acceptodds
Under review as a conference paper at ICLR 2027

Pair-Aware Self-Contrastive Learning in Spiking Neural Networks

Abstract

Self-Contrastive Forward-Forward (SCFF) learning enables layer-local representation learning from unlabeled data by constructing positive and negative pairs. However, conventional activity-based goodness evaluates the firing strength of a fused pair representation without explicitly assessing the correspondence between its constituent responses. This creates a mismatch between relation-based pair construction and activity-based goodness evaluation in spiking neural networks (SNNs). To address this limitation, we propose Pair-aware SCFF, which incorporates cross-branch spike correspondence into the local goodness function. Specifically, we evaluate observed spike co-activation relative to marginal firing activity and combine the resulting relational signal with evidence from activity-supported pairs. This formulation jointly accounts for neural response strength and cross-branch correspondence while preserving layer-wise local optimization without class-label supervision or network-wide error propagation. Experiments across six static-image and two event-based datasets demonstrate consistent improvements over a matched Spiking SCFF baseline, with accuracy gains of up to 10.71 percentage points. Controlled ablations distinguish the contributions of relation-aware goodness and pair-stream propagation, while pair-level analyses reveal more structured pair relations and improved class-level semantic organization in the learned representations. These results highlight the importance of explicitly evaluating cross-branch correspondence in local self-contrastive SNN learning.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking neural network, Non-backpropagation, Forward-Forward, Unsupervised learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gxWwOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The submission addresses layer-local, label-free representation learning in SNNs. Its point of departure is Self-Contrastive Forward-Forward (SCFF) learning, in which a positive pair comprises two views of the same sample and a negative pair comprises two different samples, and in which the local goodness is evaluated on a fused representation of the pair. The authors identify a mismatch between how such pairs are constructed and how they are evaluated: pairs are defined by the relation between two observations, whereas the goodness function quantifies only the magnitude of the fused response, so that pairs exhibiting distinct patterns of cross-branch spike agreement may nonetheless receive comparable scores. The proposes Pair-aware SCFF retains the two branch spike responses separately and evaluates a pair by combining two quantities. A stop-gradient, minibatch-level prevalence reweighting additionally suppresses channels that fire for most inputs, and a self-pairing Pair Stream propagation preserves the correspondence between the two members across depth while maintaining layer-local optimization, since propagated representations are detached. Validation consists of a controlled comparison against a matched Spiking SCFF implementation that shares the backbone, neuron model, support-view construction, optimizer, per-layer training budget, frozen linear evaluator, and seeds, conducted across six static-image and two event-based datasets with three seeds, which reports gains of up to 10.71 percentage points on CIFAR-100. The paper further reports component ablations, alternative relation formulations, threshold-robustness studies, and pair-level diagnostics demonstrating that improved layer-wise frozen-probe accuracy does not require improved separability of the raw fused pair score. Comparisons against published unsupervised and self-supervised SNN methods are also reported and are explicitly identified as cross-protocol. The authors release training code comprising both matched trainers, the component-ablation experiments, and the analysis switches underlying the diagnostic figures.

Critical strengths
  1. The paper isolates a real and previously unaddressed mismatch, and the comparison measures what it claims to measure. Positive and negative examples in self-contrastive forward-forward learning are defined by a relationship, yet the goodness that ranks them is a property of the fused response magnitude alone. Everything except the propagated stream and the goodness function is held fixed across the comparison: same backbone, neuron model, support-view construction, optimizer, per-layer training budget, frozen linear evaluator, and seeds.
  2. The paper establishes a non-obvious decoupling between a widely used diagnostic and actual representation quality. Section 4.3 and Appendix C show that improved layer-wise frozen linear-probe accuracy does not require a corresponding increase in same/different-class AUROC of the raw fused pair score, with a supporting argument for why the two quantities need not be ordered (Equations 29-36) and a controlled sweep over explicit separation pressure (Appendix C.4).
Critical weaknesses
  1. The reported improvement is established against a single, self-implemented baseline, and its magnitude is not anchored to the strongest available reference point. All eight headline comparisons use a matched Spiking SCFF implementation constructed by the authors, which holds backbone, neuron model, support-view construction, optimizer, training budget, evaluator, and seeds fixed. That control is methodologically appropriate, and the concern here is not with the fidelity of the comparison but with the breadth of the reference set: the other forward-forward SNN methods cited in the paper (FF-SCP, Terres-Escudero et al., 2025; FF-SNN, Ghader et al., 2025) are not used as baselines, and no comparison against an independently reported SCFF adaptation is given. The consequence is that the improvement is established relative to one particular reference configuration rather than against the current best alternative, which limits how much weight the headline gains can carry. This is addressable by reporting the additional baselines, or by stating explicitly why the chosen reference is the appropriate one.
  2. The contribution is narrower than the title and abstract indicate. Appendix B.1 shows that the specific functional form of the relation statistic is close to interchangeable, with centered Dice, no prevalence reweighting, and normalized covariance all falling within roughly one percentage point of the full formulation, while the component ablation shows the gain factor is essential only in interaction with the evidence term. The evidence therefore supports a conjunctive evidence and relation criterion rather than the particular relational formulation that the paper advertises as its contribution. This is a matter of claim scoping rather than technical error. The mechanism the ablations identify is broader than the specific statistic proposed.
Questions for authors
  1. Appendix B.1 shows the specific relation statistic is close to interchangeable, while the component ablation shows the gain factor matters only in interaction with the carrier. Should the contribution be characterized as a conjunctive evidence and relation criterion rather than as this particular relational statistic? A decomposition that holds the evidence term E_t fixed while varying only the relation term would make the claim precise.
  2. The fused-activity baseline corresponds to the SCFF formulation as stated. Is it also the strongest available SCFF configuration on this backbone? Specifically, was any tuning of the fused pathway explored before it was fixed, (for example, pre-fusion summation versus per-branch thresholds, or the normalization of the goodness)? And does an independently reported spiking forward-forward method provide a comparable same-task reference point? I note that the two other forward-forward SNN works cited in the paper do not serve this purpose directly: the method of Terres-Escudero et al. (2025) addresses out-of-distribution detection, and the work of Ghader et al. (2025) reports layer-wise forward-forward optimization, so my question concerns whether a stronger same-task reference exists rather than a request for a particular additional baseline.
4 comments

