Datasets:

nopperl

/

arxiv-image-text

like 3

ab2e981ff74e4c4ea4bfd23bb719e963

Self-bias behaviour and reverse breakdown of the diode.

2012.15615

multi_sine_2.jpg

cadd08ee40314445984252b598212997

A multi-user multi-antenna wireless power transfer system.

2012.15615

new_sm8.jpg

5212dc9d756e412bb5c1975f8c154651

Output DC power versus maximum transmit power when N=32.

2012.15615

power3.jpg

25d60d70f7f44056bcc801e18f5b210c

Sum output power versus ET transmit power with K=2, U=16, and N=8.

2012.15615

power_sat.jpg

2911ac699d4e4220b95de8aabd84ecaa

Output DC power versus number of frequency tones when P_T = 10 W.

2012.15615

tones3.jpg

29c6350b08c344c5ac9420761318ff0b

Typical saliency maps of the listed methods. Data samples are from PASCAL VOC and the pretrained model is VGG16.

2012.15616

faithfulness_example.jpg

82f5eeffe6914047b9a87f3b1e41cbb0

Faithfulness results with iAUC metric across VOC and COCO datasets, with the best methods highlighted. Different saliency methods are represented by different colors. The number above every bar is the overall iAUC.

2012.15616

faithfulness_results.jpg

9e949dc85ad643fc851f779ca647bbc1

Saliency maps with and without CF. "Airplane" is the CF in scene classification, which should be assigned lower salient scores. The top and the bottom figures demonstrate the results of Resnet50 and VGG16 respectively.

2012.15616

fpr1.jpg

43a12371690043e6b71010f0f1067989

Overall results for localization (IoSR).

2012.15616

localization_results_ios.jpg

004b5d88330d4482b7661d5775f21175

Localization overall results with PG across VOC and COCO datasets with two pretrained models. The best performed saliency methods are highlighted. The number above every bar is the localization result based on PG.

2012.15616

localization_results_pg.jpg

c63dcef5b14b4291a41a65805c7c3b3e

Saliency maps of the same synthetic images explained on object classifier f_o and scene classifier f_s. Objects in the images are expected to be important for f_o but irrelevant for f_s. The title above original images indicates the scene and object. The values below the saliency maps are MCR scores with MCS in parenthesis. The top and the bottom figures demonstrate the results of Resnet50 and VGG16 respectively.

2012.15616

mcs.jpg

3b3150f053644d0a9ea4f72dc01170f8

Stability results from SENS_max in small perturbation region with perturbation radius r=0.2 across two different datasets and two pretrained models. The results are averaged across 50 randomly selected examples from both datasets.

2012.15616

robustness.jpg

b20cc44b5b904e99b99c1ff8f626a500

Modeling the Doppler effect for product sampling. (A) Timing profiles of our REMPI laser pulses. (B) The Doppler factor F_Doppler(v) versus the velocity of the K_2 product. The lower and upper bounds correspond to the situations with the peak value of Γ_ion/2π at 6 and 14 MHz, respectively.

2012.15842

figure_Doppler.jpg

526e6a9e8bbb4ec0b7027e63f55ee370

Energetics and product quantum states for ultracold reactions between KRb molecules. (A) Schematic of the potential energy surface for the reaction. Reactant KRb molecules are prepared in the rovibrational ground states |ν = 0, N = 0 ⟩, and the K_2 and Rb_2 products emerge in states | ν = 0, N_K_2⟩ and | ν = 0, N_Rb_2⟩. (Insets) Angular momentum vectors of the reactants and products. Two ground-state KRb molecules collide via p-wave collision, giving the system a total angular momentum of J_tot = 1. (B) Product rotational state-pairs and their degeneracies. The light red shaded region represents state-pairs that satisfy energy conservation, while the dark red squares represent those that additionally satisfy the parity constraint imposed by the exchange statistics of identical product nuclei. The number superimposed over each allowed state-pair represents its degeneracy, which is used to construct the state-counting model.

2012.15842

figure_Schematic_v4.jpg

92547bf9e48a46b990f1e4a44d30501d

Influence of the long-range potential on product formation near the energy threshold. (A) Ab initio effective long-range potentials for different scattering channels belonging to the state-pair |12,7⟩. Each channel is defined by the quantum numbers N_prod and L_prod, associated with the coupled rotational (N⃗_prod = N⃗_K_2 + N⃗_Rb_2) and orbital (L⃗_prod) angular momenta of the products, respectively. (B) Total transmission probability (black curve) through the centrifugal barriers and its contributions from different channels (curves of other colors) as functions of the translational energy of |12,7⟩. The color-coding scheme follows that of (A). The maximum contribution of each channel is given by the inverse of the degeneracy of this state-pair, 1/23 = 0.435, and is reached when the translational energy increases sufficiently above the corresponding barrier height. We use the ratio between the measured and predicted scattering probabilities (Fig. <ref>H) of 0.046 as an estimate for the transmission probability, and allow variations of +0.046 and -0.023 to account for uncertainties in the population prior to barrier transmission. This corresponds to a translational energy of 1.3_-0.5^+0.7× 10^-3 cm^-1 (magenta lines). This value is also indicated in (A). (C,D) Potentials for the state-pairs |4, 19⟩ and |10, 13⟩. The magenta line in each sub-figure indicates the translational energy of the corresponding state-pair, which is calculated based on the reaction exoergicity determined in this study. L_min and L_max respectively represent the minimum and maximum allowed orbital angular momentum for the state-pair.

