygdala synapses (in the basolateral amygdala; see Tovote et al., 2015). These findings, together with the existence of amygdala pathways to the medial PFC, has led some investigators to suggest that the amygdala actually 9 should be viewed as the "top" region in the relationship with the medial PFC (Herry et al., 2008; see also Do-Monte et al., 2015). In fact, multiple cell groups in the amygdala project to the medial PFC, whose outputs in turn influence amygdala signals. The extinction of conditioned responses is one of the oldest and most widely known findings from psychological science (Dunsmoor et al., 2015). Despite this long history, recent research has greatly expanded our knowledge about this phenomenon, revealing that extinction, far from a simple inhibitory process, is an extremely nuanced learning process. Extinction is now understood to be a form of learning (of the new relationship between the CS and UCS) that itself involves acquisition, retrieval, and consolidation. In other words, it is not a simple inhibitory mechanism of the "fear response" but a sophisticated form of learning. In particular, following extinction, contextual information plays a critical role in determining whether the original fear memory or the new "extinction memory" controls behavior -- should the animal fear or not the CS? Accordingly, an elaborate set of neural interactions is needed to support such context sensitivity. When the CS no longer predicts an aversive outcome, it behooves the animal to take into account that information, such that features of the new environment are learned so as to predict safety. The hippocampus plays a key role in establishing context dependence during extinction learning. There are at least two anatomical routes by which the hippocampus contributes to these processes (Herry et al., 2008; Maren et al., 2013). The first involves direct projections from the hippocampus to the amygdala; the hippocampus is part of a circuit that involves amygdala neurons that are engaged when the behavioral context is different from the extinction context (this pathway thus promotes fear). The second, indirect contribution involves dense projections to the medial prefrontal cortex, which appearss to participate in a circuit with the amygdala that indicates safety (this pathway is linked to extinction behaviors). Another region in the circuit determining if fear should be switched on or off is the thalamus, which is a major player in the processing of biologically significant stimuli (Heimer et al., 2017), as well as a key subcortical -- cortical connectivity hub (Pessoa, 2017b). In the past few years, the paraventricular nucleus of the thalamus (PVT) has been established as a thalamic node that interacts with cortico-amygdala circuits for the establishment, retrieval, and maintenance of long-term fear memories (Do-Monte et al., 2015; Penzo et al., 2015). Neurons in the PVT are robustly activated by behaviorally relevant events, including novel 10 ("unfamiliar") stimuli, as well as reinforcing stimuli and their predicting cues (Ren et al., 2018). Notably, PVT responses are influenced by changes in homeostatic state and behavioral context, and inhibition of the PVT suppresses appetitive and aversive learning (Ren et al., 2018). Given that the PVT is bidirectionally connected with the medial PFC, and projects throughout the extended amygdala (central amygdala plus bed nucleus of the stria terminalis), this region is well placed to further refine processing during behavioral conditions eliciting fear extinction. More generally, during fear extinction -- and, in fact, fear acquisition and expression -- signals from the amygdala, medial PFC, hippocampus, thalamus, among others, collectively determine behavioral responses. These multi-region interactions afford greater behavioral malleability when responding to threat. A more standard approach to attempting to explain fear extinction would be to label each brain region in the following manner, for example: amygdala-valence, medial PFC-regulation, hippocampus-context, thalamus-biological significance, and so on. One could then describe observed behaviors in terms of "standard interactions" (that is, those involving separate entities) between the putative processes (valence, regulation, etc.) (Figure 1B). But if these processes are not separable, they do not encode stable variables that are simply modulated by other variables. In the end, explanations in terms of standard interactions will be found wanting -- integration is needed (Figure 1C). 4. Causation in complex systems A potentially unappealing aspect of the discussion and conclusion above is that causation is muddied -- what causes what during extinction? Dissecting phenomena in terms of their component parts seems like an unimpeachable methodology, to the extent that it can be viewed as almost an axiom of modern science (Deacon, 2011). At the broadest level, the issues at hand speak to how we should study systems as complex as minds and brains. Understanding causation has always been at the core of the scientific enterprise. To a great extent, the mission of neuroscience is to uncover the nature of the signals observed in different parts of the brain, and to attempt to disentangle the potential contributions to those signals. Consider a type of reasoning prevalent in neuroscience, what can be called the billiard ball model of causation. In this Newtonian model, force applied to a ball leads to its movement on the table until it hits the target ball (Pessoa, 2017c, 2018c). The reason the target ball moves 11 is obvious; the first ball hits it, and via the force applied to it, it moves. Translated into neural jargon, we can rephrase as follows: a signal external to a brain region excites neurons in that region, which excite or inhibit neurons in another brain region given anatomical pathways connecting them. But this general mode of thinking, which has been very productive in the history of science, is too impoverished when complex systems -- the brain for one -- are considered. Mannino and Bressler (2015) highlight two features of the brain that are problematic for standard (Newtonian) causation. First, anatomical connections are frequently bidirectional, leading to bidirectional physiological influences. But if one element causally influences another while the second simultaneously causally influences the first, sometimes called mutual causality (Frankel, 1986), the concept breaks down. Second, convergence of anatomical projections implies that multiple regions concurrently influence a single receiving node, making the attribution of unitary causal influences problematic. Along these lines, in a recent paper I outlined several principles of brain organization that impact the understanding of causality (Pessoa 2017b; see also Pessoa, 2014): 1) massive combinatorial anatomical connectivity; 2) extensive cortical-subcortical anatomical connectional systems; 3) high distributed functional connectivity4; 4) overlapping large-scale functional brain networks; and 5) dynamic large-scale functional brain networks. Taken together, the brain basis of emotion-cognition involves distributed, large-scale cortical- subcortical networks. The high degree of signal distribution and integration provides a nexus for the intermixing of information related to perception, cognition, emotion, motivation, and action. Importantly, the functional architecture consists of multiple overlapping networks that are highly dynamic and context-sensitive, such that how a given brain region affiliates with a specific network shifts as a function of task demands and brain state (Najafi et al., 2017. In all, particular cognitive-emotional behaviors can be understood in terms of dynamic functionally integrated systems, such as the one involving the amygdala and its cortical-subcortical circuits (Pessoa, 2017b). 4 Functional connectivity refers to the degree of association between time series data of two brain regions, irrespective of their anatomical connectivity status (whether directly connected or not). It is typically estimated based on Pearson's correlation coefficient, although multiple measures of association have been proposed (for example, mutual information). 12 The upshot is that Newtonian causality provides an extremely poor candidate for explanation in non-isolable systems like the brain. The shift advocated here is to move away from individual entities (like billiard balls) and consider the temporal evolution of "multi- particle systems." An example is provided by considering another analogy from physics: the motion of celestial bodies in a gravitational field. Physicists and mathematicians have been interested in the problem of stability for centuries, and this was a central problem in Newtonian (that is, non- relativistic) celestial mechanics. For example, what types of trajectories do two bodies, such as the earth and the sun, exhibit? The so-called two-body problem was completely solved by Johann Bernoulli in 1734. For more than two bodies (for example, the moon, the earth, and the sun), the problem has vexed mathematicians for centuries. Although the three-body problem cannot be solved in the same sense as the two- body problem, topological properties can be used to classify families of three-body periodic orbits. Figure 2 displays the "yin-yan II" class first reported by Figure 2. The three-body problem in Newtonian gravitation. As the problem does not admit to a general mathematical solution, researchers have sought to characterize families of periodic orbits. The figure displays the "yin-yan II" family in two dimensions. The three circles represent the three bodies. Reproduced from Li and Liao (2017). Dynamic plots can be found at http://numericaltank.sjtu.edu.cn/three-body/three-body.htm. Suvakov and Dmitrasinovic (2013) in a breakthrough study describing a large number of new orbit families. This gravitational three-body problem illustrates the approach to studying complex systems described here: determining and characterizing the temporal evolution of multi- particle systems. The idea will be further developed below after a brief comment on the use of advanced neurotechniques to study causation in neural systems. 4.1 Optogenetic causation In neuroscience, causal efficacy -- that is, causal intervention -- has become the field's gold standard, and is implicitly or explicitly equated with understanding. Technologies such as optogenetics exist that make it possible to manipulate neural circuits directly and more 13 precisely. However, causal-mechanistic explanations, while valuable, are qualitatively different from understanding how circuit elements combine to produce behavior (for recent discussions, see Krakauer et al., 2017; Fregnac, 2017). Consider the following admittedly crude example. Suppose an alien species is studying how our automobiles work by using an advanced form of technology. Unbeknownst to us, they are able to measure and manipulate our cars while we drive. Suppose they are able to pull the throttle wire linked to the gas pedal, thereby accelerating a car. By doing so enough times, they establish a causal link between the throttle wire and the car's speed (and publish their results in one of their top journals). Obviously, although they can now deduce that the gas pedal plays an important role in the car's movement, their understanding of automobile function is increased only minimally. For one, they have no idea that the system controls air inflow and therefore controls fuel injection into the engine5; not to mention the principles of the combustion engine. Now, consider the following study, where optogenetic stimulation was used to activate neurons in the ventral tegmental area (VTA) of the midbrain of mice (Pascoli et al., 2018). As in the classic self-stimulation study by Olds and Milner (1954), mice learned to press a lever, in this case for optogenetic-driven enhancement of VTA activity. After two weeks of training, upon lever pressing, mice now received a brief electric shock in addition to increased VTA activation. Approximately 60% of the mice persisted in lever pressing impulsively, despite being administered the shock. In addition, the authors found that a pathway from the orbitofrontal cortex to the dorsal striatum (in the forebrain) affected the behavior of the mice; for example, optogenetic inhibition of the pathway made them stop lever pressing. Compulsive behavior (lever pressing) could be suppressed or induced by decreasing or increasing, respectively, the strength of this neural connection. Remarkable as these results may be (they were published in Nature as a full-length article), how much closer are we to understanding "compulsive behaviors?" For one, the VTA neurons stimulated by lever pressing do not connect directly with either the orbitofrontal cortex or the dorsal striatum. Clearly, a multisynaptic circuit involving the regions studied must be involved (Keiflin and Janak, 2015), and the mechanisms of action remain unknown. The discovery of the simple causal link between the orbitofrontal cortex and the dorsal striatum via sophisticated techniques, intriguing as it is, leads to an ostensible shift of the goalpost but leaves us minimally closer to 5 It is assumed that an older car is being studied, and not a newer drive-by-wire model. 14 understanding the circuit. The objective here was not to critique this particular study, but instead to highlight shortcomings of the causal-mechanistic, interventionist approach currently widespread in neuroscience (see also Krakauer et al., 2017). 5. Transient brain dynamics The upshot is that simple ways of reasoning about causation are inadequate when unravelling the workings of a complex system such as the brain. Instead of focusing on causation as the inherent goal of explanations in neuroscience, a fruitful research avenue is to develop formal tools that describe the dynamic multivariate structure of brain data. In other words, one is interested in describing the joint state of a set of brain regions, and how this joint state evolves temporally. A major goal is then to work out how groups of regions dynamically coalesce into coherent functional units and how they dissolve when their assembly is no longer needed to meet processing demands. Consider a system of neurons, neuronal populations, or brain regions, which is characterized by their activation strengths as a function of time: 𝑥1(𝑡), 𝑥2(𝑡), ⋯ , 𝑥𝑛(𝑡). The vector 𝒙 describes the current joint state of the system (when evaluated at a given time, t), and 𝒙(𝑡) describes how this joint state evolves through time. A popular approach to thinking about brain dynamics was based on the notion of steady-state attractors, in which activity levels would converge to equilibrium (for at least some period of time). For example, when started at state 𝒙𝐼, the system would evolve temporally and settle in state 𝒙𝐴, where 𝒙𝐴 is the stable state closest to 𝒙𝐼 (Cohen and Grossberg, 1983; see also Hopfield, 1982, 1984). In such networks, an input pattern will cause activity changes until it settles into one pattern, the so-called attractor state (here, 𝒙𝐴). We can thus say that the input is associated with the properties of the entire, and specific, attractor state, which can be viewed as its representation. However, the type of dynamics in "attractor networks" is limited in the sense that the key element is the state into which the network settles (which can be represented formally by, for example, a minimum in an energy function). Importantly, the path taken to reach the attractor state does not matter. The idea of "computing with attractors" should be contrasted with the one of computing with transient dynamics (Rabinovich et al., 2008; Buonomano and Maas, 2009). Transient dynamics do not require waiting for the system to reach equilibrium, and the succession of 15 states visited by the system provide the representation for the event in question. The temporal window considered is arbitrary; for example, 300 ms after an input stimulus, 500 ms prior to movement initiation, or 20 seconds during a mental event. Figure 3A illustrates the idea in the context of recordings from neurons in the antennal lobe of the locust (Broome et al., 2006), showing the succession of states associated with the presentation of two distinct odors when projected onto a lower-dimensional three-dimensional space (those less familiar with this type of plot may benefit from Figure 4A-B). The original measurements were performed in 87 neurons, and the projection here is simply for illustrative purposes (we will discuss the issue of dimensionality further below). Whereas the trajectories might come arbitrarily close at several time points6, the entire trajectory provides a potentially unique signature for the task in question, such that the transients are input specific, and contain information about what initiated them. Furthermore, the trajectories are assumed to be stable. Thus, they are resistant to noise in that they are reliable to relatively small variations in initial conditions. Figure 3. Neural trajectories. Trajectories represent the activation state of the system at every point in time. (A) Recordings were performed in 87 principal neurons (PNs) of the antennal lobe of the locust during exposure to two odors (citral: cit; geraniol: ger) (Broome et al., 2006). Local linear embedding (LLE) was employed to reduce the dimensionality of the data. (B) Recordings were performed in premotor and motor cortex during reaching movements in the macaque monkey (Churchland et al., 2012). A principal components analysis- based algorithm was used to determine the two-dimensional representation displayed. Individual trials are represented by trajectories colored based on the extent of preparatory/pre- movement activity (from red to green). Thinking in terms of trajectories moves the emphasis away from a strictly causal interpretation. Instead of, for example, statements such as "𝑥1(𝑡) causes 𝑥2(𝑡 + 1)," the framework encourages a description that summarizes the temporal evolution of the system of interest. Experimentally, a central goal then becomes estimating trajectories robustly from available data. At this point, computational models can be tested against the data, or possibly developed to explain the data. In other words, what kind of system, and what kind of interactions between system elements -- what mechanisms -- generate similar trajectories, given similar inputs and conditions? 6 The issue of the proximity of trajectories will depend on the dimensionality of the system in question (which is usually unknown) and the dimensionality of the space where data are being considered (say, after dimensionality reduction). Naturally, points projected onto a lower- dimensional representation might be closer than in the original higher-dimensional space. 16 5.1 Trajectories during threat processing Consider a functional MRI participant experiencing alternating "safe" and "threat" blocks, where they lie passively in the former, while they may experience mild shocks during the latter. At the onset of threat blocks, brain regions of the so-called salience network (including the anterior insula and medial PFC) would be expected to respond vigorously compared to the period prior to the block transition (Menon and Uddin, 2010). Regions of the salience network would also respond to the onset of the "safe" block, but suppose that these responses are less vigorous Figure 4. Neural trajectories during threat processing. (A) Evoked responses when transitioning from safe (green) to threat (red), and vice versa. 𝑥1and 𝑥2 represent the activity of two brain regions. (B) The responses can be jointly plotted as a function of time, (𝑥1(𝑡), 𝑥2(𝑡)), to show safe and threat trajectories. The two dots in panel A correspond to the ones here (schematically only). (C) Trajectories for safe and threat conditions when evokes responses are comparable but they are more correlated during threat. (D) Corresponding to panel D, we can think of the trajectories for threat and safe as evolving through cylinders of different diameters (which correspond to the trajectory variance). (E) Individual differences in anxiety can be understood as shifting the trajectories from safe to threat. than during the transition to threat. In terms of evoked responses, this scenario can be illustrated as in Figure 4A. For two hypothetical regions, if we diagram the temporal evolution of the responses during the block transitions, the trajectories for the two conditions can be illustrated as in Figure 4B; the state-space plot describes the activity levels (𝑥1(𝑡), 𝑥2(𝑡)). Now, suppose that a group of high-anxious individuals is investigated and that they partially generalize the aversiveness experienced during threat blocks to safe ones; that is, they treat 17 every block onset as a potential transition into a threat condition.7 In this case, the trajectory observed during safe periods would look more like threat trajectories, and the more so for individuals with higher levels of anxiety (Figure 4E). In an actual functional MRI study that we performed (McMenamin et al., 2014), we found that, somewhat surprisingly, responses evoked when transitioning into safe and threat blocks were rather comparable in magnitude. Presumably, both safe and threat blocks were motivationally significant, thus evoking similar salience-related responses; perhaps, in the context of encountering threat periods, safe periods are quite noteworthy. However, although evoked responses were comparable, signals were more cohesive during threat relative to safe; that is, transitions to threat blocks were associated with evoked responses that were more correlated (for a given pair of regions in the salience network). The respective trajectories for our experiment thus can be illustrated as in Figure 4C (the trajectory linked to threat stays closer to the diagonal (𝑥1 = 𝑥2) than the one linked to safe). And if we consider multiple trials, the trajectories during threat will remain in a part of the space closer to the diagonal (Figure 4D). 6. Dimensionality reduction of neural measurements Neuronal data are inherently high dimensional. Consider, for example, the simultaneous recordings across 10-102 locations in electrophysiological grids, 102 sensors with MEG/EEG, 102-103 neurons with calcium imaging, or the 104-105 spatial locations with functional MRI. Is it possible that the information across, say, hundreds of measurements could be captured in fewer dimensions without substantial loss of information? Of course, techniques such as principal components analysis are commonplace in data analysis (and can be used, for example, for noise reduction). However, aside from practical concerns, understanding the dimensionality of the data is also important conceptually. For example, it may help uncover relationships that are not apparent in higher dimensions, thus helping to elucidate the mapping from structure to function. In particular, a parsimonious description of the data may uncover stronger relationships with experimentally manipulated variables or other behaviorally relevant variables (see also Santhanam et al., 2009). In addition, the number of dimensions of a 7 High-anxious individuals generalize conditions associated with conditioned fear, for example (see Lissek et al., 2008). 18 dynamical system is an enormously important topic in mathematics (Packard et al., 1980; Takens, 1981; Sauer et al., 1991). One of the most studied systems in terms of temporal trajectories involves odor processing in invertebrates. In the locust, odors generate distributed responses across the antennal lobe, and such responses evolve in an odor-specific manner (Broome et al., 2006). In Figure 3A, the lower-dimensional representation was obtained by nonlinear dimensionality reduction (Roweis and Saul, 2000). While the dimensionality reduction technique applied was somewhat arbitrary, it helped the investigators gain insight into the following theoretical question: what happens when one odor is being experienced and a second one is presented? One possibility is that the system would "reset," namely responses would return to baseline, then start to evolve in the direction of the new odor (Broome et al., 2006). An alternative possibility would be for the first trajectory (the one associated with the first odor) to deviate from its ongoing evolution and progress along a path corresponding to the mixture of the two odors. Based on the trajectories observed under these experimental conditions, Broome and colleagues were able to rule out the first possibility, while obtaining some support for the second. Together, dimensionality reduction helped uncover mechanisms that would be potentially hard to derive in higher dimensions. Neuronal dynamics has been investigated in nonhuman primates, too. In one study, Churchland and colleagues (2012) recorded responses in motor and premotor cortex as monkeys performed reaching movements. Data from 50-200 recordings were projected onto two dimensions, revealing a rotational structure to neural trajectories (Figure 3B). Their analysis uncovered processes at the level of the population of neurons, according to which preparatory activity (that is, prior to movement initiation) sets the initial state of a dynamical process that unfolds during movement execution. More generally, the authors proposed that motor cortex expresses a dynamical system that generates and controls movements, and that can be expressed as 𝑑𝒓 𝑑𝑡 = 𝑓(𝒓(𝑡)) + 𝒖(𝑡) where r is a vector describing the firing rate of all neurons (the population response or neural state), 𝑓 is an unknown function, and u is an external input. As in the example of the locust 19 data, dimensionality reduction helped unearth processes that would not have been evident in higher dimensions. 7 Geometry of the underlying neural space If a neural dataset is acquired in a high-dimensional space and subsequently reduced to a lower dimensionality, what should be the geometry of this space? For simplicity, the original high- dimensional space is frequently, if implicitly, considered Euclidean. But given that not all information can be preserved in fewer dimensions, the question of the nature of the lower dimensionality comes to the fore. For example, in the case of the locust data, a local linear embedding algorithm was employed that attempts to capture information about global geometry in fewer dimensions (by collectively analyzing overlapping local neighborhoods; Roweis and Saul, 2000). In the case of the monkey data, a PCA-based method was applied. We could follow a similar approach with the functional MRI data of safe and threat periods discussed in Section 5.1 (McMenamin et al., 2014). The study considered 51 brain regions of the so-called salience, executive, and task-negative (also called "default") networks, in addition to the amygdala and the bed nucleus of the stria terminalis (the latter two are particularly important during threat-related processing). We performed dimensionality reduction with the local linear embedding algorithm (Roweis and Saul, 2000) and plotted the mean trajectories for the two conditions, together with an indication of their variance (Figure 5). The two trajectories initially overlap but are quite distinct overall. Figure 5. Temporal trajectories based on functional MRI data. The original data were from safe and threat periods in the study by McMenamin et al. (2014). Trajectories for safe and threat conditions are fairly distinct in the lower- dimensional space determined by local linear embedding. (A) Mean trajectories across individuals. The colored circles indicate the starting point. (B) Surfaces provide an indication of the underlying space of trajectories, or manifold, and were created by considering the variance of the trajectories across individuals. 20 More generally, the geometry of the underlying neural space will depend on a combination of the properties of the data and the task condition of interest. We propose the following neural-dynamics space hypothesis: behaviors can be described via (they are associated with) classes of trajectories within specific neural spaces (see also Gao et al., 2017). Consider the example of the citral-related trajectory in the locust antennal lobe (Figure 3A). Multiple instances of experiencing this odor are proposed to reside within the surface schematically represented in Figure 6A. This surface defines the space within which trajectories linked with this odor naturally evolve. Such surfaces, which are mathematically called manifolds, thus serve as representations of the stimuli, tasks, or conditions in question. The examples so far assumed the use of an explicit method of dimensionality reduction (the simplest of which is perhaps PCA), a data-driven approach that is suitable in many circumstances. However, knowledge of the problem domain can guide this process, too. In fact, in the case of the Churchland et al. (2012) study, the PCA- based method the authors used was developed to extract rotational information because the authors believed that such coordinate system would be relevant to understanding the topology of neural trajectories in motor cortex during reaching movements8. To illustrate the use of domain knowledge, consider a hypothetical study that records multiple Figure 6. Trajectory manifolds. A manifold is a surface (more precisely a topological space) that near each point (that is, locally) resembles Euclidean space (circles and spheres are some of the simplest manifolds in two- and three dimensions). A non-Euclidean (Riemannian) metric on a manifold allows distances and angles to be measured. (A) Example manifold. The neural-dynamics space working hypothesis suggests that system behaviors can be characterized via classes of trajectories within neural spaces with particular geometry -- that is, manifolds. (B, C) Neural manifolds associated with transient dynamics are, by definition, non-periodic. Here, "population" indicates that the manifold is a group-level property. Two inter- related problems can be posed. One is estimating the group-level trajectory (red) from sample data (trajectories in orange); the other is learning the population manifold (green) from sample-level trajectories (orange). cells in each of three distinct brain areas, and suppose that different properties are thought to 8 The rotational structure was not due to primary features of neuronal responses, such as tuning to reach direction (Elsayed and Cunningham, 2017). 21 be important for their function. In this case, one can plot the dynamics of the system in terms of these properties (Figure 7). In our study of safe and threat periods discussed above (McMenamin et al., 2014), we found that a network-level graph-theory measure called global efficiency captured a relevant facet of threat processing. Briefly, efficiency provides a measures of how effectively a network exchanges information (Latora and Marchiori, 2001). In particular, small-world networks are systems that are both locally and globally efficient. Another graph-theory property of interest in our study was node Figure 7. Geometry of the neural space. The activity of brain areas 1-3 can be mapped onto distinct properties believed to reflect their function, including univariate properties and multivariate/network-level properties. Here, a hypothetical trajectory is illustrated during a dynamic threat scenario in which threat level gradually increases (red) and then decreases (green). centrality. Increased centrality indicates that a node participates more heavily in the interactions between other nodes -- that is, they become more of a hub. And, as discussed previously, motivationally significant events, such as blocks transitions, produced stronger responses in regions of the salience network. Accordingly, it could prove informative to project the evolution of the system along these three axes. Let's apply this idea in the case of a different experimental paradigm. Consider a scenario in which threat is manipulated dynamically. For example, two circles move on the screen in a quasi-random manner and, if they collide, a mild electrical shock is administered to the participant (Myer et al., 2019). Thus, there will be periods of increased anxious anticipation (circles approaching each other) and periods of relative safety (circles retreating from each other). Figure 8 illustrates hypothetical trajectories during approach and retreat in terms of the three dimensions discussed. The overall framework is also fruitful to describe trait- or temperament-like phenotypes. For example, if Figure 8 portrays the situation for a group of low-anxious individuals, for high-anxious individuals one could hypothesize that (i) periods of approach would be associated with higher activation of salience-network regions, (ii) higher network efficiency, and (iii) increased centrality of regions such as the bed nucleus of the stria 22 terminalis and amygdala. Importantly, these properties evolve temporally, as observed experimentally (McMenamin et al., 2014; Najafi et al., 2017; see also Pessoa and McMenamin, 2017). Dispositional negativity refers to a fundamental dimension of childhood temperament and adult personality and constitutes a prominent risk factor for the development of pediatric and adult anxiety disorders (Hur et al., 2018). Behaviorally, dispositional negativity is associated with threat-related attentional bias and deficits in executive control. Key brain systems proposed to underpin dispositional negativity include the amygdala, as well as the frontoparietal and cingulo- opercular networks (Hur et al., 2018). One could further test these ideas by investigating experimental conditions involving the performance of cognitively demanding tasks during the presence of threat. In particular, imagine Figure 8. Trajectories during threat processing. Threat-level varies dynamically and increases (approach) and decreases (retreat). The temporal evolution of the system can be described in terms of the global efficiency of the salience network, the activation (evoked responses) in the same network, as well as the centrality of the amygdala/bed nucleus of the stria terminalis regions. that during the execution of an executive task the threat level is increased from low to high. In terms of neural trajectories, one could hypothesize that there would be a shift in the state-space region occupied by the conditions at hand (Figure 9). In addition, for individuals with higher dispositional negativity two predictions could be made: (i) the transition from one region to another would take place faster; and (ii) the extent of the change would be greater (that is, the two regions would be farther apart). Irrespective of the potential of these particular predictions to advance the understanding of dispositional negativity, they illustrate how hypotheses can be formulated and tested according to the present ideas. Finally, it also encourages a move away from amygdala-centric proposals that dominate the literature. 