acceptodds
Under review as a conference paper at ICLR 2027

Spatio-Temporal Margin Regularization for Mitigating Temporal Error Accumulation in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) exploit spatio-temporal spike dynamics for low-power computation and temporal modeling, yet early misfired spikes trigger cumulative temporal errors that amplify across iterative dynamics and degrade generalization. To address this issue, we propose the dual Spatio-Temporal Margin Regularization (STMR) framework, which mitigates the temporal error accumulation from both the perspectives of spike dynamics and output representation. Temporal margin regularization quantifies class-specific spike latency via weight mapping, encouraging target-correlated neurons to fire earlier than negative-correlated ones to suppress error generation at its source. Spatial margin regularization further constrains the per-timestep confidence margin between the target and the non-target risk class, with timestep-coordinated adaptive weights that focus on highly erroneous timesteps to balance regularization strength across the temporal dimension. Extensive experiments on 1D, 2D, and 3D benchmarks with spiking VGG, ResNet, and Transformer architectures demonstrate consistent performance gains. Notably, STMR achieves 87.1% accuracy on CIFAR10-DVS, and visual analysis confirms that it effectively advances target-class spike activation and narrows negative confidence margins. This work provides new insights into spike dynamics-aware optimization for SNNs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spiking Neural Network, Margin Regularization, 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 4 reviews, above 27% of the others.
Reviewer_zVbUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spatio-Temporal Margin Regularization (STMR) to mitigate temporal error accumulation in spiking neural networks. The authors argue that incorrectly fired spikes at early timesteps may propagate through neuronal dynamics and degrade prediction performance. To address this issue, STMR introduces two components: Temporal Margin Regularization (TMR), which encourages target-related neurons to fire earlier, and Spatial Margin Regularization (SMR), which reduces confidence margins favoring incorrect classes across timesteps. Experiments on various datasets and architectures demonstrate improvements in classification accuracy without additional inference overhead.

Critical strengths

1.The paper addresses temporal error accumulation in SNNs from both spike dynamics and output representations, providing a reasonable motivation for the proposed regularization framework. 2.The method is evaluated on diverse tasks and datasets, including static images, neuromorphic vision, speech recognition, and 3D point cloud classification. The experiments also cover different architectures, such as VGG, ResNet, and Transformer.

Critical weaknesses

1.The authors use the absolute values of classifier weights to estimate class-specific spike latency. However, this formulation does not distinguish between positive and negative contributions to the target class. 2.The theoretical analysis is not sufficiently rigorous. In particular, the inequality in Eq. (13) does not appear to hold in general, since the softmax of temporally averaged logits is not equivalent to averaging the per-timestep softmax probabilities. 3.Although the experimental results are competitive, some improvements over existing methods are relatively small.

Questions for authors

1.Why are the absolute values of classifier weights used to estimate class-specific spike latency? Neurons with negative weights may decrease the target-class logit. 2.Could the authors provide an ablation study comparing the adaptive timestep weighting in SMR with uniform weighting to demonstrate its necessity?

5 comments
AuthorsOpenReview ↗

We sincerely appreciate your insightful comments

We sincerely appreciate your insightful comments. Below is our point-by-point response:

Signs of classifier weights. Using only positive or negative weights results in one type of weight being ignored, whereas using the absolute value of the weights takes into account the magnitude of contributions from both types. To verify the advantages of using absolute weight values in the proposed method, we compared it with methods that use only positive or negative weight values. As shown in the table below, using absolute values of weights consistently yields better performance.

Weighting schemeCIFAR10-DVSDVS-Gesture
Positive weights only74.387.85
Negative weights only73.790.63
Absolute value of weights75.991.67

Revision of Eq. (13). We have revised Eq. (13) using logit margins, thereby resolving the mismatch between the softmax function corresponding to the time-averaged logits and the mean of the softmax probabilities at each timestep. We emphasize that the revised analysis preserves our conclusion that SMR can improve the final probability margin. Please refer to Section B.1 of the revised submission.