2012.15842

figure_centrifugal_barrier_effect_v3.jpg

a2c8a6d4596a4ae18c4cf54820485167

State-resolved coincidence detection of reaction products. (A) Schematic of the ionization and detection process. Products in rotational states N_K_2 and N_Rb_2 are simultaneously ionized via resonance-enhanced multiphoton ionization (REMPI) pulses. The resulting ions are then velocity-map imaged onto a microchannel plate (MCP) detector. The position and TOF information of the ions are recorded and used to screen for coincidence counts. We apply a 17 V/cm electric field (E) for ion extraction and a 30 G magnetic field (B) for maintaining the nuclear spin quantization. (B) Momentum image of simultaneously detected K_2^+ and Rb_2^+ ions associated with the states N_K_2 = 6 and N_Rb_2 = 7. Each simultaneously detected ion pair is connected by a line. This image is derived from the raw image of ion impact positions via the position-momentum relations described in SM. The dashed circle corresponds to the maximum achievable transverse momentum for products in |6,7 ⟩. (C) Momentum image of the coincident product ion pairs, obtained by screening for the ion pairs in (B) that satisfy momentum conservation. (D) The number of K_2^+–Rb_2^+ pairs as a function of the number of ionization pulses by which their detections are separated. Zero on the horizontal axis corresponds to simultaneous counts, which contains both correlated (coincident) and uncorrelated (accidental) ions, while a non-zero difference correspond to ions generated by separate pulses, and are therefore uncorrelated. (E) A plot similar to (D), but with the uncorrelated counts screened away.

2012.15842

figure_detection_scheme_v7.jpg

7be8de2fda4a4b17b3ff09184c91e2d0

Modeling the geometric factor for product sampling. The plot shows the fraction of product pairs that remain within the REMPI beam volume at the time of ionization as a function of the velocity of the K_2 product within the pair. Blue and red dashed lines indicate the maximum velocities of the K_2 and Rb_2 products, respectively. (inset) the ionization geometry.

2012.15842

figure_geometric_factor.jpg

e62a1392af6842db91d3807f5355e15b

Timing diagram for product ionization. The relative timing between the ODT, REMPI, and cleanup pulses during the state-selective ionization of reaction products. (Inset) A close up view of a single modulation period. Unperturbed reactions occur during the dark phase of the period, while reactions influenced by the (1064 nm) ODT light occur during the bright phase. The numbers in parentheses indicate pulse durations.

2012.15842

figure_ionization_detection_timing.jpg

0959fc954eae42efa36afc4eba997645

Distribution of product translational energy. The measured (blue circle) and predicted (red diamond) scattering probabilities for all allowed state-pairs are plotted versus their translational energies. The two sets of data are offset horizontally by 0.014 cm^-1 for clarity. To aid in the identification of systematic deviations, we multiply each scattering probability by a normalized Gaussian function with a 1σ width of 0.25 cm^-1, and sum them up to construct broadened distributions as shown by the blue and red curves. These curves are scaled by a factor of 0.2 for convenience.

2012.15842

figure_product_distribution_vs_TE_v2.jpg

5a8eb67227db43f3ada48b802954374d

Measured product state distribution and comparison to statistical theory. (A) Measured scattering probability into various product state-pairs | N_K_2, N_Rb_2⟩. (B – H) Comparing the measured (P_sc^meas) scattering probabilities to predictions (P_sc^0) from the state-counting model. Each frame displays the probabilities for state-pairs with a particular value of N_K_2, and as a function their N_Rb_2. The error bar for each measurement includes shot noise as well as 11% relative fluctuations in experimental conditions. The state-pairs | 10, 13 ⟩ and | 10, 15 ⟩ are energetically forbidden from being populated, and have measured populations that are consistent with zero. (Inset of (H)) A close up view of the scattering probability for the state-pair | 12, 7 ⟩, which displays a strongly suppressed population compared to the prediction. (I) p-value for the hypothesis that the measured and model distributions agree, as state-pairs that deviate most significantly from prediction (labled in the figure) are successively removed. An 11% relative fluctuation in experimental conditions was considered for this calculation (SM). (Inset) A zoom-in over the boxed region. The dashed line indicates p = 0.001, a threshold below which the hypothesis should be rejected.

2012.15842

figure_product_state_distribution_v6.jpg

ad4265fe021c441bb533b99ba010f9d0

Schematic diagram of LSH. For an input, we compute hash codes and retrieve candidates from the corresponding buckets.

2012.15843

lsh.jpg

0cd47b2c2c08426da2f5e1454ac19879

The phase diagram on the self-dual line, i.e. on the line y=(1-x)/(1+x) (where x=x' and y=y').