1. Broadening the Local Learning Comparisons

Thank you for your suggestion regarding the scope of the comparisons. The matched Spiking SCFF control is used to assess the effect of the proposed design; the additional experiments introduce different local learning rules to broaden the comparisons under the same backbone.

We added adapted versions of CSDP and FF-SNN, using a common three-layer CNN, 4 time steps, Fashion-MNIST, and frozen linear evaluation. The results for label-free representation learning are:

MethodWithout the Pairing Protocol (%)With the Pairing Protocol (%)
CSDP (adapted)88.10±0.2188.10\pm0.2189.26±0.0589.26\pm0.05
FF-SNN (adapted)89.48±0.0589.48\pm0.0590.35±0.2690.35\pm0.26
Pair-aware SCFF—91.15±0.1491.15\pm0.14

Pair-aware SCFF achieves the highest accuracy in this common setting. The pairing protocol also provides consistent improvements for CSDP and FF-SNN, indicating that its benefits can extend to different local learning rules.

The gains for both CSDP and FF-SNN show that pairing transfers across local learning rules, while Pair-aware SCFF achieves the highest accuracy among the methods evaluated under this unified local-SNN protocol.

2. Core Contribution: The Joint Role of Paired Evidence and Relational Information

We agree with the reviewer's assessment of the mechanism underlying the contribution. The core contribution of this work is to jointly incorporate pair-supported evidence and cross-branch relations into local goodness, enabling the training objective to use the relational information contained in the pair construction.

With the Pair Stream and evidence term EE held fixed, replacing only the relation term yields the following results:

Relation FormPrevalence ReweightingTest Accuracy (%)
Full Pair RelationYes75.25±0.4575.25\pm0.45
Without prevalence reweightingNo74.75±0.2074.75\pm0.20
Centered DiceYes75.04±0.2075.04\pm0.20
Normalized covarianceYes73.16±0.2173.16\pm0.21

All relation forms exceed the 69.18% accuracy obtained using pair-supported evidence EE alone. The full form achieves the highest mean accuracy, while Centered Dice and the variant without prevalence reweighting also maintain strong performance.

These results support a more general design principle: combining paired evidence with relational information is the main source of the gains, and the proposed prevalence-weighted statistic is an effective implementation of this principle. The gains retained across different relation forms also indicate some flexibility in how this principle can be implemented.

3. Fusion Path and Threshold Configuration

Our baseline retains the original SCFF fusion path, and we provide additional comparisons addressing both goodness design and threshold selection.

Wth the Pair Stream held fixed, the four goodness configurations on CIFAR-10 are:

Evidence FormWithout Relation ModulationWith Relation Modulation
Fused Activity FF68.62%62.86%
Pair-Supported Evidence EE69.18%75.25%

The combination of pair-supported evidence and relation modulation achieves the highest accuracy, improving upon EE alone by 6.07 percentage points. This comparison holds the paired propagation protocol fixed and directly tests the design of the local objective.

The 13 positive and negative threshold settings on Fashion-MNIST show that Pair-aware SCFF leads in all matched settings. Using each method's best thresholds within the tested grid yields accuracies of 89.98±0.1789.98\pm0.17\\% and 91.46±0.1391.46\pm0.13\\%, respectively, retaining a gap of 1.48 percentage points. The 13-setting threshold grid was fixed in advance; test results are reported for each setting and were not used to select thresholds. This tuning retains SCFF's original fusion structure and focuses on goodness and threshold selection; other fusion structures were not included in this search. With the additional CSDP and FF-SNN controls, the available results support the effectiveness of paired evidence and relation modeling from three perspectives: local objectives, threshold configuration, and learning rules.