Advantages of the proposed method. We would like to emphasize that, unlike other methods that modify the architecture or neurons, the proposed method regularizes the spatio-temporal impulse margins of SNNs and requires no changes to existing models. It offers superior generalizability, which can further enhance the performance of these methods. Furthermore, our method can boost the accuracy of SNNs by up to 4.52% with the simple plug-and-play setup. We hope this illustrates the full significance and strength of this work.

Adaptive timestep weighting versus uniform weighting. We compared our proposed adaptive timestep weighting method with the vanilla uniform weighting method on CIFAR10-DVS and DVS-Gesture. The results in the table below once again demonstrate the advantages of our proposed adaptive weighting method.

Weighting schemeCIFAR10-DVSDVS-Gesture
Uniform74.889.24
Adaptive75.991.67

We have made revisions to the submitted manuscript and highlighted the changes in blue. We hope these responses address your concerns, and we look forward to discussing any further questions you may have.

Thank you for the response and the additional experiments. However, I believe my main concerns have not been fully addressed.

  1. My main concern is that using absolute weights may encourage not only neurons that positively contribute to the target-class logit to fire earlier, but also those that negatively contribute to it. I believe this issue has not been sufficiently explained. Using absolute weights still seems to be an empirical design choice.

  2. My point was that some improvements over existing methods are relatively small, not that the improvements are small on all datasets.

Therefore, I will keep my original rating unchanged.

AuthorsOpenReview ↗

Thank you very much for your further response.

We greatly appreciate your response, which is crucial to clarifying the core attributes of the proposed method.

  • The effect of the absolute value of classifier weights. TMR measures the strength of connections between neurons and categories based on the absolute values of the classifier weights. Optimizing the TMR loss encourages neurons strongly correlated with a specific category to fire preferentially, which indeed includes both positive and negative correlations. However, the proposed method does not rely solely on TMR. The primary loss is the cross-entropy loss, which promotes correct classification. SMR further encourages the target category to have a more significant predictive margin relative to competing error categories. When TMR triggers spikes from negatively correlated neurons, it decreases the SNN’s logit output for the target category. The CE and SMR losses then work to suppress and offset the impact of these category-negative correlated spikes. Thus, the proposed STMR promotes preferential firing of positively correlated neurons, achieving an overall gain. In the future, we will delve deeper into metrics that can more comprehensively reflect the contributions of spiking neurons to categories, mitigating the impact of negatively correlated neurons even further.

  • Improvements over existing methods. The proposed method is independent of specific architectures, neuron types, and training methods, and can therefore be applied to a variety of existing SNNs. The extent to which the proposed method improves the performance of various existing methods depends on the characteristics of those methods. Nevertheless, we would like to emphasize that even if the improvement in performance on a specific existing method is not particularly significant, the experiments in Tables 1 and 2 have fully validated the effectiveness and generality of the proposed method. Accuracy improvements range from 0.70% to 2.43% across the four methods: PLIF, CLIF, TAB, and SLT. Furthermore, experiments also demonstrate that the proposed method is fully compatible with mainstream spiking VGG, ResNet, and Transformer architectures. Taken together, these results demonstrate that STMR is a competitive, plug-and-play regularization method for enhancing other methods because it preserves the network architecture and inference workflow of existing methods. Given the increasing complexity of current SNN architectures and neuron types, we believe such a seamless performance enhancer is highly necessary.

We hope these responses address your concerns fully by clarifying the mechanism by which the proposed method works and positioning our work within the SNN community. We look forward to discussing any concerns you may still have.

Reviewer_L8aUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Early erroneous spikes in SNNs can affect subsequent network states, potentially causing classification decisions to deviate progressively. This paper proposes Spatio-Temporal Margin Regularization (STMR), which comprises two regularization terms designed to mitigate this error accumulation and improve SNN performance.

