acceptodds
Under review as a conference paper at ICLR 2027

Learning Stable Spike Representations for Robust ANN-to-SNN Conversion

Abstract

Spiking neural networks (SNNs) enable event-driven computation with the potential for energy-efficient inference, but training deep SNNs remains costly. ANN-to-SNN conversion offers a practical alternative by transferring pretrained artificial neural networks (ANNs) to spiking models. Existing work has largely focused on preserving post-conversion accuracy on clean inputs, while the robustness of converted SNNs to noises remains less understood. In this paper, we ask which properties of the source ANN determine the stability of the resulting SNN. We show that post-conversion stability is governed by the interaction between perturbation amplification and the spacing of discrete spike-count states. This analysis yields a layerwise robustness surrogate and motivates **S**pike-**S**tate **S**tability **R**egularizer (SSSR), a conversion-aware source-training regularizer that requires neither noise injection nor modification of the conversion rule. Across VGG-16 and ResNet-18 on CIFAR-10/100, SSSR improves strong-noise accuracy over the strongest standard L1/L2 baseline by up to 50.17 percentage points, while remaining within 2.16 points of clean accuracy across all four settings. These results show that post-conversion robustness can be shaped directly during the training of source-ANN models. Our code is available at https://anonymous.4open.science/r/sssr-4E4C/.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking neural networks, ANN-to-SNN conversion, Neural network regularization, Neuromorphic computing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_j9WjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the robustness of SNNs obtained through ANN-to-SNN conversion, focusing on the stability of finite-time spike-count representations under internal perturbations. The authors analyze how local state-transition probability depends on perturbation amplification and the spacing between adjacent discrete activation states. Based on this analysis, they introduce the Spike-State Stability Regularizer (SSSR), a conversion-aware regularization objective that penalizes the ratio between effective weight amplification and squared activation-state spacing. The method jointly adapts the weights of matched layers and the learnable activation thresholds during source ANN training, without requiring noise injection during source ANN training or modifications to the conversion procedure. The authors evaluate SSSR on VGG-16 and ResNet-18 using CIFAR-10 and CIFAR-100, with an additional ResNet-18 experiment on ImageNet. Comparisons with standard L1 and L2 regularization baselines, together with layerwise representation-stability analyses and ablation studies, show improved accuracy under internal Gaussian perturbations.

Critical strengths
  1. The paper provides a conceptually motivated approach to improving post-conversion robustness by explicitly accounting for both effective weight-induced perturbation amplification and the spacing between discrete spike-count states. The proposed SSSR translates this analysis into a trainable regularizer that jointly adapts weights and activation thresholds, without requiring noise injection during training or modifications to the conversion procedure. This offers a potentially useful way to incorporate the properties of finite-time spiking representations into conventional ANN training.

  2. The experiments on VGG-16 and ResNet-18 across CIFAR-10 and CIFAR-100 show substantial improvements over standard L1/L2 regularization baselines under strong internal Gaussian noise. The layerwise stability diagnostics and ablation studies also provide some evidence connecting the proposed regularization objective to improved representation stability.

Critical weaknesses
  1. Insufficient clean accuracy and unclear accuracy–robustness trade-off. SSSR consistently achieves lower clean accuracy than the best-performing L1/L2 baseline across all four main CIFAR settings. Moreover, compared with previously reported QCFS results under the same inference settings, the gap remains substantial. The paper does not adequately explain these gaps or clarify whether the robustness gains come at the cost of clean accuracy or arise from differences in training and conversion. This raises concerns about the practical value of SSSR.

  2. Limited robustness evaluation and generalization. The experiments focus almost exclusively on independent Gaussian noise injected after the input IF layer, leaving the method's effectiveness under other perturbation types unclear. Furthermore, on ImageNet, SSSR's top-1 accuracy drops from 55.80% on clean inputs to 13.75% at σ=2\sigma=2 and 1.83% at σ=3\sigma=3. Although it outperforms the evaluated baselines, its absolute performance under stronger perturbations remains limited. These results provide insufficient evidence that the proposed robustness mechanism generalizes beyond the specific perturbation setting studied.

Questions for authors
  1. The paper injects Gaussian noise after the input IF activation, where the per-timestep representations correspond to discrete spike events. However, additive Gaussian noise introduces continuous-valued perturbations that do not directly model errors in spike events. Would it be more realistic to evaluate spike-level errors, such as flipping 0 to 1 (spurious spikes) or 1 to 0 (missed spikes), to better assess the practical robustness of the proposed method?

  2. Could the authors provide controlled comparisons with prior ANN-to-SNN conversion methods in terms of both clean accuracy and robustness under a matched perturbation protocol? Where feasible, comparisons with existing robustness-oriented SNN or conversion methods would also help clarify the contribution of SSSR relative to prior approaches.

