acceptodds
Under review as a conference paper at ICLR 2027

Towards Stable and Scalable Event-Driven Training of Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) offer a promising route towards energy-efficient deep learning on account of their sparse, event-driven computation. Activation-based training achieves competitive performance but does not exploit the event-driven computation during training, leading to memory inefficiency. In contrast, event-driven (time-based) training performs credit assignment with more efficient memory utilization but remains underexplored and has not been shown to scale favorably across architectures, datasets, and inference timesteps. In this work, we develop an event-driven training approach by identifying and addressing sources of instability in existing formulations, thereby enabling stable and scalable training across architectures and inference timesteps. Our approach retains the memory-efficient formulation of event-driven learning while enabling the effective use of batch normalization, which we show to be important for SNN training stability, without relying on ad-hoc methods such as supervisory spike injection used in previous approaches. We obtain state-of-the-art results among event-driven methods on both static and neuromorphic benchmarks, with a 5% improvement in accuracy at less than half the inference timesteps on CIFAR-100, and even extend event-driven training to Tiny-ImageNet. Overall, our results establish event-driven learning as a scalable and efficient alternative to activation-based training for deep SNNs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spiking Neural Networks, Event based backpropogation, Neuromorphic Computing

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 3 reviews, above 27% of the others.
Reviewer_FzrbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper attempts to address the issues of gradient instability, reliance on supervisory spike signals, and poor scalability across architectures in the event-driven (time-based) training of spiking neural networks (SNNs). To mitigate gradient vanishing caused by silent neurons, the authors replace the final spiking layer with a non-spiking, continuous accumulation layer:si(l)(t)=∑jwij(sj(l−1)(t)+bi)s_i^{(l)}(t) = \sum_j w_{ij} \left(s_j^{(l-1)}(t) + b_i\right)while retaining spiking dynamics in the penultimate layers. In addition, they simplify the post-synaptic potential (PSP) kernel of the Leaky Integrate-and-Fire (LIF) neuron from a dual-exponential form:ϵ(t)=τmτm−τs(e−t/τm−e−t/τs)\epsilon(t) = \frac{\tau_m}{\tau_m - \tau_s} \left(e^{-t/\tau_m} - e^{-t/\tau_s}\right)to a single-exponential decay:ϵ(t)=e−t/τm\epsilon(t) = e^{-t/\tau_m}thereby eliminating the gradient reversal inherent in the rising phase of the dual-exponential kernel. Furthermore, a Batch Normalization (BN) layer is incorporated prior to spike generation to enhance training stability. Experiments across several benchmark datasets show notable accuracy improvements over prior event-driven baselines and competitive performance with lower peak GPU memory consumption compared to standard activation-based (BPTT) training.

Critical strengths

1.Gradient instability and silent neurons represent critical bottlenecks in time-based/event-driven SNN training. Investigating principled remedies for these issues is timely and holds substantial potential for neuromorphic learning.

2.Event-driven training aligns much more closely with the asynchronous, event-based nature of neuromorphic hardware than standard frame-based BPTT, making scalable exploration in this domain valuable.

3.The paper is generally well-structured and easy to follow.

Critical weaknesses

1.As revealed by the ablation study in Table 2, the primary resolution of training instability and the bulk of the accuracy gain stem almost entirely from replacing the spiking output layer with a non-spiking linear layer. However, using a non-spiking readout layer is already a standard and well-established technique in the broader SNN literature (as acknowledged by the authors). Meanwhile, the remaining modifications—simplifying the LIF kernel to a single exponential and introducing BN—yield only marginal accuracy improvements. Overall, the method appears to be a heuristic combination of existing engineering tricks rather than a fundamentally novel theoretical framework. Furthermore, turning the final layer into continuous-valued accumulation deviates from the discrete, event-driven nature of SNNs.

2.The manuscript contains concerning citation issues, including future-dated and seemingly hallucinated references (e.g., referencing Ramesh & Srinivasan (2025) and Béna et al. (2025)).

3.While event-driven training reduces peak memory consumption by constraining gradient updates and avoiding dense unrolling across time steps, this temporal simplification often incurs an information loss. In practice, achieving competitive accuracy under such approximations frequently requires more training epochs. The authors solely demonstrate peak memory savings during training, which does not provide a complete picture of resource efficiency. A thorough comparison of total training compute (e.g., wall-clock training time per epoch, throughput in samples/sec, and overall FLOPs/SOPs) against mature activation-based methods is noticeably missing.

