acceptodds
Withdrawn from ICLR 2027

Probe and Calibrate: Stochastic Firing-Boundary Adaptation for Spiking Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Distribution shifts can perturb the membrane responses of Spiking Neural Networks (SNNs), misaligning source-trained firing boundaries with target-domain activity. Threshold calibration provides a compact adaptation interface, but prediction confidence alone does not reveal how responsive neuronal firing is to boundary changes. We propose ProCal, a probe-and-calibrate framework for online test-time adaptation that uses this local responsiveness to guide threshold updates. During adaptation, each adapted layer receives a stochastic threshold with a learnable location and scale. The induced firing probability yields an analytic responsiveness score: the normalized conditional variance of the binary firing response. ProCal aggregates this score by sample and combines it with prediction confidence to weight entropy minimization. It updates only two scalars per adapted layer—38 on MS-ResNet-20—while freezing synaptic weights and requiring no adaptation-specific source retraining. Each target batch receives one stochastic update followed by prediction with the learned locations as noise-free thresholds. Across two event-corruption suites and two controlled frame-based variants, ProCal achieves the highest mean accuracy among the evaluated methods in three settings. It reaches 53.82% on CIFAR10dvs-C and 64.72% on DVS128Gesture-C; component ablations further show gains of 1.02 and 1.42 percentage points over deterministic threshold adaptation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
SNN, TTA

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_Y5sJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors propose ProCal, a probe-and-calibrate framework for online test-time adaptation that uses this local responsiveness to guide threshold updates. It reaches 53.82% on CIFAR10dvs-C and 64.72% on DVS128Gesture-C; component ablations further show gains of 1.02 and 1.42 percentage points over deterministic threshold adaptation.

Critical strengths
  1. The adaptation interface is well chosen for SNNs. ProCal treats layer-wise firing thresholds as both calibration variables and test-time probes, keeps synaptic weights frozen, and initializes the thresholds from the source model.
  2. The author investigated a very interesting issue in the field of spiking neural networks.
Critical weaknesses
  1. The empirical scope is still narrow for a CCF-A claim. All results use MS-ResNet-20, two neuromorphic corruption suites, and two controlled frame-based variants. There is no evidence on other SNN backbones, real event streams, hardware latency/energy, or larger-scale event datasets.
  2. In the main tables, adaptation is reset at known corruption boundaries, with the source checkpoint restored and corruption identity used to define the reset. That is closer to independent per-corruption TTA than online TTA under unknown, nonstationary shifts.
  3. The ImageNet dataset in Table 1 is significantly inferior to the existing methods, and this requires a reasonable explanation.
  4. It would be even better if the author could provide the actual computational cost.
Questions for authors

Please see Weaknesses!

Reviewer_XABYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes ProCal for test-time adaptation of SNNs under distribution shifts. It perturbs layer-wise firing thresholds with Gaussian noise and combines firing responsiveness with prediction confidence to weight entropy minimization, updating only each layer’s threshold location and perturbation scale. The authors conducted distribution-shift robustness evaluations on DVS-Gesture-C, CIFAR10dvs-C, CIFAR10-C-T, and ImageNet1K-C-T.