9 comments
AuthorsOpenReview ↗

Response to Reviewer j9Wj (part 1 of 4)

We thank the reviewer for the detailed review and for recognizing the conceptual motivation of SSSR and its gains over L1/L2 regularization.

W1. SSSR has lower clean accuracy than the best L1/L2 baseline and a gap to QCFS; do the robustness gains come at the cost of clean accuracy or from differences in training and conversion?

We thank the reviewer for this constructive comment. We have done further investigation and confirmed that the gap in our reported QCFS results to the QCFS paper came from our training hyperparameters. With the QCFS hyperparameters, we have successfully repeated their results, and confirmed that the robustness gain of our proposed SSSR remains unchanged. Moreover, SSSR is only 0.30.3 / 1.01.0 points below L2-wo in clean accuracy. Detailed findings are as below.

(1) Source of the gap to QCFS. The gap came from our choice of training hyperparameters. Our CIFAR VGG-16 setup used Dropout(0.5) after every layer and a batch size of 128128, whereas the official QCFS code uses no Dropout and a batch size of 300300. We re-ran every method with the QCFS hyperparameters and report their performances below.

Table R6. VGG-16 retrained with the hyperparameters of the official QCFS code (mean ± std); the first row is L2-all with our hyperparameters used in the paper at the time of submission.

MethodC10 cleanC10 σ=5\sigma = 5C100 cleanC100 σ=5\sigma = 5
L2-all, hyperparameters at submission90.59±0.1390.59 \pm 0.1310.00±0.0010.00 \pm 0.0062.27±0.3762.27 \pm 0.371.05±0.071.05 \pm 0.07
L2-all95.88±0.1295.88 \pm 0.129.99±0.239.99 \pm 0.2377.40±0.2477.40 \pm 0.241.04±0.131.04 \pm 0.13
L2-wo95.83±0.0795.83 \pm 0.0715.43±4.2315.43 \pm 4.2377.24±0.1477.24 \pm 0.144.76±1.424.76 \pm 1.42
L1-wo95.28±0.1595.28 \pm 0.1518.41±6.7118.41 \pm 6.7176.40±0.2976.40 \pm 0.296.59±2.106.59 \pm 2.10
SSSR94.89±0.1594.89 \pm 0.1593.30±0.1493.30 \pm 0.1474.66±0.1074.66 \pm 0.1072.07±0.3572.07 \pm 0.35
SSSR + weight decay95.49±0.1295.49 \pm 0.1293.85±0.3093.85 \pm 0.3076.20±0.2976.20 \pm 0.2973.71±0.4173.71 \pm 0.41
  • With the same hyperparameters as QCFS, our conclusion still holds true. With exactly the hyperparameters of the original QCFS code, SSSR remains the most robust method by a wide margin: at σ=5\sigma = 5 it keeps 93.993.9% / 73.773.7% with weight decay, while the best baseline reaches 18.418.4% / 6.66.6%. This is the same conclusion as in the submission.
  • The baselines are now in line with QCFS. L2-all reaches 95.8895.88% / 77.4077.40% on CIFAR-10 / CIFAR-100, compared with 90.5990.59% / 62.2762.27% with our submission hyperparameters.

(2) The clean-accuracy cost of SSSR is small. With weight decay of 10−410^{-4} on the weights, we are able to reduce the gap of SSSR and baseline on clean-accuracy, using both VGG-16 and ResNet-18.

  • On VGG-16 (Table R6), SSSR is 0.30.3 and 1.01.0 points below L2-wo on clean CIFAR-10 and CIFAR-100, while raising accuracy at σ=5\sigma = 5 from 15.415.4% / 4.84.8% to 93.993.9% / 73.773.7%. Without weight decay, the clean gap is 0.90.9 / 2.62.6 points with similar robustness.
  • On ResNet-18 (Table R7), SSSR is also the most robust. At σ=5\sigma = 5, SSSR keeps 94.2094.20% / 74.6274.62% and 94.6794.67% / 74.2774.27% with weight decay, while the best baseline reaches 80.1080.10% / 50.6950.69%. Weight decay narrows the clean gap on ResNet-18 as well. It reduces the clean gap of SSSR to L2-wo from 1.01.0 / 3.23.2 to 0.30.3 / 1.91.9 points, without loss of robustness at σ=5\sigma = 5.

Table R7. The same comparison on ResNet-18 (mean ± std).