4.On neuromorphic (DVS) datasets, inputs are continuous, asynchronous event streams. If these streams are simply binned into discrete time windows, the gradient calculation proposed in this paper may struggle to capture inter-window temporal correlations. The empirical evaluation on neuromorphic data is limited to DVS-CIFAR10, which is essentially recorded by displaying repeated saccadic movements of static images on an LCD screen and lacks rich, complex temporal dynamics (many models achieve acceptable accuracy on DVS-CIFAR10 by collapsing events into a single static frame). Furthermore, genuine event-based streams contain event volumes far exceeding those of static benchmarks. The paper should evaluate efficiency and scalability on more complex neuromorphic benchmarks (e.g., DVS-Gesture, NCALTECH101, or SHD) to substantiate whether the memory savings and accuracy gains genuinely persist under heavy event traffic.

Questions for authors

1.In Table 2, replacing the spiking output layer with a non-spiking linear layer delivers the decisive performance improvement. How does this continuous-valued readout affect deployment on fully asynchronous, event-driven neuromorphic processors (e.g., SpiNNaker or Loihi)? Does it introduce non-negligible MAC operations and synchronization overhead, potentially undermining the low-power premise of SNNs?

2.How does the proposed event-driven backward pass propagate credit assignment across distinct time windows when processing dynamic event streams? Please clarify how long-term dependencies are preserved if gradients are not propagated densely across temporal boundaries.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Fzrb

We thank the reviewer for their thoughtful assessment and constructive feedback. We have addressed the concerns raised in the following responses.

Deployment on neuromorphic hardware

The readout receives only binary spikes from the penultimate layer (Eq. 11). Consequently, its computation consists of accumulate operations rather than general multiply-accumulate (MAC) operations: each incoming spike triggers an addition of the corresponding synaptic weight to the output neuron's potential. The readout is an integrate-only neuron without a firing threshold. The accumulated potentials are read once per sample, at the same point where a spiking output layer would otherwise provide its spike counts. Thus, the readout does not require additional synchronization between the hidden layers or introduce a new temporal synchronization mechanism. Importantly, all hidden layers and their credit assignment remain event-driven. Regarding deployment, integrate-only neurons are simple computational primitives that can be implemented on programmable neuromorphic platforms. Non-spiking linear readouts are also commonly used in SNN training frameworks, including TET, and do not fundamentally change the low-power premise of SNNs.

Credit assignment across temporal windows

Credit crosses bin boundaries. An input spike at time (t_m) affects the membrane potential through ϵ(t−tm)\epsilon(t-t_m) until the neuron next fires (Eq. 1). Thus, the backward pass assigns credit to every input spike since the neuron's previous output spike, whichever bin it falls in, and passes it between layers through spike times (Eqs. 3–4). The formulation does not propagate dependence through the reset, i.e., on inputs before the previous output spike, similar to surrogate-gradient training with a detached reset. Both our method and TET use the same feedforward architecture.

Weaknesses

Novelty and contribution of the individual components

We would like to point out that the kernel and BN changes are not marginal. In Table 2, the original configuration (A) fails to train at (T=6), so A→B separates "trains" from "does not train" rather than measuring each component's contribution. With the linear readout alone (B), every architecture stays below the best prior event-driven CIFAR-10 result (88.60–92.30% at (T=6) vs. 93.54% for Zhu et al. (2023) at (T=12)). Adding the monotonic kernel and BN (B→E) contributes +2.3 to +3.5 points on CIFAR-10 and +4.0 to +5.5 points on CIFAR-100 for VGG16, ResNet-18, and ResNet-19. These two changes bring performance above the prior result at significantly fewer timesteps. Our key contribution is identifying why event-driven training fails at low timesteps and in deeper networks, with evidence for each mechanism.

Citation issues

One of the bibliography entries was incomplete, which may have caused confusion; we have corrected it in the revised manuscript. We would also like to clarify that both references are existing works and are not future-dated citations: Béna et al. (2025): Available on arXiv: https://arxiv.org/abs/2412.15021 Ramesh & Srinivasan: Published in TMLR: https://openreview.net/forum?id=kIdB7Xp1Iv

Computational efficiency

The computational advantage of event-driven learning over rate-based training, theoretically established by Zhu et al. through a complexity analysis, is now included in Appendix G, highlighting the reduction in computational operations when spike activity is sparse. Moreover, our experiments demonstrate a substantial reduction in GPU memory allocated during training, as reflected by the peak memory allocation.

