acceptodds
Under review as a conference paper at ICLR 2027

A Spiking Neural Network Learning Rule Integrating Multi-scale Eligibility Traces and Steady-state Adaptation

Abstract

Spiking neural networks (SNNs) hold substantial promise for temporal modelling and neuromorphic computing owing to their event-driven computation and binary spike communication. Existing training methods, however, still face difficult temporal credit assignment, rigid surrogate gradients, and limited adaptive capacity. Inspired by spike-timing-dependent plasticity and meta-plasticity, we propose multi-scale eligibility traces with meta-plasticity and steady-state adaptation (MEMSA), a spatiotemporally local learning rule. MEMSA combines multi-scale eligibility traces for multi-resolution temporal dependencies, an activity-dependent meta-plastic surrogate gradient, and steady-state threshold adaptation that stabilizes firing. A binary-matrix local learning signal removes global error backpropagation. MEMSA preserves linear memory complexity O(Ln) and computational complexity O(LCn) up to constant factors. On CIFAR-100, IBM DVS Gesture, and CIFAR10-DVS, MEMSA improves over recent local rules and matches standard Backpropagation Through Time on CIFAR10-DVS within overlapping standard deviations (76.64% versus 76.40%). These results show that fully local learning can attain parity with global backpropagation on neuromorphic event-driven benchmarks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spiking Neural Network, Local Learning Rule, Multi-scale Eligibility Trace, Adaptive Mechanism

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xRY6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MEMSA, a temporally and spatially local learning rule for spiking neural networks built upon TESS. The method combines multi-scale eligibility traces, an activity-dependent surrogate-gradient window, and homeostatic threshold adaptation. Layer-wise learning signals are generated using fixed binary projections, avoiding cross-layer error backpropagation. Experiments with VGG-9 on three datasets show an improvement over TESS on CIFAR10-DVS, a small improvement on DVS Gesture, and unchanged mean accuracy on CIFAR-100.

Critical strengths
  1. The paper addresses an important problem: improving SNN training while avoiding the storage and communication requirements of full spatiotemporal backpropagation.

  2. The three components have clear motivations and operate at complementary stages of learning: retaining temporal information, adjusting surrogate-gradient support, and regulating neuronal activity.

Critical weaknesses
  1. The contribution beyond TESS is primarily incremental. Full spatiotemporal locality and the stated asymptotic scaling are inherited from the underlying framework. The added components combine multiple exponential traces, activity-dependent surrogate-gradient widths, and homeostatic threshold control. The paper does not sufficiently establish why this particular combination constitutes a substantial methodological advance. Comparisons with tuned single-scale traces, fixed surrogate-gradient widths, and simpler threshold adaptation would help distinguish the benefits of adaptivity from those of additional tuning.

  2. The empirical gains are concentrated on one dataset. Relative to TESS, MEMSA improves CIFAR10-DVS by 1.64 percentage points, but increases DVS Gesture accuracy by only 0.18 points and leaves CIFAR-100 mean accuracy unchanged. All experiments use VGG-9, limiting evidence for generality across architectures. Moreover, sequences contain only 6–20 time steps, so the claimed benefit of multi-scale temporal dependencies would be more convincingly demonstrated through experiments varying sequence length or requiring longer temporal credit assignment.

  3. The complexity analysis does not clearly account for the complete training procedure. Equation (19) counts local projection operations, but the forward synaptic computation and weight-update operations in Equation (17) also incur costs. Likewise, the stated memory complexity describes neuronal states and traces rather than total training memory, which includes weights and Adam optimizer states. These quantities should be distinguished explicitly. Measured runtime and peak memory, including the overhead of the three additional mechanisms, are needed to substantiate the practical efficiency claims.

  4. Several implementation details necessary for reproduction are missing. Section 2.3 does not fully specify the conflict-aware blending coefficient, the reference energy, or the normalization procedure. The non-causal time constants and several surrogate-gradient and threshold-controller parameters are also insufficiently documented. For STA, the definition of the long-term error, its interaction with the short-term offset, and the annealing/freezing schedule need clarification. Complete pseudocode and a consolidated hyperparameter table would substantially improve reproducibility.