MethodC10 cleanC10 σ=5\sigma = 5C100 cleanC100 σ=5\sigma = 5
L2-all96.53±0.1696.53 \pm 0.1610.24±0.3810.24 \pm 0.3879.99±0.2279.99 \pm 0.220.98±0.040.98 \pm 0.04
L2-wo96.68±0.0796.68 \pm 0.0753.88±6.8453.88 \pm 6.8479.96±0.3279.96 \pm 0.3231.11±4.2831.11 \pm 4.28
L1-wo95.53±0.0895.53 \pm 0.0880.10±3.2080.10 \pm 3.2078.37±0.2778.37 \pm 0.2750.69±3.1950.69 \pm 3.19
SSSR95.67±0.0695.67 \pm 0.0694.20±0.7394.20 \pm 0.7376.78±0.3276.78 \pm 0.3274.62±0.3974.62 \pm 0.39
SSSR + weight decay96.41±0.1696.41 \pm 0.1694.67±0.2394.67 \pm 0.2378.06±0.1878.06 \pm 0.1874.27±0.7874.27 \pm 0.78

We have stated this trade-off explicitly in Section 4 and Appendix C.8 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer j9Wj (part 2 of 4)

W2 (part 1). The experiments focus almost exclusively on Gaussian noise after the input IF layer; is SSSR effective under other perturbation types?

We thank the reviewer for this suggestion. Yes: SSSR leads at the largest strength of every additional perturbation family we tested. We extended the evaluation to other noise distributions and to noise injected at other locations in the network (Table R8).

Table R8. Performances on various types of perturbations: Laplace and uniform noise of the same variance, Gaussian noise after every IF layer, and sparse impulse noise at the input IF layer or at every IF layer.

DatasetMethodGaussian σ=5\sigma = 5Laplace σ=5\sigma = 5Uniform σ=5\sigma = 5All layers σ=1\sigma = 1All layers σ=2\sigma = 2Impulse, input layerImpulse, all layers
C10L2-all10.010.010.010.010.010.09.89.810.310.310.110.19.99.9
C10L2-wo13.213.213.213.213.213.291.991.951.451.487.887.838.238.2
C10L1-wo27.927.928.028.028.128.191.891.858.658.689.489.445.445.4
C10SSSR + weight decay93.993.993.893.893.893.893.493.482.082.095.295.275.475.4
C100L2-all1.01.01.01.01.01.01.01.00.90.91.01.01.11.1
C100L2-wo4.64.64.84.84.94.964.764.717.117.166.466.410.210.2
C100L1-wo8.68.68.78.79.09.068.068.025.025.066.866.815.615.6
C100SSSR + weight decay74.274.274.574.574.374.370.470.441.741.776.076.032.832.8
  • SSSR leads under every noise distribution. With Laplace or uniform noise of the same variance at σ=5\sigma = 5, SSSR with weight decay keeps 93.893.8% / 93.893.8% on CIFAR-10 and 74.574.5% / 74.374.3% on CIFAR-100, while every baseline stays below 2929% and 1010%.
  • SSSR leads when noise enters every layer. With Gaussian noise after every IF layer at σ=2\sigma = 2, it keeps 82.082.0% / 41.741.7%, compared with at most 58.658.6% / 25.025.0%.
  • SSSR leads under sparse impulse noise. With impulses of ±10\pm 10 with probability 55% per neuron and timestep, it keeps 95.295.2% / 76.076.0% when only the input IF layer is hit and 75.475.4% / 32.832.8% when every IF layer is hit, compared with at most 89.489.4% / 66.866.8% and 45.445.4% / 15.615.6%.
AuthorsOpenReview ↗

Response to Reviewer j9Wj (part 3 of 4)

W2 (part 2). On ImageNet, SSSR drops from 55.8055.80% clean to 13.7513.75% at σ=2\sigma = 2 and 1.831.83% at σ=3\sigma = 3; its absolute performance under strong perturbations remains limited.

We thank the reviewer for this comment. To better characterize this, we report ImageNet results across moderate and severe noise levels and additionally evaluate a deeper ResNet-34 (Table R9, Appendix C.10).

Table R9. ImageNet top-1 accuracy under moderate (σ≤1\sigma \leq 1) and severe (σ≥2\sigma \geq 2) noise in Appendix C.10.