Critical strengths
  1. The method is well motivated.
  2. The reported experimental results are promising.
Critical weaknesses
  1. STMR may incur substantial training overhead because it requires tracking neuronal state trajectories throughout the network.
  2. The experimental evaluation is limited in scope.
  3. Hyperparameter tuning appears challenging.
Questions for authors
  1. The authors characterize error accumulation as a pervasive phenomenon in SNNs, but I am not convinced that this generalization is sufficiently justified. In my view, discrete spiking dynamics can suppress or eliminate errors in many cases. Equation (3) adopts a soft-reset mechanism, which can indeed permit error accumulation. However, I do not consider this mechanism representative of the most commonly used reset schemes. My intuition is that error accumulation may be less pronounced under the commonly used hard-reset mechanism. The paper would therefore benefit from grounding its motivation in a more careful discussion of the relationship between reset mechanisms and error accumulation. Additional experiments across different reset mechanisms would help establish whether STMR provides consistent benefits and clarify its scope of applicability.
  2. STMR requires monitoring neuronal dynamics throughout the network. What additional computational and memory overhead does this introduce during training? Please quantify the training time and memory usage relative to the baseline.
  3. STMR regularizes only samples that are ultimately misclassified. However, a correct final prediction does not necessarily imply satisfactory behavior at earlier time steps. For example, a strong corrective signal at the final time step could compensate for poor intermediate predictions. The authors should discuss this possibility and its implications for restricting regularization to ultimately misclassified samples.
  4. The experimental evaluation lacks breadth and is restricted to small-scale datasets. The effectiveness of STMR on larger-scale datasets remains unclear and warrants further investigation.
  5. Hyperparameter tuning appears challenging because the two regularization hyperparameters interact. Moreover, the settings reported in the appendix span a wide range, raising concerns about how readily suitable configurations can be identified for new datasets. The paper also lacks a joint grid-based ablation of these hyperparameters. Such an analysis would help characterize their interaction and sensitivity and provide practical guidance for applying STMR to new datasets.
3 comments
AuthorsOpenReview ↗

We sincerely appreciate your insightful comments

We greatly appreciate your thoughtful feedback. Our responses to each point are below:

Generalizability across reset mechanisms. To verify the effectiveness of STMR under different reset mechanisms, we conducted experiments using VGG-9 on the CIFAR10-DVS and DVS-Gesture datasets. The results are shown in the table below. The consistent performance gains achieved by the proposed method suggest that although hard resets are more effective than soft resets at mitigating the accumulation of errors, errors still persist. Therefore, STMR’s facilitation of the spatio-temporal margin also yields positive results.

MethodSoft reset-CIFAR10-DVSSoft reset-DVS-GestureHard reset-CIFAR10-DVSHard reset-DVS-Gesture
Vanilla72.987.1574.688.54
STMR (Ours)75.991.6776.393.06

Near-zero additional overhead. STMR calculates spike latencies based solely on the final classifier weights and spike trains rather than continuously monitoring all neurons throughout the network. Therefore, it does not introduce significant overhead. We compared the average batch time and memory overhead during training between the conventional method and STMR using the VGG-9 architecture on DVS-Gesture under identical conditions to evaluate this. The table below further demonstrates that the overhead introduced by STMR is negligible.

MethodTime per batch (10⁻² s)Memory (MB)
Vanilla8.745474
STMR (Ours)8.755476

Regularize only the misclassified samples. To prevent the over-regularization of samples that were ultimately predicted correctly, which could hinder generalization, the default STMR only applies regularization to samples that were predicted incorrectly. However, as you pointed out, we acknowledge that samples that are ultimately predicted correctly may have been incorrectly predicted at intermediate time steps. Therefore, we compared the performance of regularizing only the samples that were ultimately predicted incorrectly to the performance of regularizing all samples. As shown in the table below, regularizing only incorrect samples consistently achieved better performance. To further improve the performance of our proposed method, we could extract more information from samples incorrectly predicted at intermediate time steps yet ultimately predicted correctly by setting an appropriate, sample-specific regularization strength. We will investigate this in future work.

Regularized samplesCIFAR10-DVSDVS-Gesture
All samples74.588.19
Only misclassified samples75.991.67

Large-scale experiments. In Table 8, the proposed method has demonstrated superior performance on the CIFAR100 dataset, indicating that it is capable of covering static objects across 100 categories. Furthermore, we extended our proposed method to the larger-scale Tiny-ImageNet dataset, and the results in the table below once again demonstrate its competitive performance.