8 Causation in complex systems, again 23 In many systems, the relationships between entity-level variables cannot be studied independently of the overall system state. Deyle and Sugihara (2011; see references therein) proposed that such emergence-level view may help explain why many natural systems are so difficult to understand and predict. Building on the mathematical work by Takens (1981) and others, Sugihara et al. (2012) describe a powerful conceptualization that can be called dynamical systems causation: two variables Figure 9. Dispositional negativity. Participants perform a cognitive challenging task for an extended period of time. In one condition, there is a lower level of background threat, whereas a higher level is present in the second; the latter is anticipated to impair performance to a greater extent. Hypothetical neural trajectories are shown for the two conditions: the two trajectories will reside in separate sectors of state space, with the separation between them depending on an individual's level of dispositional negativity. are causally linked if they participate in the same dynamical system. In other words, the two variables share a common attractor manifold, such that each variable can identify the state of the other (for example, in the Lorenz attractor). Importantly, this notion can be formally and quantitatively developed (see their convergent cross mapping method). What is more, the proposal is general enough to encompass more traditional views, while also capturing the relationship between variables in many complex systems. The notion of dynamical systems causation is potentially powerful, but it relies on asymptotic behaviors of the systems in question (such as the attractor manifold). We have emphasized, instead, thinking in terms of transient dynamics, and neural events far from equilibrium and long-term properties. Whereas, conceptually, this does not present a significant impediment, in practice reliably estimating interdependencies may be data-limited. Indeed, these issues have been noted and related methods proposed to ameliorate the problem (Ye and Sugihara, 2016). Furthermore, the problem may be more tractable with group studies in which data from multiple participants is used to recover the underlying manifold. More generally, in group studies, the problem may be posed in terms of estimating a population manifold from individual-level trajectories (Figure 6B-C). Nevertheless, instead of adopting a single definition/measure of causation, at the current stage of scientific development, it would be beneficial to encourage a plurality of 24 conceptualizations, not least because elucidating complex systems will benefit from multiple vantage points. For example, Mannino and Bressler (2015) propose the notion of probabilistic causation: an event does not necessarily determine another event, but rather changes its probability of occurrence. As they state: "A causes B" may be defined as the probability of B given A is greater than the probability of B given that A does not occur: P(BA) > P(B~A). Their framework is enmeshed with establishing the "ultimate nature" of the brain, namely is it a deterministic or a probabilistic system? Irrespective of this more controversial question, their proposal offers an important way to move beyond outdated models of causes. It is worth pointing out that the popular framework of sydying causality introduced by Granger (1969) comes with serious limitations (as recognized by Granger himself). In today's terminology, variable x "Granger causes" y if the predictability of y declines when x is removed from the universe of all possible causative variables. However, a key requirement of the model is that of separability, such as observed in linear systems. Thus, information about a causative factor needs to be unique to that variable. In coupled systems like the brain, such assumption is clearly violated. Overall, the suggestion to embrace multiple conceptualizations of causality reflects the idea that the problem is dauntingly challenging. In this regard, it is just the opposite of what was asserted recently by Mehler and Kording (2018): "causality has a perfectly clean definition." Finally, no treatment of causality is probably comprehensive without considering the work of Pearl (2009). 9. Learning dynamics with reservoir computing Temporal trajectories potentially provide signatures for tasks, conditions, or states. Dimensionality reduction provides a strategy to potentially identify trajectories in lower- dimensional spaces. But can trajectories be learned from neural data? In this section, we describe an approach to learning temporal information that we recently developed in the context of functional MRI data (Venkatesh et al., 2019). Naturally, when considering the temporal information present in functional MRI data, it is necessary to consider the slow evolution of blood oxygenation responses. Accordingly, dynamics should be understood at a commensurate temporal scale -- on the order of a few 25 seconds or typically longer. Fortunately, many mental processes unfold at such time scales, such as the processing of event boundaries (Zacks et al., 2001), a gradually approaching threatening stimulus (Najafi et al., 2017), listening to a narrative (Ferstl et al., 2005), or watching a movie (Hasson et al., 2004). Figure 10. Reservoir computing and dimensionality reduction. (A) Brain data are provided to a three-layer neural network. The input layer registers activation at time t across a set of regions of interest. The reservoir layer contains units with random connections, and provides a memory mechanism such that activation at time t is influenced by past time points. The readout (output) layer indicates the category of the input, such as the binary labels "0" or "1" corresponding to task condition. Only the connections between the reservoir and the readout layer (shown in orange) are adaptable. (B) The first step of dimensionality reduction employed principal components analysis of the reservoir states. Subsequently, the dimensions were ordered based on classification information (such as the weights of a logistic regression classifier). If the top three dimensions are selected, the evolution of the reservoir can be plotted in this lower-dimensional space. In the present case, the trajectories originated from functional MRI data during the viewing of short movie clips, which were either "scary" or "funny" (Venkatesh et al., 2019). In the example, the trajectories separated quite well. Several machine learning techniques exist that are sensitive to temporal information. Among them, recurrent neural networks (RNNs) have attracted considerable attention (Williams and Zipser, 1989; Pearlmutter, 1989; Horne and Giles, 1995). However, effectively training RNNs can be challenging, particularly without large amounts of data (Pascanu et al., 2013); but for recent developments see (Martens and Sutskever, 2011; Graves et al., 2013). In our study (Venkatesh et al., 2019), we proposed to use reservoir computing to study temporal properties of brain data. This class of algorithms, which includes liquid-state machines (Maass 26 et al., 2002), echo-state networks (Jaeger, 2001; Jaeger and Haas, 2004), and related formalisms (Sussillo and Abbott, 2009), includes recurrence (like RNNs) but the learning component is only present in the read-out, or output, layer (Figure 10A). Because of the feedback connections in the reservoir, the architecture has memory properties, that is, its state depends on the current input and past reservoir states. The read-out stage can be one of many simple classifiers, including linear discrimination or logistic regression, thus providing considerable flexibility to the framework. Intuitively, reservoir computing is capable of separating complex stimuli because the reservoir projects the input onto a higher-dimensional space, making it easier to classify them. Of course, this is related to the well-known difficulty of attaining separability in low dimensions, as recognized early on with the use of perceptrons. Very briefly, the state of the reservoir can be determined as follows: 𝒙(𝑡) = 𝑓 (𝑾i𝒖(𝑡) + 𝑾𝒙(𝑡 − 1)), 𝒙(𝑡) = (1 − 𝛼)𝒙(𝑡 − 1) + 𝛼𝒙(𝑡), where 𝒙 is an intermediate state and 𝒙 is the state of the reservoir with dimensionality 𝜏𝑁, where 𝜏 is a parameter and 𝑁 is the number of input units; 𝒖 specifies the input to the system (augmented with a standard bias term of 1). The function 𝑓 is a sigmoidal function, and 𝛼 is the forgetting rate parameter. The matrix 𝑾i is the input-to-reservoir matrix and the matrix 𝑾 specifies the within-reservoir weights, both of which are generated randomly, that is, they are not learned. For more details, see (Jaeger and Haas, 2004; Lukosevicius, 2012). A central objective of our study was to investigate reservoir computing for the purposes of classifying fMRI data, in particular when temporal structure might be relevant, including both task data and data acquired during movie watching. The latter illustrates the potential of the technique for the analysis of naturalistic conditions, which are an increasing focus of research. One of the conditions we investigated was the so-called "theory of mind" task9. Participants watched 20-second clips containing simple geometrical objects (including squares, rectangles, triangles, and circles) that engaged in a potential socially relevant interaction (such as appeareing to initially fight and then make up) that unfolded throughout the duration of the clip. When watching such clips, one has the impression that the potential meaning of the 9 From the Human Connectome Project. 27 interactions gradually becomes clearer and evolves during the clip. The control condition consisted of same-duration clips using the same geometrical objects following random motion. Could the network classify theory of mind versus random clips, and in what manner was that related to temporal information? Two key parameters determine the memory properties of the reservoir: the forgetting rate, 𝛼, the and ratio of the number of reservoir-to-input units, 𝜏. Classification accuracy increased as the size of the reservoir increased, and exceeded 85% (which robustly differed from chance levels). We also trained the classifier by randomzing temporal information, namely, by randomly shuffling the data points in a block prior to training, and testing on unperturbed blocks (that is, temporally ordered). In this case, mean classification accuracy was drastically reduced to 56% correct. These and other control analyses indicated that reservoir networks were able to capture some of the temporal dynamics measured by functional MRI. We also sought to determine the dimensionality of the reservoir representation capable of classifying task conditions. The original dimensionality of our data was 360, which corresponded to the number of brain regions of interest investigated. As the goal was task classification, we selected dimensions that would contribute the most discriminative information in this regard (Figure 10B). Therefore, we performed PCA on reservoir data (that is, activation of the reservoir layer), and ordered the components based on their contributions to classification (somewhat akin to partial least squares), instead of the variance explained. Figure 11 shows classification accuracy as the number of Figure 11. Lower-dimensional representation of reservoir signals and classification accuracy. Accuracy is shown as a function of the number of dimensions. The magenta line indicates the performance using all dimensions. Classification accuracy reached within 95% of the full data with 12 dimensions. components was increased from 2 to 20 in steps of two. Remarkably, only ten principal components were required to attain classification at 95% of the level of the full dimensionality of the data. It is noteworthy that these components captured only 7% of the total variance, which should be compared to 70% if one selected components based on the amount of variance explained. Thus, only a small percentage of the original signal variance was informative for classification. 28 Using just the top three classification-related components allowed good accuracy, as shown in Figure 10A where performance is plotted as a function of time. Accuracy was initially around chance, and increased considerably between time points ~4 to ~8 seconds, Figure 12. Classification accuracy and neural trajectories. (A) Classification accuracy as a function of time during the viewing of "social" and "random" clips. (B) The mean trajectories for the two experimental conditions overlapped considerably during the initial period, but separated well with time. eventually surpassing around 80% correct. In terms of the top three dimensions, the trajectories of the social and random conditions initially overlapped, but later became quite distinct (Figure 12B). 10. Conclusions for a science of emotion and cognition Neuroscience strives to elucidate the neural underpinnings of interesting behaviors. Modern neuroscience has done so in a preponderantly reductionistic fashion for over a century and a half10. I would venture that progress has been stymied by such approach and that the time is ripe for the field to phase-transition into a period when Grossbergian themes come to the fore. By embracing head on the leitmotifs of dynamics, decentralized computation, emergence, selection and competition, and autonomy, a science of the mind-brain can be developed that is built upon a solid foundation of understanding behavior while employing computational and mathematical tools in an integral manner. At a time when the development of neurotechniques has attained a fever pitch, neuroscience needs to take stock and invest comparable energy in the conceptual and theoretical sides. Otherwise we run the risk of being able to measure every atom in the brain in a theoretical vacuum. Suppose, for that matter, that experimental physicists could measure every atom of a given galaxy. How would that advance understanding if not for a theory of 10 We can arbitrarily consider "modern neuroscience" to start with Broca's 1861 clinical report (Broca, 1861). 29 gravitation that took more than 400 years of development? The current obsession in the field with causation is equally problematic. Without theory, "causal" explanations add little to current understanding. Ultimately, to explain the cognitive-emotional brain, we need to dissolve boundaries within the brain -- perception, cognition, action, etc. -- as well as outside the brain, as we bring down the walls between biology, ecology, mathematics, computer science, philosophy, and so on11. Let's hope that the body of work by Steve Grossberg can inspire us all in this formidable endeavor. Acknowledgements The author's research is supported in part by the National Institute of Mental Health (R01 MH071589 and R01 MH112517). I thank Anastasiia Khibovska and Manasij Venkatesh for assistance with figures. I also thank Joyneel Misra and Govinda Surampudi for generating Figure 5. 11 A good example here is quantum physics which has immensely benefited from an intense, if at times strained, exchange between experimental physics, theoretical physics, and philosophy, for example. 30 Figure captions Figure 1. Fear extinction and structure-function mapping. (A) Fear extinction. (B) Conceptualization of fear extinction in terms of the top-down regulation of the amygdala by the medial prefrontal cortex, with additional variables influencing the process. (C) Schematic representation of the connections between some of the brain regions involved, emphasizing a non-hierarchical view of the processes leading to fear extinction. The descriptors "valence," "regulation," and so on, are not tied to brain areas in any straightforward one-to-one fashion. Abbreviations: CS, conditioned stimulus; MPFC, medial prefrontal cortex; OFC, orbitofrontal cortex. Figure 2. The three-body problem in Newtonian gravitation. As the problem does not admit to a general mathematical solution, researchers have sought to characterize families of periodic orbits. The figure displays the "yin-yan II" family in two dimensions. The three circles represent the three bodies. Reproduced from Li and Liao (2017). Dynamic plots can be found at http://numericaltank.sjtu.edu.cn/three-body/three-body.htm. Figure 3. Neural trajectories. Trajectories represent the activation state of the system at every point in time. (A) Recordings were performed in 87 principal neurons (PNs) of the antennal lobe of the locust during exposure to two odors (citral: cit; geraniol: ger) (Broome et al., 2006). Local linear embedding (LLE) was employed to reduce the dimensionality of the data. (B) Recordings were performed in premotor and motor cortex during reaching movements in the macaque monkey (Churchland et al., 2012). A principal components analysis-based algorithm was used to determine the two-dimensional representation displayed. Individual trials are represented by trajectories colored based on the extent of preparatory/pre-movement activity (from red to green). Figure 4. Neural trajectories during threat processing. (A) Evoked responses when transitioning from safe (green) to threat (red), and vice versa. 𝑥1and 𝑥2 represent the activity of two brain regions. (B) The responses can be jointly plotted as a function of time, (𝑥1(𝑡), 𝑥2(𝑡)), to show safe and threat trajectories. The two dots in panel A correspond to the ones here (schematically only). (C) Trajectories for safe and threat conditions when evokes responses are comparable but they are more correlated during threat. (D) Corresponding to panel D, we can think of the trajectories for threat and safe as evolving through cylinders of different diameters (which correspond to the trajectory variance). (E) Individual differences in anxiety can be understood as shifting the trajectories from safe to threat. Figure 5. Temporal trajectories based on functional MRI data. The original data were from safe and threat periods in the study by McMenamin et al. (2014). Trajectories for safe and threat 31 conditions are fairly distinct in the lower-dimensional space determined by local linear embedding. (A) Mean trajectories across individuals. The colored circles indicate the starting point. (B) Surfaces provide an indication of the underlying space of trajectories, or manifold, and were created by considering the variance of the trajectories across individuals. Figure 6. Trajectory manifolds. A manifold is a surface (more precisely a topological space) that near each point (that is, locally) resembles Euclidean space (circles and spheres are some of the simplest manifolds in two- and three dimensions). A non-Euclidean (Riemannian) metric on a manifold allows distances and angles to be measured. (A) Example manifold. The neural- dynamics space working hypothesis suggests that system behaviors can be characterized via classes of trajectories within neural spaces with particular geometry -- that is, manifolds. (B, C) Neural manifolds associated with transient dynamics are, by definition, non-periodic. Here, "population" indicates that the manifold is a group-level property. Two inter-related problems can be posed. One is estimating the group-level trajectory (red) from sample data (trajectories in orange); the other is learning the population manifold (green) from sample-level trajectories (orange). Figure 7. Geometry of the neural space. The activity of brain areas 1-3 can be mapped onto distinct properties believed to reflect their function, including univariate properties and multivariate/network-level properties. Here, a hypothetical trajectory is illustrated during a dynamic threat scenario in which threat level gradually increases (red) and then decreases (green). Figure 8. Trajectories during threat processing. Threat-level varies dynamically and increases (approach) and decreases (retreat). The temporal evolution of the system can be described in terms of the global efficiency of the salience network, the activation (evoked responses) in the same network, as well as the centrality of the amygdala/bed nucleus of the stria terminalis regions. Figure 9. Dispositional negativity. Participants perform a cognitive challenging task for an extended period of time. In one condition, there is a lower level of background threat, whereas a higher level is present in the second; the latter is anticipated to impair performance to a greater extent. Hypothetical neural trajectories are shown for the two conditions: the two trajectories will reside in separate sectors of state space, with the separation between them depending on an individual's level of dispositional negativity. Figure 10. Reservoir computing and dimensionality reduction. (A) Brain data are provided to a three-layer neural network. The input layer registers activation at time t across a set of regions of interest. The reservoir layer contains units with random connections, and provides a 32 memory mechanism such that activation at time t is influenced by past time points. The readout (output) layer indicates the category of the input, such as the binary labels "0" or "1" corresponding to task condition. Only the connections between the reservoir and the readout layer (shown in orange) are adaptable. (B) The first step of dimensionality reduction employed principal components analysis of the reservoir states. Subsequently, the dimensions were ordered based on classification information (such as the weights of a logistic regression classifier). If the top three dimensions are selected, the evolution of the reservoir can be plotted in this lower-dimensional space. In the present case, the trajectories originated from functional MRI data during the viewing of short movie clips, which were either "scary" or "funny" (Venkatesh et al., 2019). In the example, the trajectories separated quite well. Figure 11. Lower-dimensional representation of reservoir signals and classification accuracy. Accuracy is shown as a function of the number of dimensions. The magenta line indicates the performance using all dimensions. Classification accuracy reached within 95% of the full data with 12 dimensions. Figure 12. Classification accuracy and neural trajectories. (A) Classification accuracy as a function of time during the viewing of "social" and "random" clips. (B) The mean trajectories for the two experimental conditions overlapped considerably during the initial period, but separated well with time. 33 References Ahrens, M. B., Li, J. M., Orger, M. B., Robson, D. N., Schier, A. F., Engert, F., & Portugues, R. (2012). Brain-wide neuronal dynamics during motor adaptation in zebrafish. Nature, 485(7399), 471. Broca, P. (1861). Remarks on the seat of the faculty of articulated language, following an observation of aphemia (loss of speech). Bulletin de la Société Anatomique, 6, 330-57. Broome, B. M., Jayaraman, V., & Laurent, G. (2006). Encoding and decoding of overlapping odor sequences. Neuron, 51(4), 467-482. Buonomano, D. V., & Maass, W. (2009). State-dependent computations: spatiotemporal processing in cortical networks. Nature Reviews Neuroscience, 10(2), 113. Churchland, M. M., Cunningham, J. P., Kaufman, M. T., Foster, J. D., Nuyujukian, P., Ryu, S. I., & Shenoy, K. V. (2012). Neural population dynamics during reaching. Nature, 487(7405), 51. Cohen, M. A., & Grossberg, S. (1983). Absolute stability of global pattern formation and parallel memory storage by competitive neural networks. IEEE transactions on systems, man, and cybernetics, (5), 815-826. Dean, P., Redgrave, P., & Westby, G. W. (1989). Event or emergency? Two response systems in the mammalian superior colliculus. Trends in Neurosciences, 12(4), 137 -- 147. Deacon, T. W. (2011). Incomplete nature: How mind emerged from matter. WW Norton & Company. Deyle, E. R., & Sugihara, G. (2011). Generalized theorems for nonlinear state space reconstruction. PLoS One, 6(3), e18295. Do-Monte, F. H., Quinones-Laracuente, K., & Quirk, G. J. (2015). A temporal shift in the circuits mediating retrieval of fear memory. Nature, 519(7544), 460. Do-Monte, F. H., Manzano-Nieves, G., Quiñones-Laracuente, K., Ramos-Medina, L., & Quirk, G. J. (2015). Revisiting the role of infralimbic cortex in fear extinction with optogenetics. Journal of Neuroscience, 35(8), 3607-3615. Dunsmoor, J. E., Niv, Y., Daw, N., & Phelps, E. A. (2015). Rethinking extinction. Neuron, 88(1), 47-63. Elsayed, G. F., & Cunningham, J. P. (2017). Structure in neural population recordings: an expected byproduct of simpler phenomena?. Nature Neuroscience, 20(9), 1310. Ferstl, E.C., Rinck, M., Cramon, D. Y. v., 2005. Emotional and temporal aspects of situation model processing during text comprehension: an event-related fmri study. J. Cognit. Neurosci. 17 (5), 724 -- 739. Fox, E. (2018). Perspectives from affective science on understanding the nature of emotion. Brain and Neuroscience Advances, 2, 2398212818812628. 34 Fox, A. S., Lapate, R. C., Shackman, A. J., & Davidson, R. J. (Eds.). (2018). The nature of emotion: Fundamental questions. Oxford University Press. Frankel, L. (1986). Mutual causation, simultaneity and event description. Philosophical Studies, 49(3), 361-372. Frégnac, Y. (2017). Big data and the industrialization of neuroscience: A safe roadmap for understanding the brain?. Science, 358(6362), 470-477. Fuster, J. M. (2001). The prefrontal cortex -- an update: time is of the essence. Neuron, 30(2), 319-333. Gao, P., Trautmann, E., Byron, M. Y., Santhanam, G., Ryu, S., Shenoy, K., & Ganguli, S. (2017). A theory of multineuronal dimensionality, dynamics and measurement. bioRxiv, 214262. Gomez-Marin, A. (2017). Causal circuit explanations of behavior: Are necessity and sufficiency necessary and sufficient?. In Decoding Neural Circuit Structure and Function (pp. 283-306). Springer, Cham. Gomez-Marin, A., Paton, J. J., Kampff, A. R., Costa, R. M., & Mainen, Z. F. (2014). Big behavioral data: psychology, ethology and the foundations of neuroscience. Nature neuroscience, 17(11), 1455. Granger, C. W. (1969). Investigating causal relations by econometric models and cross-spectral methods. Econometrica: Journal of the Econometric Society, 424-438. Graves, A., Mohamed, A.-r., Hinton, G., 2013. Speech recognition with deep recurrent neural networks. In: Acoustics, Speech and Signal Processing (icassp), 2013 IEEE International Conference on. IEEE, pp. 6645 -- 6649. Grossberg, S. (1964). The theory of embedding fields with applications to psychology and neurophysiology. Rockefeller Institute for Medical Research, monograph (451 pp.). Grossberg, S. (2018). Desirability, availability, credit assignment, category learning, and attention: Cognitive- emotional and working memory dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal cortices. Brain and Neuroscience Advances, 2, 2398212818772179. Grossberg, S., & Levine, D. S. (1987). Neural dynamics of attentionally modulated Pavlovian conditioning: blocking, interstimulus interval, and secondary reinforcement. Applied optics, 26(23). Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., Malach, R., 2004. Intersubject synchronization of cortical activity during natural vision. Science 303 (5664), 1634 -- 1640. Heimer, L., van Hoesen, G. W., Trimble, M., & Zahm, D. S. (2007). Anatomy of neuropsychiatry: The new anatomy of the basal forebrain and its implications for neuropsychiatric illness. Burlington, MA: Academic Press. Herry, C., Ciocchi, S., Senn, V., Demmou, L., Müller, C., & Lüthi, A. (2008). Switching on and off fear by distinct neuronal circuits. Nature, 454(7204), 600. 35 Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences, 79(8), 2554-2558. Hopfield, J. J. (1984). Neurons with graded response have collective computational properties like those of two- state neurons. Proceedings of the national academy of sciences, 81(10), 3088-3092. Horne, B.G., Giles, C.L., 1995. An experimental comparison of recurrent neural networks. In: Advances in Neural Information Processing Systems, pp. 697 -- 704. Hur, J., Stockbridge, M. D., Fox, A. S., & Shackman, A. J. (2018). Dispositional negativity, cognition, and anxiety disorders: An integrative translational neuroscience framework. PsyArXiv. December, 11. Jaeger, H., Haas, H., 2004. Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science 304 (5667), 78 -- 80. Keiflin, R., & Janak, P. H. (2015). Dopamine prediction errors in reward learning and addiction: from theory to neural circuitry. Neuron, 88(2), 247-263. Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poeppel, D. (2017). Neuroscience needs behavior: correcting a reductionist bias. Neuron, 93(3), 480-490. Latora, V., & Marchiori, M. (2001). Efficient behavior of small-world networks. Physical review letters, 87(19), 198701. Li, X., & Liao, S. (2017). More than six hundred new families of Newtonian periodic planar collisionless three-body orbits, Science China Physics, Mechanics & Astronomy. DOI: 10.1007/s11433-017-9078-5 Lovett-Barron, M., Andalman, A. S., Allen, W. E., Vesuna, S., Kauvar, I., Burns, V. M., & Deisseroth, K. (2017). Ancestral circuits for the coordinated modulation of brain state. Cell, 171(6), 1411-1423. Lukoševičius, M. (2012). A practical guide to applying echo state networks. In Neural networks: Tricks of the trade (pp. 659-686). Springer, Berlin, Heidelberg. Maass, W., Natschl€ager, T., Markram, H., 2002. Real-time computing without stable states: a new framework for neural computation based on perturbations. Neural Comput. 14 (11), 2531 -- 2560. Mannino, M., & Bressler, S. L. (2015). Foundational perspectives on causality in large-scale brain networks. Physics of life reviews, 15, 107-123. Maren, S., Phan, K. L., & Liberzon, I. (2013). The contextual brain: implications for fear conditioning, extinction and psychopathology. Nature reviews neuroscience, 14(6), 417. Martens, J., Sutskever, I., 2011. Learning recurrent neural networks with hessian-free optimization. In: Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 1033 -- 1040. Citeseer. 36 McCulloch, W. S. (1945). A heterarchy of values determined by the topology of nervous nets. The bulletin of mathematical biophysics, 7(2), 89-93. McMenamin, B. W., Langeslag, S. J., Sirbu, M., Padmala, S., & Pessoa, L. (2014). Network organization unfolds over time during periods of anxious anticipation. Journal of Neuroscience, 34(34), 11261-11273. Mehler, D. M. A., & Kording, K. P. (2018). The lure of causal statements: Rampant mis-inference of causality in estimated connectivity. arXiv preprint arXiv:1812.03363. Menon, V., & Uddin, L. Q. (2010). Saliency, switching, attention and control: a network model of insula function. Brain Structure and Function, 214(5-6), 655-667. Meyer, C., Padmala, S., Pessoa, L. (2019). Dynamic threat processing. Journal of Cognitive Neuroscience, in press. Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual review of neuroscience, 24(1), 167-202. Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H., Howerter, A., & Wager, T. D. (2000). The unity and diversity of executive functions and their contributions to complex "frontal lobe" tasks: A latent variable analysis. Cognitive psychology, 41(1), 49-100. Monsell, S., & Driver, J. (2000). Banishing the control homunculus. Control of cognitive processes: Attention and performance XVIII, 3-32. Morgan, M. A., Romanski, L. M., & LeDoux, J. E. (1993). Extinction of emotional learning: contribution of medial prefrontal cortex. Neuroscience letters, 163(1), 109-113. Morgane, P. J. (1979). Historical and modern concepts of hypothalamic organization and function. Anatomy of the Hypothalamus, 1, 1-64. Najafi, M., Kinnison, J., & Pessoa, L. (2017). Dynamics of intersubject brain networks during anxious anticipation. Frontiers in human neuroscience, 11, 552. Nieuwenhuys, R., Voogd, J., & Huijzen, C. V. (2008). The Human Central Nervous System. Springer (Fourth edition). Olds, J., & Milner, P. (1954). Positive reinforcement produced by electrical stimulation of septal area and other regions of rat brain. Journal of comparative and physiological psychology, 47(6), 419. Packard, Norman H., James P. Crutchfield, J. Doyne Farmer, and Robert S. Shaw. "Geometry from a time series." Physical review letters 45, no. 9 (1980): 712. Pascanu, R., Mikolov, T., Bengio, Y., 2013. On the difficulty of training recurrent neural networks. In: International Conference on Machine Learning, pp. 1310 -- 1318. Pascoli, V., Hiver, A., Van, R. Z., Loureiro, M., Achargui, R., Harada, M., ... & Lüscher, C. (2018). Stochastic synaptic plasticity underlying compulsion in a model of addiction. Nature, 564(7736), 366-371. 37 Pearl, J. (2009). Causality. Cambridge university press. Pearlmutter, B.A., 1989. Learning state space trajectories in recurrent neural networks. Neural Comput. 1 (2), 263 -- 269. Peek, M. Y., & Card, G. M. (2016). Comparative approaches to escape. Current Opinion in Neurobiology, 41, 167 -- 173. Penzo, M. A., Robert, V., Tucciarone, J., De Bundel, D., Wang, M., Van Aelst, L., ... & Huang, Z. J. (2015). The paraventricular thalamus controls a central amygdala fear circuit. Nature, 519(7544), 455. Pereira, A. G., & Moita, M. A. (2016). Is there anybody out there? Neural circuits of threat detection in vertebrates. Current Opinion in Neurobiology, 41, 179 -- 187. doi:10.1016/j.conb.2016.09.011 Pessoa, L. (2014). Understanding brain networks and brain organization. Physics of life reviews, 11(3), 400-435. Pessoa, L. (2017a). The emotional brain. In Conn's Translational Neuroscience (pp. 635-656). Academic Press. Pessoa, L. (2017b). A network model of the emotional brain. Trends in cognitive sciences, 21(5), 357-371. Pessoa, L. (2017c). Cognitive-motivational interactions: Beyond boxes-and-arrows models of the mind-brain. Motivation Science, 3(3), 287. Pessoa, L. (2018a): Embracing integration and complexity: placing emotion within a science of brain and behaviour, Cognition and Emotion, DOI: 10.1080/02699931.2018.1520079 Pessoa, L. (2018b). Emotion and the Interactive Brain: Insights from Comparative Neuroanatomy and Complex Systems. Emotion Review, 10(3), 204-216. Pessoa, L. (2018c). Understanding emotion with brain networks. Current opinion in behavioral sciences, 19, 19-25. Pessoa, L., & McMenamin, B. (2017). Dynamic networks in the emotional brain. The Neuroscientist, 23(4), 383-396. Rabinovich, M., Huerta, R., & Laurent, G. (2008). Transient dynamics for neural processing. Science, 48-50. Ren, S., Wang, Y., Yue, F., Cheng, X., Dang, R., Qiao, Q., ... & Qin, H. (2018). The paraventricular thalamus is a critical thalamic area for wakefulness. Science, 362(6413), 429-434. Roweis, S. T., & Saul, L. K. (2000). Nonlinear dimensionality reduction by locally linear embedding. science, 290(5500), 2323-2326. 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. Journal of neurophysiology, 102(2), 1315-1330. Sauer, T., Yorke, J. A., & Casdagli, M. (1991). Embedology. Journal of statistical Physics, 65(3-4), 579-616. Shallice, T (1988). From neuropsychology to mental structure. New York: Cambridge University Press. 38 Striedter, G.F. (2005). Principles of brain evolution, Sinauer Associates. Sugihara, G., May, R., Ye, H., Hsieh, C. H., Deyle, E., Fogarty, M., & Munch, S. (2012). Detecting causality in complex ecosystems. science, 1227079. Sussillo, D., Abbott, L.F., 2009. Generating coherent patterns of activity from chaotic neural networks. Neuron 63 (4), 544 -- 557. Šuvakov, M., & Dmitrašinović, V. (2013). Three classes of Newtonian three-body planar periodic orbits. Physical review letters, 110(11), 114301. Tovote, P., Fadok, J. P., & Lüthi, A. (2015). Neuronal circuits for fear and anxiety. Nature Reviews Neuroscience, 16(6), 317-331. Venkatesh, M., Jaja, J., & Pessoa, L. (2019). Brain dynamics and temporal trajectories during task and naturalistic processing. NeuroImage, 186, 410-423. Verbruggen, F., McLaren, I. P., & Chambers, C. D. (2014). Banishing the control homunculi in studies of action control and behavior change. Perspectives on Psychological Science, 9(5), 497-524. von Bertalanffy, L. (1950). An outline of general system theory. The British Journal for the Philosophy of Science, 1(2), 134 -- 165. Williams, R.J., Zipser, D., 1989. A learning algorithm for continually running fully recurrent neural networks. Neural Comput. 1 (2), 270 -- 280. Zacks, J.M., Braver, T.S., Sheridan, M.A., Donaldson, D.I., Snyder, A.Z., Ollinger, J.M., Buckner, R.L., Raichle, M.E., 2001. Human brain activity time-locked to perceptual event boundaries. Nat. Neurosci. 4 (6), 651. 39
1202.3539
1
1202
2012-02-16T09:23:15
Multiple firing coherence resonances in excitatory and inhibitory coupled neurons
[ "q-bio.NC", "cond-mat.dis-nn", "nlin.PS", "physics.bio-ph" ]
The impact of inhibitory and excitatory synapses in delay-coupled Hodgkin--Huxley neurons that are driven by noise is studied. If both synaptic types are used for coupling, appropriately tuned delays in the inhibition feedback induce multiple firing coherence resonances at sufficiently strong coupling strengths, thus giving rise to tongues of coherency in the corresponding delay-strength parameter plane. If only inhibitory synapses are used, however, appropriately tuned delays also give rise to multiresonant responses, yet the successive delays warranting an optimal coherence of excitations obey different relations with regards to the inherent time scales of neuronal dynamics. This leads to denser coherence resonance patterns in the delay-strength parameter plane. The robustness of these findings to the introduction of delay in the excitatory feedback, to noise, and to the number of coupled neurons is determined. Mechanisms underlying our observations are revealed, and it is suggested that the regularity of spiking across neuronal networks can be optimized in an unexpectedly rich variety of ways, depending on the type of coupling and the duration of delays.