To further examine your question about threshold calibration, we evaluated the same pre-specified grid of 13 positive/negative threshold-scale pairs, (α+,α−)(\alpha_+,\alpha_-), independently for Spiking SCFF and Pair-aware SCFF on CIFAR-10 and CIFAR-100. This gives 52 independently trained method–dataset–setting runs, all with seed 730. We report the final test accuracy for every setting below.

(α+,α−)(\alpha_+,\alpha_-)C10 SCFFC10 Pair-awareΔ\DeltaC100 SCFFC100 Pair-awareΔ\Delta
(0.50, 1.00)66.8471.13+4.2937.9645.54+7.58
(0.75, 0.75)68.2074.89+6.6941.1350.05+8.92
(0.75, 1.00)70.2075.00+4.8040.5449.85+9.31
(0.75, 1.25)68.5773.55+4.9840.8448.02+7.18
(1.00, 0.50)65.5273.74+8.2237.0548.76+11.71
(1.00, 0.75)66.0474.72+8.6837.6949.45+11.76
(1.00, 1.00)67.0074.89+7.8939.6950.03+10.34
(1.00, 1.25)67.0175.36+8.3539.4950.07+10.58
(1.00, 1.50)67.0375.44+8.4139.7349.66+9.93
(1.25, 0.75)64.7573.34+8.5936.2947.68+11.39
(1.25, 1.00)66.1473.94+7.8037.9249.46+11.54
(1.25, 1.25)65.9375.08+9.1539.3349.39+10.06
(1.50, 1.00)63.6273.62+10.0035.6347.83+12.20

For seed 730, Pair-aware SCFF improves over Spiking SCFF at all 13 matched settings on both datasets. The mean paired gain across the evaluated threshold grid is 7.53 percentage points on CIFAR-10 and 10.19 points on CIFAR-100. The highest observed test accuracies are 70.20% for Spiking SCFF and 75.44% for Pair-aware SCFF on CIFAR-10, and 41.13% and 50.07%, respectively, on CIFAR-100. These maxima are descriptive summaries of the sweep. Comparing each method at its own best observed thresholds preserves gains of 5.24 and 8.94 percentage points.

With the earlier Fashion-MNIST experiments, these results directly address threshold calibration on two of the larger-gain datasets you identified. Spiking SCFF benefits from threshold adjustment, but the substantial advantage of Pair-aware SCFF remains after that adjustment and across every evaluated setting.

Thank you again for your feedback on baseline coverage and the scope of the contribution. We added comparisons with adapted CSDP and FF-SNN under a shared backbone and Fashion-MNIST evaluation protocol, where Pair-aware SCFF achieves the highest accuracy among the evaluated label-free local SNN methods. The pairing protocol also improves both adapted CSDP and FF-SNN, supporting its applicability beyond SCFF.

We further provided the ablation you requested, holding the Pair Stream and evidence term fixed while varying only the relation term, and clarified that the central contribution is the joint use of paired evidence and relational information. We also extended the threshold-sensitivity analysis to Fashion-MNIST, CIFAR-10, and CIFAR-100. Pair-aware SCFF outperformed Spiking SCFF in every evaluated matched setting.

As the discussion period draws to a close, we would appreciate your feedback on whether these additions address your concerns about baseline coverage and contribution scope, and your consideration of this evidence in your final assessment. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

The rebuttal resolves my main concerns. And I will keep my rating. Thank you for your time.

Reviewer_xe4AOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a modification to Self-Contrastive Forward-Forward (SCFF) learning for spiking neural networks (SNNs). In particular, they argue that existing SCFF scores a pair by the firing rate of a single neuron population driven by both inputs’ summed currents. This cannot directly assess whether the two inputs’ responses correspond. They propose to also run each input through its own neurons and measure how much the two branches co-fire beyond what their individual firing rates predict. This results in a new relation-aware goodness function, which they pair with propagating each input separately across layers. They validate their method by comparing it against a spiking reimplementation of SCFF on six static-image and two event-based datasets using frozen linear probes, along with component ablations, alternative relation formulas, and post hoc analyses of the learned pair scores. Results show consistent gains over the baseline on all eight datasets, similar gains across alternative relation formulas, and improved layer-wise probe accuracy.

Critical strengths

(S1) The paper is well motivated. In SCFF, pairs are defined by the relationship between two inputs, but goodness is computed only after their responses are fused. This leaves the objective with no direct way to assess whether the members’ responses correspond. The proposed fix follows naturally, adds no learnable parameters, and gives consistent gains over the baseline across all datasets.

(S2) To me a point that is particularly promising is that the gain does not rely on one given hand-picked formula. Table 3 swaps in alternative correspondence measures and finds similar performance, especially with Centered Dice for example. After removing prevalence reweighting the performance only reduced by 0.5%, and even the weakest alternative, normalized covariance, remains well above the baseline. I think this supports the broader claim that keeping cross-branch correspondence in the objective is essential.

