acceptodds
Under review as a conference paper at ICLR 2027

USHE: Variance-Controlled Probabilistic Spiking Encoding for Low-Latency SNNs

Abstract

Input encoding is a central design choice in spiking neural networks (SNNs), shaping how information and stochasticity enter both learning and inference. Widely used Bernoulli rate encoding relies on independent sampling, introducing input variance that can limit representational precision at low latency. We introduce unbiased spiking hypergeometric encoding (USHE), a tunable family of probabilistic encoders that preserves input intensities in expectation while controlling spike-rate variance and temporal dependence. By decoupling the latent population size from the simulation horizon, USHE controls encoding stochasticity without increasing network latency. Theoretically, we characterize USHE members through constrained variance-minimization problems and establish independent Bernoulli encoding as the infinite-population limit at fixed latency. Experiments on CIFAR-10 and CIFAR-100 show consistent improvements in clean accuracy over Bernoulli encoding and reveal distinct finite-population operating points under adversarial attacks and Gaussian corruption. Controlled encoder substitutions further separate training and inference effects, showing that the inference population governs a systematic clean–robustness trade-off whose magnitude depends on the trained model and architecture. Together, these results establish population-controlled temporal dependence as a principled design dimension for stochastic spike encoding, rather than treating input encoding as a fixed preprocessing step.

Primary area:
optimization
Keywords:
SNN, spiking neural networks, spiking encoders, hypergeometric encoders

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_SSP2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces USHE, a probabilistic spike encoding method designed to reduce encoding variance within a fixed number of simulation steps. The method combines stochastic rounding with sampling without replacement to reduce temporal variance while keeping the encoding unbiased. The authors provide theoretical results on variance optimality and its relationship to Bernoulli encoding. Experiments on CIFAR-10 and CIFAR-100 show improved clean accuracy compared with Bernoulli encoding, along with different trade-offs between accuracy and adversarial robustness.

Critical strengths
  1. The motivation is clear and relevant. Encoding variance can be a significant issue for SNNs, especially when only a small number of time steps is available. Addressing this problem without increasing the simulation horizon is worthwhile.

  2. The theoretical analysis is one of the stronger parts of the paper. The authors provide results on unbiasedness, variance optimality, and convergence to Bernoulli encoding, which help explain the behavior of the proposed method.

  3. USHE consistently achieves better clean accuracy than standard Bernoulli encoding in the reported experiments. The comparison between different training and inference encoding strategies also provides some useful observations.

Critical weaknesses
  1. I have some concerns about the methodological novelty. The main idea is based on stochastic rounding and sampling without replacement, both of which are well-established techniques for variance reduction. While their application to spike encoding is reasonable, it is not entirely clear how USHE differs from existing structured or variance-reduced encoding methods. More comparisons with related approaches would help clarify its contribution.

  2. The adversarial robustness evaluation is not fully convincing. The attacks mainly rely on single-sample BPDA gradient estimation. Since USHE uses stochastic encoding, this may not be sufficient to properly evaluate its robustness. Stronger adaptive attacks using EOT with multiple samples would be important to rule out possible gradient masking effects. Also, the reported robustness improvements are not consistent across different architectures and attacks, making it difficult to draw a general conclusion.

  3. The practical advantage over existing encoding strategies is somewhat limited. Although USHE improves clean accuracy compared with Bernoulli encoding, it still performs considerably worse than direct encoding. In addition, the theoretical variance optimality applies to input-rate reconstruction, but it does not necessarily translate into better classification accuracy or adversarial robustness. The connection between these theoretical results and the downstream improvements could be explained more clearly.

  4. The experimental evaluation is relatively narrow. The results are limited to CIFAR-10/100, two network architectures, and short simulation horizons. More importantly, each configuration appears to use only one training seed. The reported standard deviations mainly reflect randomness during evaluation rather than variation across independent training runs. This makes it difficult to assess how stable and reproducible the improvements are.

  5. The computational overhead of USHE is not sufficiently discussed. Although the method does not require additional simulation steps, the sampling procedure itself may introduce extra computation compared with standard Bernoulli encoding. Without actual latency or energy measurements, it is unclear whether the lower encoding variance leads to meaningful efficiency gains in practice.

Questions for authors

See weaknesses

4 comments

Thank you for recognizing the motivation, theoretical analysis and clean-accuracy gains. We have strengthened the comparisons and clarified the practical role of USHE. We address the encoding contribution and Direct comparison here, the adaptive-attack sweep in Comment 2, and training replication, broader datasets and implementation cost in Comment 3.

W1 — Contribution beyond stochastic rounding and sampling without replacement

These sampling primitives are established. Our contribution is a population-indexed family that preserves binary spikes, each timestep's Bernoulli marginal, expected input activity and the visible horizon T, while controlling temporal dependence through N. Only T timesteps are sampled, even when N>T. The endpoint N=T minimizes unbiased rate-reconstruction variance; increasing N connects it to independent Bernoulli encoding.

The analysis characterizes this entire family: variance optimality among unbiased exchangeable length-N populations observed for T steps (Theorem 3.2), global count-variance optimality at N=T without requiring exchangeability (Corollary 3.3), convex-order guarantees and quantitative convergence to Bernoulli. This identifies both the achievable variance reduction and the effect of the tuning parameter.

Matched structured-encoder comparisons. We added DRC, which deterministically rounds the spike count (generally biasing the rate) and places spikes uniformly, and cyclic burst, which uses USHE₄'s unbiased count distribution but places spikes contiguously with a random cyclic shift. These controls test count quantization and temporal placement under the same CIFAR-100 training recipes.

USHE₄ improves full-test PGD-7 accuracy over DRC by 2.97 ± 0.53 points on VGG and 4.79 ± 0.63 on ResNet across three training seeds. The adaptive comparisons in Comment 2 also support this advantage and assess the effect of alternative temporal placement. These results support USHE's practical benefit over deterministic count rounding, while the cyclic control shows comparable strong-PGD performance with another arrangement of the same counts. USHE additionally provides an explicit population parameter connecting the low-variance endpoint to Bernoulli.

W3 — Practical advantages over Direct encoding.

USHE provides a way to balance clean accuracy and adversarial performance while retaining binary input spikes. Direct encoding achieves strong clean accuracy, but stochastic encoders retain substantially higher attacked accuracy in the paper’s evaluations. The following CIFAR-100/ResNet-19 results at T=4T=4 illustrate how the USHE population parameter adjusts this balance. All accuracies are percentages, evaluated on the full test set.

EncoderTraining seedsCleanRFGSMPGD-7
Direct176.878.280.03
Bernoulli351.33 ± 0.5139.54 ± 0.3224.64 ± 0.42
USHE₄357.89 ± 0.1141.20 ± 0.4721.93 ± 0.76
USHE₁₂352.60 ± 0.5840.38 ± 0.3524.74 ± 0.29

Bernoulli and USHE entries are means ± sample SD across three independently trained models. Direct is the original single-checkpoint reference; its training variability has not been measured.

USHE₄ improves clean accuracy over Bernoulli by 6.56 percentage points, while retaining substantially higher attacked accuracy than Direct under these evaluations. Increasing the population to N=12N=12 recovers Bernoulli-level PGD-7 accuracy while improving both clean and RFGSM accuracy over Bernoulli. In particular, USHE₁₂’s paired RFGSM gain is 0.83 ± 0.05 percentage points, positive in all three training seeds. Its mean PGD-7 accuracy is also slightly higher, by 0.09 points. These results demonstrate useful operating points: the smaller population favors clean accuracy, while the larger population moves toward Bernoulli’s PGD-7 performance without giving up the clean-accuracy improvement. The stronger adaptive evaluations are reported separately in Comment 2. We also note that on certain datasets the difference between direct and USHE encodings becomes even less emphasized (see SVHN table bellow).