q-bio.NC
q-bio
Multiple firing coherence resonances in excitatory and inhibitory coupled neurons Qingyun Wang,1, ∗ Honghui Zhang,1 Matjaz Perc,2, † and Guanrong Chen3 1Department of Dynamics and Control, Beihang University, Beijing, China 2Department of Physics, Faculty of Natural Sciences and Mathematics, University of Maribor, Slovenia 3Department of Electronic Engineering, City University of Hong Kong, Hong Kong SAR, China 2 1 0 2 b e F 6 1 ] . C N o i b - q [ 1 v 9 3 5 3 . 2 0 2 1 : v i X r a The impact of inhibitory and excitatory synapses in delay-coupled Hodgkin–Huxley neurons that are driven by noise is studied. If both synaptic types are used for coupling, appropriately tuned delays in the inhibition feedback induce multiple firing coherence resonances at sufficiently strong coupling strengths, thus giving rise to tongues of coherency in the corresponding delay-strength parameter plane. If only inhibitory synapses are used, however, appropriately tuned delays also give rise to multiresonant responses, yet the successive delays warranting an optimal coherence of excitations obey different relations with regards to the inherent time scales of neuronal dynamics. This leads to denser coherence resonance patterns in the delay-strength parameter plane. The robustness of these findings to the introduction of delay in the excitatory feedback, to noise, and to the number of coupled neurons is determined. Mechanisms underlying our observations are revealed, and it is suggested that the regularity of spiking across neuronal networks can be optimized in an unexpectedly rich variety of ways, depending on the type of coupling and the duration of delays. Keywords: coherence resonance, synaptic coupling, information transmission delay, regularity of spiking, time scales I. INTRODUCTION Neurophysiological studies have revealed the existence of accurately timed patterns of spikes by a variety of cogni- tive and motoric tasks [1–6]. The timing of these spikes, or neuronal firings, is accurate to within the millisecond range, which poses great challenges with regards to the identification of mechanisms that would be able to ensure such precision. Following their initial observation in the cortex of monkeys [1, 2], the precisely timed spikes have been reported and in- vestigated for motor functions [3], the neuronal response of visual systems [4], and the complex spatial fingertip events [5], to name but a few examples. Not surprisingly, synchro- nized, precisely timed firings can be observed at virtually all neuronal processing levels, including the retina [9], the lateral geniculate nucleus [10], and the cortex [11, 12]. Since it is well known that noise can play a constructive role in different types of nonlinear dynamical systems, which arguably describe also neuronal dynamics [13], this opens the possibility of exploiting such mechanisms for explaining, or at least supporting, the aforementioned precision of neuronal firings. Stochastic resonance [14–16] and coherence reso- nance [17–19] are amongst the most prominent examples by means of which noise of appropriate intensity is able either to enhance the detection of weak deterministic signals [20] or evoke coherent response in nonlinear dynamical systems in the absence of any deterministic inputs. The potential benefits of noise range from ice ages to crayfish and SQUIDs [21], to neural systems, as most recently reviewed in [22]. Following initial advances on individual dynamical sys- tems, the focus begun shifting to spatially extended systems [23], especially also to such with complex networks describ- ing connections between the individual units [24, 25]. For ∗Electronic address: [email protected] †Electronic address: [email protected] example, coherence resonance on a small world network was investigated in [26], while array-enhanced resonances were reported in [27]. Moreover, spatial coherence resonance was observed first near pattern-forming instabilities [28], and lat- ter also in excitable media [29]. Excitable systems in general proved to be very susceptible to a multitude of noise-induced phenomena, as reviewed comprehensively in [30]. Adding spatial degrees of freedom, along with the possibilities for in- troducing other sources of heterogeneity, lead to the discovery of very interesting and quite exotic phenomena, such as the ghost resonance [31], and double as well as multiple stochas- tic [32–35] and coherence [36–39] resonances. For neural systems, a wealth of interesting and new phe- nomena was made observable by integrating realistic features of neuronal dynamics into the studied models. Information transmission delays or synaptic delays, for example, are inher- ent to the nervous system because of the finite speed at which action potentials propagate across neuron axons, and due to time lapses occurring at both dendritic and synaptic process- ing [40]. Following seminal works examining the impact of delays on excitable and other dynamical systems [41–43], the stability and attainability of synchronous oscillations [44–46] and the role of delays in shaping spatiotemporal dynamics of neuronal activity [47] were investigated. Moreover, the role of delays in coupled Hodgkin-Huxley neurons was also investi- gated for the phenomenon of coherence resonance, and it was reported that properly tuned delays can lead to the occurrence of multiple resonances [48, 49]. In this letter, we extend the scope of coherence resonance in models of neuronal dynamics by considering besides synap- tic delays also different types of synaptic coupling. While the role of chemical synapses in coupled neurons with noise has been investigated in [50], and although the general dy- namics of sparsely connected networks of excitatory and in- hibitory spiking neurons is known [51], our approach, joining these distinctive features of neuronal dynamics (synaptic de- lays, different types of synaptic coupling, and noise), allows for the identification of new ways by means of which the co- herence, and thus the accuracy of neuronal firings, can be im- proved. Most interestingly, we report the occurrence of multi- ple coherence resonance patterns in the corresponding delay- strength parameter plane when either inhibitory and excitatory or only inhibitory synapses are used for coupling. The details of these multiple firing coherence resonances, and in particu- lar the conditions at which they occur, however, depend sig- nificantly on the type of coupling. Reported results suggest that characteristic time scales related to the information trans- mission and inhibition in neuronal networks may interplay in intricate ways, and by doing so give rise to new mechanisms for optimizing spiking regularity. The remainder of this letter is organized as follows. In the next section we describe the model, then we present the main results separately for the two coupling scenarios, while lastly we summarize our findings and discuss their potential impli- cations. II. MODEL DEFINITION For simplicity, we consider two Hodgkin–Huxley neurons [13] that are coupled by inhibitory and/or excitatory synapses. Equations describing the dynamics are: C dVi dt dmi dt dhi dt dni dt = −gN am3h(Vi − VN a) − gL(Vi − VL) −gKXKn4(Vi − VK) + I + σξi(t) + I i,j syn,(1) = αmi (1 − mi) − βmi mi, = αhi(1 − hi) − βhihi, = αni (1 − ni) − βni ni, (2) (3) (4) where Vi is the transmembrane potential of the i-th neuron. Moreover, mi, hi and ni are the gating variables, where the voltage-dependent opening and closing rates are: αmi = 0.1(Vi + 10) 1 − exp[− (Vi+40) (Vi + 65) 10 , ] (cid:21) , βmi = 4exp(cid:20)− αhi = 0.07exp(cid:20)− βhi = (cid:26)1 + exp(cid:20)− 18 (Vi + 65) 20 (Vi + 35) 10 (cid:21) , (cid:21)(cid:27)−1 αni = 0.01(Vi + 55) 1 − exp[− (Vi+55) , ] βni = 0.125exp(cid:20)− 10 (Vi + 65) 80 (cid:21) , (5) (6) (7) (8) (9) (10) , The membrane capacity is C = 1 (µF/cm2), and gN a = 120 µF/cm2, gK = 36 µF/cm2 and gL = 0.3 µF/cm2 are the max- imal sodium, potassium and leakage conductances, respec- tively. The corresponding reversal potentials are VN a = 50 2 mV, VK = −77 mV and VL = −54.4 mV. Using these pa- rameter values, a single Hodgkin–Huxley neuron has a sub- critical Hopf bifurcation at the external current I = I1 = 9.8µA/cm2. Between I = I2 = 6.2µA/cm2 and I1 sta- ble limit cycles coexist with stable steady states, whereas for I < I2 (I > I1) excitable steady states (limit cycles) are the only stable solutions. If I > 155µA/cm2, on the other hand, the oscillations vanish by means of a supercritical Hopf bifur- cation. A more detailed bifurcation analysis of the Hodgkin– Huxley model was performed in [52, 53]. Here we are in- terested in the region I < I2, where neurons are unable to fire spontaneously, i.e, remain forever quiescent in the ab- sence of external stimuli. We thus set I = 6.1µA/cm2, so that both neurons are in an excitable steady state. Gaussian noise ξi(t), having mean < ξi(t) >= 0 and autocorrelation < ξi(t)ξj (t′) >= δij δ(t − t′), thus acts as the source of large- amplitude excitations, where σ determines the noise intensity. We consider two different coupling schemes. First, the two neurons are coupled in a hybrid way using inhibitory and ex- citatory synapses. The coupling terms in this case are: I 1,2 syn = −gexc I 2,1 syn = −ginh (V1 − Vexc) (1 + exp{−λ[V2(t) − Θs]}) , (11) (V2 − Vinh) (1 + exp{−λ[V1(t − τ ) − Θs]}) , (12) where the inhibitory feedback is delayed by τ. Second, only inhibitory synapses are used for coupling, in which case the coupling becomes: I i,j syn = −ginh (Vi − Vinh) (1 + exp{−λ[Vj(t − τ ) − Θs]}) , (13) where the inhibitory feedback is again delayed by τ, only that here this applies to both directions. In the above cou- pling terms ginh(exc) determines the strength of the synaptic conductance, i.e., the coupling strength, while Vinh = −80 mV and Vexc = 20 mV are the reversal potentials for the in- hibitory and the excitatory synapse, respectively. Moreover, Θs = 0 is the threshold, above which the postsynaptic neuron is affected by the presynaptic one, and λ = 10 is a constant rate for the onset of excitation or inhibition. In what follows, we will investigate the impact of the delay τ and the coupling strength ginh(exc) on the occurrence of firing coherence res- onance, and we will do so separately for the two described coupling schemes. III. RESULTS We start by presenting the results as obtained with hybrid coupling, i.e., when excitatory and inhibitory synapses are used for connecting the two neurons. Figure 1 features char- acteristic time courses of the transmembrane potential V of the excitatory neuron, from where it can be observed at a glance that the coherence of excitations depends critically on the delay of the inhibitory feedback τ. Importantly though, the relation between the coherency and the value of τ is not monotonous, but rather it is intermittent. That is to say, as τ ) t ( V 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 600 800 1000 t 1200 1400 FIG. 1: Appropriately adjusted delays τ in the one-directional in- hibition feedback enhance the regularity of spiking by hybrid cou- pling of the two neurons. Depicted are characteristic time courses of the transmembrane potential V of the excitatory neuron for different values of τ : (a) 0, (b) 8.0, (c) 20, (d) 24, (e) 35 and (f) 40. It can be observed that the regularity of spiking in panels (b), (d) and (f) (traces depicted green) is higher than in panels (a), (c) and (e) (traces depicted red). Other parameter values are: gexc = 0.11, ginh = 1.0 and σ = 1.5. increases the regularity is lost and regained intermittently as different values of τ come to determine the delay of inhibi- tion. Time courses depicted green (panels b, d and f) exhibit more coherent spiking than time courses depicted red (pan- els a, c and e). This is characteristic for multiresonant phe- nomena, and in fact these observations can be made quanti- tatively more precise by introducing a coherence measure C as follows. Let the sequence t0 < t1 < t2 < · · · < tn denote the firing times of the considered neuron. From the se- quence of {tk}, the interspike intervals (ISI) are determined as Tk = tk − tk−1(k = 1, 2, · · · , n). To characterize the co- herence of the firings, the measure C is defined as C = p< T 2 k > − < Tk >2 < Tk > . (14) where h·i is the time average. In particular, C is the ratio of the standard deviation and the average of the interspike in- tervals, and it is indeed an excellent quantity for effectively determining the occurrence of coherence resonance from neu- ronal firing. From Eq. (14) it follows that the more coherent the firing, the smaller the value of C. We would also like to note that C is the reciprocal of the coefficient of variation in a 3 ginh = 0.5 ginh = 1.0 ginh = 1.5 ginh = 2.5 0 10 20 30 40 50 b f e d c b a C 0.6 0.5 0.4 0.3 0.2 0.1 50 40 30 20 10 0 0 1 2 3 ginh 4 5 FIG. 2: Delay-induced multiresonances in case of hybrid coupling of the two neurons. Panel (a) shows the coherence measure C in dependence on τ for different values of ginh. It can be observed that the stronger the coupling the better expressed the recurrently ap- pearing minima of C. Panel (b) features the contours of C (white depicts minimal and black maximal values) on the corresponding delay-strength τ − ginh parameter plane, where multiple tongues of coherency (white) emerge due to an interplay between the synaptic delay τ and the characteristic time scale of the two Hodgkin–Huxley neurons (as determined by the characteristic excitatory time Te and the complex conjugate part of the eigenvalues of the excitatory steady state). Other parameter values are: σ = 1.5. a point process, which is widely used in the field of neuro- science [54]. Using the introduced coherence measure C, we demon- strate in Fig. 2 the occurrence of multiresonant behavior in dependence on τ. Results presented in panel (a) indicate that C has several minima in the considered interval of τ, and that these are better pronounced, i.e., less susceptible to statisti- cal deviations, for larger coupling strengths ginh. In general, however, the dependence of C on ginh is fairly insignificant, pointing towards the fact that in case of hybrid coupling the strength of the synaptic conductance of one type (e.g., the in- hibitory type) has little impact if the other (e.g., the excitatory type) remains unchanged. The contours in panel (b) confirm this, as the tongues of coherency (white regions) simply shrink in width as ginh decreases, but otherwise do not alter the de- pendence of C on the inhibition delay τ. In many ways, these results are reminiscent of delay-induced multiple stochastic resonances that were previously reported for scale-free neu- ronal networks [33], and are indicative for an interplay be- tween the time scales inherent to the system dynamics and the ) t ( V 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 40 0 -40 -80 600 f e d c b a 800 1000 t 1200 1400 a 0.5 0.4 C 0.3 4 ginh = 1.0 ginh = 1.5 ginh = 2.0 ginh = 3.0 0.2 0.1 50 40 30 20 10 0 0 10 20 30 40 50 b 0.5 1.0 1.5 2.0 ginh 2.5 3.0 3.5 4.0 FIG. 3: Appropriately adjusted delays τ in the bidirectional inhibi- tion feedback enhance the regularity of spiking by inhibitory cou- pling of the two neurons. Depicted are characteristic time courses of the transmembrane potential V of one neuron for different values of τ : (a) 0, (b) 2.0, (c) 5.0, (d) 11, (e) 15 and (f) 19. As in Fig. 1, it can be observed that the regularity of spiking in panels (b), (d) and (f) (traces depicted green) is higher than in panels (a), (c) and (e) (traces depicted red). Other parameter values are: ginh = 0.75 and σ = 1.5. time scales introduced by means of the delay. Turning to the second coupling scheme relying only on in- hibitory synapses, however, we find somewhat unexpected re- sults. While the time courses of the transmembrane potential V presented in Fig. 3 do not suggest quantitatively different behavior in that certain values of τ warrant higher coherency of spiking than other values (which is also what we can ob- serve in Fig. 1), a more accurate quantitative analysis pre- sented in Fig. 4 indicates otherwise. In particular, in panels (a) and (b) we find that the minima of C are much more fre- quent in the considered span of τ values as this was the case for hybrid coupling. While for the later a total of three minima can be observed within 0 ≤ τ ≤ 50 (see Fig. 2), for purely in- hibitory coupling twice as many minima are inferable within the same span of τ values. The origins of these multiresonant phenomena can be linked to different inherent properties of neuronal dynamics. First, it is useful to define the so called average excitatory time Te, which is the average time between two consecutive spikes. For an isolated Hodgkin–Huxley neuron driven by noise this time decreases and saturates towards Te ≈ 16 for σ ≥ 4.0 (note that this corresponds to a strong noise limit, FIG. 4: Delay-induced multiresonances in case of inhibitory cou- pling of the two neurons. Panel (a) shows the coherence measure C in dependence on τ for different values of ginh. As in Fig. 2, it holds that the stronger the coupling the better expressed the recurrently ap- pearing minima of C. However, in the considered span of τ values, twice as many minima as by hybrid coupling can be observed. Panel (b) features the contours of C (white depicts minimal and black max- imal values) on the corresponding delay-strength τ − ginh parameter plane, where the much denser tongues of coherency are clearly in- ferable. This indicates that the interplay between the synaptic delay τ and the characteristic time scale of the two Hodgkin–Huxley neu- rons is more efficient by purely inhibitory coupling. Other parameter values are: σ = 1.5. above which the system may already exhibit numerical insta- bility). Increasing the noise intensity further and lowering the time step for numerical integration, it is in principle possi- ble to arrive at even lower average excitatory times Te ≈ 12, which agrees with the theoretical prediction stemming from the imaginary parts of the complex conjugate eigenvalues Imλi,j = ±iω = ±i0.54, where Te = 2π/ω = 11.63. Since in our simulations, however, we use a comparatively low noise intensity σ = 1.5, the average excitatory time Te ≈ 16 of an isolated Hodgkin–Huxley neuron is the more accurate approx- imation for the inherent time scale of the considered neuronal dynamics. For hybrid coupling, we thus find the first mini- mum of C at Te/2, and subsequent minima at odd multiples of the half of the average excitatory time [see Fig. 2(a)], which agrees with the doubly effect of the two considered synaptic types [55]. The average excitatory time is reflected also in the time courses presented in Figs. 1(b,d,f) (note that in these the firing is accurate and ordered due to the constructive impact of τ), where the average spiking period is approximately equal a 0.5 0.4 C 0.3 0.2 0.1 0.7 0.6 0.5 0 b immediate excitatory feedback delayed excitatory feedback 10 20 30 40 = 1.0 = 2.0 = 3.0 C 0.4 0.3 0.2 0.1 0 10 20 30 40 a 0.45 0.40 0.35 C 0.30 0 b 0.25 0.20 0.15 0.35 0.30 0.25 0.20 0 C 5 gexc = 0.5 gexc = 1.0 gexc = 1.5 10 20 30 40 ginh = 1.5 ginh = 2.0 ginh = 2.5 10 20 30 40 FIG. 5: Delay-induced multiresonances in the presence of additional delay in the excitatory feedback and noise. Panel (a) features a com- parison of the coherence measure C as obtained with and without excitatory synaptic delay in dependence on τ for hybrid coupling. It can be observed that the introduction of delays in the excitatory feedback can substantially reduce delays warranting the most coher- ent response. Other parameter values are: gexc = 1.0, σ = 1.5. Panel (b) depicts C in dependence on τ for different values of the noise intensity σ in two purely inhibitory coupled neurons. It can be observed that as the intensity of noise increases the maximally at- tainable values of C decrease (yet the effect saturates for higher σ). Optimal delays, however, remain unaffected by noise, which indi- cates robustness of the observed delay-induced multiresonances. FIG. 6: Delay-induced multiresonances in a ring network consisting of 100 neurons. Panel (a) features results as obtained with delay in the excitatory feedback (gexc) and hybrid coupling. Panel (b), on the other hand, depicts C in dependence on τ as obtained with delay in the inhibitory feedback (ginh) and purely inhibitory coupling. Based on the presented results, it can be concluded that multiresonances in a ring network can be observed irrespective of the coupling and delay type, if only the delays are appropriately adjusted. However, delays warranting optimal coherence in the network with purely inhibitory coupling (b) are smaller that those in the network with hybrid cou- pling (a). Other parameter values are: ginh = 1.5 [applicable for panel (a) only] and σ = 1.5. to Te. Conversely, for inhibitory coupling, the matching of the time scales leading to the multiresonant dependence of C on τ is different. Although the average excitatory time Te ≈ 16 is likewise [as in Figs. 1(b,d,f)] reflected in the correspond- ing time courses presented in Figs. 3(b,d,f), which have the same average inter-spike interval, twice as many minima im- ply that the resonant matching occurs not just for odd multi- ples of Te/2, but in fact for odd and even multiples. However, all the minima of C are preceded by a small delay of 2s (where the first minimum occurs) that is necessary for the first reso- nant response. Since the purely inhibitory type of synaptic coupling lacks the excitatory input that is present by hybrid coupling, in the former case the matching of the time scales is twice as efficient. Finally it is of interest to examine the robustness of our findings in the presence of delayed excitatory feedback, dif- ferent levels of noise, and for different sizes of the network. In Fig. 5(a), we present the results with and without delayed ex- citatory feedback in a hybridly coupled two-neuron system. It can be observed that, while multiresonances can be observed in both cases, the introduction of delays also in the excita- tory feedback (in addition to delays in the inhibitory feedback) may substantially reduce the delays that warrant an optimal response of the system (maximal values of C). Thus, delayed excitatory feedback does affect the results quantitatively, yet it does not affect the qualitative picture. Figure 5(b) shows that different noise intensities σ have a similar impact. In particu- lar, while higher values of σ may reduce maximally attainable values of C, the multiple maxima are always clearly inferable and their positions do not shift. Hence, noise is also unable to significantly affect the results. Lastly, we present in Fig. 6 results obtained on a larger ring network for the two different coupling types. Regardless of whether the coupling is hybrid with delays introduced to both types of synapses [panel (a)] or purely inhibitory [panel (b)], the multiple coherence reso- nances are clearly inferable. Importantly, also on larger net- works the purely inhibitory mode of interneuronal communi- cation appears to be more efficient (there are more maxima of C in a given span of τ) than the hybrid mode, which fully agrees with our conclusions obtained by means of the analysis of the two-neuron system, and thus solidifies the high robust- ness of our main conclusions, which we will summarize in what follows. IV. SUMMARY Summarizing, we have demonstrated the occurrence of multiresonant elevation of firing precision, as quantified by means of a coherence measure, in synaptically coupled Hodgkin–Huxley neurons. We have separately considered hy- brid and purely inhibitory coupling, and we have discovered that the resonant matching of the different time scales that are inherent to the Hodgkin–Huxley model (and the informa- tion transmission delay) is twice as efficient in the latter case. Our results thus reveal unexpected possibilities for the reso- nant enhancement of firing precision by means of matching of different time scales of neuronal dynamics. Moreover, we have examined the robustness of our findings to the introduc- tion of delay in the excitatory feedback, to noise, and to the 6 number of coupled neurons. We have found that delayed exci- tatory feedback may substantially reduce the length of delays that ensure an optimal response of the system, yet that it does not qualitatively affect the results. Neither do noise and the size of the network, which led us to the conclusion that the reported results are highly robust, and that they are thus ex- pected to remain valid also in other related neuronal systems. We hope that our study will prove useful for facilitating the de- velopment of concepts such as function-follow-form [56, 57] and the application of methods of statistical physics for better understanding conditions such as epilepsy [58–60] and other neurodegenerative diseases, as well as for better understand- ing the mechanisms behind high-precision firing patterns in more realistic neuronal networks. Acknowledgments This research was supported by the National Science Foun- dation of China (grants 11172017 and 10832006) and by the Slovenian Research Agency (grant J1-4055). [1] R. Lestienne, B. Strehler, Time structure and stimulus depen- dence of precisely replicating patterns present in monkey corti- cal neuronal spike trains, Brain Res. 437 (1987) 214. [2] M. Abeles, H. Bergman, F. Margalit, E. Vaadia, Spatiotempo- ral firing patterns in the frontal cortex of behaving monkeys, J. Neurophysiol. 70 (1993) 1629. [3] A. Riehle, S. Grun, M. Diesmann, A. Aertsen, Spike synchro- nization and rate modulation differentially involved in motor function, Science 278 (1997) 1950. [4] M. Oram, M. Wiener, R. Lestienne, B. Richmond, Stochastic nature of precisely timed spike patterns in visual system neu- ronal responses, J. Neurophysiol. 81 (1999) 3021. [5] R. S. Johansson, I. Birznieks, First spikes in ensembles of hu- man tactile afferents code complex spatial fingertip events, Nat. Neurosci. 7 (2004) 170. [6] G. Pipa, A. Riehle, S. Grun, Validation of task-related excess of spike coincidences based on neuroxidence, Neurocomputing 70 (2007) 2064. [7] H. J. Cao, Miguel A. F. Sanju´an, A mechanism for elliptic-like bursting and synchronization of bursts in a map-based neuron network, Cognitive Processing 10 (2009) 23. [8] B. Ibarz, Jose M. Casado, A. F. Sanju´an, Map-based models in neuronal dynamics, Physics Reports, 501(1-2) (2011) 1. [9] S. Neuenschwander, W. Singer, Long-range synchronization of oscillatory light responses in the cat retina and lateral geniculate nucleus, Nature 379 (1996) 728. [10] M. Castelo-Branco, S. Neuenschwander, W. Singer, Synchro- nization of visual responses between the cortex, lateral genicu- late nucleus, and retina in the anesthetized cat, J. Neurosci. 18 (1998) 6395. [11] C. M. Gray, P. Konig, A. K. Engel, W. Singer, Oscillatory responses in cat visual cortex exhibit inter-columnar synchro- nization which reflects global stimulus properties, Nature 338 (1989) 334. [12] P. Fries, P. R. Roelfsema, A. K. Engel, P. Konig, W. Singer, Synchronization of oscillatory responses in visual cortex cor- relates with perception in interocular rivalry, Proc. Natl. Acad. Sci. USA 94 (1997) 12699. [13] A. L. Hodgkin, A. F. Huxley, A quantitative description of membrane current and its application to conduction and exci- tation in nerve, J. Physiol. 117 (1952) 500. [14] R. Benzi, A. Sutera, A. Vulpiani, The mechanism of stochastic resonance, J. Phys. A 14 (1981) L453. [15] C. Nicolis, G. Nicolis, Stochastic aspects of climatic transitions - additive fluctuations, Tellus 33 (1981) 225. [16] L. Gammaitoni, P. Hanggi, P. Jung, F. Marchesoni, Stochastic resonance, Rev. Mod. Phys. 70 (1998) 223. [17] H. Gang, T. Ditzinger, C. Z. Ning, H. Haken, Stochastic res- onance without external periodic force, Phys. Rev. Lett. 71 (1993) 807. [18] A. Longtin, Autonomous stochastic resonance in bursting neu- rons, Phys. Rev. E 55 (1997) 868. [19] A. S. Pikovsky, J. Kurths, Coherence resonance in a noise- driven excitable system, Phys. Rev. Lett. 78 (1997) 775. [20] P. Hanggi, Stochastic resonance in biology. how noise can en- hance detection of weak signals and help improve biological information processing, ChemPhysChem 3 (2002) 285. [21] K. Wiesenfeld, F. Moss, Stochastic resonance and the bene- fits of noise: from ice ages to crayfish and squids, Nature 373 (1995) 33. [22] M. D. McDonnell, L. M. Ward, The benefits of noise in neural systems: bridging theory and experiment, Nat. Rev. Neurosci. 12 (2011) 415. [23] F. Sagu´es, J. M. Sancho, J. Garc´ıa-Ojalvo, Spatiotemporal order out of noise, Rev. Mod. Phys. 79 (2007) 829. [24] R. Albert, A.-L. Barab´asi, Statistical mechanics of complex net- works, Rev. Mod. Phys. 74 (2002) 47. [25] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D. Hwang, Complex networks: Structure and dynamics, Phys. Rep. 424 (2006) 175. [26] O. Kwon, H.-T. Moon, Coherence resonance in small-world networks of excitable cells, Phys Lett. A 298 (2002) 319. [27] C. Zhou, J. Kurths, B. Hu, Frequency and phase locking of noise-sustained oscillations in coupled excitable systems: Array-enhanced resonances, Phys. Rev. E 67 (2003) 030101. [28] O. Carrillo, M. A. Santos, J. Garc´ıa-Ojalvo, J. M. Sancho, Spa- tial coherence resonance near pattern-forming instabilities, Eu- rophys. Lett. 65 (2007) 452. [29] M. Perc, Spatial coherence resonance in excitable media, Phys. Rev. E 72 (2005) 016207. [30] B. Lindner, J. Garc´ıa-Ojalvo, A. Neiman, L. Schimansky-Geier, Effects of noise in excitable systems, Phys. Rep. 392 (2004) 321. [31] P. Balenzuela, J. Garc´ıa-Ojalvo, E. Manjarrez, L. Mart´ınez, C. R. Mirasso, Ghost resonance in a pool of heterogeneous neu- rons, BioSystems 89 (2007) 166. [32] A. Zaikin, J. Garc´ıa-Ojalvo, R. B´ascones, E. Ullner, J. Kurths, Doubly stochastic coherence via noise-induced symmetry in bistable neural models, Phys. Rev. Lett. 90 (2003) 030601. [33] Q. Wang, M. Perc, Z. Duan, G. Chen, Delay-induced multiple stochastic resonances on scale-free neuronal networks, Chaos 19 (2009) 023112. [34] C. Gan, M. Perc, Q. Wang, Delay-aided stochastic multireso- nances on scale-free fitzhugh-nagumo neuronal networks, Chin. Phys. B 19 (2010) 040508. [35] Zeng, C. H., Gong, A. L., Zeng, C. P., Nie, L. R., Stochas- tic multi-resonance in an overdamped bistable system with two types of modulation signal, Eur. Phys. J. D 62 (2011) 219. [36] Y. Horikawa, Coherence resonance with multiple peaks in a coupled fitzhugh-nagumo model, Phys. Rev. E 64 (2001) 031905. [37] T. Kreuz, S. Luccioli, A. Torcini, Double coherence resonance in neuron models driven by discrete correlated noise, Phys. Rev. Lett. 97 (2006) 238101. [38] W. Bao-Hua, L. Qi-Shao, L. Shu-Juan, L. Xiu-Feng, Spatiotem- poral multiple coherence resonances and calcium waves in a coupled hepatocyte system, Chinese Physics B 18 (2009) 872. [39] X. Lin, Y. Gong, L. Wang, Multiple coherence resonance in- duced by time-periodic coupling in stochastic Hodgkin-Huxley neuronal networks, Chaos 21 (2011) 043109. [40] E. R. Kandel, J. H. Schwartz, T. M. Jessell, Principles of Neural Science, Elsevier, Amsterdam, 1991. [41] F. M. Atay, Distributed delays facilitate amplitude death of cou- pled oscillators, Phys. Rev. Lett. 91 (2003) 094101. [42] G. C. Sethia, J. Kurths, A. Sen, Coherence resonance in an ex- citable system with time delay, Phys. Lett. A 364 (2007) 227. [43] G. C. Sethia, A. Sen, F. M. Atay, Clustered chimera states in delay-coupled oscillator systems, Phys. Rev. Lett. 100 (2008) 144102. [44] E. Rossoni, Y. Chen, M. Ding, J. Feng, Stability of synchronous oscillations in a system of hodgkin-huxley neurons with de- 7 layed diffusive and pulsed coupling, Phys. Rev. E 71 (2005) 061904. [45] Q. Wang, Z. Duan, M. Perc, G. Chen, Synchronization tran- sitions on small-world neuronal networks: Effects of informa- tion transmission delay and rewiring probability, EPL 83 (2008) 50008. [46] Q. Wang, M. Perc, Z. Duan, G. Chen, Synchronization transi- tions on scale-free neuronal networks due to finite information transmission delays, Phys. Rev. E 80 (2009) 026206. [47] A. Roxin, N. Brunel, D. Hansel, Role of delays in shaping spa- tiotemporal dynamics of neuronal activity in large networks, Phys. Rev. Lett. 94 (2005) 238103. [48] Y. Gong, Y. Hao, X. Lin, L. Wang, X. Ma, Influence of time delay and channel blocking on multiple coherence resonance in Hodgkin-Huxley neuron networks, BioSystems 106 (2011) 76. [49] Y. Hao, Y. Gong, X. Lin, Multiple resonances with time de- lays and enhancement by non-gaussian noise in newman-watts networks of Hodgkin-Huxley neurons, NeuroComputing 74 (2011) 1748. [50] P. Balenzuela, J. Garc´ıa-Ojalvo, Role of chemical synapses in coupled neurons with noise, Phys. Rev. E 72 (2005) 021901. [51] N. Brunel, Dynamics of sparsely connected networks of ex- citatory and inhibitory spiking neurons, J. Comp. Neurosci. 8 (2000) 183. [52] S. Lee, A. Neiman, S. Kim, Coherence resonance in a Hodgkin- Huxley neuron, Phys. Rev. E 57 (1998) 3292. [53] Q. Wang, X. Shi, G. Chen, Delay-induced synchronization tra- sition in small-world Hodgkin-Huxley neuronal networks with channel blocking, Discrete and Continous Dynamical Systems Series B 16 (2011) 607. [54] C. Koch, Biophysics of Computation: Information Processing in Single Neurons, Oxford University Press, Oxford, 1999. [55] Q. Wang, G. Chen, M. Perc, Synchronous bursts on scale-free neuronal networks with attractive and repulsive coupling, PLoS ONE 6 (2011) e15851. [56] V. Volman, I. Baruchi, E. Ben-Jacob, Manifestation of function- follow-form in cultured neuronal networks, Physical Biology 2 (2005) 98. [57] C. Zhou, L. Zemanov´a, G. Zamora, C. C. Hilgetag, J. Kurths, Hierarchical organization unveiled by functional connectivity in complex brain networks, Phys. Rev. Lett. 97 (2006) 238103. [58] J. Gao, W. Tung, Y. Cao, J. Hu, Y. Qi, Power-law sensitivity to initial conditions in a time series with applications to epileptic seizure detection, Physica A 353 (2005) 613. [59] Y.-C. Lai, M. G. Frei, I. Osorio, L. Huang, Characterization of synchrony with applications to epileptic brain signals, Phys. Rev. Lett. 98 (2007) 108102. [60] V. Volman, M. Perc, M. Bazhenov, Gap junctions and epilep- tic seizures two sides of the same coin?, PLoS ONE 6 (2011) e20572.
