- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2(+1)
- Soundness
- 2
Read review
Paper summary
This work presents a spiking neural network architecture that pairs Kalman-style predictive fusion with an auxiliary imaginary spike stream to handle inhibitory regulation and temporal representation. To train the predictor, the authors use a local learning rule alongside sparsity regularization, and they further frame the membrane regulation mechanism through the lens of proximal optimization. On the empirical side, the model is tested on audio classification tasks for both accuracy and robustness against random noise, backed by component ablations, predictor diagnostics, and temporal reconstruction probes.
Critical strengths
The model's performance under stochastic corruption really stands out. On both SHD and ESC-50, it achieves notable gains over the competing spiking baselines across a variety of corruption families. It is also good to see the same Hilbert encoding applied uniformly across the main baselines as a controlled variable. In general, these findings nicely back up the claim that the proposed dynamics improve robustness in the tested scenarios.
Critical weaknesses
- The actual benefit of the Kalman-inspired mechanism is not convincingly demonstrated. The gain calculation is based on heuristic discrepancy traces rather than true calibrated uncertainties, and the proximal interpretation fails to show any clear advantage over much simpler fusion strategies. In practice, removing the KF component barely hurts corrupted SHD accuracy (dropping only from 81.20% to 80.12%), and freezing the predictor yields almost the exact same performance as training it with the local rule.
- Limited evaluation on long-range temporal modeling. The primary experiments on SHD, SSC, and ESC-50 operate over very short horizons of 5, 25, and 50 effective time steps, respectively. Even with the appendix forecasting numbers, it remains unclear how well the network scales to longer temporal dependencies. Testing on established benchmarks like PS-MNIST, S-CIFAR10, and Long Range Arena is needed here. The experimental section is also missing head-to-head comparisons with strong sequence-modeling SNN baselines, notably P-SpikeSSM [1], D-RF [2], and LMUFormer [3], which leaves the relative positioning of the proposed model somewhat ambiguous.
[1] Bal, M., & Sengupta, A. P-spikessm: Harnessing probabilistic spiking state space models for long-range dependency tasks. ICLR 2025.
[2] Zhang, D. et al. Dendritic resonate-and-fire neuron for effective and efficient long sequence modeling. NeurIPS, 2026.
[3] Liu, Z., et al. Lmuformer: Low complexity yet powerful spiking model with legendre memory units. ICLR 2024.
Questions for authors
- If we keep the predictor, imaginary pathway, and training budget fixed, how does the proposed discrepancy-based gain perform compared to a simple constant gain or a learnable per-neuron gain?
- Have you benchmarked the actual training time and inference latency on the same hardware against recent SNN baselines, along with a breakdown of the overhead from predictive fusion, local learning, and the imaginary pathway? Right now, Table 5 only gives analytical operation counts rather than measured wall-clock runtime. Seeing how the actual runtime scales with increasing sequence length would be very helpful for judging the practical scalability of the method.
4 comments
Reminder to reply
Dear Reviewer YCoz,
I hope this message finds you well. As the discussion period is nearing its end with less than one day, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.
Thanks again for your time and effort in reviewing our paper.
Thank you for the detailed response and the substantial revisions to the manuscript.
-
Regarding long-sequence tasks and baselines, I appreciate the authors’ transparent reporting of the results on S-CIFAR10, even though they do not reach state-of-the-art performance. My original request concerning Long-Range Arena (LRA) was intended to encourage a broader evaluation across the benchmark, including ListOps, Retrieval, Image, Pathfinder, and Path-X, rather than LRA-Text alone. Such an evaluation would provide a more complete picture of the model’s ability to capture long-range dependencies across different tasks.
-
Regarding computational efficiency, my concern is inference latency and how it scales with sequence length. The measured single-sample inference latency on SHD and the parameter counts are useful additions. However, training time and runtime scaling with sequence length have not been reported. Measurements on the SHD dataset do not establish whether the approach remains computationally practical for substantially longer sequences. Given the emphasis on noise robustness, this question is particularly relevant to potential applications in speech processing, where long audio streams and low-latency requirements often coexist. Reporting inference latency over a range of sequence lengths, alongside appropriate baselines under the same hardware conditions, would help clarify the model’s practical applicability. Training-time measurements would also provide a fuller assessment of its computational cost.
In recognition of the authors’ efforts and the additional evidence provided, I am raising my score to 2, although the concerns above remain only partially addressed.
Dear Reviewer YCoZ,
Thank you for your constructive feedback and for reconsidering your evaluation. We appreciate your suggestions regarding broader long-range benchmarks and computational efficiency.
Regarding the LRA benchmarks, due to limited computational resources and the time constraints of the rebuttal period, we selected LRA-Text as a representative task that differs substantially from our original benchmarks, aiming to provide an initial assessment of the model's cross-task generalization capability. We agree that evaluating the remaining LRA tasks would provide a more comprehensive understanding of the model's ability to capture long-range dependencies. If accepted, we will extend our evaluation to the remaining LRA tasks in the camera-ready version.
Regarding computational efficiency and latency, we believe that the measurements on SHD, despite its relatively short temporal sequences, still provide meaningful insights into computational efficiency, as the computational costs of the evaluated models generally scale approximately linearly with the number of time steps, rather than exhibiting nonlinear growth. Nevertheless, we acknowledge that direct measurements across different sequence lengths would provide stronger empirical evidence, and we will include these results in the camera-ready version.
As for training time, the main difficulty lies in securing sufficiently long, uninterrupted computational resources to train our model and all baselines under identical hardware conditions and system loads, ensuring a fair comparison. We recognize the importance of this metric and will provide the corresponding measurements in the camera-ready version.
Thank you again for your valuable feedback and suggestions.
Follow-up Response: Additional Results
Dear Reviewer YCoZ,
Following our previous response, we have added the following results:
Thank you again for your constructive feedback.