USHE also retains a computational advantage of binary input encoding. For a synaptic weight ww and spike s∈{0,1}s\in\{0,1\}, the product wsws is either zero or ww, so an event-driven first layer can use gated weight additions in place of floating-point input–weight multiplication. USHE therefore combines an adjustable clean–robustness balance with the binary input arithmetic shared by Bernoulli; its additional sampling cost is accounted for in Comment 3.

W3 - Connecting encoding theory to classification.

Let SˉN=T−1∑tSt(N)\bar S_N=T^{-1}\sum_tS_t^{(N)} be the average input spike vector and zN=WSˉN+bz_N=W\bar S_N+b the class scores of a fixed affine classifier, with weights W and biases b. For independently encoded input coordinates we have,

E[CE(zT,y)]≤E[CE(zN,y)]≤E[CE(z∞,y)],\mathbb E[\mathrm{CE}(z_T,y)]\le\mathbb E[\mathrm{CE}(z_N,y)]\le\mathbb E[\mathrm{CE}(z_\infty,y)],

where y is the true class and N=∞ denotes Bernoulli. This expected cross-entropy ordering follows from convex ordering of the count distributions, not from variance alone. We also obtain a classification-error bound through true-class-versus-competitor score margins: USHE preserves their means and reduces their variances, improving the bound when the mean margins are positive. With these results we hint how encoding variability connects to downstream loss and error for a specified classifier class.

W2 — Adaptive attacks: the 200-image sensitivity sweep

We tested the attack changes you recommend: longer optimization, multiple stochastic gradient samples, and additional restarts. The complete PGD-100 sweep below uses the same 200 class-balanced images per CIFAR dataset, training seed 1001, T=4 and ε=8/255. E denotes EOT samples; R5 denotes five restarts. Other columns use one restart; α=0.01 unless indicated. Entries are accuracies (%).

Setting / encoderE20E50E100E20, R5E20, α=2/255
C10/VGG / Bernoulli26.7626.6926.4126.6026.79
C10/VGG / USHE₄21.6621.4121.4321.2221.59
C10/VGG / USHE₁₂26.9726.7026.5226.6627.16
C10/ResNet19 / Bernoulli32.1432.1531.7032.2132.15
C10/ResNet19 / USHE₄27.0926.8826.5427.0726.91
C10/ResNet19 / USHE₁₂32.7431.9032.0532.4932.52
C100/VGG / Bernoulli11.2711.1811.2011.2011.31
C100/VGG / USHE₄10.1110.0310.0510.2110.25
C100/VGG / USHE₁₂12.3012.3012.3012.2412.46
C100/ResNet19 / Bernoulli15.1115.0415.0315.1815.08
C100/ResNet19 / USHE₄15.2515.4815.1015.2315.56
C100/ResNet19 / USHE₁₂15.0715.2315.0315.2715.18

USHE₁₂ exceeds Bernoulli under all five CIFAR-100/VGG schedules, by 1.03–1.15 points. Raising EOT from 20 to 100 changes accuracy only modestly: the average reduction across all twelve configurations is 0.26 points (maximum 0.68); five restarts reduce it by 0.07 on average. The main reduction occurs when strengthening the original attack: CIFAR-10/ResNet/USHE₁₂ falls from 46.99% under PGD-7/EOT-1 to 32.74% under PGD-100/EOT-20 on these same images.

Further study: 2,000 images and three training seeds. We then extended PGD-100/EOT-20 to a fixed, class-balanced 2,000-image cohort per dataset, disjoint from the sensitivity cohort, across all 36 Bernoulli/USHE₄/USHE₁₂ checkpoints. This study uses ε=8/255, α=0.01 and one restart. Mean ± sample SD across three training seeds:

Dataset / architectureBernoulliUSHE₄USHE₁₂
CIFAR-10 / VGG-SNN26.70 ± 0.1020.01 ± 0.2725.36 ± 0.14
CIFAR-10 / ResNet-1931.38 ± 1.0625.30 ± 0.6430.87 ± 0.21
CIFAR-100 / VGG-SNN11.97 ± 0.159.34 ± 0.4211.23 ± 0.11
CIFAR-100 / ResNet-1916.71 ± 0.4113.98 ± 0.5616.54 ± 0.26

The extension checks whether the small-cohort ranking persists across more images and independently trained models. USHE₁₂ stays much closer to Bernoulli than USHE₄, while Bernoulli has the highest mean in these four settings. The favorable CIFAR-100/VGG ranking from the sensitivity sweep does not persist here. Together with the clean results in Comment 3, this supports a tunable clean–robustness trade-off; the advantage over Bernoulli depends on the attack and setting.

For CIFAR-100, we also evaluated DRC and cyclic burst on the same 2,000 images with training seed 1001. USHE₄ exceeds DRC by 1.67/3.65 points and cyclic burst by 0.28/0.07 points (VGG/ResNet), extending the original-attack comparisons in Comment 1.

Gradient estimation and scoring. The forward pass retains actual sampled spikes. EOT averages per-realization cross-entropy gradients through the expected-rate BPDA surrogate, reducing sampling variability while leaving the backward approximation explicit. Candidate selection uses 32 separate draws and final scoring 128 fresh draws. Each scoring draw makes one prediction; we average correctness indicators, so this evaluates single-encoding predictions rather than a 128-sample ensemble.

Alternative attacks. In addition, we tested on custom APGD-CE (100 steps/EOT-10) and gradient-free square-patch search useing matched 500-image cohorts (seed 1001). On CIFAR-100 VGG/ResNet, USHE₄ exceeds Bernoulli by 5.26/5.74 pp under the search and DRC by 1.04/3.62 pp under APGD and 8.99/5.94 pp under the search.

W4 — Independent training runs and broader evaluation

We added training seeds 1001 and 1002 to seed 1000. Bernoulli, USHE₄ and USHE₁₂ now have three checkpoints per CIFAR dataset/architecture setting (36 total), with clean and classical-attack evaluation on all 10,000 test images. Dedicated clean accuracy (%) and paired gains are:

Dataset / architectureBernoulliUSHE₄Paired gain (pp)
CIFAR-10 / VGG-SNN79.32 ± 0.1784.98 ± 0.08+5.67 ± 0.18
CIFAR-10 / ResNet-1979.18 ± 0.2585.46 ± 0.15+6.28 ± 0.30
CIFAR-100 / VGG-SNN51.60 ± 0.1157.90 ± 0.04+6.30 ± 0.09
CIFAR-100 / ResNet-1951.33 ± 0.5157.89 ± 0.11+6.56 ± 0.41

Both USHE populations improve clean accuracy for every seed in all four settings; USHE₁₂ gains 1.27–1.87 points on average. SD is computed across checkpoint means after averaging inference repetitions. Some attacked gains also reproduce: CIFAR-10/ResNet-19 RFGSM gives 69.58 ± 0.39% for USHE₄ versus 66.89 ± 0.34% for Bernoulli, a paired gain of 2.69 ± 0.06 points, positive in all three seeds.

Additional datasets. SVHN and Tiny ImageNet extend the clean-accuracy evidence beyond CIFAR and add a spiking ResNet-18. At the endpoint N=T:

Dataset / architectureTDirectBernoulliUSHE N=TN=TUSHE − Bernoulli (pp)
Tiny ImageNet / ResNet-184—48.9452.43+3.49
Tiny ImageNet / ResNet-188—52.8455.67+2.84
SVHN / VGG-SNN496.1786.9992.55+5.56
SVHN / VGG-SNN896.4990.2993.58+3.28
SVHN / ResNet-19496.2288.5994.16+5.57
SVHN / ResNet-19896.6891.9495.29+3.34

All results use one training seed (1000), with repeated inference evaluations. Differences are calculated from the unrounded means and expressed in percentage points. “—” indicates that a Direct result is not available in the supplied report.