MethodArchitectureTAcc. (%)
ASGL [1]VGG-13856.81
Joint A-SNN [2]VGG-16455.39
MPS [3]VGG-13458.04
MPD-AGL [4]VGG-13458.14
LM-H [5]VGG-13459.93
LM-H [5] (radical version)VGG-13460.37
CLIF [6]VGG-13462.10
STMR (Ours)VGG-13462.58
[1] Adaptive smoothing gradient learning for spiking neural networks. ICML. 2023.
[2] Joint a-snn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition. 2023.
[3] Rethinking spiking neural networks from an ensemble learning perspective. ICLR. 2025.
[4] Adaptive gradient learning for spiking neural networks by exploiting membrane potential dynamics. IJCAI. 2025.
[5] A progressive training framework for spiking neural networks with learnable multi-hierarchical model. ICLR. 2024.
[6] Clif: Complementary leaky integrate-and-fire neuron for spiking neural networks. ICML. 2024.

Hyperparameter settings. Figure 6 in the Appendix provides a joint analysis of the hyperparameters of the VGG and QKFormer architectures. Although hyperparameters affect the performance of the proposed method, the results show that it consistently outperforms the baseline models. For setting the regularization weights, we recommend starting with smaller values, such as 0.1 or 0.01. This is because STMR directly affects the temporal order of the spike sequences, which can significantly influence spiking dynamics. In future work, we will explore adaptive adjustment of the STMR regularization strength based on spike dynamics and the evolution of temporal features in the samples to fully unleash the positive effects of STMR.

We have made revisions to the submitted manuscript and highlighted the changes in blue. We hope these responses address your concerns, and we look forward to discussing any further questions you may have.

Thank you for the effort you put into the rebuttal. I still have some reservations about the evidence for the mechanism underlying the proposed method and about the breadth and completeness of the experimental evaluation. In my view, both aspects need further development for the paper. Therefore, I will keep my original score.

AuthorsOpenReview ↗

Thank you for your reply. We would appreciate it if you could provide more specific details.

Thank you for your further reply. Since the remaining reservations concern the mechanism and experimental coverage, we would appreciate clarification on which specific questions remain unresolved after the additional analyses. To facilitate this assessment, we summarize the evidence relevant to the original concerns.

We evaluated both reset mechanisms. On CIFAR10-DVS, STMR improves accuracy from 72.9% to 75.9% under soft reset and from 74.8% to 76.3% under hard reset. On DVS-Gesture, the corresponding improvements are from 87.15% to 91.67% and from 88.54% to 93.06%. These results establish benefits under both mechanisms and clarify the applicability of STMR beyond the soft-reset setting.

Evidence concerning the design and evaluation. We directly compared regularization on all samples with regularization restricted to misclassified samples. The latter achieves 75.9% versus 74.5% on CIFAR10-DVS and 91.67% versus 88.19% on DVS-Gesture. We also quantified training overhead, with measured time changing from 8.74 to 8.75 ×10−2s\times 10^{-2} s and memory from 5474 to 5476 MB. This indicates that our method incurs almost no additional overhead. In addition, Figure 6 in the Appendix provides a joint hyperparameter analysis for VGG and QKFormer, showing that the proposed method consistently yields positive gains. On experimental coverage, we added Tiny-ImageNet, where STMR achieves 62.58% accuracy with VGG-13 at four timesteps, compared with 62.10% for CLIF at the same architecture and timestep count.

These additions directly address the reset mechanism, sample-selection policy, computational overhead, hyperparameter interaction, and experimental coverage raised in the original review. Could you specify which remaining aspect of the mechanism or evaluation you consider insufficient, and how the additional evidence falls short? This would allow us to respond precisely and help the Area Chair assess the paper on the basis of the full discussion.

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

The paper addresses temporal error accumulation in spiking neural networks (SNNs), arguing that early misfired spikes can propagate through recurrent membrane dynamics and lead to negative confidence margins over time. It proposes Spatio-Temporal Margin Regularization (STMR), a dual regularization scheme: (i) a temporal margin regularizer (TMR) that advances target-class spike timing via a class-specific spike latency surrogate computed by weight mapping, and (ii) a spatial margin regularizer (SMR) that penalizes the per-timestep negative confidence margin between the target and the most competitive non-target class using adaptive, timestep-coordinated weights. Experiments across 1D/2D/3D tasks and VGG/ResNet/Transformer SNNs show consistent gains, with qualitative analyses indicating earlier target spike activation and reduced negative margins.