ArchitectureMethodσ=0.5\sigma = 0.5σ=1\sigma = 1σ=2\sigma = 2σ=3\sigma = 3
ResNet-18L2-all0.130.130.100.100.100.100.100.10
ResNet-18L2-wo36.5736.5710.7210.720.310.310.160.16
ResNet-18L1-wo46.2146.2121.1121.110.570.570.140.14
ResNet-18SSSR50.2750.2740.5640.5613.7513.751.831.83
ResNet-34L2-all0.140.140.100.100.100.100.100.10
ResNet-34L2-wo31.6931.6911.0811.080.330.330.110.11
ResNet-34L1-wo46.9346.9331.0331.031.231.230.100.10
ResNet-34SSSR53.2053.2044.8444.8423.9923.997.507.50
  • Substantial improvements under moderate noise.** With ResNet-18, SSSR achieves 50.2750.27% and 40.5640.56% at σ=0.5\sigma = 0.5 and 11, compared with 46.2146.21% and 21.1121.11% for the strongest conventional baseline; with ResNet-34, it achieves 53.2053.20% and 44.8444.84%, compared with 46.9346.93% and 31.0331.03%.
  • Under severe noise, the baselines collapse while SSSR retains part of its accuracy. We agree that σ≥2\sigma \geq 2 is a severe regime for 1000-class classification. Here the baselines are near chance (at most 1.231.23% at σ=2\sigma = 2), whereas SSSR keeps 13.7513.75% (ResNet-18) and 23.9923.99% (ResNet-34); its margin over the best baseline grows with depth from 13.213.2 to 22.822.8 points.
  • Scientific contribution and future directions. Our primary contribution is not to fully resolve robustness under extreme noise, but to highlight an underexplored vulnerability in ANN-to-SNN conversion, identify the source-model properties governing representation stability, and introduce a principled regularizer that improves robustness without noise injection during training. On ImageNet, converted SNNs trained with L1/L2 regularization lose almost all accuracy once the internal noise reaches σ≥2\sigma \geq 2. This matters in practice, because converted SNNs are intended for mixed-signal or analog in-memory hardware and for split inference, where the activations after the input layer can be corrupted by readout noise or by an impaired channel. The ImageNet results demonstrate both the promise of this approach and the substantial room for further improvement. We view robustness under severe perturbations, particularly for large-scale deployment, as an important open challenge that warrants broader research efforts. The extra results on ResNet-34 have been included in Appendix C.10. Conclusion of the revised manuscript lists broader perturbation models and robustness in large-scale deployment settings as future work.
Reviewer_Qb3EOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the robustness of Spiking Neural Networks (SNNs) converted from Artificial Neural Networks (ANNs) that were trained solely on clean data—a relatively underexplored topic. The authors propose a regularization method, SSSR, which improves performance under noisy conditions without requiring the explicit addition of noise during training.

Critical strengths

Strengths: The topic—robustness in SNNs—is an under-researched area. However, this raises further questions: Why is this topic significant? Where does its importance stem from—is it a practical, real-world concern? Furthermore, does the improvement in robustness manifest similarly in both ANNs and SNNs? These questions seem even more critical than the points illustrated in Fig. 1b.

Critical weaknesses

Weaknesses: The discussion on existing work regarding robustness is too sparse. This might mislead readers into believing that SNN robustness has rarely been studied, potentially causing them to overlook the authors' core insight: "A more fundamental question, however, remains unresolved: when the source ANN is trained only on clean data, what properties of the source model determine the stability of the converted SNN under internal perturbations, as illustrated in Fig. 1?"

Additionally, it appears the authors do not formally define the noise added to the system within the main text.

Regarding Gaussian noise and SNN robustness, I recently encountered a paper titled "Enhancing adversarial robustness in SNNs with sparse gradients," which notes that SNNs generally exhibit greater robustness against random perturbations than against adversarial ones. Given that the authors' primary argument centers on Gaussian noise, one might question whether this focus is sufficiently robust or compelling.

Furthermore, the paper lacks comparisons with existing methods, making it difficult to gauge the effectiveness of the proposed approach. I recognize that most existing methods focus on adversarial robustness; therefore, the authors should clearly articulate why robustness against Gaussian noise is a necessary and meaningful objective.

Questions for authors

Questions: I have integrated my primary concerns into the "Weaknesses" section above to help the authors better understand my perspective.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Qb3E (part 1 of 4)

We thank the reviewer for the thoughtful review and for noting that the robustness of converted SNNs is under-researched. In response, we expanded the related work, added a formal definition of the noise, compared SSSR with conversion-error correction (COS: Hao et al., ICLR 2023; SRP: Hao et al., AAAI 2023; TPP: Bojković et al., ICML 2025) and with SNNs trained directly with SR (Liu et al., ICML 2024), and evaluated the same source ANNs as ANNs and as SNNs. Under the same internal noise, these prior methods fall to near chance at σ=5\sigma = 5, whereas SSSR with weight decay keeps 93.993.9% / 74.274.2%.

W1. The discussion of existing work on robustness is too sparse and may suggest that SNN robustness has rarely been studied.

We thank the reviewer for this suggestion and agree that the discussion should be expanded. We have expanded the Related Work with a paragraph on SNN robustness and now state how our question differs.

Revised related work. We have revised the Related Work to cover adversarial robustness of SNNs through input encoding (Sharmin et al., ECCV 2020), regularized training (Ding et al., NeurIPS 2022; Liu et al., ICML 2024), certification of rate encoding (Mukhoty et al., ICLR 2024), and robust ANN-to-SNN conversion (Özdenizci and Legenstein, TMLR 2024).

