acceptodds
Withdrawn from ICLR 2027

Selective Activation via Neuronal Dynamics for Continual Learning in Spiking Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual learning with Spiking Neural Networks (SNNs) offers a promising solution for low-power, rehearsal-free learning on resource-constrained devices. Brain-inspired selective activation, arising from neural competition and lateral inhibition, is crucial for learning sparse, task-specific representations. However, existing -Winner-Take-All (-WTA) SNNs allocate neurons statically, leaving many neurons inactive while task-relevant neurons remain vulnerable to subsequent interference. We propose STAR, a Spike-driven Task-Adaptive Resource allocation framework that manages the neuron lifecycle under a fixed capacity. During task training, Generate-and-Test (GnT) uses maturity-gated spike utility to replace low-utility neurons with mutated clones of high-utility neurons, thereby renewing plasticity. At task boundaries, Sleep Consolidation protects statistically selected experts and reinitializes redundant neurons to preserve knowledge and release capacity. Together, they form a generate–select–consolidate–release cycle without network expansion or data replay. Experiments on Split-MNIST, CIFAR-100, and TinyImageNet show that STAR consistently improves accuracy, memory retention, and neuron utilization. On Split-MNIST, STAR improves average accuracy by percentage points and increases neuron utilization from to , while reducing forgetting across all three benchmarks. Ablation and neurodynamic analyses further validate the complementarity of GnT and Sleep Consolidation in addressing the stability–plasticity dilemma.

Primary area:
applications to neuroscience & cognitive science
Keywords:
spiking neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pnJaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STAR (Spike-driven Task-Adaptive Resource allocation), a rehearsal-free method for continual learning in spiking neural networks. Built on the k-WTA SNN architecture, STAR introduces two mechanisms: Generate-and-Test and Sleep Consolidation. These mechanisms function by replacing low-utility neurons with mutated clones of high-utility neurons, and by preserving statistically selected expert neurons while resetting redundant ones, respectively. Experiments on Split-MNIST, CIFAR-100, and TinyImageNet report improvements in accuracy, memory retention, and neuron utilization.

Critical strengths

This paper proposes a well-motivated and clearly structured neuron-level resource management framework. It addresses two key issues in SNN continual learning: reclaiming under-utilized neurons and preserving important neurons. The paper also provides neurodynamic analyses to examine the behavior of the proposed method.

Critical weaknesses
  1. The paper uses cumulative spike counts as a measure of neuronal utility. Based on this criterion, high-utility neurons are treated as experts and frozen, whereas low-utility neurons are considered redundant and replaced. However, the paper lacks targeted ablation studies or direct analyses demonstrating whether spike utility truly reflects the functional importance of individual neurons.
  2. The experiments are not sufficient to fully establish STAR's competitiveness against recent SNN continual learning methods. Comparisons with several recent and closely related approaches are missing. Furthermore, on TinyImageNet, the reported results primarily compare STAR with the baseline, without a comprehensive comparison against other methods.
Questions for authors
  1. The paper states that the CIFAR-100 experiments use a frozen non-spiking CNN feature extractor, with continual learning performed on the resulting embeddings. What data were used to train this encoder? Did all comparison methods use exactly the same frozen encoder?
  2. The number of experts in the proposed method increases continuously. Does the method remain applicable for longer sequences of tasks?
  3. Although the paper claims the method is resource-efficient, it lacks empirical metrics to substantiate these claims.
  4. What is the rationale for the expert selection threshold, and how sensitive is the method to this choice?
  5. How is "neuron utilization" defined in the experimental results?
Reviewer_MyAuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper aims to address the limitations of existing k-Winner-Take-All (k-WTA) methods for enabling continual learning (CL) in the spiking neural network (SNN) domain by statically allocating neurons without rehearsal. Specifically, it proposes a Spike-driven Task-Adaptive Resource (STAR) allocation framework that manages the neuron lifecycle under a fixed capacity. It employs a Generate-and-Test (GnT) process to replace mature, low-utility neurons with perturbed copies of high-utility neurons, and a consolidation process to freeze the input weights of selected experts and reinitialize the remaining plastic neurons. The proposed idea is interesting, but incremental to the baseline method. Moreover, the experimental results for the proposed STAR framework do not always provide promising performance. For instance, the average accuracy of STAR on the Split CIFAR-100 dataset shown in Table 1 is too low. Therefore, the work still seems to be in the development stage, and it is not mature enough for publication.

Critical strengths

• Its idea of replacing dead/silent neurons with Gaussian-perturbed clones of expert neurons is interesting.

Critical weaknesses

• This work seems to be an incremental enhancement of the baseline method (i.e., k-WTA + EWC from Shen et al.@AAAI’24), hence offering limited novel contributions.