1812.03455
1
1812
2018-12-09T10:26:18
Macroscopic phase resetting-curves determine oscillatory coherence and signal transfer in inter-coupled neural circuits
[ "q-bio.NC" ]
Macroscopic oscillations of different brain regions show multiple phase relationships that are persistent across time and have been implicated routing information. Various cellular level mechanisms influence the network dynamics and structure the macroscopic firing patterns. Key question is to identify the biophysical neuronal and synaptic properties that permit such motifs to arise and how the different coherence states determine the communication between the neural circuits. We analyse the emergence of phase locking within bidirectionally delayed-coupled spiking circuits showing global gamma band oscillations. We consider both the interneuronal (ING) and the pyramidal-interneuronal (PING) gamma rhythms and the inter coupling targeting the pyramidal or the inhibitory interneurons. Using a mean-field approach together with an exact reduction method, we break down each spiking network into a low dimensional nonlinear system and derive the macroscopic phase resetting-curves (mPRCs) that determine how the phase of the global oscillation responds to incoming perturbations. Depending on the type of gamma oscillation, we show that incoming excitatory inputs can either only speed up the oscillation (phase advance; type I PRC) or induce both an advance and a delay the macroscopic oscillation (phase delay; type II PRC). From there we determine the structure of macroscopic coherence states (phase locking) of two weakly synaptically-coupled networks. To do so we derive a phase equation for the coupled system which links the synaptic mechanisms to the coherence state of the system. We show that the transmission delay is a necessary condition for symmetry breaking, i.e. a non-symmetric phase lag between the macroscopic oscillations, potentially giving an explanation to the experimentally observed variety of gamma phase-locking modes.
q-bio.NC
q-bio
Macroscopic phase resetting-curves determine oscillatory coherence and signal transfer in inter-coupled neural circuits. Gr´egory Dumont∗, Boris Gutkin † Abstract Macroscopic oscillations of different brain regions show multiple phase relationships that are persistent across time and have been implicated routing information. Various cellular level mechanisms influence the network dynamics and structure the macroscopic firing patterns. Key question is to identify the biophysical neuronal and synaptic properties that permit such motifs to arise and how the different coherence states determine the communication between the neural circuits. We analyse the emergence of phase locking within bidirectionally delayed-coupled spiking circuits showing global gamma band oscillations. We consider both the interneuronal (ING) and the pyramidal-interneuronal (PING) gamma rhythms and the inter coupling targeting the pyramidal or the inhibitory interneurons. Using a mean-field approach together with an exact reduction method, we break down each spiking network into a low dimensional nonlinear system and derive the macroscopic phase resetting-curves (mPRCs) that determine how the phase of the global oscillation responds to incoming perturbations. Depending on the type of gamma oscillation, we show that incoming excitatory inputs can either only speed up the oscillation (phase advance; type I PRC) or induce both an advance and a delay the macroscopic oscillation (phase delay; type II PRC). From there we determine the structure of macroscopic coherence states (phase locking) of two weakly synaptically-coupled networks. To do so we derive a phase equation for the coupled system which links the synaptic mechanisms to the coherence ∗ ´Ecole Normale Sup´erieure, Group for Neural Theory, Paris, France, email: [email protected] † ´Ecole Normale Sup´erieure, Group for Neural Theory, Paris, France, email: [email protected] 1 state of the system. We show that the transmission delay is a necessary condition for symmetry breaking, i.e. a non-symmetric phase lag between the macroscopic oscillations, potentially giving an explanation to the experimentally observed variety of gamma phase-locking modes. Our analysis further shows that symmetry- broken coherence states can lead to a preferred direction of signal transfer between the oscillatory networks and how this depends on the timing of the signal. Hence we propose a theory for oscillatory modulation of functional connectivity between cortical circuits. Author summary Large scale brain oscillations emerge from synaptic interactions within neuronal circuits. Over the past years, such macroscopic rhythms have been suggested to play a crucial role in routing the flow of information across cortical regions, resulting in a functional connectome. The underlying mechanism are cortical oscillations that bind together following a well-known motif called phase locking. While there is significant experimental support for multiple phase-locking modes int eh brain, it is still unclear what is the underlying mechanism that permits macroscopic rhythms to phase lock. In the present paper we take up with this issue, and to show that, one can formulate emergent macroscopic phase-locking within the mathematical framework of weakly coupled oscillators. We thus offer clarification on the synaptic and circuit properties responsible for the emergence of multiple phase locking patterns and provide support for its functional implication in information transfer. Introduction Ranging from infraslow to ultrafast, brain rhythms is a nearly omni-present phenomenon covering more than four orders of magnitude in frequency. Of this variety of rhythms, the class of gamma oscillations falling in the specific frequency band of 30 − 150 Hz - is arguably the most studied rhythmic activity pattern [1, 2]. It has been reported in many brain regions, across many species, and is associated with a variety of cognitive tasks [3, 4]. There is nowadays growing evidence that the gamma cycle results from emergent dynamics of cortical networks, a natural consequence of the interplay between interconnected pyramidal cells and subnetworks of interneurons [5, 6]. Although brain rhythms such as gamma rhythms emerge locally [6], they are known to interact in a coherent fashion across the cortical scale [7, 8]. As such, macroscopic oscillations within different brain regions show multiplt phase relationships that are persistent across time [9]. Crucial for a prominent theory 2 of how oscillations shape the information transfer within and across the cortex, the communication through coherence (CTC) hypothesis, such cross-coupling is believed to be implicated in a number of higher cognitive functions. For example, enhanced interareal gamma-band coherence is considered as the neural correlate of selective attention, in which a network receiving several informational stimulus can preferentially react to one or another depending on task relevance [7]. The CTC hypothesis provides a mechanism by which gamma rhythms can participate in regulating the information flow [2]. The rationale behind is that gamma oscillations are the consequence of rhythmic inhibitory feedback inducing an hyperpolarization of the principle cell membrane potential [5, 6]. Synaptic inputs targeting excitatory cells are then expected to cause a stronger reaction when the inhibition drops off. This gives rise to a temporal window of excitability within the oscillatory cycle during which pyramidal neurons are more likely to respond to stimulation [10]. Ongoing oscillatory firing patterns rhythmically modulate the excitability of networks, and therefore, two neural groups engaged in a rhythmic dynamic communicate more efficiently when they maintain a coherent relationship: they can consecutively send their information at the most excitable phase [4, 7]. According to the CTC hypothesis, neuronal interactions and transfer of information are dynamically shaped by the phase relationship between neuronal oscillations [11]. In fact it has been proposed that macroscopic rhythms offer a way of adjusting the effectivity of functional connectivity while leaving un- touched the anatomical connections [9] and resulting in a functional connectivity [12, 13]. This functional connectivity, often defined in correlational or information transmission terms, is determined by the relative phase relationship between the communicating networks. Note that an optimal locking mode does not always result from a zero phase lag or perfect synchrony. The reason is that, spike transmission from one network to another is not instantaneous and, depending on the distance, projection across the brain can take up to hundreds of milliseconds [14]. Therefore oscillations should be lagged in order to see their spikes arriving at the most excitable phase. This most excitable phase also depends on the biophysical properties of the constituent neurons and of the emergent rhythms (e.g. as characterised by the network-wide phase response curves [15]. An optimal phase difference will thus depend on the properties of the neural groups at work and the distance between the two [16, 17]. Recent experimental studies have reported a multiplicity a phase differences and it has been argued that such a diversity might facilitate information selectivity [17]. Over the past few years, computational studies have devoted a great deal of attention to uncovering the precise functional roles of gamma patterns and gamma interaction. Doing so, they have been able to reproduce experimental findings in support of several predictions of the CTC hypothesis. For instance, 3 modeling approaches have shown that the gamma cycle generates a temporal window of excitability [18], which is suitable to suppress irrelevant stimuli [19, 20]. Others studies have demonstrated that the mutual information between two neural groups engaged in rhythmic patterns is tuned with respect to their phase lag [21, 22], and a directionality in the flow of information emerges through a symmetry breaking in the phase relationship [12, 13]. A diversity of phase lags can then be observed which benefits information coding and stimulus reconstruction [23]. Finally, in a rather different line of thinking from the main current view of CTC, computational studies have exposed how cortical oscillations could implement a multiplexing [24 -- 26]. However, the underlying mechanisms responsible for the emergence of the multiple phase-locking modes and of the ensuing functional connectivity as proposed by the CTC are not trivial. So far, no mechanistic view to explain the observed variety of phase lags has been proposed. The question is then to identify through what synaptic mechanisms can these rhythms coordinate their temporal relationships in such a diversity of locking modes. Answering this question is crucial and knowing the chain of causation that allows for coherent oscillations is key to understanding their functional role [27, 28]. Hence, an subsequent question is how can one characterize the functional connectivity associated with the various phase locking modes and how directed signal transmission can ensue. We investigate the dynamical emergence of phase locking within two bidirectionally delayed-coupled spiking networks. Importantly, the neurons within the circuits have a relatively wide distribution of intrinsic excitability, meaning that most of them not intrinsically oscillating. Hence the gamma rhythm in our network is an emergent property of the global dynamics, as opposed to phasing of coupled oscillators (see ?? for instance). Furthermore, the design of the interconnections between our networks is inspired from previous research [13,21,22] to essentially capture multiple communicating brain regions where transfer of information takes place. Each network is assumed to be made up of pyramidal cells and interneurons, and each cell is characterized by a conductance-based neural model [29, 30]. A synaptic delay is included to account for possible long range distances separating the circuits [14]. We then take advantage of a thermodynamic approach combined with a reduction theory to simplify each network description - see [31 -- 33] - and to express the macroscopic phase resetting curve (mPRC) of their oscillatory cycle [15, 34, 35]. The network mPRC is an important causal measure which allows us to use the weakly coupled oscillator theory [36, 37] to characterise the inter-network dynamics. The fundamental assumption at the core of this theoretical setting is that synaptic projections from one circuit to another must be sufficiently weak. Please note that the weak coupling condition is not on the synaptic connections within each of the circuits, but only across them. The weak coupling condition allows one to take advantage of a variety of mathematical 4 techniques and to abbreviate the bidirectionally delayed-coupled spiking circuits description to a single phase equation [38, 39]. This simplification significantly reduces the complexity of the interacting macroscopic oscillations, making them mathematically tractable, while at the same time capturing crucial principles of phase locking. As we show below, an analysis of the phase equation sheds lights on the synaptic mechanism enabling circuits with emergent global oscillations to bind together. We give particular attention to the central role played by the conduction delay in producing symmetry-boken states of activity (with purely symmetric connectivity) , i.e to permit the emergence of a variety of non-symmetric phase lags. Such a collection of phase lags has been suggested to facilitate the control and selection of the information flow through anatomical pathways [17], and conduction delay has been at the core of recent discussion regarding the CTC hypothesis [40]. Our final goal is then to show that non-symmetric lags lead to a directed functional coupling between the networks. We indeed show that symmetry-broken states induce with a preferred direction of signal transfer between the networks, and therefore provide theoretical support for the role of oscillations in modulating functional connectivity between cortical circuits [12, 13]. The paper is structured as follows. First, we present the network and neural model which will be used throughout. We expose the low dimensional system for which we can perform a bifurcation analysis and extract the infinitesimal PRC. From there, we compute the so-called interaction function and reduce the bidirectionally delayed-coupled spiking networks to an unique phase equation. The analysis of the phase equation enables us to make several prediction on the locking states between the emerging oscillations. We support our theoretical findings with extensive numerical illustrations and discuss our results in light of the CTC hypothesis and functional connectivity. Finally, the mathematical techniques are exposed in a detailed Methods section at the end of the paper. Results The Network and its Reduced Description Our generic cortical circuit is assumed to be made up of Ne excitatory cells (E-cells) and Ni inhibitory cells (I-cells) coupled in an all-to-all fashion. Each cell is described by a well-established conductance-based model - the quadratic integrate-and-fire (QIF), see [41] - which is known to capture the essential dynamical features of the neural voltage [29]. The onset of an action potential is taken into account by a discontinuous reset mechanism. Whenever a cut off value vth is reached, the voltage is instantaneously set to vr, a reset 5 parameter. To permit analytical computations, threshold vth and reset vr are respectively taken at plus and minus infinity [29]. The QIF reads τ d dt vj = ηj + v2 j + I, (1) where v(t) is the neural voltage, j the neuron number, τ the membrane time constant, η the bias current that defines the intrinsic resting potential and firing threshold of the cell and finally I(t) the total synaptic current injected at the soma. To account for the network heterogeneity, the intrinsic parameter η is distributed randomly according to a Lorentzian distribution: L(η) = 1 π ∆ (η − ¯η)2 + ∆2 . Here ¯η stands for the mean value taken by the parameter η across the population and ∆ is the half-width of the distribution. Note that the heavy-tailed Lorentzian distribution implies a wide range of intrinsic excitability, i.e. many neurons are not intrinsically oscillating and if they do, they have different firing frequency, as opposed to the classical framework of phasing of coupled oscillators (see [42] for instance). Indeed, for a null current, the proportion of neurons not being intrinsic oscillator is given by (cid:90) 0 L(η) dη = 1 π −∞ − arctan (cid:16) π 2 (cid:16) ¯η (cid:17)(cid:17) ∆ , which can not be zero as soon as there is heterogeneity within the network. Note nonetheless that the proportion will be affected by the synaptic current I. The total synaptic current, I(t) is assumed to be the sum of an external input I ext(t) that takes into account inputs coming to the cell from sub-cortical structures or nearby cortical networks through lateral connections, and the synaptic inputs se and si which models the effect of recurrent connexions within the circuit: I = I ext + τ se − τ si. The synaptic current, s(t), depends on the synapse type, for the excitatory synapse, we have τs d dt se = −se + Jere, 6 Figure 1: Comparison between the full network and the reduced system. Left panel: Schematic illustration of a canonical cortical neural network. The parameter Jαβ denotes the connectivity strength of the population β onto the population α. The external influence on the population α is denoted I ext α . Right panels: A) Time evolution of the stimulus I ext on the E-cells. B) Spiking activity obtained from simulations of the full network, the first 800 cells are excitatory, the last 200 are inhibitory. C) Firing rate of the E-cells obtained from simulations of the full network (red line) compared with the reduced system (black line). D) Firing rate of the I-cells obtained from simulations of the full network (blue line) compared with the reduced system (black line). Parameters: Ne = Ni = 5000; ∆e = ∆i = 1; τe = τi = 10; τse = τsi = 1; ¯ηe = ¯ηi = −5; Jee = 0; Jei = 15; Jii = 10; Jie = 15; I ext i = 0; vth = 500; vr = −500. e respectively for the inhibitory synapse, τs d dt si = −si + Jiri. Here, τs the synaptic time constant, J the synaptic strength - see Fig. 1 - and r(t) the population firing rate. For the E-cells, we have: and for the I-cells, we have: Ne(cid:88) (cid:88) k=1 f Ni(cid:88) (cid:88) k=1 f δ(t − tk f ), δ(t − tk f ), re(t) = 1 Ne ri(t) = 1 Ni where δ is the Dirac mass measure and tk f are the firing time of the neuron numbered k. To get a clear picture of how the synaptic structure shapes the firing patterns, we take advantage of a thermodynamic approach combined with a reduction method. The thermodynamic framework produces a single average system written in term of partial differential equations that is valid in the limit of an infinitely 7 re = τe τe τs τs d dt d dt d dt d dt   + 2reVe ∆e πτe e + ¯ηe + Ie − τ 2 e π2r2 e, Ve = V 2 see = −see + Jeere, sei = −sei + Jeiri, τi τi τs τs d dt d dt d dt d dt ri = + 2riVi ∆i πτi i + ¯ηi + Ii − τ 2 Vi = V 2 sie = −sie + Jiere, sii = −sii + Jiiri, i π2r2 i . (2) (3) large number of neurons [43]. The reduction method allows further simplification and breaks down the mean-field system into a small set of differential equations [32, 33]. In our case, see Method for more details about the derivation, the low dimensional dynamical system reads: and for the I-cells: In here, V (t) represents the mean voltage of the population, while r(t) still stands for the firing activity. Note that the two systems are coupled via the expression of the total current arriving on each sub-population: and Ie = I ext e + τesee − τesei, Ii = I ext i + τisie − τisii. The numerical simulation presented in Fig. 1 compares the dynamics of the full network with the low dimensional system (2)-(3). It shows the time evolution of the external stimulus in the first panel (Fig. 1A), whereas the second panel gives the spiking activity obtained from a simulation of the full network (Fig. 1B). In the subsequent panels (Fig. 1C-D), the firing rate given by the reduced description is compared with the firing rate obtained from network simulations. The perfect agreement between the population activities convinced us that the reduced dynamical system captures the fundamental aspects of the population firing rate. Of course, such a reduced description provides an efficient way to carry out a study of the circuit since it can be simulated very quickly and it is amenable to mathematical analysis. 8 Figure 2: The PING interaction. Left panel: Schematic illustration of the PING (Pyramidal Interneuron Network Gamma) interaction. The parameters have similar properties than Fig. 1. Right panels: Nonlinear analysis of the The PING interaction. A-B) Bifurcation diagrams. The blue line, (respectively the red line), corresponds to the steady state of the inhibitory cells, (respectively the excitatory cells) while dots correspond to limit cycles. C) Stability region. D) The stimulus and corresponding raster plot of the spiking activity. E) Comparison between simulated and calculated mPRCs. The black line illustrates the analytical adjoint method while dots indicates direct perturbations of the full network. Red dots, perturbations are made on the E-cells, second row, with the blue dots, perturbations are made on the I-cells. F) PRC shape as a function of parameters obtained via the adjoint method for different values of the external current I ext e = 9, 10, 11. Parameters are as in Fig. 1, except Jee = 0; Jei = 15; Jii = 0; Jie = 15 and for the panel B) I e ext = 10. Direct perturbations in panels E) are made with a square wave current pulse (amplitude 10, duration 0.5). Emerging Rhythms and Phase-Resetting Curve To understand how the emergent network gamma oscillations can phase lock, it is essential to first consider their basic underlying mechanisms. To gain insights, a nonlinear analysis of the reduced system is performed. This enables us to reveal how the inhibitory feedback loop renders possible the emergence of macroscopic rhythms. Two processes can be described: PING and ING [6]. In the PING (Pyramidal Interneuron Network Gamma) interaction, see Fig. 2, the underlying synaptic machinery involves an interplay between the pyramidal cells and the fast-spiking cells. For a chosen set of connectivity parameters, the dynamical system exhibits a Hopf bifurcation (Fig. 2A), such that, enhancing 9 the external stimulus upon the pyramidal cells induces a graded progression toward an oscillatory regime. Note that this rhythmic regime disappears as the network heterogeneity is expanded (see Fig. 2B-C). The rhythmic transition is illustrated with a simulation displayed in Fig. 2D. A self-sustained oscillatory regime emerges as soon as the E-drive is strong enough. Of course, the presence of a Hopf bifurcation in the system should be put in relation with the seminal work of Wilson and Cowan [44]. In the ING (Interneuron Network Gamma) interaction, see Fig. 3, the mechanism requires an inhibitory feedback from fast-spiking cells onto themselves and the rhythm arises from this interconnected inhibitory network which in turn defines the excitatory spike times. The nonlinear analysis reveals a Hopf bifurcation as the external drive is raised (see Fig. 3A). Again, this rhythmic regime disappears with too much hetero- geneity (see Fig. 3B-C). The network activity undergoes a transition from an asynchronous regime toward an oscillatory which is displayed in Fig. 3D. Interestingly, the ING behavior can not emerge within the traditional rate equation proposed by Wilson and Cowan [44], see [45] for a more complete discussion. Note finally the frequency difference between the PING and the ING rhythm. The two interaction models are then seen as canonical descriptions of the low and fast gamma oscillations, PING for low gamma range and ING for fast gamma spectrum. Over the past decades, the Phase Resetting Curve (PRC) has become one of the fundamental concepts in theoretical neuroscience. Its usefulness has been reviewed in multiple papers [36 -- 38,46] and its outcomes are expected to impact our understanding of brain rhythms [27]. PRC measures the effects of transient stimuli upon oscillatory systems and can be obtained experimentally [47 -- 50]. In our case, the application of a short depolarizing current to the network affects the spiking activity, and the macroscopic oscillation shifts in time. The induced phase shift depends on the perturbation strength but also on the phase at which the perturbation is presented. It can either delayed or advanced depending on the onset phase of the perturbation The PRC results in plotting the advance or delay with respect to the phase onset at which the perturbation is made. Doing so, it quantifies the effect of the perturbation on the macroscopic oscillation. For the cortical network under consideration, several PRCs coexist at the same time depending on where the depolarizing input is applied. In the limit of short, weak perturbations, the shift in timing can be described by the so-called infinitesi- mally PRC (iPRC). The iPRC is mathematically expressed by a linear differential system, known as the the adjoint system [51]. This method can be applied to the low dimensional system (2)-(3) and a semi-analytical expression of the iPRC be obtained. Assuming that the reduced E-I system (2)-(3) has a stable limit cycle, 10 Figure 3: The ING interaction. Left panel: Schematic illustration of the ING (Interneuron Network Gamma) interaction. The parameters have similar properties than Fig. 1. Right panels: Nonlinear analysis of the The ING interaction. A-B) Bifurcation diagrams. The blue line, (respectively the red line), corresponds to the steady state of the inhibitory cells, (respectively the excitatory cells) while dots correspond to limit cycles. C) Stability region. D) The stimulus and corresponding raster plot of the spiking activity. E) Comparison between simulated and calculated mPRCs. The black line illustrates the analytical adjoint method while dots indicates direct perturbations of the full network. Red dots, perturbations are made on the E-cells, second row, with the blue dots, perturbations are made on the I-cells. F) PRC shape as a function of parameters obtained via the adjoint method for different values of the external current I ext = 24, 25, 26. Parameters are as in Fig. 1, except Jee = 0; Jei = 10; Jii = 15; Jie = 0 and for the panel B) Iext = 25. Direct perturbations in panels E) are made with a square wave current pulse (amplitude 10, duration 0.5). we find that, see Method for more details, the iPRC Z(t) is a periodic vector that is a solution of the adjoint equation where the matrix M(t) is given by a linearization of the E-I system (2)-(3) around the limit cycle, see Method − d dt Z(t) = M(t)T · Z(t), (4) for its precise expression. When perturbations made to the network are sufficiently small, the PRC becomes proportional to the iPRC [35, 52, 53]. We present in Figs. 2E and 3E and the iPRC obtained via a simulation of the adjoint 11 system (4) compared with direct perturbations made on the spiking network. The blue line, (respectively the red line), corresponds to the iPRC of the excitatory synapse of the I-cells (respectively the E-cells). From the simulations and semi-analytical expression of the PRC we can classify the PING and ING rhythms as having different PRC types, i.e. as having different rhythmic properties. For the PING dynamics, see Fig. 2E, a biphasic shape of the PRC is observable when perturbations are made on the I-cells. In contrast, when perturbations are on the E-cells, the PRC is monophasic. This is a classification already observed in a previous work where the synaptic dynamic was neglected and considered to be instantaneous [15]. Regarding the ING pattern, see Fig. 3E, the PRC is monophasic for perturbation targeting the I-cells. The PRC is null when perturbations are made onto the pyramidal cells, which means that any perturbations will die out after a few cycle. This comes without a surprise since in the ING interaction, pyramidal cells do not play a part in the emergence of the oscillations. PRCs are thus quite different between the ING and the PING oscillations. This is because the contribution of the cell type to the rhythmic behavior is largely different in the ING and PING mechanisms. The PRC difference between the ING and the PING oscillations have been investigated in very recent work [34]. From there, we can explore the consequences of differences of locking regimes to periodic pulsatile stimuli, and their result supports that the origin of the cell-type-specific response, already experimentally observed [10], comes from the different entrainment properties [34]. Similarly, we can also question the network sensitivity to perturbation to the excitatory cells or to the inhibitory cells from the difference in amplitudes of the respective macroscopic PRCs. As we shall see in Fig. 2F and in Fig. 3F, the PRC amplitude strongly depends on parameters such as the external current. This can be intuitively interpreted as a dependence of stability of the macroscopic oscillation. Phase equation We now turn our investigation to the dynamical emergence of phase synchrony across multiple networks (as reflecting multiple brain regions). While we do not aim at studying any specific brain interaction, however, the structure that is shown in Fig. 4 reflects the architecture of many communicating cortical and sub- cortical areas where information transmission is at play [21, 22]. Each network is assumed to be made up of interacting pyramidal cells and interneurons as presented in the previous sections (see Fig. 1). Since interneurons are known to wire on local scale, the synaptic projection from one circuit to another is made via the pyramidal cells only. A delay is added to account for finite transmission speeds and synaptic time- courses across circuits. Importantly we note that the considered structural motif is symmetric, i.e. it is 12 Figure 4: The bidirectionally neural circuits. Top panel: The two coupled circuits. The parameter Gαβ denotes the connectivity strength of the population β of one network onto the population α of the other circuit. The intrinsic parameters are unchanged and similar within each network as presented in Fig. 1. stable under permutation of the two cortical networks. Note that locking across gamma oscillations appears within the same frequency range, we will thus focus our study on two interacting schemes: the PING-PING interaction and the ING-ING interaction. The two mechanistic models of gamma generation having different oscillatory regimes, the interaction PING-ING would lead to a cross-frequency coupling. First, it is far beyond the scope of this paper to investigate the coherence between slow and fast oscillations, second, we note that, under our knowledge, cross-coupling among slow and fast gamma has not been observed so far. Our whole analysis of phase locked states is based on the assumption that synaptic interactions across circuits are sufficiently weak. Such an assumption, which guarantees that the perturbed macroscopic oscilla- tion remains close to the unperturbed one, allows to place our study within the framework of weakly coupled oscillators [38, 39]. We emphasize that within each circuit, neurons are not weakly coupled. The assumption of weak coupling is only made upon the projection from one circuits to another. Within this framework, see Methods, the bidirectionally delayed-coupled neural circuits reduce to a single phase equation: d dt θ(t) = G(θ(t)), where θ(t) is the phase difference (or phase lag) between the circuits and the G-function is the odd part of the shifted interaction function (the so-called H-function) expressed via the PRC (see Method): 13 H(θ) = (cid:90) T (cid:90) T Zsee (s)re(s − θ) ds Zsie(s)re(s − θ) ds, Gee T + 0 Gie T 0 where T is the oscillation period and Gαβ denotes the connectivity strength of the population β of one network onto the population α of the other circuit, see Fig 4. Note the involvement of the synaptic component of the PRC Zs(t) and the firing rate of the E-cells re(t) all along the oscillatory cycle. The G-function is essential for our study since it conveys knowledge about the phase-locking mode between the coupled circuits. Indeed, the zeros of the G-function correspond to steady states phase lags. The stability of a locking mode is conditioned by a negative slope (G(cid:48)(θ) < 0), while a positive slope (G(cid:48)(θ) > 0) implies instability. Locking modes To disentangle the synaptic mechanisms responsible for the dynamical emergence of cross-network phase locking, we first fix the delay to zero and focus our study on the effect of coupling weights. To put it in mathematical terms, we investigate the location of the zeros of the G-function with respect to the coupling strengths when the parameter d is set to zero. As we see from Fig. 5, modification in the network parameters affects the shape of the G-function. This is made for the PING interaction. The zeros of the G-function are located at the in-phase (synchrony) and anti-phase locking (anti-synchrony) mode. The anti-phase state is nonetheless unstable. We therefore expect the in-phase synchrony mode to emerge from the dynamic of the bidirectionally coupled circuits. This is the case for a cross-coupling targeting exclusively the E-cells (Gie = 0, Fig. 5A) or the I-cells only (Gee = 0, Fig. 5B). Since in the general case, the interaction function will result in a linear superposition of the two previously mentioned possibilities, in the non-delayed coupling scenario, only a perfect zero lag synchrony can be expected, see Fig. 5C. We illustrate this prediction by showing the network rasters in Fig 5D. The black dots correspond to the first network, whereas the colored dots to the second circuit. The spiking activity of the two circuits oscillate in phase, i.e. the two raster plots are synchronized at zero lag and thus perfectly overlap and that is the reason why only black dots can be seen. Simulation and theoretical prediction are in perfect agreement. As we can see, despite its vast simplifications, the phase equation yields amazingly accurate predictions. 