Positioning. These methods build robustness into SNN training or conversion. Our setting keeps the source ANN trained only on clean data and asks which properties of that model determine the stability of the converted SNN. We would also welcome any suggestions from the reviewer regarding additional relevant studies we may have overlooked.

W2. The noise added to the system is not formally defined in the main text.

We thank the reviewer for pointing this out. We have added a formal definition to Section 4 (Setup and Perturbation Protocol) of the revised manuscript.

Definition. At every timestep, zero-mean Gaussian noise with standard deviation σ\sigma is added to the output of the input IF layer, drawn independently across timesteps, neurons and samples. No noise is used during training.

AuthorsOpenReview ↗

Response to Reviewer Qb3E (part 2 of 4)

W3. SNNs are generally more robust to random than to adversarial perturbations (Liu et al., ICML 2024); is a focus on Gaussian noise sufficiently compelling?

We thank the reviewer for pointing us to Liu et al. (ICML 2024). We believe the focus is compelling, because our noise is internal rather than at the input, and under it SNNs trained for adversarial robustness are not robust, whereas SSSR is. We agree that SNNs are generally more resilient to random perturbations of the input than to adversarial ones. Our setting differs in where the perturbation enters.

Internal rather than input noise. The noise in our study enters after the input encoding and propagates through the network. We trained SNNs directly with the SR regularizer of Liu et al., using the official code, and evaluated them under our protocol (Table R3).

Table R3. Comparison with adversarially robust SNN training (SR) under the same internal Gaussian noise.

ArchDatasetMethodCleanσ=1\sigma = 1σ=2\sigma = 2σ=3\sigma = 3σ=5\sigma = 5
VGG-16C10SR (Liu et al., ICML 2024)91.691.675.575.519.819.811.011.010.610.6
VGG-16C10SSSR + weight decay95.595.595.395.395.295.295.095.093.993.9
VGG-16C100SR (Liu et al., ICML 2024)67.967.948.748.76.56.52.52.51.71.7
VGG-16C100SSSR + weight decay76.576.576.676.676.376.375.875.874.274.2
WRN-16-4C10SR (Liu et al., ICML 2024)87.587.574.774.727.527.512.712.710.010.0
WRN-16-4C10SSSR95.095.094.694.694.494.493.493.488.988.9
WRN-16-4C100SR (Liu et al., ICML 2024)68.668.624.724.72.92.91.51.51.31.3
WRN-16-4C100SSSR76.476.476.376.375.475.474.174.168.568.5
  • SR does not provide robustness to internal noise. At σ=1\sigma = 1, SR keeps 75.575.5% / 48.748.7% on VGG-16; at σ=2\sigma = 2 it falls to at most 27.527.5% on CIFAR-10 and 6.56.5% on CIFAR-100, on either architecture.
  • SSSR keeps its accuracy. At σ=5\sigma = 5, SSSR keeps 93.993.9% / 74.274.2% on VGG-16 and 88.988.9% / 68.568.5% on WRN-16-4, where SR is at most 10.610.6% / 1.71.7%.
  • Setting. SR uses the official code and training setting (T=8T = 8) and is evaluated at T=16T = 16; on WRN-16-4, SSSR is used without weight decay at L=T=8L = T = 8.

Complementary objectives. SR targets adversarial attacks such as PGD, whereas SSSR targets random internal noise. We therefore view the two objectives as complementary. Robustness to random internal noise is meaningful on its own because such noise arises from deployment conditions, such as noisy analog readout or transmission, without an adversary.

AuthorsOpenReview ↗

Response to Reviewer Qb3E (part 3 of 4)

W4. The paper lacks comparisons with existing methods and should articulate why robustness against Gaussian noise is a necessary and meaningful objective.

We thank the reviewer for this suggestion. We added two groups of comparisons, and neither provides robustness to internal noise, whereas SSSR does. This robustness is a meaningful objective because converted SNNs are deployed on hardware and channels that perturb their internal activations without any adversary (W3 and Strengths Q1). Both groups use the same perturbation protocol: conversion-error correction (Table R4) and adversarially robust SNN training (Table R3).

Table R4. Varying the ANN-SNN conversion methods.