2012.15845

1dphasediagram.jpg

b03b2070f5b543f39e8bba53ab54d6a1

Duality-breaking order parameter, √(⟨ A^2⟩), as a function of x on the self-dual line, for various system sizes (indicated in legend). Lines are just a guide to the eye. The deconfined phase is at larger x.

2012.15845

A_x.jpg

137f326f247044d980a993b61daf4531

Schematic: some of the configurations contributing to the correlator ⟨A(r)A(r')⟩ in the model (<ref>). Circles indicate locations r, r'.

2012.15845

Acorrelatorfig.jpg

72db9afc35c1471c8cc6cc713cd3abdf

Two-point correlators for the operators A (top) and S (bottom). The displacement between the two operators is taken parallel to a lattice direction. Dashed lines are fits of the L=64 data in the range r∈ [10, 15] to the forms in Eq. <ref> with x_A=1.224, x_S=1.506 fixed and α_A,S free, giving α_A^2 = 0.72 and α_S^2 = 0.77.

2012.15845

C2h.jpg

b47e475d7cb44f1baa5861784104e449

Main panel: Finite-r estimates of the OPE coefficient C_AAS using 3-point functions with four different geometries, using data from system size L=48 (error bars are from variation between 6 samples). Inset: Average of the four estimates for L=48 and also for L=32.

2012.15845

CAAS.jpg

95afd4dcf112457cbdc83c9ec690c476

⟨ S⟩ as a function of x for different system sizes. The legend indicates system size. The lines are just a guide to the eye.

2012.15845

S_x.jpg

bad576e820444d95b126ce22626fce95

b_4(A)=-(1/2)κ_4(A)/ Var(A)^2 as a function of x for different system sizes. The legend indicates system size. The lines are b-spline fits and are just a guide to the eye.

2012.15845

b4A_x.jpg

2a4c5200017d4b39a07e9586c7217871

b_4(A)=-(1/2)κ_4(A)/ Var(A)^2 as a function of x for three system sizes. The lines are cubic polynomial fits.

2012.15845

b4A_x_cep.jpg

f70d11e3fdac44d2975d6f51739108b7

Main panels: Scaling collapse of autocorrelation functions for ⟨ A(0) A(t) ⟩ (top), ⟨ S(0) S(t) ⟩ (center) and ⟨ S(0) A(t) ⟩ (bottom) as a function of t/L^z (using x_A=1.224 and x_S=1.506). Inset: Autocorrelation time as a function of system size for 4 different correlators: ⟨ E(0) E(t) ⟩ (blue triangle), ⟨ A(0) A(t) ⟩ (orange triangle), ⟨ S(0) S(t) ⟩ (green pentagon), ⟨ S(0) A(t) ⟩ (red star). Straight line fits all the data points to a power-law τ = A L^z with z = 2.48.

2012.15845

correlatorsz.jpg

ef7a2cfb71284ac0b16edec7f64004fe

Mass of the largest cluster (number of links) as a function of the system size at x= 0.6367. The dashed line fits system sizes as a power-law A L^d_f, with d_f=1.77(2).

2012.15845

df.jpg

f3454c5799f14c2295a4ae285ba40d4d

The relation between e particles and membranes. Left: in the Z basis, the toric code wavefunction is a superposition of strings of occupied links (bold) representing Z=-1, and an e excitation is a vertex where an odd number of occupied links meet. Right: constructing a path integral (for a generic perturbed model) in this basis, worldsurfaces of strings become membranes ℳ (grey), and worldlines of e particles form the membrane boundary, or “loops”, ∂ℳ=𝒞_e. (The figure on the right is schematic: in the model we study, membrane configurations look like Fig. <ref>.)

2012.15845

eworldlines.jpg

cd284a7518ed49c685ea0ce3b85a4c68

Scaling collapse of L^x_A√(A^2) versus scaling variable z=(x-x_c)L^1/ν, where 1/ν=3-x_S. Blue line is the fitted scaling function using B-splines with 10 degrees of freedom. Legend indicates system sizes.

2012.15845

fit-A.jpg

eeb6a4fb49de4625aa24e99766a5fc98

Scaling collapse of b_1(A) versus scaling variable z=(x-x_c)L^1/ν, where 1/ν=3-x_S. The blue line corresponds to the fitted scaling function using B-splines with 12 degrees of freedom. The legend indicates the different system sizes.

2012.15845

fit-b1A.jpg

7c5ee9bb7a624ff0b6f37fb9eb1a89cf

Scaling collapse for L^3 x_Sκ_3(S) versus scaling variable z=(x-x_c)L^1/ν, where 1/ν=3-x_S. The blue line corresponds to the fitted scaling function using B-splines with 10 degrees of freedom.

2012.15845

fit-k3S.jpg

e5e3cdff30ed477c9ac188f201d418dc

Cubes of the original and dual lattices, with e worldlines (red) and m worldlines (blue) on the former and the latter respectively. (This configuration has linking number X=1.)

2012.15845

linkedlatticesfig.jpg

d850fb11ec34434fb1180096a59e549c

Loops in a sample of system size L=32 at x=0.6367. A spanning loop is highlighted in red (note periodic B.C.s). Our definition of a “loop” allows for junctions where 4 or 6 occupied links meet, but as can be seen they are relatively rare at the critical point.