These additional datasets use one training seed (1000) per configuration, with repeated single-encoding evaluations. On SVHN, USHE at T=4 exceeds Bernoulli at T=8 by 2.25/2.22 points (VGG/ResNet), using half as many simulated steps. It also approaches Direct's clean accuracy: on SVHN/ResNet-19 at T=8, USHE reaches 95.29% versus Direct's 96.68%, a gap of about 1.4 points.

The submitted CIFAR T=8 results retain clean gains of 3.40–6.01 points over Bernoulli (one training seed). Our theory covers arbitrary horizons; experiments cover T=4 and T=8.

W5 — Encoding operations and neuromorphic cost

USHE retains binary input computation and adds a small budget counter plus conditional sampling. For intensity p, we initialize K=floor(Np)+B, with B Bernoulli of parameter Np−floor(Np). Further, at a visible step t, we set m=N−t+1, draw J uniformly from {1,…,m}, emit s=1{J≤K}, and update K←K−s. This implements probability K/m without division in the per-step threshold. The value m is shared across synchronized input coordinates; the unused N−T positions are never simulated or stored.

ResourceBernoulliUSHE
InitializationInput probability pForm Np and stochastically round
Per-step samplingUniform draw, compare with pBounded-integer draw, compare with K
Count updateNoneK←K−s
Extra count state per coordinateNone⌈log₂(N+1)⌉ bits
Visible steps / expected spikesT / TpT / Tp

The counter needs three bits for N=4 and four for N=12. Both encoders use O(Td) logical sampling/comparison operations for d inputs; USHE adds O(d log N) counter bits. Deterministic K=0 or K=m cases can skip sampling. Bounded-integer draws may require range conversion or rejection, so equal operation order does not mean equal random-number cost.

With fan-out f_j for input j, both encoders have expected first-layer event count TΣ_j f_jx_j. Binary spikes permit gated weight additions in that layer, as explained in Comment 1. This reduces the need for floating-point input–weight multiplication in an event-driven implementation.

Thank you for the detailed responses and the considerable effort put into the additional experiments. I appreciate that the revision has addressed several of my concerns, particularly regarding stronger attacks and statistical reliability. However, the new results also weaken the original robustness claims, and I still have reservations about the novelty and practical benefits of the method. I will therefore maintain my original rating and look forward to seeing these claims revised and the work further improved in a future version.

Reviewer_UPnAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes unbiased spiking hypergeometric encoding (USHE), a new probabilistic spiking input encoding scheme that controls the variance of the spiking sequence by adjusting the hidden population size under fixed time step delays. The authors demonstrate the effectiveness of their method through theoretical analysis and experimental results.

Critical strengths
  1. The authors provide a theoretical analysis of the proposed method.
  2. The authors conducted experimental validations in clean conditions and in the presence of various adversarial attacks and Gaussian noise corruption.
Critical weaknesses
  1. The theory only shows that input-level reconstruction variance is optimal. However, classification accuracy and adversarial robustness are properties learned by the downstream network. Minimal variance in input reconstruction does not equate to optimal classification performance. While this paper indirectly suggests that optimality at the input level is beneficial for downstream tasks, it does not establish a theoretical connection between the two.
  2. A major advantage of Bernoulli-rate encoding is its ability to provide random-smooth authentication robustness. However, USHE samples are no longer independently and identically distributed, meaning existing certificates cannot be directly reused. The authors did not attempt to derive any authentication bounds under USHE. When considering only empirical adversarial accuracy, the results are susceptible to gradient obfuscation and BPDA evaluation vulnerabilities, limiting their reliability.
  3. USHE requires that each input channel maintain a spike budget, perform random rounding, and conduct conditional Bernoulli sampling. Compared to simple independent Bernoulli sampling, this increases the hardware logic overhead. While the paper claims that this does not increase the number of time steps, it does not evaluate the hardware cost.
  4. The experiment is limited to static image classification using the CIFAR dataset. There is a lack of testing on event cameras, temporal tasks, and larger datasets. USHE is a coding scheme designed to convert static images into spike sequences. Its applicability to native temporal event inputs is unknown.
  5. The primary benefits of USHE are only observed at very low latency (T=4). When T=8, the advantage diminishes significantly. However, in many real-world SNN applications, time steps greater than eight are commonly used. What would happen if the proposed method were applied to larger time steps and tasks involving temporal dependencies, such as speech recognition or long-video understanding?
Questions for authors

The authors should provide a detailed explanation of the aforementioned concerns.

3 comments

Thank you for distinguishing reconstruction optimality, downstream prediction, and certification. We address W1–W2 here, stronger attacks in Comment 2, and hardware cost, additional datasets and longer horizons in Comment 3.

W1 — When does reduced encoding variability improve classification?

To initiate the study of the effect of input variance on the classification performance, we provide the following downstream result: for a linear classifier of time-averaged input spikes, replacing Bernoulli encoding by USHE cannot increase expected cross-entropy at fixed classifier weights.

For input x∈[0,1]dx\in[0,1]^d, horizon TT and population size N≥TN\ge T, let SˉN=T−1∑t=1TSt(N)\bar S_N=T^{-1}\sum_{t=1}^T S_t^{(N)} be the average input spike vector. The class scores are zN=WSˉN+bz_N=W\bar S_N+b, with fixed weight matrix WW and bias vector bb. Input coordinates use independent encoding randomness. We recall that N=∞N=\infty denotes Bernoulli. For every xx, true label yy, and N≥TN\ge T,

E[CE(zT,y)]≤E[CE(zN,y)]≤E[CE(z∞,y)].\mathbb E[\mathrm{CE}(z_T,y)]\le\mathbb E[\mathrm{CE}(z_N,y)]\le\mathbb E[\mathrm{CE}(z_\infty,y)].

The proof uses convex ordering of spike counts, which is stronger than variance ordering as replacing independent coordinate counts by Bernoulli counts increases the expectation of any separately convex function. In particular, cross-entropy of a fixed affine readout has this property. The result extends to nonnegative normalized temporal weights, including leaky integration. Taking the supremum over the same input-perturbation set preserves the expected-loss ordering for this classifier class.

Connection to classification errors. Let us write wc⊤w_c^\top for row cc of WW and bcb_c for its bias, so class cc has score zN,c=wc⊤SˉN+bcz_{N,c}=w_c^\top\bar S_N+b_c. For the true class yy and a competing class k≠yk\ne y, let dk=wy−wkd_k=w_y-w_k. The random margin MN,k=zN,y−zN,kM_{N,k}=z_{N,y}-z_{N,k} measures how far the true class's score exceeds its competitor's. Its mean and variance are

μk=EMN,k=by−bk+dk⊤x,vN,k=Var⁡(MN,k)=∑jdkj2VT,N(xj),\mu_k=\mathbb E M_{N,k}=b_y-b_k+d_k^\top x,\qquad v_{N,k}=\operatorname{Var}(M_{N,k})=\sum_j d_{kj}^2V_{T,N}(x_j),

where dkjd_{kj} is the weight difference on input coordinate jj and VT,N(xj)=Var⁡(SˉN,j)V_{T,N}(x_j)=\operatorname{Var}(\bar S_{N,j}). If every mean margin μk\mu_k is positive, the one-sided Chebyshev inequality and a union bound give, for y^N=arg⁡max⁡czN,c\hat y_N=\arg\max_c z_{N,c},

\Pr(\hat y_N\ne y)\le\min\left�0�.

USHE preserves the mean margins while reducing their variance, so this error upper bound cannot increase. Although considered in a simplified setting, we hope to hint how the encoding variability is connected to classification loss and margin stability at fixed weights.