Critical weaknesses

(W1) It’s unclear whether the SCFF baseline is fairly tuned. Both methods share the same positive and negative thresholds. However, the two goodness functions have different scales, hence I don’t think it’s appropriate to use shared thresholds here. In particular, in Fig. 5, the SCFF baseline sits at near-zero or negative normalized operating margins across the tested threshold scales, while Pair-aware SCFF sits clearly positive. While the authors test threshold scaling with shared threshold, this does not establish that the thresholds are optimally calibrated for the individual methods. I would like to see method-specific threshold tuning, since this is a critical comparison point.

Questions for authors

(Q1) What are the training-time and memory overheads relative to Spiking SCFF? The method keeps three neuron-state sets per layer and uses a batch-level prevalence statistic. Given the on-chip local learning setting that motivates the work, I think it would be good to add a discussion on this.

(Q2) The central motivation is that fused firing rate fails to capture cross-branch correspondence. While Fig. 4(a) illustrates this qualitatively, the claimed distinction is not directly quantified empirically. Can the authors show, on a trained Spiking SCFF model, that FF and RR are meaningfully decoupled?

4 comments

1. Method-Specific Threshold Tuning

Thank you for pointing out the effect of differences in goodness scale on threshold selection. We tuned the positive and negative thresholds separately on Fashion-MNIST and evaluated 13 settings. The thresholds yielding the highest accuracy within the tested grid differ between the two methods:

MethodAccuracy with Default Thresholds (%)Best Thresholds in the Grid (α+,α−)(\alpha_+,\alpha_-)Highest Accuracy in the Grid (%)
Spiking SCFF89.88±0.0989.88\pm0.09(1.25,1.25)(1.25,1.25)89.98±0.1789.98\pm0.17
Pair-aware SCFF91.15±0.1491.15\pm0.14(1.00,0.75)(1.00,0.75)91.46±0.1391.46\pm0.13

Pair-aware SCFF achieves higher accuracy in all 13 matched settings, with gains ranging from 0.98–1.61 percentage points. When comparing the best result for each method within the grid, Pair-aware SCFF retains an advantage of 1.48 percentage points.

We report test accuracy to evaluate the pre-specified threshold grid. Pair-aware SCFF is higher in all 13 tested matched settings. This result shows that, within the evaluated threshold range, the performance improvement persists and does not depend on an advantage arising from shared default thresholds.

2. Training Resources and Learning Performance Within Limited Time

We added measurements of actual training memory and performance under a training time budget. Both methods have 5.649M parameters. With identical hardware, batch size, and data loading settings, Pair-aware SCFF has a peak allocated training GPU memory of 4216 MiB, compared with 4616 MiB for Spiking SCFF, a reduction of approximately 8.7%.

To further assess the representation quality achievable within limited training time, we use a common time cap of 10.36 minutes and select each method's latest evaluated result within the budget:

MethodTest Accuracy Within the Budget (%)Corresponding Cumulative Training Time (minutes)
Spiking SCFF67.45±0.0867.45\pm0.089.63±1.189.63\pm1.18
Pair-aware SCFF75.25±0.3175.25\pm0.316.97±0.056.97\pm0.05
SSL-BPTT53.74±1.1353.74\pm1.138.01±0.068.01\pm0.06

Under the common 10.36-minute cap, Pair-aware SCFF reaches 75.25±0.3175.25\pm0.31\\% accuracy in 6.97±0.056.97\pm0.05 minutes, compared with 67.45±0.0867.45\pm0.08\\% for Spiking SCFF in 9.63±1.189.63\pm1.18 minutes and 53.74±1.1353.74\pm1.13\\% for SSL-BPTT in 8.01±0.068.01\pm0.06 minutes. Thus, at the evaluated operating points, Pair-aware SCFF achieves higher accuracy with less elapsed training time than either comparator. This result demonstrates the ability of the relation-aware local objective to learn high-quality representations under limited training resources.

For on-chip implementations, layer-local updates allow the backward computational graph of the current layer to be released promptly after its update, reducing the need for simultaneous residency of cross-layer states. These storage characteristics and measured resource results support the implementation of our method for memory-constrained training.

3. Quantifying Fused Goodness and Branch Relations in Trained Models

Using layer-1 of the trained models, we compute the Spearman correlation between fused goodness and the branch-relation correction term over 2,048 sample pairs. The relation correction term is defined as

\quad R=\operatorname{mean}_c(\tilde z_c^{(1)}\tilde z_c^{(2)}), \quad B=\operatorname{mean}_c(\tilde z_c^{(1)})\operatorname{mean}_c(\tilde z_c^{(2)}),$$ where $\tilde z_c^{(i)}=(1-p_c)z_c^{(i)}$ is the branch response weighted by its estimated firing prevalence $p_c$. We compute fused goodness $G$ and relation correction $A$ for each pair, averaging both over timesteps and spatial locations. We report Spearman correlation across the 2,048 pair-level $(G,A)$ values. | Trained Model | Spearman Correlation Coefficient | |---|---:| | Spiking SCFF | 0.068 | | Pair-aware SCFF | 0.846 | In Spiking SCFF, fused goodness and the branch-relation correction term exhibit a weak monotonic association (Spearman correlation coefficient: **0.068**). This result provides empirical support for explicitly assessing cross-branch correspondence. The correlation coefficient reaches **0.846** in Pair-aware SCFF, consistent with its explicit integration of paired evidence and relational information into the local objective. This analysis extends the original qualitative motivation with quantitative evidence from trained models.

