acceptodds
Under review as a conference paper at ICLR 2027

KFC-SNN: A Kalman-Filtered Complex-Valued Spiking Neural Network for Robust Temporal Learning

Abstract

Spiking neural networks (SNNs) offer a promising approach to energy-efficient temporal processing through intrinsic memory and sparse event communication. However, disturbances can disrupt membrane dynamics and spike generation, undermining computational reliability. Biologically inspired gating mechanisms use temporal state information to improve robustness, motivating a complementary estimation-based account of how current inputs should be combined with neuronal history. Guided by the prediction-and-correction principle of Kalman filtering, we introduce a Kalman-Filtered Complex-Valued (KFC) SNN that combines Kalman-inspired predictive fusion dynamics (KF) with complex-valued spike representations (CS). The KF neuron forms a prediction from the preceding membrane state and treats the current input as an observation. The prediction and observation are adaptively weighted following the Kalman filtering principle to regulate the membrane potential, thereby reducing the influence of input disturbances. The CS mechanism introduces imaginary spikes that regulate real-valued firing through activity-dependent inhibition, while providing an additional output stream that carries temporal information. We further develop a general framework for membrane-potential regulation that describes a class of input-state integration mechanisms through a common convex optimization formulation, with KF as a quadratic instance. Experiments on SHD, SSC, and ESC-50 demonstrate improved classification accuracy and robustness to stochastic input corruptions over the evaluated spiking baselines, including gains of 5.7 percentage points in clean accuracy on ESC-50 and 13.5 percentage points under Gaussian noise on SHD.

Primary area:
learning on time series and dynamical systems
Keywords:
Spiking Neural Networks, Neuronal Dynamics, Robust Temporal Learning, Local Learning

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_YCoZOpenReview ↗
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
  1. 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.
  2. 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
  1. 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?
  2. 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
AuthorsOpenReview ↗

Follow-up Response: Additional Results

Dear Reviewer YCoZ,

Following our previous response, we have added the following results:

  1. Fusion gain and predictor analysis: On SHD, adaptive fusion achieves the highest mean clean and perturbed accuracy among the tested gains (Appendix E.1, Table 9). The layerwise analysis also shows similar mean gains for the two predictors, while the sparse local predictor reduces current reconstruction error in layers 3 and 4 (Appendix E.3, Table 10).
  2. Long-sequence tasks and baselines: We added comparisons with D-RF, P-SpikeSSM, and Spiking LMUFormer. KFC shows limited gains on S-CIFAR10 but achieves the highest clean accuracy on the byte-level LRA-Text variant among the compared models (Section 5.1, Table 2).
  3. Inference latency: We report measured single-sample inference latency on SHD, together with parameter counts (Section 5.2, Figure 3; Appendix A.3). Training time and runtime scaling with sequence length are not included in the current results.
  4. PCA analysis: The feature distribution analysis mentioned in our previous response is provided in the accompanying attachment.

Thank you again for your constructive feedback.

AuthorsOpenReview ↗

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.

  1. 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.

  2. 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.

Reviewer_p4t9OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a Kalman-filtered complex-valued spiking neural network (KFC-SNN). It leverages predictive fusion and dual-channel spike representation to improve the model’s temporal learning capability and strengthen its robustness against random input perturbations. The model predicts the current input using the membrane potential from the previous time step. Combining temporal statistics of input variations and prediction errors, it adaptively fuses predicted results and real observations. The prediction matrix is updated via a local learning rule with sparsity constraints. This paper also designs two spike channels for the real and imaginary components: neural activity in the imaginary channel suppresses spike firing in the real channel, and information from both channels is aggregated for final classification. This paper conducts experiments on the SHD, SSC, and ESC-50 datasets. Multi-seed classification tests, random perturbation experiments, and component ablation studies are performed to verify the efficacy of the proposed method.

Critical strengths
  1. The proposed method achieves substantial improvements under noisy conditions and maintains consistent performance across multiple tasks, demonstrating its generalization ability. For instance, the results in Table 1 show notable gains, indicating that the presented design improves classification accuracy and perturbation tolerance within the evaluated framework.