Critical strengths

1. Well-motivated regularization design.

The paper addresses temporal error accumulation in SNNs from both spike timing and output confidence perspectives. The class-specific spike latency metric provides a practical way to regularize firing behavior, while the adaptive timestep weighting focuses on more error-prone predictions. These two components complement each other and can be integrated into existing SNN training pipelines.

2. Extensive experimental evaluation.

The method is evaluated across 1D, 2D, and 3D datasets using different SNN architectures, including VGG, ResNet, and Transformers. The ablation studies show that both regularization terms contribute to the improvements. Additional compatibility experiments and visualizations of spike timing and confidence margins further support the effectiveness of the proposed method.

Critical weaknesses

1. Some methodological choices need further justification.

The class-specific spike latency metric uses absolute classifier weights, which may not fully distinguish positive and negative contributions to the target class. In addition, the handling of neurons that never spike is unclear, as the average spike-time calculation involves division by the total spike count. These details could affect the interpretation and stability of the proposed temporal regularization.

2. Limited comparisons with related temporal training methods.

Although the experiments cover multiple datasets and architectures, direct comparisons with related temporal training strategies, such as TET and other timestep-level regularization methods, are limited. Such comparisons would help clarify the advantages of the proposed margin-based approach. The impact on firing rates and computational efficiency is also not evaluated.

Questions for authors
  1. Class-specific spike latency: Why does TMR use absolute classifier weights instead of sign-aware weights? Since negative weights may indicate negative contributions to the target class, could the authors clarify this choice or provide a comparison with positive-only weighting?

  2. Handling non-spiking neurons: How is the average spike time in Eq. (5) computed when a neuron produces no spikes? Since the denominator becomes zero, clarification of the implementation would help assess the numerical stability of TMR.

  3. Comparison with related methods: How does STMR compare with temporal training methods such as TET, ETC, and TRT under matched experimental settings? Such comparisons would help clarify the benefits of margin-based regularization over existing timestep-level training strategies.

1 comment
AuthorsOpenReview ↗

We sincerely appreciate your insightful comments

We sincerely appreciate your insightful comments, especially your recognition of our motivation and the extensive experiments we conducted. Below is our point-by-point response to your concerns:

Signs of classifier weights. Classifier weights reflect a neuron's influence on a specific class cc. Positive weights increase the corresponding logit, and negative weights decrease it. The absolute value captures the strength of influence from both types of weights. To verify the advantages of using absolute weight values in the proposed method, we compared it with methods that use only positive or negative weight values. As shown in the table below, using absolute values of weights consistently yields better performance.

Weighting schemeCIFAR10-DVSDVS-Gesture
Positive weights only74.387.85
Negative weights only73.790.63
Absolute value of weights75.991.67

Processing of non-firing neurons. To prevent any influence of non-firing neurons on numerical stability, Equations (4) and (5) exclude the weights of these permanently silent neurons from the numerator and denominator when calculating the spike latency, and add a minimum value of 1e-8 to the denominator of Eq. (4). We have clarified this in the revised version, though it does not affect the present paper or the core method.

Comparison with other temporal training methods. Unlike previous methods, such as TET, ETC, and TRT, which extend the cross-entropy loss to each timestep and promote temporal consistency, this work identifies the timing of spikes across different categories as the root cause of suboptimal performance. Therefore, this paper is driven by a completely different motivation than previous methods. Additionally, the table below presents comparative results against TET, ETC, and TRT, further validating the superior performance of the proposed method.

MethodCIFAR10-DVSDVS-Gesture
TET75.391.32
ETC74.990.97
TRT75.190.63
STMR (Ours)75.991.67

Spike firing rate and energy efficiency. To evaluate the influence of the proposed method on the spike firing rate and energy efficiency of SNNs, we calculated the spike firing rate of VGG-9 across all test samples on CIFAR10-DVS and computed the inference energy, following the widely adopted method in the SNN community. The results demonstrate that the proposed method optimizes spike trains and significantly reduces the spike firing rate and overall inference energy consumption. These improvements further enhance the deployment and inference advantages of the proposed method.