W2 — Certification despite dependent spike times

For x,x′∈[0,1]dx,x'\in[0,1]^d, let PxP_x be the full USHE spike-tensor distribution. For every N≥TN\ge T,

dTV(Px,Px′)≤min⁡{1,T∥x−x′∥1}.d_{\mathrm{TV}}(P_x,P_{x'})\le\min\{1,T\|x-x'\|_1\}.

Proof idea. We couple one-coordinate inputs p≤qp\le q using the same uniform UU in K(p)=⌊Np+U⌋K(p)=\lfloor Np+U\rfloor and the same random permutation of the latent population. The emitted bits are ordered and each differs with probability q−pq-p. Then, a union bound over input coordinates and timesteps proves the result without assuming independent spike times. The coefficient TT is sharp, as for example p=0p=0 and 0<q<1/N0<q<1/N give total variation exactly TqTq.

For any fixed downstream classifier ff, we define pc(x)=Pr⁡(f(SN(x))=c)p_c(x)=\Pr(f(S_N(x))=c) and g(x)=arg⁡max⁡cpc(x)g(x)=\arg\max_c p_c(x). If AA and BB are the most and second-most probable classes, then

g(x′)=Aif∥x′−x∥1<pA(x)−pB(x)2T.g(x')=A\quad\text{if}\quad\|x'-x\|_1<\frac{p_A(x)-p_B(x)}{2T}.

Each class probability changes by at most T∥x′−x∥1T\|x'-x\|_1, preserving the leading class within this radius. We note that this certificate applies arbitrary fixed downstream networks.

Relation to Bernoulli certification and latency. Mukhoty et al. (ICLR 2024, Theorem 5) give the Bernoulli radius (pA−pB)/(2T3T−1)(p_A-p_B)/(2T3^{T-1}). Our sharper TT coefficient also holds for Bernoulli by coupling its independent bits. Thus USHE's temporal dependence incurs no extra penalty in this universal bound. At a fixed margin, the guaranteed radius decreases as 1/T1/T, which somewhat was our original motivation to focus on low latency setting.

W2, continued — Stronger adaptive attacks and a gradient-free check

To address the empirical concern separately from certification, we completed PGD-100/EOT-20 for Bernoulli, USHE₄ and USHE₁₂ across both datasets and architectures, with three independently trained checkpoints per configuration (36 total). Each dataset uses a fixed, class-balanced 2,000-image cohort; ε=8/255, step size 0.01 and one restart. Accuracy (%) is mean ± sample SD across training seeds:

Dataset / architectureBernoulliUSHE₄USHE₁₂
CIFAR-10 / VGG-SNN26.70 ± 0.1020.01 ± 0.2725.36 ± 0.14
CIFAR-10 / ResNet-1931.38 ± 1.0625.30 ± 0.6430.87 ± 0.21
CIFAR-100 / VGG-SNN11.97 ± 0.159.34 ± 0.4211.23 ± 0.11
CIFAR-100 / ResNet-1916.71 ± 0.4113.98 ± 0.5616.54 ± 0.26

These attacks materially lower the original PGD-7 accuracies. USHE₁₂ retains substantially more attacked accuracy than USHE₄, while it remains competitive with Bernoulli. Together with the replicated clean gains, we argue that this supports a population-controlled clean–robustness trade-off of USHE encoders.

Attack-budget sensitivity. The complete PGD-100 sweep below uses a separate 200-image, class-balanced calibration cohort and training seed 1001. E denotes EOT samples; R5 denotes five restarts. Other columns use one restart; α=0.01 unless indicated. All encoders within a setting use the same images.

Setting / encoderE20E50E100E20, R5E20, α=2/255
C10/VGG / Bernoulli26.7626.6926.4126.6026.79
C10/VGG / USHE₄21.6621.4121.4321.2221.59
C10/VGG / USHE₁₂26.9726.7026.5226.6627.16
C10/ResNet19 / Bernoulli32.1432.1531.7032.2132.15
C10/ResNet19 / USHE₄27.0926.8826.5427.0726.91
C10/ResNet19 / USHE₁₂32.7431.9032.0532.4932.52
C100/VGG / Bernoulli11.2711.1811.2011.2011.31
C100/VGG / USHE₄10.1110.0310.0510.2110.25
C100/VGG / USHE₁₂12.3012.3012.3012.2412.46
C100/ResNet19 / Bernoulli15.1115.0415.0315.1815.08
C100/ResNet19 / USHE₄15.2515.4815.1015.2315.56
C100/ResNet19 / USHE₁₂15.0715.2315.0315.2715.18

Increasing EOT from 20 to 100 reduces accuracy by 0.26 percentage points on average (maximum 0.68) while five restarts at EOT-20 reduce it by 0.07 on average. USHE₁₂ exceeds Bernoulli by 1.03–1.15 points across the five CIFAR-100/VGG schedules on this cohort.

Forward/backward and scoring. Attacks retain the actual sampled encoder in the forward pass and use the expected-rate BPDA surrogate in backward computation. EOT averages per-realization cross-entropy gradients as it reduces sampling variability without making BPDA exact. Candidate selection uses 32 separate draws and final scoring 128 fresh draws. Each scoring draw produces one prediction. Finally, the correctness indicators are averaged, rather than ensembling logits or votes.

Additional attack strategies. We completed a custom APGD-CE-style optimizer (100 steps/EOT-10) and a gradient-free square-patch search on common 500-image cohorts, using seed 1001. The latter covers CIFAR-100 and uses no BPDA. The black-box search uses 300 proposals and eight draws per incumbent/proposal score. In this setting neither produced lower aggregate accuracy than PGD on the same images.

We also provided positive baseline comparisons beyond Bernoulli. On CIFAR-100, USHE₄ exceeds deterministic rounded-count encoding (DRC) by 1.67/3.65 points under strong PGD, 1.04/3.62 under APGD, and 8.99/5.94 under gradient-free search (VGG/ResNet, seed 1001, matched within each attack). The three-seed, full-test PGD-7 gains over DRC are 2.97 ± 0.53/4.79 ± 0.63 points. Cyclic-burst results closer to USHE₄ in clean accuracy (indicating that low count variance can remain useful with a different temporal arrangement), however the comeptiteveness is lost in the attacks settings.

The replicated full-test RFGSM comparison also favors USHE₄ on CIFAR-10/ResNet-19: 69.58 ± 0.39% versus 66.89 ± 0.34% for Bernoulli. These attack-dependent advantages are reported alongside the stronger-PGD trade-off, to provide a more precise picture.

W3 — What extra logic does neuromorphic USHE encoding require?

We give an operation/state account of the encoding front-end, separate from downstream spiking computation.

For an input intensity p, initialize K=floor(Np)+B, with B Bernoulli of parameter Np−floor(Np). At step t, let m=N−t+1, draw J uniformly from {1,…,m}, emit s=1{J≤K}, and set K←K−s. This realizes probability K/m without division in the per-step probability calculation. Here, the denominator m is shared across synchronized coordinates.

ResourceBernoulliUSHE
InitializationInput probability pForm Np and stochastically round
Per-step samplingUniform draw, compare with pBounded-integer draw, compare with K
Count updateNoneK←K−s
Extra count state per coordinateNone⌈log₂(N+1)⌉ bits
Visible steps / expected input spikesT / TpT / Tp

The counter needs three bits at N=4 and four at N=12. Beyond initial rounding, at most T conditional draws are needed as deterministic K=0 or K=m cases can bypass sampling. Both encoders use O(Td) sampling/comparison operations for d inputs ehilr USHE adds O(d log N) counter bits.

For input fan-outs f_j, unbiased marginals give the same expected first-layer synaptic-event count:

E[input synaptic events]=T∑jfjxj.\mathbb E[\mathrm{input\ synaptic\ events}]=T\sum_j f_jx_j.

Total energy also depends on initialization precision, sampling circuitry, routing and hidden-layer activity, which temporal correlations may change. The added logic resides in the front-end if encoding is off-chip, and requires device support if on-chip. We have not implemented USHE on a neuromorphic chip, so this analysis specifies operations and state without asserting a measured energy ratio.