14 Figure 5: locking modes of PING. A) The panel gives the G-function for different parameter Gee when Gie = d = 0. B) The panel gives the G-function for different parameter Gie when Gee = d = 0. The circles are filled for stable fixed point and empty for the unstable points. C) Resulting locking mode when there is no delay. D) Raster plot of the spiking activity of the two neural networks, black dots indicate the spike timing of the first network, colored dots indicate the spike timing of the second network. E) The panel gives the G-function for d = 2, Gee = 0.1; Gie = 0.5;. F) The panel gives the G-function for d = 10; Gee = 0.1; Gie = 0.5;. The circles are filled for stable fixed point and empty for the unstable points. G) Resulting locking mode for short and large delay. H) Raster plot of the spiking activity of the two neural networks, black dots indicate the spike timing of the first network, colored dots indicate the spike timing of the second network. Parameters are as in Fig. 9. The fact that two oscillatory networks (two oscillators) synchronize at zero lag when delay is neglected was to be expected. However, in real settings, neuronal signals travel at finite speeds across the brain and a wide range of delays between neuronal populations has been reported [14]. How the presence of transmission delay reshapes the phase relationship between macroscopic oscillations has remained elusive so far. This is a central issue since recent studies have proposed an updated formulation of the CTC hypothesis where delay between communicating sites plays a critical role [40]. To put it into a mathematical perspective, we expect that distinct delays lead to different fixed-points in 15 the G-function, and to illustrate this expectation, we plot the G-function obtained for two different delays (Fig. 5E-F). As we can see, the stability of the locking modes are reversed, and the anti-phase mode, which was unstable, becomes stable. In contrast, the in-phase mode turns into an unstable state. Two possible phase locking modes are then possible, the in phase mode for short delay and the anti-phase mode for large delay (Fig. 5E-F). We illustrate this prediction by showing the network rasters in Fig 5H. As we can see, for large delay value, the spiking activity of the two circuits oscillate in an out of phase mode. We push further the analysis by investigating the transition between the two phase locking modes. In Fig. 6A we plot the G-function obtained several delays, black lines correspond to small delays while grey lines to bigger ones. A continuous deformation is seen, leading the G-function to slip over the phase-axis. To get a better visualization, we plot a bifurcation diagram (Fig. 6B) which gives access to the phase mode with respect to parameter change. Such a diagram is very helpful to anticipate the locking states in the bidirectionally delayed-coupled networks. We note that the stability of the in-phase mode is kept for small delays. On the other hand, for larger transmission time, a switch of stability between the in-phase and anti-phase locking modes is observed. Importantly, a wide region of delays for which the phase lag goes over all the possibilities appears in the diagram. This result confirms the role of delay in the emergence of a complete variety of phase shifts across gamma interaction in the cortex [17, 54]. In Fig. 6D we validate this theoretical prediction by showing rasters of the spiking circuits that reflects the modulation of the emerging phase lag by the delay. As we see from Fig 6D, the spiking activity of the two networks oscillate with a small phase lag. Increasing slightly the delay leads the spiking activity of the two networks to oscillate with a bigger phase lag. Simulation and theoretical prediction are again in perfect agreement. This result shows that it is normal to observe persistent phase relationship across time that are so diverse across space. As already noticed in [55], this case corresponds to a spontaneous symmetry breaking. We talk about symmetry breaking because those variety of phase lag states do not share the symmetric feature with the full system. Note that when the delay is kept fixed, and sufficiently large, a variation of the synaptic strength onto the E-cells in Fig. 6F leads to a transition from the in-phase state to the out-of-phase locking. As a part of this transition, a variety of stable phase lags appear. A reverse situation is depicted in Fig. 6G: when the coupling onto the I-cells is varied the in-phase mode transitions to an anti-phase mode. As we can see from these diagrams, we can tune the phase shifts across brain oscillations at least for the PING rhythm. Indeed, this only possible for the PING rhythm since in that case two non-zero PRCs exist at the same time. A similar situation emerges for the ING interaction. In Fig. 7, we show the interaction function and 16 Figure 6: Bifurcation analysis of locking modes with respect to delay for the PING. A) The panel gives the G-function for different parameter d. B) Bifurcation analysis. The circles are filled for stable fixed point and empty for the unstable points, red dots represent the illustrations parameters C) Resulting locking modes. D) Raster plot of the spiking activity of the two neural networks, black dots indicate the spike timing of the first network, colored dots indicate the spike timing of the second network. Parameters are as in Fig. 9 with Gee = 0.1; Gie = 0.5; A) d = 6, B) d = 7. corresponding locking modes. While short delays induces only an in-phase locking mode Fig. 7C-F, larger delays will reverse the interaction function and induce an out of phase locking scheme Fig. 7E-F. Once again, notice the spontaneous symmetry breaking implying the existence of a variety of phase lags for moderate values of the delay Fig. 7G-H-I. Not that for the ING-ING interaction, modification of the synaptic coupling Gαβ will not affect the locking modes since the coupling is through the pyramidal neurone and these do not affect the macroscopic oscillatory phase. 17 246810Delay (ms)0/23/22Phase E0.050.10.150.20.250.3Gee (mV)0/23/22Phase F0.20.30.40.50.6Gie (mV)0/23/22Phase G Figure 7: Locking modes of the ING. Effect of coupling strength on locking modes without delay. A) The panel gives the G-function for different parameter Gee when Gie = d = 0. B) The panel gives the G-function for different parameter Gie when Gee = d = 0. The circles are filled for stable fixed point and empty for the unstable points. C) Resulting locking mode when there is no delay. D) The panel gives the G-function for d = 0.1. E) The panel gives the G-function for d = 0.6. The circles are filled for stable fixed point and empty for the unstable points. F) Resulting locking mode for short and large delay. G) The panel gives the G-function for different parameter d. H) Bifurcation analysis. The circles are filled for stable fixed point and empty for the unstable points, red dots represent the illustrations parameters I) Resulting locking modes. Parameters are as in Fig. 3 with Gee = 0.; Gie = 0.3; A) d = 0.2, B) d = 1. Emerging causal directionality We now turn to the functional role that could be supported by a breaking of symmetry between the two structurally symmetrically connected brain circuits, each with a emergent oscillatory activity. Recent studies have associated spontaneous symmetry breaking with an unidirectional effective transfert of information despite a symmetric structural connectivity motif [12, 13]. We therefore expected to characterize a similar phenomenon but with a direct causal measure . To prove that there is indeed a causal directionality of signaling, we compute a global PRC of the full delay system. Our intention is to measure how an impact on one of the two networks affects the other circuit and the system as a whole. In the symmetry broken state, there is a leader circuit and a follower circuit. We posited that should we find that the global PRCs are identical for perturbations to either the leader or the follower, signal transfer is symmetric, should the two PRCs differ, signal transfer is directed. While we seek a fully analytical approach, the difficulty with the presence of the delay is that it hard to offer an analytical treatment. An analytical method for such a global PRC is not easy to implement and its convergence is not guarantee. We therefore follow a semi-analytical approach, where we use the direct method to compute the global PRC for the reduced model, which makes 18 the computations efficient (see Method Eqs. (9)-(12)). We this perturb one network (the leader or the follower) and observe the resulting asymptotical phase shift of the second network. Fig. 8 illustrates the perturbation effects. As expected, when the two networks are in phase, perturbing one or the other has similar outcomes. When the two networks are out of phase, the resulting global PRCs are only shifted with respect to one another. This is a natural consequence of the symmetry in the oscillatory modes of the macroscopic oscillations. The most interesting scenario is when the resulting phase locking mode is not symmetric. In this situation, perturbing the leader or the follower does not give the same phase shift. As we can see, PRCs are almost reverse, i.e. while a perturbation of the leader induces a phase advance, a perturbation on the follower implies a phase delay. Importantly, the amplitude of the PRCs have also different order of magnitude. Perturbation of the leader have stronger impact than on the follower. Furthermore, we see that for each of the perturbations, phase shifts depend on the phase at which the external "signal" arrives: e.g. in cases where an excitatory perturbation on either networks advances the oscillations, and phases where perturbing the leader advances the phase, while exciting the follower delays the oscillation. As a note, it has been previously shown that the post-stimulus spike-time histogram (PRC) can be directly related to the PRC [56, 57]. Hence, the asymmetric PRCs for the leader and the follower predict non-symmetric PSTHs, once again giving a direct and causal measure of how broken-symmetry states can induce a directional functional connectivity despite complete structural symmetry. In summary, we can interpret our results giving a causal directionality in the communication between the two circuits which has been recently showed using using correlative statistical measure such as transfer entropy [12, 13]. Discussion The omnipresence of oscillations in the brain gives significant support to the hypothesis that rhythmic firing patterns are well suited to specific cognitive functions [1, 2]. In particular, recent physiological experiments have indeed acknowledged that coherent gamma rhythms play a determinant part in the transfer of informa- tion across cortical areas [7,9,40]. As this communication depends on stable phase-relationships between the oscillatory cortical networks, a key question has been to determine the conditions under which two oscillatory brain circuits phase lock, what is the resulting phase lag between them and how the phase lag relates with delay and synaptic couplings [28]. Here, we have outlined and tested a new analytical approach to deal with the dynamical rise of phase 19 Figure 8: Global PRC and emerging directionality. Black dots illustrate direct perturbations of the leader, color dots indicate direct perturbations of the follower. A) Perturbations are made on the E-cells. B) Perturbations are made on the I-cells. The network parameters are the same as in Fig. 2 with I ext e = 10, direct perturbations are made with a square wave current pulse (amplitude 1, duration 0.2), the first panels d = 6, the second panel d = 7, the third panel d = 10. synchrony between multiple spiking neural circuits. Making use of a mixture of mathematical techniques - mean-field theory, reduction methods, PRC measures and the framework of weakly coupled oscillators - we have been able to reduce the complexity of the problem to a single phase equation. The obtained dynamical equation reflects the contribution of cortical structure to the coordination of the macroscopic firing patterns: notably the role of the effective distance between the circuits, the native of the emergent local rhythms and the structure of the synaptic coupling across the circuits. We have shown that this level of abstraction suffices to qualitatively reproduce and explain experimentally observed oscillatory patterns. For instance, our synaptic theory allows us to clarify on the observed diversity of phase lags between multiple cortical gamma rhythms that have been proposed to play a crucial role in controlling and selecting information through anatomical pathways [17]. Furthermore, our technique allows us to determine the directionality of cause signal transfer between multiple interacting neural circuits with emergent gamma oscillations. Using the PRC technique, we first confirmed that the signal transfer is undirected in dynamical states with full symmetry: the global PRCs were identical or just phase shifted for in-phase and anti-phase synchrony. For dynamical symmetry-broken states, where the circuits separate into a leader and a follower (also sometime called stuttering states), the global PRCs depend qualitatively on where the signal originates (e.g. in the leader) and where it propagates (e.g. to the follower). Our results show that depending on this and on the timing of the external signal perturbations, 20 0/23/22Phase -0.500.5Global E-PRC0/23/22Phase -1-0.500.51Global E-PRC0/23/22Phase -0.200.2Global E-PRCA0/23/22Phase -0.500.5Global I-PRC0/23/22Phase -0.200.2Global I-PRC0/23/22Phase -0.100.1Global I-PRCB the neural activity can be either advanced or delayed. Once again, this causal functional directionality in the communication between neural circuits appears as a consequence of the system dynamics and despite a completely mirror symmetric structural connectivity and the individual network properties. We believe that these results give a causal basis for the recent statistical directed functional connectivity measures. In the end, the series of mathematical arguments leads to a simple visualization technique - a bifurcation diagram - which compiles all the relevant information about circuit phase relationships when parameters are changed. Such graphical representation demonstrate that, in multiple delayed-coupled spiking networks, phase locking of the emergent macroscopic oscillatory rhythms are natural features that can be controlled. Our synaptic theory sheds new light on the long range cortical circuit interactions, and importantly, it offers a way to make strong predictions that can be tested against experimental data. For instance, one can compare the phase locking modes generated by different brain areas with distinct synaptic organization of the model. The formalism employed within the paper requires pyramidal neurons to work in a regime where projec- tions across circuits are weak. Within this parameter regime, the presented sequence of theoretical arguments are fully valid. How our results extend to the strongly coupled regime remains a challenging topic for future studies. Although we have restricted our study to considering networks with homogenous synaptic weights and current-based synaptic interaction, the mathematical strategy that served throughout this paper is adjustable and easily accepts the inclusion of conductance-based synaptic description with a certain level of synaptic heterogeneity [33, 58]. Similarly the accommodation of delay within the circuits themselves would not bring difficulty, neither for the reduction method [59], nor for the PRC computation [52]. This could be an interesting subject of research for future works as well as the study of locking to an external periodic modulation for which the PRC offers several path of investigation [34, 60]. All along the paper, we have only been interested to study the locking of oscillations having identical properties, however, several studies have reported coupling across different frequency band of oscillations [54]. Termed as cross-frequency coupling, the locking of brain regions with different frequencies is an open subject of research. A promising extension would then be to generalize the phase locking analysis to a network of subsequent layers made up of diversified interneuron types [61 -- 63]. Such a consideration would clarify the specific roles of each layer and cell types in the generation of locking and elucidate the underlying synaptic mechanism and functional roles of cross-frequency coupling observed in slow-fast oscillations [54]. To that end, one would need to study the case of interacting circuits with different intrinsic frequency which remains 21 for us an open issue to be investigated. Method Mean-Field We consider an all-to-all coupled network made up of N spiking cells characterized by the quadratic integrate- and-fire (QIF) model: τ d dt vj(t) = ηj + v2 j (t) + I(t), where v(t) represents the time evolution of the membrane potential, τ is the membrane time constant, I(t) is the total current, and we assume the intrinsic parameter η being randomly distributed across the network according to a Lorentzian distribution: L(η) = 1 π ∆ (η − ¯η)2 + ∆2 . with ¯η the mean value and ∆ the half-width of the distribution. The onset of an action potential is taken into account by a discontinuous mechanism with a threshold vth and a reset parameter vr respectively set at plus and minus infinity [29]. The population firing rate is then given by the sum of all the spikes: N(cid:88) (cid:88) δ(t − tk f ) r(t) = 1 N k=1 f where δ is the Dirac mass measure and tk f are the firing times of the neuron numbered k. In the mean-field limit, that is, when the number of cells is taken infinitely large, see [43] for instance, the system is well represented by the probability of finding the membrane potential of any randomly chosen cell at potential v at time t knowing the value η of its intrinsic parameter. The dynamic of this density, which we denote p(t, vη), is given by a continuous transport equation written in the form of a conservation law: p(t, vη) + τ ∂ ∂t ∂ ∂v J(t, vη) = 0, (5) where the total probability flux is defined as J (t, vη) = (η + v2 + I(t))p(t, vη). 22 A boundary condition, consistent with the reset mechanism of the QIF model, is imposed: v→−∞J (t, vη) = lim v→+∞J (t, vη). lim One can check easily the conservation property of the equation: (cid:90) +∞ −∞ p(t, vη) dv = L(η). Importantly, the firing rate of the population r(t) can be extracted from the mean-field equation, defining: the firing rate is then given by the total probability flux crossing the threshold: r(t, η) = lim v→+∞J (t, vη), (cid:90) +∞ L(η)r(t, η) dη. −∞ r(t) = lim v→+∞ Reduction The reduction method, see [33], consists in assuming that the solution of the mean-field equation (5) has the form of a Lorentzian distribution: p(t, vη) = 1 π x(t, η) (v − y(t, η))2 + x(t, η)2 . (6) The mean potential and the firing rate are related to the Lorentzian coefficients: and r(t, η) = 1 π x(t, η), (cid:90) +∞ −∞ y(t, η) = vp(t, vη) dv. Thus the mean membrane potential of the network is V (t) = L(η)y(t, η) dη. (cid:90) +∞ −∞ 23 Note that integrals are defined via the Cauchy principal value, the reason being that the Lorentz distribution only has a mean in the principal value sense. After algebraic manipulation, see [33], the transport equation (5) reduces to the dynamical system: τ τ d dt d dt ∆e πτ + 2rV r = V = V 2 + ¯η + I − τ 2π2r2, Such a reduced description has the tremendous advantage to be low dimensional. E-I Interaction Considering now a network of two interacting neural populations of excitatory cells and inhibitory cells, the system is then represented by two probability density functions, one for the excitatory population, which we denote pe(t, vη), and one for the inhibitory neurons, which we denote pi(t, vη). Each density function follows a continuous transport equation similar to (5). In our case, the dynamic of the two coupled PDEs that describe the time evolution of pe(t, vη) and pi(t, vη) reduces to a set of differential equations. For the E-cells, we have: re = + 2reVe ∆e πτe e + ¯ηe + Ie − τ 2 Ve = V 2 e π2r2 e, and for the I-cells: d dt d dt τe  τe  τi d dt d dt ri = τi Vi = V 2 + 2riVi ∆i πτi i + ¯ηi + Ii(t) − τ 2 i π2r2 i . Note that the two systems are in interaction via the expression of the currents Ie and Ii which include self-recurrent connections and synaptic projections, see Fig. 1 for a schematic view. For the E-cells, the total current has the following form: and for the I-cells: Ie(t) = I ext e (t) + τesee(t) − τesei(t), Ii(t) = I ext i (t) + τisie(t) − τisii(t), 24 τe τe   re = ∆e πτe + 2reVe e + τesee − τesei − τ 2 e π2r2 e e + ¯ηe + I ext Ve = V 2 see = −see + Jeere sei = −sei + Jeiri, d dt d dt d dt d dt τs τs τi τi τs τs d dt d dt d dt d dt ri = ∆i πτi + 2riVi i + τisie − τisii − τ 2 i π2r2 i Vi = V 2 i + ¯ηi + I ext sie = −sie + Jiere sii = −sii + Jiiri, (7) (8) here, sαβ(t) represents the time evolution of the synaptic current of the population β projected on the population α and is given by an exponential filter of the firing activity. In the end, we get that the dynamic of the cortical network is well described by the following set of eight differential equations: and where τs is the synaptic time constant, and Jαβ is the synaptic strength of the population β projecting on the population α. Phase Response Curve The infinitesimal phase resetting curve (iPRC) is defined mathematically for infinitesimally small perturba- tion, and it is computed in a perfectly rigorous way via the adjoint method [51]. Let us consider a general dynamical system: d dt x(t) = F (x(t)), where x ∈ Rn. Assuming that the system admits a stable limit cycle x0(t), then if the system is perturbed by a small perturbation, the solution can be written as where p(t) is the small deviation from the limit cycle. Up to a linearization, we get that x(t) = x0(t) + p(t), 25 p(t) = DF (x0(t)) · p(t), d dt where DF (x0(t)) is the time dependent Jacobian matrix. The iPRC is then defined as (Z(t) · p(t)) = 0, d dt which is equivalent to (Z(t) · p(t)) = d dt = = = p(t) Z(t) · p(t) + Z(t) · d dt Z(t) · p(t) + Z(t) · DF (x0(t)) · p(t) Z(t) · p(t) + DF (x0(t))T · Z(t) · p(t) · p(t) Z(t) + DF (x0(t))T · Z(t) (cid:19) d dt d dt d dt (cid:18) d dt = 0. Since the last equation is valid for every perturbation p(t), we get that the iPRC is solution of the adjoint equation: Z(t) = −DF (x0(t))T · Z(t). d dt This method can be applied on the low dimensional system (7)-(8) and a semi analytical expression of the iPRC can be extracted. Assuming that O(t) = (reo(t), Veo(t), see(t), sei(t), rio(t), Vio(t), sie(t), sii(t)) , is a stable limit cycle of the E-I system (7)-(8) of period T , that is, O(t) = O(t + T ) we find that the iPRC Z(t) is a periodic vector of eight components Z(t) = (Zre(t), Zve(t), Zsee (t), Zsei(t), Zri(t), Zvi (t), Zsie(t), Zsii(t)) , 26 that is a solution of the adjoint equation Z(t) = M(t)T · Z(t), − d dt where the matrix M(t) is given by a linearization of the E-I system (7)-(8) around the limit cycle: 2Veo(t) τe −2τeπ2reo(t) M(t) =  Jee τs 0 0 0 Jie τs 0 2reo(t) τe 2Veo(t) τe 0 0 0 0 0 0 0 1 − 1 τs 0 0 0 0 0 0 −1 0 − 1 τs 0 0 0 0 0 0 0 Jei τs 2Vio(t) τi −2τiπ2rio(t) 0 Jii τs 0 0 0 0 2rio(t) τi 2Vio(t) τi 0 0 0 0 0 0 0 1 − 1 τs 0 0 0 0 0 0 −1 0 − 1 τs  . The iPRC Z(t) is given by the unique periodic solution that satisfies the normalization condition Z(t) · O(t) = 2π/T. The iPRC can be compared with a direct method which consist in presenting perturbation to the network. Depending on the phase onset of the perturbation, the network activity is going to shift. Raster plots from numerical simulations of the full network (Fig. 9) illustrate the shift. Here the black dots correspond to the unperturbed network, whereas the colored dots to the perturbed circuit. Before the stimulus onset, the two rasters overlap perfectly. After the stimulus presentation, spikes of the perturbed network are shifted: either delayed (Fig. 9A) or advanced (Fig. 9B) depending on the onset phase of the perturbation. In the Result section, the two approches - direct perturbation and the adjoint method - are compared for the PING interaction in Fig. 2 and the ING interaction in Fig. 3. Bidirectionally Coupled Networks Considering now two bidirectionally delayed coupled networks where the coupling is made via long projections of the pyramidal cells from one network to another, the whole system reduces to a set of sixteen differential equations. For the first network, we have 27 Figure 9: Phase shifting. A-B) Spiking activity obtained from simulations of the full spiking network. The black dots illustrate the ongoing activity and the colored dots (blue for the I-cells and red for the E-cells) the activity of the perturbed network. Lower panels: Illustration of the stimulus onset. Perturbations are made on the I-cells. The network parameters are the same as in Fig. 2 with I ext e = 10, direct perturbations are made with a square wave current pulse (amplitude 10, duration 0.5). and τe τe τs τs   re1 = ∆e πτe + 2re1Ve1 e + τesee1 − τesei1 − τ 2 e π2r2 e1 + ¯ηe + +I ext Ve1 = V 2 e1 see1 = −see1 + Jeere1 + Geere2(t − d) sei1 = −sei1 + Jeiri1 , d dt d dt d dt d dt τi τi τs τs d dt d dt d dt d dt ri1 = ∆i πτi + 2ri1 Vi1 i + τisie1 − τisii1 − τ 2 i π2r2 i1 + ¯ηi + I ext Vi1 = V 2 i1 sie1 = −sie1 + Jiere1 + Giere2(t − d) sii1 = −sii1 + Jiiri1, (9) (10) and for the second network: 28 and τe τe τs τs   re2 = ∆e πτe + 2re2Ve2 e + τesee2 − τesei2 − τ 2 i π2r2 e2 + ¯ηe + +I ext Ve2 = V 2 e2 see2 = −see2 + Jeere2 + Geere1(t − d) sei2 = −sei2 + Jeiri2 , d dt d dt d dt d dt τi τi τs τs d dt d dt d dt d dt ri2 = ∆i πτi + 2ri2 Vi2 i + τisie2 − τisii2 − τ 2 i π2r2 i2 + ¯ηi + I ext Vi2 = V 2 i2 sie2 = −sie2 + Jiere2 + Giere1(t − d) sii2 = −sii2 + Jiiri2, (11) (12) Note the presence of long range projections between circuits, see Fig. 3 for a schematic view. Here Gαβ denotes the connectivity strength of the population β of one network onto the population α of the other circuit, and the parameter d is the conduction delay between the two networks. Phase Equation Assuming that the two networks are oscillating and placing our study within the framework of weakly coupled oscillators, that is, if we assume that Gαβ << 1, we can reduce the bidirectionally delayed-coupled neural circuits description (9)-(10)-(11)-(12) to a single phase equation: d dt θ(t) = G(θ(t)). Here θ(t) is the phase difference (or phase lag) between the circuits and the G-function is the odd part of the shifted interaction function (the H-function), see [39] for instance: G(θ) = H(θ − d) − H(−θ − d), 29 with d, the time delay between the two circuits. In our case, the interaction function is mathematically described as H(θ) = (cid:90) T (cid:90) T Zsee (s)re(s − θ) ds Zsie(s)re(s − θ) ds, Gee T + 0 Gie T 0 where T is the oscillation period. Note the involvement of the synaptic component of the PRC Zs(t) and the firing rate of the E-cells re(t) all along the oscillatory cycle in the expression of the G-function. Acknowledgments This study was funded by CNRS, INSERM, and partial support from LABEX ANR-10-LABX-0087 IEC and IDEX ANR-10-IDEX-0001-02 PSL*. The study received support from the Russian Science Foundation grant (Contract No. 17- 11-01273). References [1] Buzs´aki G, Draguhn A. Neuronal Oscillations in Cortical Networks. Science. 