Questions for authors
  1. How does MET compare with a tuned single-scale trace under the same hyperparameter-search budget? Does its advantage increase with sequence length?

  2. What are the results for the three pairwise component combinations? Is the full configuration consistently better across seeds?

  3. Can the authors provide a complete update algorithm, including trace normalization, conflict-aware blending, threshold adaptation, and the timing of optimizer updates?

  4. What are the measured training time and peak memory of MEMSA, TESS, and BPTT under identical settings? Which operations and stored quantities are included in the complexity comparison?

3 comments
AuthorsOpenReview ↗

Authors' Response

Thank you very much for taking the time to review and for your support We try our best to address your questions as follows.

Q1:How does MET compare with a tuned single-scale trace under the same hyperparameter-search budget? Does its advantage increase with sequence length?

A1: Our single-scale baseline uses the optimal configuration reported by the TESS authors, not an arbitrarily weakened one. Adding only MET to MEMSA improves the average accuracy on CIFAR10-DVS from 75.00% to 75.93% (+0.93 percentage points).

MEMSA's main contribution is not to claim that multi-scale traces universally dominate, but to jointly address three distinct issues. MET provides multi-scale historical evidence, MSG modulates the activity-dependent surrogate gradient window, and STA regulates neuronal excitability.

Regarding MET's advantage, it preserves activity history across different decay rates and forms the temporal information required for the current update through controlled fusion. When both recent and earlier activities are informative for classification, it is better positioned to capture both than a fixed single scale. However, a longer sequence does not guarantee a greater advantage. If added time steps are redundant or irrelevant, longer traces may introduce interference; the effective scale must cover the historical range truly required by the task.

Q2: What are the results for the three pairwise component combinations? Is the full configuration consistently better across seeds?

A2: MET, MSG, and STA target temporal information preservation, gradient-sensitive regions, and neuronal activity regulation, respectively, providing a functional complementarity rationale for our design. However, we agree that functional complementarity alone does not guarantee higher accuracy when combined, nor does it substitute for a comprehensive combinatorial ablation. Currently, we lack fully comparable results under identical settings and random seeds for all pairwise and full configurations. Therefore, the existing evidence is insufficient to support the claim that "the complete MEMSA outperforms all combinations across every random seed," and we do not regard this as a necessary condition for the method's validity.

Mechanistically, the three components interact: STA alters subsequent firing activity, which in turn affects MSG's activity statistics and MET's input history, while MSG modulates how eligibility traces participate in weight updates. Consequently, the combined effect is not necessarily a simple summation of individual module gains. In future work, we will complete the full ablation studies to verify these interactions among the three modules.

Q3: Can the authors provide a complete update algorithm, including trace normalization, conflict-aware blending, threshold adaptation, and the timing of optimizer updates?

A3: Regarding the four key mechanisms you are concerned about, their specific implementations in the algorithm are as follows: (1)Trace normalization: The algorithm respectively calculates the exponentially decaying accumulations of the pre-synaptic trace and the post-synaptic trace, and completes the normalization and extraction of trace features through a non-linear transformation function. (2)Conflict-aware blending: The gradient is composed of the contributions of the pre- and post-traces (weighted by their respective coefficients), and interacts with the learning signal through element-wise multiplication, thereby achieving the perception and balancing of potential conflicts. (3)Threshold adaptation: Involves lines 7 and 14. Line 7 calculates the dynamic threshold as the maximum value between the sum of the base threshold and the threshold offset (constrained to a minimum of 0.5); Line 14 utilizes a specific function to update the threshold offset and the average firing rate based on the current output. (4)Timing of optimizer updates: As shown in lines 3 and 19-20. The gradient is accumulated step-by-step throughout the entire time sequence. Only after this temporal loop ends (i.e., when the sequence processing is complete), is a single Adam optimizer step executed at line 20.