2012.15845

links_c32.jpg

78e7b2c3308f4f98964a4c6cd77ff96f

Membrane patching process (schematic). A configuration ℳ of membranes, with a non-empty boundary ∂ M made up of finite clusters of occupied links (shown red) is “patched” by attaching finite surfaces (blue) to the components of ∂ℳ. In the resulting ensemble of closed membranes, line operators may be defined, specifying an emergent one-form symmetry. A “fictitious” scalar field ϕ=± 1 may also be introduced, whose domain walls are the patched membranes. The Ising^* transition is the ordering transition for this fictitious field.

2012.15845

membranepatching.jpg

45e250bb9f864f4188246d90b3d1c17b

Parameter-free scaling collapse for b_4(A) as a function of b_1(A) for several system sizes (colored). Black stars mark the two phases: the deconfined phase at (0,0) and the broken-duality phase (i.e. first-order coexistence) at (1,1). Inset shows a zoom of the left lower corner of the main panel. Two selected x values (0.6364 and 0.6370) are highlighted and arrows are drawn in-between consecutive system sizes. Close to the critical point, x ≈ 0.6367, values remain in a small region around (b_1(A), b_4(A)) ≈ (-0.109, -0.285).

2012.15845

parametric.jpg

8c023747083a4cf8a0beb851a1004d05

Sketch of phase diagram in the plane (x,y). The gauge stiffness K=tanh^-1 x increases to the right, and the matter field coupling J=tanh^-1 y increases upwards (self-dual line shown dashed). Inset: same phase diagram in the (y,y') coordinates, where duality acts as a reflection. (The exponents below imply that the e and m condensation lines are asymptotically parallel as they approach the self-dual critical point, though the curvature is not visible at this scale.)

2012.15845

phdiagram.jpg

9e0cc5fe17de493d90353d8bdb733968

Phase diagram showing the phase boundary for percolation of e worldlines (defined as clusters of links in ∂ℳ, see Eq. <ref>). The deconfined phase lies within the non-percolating phase. The self-dual line is shown dashed.

2012.15845

phdiagram_perco.jpg

2d1181db96af4c53ae69cfaec378e944

Membranes with boundary: part of a configuration, close to the critical point on the self-dual line. Occupied plaquettes are shown coloured (colours have no meaning). Occupied links, where an odd number of plaquettes meet, are shown in red, thick.

2012.15845

plaquettesigt1.jpg

6e628166a17045c59b8d51e360a066b8

Test of conformal invariance: ratios of the three-point correlators defined in the text. Assuming exponent values x_A= 1.224, x_S = 1.506, conformal invariance requires these ratios to converge at large r to the values indicated by dashed lines (error bars are from variation between 6 samples). We find agreement with the predicted value, within error bars, once r≳ 3. Error bars become too large for a useful comparison once r>4.

2012.15845

ratiosC3h.jpg

77faf75a39984299b0d2ec72d23fc07c

Scaling collapse of P_s as a function of (x-x_c)L^1/ν with x_c = 0.63666 fixed. Inset: flow of P_s as a function of the system size for several x values close to the multicritical point, showing approximate scale invariance at x=0.6367.

2012.15845

scaling_ps_xcfixed.jpg

b1e34db9e2a74ca88e603f30d0b5ac4e

Topology of the phase diagram of ℤ_2 gauge theory with matter. The shaded region is the trivial phase. The double line () represents a first-order line, ending at a standard critical endpoint (). The Higgs and confinement transitions have Ising exponents. The question mark shows the region studied in this paper. We give evidence for a scale-invariant, self-dual critical point.

2012.15845

schematicphasediagram.jpg

f8a7f8dc6a2849e7a57db034757d05fc

Main Panel: scaling collapse of P_s as a function of (y-y_c,I)L^1/ν_I, for x=1, with y_c,I = 0.21809 and ν_I = 0.63012. Inset: M_max as function of the system size L at x = 1, y=0.218. Straight-line shows a power-law using the fractal dimensions of Ising worldlines d_f,I = 1.7349.

2012.15845

x1A.jpg

5dbaed56c8924039995f748ac57d635c

Main Panel: scaling collapse of P_s as a function of (x-x_cP)L^1/ν_P, for y=1, with x_cP = 0.0865 and ν_P = 0.8762 Inset: M_max as function of the system size L at at x = 0.087. Straight line shows power-law with fractal dimension of percolation universality class, d_f,P = 2.53.

2012.15845

y1A.jpg

1d9fd4a78a974e9c8d6611fa03f8f25b

We define the low energy couplings through a contour integral on the complex s-plane, where the contour 𝒞_0 incircles the origin. On the complex plane, if the amplitude only has singularities on the real-s axes, either poles or branch points, then we can deform to contour 𝒞_∞.

2012.15849

Contour.jpg

849353d95e2e456f9d752e526529debd

Different origins for the EFT: (I) Integrating away massive states in tree exchanges, for example the Higgs for the Sigma model and the infinite tower of higher spin states in string amplitudes, (II) or massive states in the loop, for example the φ X^2 coupling.