  2. The component analysis is comprehensive and supports concrete understanding of the architectural design. Beyond the basic ablation of KF and CS, the paper further presents a full four-factor ablation, comparisons with Hilbert transform and repeated input baselines, as well as evaluations among local, global, and fixed predictors, rendering the results persuasive.

Critical weaknesses

The predictor achieves more accurate predictions, yet I argue this alone cannot explain the improved robustness. Specifically, in Table 3, local learning reduces the prediction error from 1.18 for the fixed predictor to 0.38, while the average perturbed accuracy only rises from 80.9% to 81.3%. The global learning scheme yields a prediction error of 10.42 but still maintains a perturbed accuracy of 80.7%. Can this result be interpreted to mean that the model can attain comparable classification performance under vastly different prediction quality? The authors may provide an in-depth analysis of this observation.

Questions for authors
  1. Table 2 jointly disables inhibition and the imaginary-channel readout, making it difficult to distinguish their individual contributions. I suggest incorporating the relevant results from the full ablation in Table 7 into Table 2.

  2. What are the distributions of K for the different predictor variants in Table 3? In particular, when global learning yields a large prediction error, is the prediction pathway largely suppressed? For the same perturbed trajectory, could the authors compare the errors of the observation, prediction, and fused current against the corresponding current from a clean-input run? This would help determine whether the gains arise from predictive denoising.

2 comments
AuthorsOpenReview ↗

Follow-up Response: Additional Results

Dear Reviewer p4t9,

Following our previous response, we have revised the ablation presentation and added a predictor analysis:

  1. Ablation presentation: Table 3 now distinguishes cross-channel inhibition and imaginary-channel readout; the complete four-factor comparison is in Appendix D, Table 8.
  2. Gain and current reconstruction: The two predictors have broadly similar layerwise mean gains. With the sparse local predictor, fusion reduces current reconstruction error in layers 3 and 4, while the task-gradient predictor has higher fused-current error than observation error in all four layers (Appendix E.3, Table 10). The table reports mean gains and observation/fusion errors; it does not include full gain distributions or separate prediction errors.

Thank you again for your helpful suggestions.

AuthorsOpenReview ↗

Reminder to reply

Dear Reviewer p4t9,

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.

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

This paper proposes KFC-SNN, a spiking neural network combining Kalman-inspired predictive fusion (KF) with complex-valued spike representations (CS) for robust temporal learning. The KF mechanism uses a locally learned predictor and adaptive weights based on input discrepancies to integrate historical membrane information with current input. The CS mechanism introduces an auxiliary imaginary spike stream that regulates real-valued firing through cross-channel inhibition and provides additional information for classification. The authors also present a proximal optimization framework to interpret the proposed membrane regulation mechanism. Experiments on SHD, SSC, and ESC-50 evaluate classification accuracy and robustness under stochastic input corruptions. The paper includes ablation studies, predictor-learning analyses, input-encoding comparisons, and supplementary forecasting experiments. The authors report improvements over several evaluated SNN baselines, particularly under stochastic input corruption.

Critical strengths
  1. Interesting methodological design.

(1) The paper introduces a Kalman-inspired predictive fusion mechanism for membrane regulation, providing an estimation-based perspective on SNN dynamics.

(2) The auxiliary spike pathway combines cross-channel inhibition with additional temporal representation, offering a reasonable extension to conventional spiking neurons.

  1. Encouraging experimental results.

(1) KFC achieves improved classification accuracy and robustness under stochastic input corruptions compared with the evaluated SNN baselines, particularly on SHD and ESC-50.

(2) The ablation studies provide evidence for the effectiveness of the proposed components, especially the cross-channel inhibitory mechanism.

Critical weaknesses
  1. Limited contribution of KF. The ablation results show relatively modest improvements from KF in the complete model. Its individual contribution could be further clarified.