Regarding convergence efficiency, our experiments typically require 200–400 epochs to converge, which is comparable to the 300-epoch training schedule used by TET. Thus, we do not observe a substantial additional epoch requirement in our experiments that would negate the per-epoch computational advantages of event-driven training.

We additionally measured wall-clock training time per epoch at (T=6). Our method takes 71 s/epoch compared with 172 s/epoch for TET on VGG16, and 52 s/epoch compared with 85 s/epoch for TET on ResNet-18.

Evaluation on more complex neuromorphic datasets

As requested, we conducted additional experiments on DVS-Gesture. Using VGG11 with (T=20), our method achieves 97.45±0.40%97.45 \pm 0.40\% accuracy over three random seeds. We also compared peak training memory against TET under identical settings (dataset, batch size, model, and timesteps) on the same RTX 4090 GPU. Our method requires 4.7 GB, compared with 6.7 GB for TET, corresponding to approximately 30% lower peak memory consumption. These results demonstrate that the memory benefits persist beyond DVS-CIFAR10 on a more temporally demanding event-based benchmark. We have included these results in the revised manuscript.

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

This paper studies event-driven, spike-timing-based training of deep spiking neural networks and focuses on several practical instability issues observed in prior formulations. The proposed framework replaces the spike-based output layer and supervisory spike injection with a continuous linear readout, uses a monotonic single-exponential neuronal kernel to avoid gradient reversal, and introduces batch normalization before spike generation. The authors evaluate the resulting training recipe across several architectures, datasets, and timestep budgets, reporting improved performance over previous event-driven methods and reduced training memory relative to activation-based BPTT-style SNN training.

Critical strengths
  1. The paper identifies practically important failure modes in existing event-driven SNN training. The discussion of dead output neurons, forward/backward inconsistency caused by supervisory spike injection, and gradient reversal provides a useful diagnosis of why previous time-based training formulations can become unstable, especially at low timestep budgets.

  2. The empirical results demonstrate substantially improved scalability of event-driven training. The proposed recipe works across VGG and ResNet architectures, reaches competitive accuracy at substantially reduced timesteps, and provides concrete model-scale evidence of lower training memory consumption than an activation-based baseline.

Critical weaknesses
  1. The algorithmic novelty appears limited relative to the strength of the claims. The proposed framework mainly combines a non-spiking linear readout, a monotonic single-exponential kernel, batch normalization, and conventional optimization/augmentation choices. The paper itself notes that non-spiking readouts have been used in prior SNN training, while batch normalization is a standard technique. Although the resulting recipe is empirically useful, I am not yet convinced that the work establishes a sufficiently new event-driven learning principle or algorithmic contribution for ICLR, rather than a careful stabilization and engineering of an existing framework.

  2. The current experimental design does not cleanly attribute the reported gains, and the supplied reproducibility package appears incomplete. The largest improvement in Table 2 occurs between the original implementation and the authors’ “Baseline,” but this baseline simultaneously replaces the output layer and changes the optimizer to Adam, so the contribution of the proposed output-layer fix cannot be isolated. Some individual BN ablations also reduce accuracy, which makes the general claim about its stabilizing contribution less clear. In addition, the paper states that the supplementary material contains the complete code and configurations for the main results, whereas the supplied package appears to contain only CIFAR-10/100 loaders/configurations and does not provide the DVS-CIFAR10 or Tiny-ImageNet setups or configurations corresponding to all reported ablations.

Questions for authors
  1. In Table 2, the transition from the original implementation (A) to the baseline (B) changes both the output layer and the optimizer. Could the authors provide a factorial ablation separating the effect of replacing the spiking output layer from the effect of switching to Adam? This appears important because A→B accounts for the largest performance recovery.

  2. Could the authors clarify the “5% improvement” claim in the abstract? On CIFAR100, the non-augmented result improves from 70.50% to 71.60% relative to Zhu et al. (2023), while the 76.9% result uses augmentation and a different architecture. What exact comparison is intended by the 5% figure?

  3. How should the role of batch normalization be interpreted given that the BN-only condition decreases accuracy for some architectures, e.g., ResNet-19 on both CIFAR10 and CIFAR100? Is the benefit primarily an interaction between BN and the proposed kernel rather than an independently robust stabilization effect?