W4 — Additional datasets and the scope of temporal inputs

We can now add matched SVHN and Tiny ImageNet evaluations, extending the static-image evidence beyond CIFAR. Clean accuracy (%) is shown for the common endpoint N=T:

Dataset / architectureTBernoulliUSHE N=TGain (pp)
Tiny ImageNet / ResNet-18448.9452.43+3.49
Tiny ImageNet / ResNet-18852.8455.67+2.84
SVHN / VGG-SNN486.9992.55+5.56
SVHN / VGG-SNN890.2993.58+3.28
SVHN / ResNet-19488.5994.16+5.57
SVHN / ResNet-19891.9495.29+3.34

These use one training seed (1000) per configuration, averaging repeated single-encoding evaluations. All tested USHE population sizes improve clean accuracy over Bernoulli on both datasets. On SVHN, USHE at T=4 also exceeds Bernoulli at T=8 by 2.25/2.22 points (VGG/ResNet), using half as many simulated steps.

USHE targets static intensities converted into short binary spike trains. Native events already carry timing information. Speech/video or time-varying intensities require a specified window representation and budget-reset rule, followed by temporal-task evaluation. We have not tested those extensions.

W5 — Do the benefits disappear beyond T=4?

The additional T=8 results above retain clean gains of 2.84–3.34 points. The submitted CIFAR T=8 results also retain gains in all four settings (one training seed; full-test accuracy, %):

Dataset / architectureBernoulliUSHE₈Gain (pp)
CIFAR-10 / VGG-SNN82.9086.30+3.40
CIFAR-10 / ResNet-1983.7388.31+4.58
CIFAR-100 / VGG-SNN56.4261.33+4.91
CIFAR-100 / ResNet-1957.4663.47+6.01

Our new CIFAR T=4 clean gains of 5.67–6.56 points remain positive across all three training seeds. We therefore observe clean benefits at both tested horizons.

The reconstruction guarantee holds for every integer horizon. With r_T the fractional part of Tp,

VT,T(p)=rT(1−rT)T2≤14T2,VT,∞(p)=p(1−p)T.V_{T,T}(p)=\frac{r_T(1-r_T)}{T^2}\le\frac{1}{4T^2},\qquad V_{T,\infty}(p)=\frac{p(1-p)}{T}.

For fixed 0<p<1, USHE's endpoint variance bound decreases faster with T. Both codes become more accurate rate estimators; downstream classification still depends on the model and task.

Reviewer_zRs4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes USHE, a family of stochastic spike encoders for static images that keeps the binary alphabet, the per-step Bernoulli marginals and the simulation horizon fixed, but controls the temporal dependence between spikes. For an intensity pp, USHEN\text{USHE}_N stochastically rounds NpNp to an integer budget KK and then samples the TT visible spikes without replacement from a latent population of N≥TN \ge T positions containing KK ones; a streaming implementation only needs the remaining budget per pixel. The theory derives the exact rate variance, shows that USHEN\text{USHE}_N is the unique variance minimizer among exchangeable length-NN laws with mean pp (and is globally optimal at N=TN = T), extends this to all convex rate losses and to a minimax statement, and proves convergence in total variation to independent Bernoulli coding as N→∞N \to \infty at fixed TT.

Critical strengths
  • The idea is simple and clean. USHE keeps the binary alphabet, every step's Bernoulli marginal, the expected input activity and the latency fixed, and changes only the temporal dependence through one integer knob NN. It's a drop-in encoder with no learned parameters and a cheap streaming sampler (O(Td)O(Td) work, O(d)O(d) state), and it isolates temporal dependence as a variable in a way that signed coding, Gaussian input noise or learned input transforms don't.
  • The theory is careful and complete for what it claims: exact variance, uniqueness of the minimizer within the exchangeable class, global optimality at N=TN = T, convex-order and minimax versions, and an explicit total-variation bound of order T2/(2N)T^2/(2N) for the Bernoulli limit. The authors are also clear that these guarantees concern rate reconstruction, not classification or robustness.
Critical weaknesses
  • The robustness evaluation isn't strong enough for a stochastic encoder, and robustness is half of the paper's story. Gradients come from a single stochastic realization through an expected-rate BPDA surrogate, PGD uses only 7 steps with one restart, APGD-CE has a single seed and is described as still being extended, and there's no EOT, no randomized AutoAttack, and no attack tailored to the hypergeometric sampling. The observed pattern, where more randomness (larger NN) gives higher PGD accuracy but lower clean accuracy, is exactly what you'd expect if randomness is mainly weakening a weak attack. Before drawing conclusions about intermediate-NN robustness, I'd want EOT over many samples (e.g., 20–100), PGD with 50–100 steps and restarts, a gradient-free attack such as Square, and the usual sanity checks (accuracy going to zero as ϵ\epsilon grows).
  • The statistical support is thin. Every configuration uses one training seed, and the reported standard deviations only reflect encoder sampling at evaluation time. That's fine for the clean-accuracy gains, which are large, but many of the robustness differences the text interprets are 0.1–2 points (e.g., R-FGSM gains of 0.08–0.10 points on CIFAR-100/VGG-SNN, or 69.90 vs. 69.78 for ResNet-19 at T=8T = 8), well within typical training-seed variance on CIFAR. Claims such as "intermediate populations improve attacked accuracy" and the architecture dependence need at least three training seeds.
  • Table 3 overstates USHE's advantage over robust SNN methods, and key baselines are missing. Plain Bernoulli coding isn't listed, yet under the paper's own protocol it reaches 44.89% and 24.16% PGD-7 accuracy on ResNet-19 at T=4T = 4, already above every listed baseline, and on the two VGG blocks it beats USHE12\text{USHE}_{12} (38.13 vs. 36.89 and 18.26 vs. 17.22). So the "highest PGD accuracy in all four blocks" claim comes from stochastic rate coding plus a weak attack, not from USHE, and the table also mixes published numbers from different architectures (VGG-11 vs. VGG-SNN), horizons and attack protocols. What's really missing is a frontier comparison: other low-variance binary codes (deterministic round(Tp)\mathrm{round}(Tp) counts with random ordering, sigma-delta or error-diffusion coding) and other randomness knobs (Bernoulli or direct input with Gaussian noise, RSC-SNN), traced on the same clean-vs-robust plot, to show that tuning NN gives a better trade-off than simply dialing the noise level.
Questions for authors
  1. With EOT over many samples (e.g., 20–100), PGD with 50–100 steps and restarts, and a randomized AutoAttack, does the intermediate-NN advantage over Bernoulli survive?
  2. How exactly is the expected-rate BPDA surrogate applied? Since E[st]=p\mathbb{E}[s_t] = p for every NN, is the backward pass identical across encoders, and is the sample used for evaluation the same one used to compute the gradient?
  3. Can you report at least three training seeds for Tables 1 and 2?
2 comments

Thank you for the detailed evaluation recommendations and for recognizing that USHE isolates temporal dependence while preserving binary marginals and latency. We followed your proposed attack-budget checks and added training replication and alternative encoders. This comment answers Q1–Q2; the next addresses Q3 and the comparison concerns.

Q1 / W1 — Does the intermediate-population advantage survive stronger attacks?

We tested whether more optimization, more stochastic gradient samples, and additional restarts change the results. The table below shows the complete PGD-100 sensitivity sweep on the same 200 class-balanced images per dataset, using training seed 1001 and T=4. E denotes EOT samples; R5 denotes five restarts. All other columns use one restart; ε=8/255 and α=0.01 unless indicated. Entries are accuracies (%).