  2. Limited empirical validation of the robustness mechanism. Although theoretical analyses are provided, additional experimental evidence on membrane perturbations and spike-event changes would help support the proposed mechanism.

  3. Limited evaluation across temporal resolutions. The experiments use relatively few effective time steps, particularly on SHD. It remains unclear whether the improvements hold at finer temporal resolutions.

  4. Unclear benefit of the complex-valued formulation. The ablation studies support the auxiliary spike pathway, but a comparison with a comparable real-valued dual-stream design would help clarify the role of the complex-valued representation.

Questions for authors
  1. Contribution of KF. Could the authors further clarify the advantages of the proposed adaptive gain over simpler fusion strategies, such as fixed weights or learned gates?

  2. Robustness mechanism. Could the authors further explain how the theoretical analysis of membrane regulation relates to the observed robustness improvements? Additional evidence on membrane perturbations or spike-event changes would be helpful.

  3. Role of complex-valued representations. The ablation studies support the auxiliary spike pathway. Could the authors clarify what benefits arise specifically from the complex-valued formulation, rather than from the additional channel and inhibitory connection?

  4. Temporal resolution. What motivates the choice of five effective time steps for SHD? Could the authors discuss how this temporal resolution may influence the evaluation of the proposed neuron dynamics and the generality of the results?

2 comments
AuthorsOpenReview ↗

Follow-up Response: Additional Results

Dear Reviewer Xcqy,

Following our previous response, we have added the following results to the revised manuscript:

  1. Fusion gain comparison: On SHD, the adaptive gain achieves the highest mean clean and perturbed accuracy among the tested fixed, learnable, and adaptive gains (Appendix E.1, Table 9).
  2. Membrane and spike stability: Enabling KF reduces membrane deviation and spike flip rates across all four layers under input perturbations (Appendix E.2, Figure 6).
  3. Long-sequence tasks: On S-CIFAR10, KFC shows limited gains; on the byte-level LRA-Text variant, it achieves the highest clean accuracy among the compared models (Section 5.1, Table 2).
  4. PCA analysis: The feature distribution analysis discussed in our previous response is provided in the accompanying attachment.

Thank you again for your suggestions.

AuthorsOpenReview ↗

Reminder to reply

Dear Reviewer Xcqy,

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.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

Dear Area Chair and Reviewers,

We thank all three reviewers for their constructive feedback. We have provided point-by-point responses and uploaded a revised manuscript. Our work introduces Kalman-inspired predictive fusion (KF) and complex-valued spikes (CS) within a general framework for membrane-potential regulation. We highlight three revisions for your consideration.

1. Clarifying the robustness mechanism

With cross-channel inhibition retained and imaginary-channel readout disabled, KF improves mean SHD accuracy under corruption by 4.55 percentage points (Table 8). In a comparison of fusion rules, adaptive KK achieves the highest mean accuracy on both clean and perturbed inputs among the three fixed gains, a learnable per-neuron gain, and the KF-disabled setting (Table 9). KF also reduces membrane-potential deviations and spike-flip rates in all four layers (Figure 6). TThe supplementary material uses PCA to illustrate how predictor choice shapes the learned representations through the adaptive fusion gain KK.

2. Expanding the benchmarks and baselines

We extended our evaluation to two long-sequence classification tasks, sequential CIFAR-10 and the Long Range Arena (LRA) text classification benchmark, with comparisons against D-RF, P-SpikeSSM, and Spiking LMUFormer (Table 2). On LRA-Text, KFC achieves the highest accuracy among the evaluated models on clean inputs (84.8%) and at every tested Gaussian noise level.

3. Quantifying computational cost and inference latency

On SHD, KFC uses fewer effective parameters and fewer estimated linear operations than standard LIF (Table 6). Its measured inference latency is 0.179 ms per sample, compared with 0.245 ms for standard LIF, 0.354 ms for P-SpikeSSM, and 1.049 ms for Spiking LMUFormer (Figure 3).

We respectfully ask the Area Chair and reviewers to consider these findings alongside our individual responses and the revised manuscript in their final assessment.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.