Conversion MethodSource ANN RegularizationC10 cleanC10 σ=5\sigma = 5C100 cleanC100 σ=5\sigma = 5
QCFSL2-all95.6495.6410.010.077.0477.041.01.0
COS (Hao et al., ICLR 2023)L2-all95.6995.6910.010.077.5377.531.01.0
SRP (Hao et al., AAAI 2023)L2-all95.6995.6910.010.077.3677.361.01.0
TPP (Bojković et al., ICML 2025)L2-all95.6995.6910.010.077.1177.111.01.0
QCFSSSSR + weight decay95.4695.4693.993.976.5276.5274.274.2
  • Conversion-error correction does not provide robustness to internal noise. COS, SRP and TPP, applied to the same ANN trained with QCFS (Bu et al., ICLR 2022) at a matched total number of timesteps, reach high clean accuracy (95.795.7% / 77.577.5%, 95.795.7% / 77.477.4% and 95.795.7% / 77.177.1%) but drop to 10.010.0% / 1.01.0% at σ=5\sigma = 5, where SSSR with weight decay keeps 93.993.9% / 74.274.2%. These methods reduce conversion error on clean inputs, not the effect of internal noise.
  • Adversarially robust SNN training does not either. SNNs trained directly with SR, using the official code, reach 10.610.6% / 1.71.7% at σ=5\sigma = 5 on VGG-16 and 10.010.0% / 1.31.3% on WRN-16-4, whereas SSSR keeps 93.993.9% / 74.274.2% on VGG-16 (with weight decay) and 88.988.9% / 68.568.5% on WRN-16-4 (without weight decay) (Table R3).

We have added these comparisons to Appendix C.11 of the revised manuscript.

Reviewer_pYJHOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a Spike-State Stability Regularizer (SSSR) for ANN-to-SNN conversion to improve the robustness of the resulting SNN model against internal perturbations. The authors argue that injected noise can alter spike counts by shifting preactivations across discrete activation boundaries, thereby affecting the model's performance. Based on their analysis, they introduce a regularization term that controls the effective weight magnitudes and learns activation thresholds during source-ANN training, without injecting noise during training. During evaluation, Gaussian noise is injected after the input IF layer at each timestep. Experiments on VGG and ResNet architectures using CIFAR-10, CIFAR-100, and ImageNet demonstrate improved classification accuracy under internal noise compared with standard L1/L2 regularization, while largely preserving clean accuracy.

Critical strengths

This work provides an intuitive explanation and empirical evidence of how injected noise affects the behavior of converted SNN models. It also proposes a theoretically motivated regularizer to improve robustness. The method is simple to incorporate into ANN training and does not require noise injection during training. The experimental results demonstrate significant improvements in robustness across different architectures and datasets.

Critical weaknesses
  1. Lack of practical motivation: The paper lacks a clear motivation regarding the types of internal noise that could arise in SNNs after deployment. In practice, what kind of noise can be introduced at the activation level, and how large can its magnitude be? The authors use relatively high noise levels to demonstrate the effectiveness of their method. Are such noise levels realistic in practical applications? It would be helpful if the authors could provide some justification for the noise levels considered in their experiments.

  2. Noise injection into spike activations: Since Gaussian noise is added directly to the spike outputs, wouldn't this result in continuous-valued activations, including negative values? How are such values handled during inference, and how does this noise model relate to practical SNN implementations?

  3. Misleading illustration in Figure 1: From Figure 1, it appears that noise is added at the input image level. However, in the experiments, noise is injected at the activation level. I suggest that the authors modify the figure to clearly indicate where the noise is introduced, as the current illustration could be misleading.

Questions for authors

Check the weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer pYJH

We thank the reviewer for the careful review and for recognizing that SSSR is simple to add to ANN training and improves robustness across architectures and datasets. In response, we explain where activation-level noise arises in practice and how large it can be, clarify how continuous and negative values are handled, and redraw Figure 1 to mark the injection point. SSSR also improves accuracy at moderate noise, not only at severe noise.

W1. What kind of noise can be introduced at the activation level after deployment, how large can it be, and are the noise levels used realistic?

We thank the reviewer for highlighting the need to better motivate the practical significance of the proposed method. Activation-level noise arises on mixed-signal and analog in-memory hardware and when intermediate spikes are transmitted in split inference; its magnitude depends on the device and the channel, so we evaluate noise from moderate to severe levels, and SSSR helps across this range.

Practical origins and deployment scenarios. Activation-level perturbations can arise in two relevant settings. First, mixed-signal neuromorphic and analog in-memory hardware can introduce computation noise through analog readout, circuit fluctuations, and imperfect signal conversion (Rasch et al., Nat. Commun. 2023). Second, in distributed or split SNN inference, intermediate spike representations are transmitted between devices and may be corrupted by channel noise, fading, or interference (Chen et al., IEEE TCCN 2023; Wu et al., IEEE TMLCN 2025).

The second setting is particularly relevant: when an SNN is partitioned immediately after its encoding layer, the transmitted representation corresponds to the post-input-IF activations where we inject noise. We model additive errors at this interface using a controlled Gaussian perturbation model. This abstraction allows us to study the SNN robustness.