  4. The reproducibility statement says that the supplementary material contains the complete source code and configurations for the main results. However, the supplied package appears to contain only CIFAR10/CIFAR100 dataset support and does not include the DVS-CIFAR10 or Tiny-ImageNet configurations, nor clear configurations for all Table 2 ablations. Could the authors provide the missing configurations/code and specify which configuration reproduces each reported result?

1 comment
AuthorsOpenReview ↗

Response to Reviewer XGsf

We thank the reviewer for their thoughtful assessment and constructive feedback. We have addressed the concerns raised in the following responses.

Output-layer and optimizer ablation

To isolate the optimizer effect, we evaluated both configurations with Adam and AdamW, keeping the kernel and normalization unchanged. The original configuration achieves only 4.18%/5.08% (ResNet-18) and 1.30%/1.10% (VGG-16) with Adam/AdamW, respectively. In contrast, the linear-readout configuration achieves 61.41%/62.08% and 64.80%/64.27%, respectively. Thus, changing the optimizer alone does not resolve the original configuration's instability. Further details are provided in Appendix Section F.

Clarification of the 5% improvement

The approximately 5% improvement stated in the abstract refers to the difference between our best ResNet-19 result with augmentation (76.9%) and the previously reported best result (70.5%), corresponding to a 6.4 percentage-point gain. We appreciate the reviewers' concern that this comparison involves different augmentation settings and have revised the abstract to clarify this distinction. Importantly, under a controlled comparison at the same timestep budget ((T=6)) and without augmentation, our method still improves over the baseline by 4.69 percentage points for ResNet-18 and 4.0 percentage points for ResNet-19 (Table 2, E vs. B).

Role of batch normalization

We conducted additional experiments with different random seeds for the BN-only configuration. These experiments further reduced the observed degradation: on CIFAR-10, accuracy increased from the initially reported 89.16% to 90.44% and 90.87% (mean: 90.16%), while on CIFAR-100, it increased from 65.30% to 65.94% and 65.48% (mean: 65.57%). Although performance remains below the corresponding baseline, these results suggest that the magnitude of the degradation is sensitive to the random seed. Thus, while combining BN with the proposed kernel provides more consistent benefits, there is insufficient evidence to conclude that BN is beneficial only when used with the proposed kernel, as BN alone improves accuracy in several other evaluated configurations.

Reproducibility and code

As stated in our reproducibility statement, we had provided the code required to reproduce only the main results of the paper (CIFAR10 and CIFAR100). We now we have updated the package with the missing DVS-CIFAR10, DVS-Gesture and Tiny-ImageNet configurations.

Nature of the contribution

The primary contribution of our work goes beyond combining linear readout, single-exponential kernels, batch normalization, and data augmentation. Prior event-driven training methods were demonstrated on a relatively limited set of architectures, generally at larger timestep budgets (e.g., T=16), and exhibited instability or failure under several settings, including deeper architectures and reduced timestep budgets. Their evaluation was also largely restricted to CIFAR-scale benchmarks. Our work systematically identifies the underlying failure modes and addresses them to obtain stable training across VGG and ResNet architectures, at substantially lower timestep budgets (down to T=2), and on Tiny-ImageNet. On CIFAR-100, we improve over prior event-driven results while reducing the timestep budget from 16 to 6. Thus, our contribution is a systematic stabilization and extension of event-driven, spike-time-based training.

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

This paper revisits time/event-driven gradient-based training of deep spiking neural networks (SNNs), which can reduce training-time temporal state storage but has historically been difficult to optimize at low timestep counts and across architectures. Building on Zhu et al. (2022, 2023), the authors identify silent output neurons, sign-changing derivatives of the double-exponential synaptic kernel, and unstable activity distributions as obstacles. They introduce a non-spiking linear classifier to avoid supervisory spike injection, use a single-exponential (or integrate-and-fire) kernel, and place batch normalization before spiking units while retaining event-time credit assignment. Experiments include CIFAR10/100, DVS-CIFAR10, and Tiny-ImageNet on VGG and ResNet variants.

Critical strengths
  1. Substantive empirical extension of event-time training. The paper goes well beyond one architecture: the VGG11/VGG16/ResNet18/ResNet19 ablation grid (Table 2), low-T CIFAR results, DVS benchmark, and Tiny-ImageNet run provide evidence of practical applicability. The strong VGG11 CIFAR10 result at T=6 (94.35 +/- 0.22%) compared with Zhu et al. (2023) at T=12 (93.54%) is encouraging, with caveats about implementation/training comparability.

