acceptodds
Under review as a conference paper at ICLR 2027

Functional Prox: Decoupling Importance Estimation from Optimization for Continual Learning in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) offer a promising path toward energy-efficient intelligent computing, yet retaining prior knowledge while continually learning new tasks remains a key challenge for continual adaptation. To mitigate catastrophic forgetting, importance-based regularization methods typically use gradients to estimate parameter importance and consolidate prior knowledge by constraining changes to important parameters. However, the non-differentiability of spike generation makes these estimates depend on the choice of surrogate gradient. Even when network parameters and forward predictions remain unchanged, changing only the backward surrogate can substantially alter hidden-parameter importance rankings or even cause importance values to explode, resulting in misplaced or excessive parameter constraints. To address this issue, we propose Functional Prox, a continual learning method that decouples functional importance estimation from surrogate-gradient optimization. Functional Prox estimates importance solely through forward spiking computations, measuring changes in predictive distributions under channel-level interventions to quantify functional contributions. These scores are converted into layer-normalized protection weights and combined with a closed-form, non-expansive proximal update to control constraint magnitudes and parameter displacement. Experiments on CIFAR-100, Tiny-ImageNet, and ImageNet-100 under class-incremental learning (CIL) setting show that Functional Prox significantly outperforms EWC and other representative regularization methods, improving both knowledge retention and new-task learning.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual learning, spiking neuron network

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yNLyOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Functional Prox, a continual learning method for spiking neural networks. The authors identify that existing gradient-based importance estimation methods are sensitive to the choice of surrogate gradients, which may lead to unreliable parameter protection. To address this issue, the proposed method estimates channel importance using JS divergence through forward interventions and applies layer-wise sparse protection with proximal updates to mitigate catastrophic forgetting.

Critical strengths

1.The paper addresses an important problem in continual learning for SNNs. The observation that different surrogate gradients can produce inconsistent importance estimates is interesting and provides a reasonable motivation for the proposed method. 2.The proposed forward-only importance estimation is intuitive and avoids dependence on surrogate gradients when evaluating a fixed network. The combination with proximal updates provides a reasonable approach to protecting previously learned knowledge.

Critical weaknesses

1.The paper is difficult to follow. In the Related Work section, the authors divide the discussion into two parts, but both parts mainly list existing studies without sufficient analysis. The connections between these studies and the proposed method are not clearly explained. The authors should better discuss how their work differs from and improves upon existing approaches. 2.In the Method section, the authors introduce several components and mathematical formulations without clearly explaining the motivation behind each component, the purpose of each equation, or how these components are connected. The overall presentation and organization of the method need significant improvement. 3.Could the authors compare JS-divergence-based importance with simpler methods, such as firing-rate-based importance and random channel selection, to better demonstrate its effectiveness? 4.The experiments are limited to conventional image datasets. More experiments on event-based datasets would help validate the proposed method. 5.Overall, the writing of the paper should be improved. Humans should review AI-generated content to ensure that it is logical and coherent.

Questions for authors

See weaknesses

Reviewer_BN75OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper explores representational regularisation for CIL. Here, the author's main contribution is to decouple functional importance from surrogate gradient optimisation and to show that changes in the surrogate gradient affect hidden-layer importance. The current SOTA methods only study importance with respect to the network's final representation.

Critical strengths
  1. Normalising protection budgets locally per layer prevents deep architectural layer imbalances and focuses constraints strictly on functionally important channels, maintaining network plasticity for future tasks

  2. Utilising a closed-form proximal operator avoids the numerical instability or overshooting common in explicit gradient penalty updates when regularisation coefficients become large.

Critical weaknesses
  1. The main paper, except for the appendix, doesn't contain any tables for comparison. Tables 1, 2, and 3 don't show or compare against the SOTA SNN methods such as [1], [2], and [3].

  2. Spiking vision transformers may be better at capturing features for continual learning. We would like to know whether the method is invariant to the backbone; i.e., extend the experiments to different backbones to analyse its sensitivity to the backbone.

  3. There are some papers claiming that different surrogate derivative shapes can lead to similar final classification performance in SNNs, after appropriate hyperparameter tuning. Would like to know how the authors justify their claim with respect to these papers [5].

  4. No comparison on larger or smaller numbers of timesteps with different even-based datasets [4]. Does this increase complexity, since the algorithm depends on the number of channels, but is it also affected by the number of timesteps?

  5. There is no forgetting, latency or even energy analysis given in the paper.

  6. There is a small calibration set that's stored; does the quality of the set make any difference to the learning? At some point, there could be noisy data points/features stored in this set. Does an increase or decrease in the size of this set make a difference to the result?


References

[1] Ni, Wenyao, et al. "Alade-snn: Adaptive logit alignment in dynamically expandable spiking neural networks for class incremental learning." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 18. 2025

[2] Han, Bing, et al. "Enhancing efficient continual learning with dynamic structure development of spiking neural networks." In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI '23). Article 334, 2993–3001. https://doi.org/10.24963/ijcai.2023/334

[3] Shen, Jiangrong, et al. "Robust selective activation with randomized temporal k-winner-take-all in spiking neural networks for continual learning." International Conference on Learning Representations. Vol. 2026. 2026.

