- 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: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:to a single-exponential decay: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.
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) 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% 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.