W1 (threshold tuning): I thank the authors for updating the tuning procedure, and I am now convinced that on Fashion-MNIST the improvement does not depend on shared default thresholds. However, I believe my concern applies to the baseline comparison as a whole. Ideally, the main comparison would use method-specific thresholds on all evaluated dataset. At minimum, it would help to see this for one of the datasets with the largest reported gains (CIFAR-10/100 or STL-10). Since this is not yet established in the current version, I maintain my score for now, though I’m happy to be corrected if I’ve misunderstood how the thresholds were chosen.

Q1 and Q2: I appreciate the new resource measurements and the correlation analysis. The latter directly supports the paper’s motivation, and both address my questions. Could the authors briefly explain why Pair-aware SCFF uses less memory and training time than Spiking SCFF, given the additional branch-wise states and per-member propagation?

Thank you for the follow-up and for recognizing that the correlation analysis supports the paper’s motivation. Following your suggestion, we extended the threshold evaluation to both CIFAR-10 and CIFAR-100, two datasets with substantial reported gains. We also updated the training-memory measurements after aligning implementation-level optimizations.

  1. Method-specific threshold evaluation on CIFAR-10/100

We evaluated 13 positive/negative threshold pairs separately for both methods on both datasets, giving 52 independently trained configurations with seed 730. The thresholds were scaled by (α+,α−)(\alpha_+,\alpha_-). “Best” denotes the highest observed accuracy within the predefined grid.

DatasetBest Spiking SCFF (%)Best Pair-aware SCFF (%)Best-to-best gain (pp)Mean gain across matched settings (pp)Settings with improvement
CIFAR-1070.2075.44+5.24+7.5313/13
CIFAR-10041.1350.07+8.94+10.1913/13

The best-performing thresholds differ between the methods. On CIFAR-10, Spiking SCFF performs best at (0.75,1.00)(0.75,1.00), whereas Pair-aware SCFF performs best at (1.00,1.50)(1.00,1.50). On CIFAR-100, the respective settings are (0.75,0.75)(0.75,0.75) and (1.00,1.25)(1.00,1.25).

These differences support your point that goodness-scale calibration matters. Nevertheless, comparing each method at its own best observed thresholds preserves gains of 5.24 and 8.94 percentage points.

Across the full scan, the lowest Pair-aware SCFF accuracy exceeds the highest Spiking SCFF accuracy: 71.13% versus 70.20% on CIFAR-10 and 45.54% versus 41.13% on CIFAR-100. Full scan below; Δ\Delta denotes the accuracy difference in percentage points.

(α+,α−)(\alpha_+,\alpha_-)C10 SCFFC10 Pair-awareΔ\DeltaC100 SCFFC100 Pair-awareΔ\Delta
(0.50, 1.00)66.8471.13+4.2937.9645.54+7.58
(0.75, 0.75)68.2074.89+6.6941.1350.05+8.92
(0.75, 1.00)70.2075.00+4.8040.5449.85+9.31
(0.75, 1.25)68.5773.55+4.9840.8448.02+7.18
(1.00, 0.50)65.5273.74+8.2237.0548.76+11.71
(1.00, 0.75)66.0474.72+8.6837.6949.45+11.76
(1.00, 1.00)67.0074.89+7.8939.6950.03+10.34
(1.00, 1.25)67.0175.36+8.3539.4950.07+10.58
(1.00, 1.50)67.0375.44+8.4139.7349.66+9.93
(1.25, 0.75)64.7573.34+8.5936.2947.68+11.39
(1.25, 1.00)66.1473.94+7.8037.9249.46+11.54
(1.25, 1.25)65.9375.08+9.1539.3349.39+10.06
(1.50, 1.00)63.6273.62+10.0035.6347.83+12.20

Together with the earlier Fashion-MNIST experiments, these results directly address threshold calibration on two of the larger-gain datasets you identified. Spiking SCFF benefits from threshold adjustment, but the substantial advantage of Pair-aware SCFF remains after that adjustment and across every evaluated setting.

  1. Training memory and training-time interpretation

Thank you for asking about the overhead associated with the additional branch states. In the initial implementation, Pair-aware SCFF already incorporated memory-management optimizations that avoid retaining unnecessary autograd graphs and intermediate states for computations that do not require backpropagation. To further align the resource comparison, we have now applied the same optimization principles to Spiking SCFF, while keeping each method’s learning objective, parameter-update rules, and experimental settings unchanged.