SFRLayer 1 (%)Layer 2 (%)Layer 3 (%)Layer 4 (%)Layer 5 (%)Layer 6 (%)Layer 7 (%)Layer 8 (%)Energy (10⁶ pJ)
Vanilla12.314.443.843.532.511.801.308.97260.00
Ours7.132.852.242.151.491.090.965.96174.91

We have made revisions to the submitted manuscript and highlighted the changes in blue. We hope these responses address all your concerns. We look forward to further discussion with you.

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

The paper suggests that SNN errors often stem from early spikes whose effects build up over time. STMR tackles this by adding two losses for misclassified samples. The first loss ensures the target class spikes sooner than its earliest rival, and the second focuses on time steps where a competitor is more confident. Notably, the experiments on vision, speech, and point cloud data appear to support the training objective itself more strongly than the underlying mechanism.

Critical strengths
  1. Table 1 tests both losses separately and together across three architectures and two event-vision datasets at four time steps. Each loss improves on the baseline in all six settings, and their combination gives the highest reported accuracy in each.
  2. Figure 3 separates corrected errors from persistent errors. In the corrected group, the median target latency gap falls from 0.065 to 0.029 but remains positive, showing that correct classification can coexist with a target class that still lags its earliest competitor by this measure.
Critical weaknesses
  1. The paper claims in Section 5.3 that a smaller latency gap implies fewer errors. Yet, this gap might decrease simply because the model weighs early-firing neurons more heavily, rather than because the spike trains have improved. The confidence trajectories in Figure 4 and the assumed perturbation gains in Appendix B.2 do not clearly resolve this ambiguity.
  2. If per-step confidence is a softmax probability, Eq. (13) fails for the predictor in Eq. (12). For binary target-minus-competitor logit differences of 3,3,-5, the final probability margin is approximately 0.165, below the mean per-step margin of 0.275. The proposed connection to the generalization bound therefore needs a different argument.
Questions for authors
  1. Can the Figure 3 analysis separate changes in spike timing from changes in readout weights, for example by holding the readout fixed where neurons can be matched?
2 comments
AuthorsOpenReview ↗

We sincerely appreciate your insightful comments

We sincerely thank you for your efforts in reviewing this paper and for your insightful feedback. Below is our point-by-point response to your concerns:

The proposed method optimizes spike trains. To isolate the influence of the classifier’s readout weights on the time margin, we initialized the VGGSNN with the same frozen classifier weights and trained it using both the standard cross-entropy loss and the proposed STMR method, respectively. We then compared the pulse time margins under fixed weights and identical sets of activated neurons. The results in the table below show that the proposed method consistently demonstrates a significant advantage in time margin, proving its effectiveness in optimizing spike trains. We have included these analysis results in the revised PDF, specifically in Figure 4 of Section 5.3 and Figure 7 of Appendix E.3.

Time median margin→Persistently misclassified samples — VanillaPersistently misclassified samples — STMRSamples correctly classified only by STMR — VanillaSamples correctly classified only by STMR — STMR
Same readout weights0.0500.0390.0620.038
Same readout weights and activated neuron sets0.0490.0400.0560.037

Revision of Eq. (13). Thank you for pointing out the mismatch between Eq. (13) and the temporal aggregation in Eq. (12). In the revised Eq. (13), we use logit margins and show that the final logit margin is at least the average of the per-timestep logit margins. We emphasize that the revised analysis preserves our conclusion that SMR can improve the final probability margin. In particular, increasing the target-class logit at a timestep with a probability deficit, while holding all other logits fixed, reduces that timestep’s deficit and increases the final probability margin after temporal aggregation. We have made this sufficient condition explicit in Appendix B.1 to clarify how correcting a per-timestep deficit targeted by SMR can improve the final prediction margin.

Finally, we hope these responses address your concerns, and we look forward to receiving your further feedback.

Thank you for your detailed response and the additional controlled experiments. While the new results with frozen readout weights provide useful evidence that the method optimizes spike timing and the revision to Eq. (13) resolves the mathematical discrepancy, the connection between the observed latency reduction and the causal mechanism of error correction remains somewhat indirect, as the persistent error group exhibits a similar reduction in time margin. Therefore, I will maintain my current rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.