  2. Concrete model-level memory measurements and transparent limitations. Appendix Table 6 reports 10.49 vs 15.29 GB for CIFAR10/VGG16, 9.84 vs 20.70 GB for CIFAR10/ResNet18, and 25.54 vs 32.32 GB for DVS-CIFAR10/VGGSNN (event-driven vs activation-based). The manuscript also discloses lack of ImageNet-scale evaluation and reports that non-augmented performance is still below the TET baseline on important settings, e.g., CIFAR100 ResNet19 at T=6 (71.60% vs 74.72%).

Critical weaknesses

The memory reqs are

Missing comparisons with state-of-the-art single-timestep SNNs. The experimental comparison overlooks important low-latency SNN literature. In particular, Chowdhury et al. (ECCV 2022) demonstrated accuracies of 93.05%, 70.15%, and 69.00% on CIFAR10, CIFAR100, and ImageNet, respectively, using only a single timestep. Similarly, Yao et al. (ACML 2023) reported 72.41% accuracy on CIFAR100 with one timestep. In contrast, the proposed method achieves 67.57% at two timesteps and 71.60% at six timesteps on CIFAR100 without additional augmentation. Although these approaches use different architectures and training paradigms, they are relevant to positioning the proposed method within the broader accuracy-latency trade-off.

Limited evidence of scalability to ImageNet-scale datasets. A central claim of the manuscript concerns stable and scalable event-driven SNN training. However, the largest image classification dataset evaluated is Tiny-ImageNet, and Appendix F explicitly acknowledges that ImageNet-scale experiments were not completed.

Relatively weak performance on Tiny-ImageNet. The reported Tiny-ImageNet accuracy of 48.52% with ResNet-18 at four timesteps is not particularly competitive with existing SNN results. Even the TET baseline evaluated in the manuscript achieves 52.40% at the same timestep count. Furthermore, Yao et al. (ACML 2023) report 57.91% accuracy with a single timestep, while Ding et al. (ICLR 2025) report 58.04% with four timesteps. Kim and Panda (2021) also reported 57.8% using BNTT at 30 timesteps.

Unresolved consistency of spike-time gradients and normalization folding (Methods 4.2-4.3; Appendix C-D). The causal kernel epsilon(t)=exp(-t/tau_m) 1[t>=0] has an instantaneous onset and decays between input events. For ordinary positive-threshold neurons, upward threshold crossings can therefore occur at incoming spike discontinuities, rather than at differentiable crossings. It is not clear how the implicit-time derivative using 1/u_dot(t_k) (Appendix Eq. 28), or the claim of using the 'true derivative' (Eq. 12), applies at those points.

Ablation attribution and fair efficiency comparison remain incomplete (Tables 1-2, Fig. 2, Appendix Tables 3 and 6). In Table 2, the largest original-to-baseline change simultaneously removes output-spike injection, changes the output layer, and changes AdamW to Adam, so the reason for the extreme failure of the original variant at T=6 is not isolated. Later B/C/D/E ablations are useful but cannot resolve this initial confound. Likewise, part of the strongest CIFAR100 gain uses an augmented setting (76.9%, without a three-seed SD), whereas the comparable non-augmented setting is 71.60%. GPU memory is compared across implementation pathways that may differ in backend optimization, and an isolated five-fold spike-activation test cannot alone establish training-system scaling. Full-model values help, but the manuscript should give matched measurement conditions and clarify whether savings are attributable to the event-driven gradient representation versus implementation choices. The unaugmented method remains below TET on CIFAR100 ResNet19 (71.60 vs 74.72%) and Tiny-ImageNet ResNet18 (48.52 vs 52.40%), so 'competitive' should be qualified by setting.

Questions for authors

Please see the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer amNk

We thank the reviewer for the detailed and constructive feedback. We have uploaded a revised manuscript and an updated code package. We address the concerns below.

Our work improves event-driven SNN training using spike-time gradients, rather than proposing a general low-latency SNN architecture. We therefore primarily compare with prior event-driven methods sharing the same training formulation. We include ANN and TET (a representative rate-based SNN training method) as reference points against commonly used alternatives, not as primary comparisons.

The low-latency methods highlighted by the reviewer, such as Chowdhury et al. and Yao et al., pursue a different objectives. Chowdhury et al. obtain single-timestep inference by temporally pruning an ANN-converted multi-timestep SNN, whereas Yao et al. recurrently reuse group-wise membrane potentials within a timestep. Neither uses our event-driven, spike-time-gradient formulation.