Setting / encoderE20E50E100E20, R5E20, α=2/255
C10/VGG / Bernoulli26.7626.6926.4126.6026.79
C10/VGG / USHE₄21.6621.4121.4321.2221.59
C10/VGG / USHE₁₂26.9726.7026.5226.6627.16
C10/ResNet19 / Bernoulli32.1432.1531.7032.2132.15
C10/ResNet19 / USHE₄27.0926.8826.5427.0726.91
C10/ResNet19 / USHE₁₂32.7431.9032.0532.4932.52
C100/VGG / Bernoulli11.2711.1811.2011.2011.31
C100/VGG / USHE₄10.1110.0310.0510.2110.25
C100/VGG / USHE₁₂12.3012.3012.3012.2412.46
C100/ResNet19 / Bernoulli15.1115.0415.0315.1815.08
C100/ResNet19 / USHE₄15.2515.4815.1015.2315.56
C100/ResNet19 / USHE₁₂15.0715.2315.0315.2715.18

USHE₁₂ exceeds Bernoulli on CIFAR-100/VGG under all five schedules, by 1.03–1.15 percentage points. More broadly, increasing EOT from 20 to 100 reduces accuracy by 0.26 points on average over the twelve configurations (maximum 0.68); five restarts at EOT-20 reduce it by 0.07 points on average. The major reduction occurs earlier: CIFAR-10/ResNet/USHE₁₂ falls from 46.99% under PGD-7/EOT-1 to 32.74% under PGD-100/EOT-20 on the same cohort.

Larger, independently replicated evaluation. We then evaluated PGD-100/EOT-20 on 2,000 fixed, class-balanced images per dataset, disjoint from calibration, across all 36 checkpoints. Accuracy is mean ± SD over three training seeds:

Dataset / architectureBernoulliUSHE₄USHE₁₂
CIFAR-10 / VGG-SNN26.70 ± 0.1020.01 ± 0.2725.36 ± 0.14
CIFAR-10 / ResNet-1931.38 ± 1.0625.30 ± 0.6430.87 ± 0.21
CIFAR-100 / VGG-SNN11.97 ± 0.159.34 ± 0.4211.23 ± 0.11
CIFAR-100 / ResNet-1916.71 ± 0.4113.98 ± 0.5616.54 ± 0.26

The two encoders Bernoulli and USHE12 remain competitive in this setting, while we note lower variability of the USHE12.

Alternative attacks. We completed custom APGD-CE-style optimization (100 steps/EOT-10) and gradient-free square-patch search (300 proposals, eight draws per score), using seed 1001 and common 500-image cohorts; the latter covers CIFAR-100. Neither lowers aggregate accuracy below PGD on identical images. Under gradient-free search, USHE₄ exceeds Bernoulli by 5.26/5.74 points on VGG/ResNet, although this search is weaker than PGD.

Q2 — What exactly is shared in BPDA, and is scoring independent?

For sampled spikes S_N(x,ω), let r(x) repeat x over T timesteps. We use

S~N=sg⁡(SN)+[r−sg⁡(r)],\widetilde S_N=\operatorname{sg}(S_N)+[r-\operatorname{sg}(r)],

where sg stops gradients. The forward pass retains the actual sampled spikes. The encoder's surrogate Jacobian is therefore identical for Bernoulli and every USHE population. The complete input-loss gradient is not identical as it is evaluated through the network at different sampled spike trajectories, activations and membrane states, and separately trained weights also differ.

At each step, EOT averages gradients of per-realization cross-entropy over independently sampled encodings. This reduces Monte Carlo variability but does not make BPDA an exact derivative of the expected discrete loss. We retain the actual hypergeometric forward sampling.

Optimization, candidate selection (32 draws), and final scoring (128 fresh draws) use separate random streams. Each scoring draw averages logits over T timesteps, makes one prediction, and contributes a correctness indicator. Finally, we do not average logits or votes across the 128 draws, i.e. the evaluated classifier still uses one encoding per prediction, independently of the attack samples.

Q3 / W2 — Three independently trained models for Tables 1 and 2

Yes. Bernoulli, USHE₄ and USHE₁₂ now have training seeds 1000/1001/1002 at T=4 across both datasets and architectures, in total 36 checkpoints. Clean, FGSM, RFGSM, PGD-7 and Gaussian-corruption evaluations use all 10,000 test images. The strong-PGD table provided above uses these same checkpoints.

Both USHE populations improve dedicated full-test clean accuracy for every seed: mean gains are 5.67–6.56 points for USHE₄ and 1.27–1.87 for USHE₁₂. Below are paired gains over Bernoulli (percentage points; mean ± sample SD across three training-seed pairs):

SettingUSHE₄ cleanUSHE₄ RFGSMUSHE₁₂ RFGSM
C10 / VGG-SNN+5.67 ± 0.18+0.70 ± 0.09+0.74 ± 0.33
C10 / ResNet-19+6.28 ± 0.30+2.69 ± 0.06+1.17 ± 0.57
C100 / VGG-SNN+6.30 ± 0.09+0.39 ± 0.47+0.08 ± 0.11
C100 / ResNet-19+6.56 ± 0.41+1.65 ± 0.59+0.83 ± 0.05

SD is across checkpoint means after averaging inference repetitions. Classical attacks use one attack per checkpoint and three fresh scoring draws, so SD includes residual evaluation noise.

Additional datasets. Using one training seed (1000), USHE at N=T=4 improves clean accuracy over Bernoulli by 3.49 points on Tiny ImageNet/ResNet-18 and 5.56/5.57 on SVHN/VGG-SNN/ResNet-19. Gains at N=T=8 remain 2.84–3.34 points.

W3 — Table 3 and the requested clean–robustness comparison

Omitting Bernoulli results in Table 3 was a blunder on our side.

In addition, we tested three complementary alternatives under the same CIFAR-100 training recipes: DRC fixes K=floor(Tp+1/2) and uniformly places its spikes while cyclic burst uses USHE₄'s unbiased count law but a randomly shifted contiguous placement. Finally, Gaussian-Bernoulli perturbs the input probability before Bernoulli sampling during both training and inference.

The following clean/strong-PGD operating points use seed 1001 and exactly the same 2,000 images for every column. PGD uses 100 steps/EOT-20, ε=8/255, α=0.01 and one restart. These are single-checkpoint estimates, not three-seed means.

EncoderVGG cleanVGG PGDResNet cleanResNet PGD
Bernoulli51.0511.9050.5016.62
USHE₄57.479.6457.8714.03
USHE₁₂52.3211.1352.3316.31
DRC56.477.9857.0310.38
Cyclic burst56.509.3657.5613.96
Gaussian-Bernoulli, σ=8/25550.9412.7550.9217.79

Deterministic rounding: USHE₄ improves both coordinates over DRC here; its strong-PGD gains are +1.67/+3.65 points. The three-seed, full-test PGD-7 gains are +2.97 ± 0.53/+4.79 ± 0.63 points. The advantage also persists under APGD (+1.04/+3.62) and the gradient-free search (+8.99/+5.94), both on matched 500-image cohorts with seed 1001.

Temporal ordering: cyclic burst is close to USHE₄. Thus this particular rearrangement preserves much of the count-controlled benefit, however the competitiveness is lost in the attacks.

Added input noise. We compared Gaussian-Bernoulli at σ=4/255, 8/255 and 16/255 on the full CIFAR-100 test set, using training seed 1001 for every encoder. Added noise improves PGD accuracy in several settings, but it does not dominate USHE across metrics. On ResNet-19, USHE₄ achieves 41.41% RFGSM accuracy versus 39.19%, 39.55% and 39.18% for the three noise levels; its clean accuracy is 57.85% versus 50.33–50.92%. Against σ=8/255, USHE₄ gains 1.86 points under RFGSM and 6.35 under external Gaussian corruption, while trailing by 1.70 points under FGSM and 3.22 under PGD-7.