Noise magnitude and practical relevance. We agree that the magnitude of activation-level noise depends on the underlying device, activation scale, and operating conditions. Severe operating conditions, such as interference-limited wireless links or highly attenuated communication channels, can lead to substantial corruption of transmitted intermediate representations.

We also want to highlight that our evaluation therefore covers a range of perturbation magnitudes, including moderate conditions and severe degradation regimes. The benefits of SSSR are not limited to the severe degradation regimes. For example, on ImageNet with ResNet-18, SSSR achieves 50.2750.27% and 40.5640.56% accuracy at σ=0.5\sigma = 0.5 and 11, compared with 46.2146.21% and 21.1121.11% for the strongest conventional baseline, respectively. The larger perturbations allow us to further characterize robustness degradation and identify differences between methods under severe impairment.

W2. Gaussian noise added to the spike outputs gives continuous and negative values; how are they handled, and how does this noise model relate to practical SNN implementations?

We thank the reviewer for this observation. Such values are treated as a perturbation of the next layer's synaptic current, and the following IF layer still emits valid, nonnegative spike counts. We clarify that Gaussian noise is added after the input IF layer to model perturbations in the effective synaptic input (which can go negative even without noise), rather than changes to the binary spike-generation mechanism.

Handling of continuous and negative values. Although Gaussian perturbations can produce continuous or negative values, these are passed through the subsequent weighted transformation and interpreted as perturbations to the synaptic current. The following IF layer still generates valid, nonnegative spike counts through the clipping operation described in Appendix B.1 in our noise model.

Practical relevance. This model approximates analog disturbances arising from noisy synaptic readout, mixed-signal computation, or intermediate feature transmission in distributed inference. It does not imply that physical spikes become continuous or negative.

W3. Figure 1 suggests that noise is added to the input image; please indicate where the noise is introduced.

We thank the reviewer for pointing out this ambiguity. We have revised Figure 1 to explicitly mark the noise injection point between the input IF layer and the subsequent network layers, keeping the input image unchanged. This ensures that the illustration accurately reflects our experimental protocol.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer pYJH,

As the discussion period draws to a close, we hope that our response has addressed your concerns. Following your suggestions, we provided:

  1. Practical motivation and noise severity. We discussed how analog readout and transmission in split inference can perturb internal signals. Because the noise magnitude depends on the device and the channel, we reported results from moderate to severe noise levels. SSSR also improves accuracy at moderate noise: on ImageNet with ResNet-18, it reaches 50.2750.27% / 40.5640.56% at σ=0.5\sigma = 0.5 / 11, compared with 46.2146.21% / 21.1121.11% for the strongest baseline. The larger values serve as stress tests and are not calibrated to a specific device or channel.
  2. Handling of continuous and negative values. We clarified that these values are treated as perturbations to the next layer's effective synaptic input, not as physical spikes. The next IF layer still emits valid, nonnegative spike counts.
  3. A revised Figure 1. The figure now shows that the input image is unchanged and that the perturbation is introduced after the input IF layer.

We have incorporated these clarifications into the revised manuscript. Could you kindly let us know whether they address your concerns? We would be happy to clarify any remaining points during the discussion.

Thank you again for your detailed and constructive review.

Thanks for your detailed response, and I raised my score accordingly.

AuthorsOpenReview ↗

Response to Reviewer pYJH

Thank you wholeheartedly for your thorough reassessment and for raising the score. We are delighted that the additional analyses addressed your main concerns. We will incorporate the new results along with new discussions presented in all rebuttals into the revised version.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to all reviewers (part 3 of 3)

Revisions in the PDF

  • The VGG-16 CIFAR results (main table, ANN–SNN comparison, layerwise diagnostics, energy) now use the hyperparameters of the official QCFS code, and the CIFAR results of SSSR use weight decay of 10−410^{-4} on the weights. ResNet models have no Dropout, so their baseline results are unchanged.
  • Section 4 (Setup and Perturbation Protocol) formally defines the perturbation.
  • Fig. 1 is redrawn with the noise added to the feature map after the input IF layer and before the second IF layer.
  • Related Work adds a paragraph on SNN robustness: adversarial robustness of directly trained and converted SNNs (including SR), and how our clean-trained setting differs.
  • The Conclusion now lists broader perturbation models and large-scale deployment as future work, and the clean-accuracy cost of SSSR is stated in the main text.
  • The appendix adds the further perturbations, ImageNet ResNet-34, the comparisons with COS, SRP, TPP and SR, the ANN–SNN comparison and SSSR with weight decay.
AuthorsOpenReview ↗

General response to all reviewers (part 2 of 3)

Main key results

Table R1. Comparison with more baseline by varying the ANN-SNN conversion methods, as well as an SNN trained directly with the adversarial-robustness regularizer SR (VGG-16).