We aim to narrow the gap over prior event-driven SNN methods while retaining event-driven gradient computation. On CIFAR-100, we improve over prior event-driven results while reducing the timestep budget from 16 to 6 (62.5%), and demonstrate training with as few as 2 timesteps. We do not claim superiority over all low-latency or rate-based SNN methods.

Our experiments do not extend to ImageNet. To our knowledge, this is the first demonstration of event-driven training on Tiny-ImageNet. Despite lower accuracy than some SNN results, this extends evaluation beyond CIFAR-scale. Our scalability claim is limited to progress over prior event-driven methods.

Derivative at discontinuities

Our implementation operates in discrete time, the term 1/u˙(tk)1/\dot{u}(t_k) in the spike-time gradient expression is approximated by the reciprocal of the membrane-potential increment between consecutive simulation steps at output spike events. Specifically, we use

−1utk−utk−1-\frac{1}{u_{t_k}-u_{t_k-1}}

clipped to ([-4,0]) to stabilize the gradient. Thus, this term is a discrete-time estimate of the inverse membrane-potential slope, rather than the exact continuous-time derivative at the threshold crossing. We acknowledge that the membrane trajectory is not differentiable at presynaptic input discontinuities. Our use of the term true derivative refers specifically to differentiating the same synaptic kernel used in the forward dynamics, rather than substituting an independently chosen surrogate function for its derivative in the backward pass as done in the prior approach we consider.

We further clarified the BatchNorm folding formulation, including the role of the fixed inference statistics and the sign-dependent firing condition in the appendix C-D.

Ablation and optimizer choice

To isolate the optimizer effect, we evaluated both configurations with Adam and AdamW, keeping the kernel and normalization unchanged. The original configuration achieves only 4.18%/5.08% (ResNet-18) and 1.30%/1.10% (VGG-16) with Adam/AdamW, respectively. In contrast, the linear-readout configuration achieves 61.41%/62.08% and 64.80%/64.27%, respectively. Thus, changing the optimizer alone does not resolve the original configuration's instability. Further details are provided in Appendix Section F.

Memory measurements and comparison conditions

We ensured a matched comparison by keeping the dataset, architecture, batch size, and other experimental settings identical across both methods. The memory savings primarily arise from the event-driven gradient representation: gradients are stored and propagated at spike events rather than retaining intermediate activations across the full temporal unrolling required by BPTT. As shown in Fig. 2, this reduces the peak number of active tensors from 118 to 19 and peak GPU memory from 4.2 GB to 2.4 GB. The CUDA backend used for event-based backpropagation contributes to runtime acceleration but does not account for the observed memory reduction. We also verified that the reduction persists when using the Python backend, indicating that the memory savings arise from the event-driven gradient representation rather than being solely attributable to backend-specific optimizations. Full-model memory measurements are additionally reported in the Supplementary Material.

Competitive claim

Although our method does not consistently match TET across all settings, particularly on Tiny-ImageNet, the gap is considerably smaller on other benchmarks: 0.15 percentage points on CIFAR-10 and 0.76 on DVS-CIFAR10. We additionally evaluated TET under the augmented CIFAR-100 setting, where our method achieves 76.9% against TET's 78.6% (1.7 percentage-point gap; single seed). All comparisons use an iso-time training budget, with our method offering significant memory savings. We acknowledge the larger gap on Tiny-ImageNet, but emphasize that this is also the first demonstration of our approach on this benchmark. We will qualify our competitiveness claims accordingly.

We thank the authors for their responses. However, upon examining their contributions I have further comments.

This paper uses a non-spiking final readout layer, hence comparing with fully Spiking counterparts is unfair. This might have also contributed to the reduction of T.

The contributions are limited - spiking BN has been proposed before (BNTT, Priya Panda et al.), 4.1 adds an extra layer, 4.2 NEURON KERNEL is not super novel either.

Overall, I feel, the paper is below the ICLR acceptance contribution level, and hence my score is same. Thanks!

Comments on the paper

1 comment
Area_Chair_7CspOpenReview ↗

Please Engage in the Author–Reviewer Discussion

Dear Reviewers,

As the author–reviewer discussion period is approaching its deadline (November 18, AoE), I would like to kindly remind you to review the authors' rebuttals and actively engage in the discussion as soon as possible.

Please address any outstanding questions or concerns and update your assessments and ratings if appropriate.

Thank you very much for your time and efforts!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.