2004;304(5679):1926 -- 1929. doi:10.1126/science.1099745. [2] Buzsaki G. Rhythms of the Brain. Oxford University Press; 2006. [3] Buzs´aki G, Logothetis N, Singer W. Scaling Brain Size, Keeping Timing: Evolutionary Preservation of Brain Rhythms. Neuron;80(3):751 -- 764. doi:10.1016/j.neuron.2013.10.002. [4] Fries P, Nikolia D, Singer W. The gamma cycle. Trends in Neurosciences. 2007;30(7):309 -- 316. doi:http://dx.doi.org/10.1016/j.tins.2007.05.005. [5] Buzsaki G, Wang XJ. Mechanisms of Gamma Oscillations. Annual Review of Neuroscience. 2012;35(1):203 -- 225. doi:10.1146/annurev-neuro-062111-150444. [6] Bartos M, Vida I, Jonas P. Synaptic mechanisms of synchronized gamma oscillations in inhibitory interneuron networks. Nat Rev Neurosci. 2007;8(1):45 -- 56. 30 [7] Fries P. A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. Trends in Cognitive Sciences. 2005;9(10):474 -- 480. doi:http://dx.doi.org/10.1016/j.tics.2005.08.011. [8] Struber M, Sauer JF, Jonas P, Bartos M. Distance-dependent inhibition facilitates focality of gamma oscillations in the dentate gyrus. Nature Communications. 2017;8(1):758. doi:10.1038/s41467-017-00936- 3. [9] Fries P. Neuronal gamma-band synchronization as a fundamental process in cortical computation. Annual review of neuroscience. 2009;32:209 -- 224. [10] Cardin JA, Carlen M, Meletis K, Knoblich U, Zhang F, Deisseroth K, et al. Driving fast-spiking cells induces gamma rhythm and controls sensory responses. Nature. 2009;459(7247):663 -- 667. [11] Womelsdorf T, Schoffelen JM, Oostenveld R, Singer W, Desimone R, Engel AK, et al. Modulation of neuronal interactions through neuronal synchronization. science. 2007;316(5831):1609 -- 1612. [12] Battaglia D, Witt A, Wolf F, Geisel T. Dynamic Effective Connectivity of Inter-Areal Brain Circuits. PLOS Computational Biology. 2012;8(3):1 -- 20. doi:10.1371/journal.pcbi.1002438. [13] Palmigiano A, Geisel T, Wolf F, Battaglia D. Flexible information routing by transient synchrony. Nature Neuroscience. 2017;20:1014 EP -- . [14] Swadlow HA, Waxman SG. Axonal conduction delays. Scholarpedia. 2012;7. [15] Dumont G, Ermentrout GB, Gutkin B. Macroscopic phase-resetting curves for spiking neural networks. Phys Rev E. 2017;96:042311. doi:10.1103/PhysRevE.96.042311. [16] Deco G, Kringelbach ML. Metastability and Coherence: Extending the Communication through Coher- ence Hypothesis Using A Whole-Brain Computational Perspective. Trends in Neurosciences;39(3):125 -- 135. doi:10.1016/j.tins.2016.01.001. [17] Maris E, Fries P, van Ede F. Diverse Phase Relations among Neuronal Rhythms and Their Potential Function. Trends in Neurosciences;39(2):86 -- 99. doi:10.1016/j.tins.2015.12.004. [18] Knoblich U, Siegle J, Pritchett D, Moore C. What do We Gain from Gamma? Local Dynamic Gain Modulation Drives Enhanced Efficacy and Efficiency of Signal Transmission. Frontiers in Human Neu- roscience. 2010;4:185. doi:10.3389/fnhum.2010.00185. 31 [19] Borgers C, Kopell NJ. Gamma Oscillations and Stimulus Selection. Neural Computation. 2007;20(2):383 -- 414. doi:10.1162/neco.2007.07-06-289. [20] Borgers C, Epstein S, Kopell NJ. Gamma oscillations mediate stimulus competition and atten- tional selection in a cortical network model. Proceedings of the National Academy of Sciences. 2008;105(46):18023 -- 18028. doi:10.1073/pnas.0809511105. [21] Buehlmann A, Deco G. Optimal Information Transfer in the Cortex through Synchronization. PLOS Computational Biology. 2010;6(9):1 -- 10. doi:10.1371/journal.pcbi.1000934. [22] Barardi A, Sancristobal B, Garcia-Ojalvo J. Phase-Coherence Transitions and Communication in the Gamma Range between Delay-Coupled Neuronal Populations. PLOS Computational Biology. 2014;10(7):1 -- 17. doi:10.1371/journal.pcbi.1003723. [23] Lowet E, Roberts M, Hadjipapas A, Peter A, van der Eerden J, De Weerd P. Input-Dependent Frequency Modulation of Cortical Gamma Oscillations Shapes Spatial Synchronization and Enables Phase Coding. PLOS Computational Biology. 2015;11(2):1 -- 44. doi:10.1371/journal.pcbi.1004072. [24] Akam T, Kullmann DM. Oscillations and Filtering Networks Support Flexible Routing of Information. Neuron. 2010;67(2):308 -- 320. doi:https://doi.org/10.1016/j.neuron.2010.06.019. [25] Akam TE, Kullmann DM. Efficient ?Communication through Coherence? Requires Oscillations Structured to Minimize Interference between Signals. PLOS Computational Biology. 2012;8(11):1 -- 15. doi:10.1371/journal.pcbi.1002760. [26] Akam T, Kullmann DM. Oscillatory multiplexing of population codes for selective communication in the mammalian brain. Nature Reviews Neuroscience. 2014;15:111 EP -- . [27] Canavier CC. Phase-resetting as a tool of information transmission. Current Opinion in Neurobiology. 2015;31:206 -- 213. doi:http://dx.doi.org/10.1016/j.conb.2014.12.003. [28] Kopell NJ, Gritton HJ, Whittington MA, Kramer MA. Beyond the Connectome: The Dynome. Neuron;83(6):1319 -- 1328. doi:10.1016/j.neuron.2014.08.016. [29] Izhikevich EM. Dynamical Systems in Neuroscience. Sejnowski TJ, Poggio TA, editors. The MIT Press; 2007. [30] Ermentrout GB, Terman D. Mathematical foundations of neuroscience. Springer; 2010. 32 [31] Ott E, Antonsen TM. Low dimensional behavior of large systems of globally coupled oscillators. Chaos. 2008;18(3). doi:http://dx.doi.org/10.1063/1.2930766. [32] Luke TB, Barreto E, So P. Complete Classification of the Macroscopic Behavior of a Heterogeneous Network of Theta Neurons. Neural Computation. 2013;25(12):3207 -- 3234. doi:10.1162/NECO a 00525. [33] Montbri´o E, Paz´o D, Roxin A. Macroscopic Description for Networks of Spiking Neurons. Phys Rev X. 2015;5:021028. doi:10.1103/PhysRevX.5.021028. [34] Akao A, Ogawa Y, Jimbo Y, Ermentrout GB, Kotani K. Relationship between the mechanisms of gamma rhythm generation and the magnitude of the macroscopic phase response function in a population of excitatory and inhibitory modified quadratic integrate-and-fire neurons. Phys Rev E. 2018;97:012209. doi:10.1103/PhysRevE.97.012209. [35] Kotani K, Yamaguchi I, Yoshida L, Jimbo Y, Ermentrout GB. Population dynamics of the mod- ified theta model: macroscopic phase reduction and bifurcation analysis link microscopic neuronal interactions to macroscopic gamma oscillation. Journal of The Royal Society Interface. 2014;11(95). doi:10.1098/rsif.2014.0058. [36] Stiefel KM, Ermentrout GB. NEURONS AS OSCILLATORS. Journal of Neurophysiology. 2016;doi:10.1152/jn.00525.2015. [37] Ashwin P, Coombes S, Nicks R. Mathematical Frameworks for Oscillatory Network Dynamics in Neu- roscience. The Journal of Mathematical Neuroscience. 2016;6(1):2. doi:10.1186/s13408-015-0033-6. [38] Nakao H. Phase reduction approach to synchronisation of nonlinear oscillators. Contemporary Physics. 2016;57(2):188 -- 214. doi:10.1080/00107514.2015.1094987. [39] Ermentrout B, Ko TW. Delays and weakly coupled neuronal oscillators. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences. 2009;367(1891):1097 -- 1115. doi:10.1098/rsta.2008.0259. [40] Fries P. Rhythms for Cognition: Communication through Coherence. Neuron. 2015;88(1):220 -- 235. doi:https://doi.org/10.1016/j.neuron.2015.09.034. [41] Ermentrout B. Ermentrout-Kopell canonical model. 2008;3(3):1398. 33 [42] Strogatz SH. From Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators. Physica D: Nonlinear Phenomena. 2000;143(1):1 -- 20. doi:https://doi.org/10.1016/S0167-2789(00)00094-4. [43] Deco G, Jirsa VK, Robinson PA, Breakspear M, Friston K. The Dynamic Brain: From Spiking Neurons to Neural Masses and Cortical Fields. PLoS Comput Biol. 2008;4(8):1 -- 35. doi:10.1371/journal.pcbi.1000092. [44] Wilson HR, Cowan JD. Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons. Biophysical Journal;12(1):1 -- 24. doi:10.1016/S0006-3495(72)86068-5. [45] Devalle F, Roxin A, Montbri´o E. Firing rate equations require a spike synchrony mechanism to correctly describe fast oscillations in inhibitory networks. ArXiv e-prints. 2017;. [46] Smeal RM, Ermentrout GB, White JA. Phase-response curves and synchronized neural networks. Philosophical Transactions of the Royal Society of London B: Biological Sciences. 2010;365(1551):2407 -- 2422. doi:10.1098/rstb.2009.0292. [47] Reyes AD, Fetz EE. Two modes of interspike interval shortening by brief transient depolarizations in cat neocortical neurons. Journal of Neurophysiology. 1993;69(5):1661 -- 1672. doi:10.1152/jn.1993.69.5.1661. [48] Akam T, Oren I, Mantoan L, Ferenczi E, Kullmann DM. Oscillatory dynamics in the hippocampus support dentate gyrus-CA3 coupling. Nat Neurosci. 2012;15(5):763 -- 768. [49] Stiefel KM, Gutkin BS, Sejnowski TJ. The effects of cholinergic neuromodulation on neuronal phase- response curves of modeled cortical neurons. Journal of computational neuroscience. 2009;26(2):289 -- 301. doi:10.1007/s10827-008-0111-9. [50] Phoka E, Cuntz H, Roth A, Hausser M. A New Approach for Determining Phase Response Curves Reveals that Purkinje Cells Can Act as Perfect Integrators. PLOS Computational Biology. 2010;6(4):1 -- 14. doi:10.1371/journal.pcbi.1000768. [51] Brown E, Moehlis J, Holmes P. On the Phase Reduction and Response Dynamics of Neural Oscillator Populations. Neural Computation. 2004;16(4):673 -- 715. doi:10.1162/089976604322860668. [52] Kotani K, Yamaguchi I, Ogawa Y, Jimbo Y, Nakao H, Ermentrout GB. Adjoint Method Pro- vides Phase Response Functions for Delay-Induced Oscillations. Phys Rev Lett. 2012;109:044101. doi:10.1103/PhysRevLett.109.044101. 34 [53] Nakao H, Yanagita T, Kawamura Y. Phase-Reduction Approach to Synchronization of Spatiotemporal Rhythms in Reaction-Diffusion Systems. Phys Rev X. 2014;4:021032. doi:10.1103/PhysRevX.4.021032. [54] Canolty RT, Knight RT. The functional role of cross-frequency coupling. Trends in Cognitive Sciences. 2010;14(11):506 -- 515. doi:10.1016/j.tics.2010.09.001. [55] Battaglia D, Brunel N, Hansel D. Temporal Decorrelation of Collective Oscillations in Neu- ral Networks with Local Inhibition and Long-Range Excitation. Phys Rev Lett. 2007;99:238106. doi:10.1103/PhysRevLett.99.238106. [56] Gutkin BS, Ermentrout GB, Reyes AD. Phase-Response Curves Give the Responses of Neurons to Transient Inputs. Journal of Neurophysiology. 2005;94(2):1623 -- 1635. doi:10.1152/jn.00359.2004. [57] Ermentrout GB, Galan RF, Urban NN. Relating Neural Dynamics to Neural Coding. Phys Rev Lett. 2007;99:248103. doi:10.1103/PhysRevLett.99.248103. [58] Ratas I, Pyragas K. Macroscopic self-oscillations and aging transition in a network of synaptically coupled quadratic integrate-and-fire neurons. Phys Rev E. 2016;94:032215. doi:10.1103/PhysRevE.94.032215. [59] Paz´o D, Montbri´o E. From Quasiperiodic Partial Synchronization to Collective Chaos in Populations of Inhibitory Neurons with Delay. Phys Rev Lett. 2016;116:238101. doi:10.1103/PhysRevLett.116.238101. [60] Kuhn T, Helias M. Locking of correlated neural activity to ongoing oscillations. PLOS Computational Biology. 2017;13(6):1 -- 32. doi:10.1371/journal.pcbi.1005534. [61] Jiang X, Shen S, Cadwell CR, Berens P, Sinz F, Ecker AS, et al. Principles of connectivity among mor- phologically defined cell types in adult neocortex. Science. 2015;350(6264). doi:10.1126/science.aac9462. [62] Potjans TC, Diesmann M. The Cell-Type Specific Cortical Microcircuit: Relating Structure and Activity in a Full-Scale Spiking Network Model. Cerebral Cortex. 2014;24(3):785 -- 806. doi:10.1093/cercor/bhs358. [63] Bos H, Diesmann M, Helias M. Identifying Anatomical Origins of Coexisting Oscillations in the Cortical Microcircuit. PLOS Computational Biology. 2016;12(10):1 -- 34. doi:10.1371/journal.pcbi.1005132. 35
1706.03839
1
1706
2017-06-12T20:27:48
A scalp-EEG network-based analysis of Alzheimer's disease patients at rest
[ "q-bio.NC" ]
Most brain disorders including Alzheimer's disease (AD) are related to alterations in the normal brain network organization and function. Exploring these network alterations using non-invasive and easy to use technique is a topic of great interest. In this paper, we collected EEG resting-state data from AD patients and healthy control subjects. Functional connectivity between scalp EEG signals was quantified using the phase locking value (PLV) for 6 frequency bands. To assess the differences in network properties, graph-theoretical analysis was performed. AD patients showed decrease of mean connectivity, average clustering and global efficiency in the lower alpha band. Positive correlation between the cognitive score and the extracted graph measures was obtained, suggesting that EEG could be a promising technique to derive new biomarkers of AD diagnosis.
q-bio.NC
q-bio
A scalp-EEG network-based analysis of Alzheimer's disease patients at rest Aya Kabbara a,b,c,d* Wassim El Falou a,b Mohamad Khalil a,b Hassan Eide Mahmoud Hassan c,d 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 e Mazloum Hospital, Tripoli, Lebanon [email protected] Abstract-Most brain disorders including Alzheimer's disease (AD) are related to alterations in the normal brain network organization and function. Exploring these network alterations using non-invasive and easy to use technique is a topic of great interest. In this paper, we collected EEG resting-state data from AD patients and healthy control subjects. Functional connectivity between scalp EEG signals was quantified using the phase locking value (PLV) for 6 frequency bands, θ (4– 8 Hz), α1(8–10 Hz), α2(10–13 Hz), ß(13–30 Hz), γ(30– 45 Hz), and broad band (0.2-45 Hz). To assess the differences in network properties, graph-theoretical analysis was performed. AD patients showed decrease of mean connectivity, average clustering and global efficiency in the lower alpha band. Positive correlation between the cognitive score and the extracted graph measures was obtained, suggesting that EEG could be a promising technique to derive new biomarkers of AD diagnosis. Keywords- Alzheimer's disease; EEG connectivity; resting state; graph theory I. INTRODUCTION Here, we explored the topological differences between the networks of AD and healthy controls using EEG. For this end, we collected EEG data from 14 subjects at rest (eyes closed). We then reconstructed the functional networks at the scalp- level using the phase synchronization method [10] for θ, α1, α2, ß, γ, and broad band. This step has been followed by a graph quantification of the constructed networks. Our results showed that AD networks are characterized by reduced global and local connectivity compared to healthy controls in alpha1 frequency band. In addition, a significant correlation between the clinical test (MMSE score) and the EEG graph metrics was obtained. II. MATERIALS AND METHODS The full pipeline of the study is illustrated in Figure 1. Around 30,000 people suffered from Alzheimer's disease (AD) or related dementia in Lebanon. AD is generally characterized by a progressive decline of memory and cognitive functions [1]. Nowadays, many evidences suggest that AD is associated with alterations of large-scale brain networks due to synaptic dysfunction [2]. Therefore, the identification of the alterations occurred in the brain network topology of AD patients is a topic of great interest [3]. Moreover, in some cases, AD patients show only moderate clinical symptoms which impede their diagnosis [4], [5]. This highlights the need of objective, non-invasive and easy to use biomarkers for an early detection of the disease. In this context, from electroencephalography (EEG) may be a key feature to explore the network properties of AD. To quantify brain network, graph theory-based analysis is usually used. In graph-theory approaches, brain networks are represented as graphs composed of nodes (electrodes or brain regions) connected by edges (functional connectivity) [6]. Once nodes and edges are identified from the neuroimaging data, network topological properties can be explored by extracting specific graph metrics. In the context of resting- state analysis of AD, numerous functional and structural studies and have been conducted [7]–[9].. resting-state networks derived Figure 1. The full pipeline of the study A. Participants Seven healthy controls (5 males and 2 females, age 64–74 y) and seven patients diagnosed with AD (4 females and 3 males, age 69–81 y) participated in this study. AD patients were recruited at the memory clinic of Al-Ajaza Hospital and Mazloum Hospital, Tripoli, Lebanon. None of the patients was suffering from another neurological disease. Controls were recruited from the local community, and from Al-Ajaza Hospital. Prior to their participation, all subjects have provided informed consent. The study was approved by the local institutional review boards (CE-EDST-3-2017). B. Data acquisition EEG signals were recorded using a 32-channel EEG system (Twente Medical Systems International (TMSi), Porti system). Electrodes were placed in accordance with the revised international 10/20 system [11] at Fpz, Fp1, Fp2, Fz, F3, F4, FC1, FC2, FC5, FC6, C3, C4, Cz, CP1, CP2, CP5, CP6, P3, P4, P7, P8, Pz, PO3, PO4, PO7, PO8, POz, O1, O2, Oz, T7 and T8. The wristband was used as ground. EEG signals were sampled at 512 Hz, bandpass filtered within 0.1– 80 Hz, and stored in GDF format. Subjects were asked to relax with their eyes-closed without falling asleep for 10 minutes. We also screened their cognitive score using the mini-mental state examination (MMSE) [12]. C. EEG preprocessing EEG signals were filtered between 0.1 and 45 Hz, and segmented into 40s epochs. Segments with amplitude ±90μV were considered as artifactual and were eventually discarded from the analysis after visual inspection. For some cases, the electrodes detected as noisy were interpolated. For each participant, we selected the maximum number of non- artifactual segments. D. Functional connectivity analysis representing functional connectivity matrices The the estimates of interaction between all pairs of EEG channels were computed in the different frequency bands (θ(4–8 Hz), α1(8–10 Hz), α2(10–13 Hz), ß(13–30 Hz), γ(30–45 Hz), and broad band (0.2-45 Hz)). Here, the connectivity was assessed using the phase synchronization (PS) method as it is less susceptible to the effects of artifacts [13] and was successfully used in several scalp-level EEG connectivity analyses [14], [15] such as Brain Computer Interfaces . In particular, we used the phase locking value (PLV) proposed in [16] and defined as: 𝑃𝐿𝑉 𝑡 = 𝑒𝑥𝑝(𝑗(𝜑𝑦 (𝑡) − 𝜑𝑥 (𝑡))𝑑𝜏 1 𝛿 𝑡+𝛿/2 𝑡−𝛿/2 where 𝜑𝑦 (𝑡) and 𝜑𝑥 𝑡 are the unwrapped phases of the signals x and y at time t. The Hilbert transform was used to extract the instantaneous phase of each signal. 𝛿 denotes the size of the window in which PLV is calculated. The connectivity matrices were thresholded by using a 10% proportional threshold. For each subject, we obtained the frequency-specific networks, in which, nodes signify the 32 electrodes and edges represent the connectivity value between channels (PLV). E. Graph metrics extraction In order to study the difference between healthy and AD networks, graph theoretical analysis was used [17]. Here we used three graph metrics: 1- Clustering coefficient: The clustering coefficient of a node quantifies how close its neighbors are to being a clique [18]. This parameter is one of the most elementary measures of local segregation (i.e. the extent to which the network is organized into specialized functional regions). [17] The average of the clustering coefficients for each individual node is the clustering coefficient of the graph. 2- Global efficiency: The global efficiency is computed as the average inverse shortest path length [19]. A short path length indicates that, on average, each node can be reached from any other node along a path composed of only a few edges [6]. The global efficiency measures the global integration of a network (i.e the capacity of the network to combine information from many distributed areas). 3- Strength: It evaluates the importance of a node within a network. This measure is defined as the sum of the edges' weights linked to the node [20]. III. RESULTS First, for each subject, the connectivity matrices of all segments were averaged, and then the different graph metrics (mean PLV value, average clustering, and global efficiency) were extracted. Figure 2.A illustrates the difference between healthy and AD subjects at each frequency band. The mean functional connectivity as well as the average clustering and the global efficiency in the AD group were decreased in the lower alpha band compared to controls (p <0.001, Wilcoxon test) (Figure 2.A). As illustrated, no between group differences were obtained for theta, alpha2, beta, and gamma frequency bands (p>0.01). Second, we assessed the correlation between the cognitive state of each subject (MMSE score) and the extracted graph measures (Figure 2.B). Results show a positive significant correlation for the mean connectivity (R=0.81; p<0.001), the average clustering the global efficiency (R= 0.81, p<0.001). We were also interested in identifying the nodes that show decreased activity in AD patients compared to controls, in terms of strength and clustering coefficient. For this aim, we concatenated the metrics distributions of each node across subjects at alpha1 frequency band. Then, we computed the difference between the two groups on each individual node using Wilcoxon test. Results depict that the nodes that show significantly reduced functionality in the AD group (p<0.001) with regard to the clustering coefficient are: Oz, Fp1, Fp6, FC6, FC1, POz. According to the strength, the disrupted nodes are: Oz, O2, C3, C4, Fp2, Fpz, Fp1, FC5, FC1, FC2, F4, POz, PO8, Fz. (R=0.82, p<0.001), and IV. DISCUSSION Results show that AD alters the global and local properties of the scalp-level brain network. The decreased functional synchronization obtained in AD patients compared to healthy controls agrees with previous finding [8], [21]–[23]. Similarly, the decreased global efficiency of the AD network is consistent with [7], [9]. Results revealed an association of the alterations in brain network organization with cognitive deficits (MMSE clinical test) in AD. This was also observed in several studies [7]–[9], [21], [22]. However, the reduced segregation (low clustering coefficient) associated mainly to the frontal regions is in contradiction with multiple studies as reported in this review [23]. A possible explanation of this difference is the use of different neuroimaging techniques (with different spatial resolutions) and different methodologies (connectivity measure, graph measure). Figure 2. A) The difference between AD patients and healthy networks in terms of 1) mean functional connectivity, 2) average clustering and global efficiency for the six frequency bands and B) The correlations between the MMSE score and the graph metrics in the alpha1 band. This preliminary study has some limitations that should be addressed. First, the functional connectivity was computed on the scalp (electrodes level) that could be generally corrupted by the volume conduction problem. To overcome this problem and to improve the spatial resolution of the networks, ‗EEG source connectivity' method is an emerging technique used in cognition [24]–[26] and brain disorders [27][28]. However, in the current study, we were cautious to reconstruct the brain sources because of the relatively low number of available channels (32 electrodes) which may affect the accuracy of the results [24]. Therefore, a further work is to take into account this technical issue and adapt the EEG source connectivity method in order to accurately identify the cortical networks using relatively low number of channels. Second, we used a proportional threshold to remove weak connections of the functional connectivity matrices. This threshold was used to ensure equal density across patients and control subjects. Nevertheless, one should test the reproducibility of results by applying other proportional thresholds. Finally, the analysis should be extended on a larger dataset to obtain consistent results across subjects. V. CONCLUSION In this paper, we used EEG scalp connectivity method to explore the alterations in the networks organization of AD patients. The synchronization was assessed using the PLV measure, and the networks were characterized in terms of average clustering, global efficiency and mean functional J. P. Lachaux, E. Rodriguez, J. Martinerie, and F. J. Varela, ―Measuring phase synchrony in brain signals,‖ Hum. Brain Mapp., vol. 8, pp. 194–208, 1999. E. Bullmore, E. Bullmore, O. Sporns, and O. Sporns, ―Complex brain networks: graph theoretical analysis of structural and functional systems,‖ Nat Rev Neurosci, vol. 10, no. 3, pp. 186–198, 2009. D. J. Watts and S. H. Strogatz, ―Collective dynamics of ‗small- world' networks.,‖ Nature, vol. 393, no. 6684, pp. 440–2, 1998. V. Latora and M. Marchiori, ―Efficient Behavior of Small World Networks,‖ Phys. Rev. Lett., vol. 87, p. 198701, 2001. A. Barrat, M. Barthélemy, R. Pastor-Satorras, and A. Vespignani, ―The architecture of complex weighted networks.,‖ Proc. Natl. Acad. Sci. U. S. A., vol. 101, no. 11, pp. 3747–3752, 2004. L. Canuet, I. Tellado, V. Couceiro, C. Fraile, L. Fernandez-Novoa, R. Ishii, M. Takeda, and R. Cacabelos, ―Resting-State Network Disruption and APOE Genotype in Alzheimer's Disease: A lagged Functional Connectivity Study,‖ PLoS One, vol. 7, no. 9, 2012. M. Hata, H. Kazui, T. Tanaka, R. Ishii, L. Canuet, R. D. Pascual- Marqui, Y. Aoki, S. Ikeda, H. Kanemoto, K. Yoshiyama, M. Iwase, and M. Takeda, ―Functional connectivity assessed by resting state EEG correlates with cognitive decline of Alzheimer's disease - An eLORETA study,‖ Clin. Neurophysiol., vol. 127, no. 2, pp. 1269– 1278, 2016. P. Vemuri, D. T. Jones, and C. R. Jack, ―Resting state functional MRI in Alzheimer's Disease.,‖ Alzheimers. Res. Ther., vol. 4, no. 1, p. 2, 2012. 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. A. Kabbara, W. El Falou, M. Khalil, F. Wendling, and M. Hassan, ―Graph analysis of spontaneous brain network using EEG source connectivity,‖ arXiv Prepr. arXiv1607.00952, 2016. 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. M. Hassan, I. Merlet, A. Mheich, A. Kabbara, A. Biraben, A. Nica, and F. Wendling, ―Identification of Interictal Epileptic Networks from Dense-EEG,‖ Brain Topography, pp. 1–17, 2016. M. Hassan, L. Chaton, P. Benquet, A. Delval, C. Leroy, L. Plomhause, A. J. H. Moonen, A. A. Duits, A. F. G. Leentjens, V. van Kranen-Mastenbroek, L. Defebvre, P. Derambure, F. Wendling, and K. Dujardin, ―Functional connectivity disruptions correlate with cognitive phenotypes in Parkinson's disease,‖ NeuroImage Clin., vol. 14, pp. 591–601, 2017. [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] connectivity. Our results show a correlation between these measures and the MMSE score in the lower alpha band. This preliminary study opens insights on the capacity of EEG-based functional connectivity method to identify the pathological brain networks during rest. ACKNOWLEDGMENT This work has been funded with support from the National Council for Scientific Research in Lebanon. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] REFERENCES J. L. Cummings, ―Alzheimer's Disease.,‖ N. Engl. J. Med., vol. 351, no. 1, pp. 56–67, 2004. S. H. Joo, H. K. Lim, and C. U. Lee, ―Three Large-Scale Functional Brain Networks from Resting-State Functional MRI in Subjects with Different Levels of Cognitive Impairment,‖ Psychiatry Investig, vol. 13, no. 1, pp. 1–7, 2016. A. Fornito, A. Zalesky, and M. Breakspear, ―The connectomics of brain disorders,‖ Nat. Rev. Neurosci., vol. 16, no. 3, pp. 159–172, 2015. S. L. Rogers, M. R. Farlow, R. S. Doody, R. Mohs, and L. T. Friedhoff, ―A 24-week, double-blind, placebo-controlled trial of donepezil in patients with Alzheimer's disease. Donepezil Study Group.,‖ Neurology, vol. 50, no. 1, pp. 136–45, 1998. P. N. Tariot, P. R. Solomon, J. C. Morris, P. Kershaw, S. Lilienfeld, and C. Ding, ―A 5-month, randomized, placebo-controlled trial of galantamine in AD. The Galantamine USA-10 Study Group.,‖ Neurology, vol. 54, no. 12, pp. 2269–76, 2000. O. Sporns, Networks of the brain. 2010. C.-Y. Lo, P.-N. Wang, K.-H. Chou, J. Wang, Y. He, and C.-P. Lin, ―Diffusion Tensor Tractography Reveals Abnormal Topological Organization in Structural Cortical Networks in Alzheimer's Disease,‖ J. Neurosci., vol. 30, no. 50, pp. 16876–16885, 2010. C. A. Mallio, R. Schmidt, M. A. de Reus, F. Vernieri, L. Quintiliani, G. Curcio, B. Beomonte Zobel, C. C. Quattrocchi, and M. P. van den Heuvel, ―Epicentral Disruption of Structural Connectivity in Alzheimer's Disease,‖ CNS Neurosci. Ther., vol. 21, no. 10, pp. 837–845, 2015. C. J. Stam, B. F. Jones, G. Nolte, M. Breakspear, and P. Scheltens, ―Small-world networks and functional connectivity in Alzheimer's disease,‖ Cereb. Cortex, vol. 17, no. 1, pp. 92–99, 2007. J.-P. Lachaux, E. Rodriguez, M. Le van Quyen, A. Lutz, J. Martinerie, and F. J. Varela, ―Studying single-trials of phase synchronous activity in the brain,‖ Int. J. Bifurc. Chaos, vol. 10, no. 10, pp. 2429–39, 2000. J. N. Acharya, A. Hani, J. Cheek, P. Thirumala, and T. N. Tsuchida, ―American Clinical Neurophysiology Society guideline 2: guidelines for standart electrode position nomenclature,‖ J. Clin. Neurophysiol., vol. 33, no. 4, pp. 308–311, 2016. M. Galea and M. Woodward, ―Mini-mental state examination (MMSE),‖ Aust. J. Physiother., vol. 51, no. 3, p. 198, 2005. W. A. Truccolo, M. Ding, K. H. Knuth, R. Nakamura, and S. L. Bressler, ―Trial-to-trial variability of cortical evoked responses: Implications for the analysis of functional connectivity,‖ Clin. Neurophysiol., vol. 113, no. 2, pp. 206–226, 2002. A. Kabbara, M. Hassan, M. Khalil, H. Eid, and W. El-Falou, ―An efficient P300-speller for Arabic letters,‖ in Advances in Biomedical Engineering (ICABME), 2015 International Conference on, 2015, pp. 142–145. A. Kabbara, M. Khalil, W. El-Falou, H. Eid, and M. Hassan, ―Functional brain connectivity as a new feature for P300 speller,‖ PLoS One, vol. 11, no. 1, 2016.