2012.15849

Diagrams0.jpg

74dfc3487a1d429891eca921e5c9a0ae

We compare our Gegenbauer sum expression in eq.(<ref>), truncating at ℓ=10, with the explicit discontinuity in eq.(<ref>). We've normalized s/4m^2→ s, so that the discontinuity begins at s=1 to ∞. We've compared the result of eq.(<ref>), in red dots, to eq.(<ref>) which is the colored curve. The brown curve is for cosθ=1/2, and the blue curve for cosθ=1/6. Both exhibit perfect matching.

2012.15849

DisCom.jpg

9cb1eecb96d0497285e6069cf3d7cfac

An operator of four fields will contribute to the four-point amplitude as a polynomial, and the six-point amplitude as a rational term.

2012.15849

LocalOp.jpg

6e5abd055f5c4d2cb8e0abfc0efe8a01

For a given state, its contribution to each row is the same vector (v_ℓ_a,0, v_ℓ_a,1) after scaling away the moment factor x_a^k.

2012.15849

Organize2.jpg

6b14602cad084ba9b3f3c36e4648a72d

We organize the information that each state contributes to the determination of a_k,q. For each fixed row (fixed k), for example the red box, each state's contribution is proportional to a Gegenbauer vector multiplied by a universal factor x_a^k. For a fixed column (fixed q), the purple box, each state contributes to a point on a half moment curve multiplied by universal factor v_ℓ_a,q.

2012.15849

OrganizedPlot.jpg

6d22b21601cf4cc6b65cdb8cc6c721a4

The intersection of the ℙ^4 polytope defined by the boundaries in eq.(<ref>) with 𝐗_sym.

2012.15849

Photon2DPlot.jpg

a4134b2cfce7432fb12a15b6c254517a

We take the ratio of the four determinants in the second and third line in eq.(<ref>), denoted as m_i(ℓ_1,ℓ_2,ℓ_3), for i=1,⋯,4. We plot m_1(1+n,2+n,3+n)/m_2(1+n,2+n,3+n) (red), m_1(1+n,2+n,3+n)/m_3(1+n,2+n,3+n) (blue), and m_1(1+n,2+n,3+n)/m_4(1+n,2+n,3+n) (green), with n=1,⋯. As we can see, m_1(ℓ_1,ℓ_2,ℓ_3) is the largest and the ratio is an increasing function with spins.

2012.15849

Plots.jpg

1bd3d415675946f98a7f760fe3350904

The region allowed for 𝐬_ℓ and 𝐭_ℓ by unitarity. At weak coupling this constraint is only reflected in Re[𝐬_ℓ]≤1 and Im[𝐭_ℓ]≥0.

2012.15849

UnitCircle.jpg

9d4ac94bb310428f9efe57f9c75fa278

The s-channel geometry at fixed k. The vector 𝐚_k must live on the intersection between the cyclic plane 𝐗_ cyc with the unitary polytope U_k.

2012.15849

Xplane0.jpg

2c50a51cb9c845679efe083355d23e15

In the presence of massless cuts, we can either (I) introduce a small mass regulator and push the cut slightly away from the origin, or (II) we analytically define our generalized couplings by moving the contour at origin onto to the complex plane to s=± i μ in a way that the integration measure is positive definite. After deformation the contour picks up the discontinuity on the real s-axes, which for |t|<|s|, is controlled by unitarity. We can analytically continue to |s|<|t| for theories with well behaved soft limits.

2012.15849

masslesscut.jpg

94105a8a54d0443ba18ad37f25cbd99f

We represent the internal loop momentum in the center of mass frame. The angle between the loop momentum and p⃗_1=-p⃗_2 is θ', while the angle between the plane spanned by (p⃗_1,ℓ⃗_1) and the plane (p⃗_1,p⃗_2) is ϕ'. θ is then the usual scattering angle.

2012.15849

phase.jpg

4d3f45fa1a6f40e6b59efcd64d074ed8

The convex hull of these three vectors encloses the origin, and hence trivially covers the entire two-dimensional plane.

2012.15849

trivial.jpg

45704892a4ea4d9e8049d64abca82746

Phase Response. (a) The measured relative phase response ΔΦ(ω) = Φ (2ω) - 2Φ (ω) extracted from modulation with φ=0 (blue circles) and φ=π/2 (red triangles). Shown in black is the rolling average. The two sets of φ data are plotted in the frequency ranges where they are sensitive to ΔΦ(ω). (b) Cavity scan (black dots) and the fitted curve (orange dashed) of a representative interferometric spectrum at the frequency indicated by the orange square in (a). (c) Phase response Φ(ω) (black solid) reconstructed from the measured ΔΦ (ω). (d) Comparison between the measured relative phase ΔΦ(ω) (blue solid) and the relative phase obtained from the reconstructed phase response Φ(ω) (red dashed).

2012.15850

Phase_xfn_v4.jpg

d6783132b9d14061875a60aa84a64d2d

The A_3 Stasheff associahedron, relevant for six-particle scattering. The positive region in the space of kinematics amounts to the interior of the polytope, whereas pentagonal faces correspond to double-scaling limits, and square faces to soft (or equivalently multi-Regge) limits. The two red pentagonal faces represent the two possible double scaling limits DS1 and DS2 that are considered in this work.