Reviewer_G4PKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Unbiased Spiking Hypergeometric Encoding (USHE), a probabilistic input encoding framework for low-latency spiking neural networks (SNNs). USHE introduces a latent population size to control temporal dependence and spike-rate variance while preserving unbiased input representation, Bernoulli marginals, and a fixed simulation horizon. Experiments demonstrate consistent improvements in clean accuracy over Bernoulli encoding and investigate the effects of population size on adversarial robustness.

Critical strengths

S1. The paper addresses a meaningful problem in low-latency SNNs. Controlling the stochasticity of input spike encoding without increasing the simulation horizon is an interesting and important direction.

S2. The proposed encoding framework is mathematically well characterized. The analysis results provide a solid theoretical foundation for the proposed encoding family.

S3. The experimental results demonstrate consistent improvements in clean accuracy. USHE outperforms Bernoulli encoding across the evaluated datasets, architectures, and simulation horizons.

Critical weaknesses

W1. The adversarial robustness evaluation does not sufficiently account for the stochastic nature of the proposed encoding. Although the paper considers several attacks, Appendix B.2 indicates that gradient estimation generally relies on a single stochastic encoding realization with an expected-rate BPDA surrogate. For a stochastic encoder such as USHE, this approximation may not adequately capture the gradient of the expected attack objective. Consequently, the reported adversarial accuracies may be affected by gradient estimation errors. A stronger evaluation using Expectation over Transformation (EOT) with multiple stochastic samples, ideally complemented by an alternative attack strategy, would provide more reliable evidence for the robustness claims.

W2. The distinct benefits of USHE over alternative low-variance spike encoding strategies are not sufficiently established. The theoretical analysis demonstrates that USHE achieves minimum spike-rate reconstruction variance under the specified constraints. However, as acknowledged in Corollary 3.3, the optimal spike-count distribution does not uniquely determine the temporal arrangement of spikes. Different encoding strategies could therefore achieve the same reconstruction variance while producing different temporal patterns. Since SNN dynamics depend on the timing of incoming spikes, it remains unclear whether the observed accuracy improvements arise primarily from reduced spike-count variance or from the particular temporal distribution induced by USHE.

W3. As stated in Section 4.1, each configuration is evaluated using a single trained checkpoint, with multiple evaluation seeds used to measure stochastic inference variability. Therefore, the reported standard deviations do not capture variability arising from model initialization and training. While the clean accuracy improvements over Bernoulli encoding are relatively substantial and consistent, several adversarial robustness improvements are considerably smaller. Without independent training runs, it is difficult to determine whether these differences are stable.

Questions for authors

Q1. How do the robustness results change under stronger adaptive attacks using EOT with multiple stochastic realizations? Can the authors also clarify the different prediction aggregation rules used for PGD and APGD? Q2. How does USHE compare with alternative spike encoding strategies that achieve the same spike-count variance but use different temporal arrangements? Q3. Are the reported improvements, particularly in adversarial robustness, consistent across independently trained models with different random seeds?

2 comments

Thank you for recognizing the motivation, mathematical analysis, and clean-accuracy gains. We address your three questions in two comments: this comment covers adaptive evaluation and aggregation (Q1/W1); the next covers temporal controls and training replication (Q2/W2, Q3/W3).

Q1 / W1 — Stronger adaptive attacks

We tested whether more optimization, more stochastic gradient samples, and additional restarts change the results. The table below shows the complete PGD-100 sensitivity sweep on the same 200 class-balanced images per dataset, using training seed 1001 and T=4. E denotes EOT samples; R5 denotes five restarts. All other columns use one restart; ε=8/255 and α=0.01 unless indicated. Entries are accuracies (%).

Setting / encoderE20E50E100E20, R5E20, α=2/255
C10/VGG / Bernoulli26.7626.6926.4126.6026.79
C10/VGG / USHE₄21.6621.4121.4321.2221.59
C10/VGG / USHE₁₂26.9726.7026.5226.6627.16
C10/ResNet19 / Bernoulli32.1432.1531.7032.2132.15
C10/ResNet19 / USHE₄27.0926.8826.5427.0726.91
C10/ResNet19 / USHE₁₂32.7431.9032.0532.4932.52
C100/VGG / Bernoulli11.2711.1811.2011.2011.31
C100/VGG / USHE₄10.1110.0310.0510.2110.25
C100/VGG / USHE₁₂12.3012.3012.3012.2412.46
C100/ResNet19 / Bernoulli15.1115.0415.0315.1815.08
C100/ResNet19 / USHE₄15.2515.4815.1015.2315.56
C100/ResNet19 / USHE₁₂15.0715.2315.0315.2715.18

USHE₁₂ exceeds Bernoulli on CIFAR-100/VGG under all five schedules, by 1.03–1.15 percentage points. More broadly, increasing EOT from 20 to 100 reduces accuracy by 0.26 points on average over the twelve configurations (maximum 0.68); five restarts at EOT-20 reduce it by 0.07 points on average. The major reduction occurs earlier: CIFAR-10/ResNet/USHE₁₂ falls from 46.99% under PGD-7/EOT-1 to 32.74% under PGD-100/EOT-20 on the same cohort.

Larger, independently replicated evaluation. We then evaluated PGD-100/EOT-20 on 2,000 fixed, class-balanced images per dataset, disjoint from calibration, across all 36 checkpoints. Accuracy is mean ± SD over three training seeds:

Dataset / architectureBernoulliUSHE₄USHE₁₂
CIFAR-10 / VGG-SNN26.70 ± 0.1020.01 ± 0.2725.36 ± 0.14
CIFAR-10 / ResNet-1931.38 ± 1.0625.30 ± 0.6430.87 ± 0.21
CIFAR-100 / VGG-SNN11.97 ± 0.159.34 ± 0.4211.23 ± 0.11
CIFAR-100 / ResNet-1916.71 ± 0.4113.98 ± 0.5616.54 ± 0.26

The two encoders Bernoulli and USHE12 remain competitive in this setting, although we do note lower variability of the USHE12 encoder.

Alternative attacks. We also completed custom APGD-CE-style optimization (100 steps, EOT-10) and a gradient-free square-patch search on matched 500-image cohorts with seed 1001; the gradient-free check covers CIFAR-100. Neither produced lower aggregate accuracy than PGD on the same images. The next comment reports the encoder comparisons, including DRC and cyclic burst. EOT reduces sampling variability in the BPDA gradient estimate. In this setting the backward surrogate remains an approximation.

Prediction aggregation. The original PGD used mean logits across encoding draws, while APGD used mean probabilities. With one encoding draw per prediction, these rules have the same top-1 decision: argmax(z)=argmax(softmax(z)); logits are first averaged over simulation timesteps. The original summed-logit and individual-CE objectives also coincide at one draw. In the new gradient attacks, EOT averages per-realization cross-entropy, not cross-entropy of averaged logits. PGD and APGD use 32 separate draws for candidate selection and 128 fresh draws for final scoring. Each scoring draw makes one prediction, and we average correctness indicators—not logits or votes. The evaluated classifier therefore still uses one encoding per prediction.

Q2 / W2 — What changes when the count law is held fixed?

Thank you for highlighting the distinction between spike counts and temporal order. We trained a cyclic-burst control using the same stochastic-rounding law for the count as USHE₄, but placing the spikes contiguously with a uniform random cyclic shift. This preserves the binary alphabet, each timestep's Bernoulli marginal, and the entire count distribution, while changing temporal arrangement. The training recipe is matched.

Full-test clean accuracy on CIFAR-100 (mean ± SD over three training seeds, dedicated clean evaluations) is:

EncoderVGG-SNNResNet-19
USHE₄57.90 ± 0.0457.89 ± 0.11
Cyclic burst57.59 ± 0.0857.79 ± 0.31

The strong-attack comparison below includes Bernoulli, both USHE populations, cyclic burst, and deterministic rounded-count encoding (DRC). DRC fixes K=round(Tp) and distributes its spikes uniformly; unlike cyclic burst, it changes the count law and generally introduces bias.