Q4: What are the measured training time and peak memory of MEMSA, TESS, and BPTT under identical settings? Which operations and stored quantities are included in the complexity comparison?

A4: Experimental results show that TESS achieves a throughput of 126.42 samples/s, an average training time of 71.19 seconds/epoch, and a peak allocated memory of 2060.14 MiB; MEMSA achieves a throughput of 89.35 samples/s, an average training time of 100.72 seconds/epoch, and a peak allocated memory of 2478.45 MiB; BPTT has the highest throughput of 273.99 samples/s and the lowest average training time of 32.85 seconds/epoch, but its peak allocated memory is the highest at 7610.18 MiB.

Reviewer_xRY6OpenReview ↗

Review

Thank you for the detailed response and additional experiments.

Table 2 shows that MEMSA’s largest gain over TESS is on CIFAR10-DVS (+1.64 percentage points), with a small gain on DVS Gesture (+0.18 points) and unchanged mean accuracy on CIFAR-100. The response confirms that matched pairwise ablations are still missing, so it remains unclear whether all three mechanisms are needed. The MET comparison uses TESS’s reported configuration, leaving its advantage over a single-scale trace tuned with the same search budget, and across different sequence lengths, unresolved. The new measurements clarify the efficiency trade-off: MEMSA uses less peak GPU memory than BPTT, but takes about 41% longer per epoch and uses about 20% more peak memory than TESS. For reproducibility, please also specify how the conflict-aware blending coefficient and reference energy are computed. I maintain my original rating.

Q1: Table 2 shows that MEMSA’s largest gain over TESS is on CIFAR10-DVS (+1.64 percentage points), with a small gain on DVS Gesture (+0.18 points) and unchanged mean accuracy on CIFAR-100. The response confirms that matched pairwise ablations are still missing, so it remains unclear whether all three mechanisms are needed. The MET comparison uses TESS’s reported configuration, leaving its advantage over a single-scale trace tuned with the same search budget, and across different sequence lengths, unresolved. The new measurements clarify the efficiency trade-off: MEMSA uses less peak GPU memory than BPTT, but takes about 41% longer per epoch and uses about 20% more peak memory than TESS. For reproducibility, please also specify how the conflict-aware blending coefficient and reference energy are computed. I maintain my original rating. A1: Thank you for identifying the remaining evidence and reproducibility gaps. We agree that matched ablations and equal-budget comparisons are important. At the same time, we would like to distinguish the benefits supported by the current results from the broader claims requiring further validation.

Contribution and existing evidence. MEMSA coordinates three distinct aspects of the learning update: MET preserves activity histories across decay ranges, MSG adjusts gradient sensitivity according to neuronal activity, and STA regulates subsequent firing through feedback. Importantly, MET does not simply replace the reference trace with a multi-scale mixture. Magnitude matching and deviation-aware residual fusion constrain the influence of additional temporal information. This controlled integration is a concrete design feature, rather than merely an increase in tunable parameters.

The clearest accuracy benefit is the 1.64-point improvement on CIFAR10-DVS. DVS Gesture and CIFAR-100 provide evidence of comparable baseline performance, not significant improvements. Accordingly, our claim is task-dependent improvement while retaining competitive performance on the other evaluated tasks—not uniform superiority.

Commitment to additional comparisons. We will complete matched pairwise and full-configuration ablations across multiple seeds, using consistent backbones, preprocessing, training budgets, and evaluation protocols. We will report means and variability rather than only the best runs. These experiments will assess both beneficial interactions and combinations whose effects may offset one another. We do not assume that the full configuration must rank first for every seed; the objective is to establish the empirical justification for combining all three components.

We will also compare MET with single-scale traces tuned under the same search budget and examine different sequence lengths. TESS’s reported configuration is a direct baseline, but does not eliminate the possibility of an advantage from additional tuning. MET is intended to retain useful activity across heterogeneous temporal ranges, not to guarantee larger gains whenever sequences become longer. The additional comparisons will test this specific hypothesis.

