acceptodds
Withdrawn from ICLR 2027

Event Actions Predict Loss Jumps at Spike Deletion Boundaries

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spiking neural networks combine continuous dynamics with discrete events triggered by threshold crossings. A finite parameter change can remove a spike, switching the network to a different event sequence and producing a loss change that is not represented by the gradient evaluated before the spike disappears. We show that this loss jump can be predicted from the EventProp backward pass while the spike is still present. Deleting a spike removes both its membrane reset and its downstream synaptic transmission. We sum the corresponding backward terms and call the result the event action. Across 120 deletions of single spikes in a learned LIF network trained on the Spiking Heidelberg Digits (SHD) dataset, the event action computed before deletion closely matches the signed loss jump. The result replicates in an independently trained network and on examples from the speaker held out from training. Replacing the target spike with other spikes reduces prediction accuracy, and spike time, boundary location, spike position, and source neuron activity do not explain the relation. The prediction magnitude also separates deletions with small and large effects: the highest predicted impact quintile has a median realized loss change about 803 times larger than the lowest. For 20 parameter directions that cross a deletion boundary, adding the predicted jump raises the rank correlation with the realized loss change from 0.281 to 0.997 and selects the candidate with the lowest loss. The backward pass contains information about the loss change caused by losing an event that the parameter gradient before deletion does not capture.

Primary area:
learning on time series and dynamical systems
Keywords:
spiking neural networks, gradient-based learning, EventProp

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

This paper proposes an event action to predict the loss jump caused by the deletion of a spike in spiking neural networks. The method uses EventProp adjoints to combine the contributions of the membrane reset and downstream synaptic transmission. The authors evaluate the prediction on 120 spike deletions in a trained LIF network and report a Pearson correlation of 0.99995 between the predicted and measured loss jumps. They also show that the prediction improves the ranking of finite parameter updates. The problem is relevant to gradient-based learning in SNNs, although the theoretical basis and practical benefit of the proposed method remain unclear.

Critical strengths
  • The paper addresses an important problem in SNN learning, where finite parameter updates can change the spike sequence and cause loss changes that are not captured by local gradients.

  • The proposed approach keeps the original loss and neuron dynamics unchanged, which makes it an interesting alternative to existing methods based on loss shaping or modified dynamics.

  • The reported prediction accuracy is very high. In addition, the authors provide controls, an independently trained network, and an evaluation on examples from a held-out speaker.

  • The finite-update experiment provides initial evidence that predicting the deletion jump can improve the comparison of candidate parameter updates.

Critical weaknesses
  • The novelty is not sufficiently established. The event action is constructed from existing EventProp adjoint terms, and the paper does not clearly explain what is theoretically new beyond combining these terms to predict a finite loss change. In particular, it remains unclear whether the relationship follows from hybrid-system adjoint theory under specific assumptions or is only an empirical observation.

  • The measured loss jump is computed at two points around the deletion boundary, but the paper does not clearly report the boundary tolerance or demonstrate numerical convergence. Therefore, it is unclear whether the measurement isolates the discontinuous loss change from the continuous loss variation near the boundary.

  • The selection criterion checks changes in the hidden-layer spike-count matrix, but it does not clearly establish that no other discrete event changes contribute to the measured loss jump.

  • The evaluation is limited to a shallow feedforward network. The replication experiments use only 20 deletions each, and the finite-update analysis uses 20 directions. Moreover, the statistical analysis does not clearly account for dependencies between deletions from the same network and examples.

  • The practical relevance remains limited. The paper shows improved ranking of finite parameter updates, but does not demonstrate that incorporating the event action into training improves convergence, loss, or classification accuracy.

Questions for authors

1- Can the authors derive the relationship between the event action and the limiting loss jump, and clarify under which assumptions it is exact or approximate?

2- What boundary tolerance is used, and do the results remain stable when the boundary search and numerical simulation tolerances are varied?