The updated measurements under this aligned setup are:

MetricSpiking SCFFPair-aware SCFF
Parameters (M)5.6495.649
Peak allocated GPU memory during training (MiB)4059.424216.10

Pair-aware SCFF requires 156.68 MiB of additional peak training memory, approximately 3.86%, without adding learnable parameters.

Regarding training time, the figures in our earlier response referred to cumulative training time at selected evaluated checkpoints within a common budget, using the earlier implementations. They described the representation quality attained at those checkpoints, rather than the time required to complete the full training schedule. This earlier attainment of a high-quality representation is particularly relevant to on-chip training, where SRAM capacity, energy, and online training time are all constrained.

The resource results and threshold experiments present a clear trade-off: Pair-aware SCFF introduces no additional parameters, incurs approximately 3.86% additional peak training memory in the aligned measurement, and retains substantial accuracy gains when each method is evaluated at its own best observed threshold setting. The new CIFAR-10/100 results directly address the remaining concern that the reported improvements might depend on shared default thresholds.

Thank you again for suggesting that we extend the threshold evaluation to datasets with larger reported gains. We have completed the additional experiments on both CIFAR-10 and CIFAR-100 and provided the full 13-setting results in our latest response.

Comparing each method at its own best observed setting within the evaluated grid, Pair-aware SCFF retains gains of 5.24 and 8.94 percentage points, respectively. It also outperforms Spiking SCFF at all 13 matched threshold settings on each dataset, further supporting that the gains persist after baseline threshold tuning.

As the discussion period draws to a close, we would appreciate your feedback on whether these results address the remaining concern about baseline threshold calibration, and your consideration of this additional evidence in your final assessment. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

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

This paper introduces Pair-aware Self-Contrastive Forward-Forward (Pair-aware SCFF) learning for spiking neural networks (SNNs). It addresses a key limitation of conventional activity-based goodness in self-contrastive FF learning: fusing paired representations before evaluation obscures cross-branch correspondence. The authors propose a relation-aware goodness function that explicitly models spike co-activation relative to marginal firing activity, combining this relational signal with pair-supported evidence. The method preserves layer-local optimization without class-label supervision or network-wide error propagation. Experiments on six static-image and two event-based datasets demonstrate consistent improvements over a matched Spiking SCFF baseline, with accuracy gains of up to 10.71 percentage points. Controlled ablations, robustness tests, and pair-level analyses further validate the contribution.

Critical strengths

The paper clearly identifies a mismatch between relation-based pair construction and activity-based goodness in SNNs, and proposes a novel, mathematically grounded relation-aware goodness that explicitly captures cross-branch spike correspondence. This is a meaningful step toward more effective local representation learning in SNNs.

Critical weaknesses
  1. Despite improvements over the local Spiking SCFF baseline, the performance of Pair-aware SCFF still trails significantly behind global backpropagation-based self-supervised methods (e.g., 75.25% vs. 85.60% on CIFAR-10 in Table 1). The paper does not adequately discuss this performance gap or provide concrete insights into how local learning could narrow it, which limits the practical impact and general significance of the work.

  2. The method requires maintaining separate branch-wise spike responses and additional computations (channel-wise products, prevalence reweighting, etc.), increasing memory and compute overhead. This partially offsets the energy-efficiency advantages typically associated with local SNN learning, yet the paper lacks a quantitative analysis of this trade-off (e.g., energy, latency, or parameter footprint).

Questions for authors
  1. Could you provide a direct, protocol-matched comparison with other local learning methods such as CSDP (Ororbia, 2024) to better situate Pair-aware SCFF within the local SNN learning landscape?

  2. How does the extra computational overhead of relation-aware goodness affect the energy efficiency and inference latency of the SNN? A brief analysis or measurement would strengthen the paper.

  3. The theoretical analysis in Appendix C explains the decoupling of probe quality and fused-pair AUROC, but it remains abstract. Could you give a more intuitive geometric interpretation of how relation-aware goodness reshapes the feature space?

3 comments

1. Comparison with Global BP: Training Computational Budget and Advantages for On-Chip Training

We thank the reviewer for this constructive feedback. Full-network BPTT computes temporally unrolled weight gradients and propagates input gradients across layers, whereas Pair-aware SCFF updates each layer using a local objective. We added an SSL-BPTT control using the same CIFAR-10 split, backbone, time steps, batch size, and frozen linear evaluation; it applies symmetric NT-Xent to the final-layer two-view representations. We group the principal structural operations performed during training into forward computation and update-related computation. For one minibatch, the structural-operation cost model is:

\begin{aligned} C_{\mathrm{train}}^{(m)} = C_{\mathrm{fwd}}^{(m)} + C_{\mathrm{update}}^{(m)}, \end{aligned}