Efficiency trade-off. The current measurements show approximately 67% lower peak allocated GPU memory than BPTT, but approximately 20% higher peak memory and 41% longer training time per epoch than TESS. We therefore emphasize retaining a memory advantage over BPTT while introducing richer temporal and activity regulation, not universal acceleration. The asymptotic properties are inherited from TESS, and the practical overhead of the additional mechanisms must be stated explicitly.

Reproducible fusion definitions. Let AA denote the central-index reference trace, ee the driving activity, and MM the Gaussian-weighted, area-corrected mixture. Define a(Z)a(Z) as the mean absolute magnitude of ZZ. The reference magnitude is

L=max(a(A),0.25a(e),10−5).L = max(a(A), 0.25a(e), 10^{-5}).

The matching gain and matched mixture are

g=clip(L/max(a(M),10−5),0.75,1.25),M′=gM.g = clip(L/max(a(M),10^{-5}), 0.75, 1.25), \qquad M' = gM.

The deviation and effective blending coefficient are

C=min(a(M′−A)/L,3),beff=b(0.5+0.5/(1+C)).C = min(a(M'-A)/L,3), \qquad b_{eff} = b(0.5+0.5/(1+C)).

The fused trace is A+beff(M′−A)A+b_{eff}(M'-A). Thus, “reference energy” means mean absolute magnitude, not squared energy. “Conflict” means trace-magnitude deviation, not gradient-direction conflict. Larger deviations reduce the multi-scale correction. Taken together, the current results support a task-specific accuracy benefit and a measured memory advantage over BPTT. The additional comparisons will assess whether these benefits persist under matched tuning budgets and whether the three components provide complementary gains. We will align the final claims with those outcomes.

Reviewer_X7MgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MEMSA, a SNN learning rule that is spatiotemporally local. MEMSA combines bio-inspired plasticity mechanisms with local supervision. It extends TESS, the previous baseline method, with multi-scale eligibility traces, activity-dependent surrogate gradients, and adaptive firing thresholds. Experiments on CIFAR-100, CIFAR10-DVS, and DVS Gesture show competitive accuracy, with improvements over TESS mainly on CIFAR10-DVS.

Critical strengths
  1. Exploring biologically inspired plasticity for effective local SNN learning is meaningful.
  2. The proposed components address complementary aspects of temporal memory, learning sensitivity, and firing-rate regulation.
Critical weaknesses
  1. Insufficient evaluation. Evaluation is limited to VGG-9 and three benchmarks with relatively small numbers of time steps (TT).
  2. Incomplete reproducibility details. No source code or supplementary information is provided.
  3. Although theoretical complexity analysis is provided, there are no actual measurements of training memory and time cost.
  4. Limited discussion about related work. The connections to other local/online learning rules and adaptive-plasticity methods need further discussion to clarify MEMSA’s novelty and contribution. Also, OTTT and DECOLLE appear in Table 3 without corresponding references.
  5. Improvement over TESS on DVSGesture is marginal.
Questions for authors
  1. (Weakness 1) Please evaluate generalization to another backbone or a task with rich temporal dependencies.
  2. (Weakness 2) Would you publish the source code to make the work more reproducible?
  3. (Weakness 3) Please report peak memory and training throughput against TESS and BPTT.
  4. (Weakness 4) Please expand the discussion of related local/online learning and adaptive-plasticity methods, clarify their relationship to MEMSA, and provide references for the compared baselines.
  5. Why use fixed Gaussian trace mixing? Comparisons with uniform mixing and a tuned single-scale baseline would help justify this choice.
  6. Inconsistency: the peak of the surrogate function in Equation (13) is fixed at 0.3, whereas Figure 3 shows varying amplitudes.
2 comments
AuthorsOpenReview ↗

Authors' Response

Thank you very much for taking the time to review and for your support We try our best to address your questions as follows.

Q1: Please evaluate generalization to another backbone or a task with rich temporal dependencies.