[4] Qian, Yuanbin, et al. "Ucf-crime-dvs: A novel event-based dataset for video anomaly detection with spiking neural networks." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 6. 2025.

[5] Zenke, Friedemann, and Tim P. Vogels. "The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks." Neural computation 33.4 (2021): 899-925.


Questions for authors
  1. Add forgetting, memory, energy, latency and wall-clock analysis.

  2. Is this method scalable with respect to different backbones and large event-based datasets [4]? Also, does this method extend to large foundation models too and/or multimodal SNNs?

  3. Does an increase or decrease in the size of this calibration set make a difference to the result?

  4. Is the method affected by the choice of gamma? Since 0.5 has no justification and doesn't make sense at times, the model might need to give more weight to the new weight matrix than to the previous one. Would like to understand how the model changes wrt the gamma value; is it sensitive to it?

  5. Exponential moving average might forget or give very minimal importance to the very initial task matrices when the task length becomes too large. The author presents the following experiments with a maximum task length of 20. CIFAR-100: 20 tasks (5 classes per task). ImageNet-100: 20 tasks (5 classes per task) Tiny-ImageNet: 20 tasks (10 classes per task). Can the method scale up to larger task counts and smaller task counts ?

  6. It appears that the method requires individual channel interventions during forward passes, and this might create huge overhead in cases of larger networks. Because the number of required interventions grows linearly with the network's channel count, it would be great to know how this is handled or whether it is even scalable to larger networks.

These are the major questions. For the ones not mentioned here, please refer to the weakness section.

Reviewer_u75hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the sensitivity of surrogate-gradient-based importance estimation in continual learning with SNNs and proposes Functional Prox. It measures channel importance through changes in predictive distributions when channel spike outputs are suppressed. These scores guide layer-wise sparse protection and proximal updates. The method is evaluated against importance-based baselines across 15 settings on CIFAR-100, Tiny-ImageNet, and ImageNet-100.

Critical strengths

Consistent performance improvements across multiple settings. The method achieves higher final and average accuracy than the evaluated baselines, including EWC-DR, in 14 of 15 settings.

Critical weaknesses
  1. The causes and practical consequences of surrogate-dependent importance estimates are insufficiently established. Figure 1(b,c) shows low rank correlations for some surrogate pairs but high consistency for others. These differences alone do not establish inaccurate importance estimates or inappropriate parameter protection. Their effects on forgetting and new-task learning need to be demonstrated. The exact formulas, parameters, and normalization procedures should be specified for ZIF in the main experiments and Tri, FS, and Sig in Figure 1. Quantitative analysis should explain how each surrogate acts on the same membrane potentials at a fixed checkpoint. Distributions of (u-V_{\mathrm{th}}) and surrogate-gradient values, together with gradient norms and importance estimates at representative layers and timesteps, could distinguish the effects of effective gradient range and scale. The claimed problems should also be examined under appropriately tuned surrogate settings, including for EWC-DR. Table 5 supports the utility of functional scores and proximal updates, but does not isolate whether removing surrogate dependence explains the improvement over EWC-DR. Comparing both importance scores under matched protection selection, normalization, and update rules would clarify the source of improvement.

  2. Evidence for the contribution beyond prior work and the performance–cost advantage is insufficient. AGS-CL already combines importance estimation from forward activations with proximal optimization. Li et al. measure component importance using the JS divergence between output distributions before and after component removal. The paper should therefore clarify the additional value of combining these principles for SNN continual learning. A controlled comparison with simpler activation-based importance estimates is needed to establish the benefit of intervention-based scoring. Although the shared SNN backbone and training schedule are specified, baseline-specific importance implementations and hyperparameter selection procedures are insufficiently documented. The number of runs and standard deviations should also be reported for the main comparisons and ablations. Some improvements are small, making their reliability difficult to assess. Table 4 reports a 4.1% overhead for channel-level intervention, but only compares intervention granularities. Training costs are not compared with EWC-DR, EWC, or MAS. Total training time and peak memory, including importance estimation and parameter updates, should be compared under matched conditions.

  • Jung et al. Continual Learning with Node-Importance based Adaptive Group Sparse Regularization. NeurIPS, 2020.
  • Li et al. How Does Attention Work in Vision Transformers? A Visual Analytics Attempt. IEEE TVCG, 2023.
Questions for authors
  1. What are the exact formulas and parameters for ZIF and the surrogates in Figure 1? Does EWC-DR use the same surrogate for training and importance estimation?

  2. Are there existing analyses relating the importance differences in Figure 1 to surrogate range and scale relative to the membrane-potential distribution? Please provide any available results connecting these differences to forgetting or new-task accuracy.

  3. How were regularization coefficients and other key hyperparameters selected for each baseline? Please specify the number of runs, class orders, and variability across runs.

  4. What reference runtime and hardware were used to obtain the 4.1% overhead in Table 4? Does it include both importance estimation and proximal updates? Please provide any available comparisons of total training time and peak memory with the baselines under the same conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.