3- How do the authors ensure that the measured loss jump is caused only by the target spike deletion, rather than by additional event changes?

4- How does the proposed method differ theoretically and practically from existing EventProp adjoint-based analyses and methods addressing spike creation and deletion?

5- Does incorporating the predicted jump into an actual training procedure improve learning compared with standard EventProp and existing approaches?

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

This paper investigates loss discontinuities in spiking neural networks (SNNs) when parameter changes cause individual spikes to disappear. The authors introduce event action, defined as the sum of membrane-reset and synaptic-transmission terms from the existing EventProp backward pass, to predict the resulting loss jump before spike deletion occurs.

Experiments are conducted on a single-hidden-layer LIF network trained on the Spiking Heidelberg Digits (SHD) dataset. Across 120 single-spike deletion cases, the predicted and measured loss jumps achieve a Pearson correlation of 0.99995 and a median relative error of 3.92%. Additional experiments on an independently trained network and held-out-speaker examples report similar results. The authors further show that incorporating the predicted loss jump improves the ranking of 20 candidate parameter changes according to their realized loss.

The paper presents this approach as a way to estimate loss changes at spike-deletion boundaries that are not captured by conventional parameter gradients.

Critical strengths

The paper demonstrates a clear relationship between EventProp backward information and the finite loss jump caused by spike deletion. Across 120 cases, the prediction achieves a Pearson correlation of 0.99995 and a median relative error of 3.92%, without fitting coefficients to the deletion outcomes. Similar results are obtained with an independently trained network and held-out speaker data. These findings provide convincing evidence that the proposed predictor works under the studied conditions.

The authors compare the target spike's event action against alternative spike-based predictors, showing that the target event provides more accurate estimates. They also examine the reset and transmission contributions separately. In the finite-update experiment, incorporating the predicted loss jump improves the ranking correlation from 0.281 to 0.997 and identifies the lowest-loss candidate among 20 parameter directions. Although this does not establish benefits for actual training, it demonstrates a potential use of the prediction beyond simply measuring loss discontinuities.

Critical weaknesses

1. Unclear significance of the research problem.

The paper does not establish why single-spike deletion boundaries are an important limitation of SNN training. In Section 3.1, the authors deliberately scale neuron weights and use bisection to locate these boundaries. They do not examine how often such discontinuities affect normal training or whether they cause meaningful optimization problems. SNN computation depends on population-level spatiotemporal activity, yet the analysis focuses on isolated deletion events.

2. Limited methodological novelty.

The proposed event action (Eq. 5) combines two existing EventProp backward terms. The paper does not introduce a new learning rule or establish a broader theoretical result. Moreover, Table 4 shows that the transmission term alone performs slightly better than the full event action (3.60% versus 3.92% median relative error). The necessity of the proposed combination remains unclear.

3. Prediction accuracy does not demonstrate practical value.

The reported Pearson correlation of 0.99995 is impressive, but the optimization experiment only ranks 20 selected parameter directions that cross deletion boundaries. Table 5 shows that the baseline regret is just 1.02 × 10⁻⁵, while the predicted jump alone already selects the best candidate. No experiment demonstrates improved training convergence, accuracy, or stability under actual optimization.

4. Narrow experimental scope.

All experiments use SHD and a shallow feedforward network with 256 LIF neurons. The replications retain the same dataset and architecture. The study does not evaluate recurrent networks, spike creation, or simultaneous changes involving multiple spikes. It remains unclear whether the observed relationship holds under more complex network dynamics and realistic parameter updates.

5. Limited relevance to collective spiking dynamics. SNN computation depends on the spatiotemporal activity of neuronal populations rather than isolated spike events. Although the proposed event action accounts for downstream effects on the network loss, the experiments focus on carefully selected boundaries where a single spike disappears. In realistic training, parameter updates may simultaneously alter multiple spikes and their temporal relationships. The paper does not establish whether its single-event analysis extends to such interactions or whether the predicted loss jumps have a meaningful impact on network optimization. Consequently, the practical significance and generality of the findings remain unclear.