A1: Network Architecture: We employ only VGG-9 to isolate learning rule effects and ensure a controlled comparison with TESS. While this does not replace cross-architecture validation, MEMSA's mechanisms are backbone-agnostic, supporting transferability to other architectures.

Design Objective: MEMSA aims to coordinate fixed traces, surrogate gradient windows, and firing thresholds rather than guarantee uniform accuracy gains. Specifically, MET preserves multi-scale temporal information, MSG adapts the gradient window to current activity, and STA modulates the threshold for activity homeostasis. Its main contribution is the coordinated regulation of these learning processes.

Time Scale: Multi-scale traces involve varying decay constants—not extended sequences—allowing distinct weighting of recent versus earlier activity within a finite window. This justifies MET's use over 6–20 steps. However, these experiments only demonstrate limited-window integration; they cannot fully prove long-term credit assignment or scalability with sequence length.

Q2: Would you publish the source code to make the work more reproducible?

A2: We will subsequently make the source code used in this paper publicly available on GitHub. This will include the three modules of MEMSA, the training entry point, and the runtime configurations for the main comparative experiments. Additionally, we will provide environment dependencies, data preprocessing instructions, random seeds, and result summarization scripts.

Q3: Please report peak memory and training throughput against TESS and BPTT.

A3: Experimental results show that TESS achieves a throughput of 126.42 samples/s, an average training time of 71.19 seconds/epoch, and a peak allocated memory of 2060.14 MiB; MEMSA achieves a throughput of 89.35 samples/s, an average training time of 100.72 seconds/epoch, and a peak allocated memory of 2478.45 MiB; BPTT has the highest throughput of 273.99 samples/s and the lowest average training time of 32.85 seconds/epoch, but its peak allocated memory is the highest at 7610.18 MiB.

Q4: Please expand the discussion of related local/online learning and adaptive-plasticity methods, clarify their relationship to MEMSA, and provide references for the compared baselines.

A4: TESS is our direct baseline; MEMSA retains its learning signal but modifies three aspects of the synaptic update: MET integrates eligibility traces with varying decay rates to capture both recent and earlier spike information; MSG dynamically modulates the surrogate gradient window based on recent firing rates, allowing channels with identical distances to receive different gradients; and STA adapts the threshold according to firing-rate deviations, thereby influencing subsequent spikes and gradient computation. Together, MET determines which history to retain, MSG regulates its impact on updates, and STA alters the conditions for subsequent spikes.

Unlike DECOLLE (layer-wise losses) or Dspike (gradient shape adjustment), MEMSA's novelty lies in the synaptic update itself—specifically, MSG directly modulating the window via firing rates in conjunction with forward threshold feedback. Compared to OTTT (instantaneous gradients) and S-TLLR (STDP-inspired traces), MEMSA further introduces multi-scale traces and incorporates firing activity into surrogate gradient modulation.

Q5: Why use fixed Gaussian trace mixing? Comparisons with uniform mixing and a tuned single-scale baseline would help justify this choice.

A5: We adopt fixed Gaussian weights to assign higher initial weights to the intermediate scale among the three time-ordered eligibility traces, while retaining information from both shorter and longer scales. This approach also avoids introducing extra trainable parameters for fusion. We acknowledge that current ablations cannot yet isolate whether the gains stem from the multi-scale traces themselves or from the Gaussian weighting. We plan to prioritize comparing Gaussian versus uniform mixing on CIFAR10-DVS, and evaluate a tuned single-scale baseline under the same validation budget.

Q6: Inconsistency: the peak of the surrogate function in Equation (13) is fixed at 0.3, whereas Figure 3 shows varying amplitudes.

A6: The blue line depicts TESS’s fixed surrogate gradient window (peak = 0.3). MSG improves this by adaptively modulating the window width based on firing rate. Under area normalization, the peak varies with the width, yielding the distinct peaks observed across the adaptive curves.

Thank you for your effort and detailed response. My questions on related works and design choices have been addressed. However, there is no experiment results supporting the generalizability of the proposed method to other network architectures. Also, compared to TESS, the baseline method, MEMSA is >1.38×>1.38\times slower. Considering these aspects, I will keep my original rating.