2012.15855

assoc2.jpg

fe289916247440258069916d15efe408

Integration path γγ_1 γ_2 connecting the base point of integration 0, 0 and a general kinematical point x, y inside the unit square.

2012.15855

int-path.jpg

8dfcd27d948842e6937466422ea3ebe3

A graphical depiction of the eight regions of the _2_- summation variables presented on the j_1, j_2-plane, parametrized by the helicities l_1, l_2 and listed in tab:sign.

2012.15855

regions.jpg

33f84c183eca479ca5b3dedff5d0066b

Motivating Example. Intermediate pre-training using salient span masking is helpful in improving closed-book QA performance <cit.>. We propose to use a learned masking policy during intermediate pre-training to pack more task-relevant knowledge into the parameters of LM.

2012.15856

framework.jpg

ae02454c16ba4432b75403f03930c932

Stages in the proposed work. (1) Learn a Masking Policy: Learn a masking policy on one or a few anchor tasks; the policy learns to mask important information in the context. (2) Intermediate Pre-train: Pre-train the language model by using the learned policy to mask generic corpus, thus knowledge of interest is implicitly encoded into the LM. (3) Fine-tune: Fine-tune a pre-trained checkpoint on downstream closed-book QA tasks.

2012.15856

stages.jpg

df616b98edf44daeb5ffff2f21d5213e

(a) Schematic of a traveling phonon in a 2D phononic crystal (PnC). In case of metallic PnCs, the engineered phonons are expected to mediate abnormal 2D e–e interactions. (b) Optical micrograph of PnC–Nb used in this study; the periodically perforated area is PnC (0.3×0.5 mm2); the thickness is 150 nm, self-standing. The pathway of the excitation current (i.e., I+ toward I-) for R–T measurements is unidirectional, and the distance between electrical pads for voltage readings is 0.3 mm. (c) SEM of the region surrounded by the yellow dotted square in (b); the array of Nb islands and Nb bridges forms a 2D square lattice with a lattice constant of 20 m. (d) R–T curves; only the initial six temperature cycles are shown here. Inset, zoomed-in plot. (e) R at 300 K in (d) versus the number of temperature cycles, z. (f) 4d-orbital filled model, based on the Friedel sum rule indicated in (e); after the first five cycles, the Nb bridges undergo a transition to Mott insulators.

2012.15858

figure_1.jpg

e9a6c768e0c044f7b21293ecbd64ad11

R–T curves; only the first, fourth, and 25th cycles are shown. For all other R–T curves, see Fig. S1 in the [supplmater]supplementary material.

2012.15858

figure_2.jpg

44d1119926ce412eb11968d1cb512608

V–I results at 300 K in various perpendicular magnetic fields. Inset, zoomed-in plot.

2012.15858

figure_3.jpg

286acf02fa9d4c8fa51b2de36409cac0

Measurement of the DC critical current I_C of the RTSC at 300 K in zero magnetic field; first, an applied direct current is increased in the positive direction (orange); after V jumps to the set compliance level, V–I is measured again (red). Inset, V–I results in linear scale.

2012.15858

figure_4.jpg

ed9b2c3248a444c1890392c162d959e0

Principle of the Mott ring; a circulating supercurrent path and Cooper pairs (twin blue dots) are illustrated using a false-color SEM.

2012.15858

figure_5.jpg

f41177e978214c5eaca2d2ae7501d6cb

Additional R–T cycles of the sample after its I_C was measured in Fig. <ref>; that is, before the first cooling, the sample was in the indeterminate state (i.e., the asymmetric mixture of metallic and insulating states). The second warming process (orange) realizes zero resistance at 300 K again; the new condensed phase is named the superconductor requiem (SCRQ).

2012.15858

figure_6.jpg

3c5ebbfbe5754c33b01d2c9a0a20db9f

(a) V–I results of the SCRQ at 300K. As shown in the circuit, a pulsed current is applied; first, the pulsed current is swept from +0 to +3 mA (red); second, from +3 to –3 mA (blue). The I_C of the SCRQ is indicated (red arrow: 0.5 mA), and the diffusion length Λ of nonequilibrium quasiparticles in the vicinity of the phase slip center (PSC) are denoted for each voltage step. Inset, zoomed-in plot of the negative current region; V is normalized by the product of the flux quantum Φ_0 and the Josephson plasma frequency f_J. (b and c) Schematic representation of the beating heart of the PSC in the positive current region and the quantum phase slip (QPS) in the negative current region, respectively, using false-color SEMs.