Conversion MethodRegularizationC10 cleanC10 σ=5\sigma = 5C100 cleanC100 σ=5\sigma = 5
QCFSL2-all95.6495.6410.010.077.0477.041.01.0
COS (Hao et al., ICLR 2023)L2-all95.6995.6910.010.077.5377.531.01.0
SRP (Hao et al., AAAI 2023)L2-all95.6995.6910.010.077.3677.361.01.0
TPP (Bojković et al., ICML 2025)L2-all95.6995.6910.010.077.1177.111.01.0
SR, direct training (Liu et al., ICML 2024)SR91.6291.6210.610.667.8667.861.71.7
QCFSSSSR + weight decay95.4695.4693.993.976.5276.5274.274.2
  • Prior methods do not provide robustness to internal noise. COS, SRP and TPP reach clean accuracy on par with or above QCFS but fall to 10.010.0% / 1.01.0% at σ=5\sigma = 5, where SSSR with weight decay keeps 93.993.9% / 74.274.2%.
  • Adversarially robust SNN training does not either. SNNs trained directly with SR fall to 19.819.8% / 6.56.5% at σ=2\sigma = 2 and 10.610.6% / 1.71.7% at σ=5\sigma = 5 on VGG-16, and to 10.010.0% / 1.31.3% at σ=5\sigma = 5 on WRN-16-4, where SSSR keeps 88.988.9% / 68.568.5%.

Table R2. Performances on various types of perturbations: Laplace and uniform noise of the same variance, Gaussian noise after every IF layer, and sparse impulse noise at the input IF layer or at every IF layer.

DatasetMethodGaussian σ=5\sigma = 5Laplace σ=5\sigma = 5Uniform σ=5\sigma = 5All layers σ=1\sigma = 1All layers σ=2\sigma = 2Impulse, input layerImpulse, all layers
C10L2-all10.010.010.010.010.010.09.89.810.310.310.110.19.99.9
C10L2-wo13.213.213.213.213.213.291.991.951.451.487.887.838.238.2
C10L1-wo27.927.928.028.028.128.191.891.858.658.689.489.445.445.4
C10SSSR + weight decay93.993.993.893.893.893.893.493.482.082.095.295.275.475.4
C100L2-all1.01.01.01.01.01.01.01.00.90.91.01.01.11.1
C100L2-wo4.64.64.84.84.94.964.764.717.117.166.466.410.210.2
C100L1-wo8.68.68.78.79.09.068.068.025.025.066.866.815.615.6
C100SSSR + weight decay74.274.274.574.574.374.370.470.441.741.776.076.032.832.8
  • SSSR leads under every noise distribution. With Laplace or uniform noise of the same variance at σ=5\sigma = 5, SSSR with weight decay keeps 93.893.8% / 74.574.5%, while every baseline stays at or below 28.128.1% / 9.09.0%.
  • SSSR leads when noise enters every layer. With Gaussian noise after every IF layer at σ=2\sigma = 2, it keeps 82.082.0% / 41.741.7%, compared with at most 58.658.6% / 25.025.0%.
  • SSSR leads under sparse impulse noise. With impulses at every IF layer, it keeps 75.475.4% / 32.832.8%, compared with at most 45.445.4% / 15.615.6%.
AuthorsOpenReview ↗

General response to all reviewers (part 1 of 3)

We thank all reviewers for their careful and constructive reviews. Below we list the new experiments, the main results and the changes in the revised PDF. Each reviewer's questions are answered in full in the individual replies.

New experiments

  1. Expand robustness experiments using more types of noises: Laplace and uniform noise, Gaussian noise after every IF layer, and sparse impulse noise after the first or after every IF layer.
  2. Expand comparison to include more prior methods: including COS (Hao et al., ICLR 2023), SRP (Hao et al., AAAI 2023) and TPP (Bojković et al., ICML 2025), and SNNs trained directly with the robustness regulariser SR (Liu et al., ICML 2024) using its official code, on VGG-16 and WRN-16-4.
  3. Add results on ImageNet with ResNet-34 (in addition to ResNet-18).
  4. New results using QCFS hyperparameters. For CIFAR VGG-16 we adopted the hyperparameters of the official code of QCFS (Bu et al., ICLR 2022): no Dropout and batch size 300300; we retrained every method, baselines included.
  5. SSSR with weight decay (10−410^{-4} on the weights): VGG-16 and ResNet-18 on CIFAR-10/100; ResNet-18 on ImageNet is in progress.
  6. ANN versus SNN: the same source ANNs evaluated as ANNs and as SNNs.
  7. Layerwise diagnostics and energy re-measured with the QCFS hyperparameters.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.