Reviewer_RC6YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a local learning rule for SNNs called MEMSA, which is an extension of TESS. It combines multi-scale eligibility traces for representing temporal information at different decay rates, a meta-plastic spike gradient (MSG) that adapts the surrogate-gradient window according to firing activity, and steady-state threshold adaptation (STA) for regulating neuronal firing. The method retains layer-local learning signals based on fixed binary projections to avoid cross-layer error backpropagation. Experiments on small-scale tasks show performance comparable to BPTT.

Critical strengths
  1. The topic is a motivated research direction for resource-constrained SNN learning.
  2. The local learning results are comparable to the reported BPTT results. This is difficult to achieve.
Critical weaknesses
  1. The complementary contributions of the three adaptive mechanisms could be better established. The full MEMSA model achieves only a modest additional gain over the strongest individual component, STA. The current experiments do not fully clarify how the three mechanisms interact or whether their combination provides consistent benefits across different temporal learning conditions. Additional analysis of the interactions among MET, MSG, and STA would strengthen the experimental justification for the complete framework.

  2. The implementation details (including the hyperparameters and the trace fusion procedure) and practical benefits of the adaptive mechanisms require further clarification.

  3. Empirical measurements of memory usage and training overhead relative to TESS should be provided.

Questions for authors

How does MET compare with a properly tuned single-scale trace?

1 comment
AuthorsOpenReview ↗

Authors' Response

Thank you very much for taking the time to review and for your support We try our best to address your questions as follows.

Q1: How does MET compare with a properly tuned single-scale trace?

A1: Our single-scale baseline adopts the optimal configuration reported by the original authors of TESS, rather than an arbitrarily chosen weaker configuration. With only MET introduced in MEMSA, the average accuracy on the CIFAR10-DVS dataset increases from 75.00% to 75.93%, a difference of 0.93 percentage points. Furthermore, the main contribution of MEMSA is not to claim that multi-scale traces independently dominate in all tasks, but rather to jointly address three different issues: MET provides historical evidence at multiple time scales, MSG modulates the activity-dependent surrogate gradient window, and STA regulates neuronal excitability. MET, MSG, and STA respectively address "what historical information to retain," "when this information can effectively participate in learning," and "what activity state the neuron maintains":

(1) MET retains input evidence across different time ranges through multi-scale eligibility traces. (2) MSG adjusts the surrogate gradient window according to the firing rate, allowing this evidence to be converted into weight updates under appropriate membrane potential states. (3) STA adjusts the firing threshold, affecting the subsequent firing rate and the distance from the membrane potential to the threshold, which in turn feeds back to MSG and the eligibility traces of the next layer.

Therefore, the three form a complementary relationship of "temporal information preservation, plasticity regulation, and activity homeostasis control." MET and MSG provide historical eligibility traces and activity-dependent surrogate gradients, respectively, which multiply in the weight update term to jointly determine whether historical information can effectively participate in learning. MSG and STA share firing activity feedback; the former modulates the plasticity window, while the latter changes the threshold distance and subsequent firing rate, so STA further influences the regulation of MSG. MET and STA interact indirectly through the cycle of "weight update, subsequent firing, and threshold adjustment," and the output spikes altered by STA in turn become the input for MET in the next layer. The three thus constitute a feedback closed-loop of "temporal evidence preservation, plasticity expression, and activity homeostasis regulation."

Experimental results show that TESS achieves a throughput of 126.42 samples/s, an average training time of 71.19 seconds/epoch, and a peak allocated memory of 2060.14 MiB; MEMSA achieves a throughput of 89.35 samples/s, an average training time of 100.72 seconds/epoch, and a peak allocated memory of 2478.45 MiB; BPTT has the highest throughput of 273.99 samples/s and the lowest average training time of 32.85 seconds/epoch, but its peak allocated memory is the highest at 7610.18 MiB.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.