2012.15858

figure_7.jpg

d6de7734b2014b71979b7a2e552544b0

(a) C–V results of the SCRQ in the atmosphere. As shown in the circuit, for each biased DC voltage, C is determined by temporally integrating the measured current. Characteristic voltages are denoted. Inset, Mott–Schottky plot, i.e., 1/C^2 vs. V plot; the deep flat-band potential V_FB is indicated (red arrow). (b) dI/dV–V results of the SCRQ in the atmosphere. As shown in the circuit, a DC current is applied, and the voltage is measured. The dI/dV is obtained by differentiation. Characteristic voltages are denoted. (Bottom-left inset) Zoomed-in plot near the origin; dI/dV is normalized by the conductance quantum, G_0 = 2e^2/h. (Bottom-right inset) V^2-dependent dI/dV; the voltage regions denoted as (I)–(III) correspond to those in the main panel. (c) Constructed semiconductor model by unifying all experimental results in (a) and (b); that is, a special Schottky junction consisting of the hole-enriched semiconductor (p-Semi) and SC is formed in the SCRQ; left side, p-Semi with the band gap E_g = 1.1 eV; right side, SC with the gap Δ = 0.8 eV. Shades: light blue, valence band (V.B.); light orange, conduction band (C.B.). The Fermi level of p-Semi is assumed to be 0.05-eV above the edge level of its V.B.; shown here is the case of zero-bias voltage, i.e., which is equivalent to that of SC, ε_F. Dots: empty, a hole; blue, an electron; twin, Cooper pair. The Cooper pair is confined in the deep potential well because of the large V_FB. Solid arrows: transport of a hole (red) and an electron (blue). Broken arrow, the built-in electric field E_D; direction, polarity. For clarity, the density of states of the d-wave SC is designedly changed to that of an s-wave SC, and thermally excited carriers are disregarded because Δ≫ k_BT at room temperature.

2012.15858

figure_8.jpg

07d99a8186cf4b4e9100fd7d0e6dd8e2

(a) Photograph of the experimental setup for the discharge test. During measurements, the SCRQ is enclosed in a die-cast box to prevent Vis-light irradiation. R_L is a commercial metal film resistor, and the voltage, V_L, across R_L is measured continuously using a nanovoltmeter. Inset, equivalent circuit for the discharge test configuration. (b) Time evolution of V_L for R_L = 10 k in the atmosphere. (c) Time evolution of the time-averaged consumed energy for several R_L. (d) Schematic representation of the sustainable process using the Ginzburg–Landau free energy function f, where the horizontal axis is the phase of the macro wavefunction Ψ; fp-Semi, f in p-Semi; fSC, f in SC. δF, the condensation energy. Circles, charged particles.

2012.15858

figure_9.jpg

04c3427878d3421586cd0c9be434f723

Coreference results (en), gender bias

2012.15859

coref_new_clean.jpg

0bb5fb97567e40239675a127cc502f70

Coreference (en) results broken out by bias modification method (pre- vs. post-processing).

2012.15859

coref_new_clean_relplot.jpg

8bc8f650e653466a91a1ee66e506bddc

Coreference (en) results broken out by type of Winobias challenge, Type 1 is more difficult as there are only semantic cues to correct coreference, Type 2 has also syntactic cues.

2012.15859

coref_new_clean_relplot_type.jpg

193542f3e5c6437fba0b56b11dd87280

Coreference (en) results broken out by different standard WEAT tests for gender bias.

2012.15859

coref_new_clean_relplot_weat.jpg

1a6d1e4edb3845dfacf8cb6a3d9c056c

Sample development data from our downstream coreference and toxicity detection systems. Extrinsic bias metrics measure disparities in performance, such as rates of false negatives (a, d) and false positives (b) between different gender (or other vulnerable identity) groups. These examples display performance disparity between examples referencing females vs. males.

2012.15859

extrinsic_bias.jpg

da264c91ec2a449cac5590d247e4ff74

Hatespeech (en) results broken out by different standard WEAT tests for gender bias.

2012.15859

hsd_en_all_reverse_testhue_relplot.jpg

5bd954915b4a492db77a8c8daa4abc8c

Hatespeech (en) results broken out by bias modification method (pre- vs. post-processing).

2012.15859

hsd_en_all_reverse_testhue_relplot_method_relplot.jpg

5e5dfcf608324736ac1e1745d9e8da16

Hatespeech detection results (en), gender bias

2012.15859

hsd_en_new.jpg

c20182205f15460386afa5ea417ca414

Hatespeech (es) results for gender bias metrics broken out by bias modification method.

2012.15859

hsd_es_all_gender_relplot.jpg

87fc684ec11c41698a821edee822dddd

Hatespeech detection results (es), migrant bias

2012.15859

hsd_es_all_migrant.jpg

0b000d49d4ae42dea5aa3912a9883350

Hatespeech (es) results for migrant bias metrics broken out by bias modification method.

2012.15859

hsd_es_all_migrant_relplot.jpg

6495ccc6ea864fdea1646941d24c5ab7

Hatespeech detection results (es), gender bias

2012.15859

hsd_es_gender_new.jpg

473c28ff752f4b09b60853e6df1b2112

Intrinsic bias metrics measure the geometry of embedding spaces and reveal pervasive imbalances between types of identity words and their neighbouring words. This displays gender bias in our embeddings via the much larger distance of the female terms from the career-related terms.

2012.15859

weat_viz.jpg

4e33c2f2f62245eaaf0e7ac7f55c8a4e

The argument for Theorem <ref> illustrated for D = 2. At top left we have the special bubble introduced near the boundary. From left to right, top to bottom, this bubble is absorbed by the boundary. In purple we have indicated where absorbing a point-like junction causes a recombination of the boundary defects, and produces compensating phases. The condition that the boundary correlation function is non-vanishing requires that these phases precisely cancel e^2π iω of the original bubble. This requires ω be exact in group cohomology, hence that there is no anomaly.