\begin{aligned} C_{\mathrm{fwd}}^{(m)} = K\sum_{l=1}^{L}\sum_{c\in\mathcal{F}_{m,l}} B_c(d_l f_l+d_l), \end{aligned}

\begin{aligned} C_{\mathrm{update}}^{\mathrm{SCFF}} = C_{\mathrm{local}}^{(S)}+C_{\mathrm{goodness}}^{(S)} = K \left[ \sum_{l=1}^{L} \sum_ {c\in\mathcal{U}_ {S,l}} B_cd_l(f_l+1) +\sum_{l=1}^{L}\sum_ {c\in\mathcal{G}_{S,l}}B_cd_l \right] \end{aligned}

\begin{aligned} C_{\mathrm{update}}^{\mathrm{Ours}} =C_{\mathrm{local}}^{(P)}+C_{\mathrm{goodness}}^{(P)}+C_{\mathrm{relation}}^{(P)} = K\left[\sum_{l=1}^{L}\sum_ {c\in\mathcal{U}_ {P,l}}B_cd_l(f_l+1) +\sum_ {l=1}^{L}\sum_ {c\in\mathcal{G}_ {P,l}}B_cd_l +\sum_ {l=1}^{L}\sum_ {c\in\mathcal{R}_ {P,l}}B_cd_l \right], \end{aligned}

\begin{aligned} C_{\mathrm{update}}^{\mathrm{BP}} =C_{\mathrm{global\text{-}grad}}^{(B)} +C_{\mathrm{NTXent}}^{(B)} = K\left[ \sum_ {l=1}^{L}\sum_ {c\in\mathcal{W}_ {B,l}}B_cd_lf_l +\sum_ {l=1}^{L-1}\sum_ {c\in\mathcal{X}_ {B,l}}B_cd_lf_l \right] +3(2B)^2d_ {\mathrm{emb}} . \end{aligned}

Here, BB is the minibatch size, K=4K=4 is the number of time steps, BcB_c is the number of examples processed by call cc, dld_l is the number of output neurons across channels and spatial positions, and flf_l is the convolutional fan-in. The call sets F\mathcal{F}, U\mathcal{U}, G\mathcal{G}, R\mathcal{R}, W\mathcal{W}, and X\mathcal{X} denote the executed forward, local-update, goodness, relation, BPTT weight-gradient, and cross-layer input-gradient calls. The superscripts SS, PP, and BB denote Spiking SCFF, Pair-aware SCFF, and SSL-BPTT.

Update-related computation is categorized according to the actual computational graph: ClocalC_{\mathrm{local}} includes local weight-gradient and within-layer feedback operations; Cglobal-gradC_{\mathrm{global\text{-}grad}} includes weight gradients for all layers, cross-layer input gradients, and within-layer differentiation operations within the accounting scope. CgoodnessC_{\mathrm{goodness}}, CrelationC_{\mathrm{relation}}, and CNTXentC_{\mathrm{NTXent}} account for evaluating the corresponding objectives and computing their derivatives with respect to layer outputs. Shared intermediate results and backward computations are counted according to their actual execution frequency to avoid double counting.

We accumulate forward and update costs according to the layer-wise training schedule actually executed by each method, then normalize by the number of samples in the training set to obtain cumulative GOp/sample. Under a common computational budget cap, we select each method's latest evaluated model for each random seed. The results are as follows:

MethodCumulative Training Computation (GOp/sample)Test Accuracy Within the Budget (%)
Spiking SCFF49.42967.43±0.1167.43\pm0.11
Pair-aware SCFF49.45575.25±0.4575.25\pm0.45
SSL-BPTT51.56252.65±1.3052.65\pm1.30

At nearly matched counted cost, Pair-aware SCFF improves accuracy over Spiking SCFF by 7.82 percentage points; relative to SSL-BPTT, it uses 4.1% fewer counted operations and improves accuracy by 22.6 points, showing stronger accuracy under the same computational cap. Pair-aware SCFF releases each layer's backward graph after its local update, reducing the need to retain cross-layer activations and membrane states simultaneously. When active layer states fit in SRAM, this can reduce off-chip transfers and support memory-constrained on-chip training.

Thank you again for your interest in the practical value and performance gaps. We would like to further emphasize, beyond SCFF, pairing improves CSDP from 88.10±0.21%88.10\pm0.21\% to 89.26±0.05%89.26\pm0.05\% and FF-SNN from 89.48±0.05%89.48\pm0.05\% to 90.35±0.26%90.35\pm0.26\% under a unified protocol. It also outperforms matched Spiking SCFF on six static-image and two event datasets, with gains of up to 10.71 percentage points. On Fashion-MNIST, it reaches 91.15±0.14%91.15\pm0.14\%, the highest among the evaluated label-free local SNN methods under the unified CNN and frozen-readout protocol. These results show transfer across local learning rules and datasets; the common-budget results above show higher accuracy than both SCFF and SSL-BPTT at comparable counted computation.