Questions for authors

Why does this narrowly defined spike-deletion prediction problem warrant a standalone research contribution? What evidence shows that predicting such loss jumps addresses a practically significant limitation of SNN training? How could the proposed event action be used in practical SNN training without explicitly searching for spike-deletion boundaries?

Beyond recombining existing EventProp adjoint terms, what is the fundamental methodological or theoretical novelty of the proposed event action?

Why is SHD the only benchmark considered? Does the proposed relationship hold across different datasets, network architectures, and more complex spiking dynamics, particularly when multiple spikes are affected by parameter updates?

Reviewer_HNn3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses the discontinuous loss jump at spike‑deletion boundaries in spiking neural networks (SNNs). It proposes the event action, defined as the sum of two EventProp adjoint terms corresponding to membrane reset and downstream synaptic transmission for an individual spike. Evaluated on the trajectory just before spike disappearance, this quantity is used to predict the finite loss jump induced by removing that spike. Experiments on a shallow feedforward LIF network trained on the SHD dataset show very high correlation between predicted and measured loss jumps. The proposed quantity also substantially improves rank consistency of candidate finite‑magnitude parameter‑update directions.

Critical strengths
  1. This paper identifies event‑sequence switching and associated loss discontinuities caused by spike creation or deletion. This issue constitutes an inherent blind spot for exact‑gradient methods such as EventProp, and represents a potential source of instability during SNN training.
  2. This work links per‑spike adjoint quantities from backward propagation to loss changes across different event trajectories. It offers a fresh perspective for analysing discrete boundary effects in spiking‑network dynamics.
Critical weaknesses
  1. The proposed quantity is purely a derivative application built upon the existing EventProp backward‑propagation framework. No novel backward‑propagation mechanism or revised training paradigm is introduced. The core operation simply sums two adjoint terms associated with a single spike.
  2. The approach only works for single‑spike‑deletion scenarios. It cannot handle cases with simultaneous multiple‑event modifications, nor can it address spike creation, which severely restricts its practical applicability.
  3. All empirical evaluations are conducted exclusively on shallow feedforward networks. There is no evidence demonstrating that the findings transfer to deep SNN architectures, which greatly diminishes real‑world utility.
  4. Out of 128 tested weight‑perturbation directions, only 120 cases satisfying the strict criterion “exactly one spike disappears and subsequent network activity persists” are retained for analysis. Eight challenging cases featuring multi‑spike alterations or terminal‑event removal are discarded. These excluded cases are precisely the scenarios where gradient failure is most severe and correction is most needed; the method’s behaviour under these hard cases remains entirely unknown. The reported high correlation is observed only on this filtered subset and cannot be generalised.
  5. Observed loss jumps lie in the tiny magnitude range of 10−7∼10−510^{-7} \sim 10^{-5}. The manuscript provides no evidence demonstrating that loss perturbations of this scale produce measurable impacts on practical training outcomes, including convergence speed or final task accuracy.
  6. Experiments on finite parameter updates draw merely on 20 perturbation directions following a single input‑weight‑scaling scheme. These do not reflect stochastic gradient directions encountered in real training workflows, limiting representativeness.
  7. The strong correlation between event‑action predictions and measured jumps is an empirical observation only. The paper provides no formal theoretical justification explaining why adjoint variables computed on one trajectory can predict loss differences across a different event trajectory. Error‑order bounds and precise conditions of validity are also absent.
  8. Calculating per‑spike event‑action values requires full EventProp backward passes plus storage of adjoint variables for every individual spike. For deep, large‑scale SNNs with massive spike counts, this imposes substantial computational and memory overheads. Complexity analysis for such realistic settings is completely missing from the paper.
Questions for authors

Please see Weakness.

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

Related papers

Discussion (0)

Sign in to comment.