CIFAR-100, T=4, seed 1001. PGD: 100 steps/EOT-20, 2,000 images. APGD: custom CE-style optimizer, 100 steps/EOT-10, 500 images. GF: custom gradient-free square-patch search, the same 500 images, 300 proposals and eight stochastic draws per incumbent/proposal score. All use ε=8/255. Every within-column comparison uses identical images across encoders; PGD uses the larger cohort. Results average 128 fresh single-encoding predictions, not training seeds.

EncoderVGG: PGDVGG: APGDVGG: GFResNet: PGDResNet: APGDResNet: GF
Bernoulli11.9011.5236.8716.6215.4940.74
USHE₄9.649.6142.1314.0313.8246.48
USHE₁₂11.1310.6037.3316.3115.9942.62
DRC7.988.5733.1410.3810.2140.55
Cyclic burst9.368.7841.9713.9613.4945.96

USHE versus deterministic rounding. USHE₄ exceeds DRC under strong PGD by 1.67 points on VGG and 3.65 on ResNet. Its advantage also appears under APGD (+1.04/+3.62) and gradient-free search (+8.99/+5.94). In the three-training-seed, full-test classical evaluation, the paired PGD-7 gains are +2.97 ± 0.53 and +4.79 ± 0.63 points. Thus, the DRC contrast persists across the tested attack methods.

USHE versus count-matched rearrangement. Cyclic burst remains much closer to USHE₄: strong-PGD differences are +0.28/+0.07 points. This particular rearrangement preserves much of the clean gain over Bernoulli, so those gains do not require USHE's uniform temporal placement. We note that USHE additionally supplies the exchangeable, population-indexed family with exact distributional guarantees and tunable variability at fixed latency.

Attack-dependent advantages over Bernoulli. Under the gradient-free search, USHE₄ exceeds Bernoulli by +5.26/+5.74 points and USHE₁₂ by +0.45/+1.88. These results complement, rather than replace, the stronger-PGD findings.

Q3 / W3 — Do the gains persist across training seeds?

We added independent training seeds 1001 and 1002 to seed 1000. For Bernoulli, USHE₄ and USHE₁₂ this gives 36 checkpoints at T=4, with full-test clean/classical-attack evaluation and the replicated strong-PGD evaluation in our first comment. DRC and cyclic burst each also have three training seeds on both CIFAR-100 architectures for clean/classical attacks; their strong-attack checks above use seed 1001.

Both USHE populations improve full-test clean accuracy for every seed in all four settings. Mean gains are 5.67–6.56 points for USHE₄ and 1.27–1.87 for USHE₁₂. We average scoring repetitions within a checkpoint and then report sample SD across the three checkpoint means, separately from inference variability.

Additional datasets. Using one training seed (1000), USHE at N=T=4 improves clean accuracy over Bernoulli by 3.49 points on Tiny ImageNet/ResNet-18 and 5.56/5.57 on SVHN/VGG-SNN/ResNet-19. Gains at N=T=8 remain 2.84–3.34 points.

Some adversarial gains also reproduce: under RFGSM, CIFAR-10/ResNet-19 gives 69.58 ± 0.39% for USHE₄ versus 66.89 ± 0.34% for Bernoulli, with a paired gain of +2.69 ± 0.06 points, positive in all three seeds. USHE₁₂ has higher mean RFGSM accuracy in all four settings. In general, our conclusions remain for USHE family in form of an attack-dependent clean–robustness trade-off.

Comments on the paper

1 comment

Thank you for the careful reviews and constructive questions. Your concerns about stochastic evaluation, training variability, and alternative encoders motivated additional experiments and theoretical analyses. We summarize the shared updates here and address each review separately.

1. Training replication and additional datasets. We evaluated Bernoulli, USHE₄, and USHE₁₂ at T=4 across CIFAR-10/100 and VGG-SNN/ResNet-19 using three training seeds per configuration (36 checkpoints). Clean accuracy and the original FGSM/RFGSM/PGD-7/Gaussian-corruption evaluations cover the full test sets. We distinguish variation across trained models from stochastic inference variation. USHE₄ improves mean clean accuracy over Bernoulli by 5.67-6.56 percentage points across the four settings, with positive gains for every training seed.

In addition, we trained on new datasets SVHN and Tiny ImageNet for one training seed (1000). In this setting clean gains over Bernoulli at N=T=4 are 3.49 points on Tiny ImageNet/ResNet-18 and 5.56/5.57 on SVHN/VGG-SNN/ResNet-19. At N=T=8, gains remain 2.84–3.34 points. On SVHN, USHE at T=4 also exceeds Bernoulli at T=8.

2. Stronger stochastic-model evaluation. We completed PGD-100/EOT-20 on a common, class-balanced set of 2,000 images per CIFAR dataset for all 36 checkpoints. A separate 200-image sensitivity study evaluates EOT-50/100, five restarts, and an alternative step size. We also completed supplementary APGD-CE-style optimization and a custom gradient-free square-patch search on matched 500-image cohorts (these supplementary checks use training seed 1001, and the gradient-free check covers CIFAR-100). Attack generation and final scoring use separate random streams. Stronger PGD substantially lowers the original attacked accuracies and shows a clean-robustness trade-off, rather than consistent adversarial superiority over Bernoulli. We report this alongside the replicated clean gains and attack-dependent results.

3. Alternative encoding mechanisms. We evaluated deterministic rounded-count encoding with uniform temporal placement, count-matched cyclic bursts, and Gaussian-randomized Bernoulli encoding. DRC and cyclic burst have three-seed CIFAR-100 clean and original-attack results; their supplementary strong attacks use seed 1001. On CIFAR-100, USHE₄ exceeds DRC under PGD-100/EOT-20 by 1.67 points on VGG-SNN and 3.65 points on ResNet-19. Cyclic burst performs similarly to USHE₄ under this attack. These controls distinguish deterministic count quantization, temporal arrangement, and added input randomness rather than attributing all downstream behavior to reconstruction variance.

4. When encoding guarantees imply a classification benefit. For a linear classifier applied to time-averaged input spikes, with its weights held fixed, we prove that replacing Bernoulli encoding by USHE cannot increase expected cross-entropy. We also show that lower encoding variance reduces the variance of its decision margins, yielding a tighter upper bound on misclassification probability when the mean margins are positive. This gives an explicit link from encoding statistics to prediction for a specified classifier class, rather than assuming that reconstruction optimality guarantees better performance for every deep SNN.

5. Certified robustness despite temporal dependence. We prove that USHE's temporal dependence does not prevent certification: for any fixed downstream classifier, the most probable predicted class is unchanged by any input perturbation whose ℓ₁ norm is smaller than the gap between the two largest class probabilities divided by 2T. This certifies the probability-majority classifier, rather than each individual stochastic prediction.

6. Neuromorphic encoding cost. We analyze USHE through its encoding operations and state requirements. Each input coordinate needs one initial stochastic rounding and a remaining-spike counter with ⌈log₂(N+1)⌉ bits. At each visible timestep, an implementation can draw a uniform index from the remaining population, compare it with the spike budget, and decrement the budget when a spike is emitted. The remaining population size is shared across coordinates. USHE keeps the same T simulation steps and expected input spike count as Bernoulli, hence the same expected first-layer synaptic-event count at matched fan-out. Actual latency and energy depend on the target hardware and its random-number implementation.

The additional evidence preserves the central contribution: an analytically characterized family controlling binary temporal dependence at fixed latency, with reproducible clean-accuracy gains and explicitly evaluated downstream trade-offs. We appreciate the opportunity to discuss the detailed results and proofs in the individual reviewer threads.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.