2. A Concrete Direction for Narrowing the Local Learning Performance Gap

Global BP can coordinate all layers through the final objective, whereas local learning relies on each layer's own objective to develop useful representations. Our improvement addresses the design of these objectives: because paired samples are constructed from the relationship between them, local goodness should reflect both paired evidence and cross-branch correspondence.

With the Pair Stream held fixed, the CIFAR-10 ablation results are:

Evidence FormWithout Relation Modulation RRWith Relation Modulation RR
Fused Activity FF68.62%62.86%
Pair-Supported Evidence EE69.18%75.25%

Combining relation modulation with pair-supported evidence improves accuracy by 6.07 percentage points; applying the same modulation directly to fused activity provides no benefit. This indicates that the key is to apply relational information to an evidence term that preserves pair support, thereby producing a more effective local learning objective.

This design improves upon the matched Spiking SCFF baseline on all six static-image datasets and both event datasets, with gains of up to 10.71 percentage points. The additional ablations further attribute these gains to a specific mechanism, providing a concrete direction for improving local self-supervised learning.

3. Training Memory and Inference Overhead

Paired goodness and relation computations are used during representation training. Deployment uses a single-stream feature extraction path and requires no additional paired-relation computation. Both methods have the same 5.649M parameters. For a consistent comparison, we remeasured peak training GPU memory after applying the same gradient-management principles to computations that do not require backpropagation in both methods, while preserving their parameter-update rules.

The measurements were obtained on the same NVIDIA A100 PCIe 80 GB GPU using PyTorch 2.4.1+cu124, CUDA 12.4, and cuDNN 9.1.0. The training batch size was 100, and both data loaders used 4 workers with identical configurations. The updated training-memory results and previously reported inference measurements.

MetricSpiking SCFFPair-aware SCFF
Parameters (M)5.6495.649
Peak Allocated GPU Memory During Training (MiB)4059.424216.10
Batch-1 Inference Latency (ms)4.17 ± 0.613.87 ± 0.07
Inference Throughput (images/s)2818 ± 1632882 ± 173

Pair-aware SCFF achieves higher classification accuracy while maintaining the same parameter count and comparable inference latency and throughput, with 156.68 MiB (approximately 3.86%) of additional peak training memory.

4. Comparison with Other Local SNN Learning Methods

We added adapted versions of CSDP and FF-SNN, using a common three-layer CNN, 4 time steps, Fashion-MNIST, and frozen linear evaluation. The paired variants share the branch processing and pairing protocol while retaining each method's own learning rule and goodness definition.

The results for label-free representation learning are:

MethodWithout the Pairing Protocol (%)With the Pairing Protocol (%)
CSDP (adapted)88.10±0.2188.10\pm0.2189.26±0.0589.26\pm0.05
FF-SNN (adapted)89.48±0.0589.48\pm0.0590.35±0.2690.35\pm0.26
Pair-aware SCFF—91.15±0.1491.15\pm0.14

The pairing protocol improves CSDP and FF-SNN by 1.16 and 0.87 percentage points, respectively, showing that it can benefit different local learning rules. Pair-aware SCFF achieves the highest accuracy under this common protocol, exceeding FF-SNN + pair and CSDP + pair by 0.80 and 1.89 percentage points, respectively.

These additional results broaden the local learning comparisons, demonstrate the applicability of the pairing mechanism across different learning rules, and show the performance advantage of Pair-aware SCFF in this unified comparison.

5. A Geometric Interpretation of Representation Changes

We measure same-class neighborhood purity using cosine nearest neighbors in the original CIFAR-10 feature space:

Number of Neighbors kkSpiking SCFFPair-aware SCFF
129.40%40.93%
1023.82%33.67%
5020.24%26.22%

Pair-aware SCFF increases the proportion of same-class samples at all evaluated neighborhood scales. For the group of samples with the lowest baseline neighborhood purity, the improvement reaches 13.61 percentage points.

These results provide an intuitive geometric explanation: the relation-aware objective improves local semantic organization in the feature space, making same-class samples more prevalent in one another's neighborhoods, particularly in regions where classes were initially mixed. This change is consistent with the higher accuracy of the frozen linear readout and provides direct evidence of improved representation quality.

Thank you again for your constructive feedback. We would like to highlight the additional comparisons and analyses addressing your concerns about the practical value of our method.

At comparable analytically estimated cumulative training costs of approximately 49–52 GOp/sample, Pair-aware SCFF achieves 75.25% accuracy, compared with 67.43% for Spiking SCFF and 52.65% for the same-backbone SSL-BPTT control. We also added protocol-matched comparisons with adapted CSDP and FF-SNN, together with quantitative feature-neighborhood analyses to provide a geometric interpretation of the representation improvements.

As the discussion period draws to a close, we would appreciate your feedback on whether these additions address your concerns about the performance–resource trade-off and the significance of the contribution, and your consideration of this evidence in your final assessment. We would be happy to clarify any remaining questions.

Thank you 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.