1803.08362
1
1803
2018-03-04T17:10:03
Consciousness: From the Perspective of the Dynamical Systems Theory
[ "q-bio.NC" ]
Beings, animate or inanimate, are dynamical systems which continuously interact with the (external and /or internal) environment through the physical or physiologic interfaces of their Kantian (representational) realities. And the nature of their interactions is determined by the inner workings of their systems. It is from this perspective that this work attempts to address some of the long-held philosophical questions, major one among them consciousness-- in the context of the physicality of the dynamic systems.
q-bio.NC
q-bio
Consciousness: From the Perspective of the Dynamical Systems Theory By Jahan N. Schad, PhD Retired LBNL (UCB) Scientist 376 Tharp Drive, Moraga, Ca 94556 Email: [email protected] Phone: (925) 376-4126 1 Abstract Beings, animate or inanimate, are dynamical systems that continuously interact with the (external and /or internal) environment through the physical or physiologic interfaces of their Kantian (representational) realities. And the nature of their reactions is determined by their systems' inner workings. It is from this perspective that this work attempts to address some of the long held philosophical questions; major one among them consciousness, in the context of the physicality of such systems. And to this end, the approach relies upon the appropriate governing mathematical formalisms of system's operations (behavior): For higher beings, the computational theory of the brain (Edelman 1987; Kandel 2013), despite lack of details, provides the necessary insights into the likely mathematical processes which must be behind the operations of the system. For inanimate matter, the process is gravely simpler: the responses to environmental (initial and boundary) conditions (inputs) are governed by its field equation (constitutive properties, and constitutive and conservation laws), which render physical changes, which are the expressions (outputs) some of which may appear on their interfaces with their external environment. In the former, that is the case of the higher beings, their systems' operations are generally very complex and inevitably would involve brain (computed) solutions of discerned complexities (from sense organ inputs) and streaming downloads of the results (perceptions/conceptions outputs), through the nervous system, to the body physiologic interfaces, for their expressions. The latter expressions are animate functions and characteristics such as biological sustenance; maintenance, behavior, thoughts and vocalizations; and the seemingly awareness of sentience, and other associated phenomena, which together define the consciousness, albeit with some reporting shortcoming due to interface display limitations. Prima facie, the genesis of the consciousness, from the view point of dynamic system theory, -- being simply expressions of some of the results of their interaction with the environment -- allows for the generalization of this phenomenon, which is considered only a higher animate peculiarity, to all matter with spatial representation, --animate or otherwise -- granted with vast differences of the nature, and complexities of the related expressions, some of which in humans are referred to as the "experiences of consciousness." In such realm, consciousness is fundamental to all matter with objective and subjective aspects to it: the potential to react signifies the "objective consciousness;" and the nature of the reaction defines the "subjective consciousness." And it is the specificity of the latter, of whatever nature, which separates the animate from the inanimate existence. Within the laid out framework of the present theory, the big baffling question of philosophy, as well as how and where the human subjective experiences of consciousness happen, the hard problem (Chalmers, 1995), find plausible answers: All aspects of 2 Computational Brain; Cognition; Consciousness; Unconscious; human consciousness, are renditions of the results of some of the brain computed events (perceptions/conceptions), -- in response to external and internal stimuli –- by neural mechanisms (Schad, 2016), as functions and expressions, in different modes, through various physiological body interfaces. In humans, the utterance interface displays two of the major components of consciousness of special interest to this work; the thought and the vision: they are certain streamed downloads of perceptions, which are expressed by this interface (Schad, 2016), mostly inaudibly; though occasionally sounded off, as loud thinking. However, at times, the complexities of the thought and vision (download) contents, – likely involving an extensive Lexicon – render their occurring audible reporting deficient due to vocal instrument filtering. And this inevitable physiologic shortfall (caused by vocal frequency bandwidth limitation),-- the incompleteness of the audible expressions of subjective consciousness-- recognized as the hard problem, is very likely sanctioned by the evolutionary processes due to the absence of any survival value. This proposed system theory approach to the understanding of the human sentience and other facets of the brain (mind), follows and complements the (generally accepted) cognitive sciences reductionist (experimentally based) consensus of absence of free will. Key Words: Panpsychism Introduction "….Perhaps it will take the thinking in a science such as biology, which is of a more general order than the three with which we have been concerned, psychology, medicine, and sociology, to provide the answer all three are seeking" Samuel J. Beck and Herman B, Molish (1959) Today's much advanced state of knowledge owes much to the symbiotic efforts of the fields of philosophy and sciences, which have continued throughout all ages. However the rapid development of natural sciences, which had started since ninetieth century, and that of the advances of the sciences of the brain that had taken roots early in the twentieth century, have been increasingly influencing philosophy; and been of prodigies help in search for answers to its long held big questions. Nonetheless, as it has always been the case with all the frontiers of knowledge, philosophy remains to continue its synthesis of the facts of the mind phenomena to finally trigger the development of relevant scientific theories. As it stands, the philosophical world still remains with all of its big question such as the nature of reality; mind-body problem; free will, etc.; and then the most important one, the beholder of them all; the main characteristic of sentience, the consciousness. The very phenomenon of consciousness, at least in case of human beings, has been behind whatever sense of life they have, in general, and, in particular, driving the efforts to divulge its very own rendering. And despite the knowledge of sciences and humanities embedded in it collectively, the puzzle of the promoter itself (the consciousness) remains the major challenge that philosophy, and some of its recent daughter sciences, are facing. The perplexed state of the knowledge, in regards to the questions of the mind, is evinced in the opinion polls taken from philosophers across many world institutions, over past few years (Bourget, and Chalmers 2013): the apparent 3 stagnation, has led to examination of philosophy's reasoning structure; by some of its (today's) brilliant and ardent torch bearers (e.g., Chalmers talk 2014). In such evaluations, the lack of progress is being attributed to the weaknesses in the philosophy's arguable premises for the mind; and for how sentience is to be unraveled. Recent Phil paper surveys results are reflected in David Chalmers (2014) statement:" There is no collective convergence about truth of the big questions of philosophy such as mind body, Free will, etc., because there are no indisputable premises (axioms, or a more fitting term "postulate") to base the argumentative approach of philosophy on them, in order to come up with their compelling proof." However, there is an exception in this finding and that is the fact that opinions on consciousness are seemingly converging: there is consensus in parsing the difficulties of understanding consciousness into hard and easy problems, -- from the view point of the involved complexities – which could render them more tractable: The first category is ascribed to the subjective phenomena, "which result from physical processes and yet not explainable by them (Chalmers, 1995);" and therefore, the experiences of consciousness (the hard part), such as feelings, emotions, thought, etc., not explainable because of the absence of any functional attributes, have remained an enigma so far. But the functional events of consciousness (the easy problems), are thought to be possibly explainable by cognitive process of the brain (Chalmers, 1995). The latter designation does not by no means imply that the detail of the related mental operation are known, but the prospects are thought to be favorable, and much better understanding of them is likely to happen within this century (Chalmers 2013). The optimism is based on recent progress in sciences that has opened up the possibilities of achieving some insight into the mysteries of the brain operations in general, and of consciousness phenomena in particular. This opportunity is specifically owed to new understandings of brain functions due to the extensive neurosciences research (Kandel, Schwartz and Jessell 2000), on the one hand, and artificial intelligence successes through deployments of the traditional and the neural/neuronal network computers (Turing 1948, McCulloch 1993; Von Neumann 1958; Churchland 1992; Edelman 1987; Arbib 1989), on the other, which together have led to the formulation of the concept of the computational brain theory. This brain theory is presently widely accepted (Kandel 2013; Churchland 1992; Wilczek, 2008), and it's standing according to cognitive neuroscientist Jack Gallant (2016) is the following: "Brain is a hierarchically parallel distributed network of tightly interconnected areas feeding forward and feeding backward information all over the place and we really have no concept how such a network should compute information." The experimentally established computational brain theory has helped to consolidate some of the divided philosophical schools on the side of determinism (works such as Soon et al, 2008 and Fried et al., 2011), though still falling short of explicating how and where the experiences of consciousness, such as vision and thought, occur. The laborious work of Eric B, Baum entitled in" what is thought," published in 2004, exploits the computational brain theory to address the big questions of the mind, with consciousness among them: The following quotes: "…computer scientists are confident that thought, and for that matter life, arises from execution of a computer program;" and "Mind is flow of information, and consciousness is the experience of the information" capture some of his very valuable insights, expressed in the work. However, this substantial 4 work, relying on speculative evolutionary biologic claims, and computational principles, falls short of providing indisputable arguments towards achieving its very goal: Compatibilism shortcomings and La Mettrie type confusion persists; and "Cogito Ergo Sum", remains even more vague despite the heroic effort! Different schools of philosophy such as Materialism, Idealism, Dualism, and Panpsychism, have various palatable takes of the consciousness dilemma; and panpsychism (the belief that everything has a mind), holds that consciousness is an intrinsic property of matter. According to Stanford Philosophy Encyclopedia "the underlying premise in panpsychism, at its very microphysical levels, somehow, builds into animate beings' subjective consciousness experiences." However, all the ideas still remain speculative at their core! As a way out of the conundrum, Chalmers (1995) suggests a theory of consciousness that takes conscious experience as a fundamental property of the brain; and further claims that "we might explain familiar consciousness phenomena involving experience in terms of more basic principals involving experience and other entities." He asserts that taking experience, the inseparable feature of life, as an axiom, may provide the basis for a general theory of consciousness. To this end, Chalmers' speculative theory relies on (personal) subjective experience data and on the subjects' verbal reports relating to their experience (Chalmers, 1996). However, the thesis as skillfully as it is put forward, aligned partly with panpsychism philosophy, -- the latter unlike its past is being taken more seriously by other schools of philosophy -- similarly suffers from the arguable premises syndrome. As such, need for a robust theory that can address the mind and all its attributes, still persists -- a philosopher's stone to be found! Present work is an attempt in meeting the challenge by deploying the functional knowledge of the brain -- what facts the cognitive sciences have established so far -- in a radically different context: that of the all inclusive computational nervous (central and peripheral) system machinery, in the context of dynamic system theory; it is in such context that the development of a general inferential theory of the mind, with emphasis on consciousness, is aimed at: To this end two available works will be heavily drawn upon, namely, Schad (2016) and Schad (2016), where the natures of the brain computer and its dynamics; and how and where perceptions of thought, vision and other facets of mind occur, are theorized. The general framework of the approach has precedence in the field of cognitive psychology, in what is called the "Embodied brain" approach (see Kiverstein and Miller, 2015), which only serves to indicate the coming to terms of the "Field" with the role of the initial and boundary conditions; ("the skilled organism environment interactions"), – familiar to applied physic and engineering community -- on the cognition (computational) processes of the brain, implications of which is of importance to various aspects of human psychological mental states. In summary, the present work which is based on the scientific inferences drawn from the computational functioning of the central nervous system with brain at the helm, explores brain perception processes, on the basis of the nature of the brain computer and its mathematical computations formalism; and how they give rise to the consciousness; and other philosophically contentious sentience related phenomena -- such as free will, 5 nature of reality, and etc., in the context of a new understanding of the brain operations. Specifically, consciousness is reasoned to be the innate characteristic of all systems, animate or otherwise, which in case of higher beings, humans in particular, draw its essence autonomously from their brains. And the work well accords -- contrary to what philosopher D.C. Dennett (1996) suggests– with the take of Charles Darwin (1859), reflected in the following statement: "Nevertheless, the differences in mind between man and higher animals, as great as it is, certainly is one of degree and not kind." The Theory …" I will write about human beings as though I were concerned with lines and plains and solids" Baruch Spinoza (1632-1677 A.D.) "Life and soul are one, an animating and expansive force present in everything everywhere" Anaximenes (585-528 B.C.) All Beings animate or inanimate are dynamical systems that continuously interact with their (external/internal) environments through physiological or physical interfaces of their systems' (Kantian) realities. And the nature of their reactions (functions and interface expressions, which evolution deemed necessary), are determined by their systems' inner workings. Given the complexities of most systems, clear understanding of most systems' inner working details, is not generally possible. However, the system functional generalities which have already been established can provide the basis for the development of concepts from the perspective of the dynamical system theory; and such is the basis of the work presented here: In case of inanimate matter, the field equations (constitutive properties, and constitutive and conservation laws) generally allow (analytic/numerical) determination of their systems' reactions to the variations of the environmental conditions, and hence the resulting expressions (behavior). Their systems' physical changes (interface displays), that is, the expressions of their varying reactions, indicate the dynamics of their existence. For higher animate beings, emphatically humans, the systems' behaviors (functional operations) are governed by the mathematical formalisms, which must be (inherently) geared in their computational brains (Kandel, Schwartz and Jessell 2000), -- based on cognitive neurosciences understanding – though, the details, which would be the key to the development of a fundamental theory of the mind in general, and consciousness in particular, is not still known. And such details, if known, would encompass knowledge of the nature of the brain computer, i.e., what kind of computer it is; and what mathematical formalism underlies its operations -- considering the obvious complexities involved in reaching this end, it is not likely that any solid understanding of the dynamics of the brain computational operations will be established in the foreseeable future. However, as in all challenges sciences face, the immensity of the task, same also in this case, is not a barrier to a first order attempt of conceiving a plausible theory for the brain functional (computational) operations. The 6 the advanced intelligence levels of artificial present work embarks toward this goal by the presumption of likely semblance of the brain computer – from ground level operational perspective -- to those of the (brain inspired) scientific neural and neuronal networks (Edelman, 1987; Arbib 1989) – the similitude principal (Rayleigh, 1915) adds further credence to this extrapolation. The scientific neural and neuronal networks are well known statistical computational methods -- presently augmented by deep learning (supervised or reinforced) processes -- for the development of (AI). Grand multidisciplinary projects, such as Machine Intelligence from Cortical Networks (MCrONS), project (David Cox, Harvard University), are efforts to approach creation of human like intelligence. From the underlying presumptions in such endeavors, it further follows that the essence of the brain neuronal computation (solution) scheme, at its very fundamental levels, could be likened to that of the most basic scheme (implicitly) involved in the computational operations of the scientific neural networks. In such layout the brain and the rest of the nervous system, are posited to discern (resolve) any sense stimulating natural phenomena, -- that it to (implicitly) algorithmize them in the infinitude of the discrete synaptic nodal domain of the brain -- and to solve (mostly by trial-and error) the resulting equations (Shad, 2016), to render perceptions, which define some aspects of the mind and the consciousness: This presumed dynamics behind the operations of the central nervous system, would expectedly accord with the following premises:1) the autonomous data processing and computations in the brain, would provide possible solutions for various complexities, discerned in various states of their manifestations; 2), streaming downloads of the results via the nervous circuitry, would render expressions -- at the human interfaces-- of animation; functions such as biological sustenance, maintenance and behavior; and the awareness of sentience, and its associated phenomena, which define the consciousness phenomenon; and 3), expressions of consciousness would be limited due to the inevitable interface (frequency bandpass) filtering of the massive volume of streamed downloads of the results of the brain computations – such drawbacks are normally expected in any input/output system. Consciousness development relies on the simulations (computations) in the brain, which renders recreations of events, phenomena and the world, which all together make up the experiences of being. The simulations, of whatever nature, are most likely the processing and executions of the life span learned, and evolution hacked, neuronal ciphers (patterns, i.e., software and firmware), prompting potentiation, induced in turn by proper expression of genes, at synapses, -- beholden by some of the known and perhaps (98%) unknown (if they are not rubbish junk according to Brenner, 2013) segments of neural DNA -- rendering the excitatory and inhibitory tasks that produces what "is not a cause, it is an effect," as Dennett (20016) puts it. Some of results (outputs), which find syntactical expressions in thought, -- in the (possibly vast) lexicon of its language -- are not necessarily fully available for the efflorescence of talk -- as known; some are reportable in speech and loud thinking, and some as in feelings, emotions and other experiences, which are not satisfactorily reportable. Within the context of the system approach, it is mostly in thought and talks that consciousness debuts itself. Since thought is a tangible event on which humans have seemingly some controls over, the questions of where it occurs, perhaps is one that hits closer to home! Given that thought is the result (output) of the brain computations, there must be an interface where its expression 7 happen: only movable body parts (including facial muscles) and vocal cords are the apparent candidate interfaces to which somatic and visceral output (efferent) signals from the brain, reaches -- by means of motor sensory neurons -- displaying the features of consciousness. Such, has long been recognized as explicitly indicated in Confessions (St. Augustine, 397 AD): "And that they meant this thing and no other was plain from the motion of their body, the natural language, as it were, of all nations, expressed by the countenance, glances of the eye, gestures of the limbs, and tone of voice, indicating the affections of the mind, as it pursues, possesses, rejects, or shuns." The fact that thoughts are not always vocalized (reported), makes it possible to suggest – drawing upon Ockham's razor principal -- a silent muffled mode for the vocal system -- where the computed thought and thinking appear mainly inaudibly (other physiological displays aside), thus introducing a bi-modal utterance system (activated in either mode by a preamble signal), which act as the display medium; vocal mode for language (Chomsky, 2013) and subvocal mode for thought. Putting it succinctly: Utterance (vocal) Interface is the main venue for outputting brain's results of simulations of the real world and some aspects of interactions with it; in audible (referential or verbal) and mostly inaudible (thought) displays. And the fact that thoughts can always be verbalized amid thinking adds enough confidence to the above claim –- the latter can simply be tested. Another, and further, indication of it is the presence of subvocal activity during thinking that apparently behaviorist took note of long ago, and even went as far as to claim the possibility of decoding it: it is anecdotally reported by Will Durant (1991) in a quote from Spinoza: "Have not the behaviorist proposed to detect a man's thought by recording those involuntary vibrations in his vocal cord that seem to accompany all thinking." Subvocal Laryngeal (electromyography) recording has been deployed in psychiatric patients for clues for behavioral treatments of Hallucination (Green and Kinsbourne, 1998). Of course detecting (decoding) thoughts from subvocal activities is an enormous task involving stochastic/Neuronal, and more, along the line of the recent work by Nishimoto et al., (2011), which is aimed at "Reconstructing visual experiences from brain activity evoked by natural movies." Quoting Gallant (2016):" If there is something in emergent working cognitive memory space, potentially it is decodable information." Perhaps supplementing anatomic MRI (diffusion and functional) efforts, along with the very non-smoothed signals (as opposed to those of voxels) from the subvocal activity, should be a boon to semantic extraction that is pursued in decoding research. Besides thought, vision is the other very important phenomenon of consciousness: Other than attracting questions about its experience, the complexity of its unconscious development, has led to the general collective assumption of it being a fundamental property of Beings who have eyes; also seemingly, the knowledge of the anatomy and physics of vision's physiologic embodiment (Schwarz S.H. 1999), has served as the convincing rational for the assumption. However, vision, as in thought, begets the questions of how does one see what is seen and where it happens? Addressing the Tactile Vision (Bach-y-Rita 2006), and the Mirror Neuron (Keysers, et. al. 2004) phenomena, Schad (2016) theorizes that brain processes for vision sensation, are basically similar those of tactile sensations, except for involvements of more of the brain's neuronal 8 network and constructs (patterns), in the former. And seeing is but an autonomous recitation (reporting) of the sensation, at the utterance interface -- in the absence of such facility other bodily interfaces, as in case of other beings that lacks it, would perform the task. Also additional evidence pointing at the tactile nature of vision is the fMRI activity of parts of visual cortex during Brail reading by blind subjects (Lipman, 2014). Schad (2016) putts it summarily in the following:: The experience of vision is in reality just inaudible, and occasionally audible, recital (of its semantics) of the likes of reporting in case of massive cutaneous sensing, and apperception of the environment; and, it is the same, in essence, for all, blind and otherwise, which in the former is understandably drastically limited. Therefore, Vision thought, speech and what defines consciousness with all its bells and whistles, and everything that relates to the activities of the mind, are brain perceptions, -- being results of execution of brain programs -- which are broadcasted as expression on physiologic interfaces. The reasoning so far, having laid out a possible rational for how computational brain and the rest of the nervous system-- in the context of dynamic system theory – can account for many aspects of the mind in general, and processes of consciousness, in particular, -- of how and where they materialize, and also provides a plausible logic for resolving the hard problem of consciousness; its subjective aspect, as put by Chalmers (1995): "A phenomena which is physically based and yet not explainable by it." In the context of the present work, the hard problem finds the following simple explanation: Subjective consciousness is the thought expressions of the streaming downloads of the constructed brain perceptions (concomitant with memory registrations), which may only be partially reportable (i.e., some not utterable); due to the complexity of the contents, -- perhaps because of the richness of its lexicon -- and the bandpass limitation of the vocal cords, which could filter them. At much simpler levels, the inability of verbalization in reproducing of some natural sounds one hears is well known in all languages. This vocal reporting shortcoming could have very likely been sanctioned by the evolutionary processes -- perhaps because reporting to other fellow humans of "what is it to be me," or of "the color perceptions," has had no survival value, at least in the eye blink of time since our appearance on this planet. Despite the complexity of the environmental interactions of animate matter, and the innate simplicity of the in animate matter, the concept of consciousness can be generalized, to both, from the perspective of dynamic system theory in that they all react (respond) to the environmental inputs. This common characteristic, this intrinsic property of all objects with spatial representation, may be referred to as the objective consciousness; and the nature of the reaction which separates animate from inanimate, designates their subjective consciousness < Fig 1>. 