• The paper considers only one main comparison partner (i.e., k-WTA + EWC from Shen et al.@AAAI’24), which limits the evaluation coverage, considering that many CL methods for SNNs exist in the literature.

• The experimental results of the proposed STAR framework do not always provide promising performance. For instance, the average accuracy on the Split CIFAR-100 dataset shown in Table 1 is still low.

• The overhead costs (e.g., latency, memory, and power/energy consumption) for executing the proposed STAR framework are missing.

• The literature review seems to be limited, since it only considers the baseline method as the comparison partner and does not account for the recent SNN-based CL survey paper.

Questions for authors

• It is suggested to clarify whether the STAR framework targets task-incremental learning (TIL) or class-incremental learning (CIL), because the overall discussion in the paper considers TIL, but Figure 1 shows CIL.

• Can the authors clarify the experimental setup and how the datasets are split into tasks?

• Can the authors clarify why the STAR framework does not always provide high accuracy, especially on the Split CIFAR-100 dataset?

• Can the authors clarify why the STAR framework achieves lower performance on the Split CIFAR-100 dataset than on the Split TinyImageNet, which has more classes?

• It is suggested to present the overhead costs in terms of latency, memory, and power/energy consumption required for executing the proposed STAR framework.

• It is suggested to consider more recent state-of-the-art NCL methods as comparison partners. These recent methods can be found in the following NCL survey paper: "Continual Learning With Neuromorphic Computing: Foundations, Methods, and Emerging Applications," in IEEE Access, vol. 13, pp. 124824-124873, 2025, doi: 10.1109/ACCESS.2025.3588665.

• It is recommended to increase the quality (resolution) and font size of Figure 1.

Reviewer_1mDZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STAR proposes a dynamic neuron lifecycle management framework for SNN-based continual learning under fixed capacity. Two mechanisms operate at different timescales: Generate-and-Test (GnT) replaces low-utility neurons with mutated clones of high-utility neurons during task training, while Sleep Consolidation freezes selected expert neurons and reinitializes redundant ones at task boundaries. The combined cycle is evaluated on Split-MNIST, CIFAR-100, and TinyImageNet, reporting improved accuracy, reduced forgetting, and higher neuron utilization without data replay or network expansion.

Critical strengths

The generate-select-consolidate-release cycle is a clean conceptual framework. Separating intra-task plasticity renewal from inter-task knowledge protection is architecturally sensible. The two mechanisms address different aspects of the stability-plasticity dilemma at different timescales, and the ablation reportedly confirms their complementarity. This decomposition is more principled than applying a single regularization or replay strategy uniformly.

Critical weaknesses

The biological framing does substantial rhetorical work that the actual operations do not support. "Generate-and-Test" is cloning with Gaussian perturbation. "Sleep Consolidation" is weight freezing plus random reinitialization. Neither operation bears a meaningful mechanistic resemblance to neurogenesis or sleep-dependent consolidation in the biological sense. Neurogenesis involves integration of new neurons into existing circuits with specific connectivity constraints; sleep consolidation involves hippocampal replay and systems-level transfer. The paper invokes these phenomena for motivation but the actual algorithm does not implement any of their key features. This is not a fatal flaw, but the authors should either develop the biological analogy more faithfully or drop it and present the method on its own engineering merits. As written, the framing risks misleading readers about what the method actually does.

The evaluation benchmarks are insufficient to support the claims. Split-MNIST is a near-solved problem in continual learning; a 6% improvement there is not informative about whether the method scales to realistic settings. CIFAR-100 and TinyImageNet are more appropriate, but the paper needs to compare against strong continual learning baselines adapted for SNNs, not just other k-WTA methods. Specifically: how does STAR compare against EWC, LwF, or PackNet applied to the same SNN architecture? Without these comparisons, it is unclear whether the gains come from the neuron lifecycle management or simply from having a better training recipe. The fixed-capacity constraint is interesting, but the paper should also report what happens with a larger static network of equivalent total parameter count, to establish that dynamic allocation is genuinely superior to static over-provisioning.

Questions for authors

What is the precise definition of "maturity" in the maturity-gated spike utility? How is the gating threshold determined, and how sensitive is performance to this hyperparameter?

How does STAR compare against a static SNN with equivalent total neuron count but no lifecycle management? This would isolate the contribution of dynamic allocation from the contribution of having more neurons.

The Gaussian mutation in GnT introduces stochasticity. How many mutation attempts are needed before a replacement is accepted? Is there a rejection rate, and does it vary across tasks in the sequence?

For Sleep Consolidation, what statistical criterion selects "experts"? Is this a fixed percentile, a significance test, or something else? How sensitive is performance to this selection threshold?

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

Related papers

Discussion (0)

Sign in to comment.