Critical strengths
  1. This paper proposes a boundary responsiveness score that uses forward propagation with a random threshold to determine how sensitive each sample is to threshold perturbations.
  2. The proposed method introduces fewer additional parameters.
Critical weaknesses
  1. The boundary responsiveness score assumes that the membrane potential remains constant and that threshold perturbations only affect whether or not a neuron fires. However, this assumption does not reflect the actual temporal evolution of SNNs.
  2. The boundary responsiveness score only measures a neuron's local flip probability in response to a threshold perturbation. A high responsiveness score does not imply that the sample provides gradient signals that are beneficial to the task. The sample neuron may be extremely sensitive to the threshold, but the sample itself may be an outlier or noisy. Therefore, the proposed method is a heuristic combination rather than a rigorously theoretically guaranteed approach.
  3. TM is the closest SNN-TTA baseline to this work and is also adaptive during threshold testing. Therefore, it is unreasonable for the authors to omit it from the comparison entirely, as this prevents an assessment of the gains achieved by the proposed method. The authors should reproduce source-domain pretraining with MPBN, including TM in the comparison under a fully unified protocol. Furthermore, they should conduct ablation studies by porting TM’s core algorithm to the model described in this paper and removing MPBN while retaining only the threshold update during testing.
  4. To fully validate the source of the performance improvement, the authors should include experiments in which sample weights are adaptively increased based on a deterministic threshold. This should be done using only confidence levels or manually designed weights, without introducing random probing.
  5. Current experimental results suggest that the proposed method provides only a minimal performance enhancement, with improvements ranging from 0.3% to 1.4%. In some cases, its performance is on par with or worse than SAR. The authors should analyze the reasons behind this failure.
  6. The proposed method requires tuning multiple hyperparameters, increasing its complexity. Additionally, the absence of a sensitivity analysis for these hyperparameters prevents verification of their impact on the method's performance.
  7. The proposed method requires two complete SNN temporal forward passes and one backpropagation for each batch. However, the paper only mentions that the method has 38 scalar parameters and fails to emphasize the computational cost of the double forward passes during inference. Additionally, during the probe phase, Gaussian noise is sampled at every simulation time step for each layer. It is difficult for traditional neuromorphic chips to perform independent Gaussian random sampling for each layer. The adaptive phase requires full backpropagation and surrogate gradients, but most low-power neuromorphic hardware does not support online backpropagation, which hinders the practicality of the proposed method. These factors further hinder the practicality of the proposed method.
Questions for authors

See Weakness.

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

This paper proposes ProCal, a test-time adaptation (TTA) method for spiking neural networks (SNNs) that calibrates firing thresholds using unlabeled data under distribution shift. The key idea is to introduce per-layer stochastic thresholds during a single stochastic rollout to analytically compute local firing probabilities and a boundary-responsiveness score (normalized Bernoulli variance), then combine this score with output confidence to weight entropy minimization. ProCal updates only two scalars per adapted spiking layer (location μ and scale σ), removes noise for prediction, and shows consistent gains over strong TTA baselines on three of four benchmarks, as well as improvements over deterministic threshold adaptation with comprehensive ablations.

Critical strengths

1. Well-motivated and parameter-efficient adaptation.

The paper introduces stochastic firing-boundary probing and uses the resulting responsiveness score together with prediction confidence to guide test-time adaptation. This is a reasonable way to incorporate SNN-specific dynamics into sample weighting. The method also updates only two parameters per spiking layer while keeping the original network weights frozen.

2. Solid experimental validation.

The method is evaluated on four corruption benchmarks and achieves the best average accuracy in three of them. The component ablations clearly show the contributions of stochastic probing, learned threshold scales, and sample weighting. Additional firing-statistic analyses and recurring-domain experiments further support the effectiveness of the approach.

Critical weaknesses

1. Limited analysis of the sample weighting strategy.

The proposed responsiveness score is computed from a single stochastic rollout, and the sample weights are not normalized. As a result, it is unclear how much of the improvement comes from selecting informative samples versus changes in gradient magnitude. A more detailed analysis would help clarify the contribution of the weighting strategy.

2. Limited evaluation of computational efficiency.

Although ProCal updates only 38 parameters, each adaptation step requires two forward passes and one backward pass. The paper does not report runtime, memory usage, or energy consumption compared with existing TTA methods. This makes it difficult to assess the practical efficiency of the approach, particularly for resource-constrained deployment.

Questions for authors
  1. Sample weighting: Have the authors tested normalizing the sample weights (e.g., to have unit mean)? Since the current weights also affect the overall gradient magnitude, this comparison would help clarify whether the improvement mainly comes from emphasizing informative samples or changing the effective update size.

  2. Computational efficiency: Could the authors provide runtime and memory comparisons with existing TTA methods? Although ProCal updates very few parameters, each batch still requires two forward passes and one backward pass. Understanding this overhead would help assess its practical benefits.

  3. Robustness under different shifts: ProCal performs worse than the source model on some corruptions, such as DropArea, and trails SAR on ImageNet1K-C-T. Could the authors explain these cases and whether the firing-boundary responsiveness score behaves differently under such shifts?

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

Related papers

Discussion (0)

Sign in to comment.