9 Dynamic System Theory Rendition of Matter Consciousness Boundary Condition ConditionsConditions Initial Conditions Conditions MATTER (Ultimate Standard Model) OBJECTIVE CONSCIOUSNESS (Physical, Physiological and or Mental Display) SUBJECTIVE CONSCIOUSNESS (Nature of Display) ANIMATE MATTER (Mental & Physiological) INANIMATE MATTER (Physical) MENTAL EXPERIENCES Hard Problem PHYSIOLOG EXPERIENCES Easy Problem Figure 1- Matter consciousness depiction 10 to This generalized concept of consciousness, accords partly with panpsychism claim; an idea that perhaps has roots in very early thinking: the following quote from Aristotle (Durant, 1991), well speaks the concept of universality of consciousness, notwithstanding the obvious error: "Nature makes so gradual a transition from the inanimate to animate kingdom that the boundary lines which separates them are indistinct and doubtful." The generalized consciousness theory is anecdotally evinced in the claims of the experiences of unison with nature by subjects in (authority sanctioned) hallucinogenic drug experiments; and by practitioners of intense meditation. In both cases, subjects' brain circuitry gets extensively engaged (chemically affected or overwhelmed), –- evinced by multichannel EEG data (Nader, 2014), and fMRI (NPR's Radiolab report on LSD experiments (2016), and Marina, 1999). And during this process a "hang" situation is very likely occurring in the brain operations: most perception computation come to a halt -- at least in case of meditation with a nonsensical mantra (irresolvable problem), it is likely that the futile brain attempt for solution heavily taxes most of the brain circuitry. Therefore, in such state, output streaming become very limited to a great extent. And of course a time-lapse memory track of the period is registered: perhaps a record of what the universal objective consciousness engenders; the effects of the matter world interaction and perhaps even microphysical entanglement with the environment, when normal sense interactions in the context of physiologic animation are gravely suppressed, or altogether are absent. Obviously, the (ever present unconscious computation operations) circuitry for biological sustenance is not affected during such experiences. Finally, this perspective of consciousness adds much credibility to panpsychism philosophy; perhaps their philosopher's stone is found! Also this work provides the proper ground for establishing the idea of Man Machine. The concept has its roots in Descartes (1960), and later taken up by La Mettrie, in his "L'Homme Machine (1747)", -- understandably facing insurmountable difficulties -- and shared vigorously by Schopenhauer (1855), and by many (to some degree) in recent years (e.g. Baum (2004); Mlodinow 2011). However, in the context of the present work, the idea proves seemingly very plausible, since it considers higher beings as biological dynamic systems -- with brains (the puppeteer, according to Chomsky, 2016) and the rest of the nervous system, as the control system-- with physiologic mouthpieces, which simply broadcasts their presence. Following quote from the Nobel Laureate Sydney Brenner (Woodham 2014), who in a recorded gathering of scientist, puts the overall claim in the proper context: "1) How do the genes specify and build a machine that performs the behavior, and 2) how does the machine perform the behavior? The answer to the first one is we do not know, but the answer to the second one is that it would depend on the queued memory and boundary condition, like any readymade machine." The following statement by Philosopher David Hume (1739) sums up the sense of being in the followings: "We are nothing but a bundle or collection of perceptions which succeed each other with inconceivable rapidity and are in perpetual flux and movement" Experimental Support of the Theory 11 This proposed approach to the understanding of the human sentience and other facts of life in the context of the dynamic system theory, is partially, though not conclusively, supported by 1), the experimental works of Soon et al, 2008 and Fried et al., 2011, in addressing the experience of will: the latter research summarize its findings, as "… that the experience of will emerges as the culmination of premotor activity (probably in combination with networks in parietal cortex) starting several hundreds of ms before awareness," which purports to the underlying essence of the theory; and 2), the result of the analysis of multichannel EEG recordings of subjects during transcendental meditation experiments (Stanford Higher State Lectures), which verifies the reported claims of absence of space and time and body sense, by the coherent Alpha waves; seemingly a no download episode in wakeful healthy subjects, while the brain is at full computation capacity, -- resulting from irresolvability of the submitted problem (a nonsensical Mantra) -- causes a "hang" state, when much of the characteristics of sentience disappears. Conclusion The computational theory of the brain has been deployed to find answers to some of the long held major questions of philosophy: To this end, three propositions were put forward: 1) that Given the theory, it is the computational outputs of the brain which are relayed through motorsensory neurons to the body's physiological interfaces, which render animation and, in case of many beings, vocalizations, thought, vision, and other effects, together defining the phenomenon of consciousness; and, 2), that, in the case of higher beings, it is the vocal interface, referred to here as utterance interface, which is the main medium of expressions of perceptions that broadcasts the conscious mental states in bi-modal, audible and inaudible, modes of the vocal box.; and 3), that brain is in essence an equation solver, which discerns being's dynamic environment (as sets of parametric linear equations), through stimulation of body's senses, and solves them (by trial-and- error); and outputs the results as expressions at body's extremities. And that through heredity and learning brain engenders many such equations as readymade patterns (neuronal constructs), available for immediate or fast solutions of discerned problems -- in the likeness of today's deep learning (supervised or reinforced) in Artificial Intelligence developments. The computational brain theory is further engaged to infer that, 1) despite the complex functional operations of the central nervous system, which render (animate) consciousness (other physiologic activities aside), all matters, animate or inanimate, in the paradigm of the input output systems, have consciousness; and 2), that the mere fact of interaction with the environment, defines its objective aspect; and what make appearance at the interfaces (the expression), determines its subjective aspects, which is a function of the matter's inner workings (physical, or biophysical governing rules); while limited by the medium of its broadcast. In this light, the question of consciousness of all higher beings is also settled. The imbedded consistency of the approach in the analysis of the nature of consciousness of higher animate matter (much explored by Baum 2004), ), also allows for all aspects of Qualia (regardless of different philosophical takes), as well as providing inroads for all 12 big questions of philosophy. Finally, the reasoning for the concept of universality of consciousness, which also accords with Yogi's claims (based on their repeatable experiences during intensive TM meditation), supports, as well, the main axiom of the Panpsychism theory, and provides philosophy with grounds for unarguable premises. References Arbib A. M. (1989) The metaphorical brain 2, Neural network and beyond. New York: John Wiley & Sons Beck J. S. and Molish B. H. (1959) Reflexes To Intelligence, A reader in clinical psychology; Free Press: Glencoe Illinois Bach-y-Rita (2006) Tactile substitution studies. Annals of New York Academy of Sciences 1013, pp. 83-91 Baum B. E. (2004) What is thought?. Cambridge: A Bradford Book, MIT Press Bourget D. and Chalmers J. D. (2013) What Do Philosophers Believe? Philpapers Survey Broad, C. D. (1978) Kant: an Introduction. Cambridge: Cambridge University Press Chalmers D. J. (1995) Facing Up to the Problems of Consciousness, Journal of Consciousness studies, 2(3), pp. 200-219 Chalmers D. J. (1996) The Conscious Mind: In Search of a Fundamental theory, Oxford University Press Chalmers D. J. (1997) Moving Forward on the Problem of Consciousness, Journal of consciousness Studies, 4(1), pp. 3-46 Chalmers D. J. (2014) The Hard Problem of Consciousness, 342 yeas on; 20th Anniversary Conference, Towards a Science of Consciousness. Tucson Arizona Chomsky N. (2007) On language. New York: New York Press Chomsky N. (2012) Language and Other Cognitive Processes, Recorded talk, WGBH Forum (online) https://www.youtube.com/watch?v=6i_W6Afed2k [Nov. 2014] Chomsky & Krauss (2015) An Origin Projects Dialogue. [Online] ,https://www.youtube.com/watch?v=uBVb6wRdwV4[Jan 2016] Chun Siong Soon; Marcel Brass; Hans-Jochen Heinze & John-Dylan Haynes (2008) Unconscious Determinants of Free Decisions in the Human Brain, Nature Neuroscience, 11, pp. 543 - 545 13 Itzhak; Mukamel, Roy; Kreiman, Gabriel Church A. (1936) An Unsolvable Problem of Elementary Number Theory, Journal of Symbolic Logic, 4. pp. 53–60 Churchland P. S. and T. J. Seinowski (1992) The computational brain. Computational Neuroscience, Cambridge MA: The MIT Press Darwin, C. (1859) The origin of species, New York: P. F. Collier & Sons Company Descartes, Rene (1960) Discourse on Method (1637), trans. Laurence J. LaFleur, 3d ed. New York: Bobs Merrill Durant, W. (1991) The Story of Philosophy. Simon and Schuster: New York Edelman, G. (1987) Neural Darwinism: the Theory of Neuronal Group Selection, New York: Basic book (2011) Internally Generated Fried, Preactivation of Single Neurons in Human Medial Frontal Cortex Predicts Volition," Neuron. 69 (3), pp. 548–62 Gallant J. (2016) Mapping, modeling and decoding the human brain, [Online],https://www.youtube.com/watch?v=pzxQbdUryHU[Dec. 2016] Green M.F. and M Kinsbourne (1990) subvocal activity and auditory hallucinations: Clues for behavioral treatments? Schizophrenia Bulletin,16 (4), pp. 617-625 Haldane B. R. and Kemp J. (1908) Schopenhauer A. (trans.). The World As Will And Idea; Kegan Paul, Trench, Trubner &Co Hoffman D. (2014) Consciousness Agents: A Formal Theory of Consciousness; 20th Anniversary Conference, Towards a Science of Consciousness; Tucson Arizona, April 21-26 Hume David (1739) A Treatise of Human Nature. John Noon: London Kandel R. C. Schwartz H. J. Jessell T. M. (2000) Principles of. Neurosciences. New York: McGraw-Hill Kandel R. E. (2013). The New Science of the Mind, Grey Matter, The New York Times, Sunday Review, Opinion Page Dec. 1st Kant I.. (1902). Critique of Pure Reason, New York: Macmillan 14 Keysers C., Wicker B., Gazzola V., Anton J-L., Fogassi, L. Gallese V. (2004) A touching sight SII/PV activation during the observation and experience of touch. Neuron. 42 (2), pp. 335-346 Kiverstein J. and Miller M. (2015).. The embodied brain: toward a radical embodied cognitive neuroscience. Front Hum. Neurosci.. 9, pp. 237-247 Lipman J. (2014) Stanford Lectures: Higher States, [Online], https://www.youtube.com/watch?v=FysEMS1rgrw Mariña J. (1999) LSD Experiments, The Journal of Religion 79 (2), pp. 193-215 McCulloch S. W. (1947) How We Know Universals: The Perception of Auditory and Visual Forms, Bulletin of Mathematical Biophysics, 9, pp.127 Metzner, Ralph (2008). The Expansion of Consciousness. Berkeley, CA: Green Earth Foundation & Regent Press Mlodinow L. (2011). Subliminal: How your unconscious minds rules your behavior, First Vintage Books: New York Nader T. (2014) Stanford lectures: Higher States [Online], https://www.youtube.com/watch?v=PE0ju4_y5R8 Penrose R. (1990) Emperor's new Mind. Oxford: Oxford University Press Schad J. N. (2016). Brain Neurological Constructs: The neuronal Computational Schemes for Resolutions of Life Complexities, J. Neurol Neurophysiol., 7 (1) Schad J. (2016) Neurological Natures of Thought and Vision, and Mechanisms of Perception Experiences, Neurol Stroke, 4 (5) Schwarz S. H. (1999), Visual Perception, McGraw Hill: New York Shinji Nishimoto; An T. Vu; Thomas Naselaris; Yuval Benjamini; Bin Yu; Jack L. Gallant (2011), Current Biology, 21 (19), pp. 1641-1646 Turing Alan (1948), Machine Intelligence, in Copeland, B. Jack (eds.), The Essential Turing: The ideas that gave birth to the computer age, Oxford: Oxford University Press, ISBN 0-19-825080-0 Vartanian A. (1960) (tran.), La Mettrie's L'Homme Machine (1748): A Study in the Origins of an Idea, Princeton University Press, 15 Von Neumann J. 1958. The Computer & The Brain, Yale University Press: London Woodham B. (2014), Understanding and Designing Cognitive Systems. Lecture Notes 2013/2014 Term 2, Department of Computer Science, University of British Columbia : January–April, 2014 . Acknowledgement Thanks are due to Dr. A. Hindi, for his enlightening insights and discussions, and to Professor S. Hedayat, for his review of the manuscript and many helpful suggestions. 16
1002.4368
1
1002
2010-02-23T18:25:00
On the scaling law for cortical magnification factor
[ "q-bio.NC" ]
Primate visual system samples different parts of the world unevenly. The part of the visual scene corresponding to the eye center is represented densely, while away from the center the sampling becomes progressively sparser. Such distribution allows a more effective use of the limited transfer rate of the optic nerve, since animals can aim area centralis (AC) at the relevant position in the scene by performing saccadic eye movements. To locate a new saccade target the animal has to sample the corresponding region of the visual scene, away from AC. In this work we derive the sampling density away from AC, which optimizes the trajectory of saccadic eye movements. We obtain the scaling law for the sampling density as a function of eccentricity, which results from the evolutionary pressure to locate the target in the shortest time under the constraint of limited transfer rate of the optic nerve. In case of very small AC the visual scene is optimally represented by logarithmic conformal mapping, in which geometrically similar circular bands around AC are equally represented by the visual system. We also obtain corrections to the logarithmic scaling for the case of a larger AC and compare them to experimental findings.
q-bio.NC
q-bio
1 On the scaling law for cortical magnification factor. Alexei A. Koulakov Cold Spring Harbor Laboratory, Cold Spring Harbor, NY 11724 Primat e visual syst em samples d ifferent parts of the world unevenly. The part of the visual scene corresponding to th e eye center is represented densely, while away from the center the sampl ing becom es progressively sparser. Such distribution allows a m ore effective use of the l im ited t ransfer rate of the optic nerve, since anim als can aim area cent ralis (AC) at the relevant posit ion in the scene by per form ing saccad ic eye movem ents. To locate a new saccad e target the an imal has to sample the corresponding region of the visual scene, away from AC. In this work we der ive th e sam pling density away from AC, which optim izes the trajectory of saccad ic eye m ovem ents. W e obtain the scaling law for the sampl ing density as a function of eccent ricity, which resu lts from the evolut ionary pressure t o locate the target in the shortest time under the constraint of lim ited t ransfer rate of the opt ic nerve. In case of very sm all AC the visua l scene is opt imal ly represented by logarithm ic conform al m apping, in which geom etrically sim ilar circu lar bands aroun d AC are equally represented by the visual system . W e also obtain corrections to the logarithm ic scaling for the case of a fin ite AC and com pare them to experim ental find ings. The degree of sampling by the visual system is described by areal cortical magnificat ion factor (M). It specifies the area in cortex (in mm2 ), which represents a unit square of the image ( in deg2) (Daniel and Whitteridge, 1961) . In primates the magnification factor is large in the AC and decays away from AC as a power law of eccentricity (E), measured in degrees (1) 2EAM  Here A is a constant , and α ≈ 1 is the scaling exponent determined from experiment (Van Essen et al., 1984). This paper is the first theoret ical at tempt to explain why the scaling exponent α has this value, and why in some cases it deviates from it . An alternat ive measure of sampling is the density of ret inal ganglion cells (RGC). It is natural therefore that magnification factor is related to RGC density. Away from AC the areal magnification factor M and the density of ganglion cells ng are proport ional, i.e. n g  2EB (2) where B is a constant , independent of eccentricity. The exponent α here is the same as in Eq. (1). This implies that each RGC projects to the same area in the pr imary visual cortex (V1). Since each RGC projects to higher visual centers, the total number of RGC determines the thickness of opt ic nerve. Because a thick optic nerve impedes eye movements, the total number of ganglion cells is subject to a constraint (Meister, 1996). We assume that the total number of GC has reached its maximum possible value N, which does not substantially impair eye movements . Since this value is given by the integral of RGC density over the whole ret ina, it depends on two parameters in Eq. (2): B and α. Thus, this constraint does not allow calculating each of them individually: it only provides one cond ition on their combination (see Methods for more detail) . To determine both of the two parameters unambiguously one has to find another condition. Before formulating the other condition on B and α, we would like to provide an alternative motivat ion for the former anatomical constraint. Instead of f ixing the total number of RGC we could fix the total area of cortex, represented by the integral of the magnificat ion factor , given by Eq. (1) , over the retinal space. Since Eqs. (1) and (2) are propor tional the result ing anatomical constraints are equivalent . What is the additional condition, which can fix both α and B in Eq. (2) unambiguously? It is known that the oculomotor system in humans uses highly optimized strategy in the game of cricket, so that the cricket player’s eye movement strategy contributes to his skill i n the game (Land and McLeod, 2000). Eye movements also exhibit highly organized strategy during such everyday activities as tea making (Land et al., 1999) and driving (Land and Lee, 1994). Similarly, we suggest that the visual and oculomotor systems use mutually opt imized strategy, which allows animals to detect new targets in the fastest way, whereby resulting in a fit organism, successful in the course of evolut ion. Figure 1 Tra jectory of AC in the v isual space af ter an object suddenly appears at the target locat ion. One primary and two corrective saccades are shown. The lighter gray disc schematically shows a region where the pr imary saccade lands in many t rials due to a finite sampling density . The darker gray disk shows precision of the second saccade. Consider the follow ing basic task cooperat ively performed by the visual and oculomotor systems (see Figure 1). Assume that the animal’s AC is at certain posit ion in the visual scene (or iginal eye position) . Then a new object (target) suddenly appears in the visual scene. We assume that the object is point -like for simplicity. The goal of the animal is to aim AC at the object for correct ident ification, performing saccadic eye movements. However, this cannot be done in one take, due to a sparse sampling on the periphery of visua l scene. Indeed both visual acuity, represented by minimum angular resolution Δ, and saccade precis ion depend on eccentricity as Δ=CEβ, β ≈ 1 (Weymouth, 1958; McKee and Nakayama, 1984). Thus, the saccade precision gets better when the target is closer to the AC, roughly linearly with eccentricity. After the first unsuccessful at tempt to aim AC at the target the animal has much better chance with the second saccade, since the eccentr icity of the target has decreased. The process repeats iterat ively, unti l AC is at the target. 2 Compar ing the sampling densit y, given by Eqs. (1) and (2) to the expression for angular resolution we notice that they are controlled by the same exponents α ≈ β ≈ 1. This implies that both minimum angular resolution and saccadic precision scale as distance between gn/1 nearest ganglion cells : . Thus we can use the same exponent α to describe both the anatomical cons traint derived from Eq. (1) and the iterative saccadic process shown in Figure 1. That we use the same scaling exponents for both saccadic precision (β) and magnif ication factor (α) implies that the same fraction of visual information flaw is dedicated to establishing the correct saccadic target locat ions for every target eccentr icity. This assumpt ion is based on the general uniformity of the visual system and is confirmed by the approximate equality between two scaling exponents, α and β. We then minimize the total durat ion of the iterative process with respect to exponent α, subject to the anatomical constraint. In doing so we assume that saccades occur very fast and most of the t ime is consumed by a saccade preparat ion. Thus, we minimize the total number of saccades averaged over all possible target locations. a b Figure 2 a The total number of saccades as a funct ion of the scaling exponent. Parameters used are: N=600 and . The optimum is 01.0 shown by the circle. The optimum value of in which case the 9.0opt average number of saccades is about 3. b The optimum scaling exponent as a funct ion of parameter  . The solid line is the resu lt of our theory for N=600, and the dashed l ine is given by Eq. (3). The markers show experimental results for different species (see list in Methods) This average number of saccades is shown in Figure 2a as a funct ion of scaling exponent α. As seen from the figure the number of saccades diverges when α approaches 1.5. This is because in case of large α most of the ganglion cells are located near AC and the periphery of visual scene is undersampled. In this condit ion the objects on the per iphery cannot be located reliably and the described iterative process includes a very large number of saccades. In the  opt 2   (3) 1  other extreme, when α ≈ 0.5, the per iphery is well represented, but an object's locat ion cannot be pinpointed exactly due to poor resolution near AC . The average saccade number has a minimum at αopt ≈ 0.9, marked by a dot. Our theory has two dimensionless parameters (see Methods) . The first parameter descr ibes the total number of “saccade pixels” in the visual scene, determined by the saccade precision. Since saccade precision is about 20% in humans, this parameter N is about 600, as estimated in Methods. This parameter descr ibes the anatomical constraint and is controlled by the total number of ganglion cel ls. The second parameter is η = r/R << 1, where r ≈ 1º is the relative radius of AC and R ≈ 90º, is a measure of inhomogeneity introduced into the sampling density by the presence of AC. It therefore quantifies the finite size effects produced by AC. We derive expression for the value of α which optimizes the average saccade number for the most interesting case η << 1 g 1 3 1ln       g 1ln 2 2 ln 1ln     Here we introduced g=N/2π ≈ 100 for brevity. The limit η → 0 corresponds a very small AC. When AC size (parameter η) becomes small, the second term in this expression becomes small too, and the opt imum value of the exponent α approaches 1. This behavior is i n agreement w ith the experimental observation of α ≈ 1. The second term in Eq. (3) describes the impact of f inite (non zero) size of AC. I n the simple theory presented here the correction is always negative. The comparison between theoretical and experimental results for some anima ls with finite AC is presented in Figure 2. The value α = 1 corresponds to the logar ithmic mapping of the visual scene. In this map the cor tical posit ion corresponding to eccentr icity E is equal to ln E. The linear magnification factor resulting from the logarithmic sampling decays as . The linear d E dE E ln / 1 / magnification factor for the second angular coordinate can be predicted from the assumption of conformal mapping, i.e. that small circles in the visual wor ld are represented by circles in the brain and, therefore, the shapes of the objects are preserved by the brain map. The second linear magnification factor is therefore bound to be 1 / E too, leading to the areal magnificat ion factor M proportional to 1/ E2. This reasoning results in α = 1 in Eq. (1). The conformal logarithmic mapping is optimal because the information lear ned about an object 's location is equal between different saccades invo lved in target ing the object. In conclusion, we suggest that distr ibution of sampling density existing in many primates is optimum for fast peripheral target location. The opt imum is found under the constraint of limited transfer rate thorough the optic nerve. Methods Finding the opt imum scaling exponent We describe finding a new ta rget by an iterative proce ss, in which the length of the subsequent saccade is dete rmined by the preci sion of the previous one: l l ( )  . The expre ssion for the dependence of saccadic precision on n n 1 C n l l )( )( /   eccentricity is related to the gangl ion cell densi ty: g 1C determine s the fract ion of the vi sual bandwidth used to saccadic target. Using Eq. (2) we obtain: , C~ ~ BCl lC /   has the meaning of . Since 1 where locate l )(   2/1  ln  , 3 1 n  , n  1 l n 1   1   ~ C relative saccade precision and i s therefore about 0.2. The ite rative sequence ~ lC l  1  can be solved exact ly to give n ~/ Cl 0       ,...2,1,0n 0l is the length of the fi rst saccade, which i s where and approximately equal to the eccentricity of the target . The numbe r of saccades r l n  , whe re needed to put fovea on the target i s determined by a ssigning r is the radius of fovea. Averaging thi s expre ssion over the target position between r and R , which is the radius of retina in degrees, we obtain 0l lr ln 1 lR ln ln  (M1) ln n  . f f Here l f  Nr        Rr 1   2   1     1  1 a  2  1 , where  (M2) R 2  N ln ~ r C 2 is the total number of saccadic “pixel s”, which is kept fixed during the optimization proce ss. It represents a fraction of the total visual bandwidth, which can be used for locating targets, due to complexi ty of the task. The n  (n ) funct ion i s shown in Figure 2a. To derive Eq. (3) we expand Eq. 1 to the second degree of paramete r (4) in a Taylor series around point  n . We therefore find the parabol ic approximation of there. 1 The minimum of the re sult ing parabola is easi ly found, which results in Eq. (3). Saccadic precision To obtain a realist ic e stimate for the saccadic precision one should mea sure saccadic errors in the natural condi tions, in which animals compete for survival, such a s pre sence of di stracters, noi sy backgr ound, weak targe t luminance etc. To the be st of our knowledge such measurement s have not been done. We estimate the lowe r bound for the saccadic errors from the measurement s done in laboratory condit ions. To mimic natura l unpredictability of the ta rget location one ha s to use experimental set up, in which both target eccentricity and direction are random and vary in a wi de range . Such mea surement s have bee n done for huma n subject s (Deubel, 1985). The saccadic precision, which f ollows fr om this study, is about 20% of the targe t eccentricity (see Figure 3). This estimate is different from other est imates of about 10% precision (Becker and Fuch, 1969), due to difference in the task (random di rections). We therefore adopt the following e st imate to E E ) ( 2.0   saccadic prec ision . The number of saccadic “pixels” N 700 which follows from thi s estimate [see Eq. ()] is . a b Species Galago1 Human2 Flying fox3 Owl monkey4 Cebus5 r [deg] R [deg] 2 1 10 2 1 90 90 110 100 90 r R  0.022 0.011 0.09 0.02 0.011  0.8 0.96 0.53 0.7 0.94 histogram. The mean saccadic error, determining the saccadic precision is 17% of eccentricity. The data are from Deubel (1985). Comparison of the theory to primate data We use the fol lowing paramete rs for different primate species in Fig. 2b. _______________________________________________________ Deubel H (1985) Adaptivity of gain and direction in oblique saccades. In: Eye Movements: from physiology to cognition, pp 181-190. New York: Elesevier . Fr itsches KA, Rosa MG (1996) Visuotopic organisation of str iate cortex in the marmoset monkey (Callithr ix jacchus). J Comp Neurol 372:264-82. Gattass R, Sousa AP, Rosa MG (1987) V isual topography of V1 i n the Cebus monkey. J Comp Neurol 259:529-48. Kowler E, Blaser E (1995) The accuracy and precision of saccades to small and large targets. Vision Res 35:1741-54. Land MF, Lee DN (1994) Where we look when we s teer. Nature 369:742-4. McKee SP, Nakayama K (1984) The detect ion of motion in the peripheral visual field. V ision Res 24:25-32. Meister M (1996) Multineuronal codes in ret inal signa ling. Proc Natl Acad Sci U S A 93:609-14. O'Regan JK, Levy-Schoen A (1987) Eye movements: from physiology to cognition : selected/edited proceedings of the Third European Conference on Eye Movements, Dourdan, France, September 1985. Amsterdam ; New York, New York, N.Y ., U.S.A.: North-Holland; Sole distributors for the U.S.A. and Canada Elsevier Science Pub. Co. Rodieck RW (1998) The first steps in seeing. Sunderland, MA: Sinauer Associates. Rosa MG, Schmid LM, Krubitzer LA, Pett igrew JD (1993) Retinotopic organization of the primary visual cortex of flying foxes (Pteropus poliocephalus and Pteropus scapulatus). J Comp Neurol 335:55-72. Rosa MG, Casagrande VA, Preuss T, Kaas JH (1997) V isual field representation in striate and prestriate cort ices of a prosimian primate (Galago garnett i). J Neurophysiol 77:3193-217. Rumberger A, Tyler CJ , Lund JS (2001) Intra- and inter -areal connections between the primary visual cor tex V1. Neuroscience 102:35-52. Schwartz EL (1977) Spatial mapping in the primate sensory projection: analytic structure and relevance to perception. B iol Cybern 25:181-94. Figure 3 a Var ious posit ions at which saccades of a human subject landed (dots) relative to the target. The target is shown by a circle with a cross, red lines indicate saccadic errors. b The saccadic errors from a are shown in the 1 Rosa et . al ., 1997 2 Weymouth, 1958; McKee and Nakayama, 1984 3 Rosa et . al ., 1993 4 Silveira et. al., 1993 5 Gattass et. al., 1987 4 Schwartz EL (1980) Computat ional anatomy and functional architecture of striate cortex: a spatial mapping approach to perceptual coding. V ision Res 20:645-69. Silveira LC, Perry VH, Yamada ES (1993) The retinal gangl ion cell distr ibution and the representation of the visual f ield in area 17 of the owl monkey, Aotus tr ivirgatus. Vis Neurosci 10:887-97. Van Essen DC, Newsome WT, Maunsell JH (1984) The visual field representat ion in str iate cortex of the macaque monkey. Vision Res 24:429-48. Weysmouth FW (1958) V isual sensory units and the minimal ang le of resolution. American Journal of Ophtalmology 46:102-113.