2012.15861

boundaryabsorb.jpg

65023c9219d74f669c4c91914fa0db5b

Here we have a spatial picture of a 2+1D fermionic system with a unitary symmetry C^2 = (-1)^F. At the white circle, two co-oriented C defects (green) fuse to a fermion parity defect (dashed). This junction may trap a Majorana fermion. However, if it does, then by layering a p+ip superconductor (so taking β_3 = 1 ∈Ω^3_Spin), since the fermion parity defect ends at the fusion junction, the p+ip superconductor sees a vortex there (red star), which also traps a Majorana fermion <cit.>. This Majorana may be paired with the other one to create a featureless fusion junction. Thus, this decoration does not actually contribute an anomaly in this symmetry class. With more work, one can show that for this symmetry class and dimension, there are in fact no nontrivial anomalies (see Appendix C.4 of <cit.>).

2012.15861

pwavevortex.jpg

2d1617f4a3db4e068d5b521f7715f2b5

The elementary recombination for D=2 (sometimes called an F-move or crossing relation), drawn here as relating the red foam at bottom to the red foam at top, is captured by a point-like singularity of a three dimensional foam which is Poincaré dual to a tetrahedron and for which the initial and final foams form the top and bottom boundaries.

2012.15861

singularity.jpg

d4fc49ab74db4cf1825c20af6295f24f

A bubble halfway absorbed into a symmetric boundary condition drawn for D = 2. The bulk G_-foam is drawn in green. Where it ends on the boundary (white circles) there is a possible gravitational anomaly, described by a function β_1: G_→Ω^1, which in a spin theory could be a fermionic operator, Ω^1_spin = _2. This gravitational anomaly means that the G_-foam with boundary is not isotopy invariant, and so we cannot apply the argument we gave in Section <ref> for Prop. <ref>. However, by introducing the subliminal G_-foam (orange) carrying the invertible phase corresponding to β_1 (for Ω^1_spin it is a fermionic world line) the gravitational anomaly is cured and the entire object is isotopy invariant.

2012.15861

subliminal.jpg

bd50adc1133749568081005520794aa8

Networks governed by mixtures of Mechanisms. NMDS projections (each with identical axis scaling) of the ensemble distances between networks, as in Fig. <ref>B, show State Spaces of Mixture Mechanisms. Networks (pies) were made with the `grow_Mixture' algorithm in <cit.>. The colors in each pie are proportional to the number of nodes in that network governed by the corresponding mechanism. Mixture Identifiability Smaller mixtures (e.g. two mechanisms) may be identifiable, but de novo identification of mixture mechanisms across larger numbers of mechanisms (e.g. five) or when the number of mechanisms is unknown (far right column) is not possible with our approach. Mixture Similarity Different mixtures of mechanisms can produce functionally equivalent systems, here quantified by normalized Ascendency<cit.>. Pies are example networks generated by mixtures of the five mechanisms, grouped by their shared Ascendency. Function Identifiability Leave-one-out cross validation of system development (normalized Ascendency). Each point is a single mixture network, with randomly chosen proportions of the five mechanisms and governing parameters within each mechanism.

2012.15863

Fig_EmpiricalPies_Updated-NoDirected_Fall2020_MixtureIdentifiability_ACratio_MiniLegend_PredictedACratio_Rearranged.jpg

002e6eb877ad41dea6a82aac529d6b3f

The classification of five common network Mechanisms. 100 networks were simulated from each mechanism, systematically varying one governing parameter for each network, with 3 replicates of each value. This parameter is listed below the name of each mechanism respectively. Note that our approach is not limited to these five mechanisms. A Network State Space was made from the 1500 simulated networks by pairwise comparing 18 properties of each network in a stacking ensemble method. A two-dimensional NMDS projection of this state space (with identical axis scaling) shows that networks cluster by mechanism. In this plot circles are each a network, with their radius proportional to the governing parameter. Classification ROC curves and corresponding AUC values (inset numbers) for the ability to classify these 1500 networks. When AUC = 1 the classifier can identify all true positives without including any false positives. Random classifiers produce AUC = 0.5. Each network was classified with an independently simulated state space. Empirical Classifications of 373 empirical networks spanning 9 kinds of systems. Each circle represents a mechanism. Inset numbers are the number of networks classified as that mechanism. If a network was classified as more than one mechanism the resulting pie graph was made with equal splits for each mechanism because we cannot yet confidently assign probabilities to mixture networks as shown in Fig. <ref>.

2012.15863

Fig_StateSpaces_Direct_Classify_MechanismDescription_Updated_Fall2020_ExcludeLargeNetworks.jpg

31a3139a144b437f9df1c7790c485c6f

EC-GAN has an external classifier trained on generated images from the generator and real images simultaneously. The discriminator and generator are trained conventionally, and the discriminator's sole output is the predicted probability of the image being real. The datasets are fully-supervised, but the classifier is trained on a semi-supervised algorithm.

2012.15864

ECGAN.jpg