acceptodds
Withdrawn from ICLR 2027

Learning Robust Event Representations through Finite-Time Dynamical Stability

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Event cameras are a class of bio-inspired vision sensors with advantages such as high temporal resolution, low latency, and high dynamic range. Existing state-of-the-art event processing methods mainly rely on deep networks to learn features, but these methods do not consider the internal temporal dynamics of the model, causing models to learn fragile temporal features and thus suffer significant performance degradation under distribution shifts such as event dropout, temporal jitter, and polarity perturbations. To address this issue, incorporating recent findings in the primary visual cortex, we propose Lyapunov-Regularized Dynamical Representation Learning (LRDR) for event information processing. Specifically, this method maps event streams into a trainable nonlinear dynamical system and uses finite-time Lyapunov exponents to constrain the dynamical stability of different samples, encouraging the same sample to maintain consistency under perturbations while enhancing dynamical separability between different classes. Across four event-recognition benchmarks, LRDR achieves state-of-the-art classification accuracy of 85.7% and 99.9% on N-Caltech101 and ASL-DVS, respectively. Our results demonstrate that finite-time stability is not merely an auxiliary property for analyzing the dynamics of neural networks after training, but can be directly optimized as a learnable representational geometry, thereby providing more stable and robust temporal representations for event-based vision models.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
event-based vision, event representation learning, dynamical systems, brain-inspired neural networks

Reviews

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

This paper proposes LRDR, a Lyapunov-regularized representation learning method for event data classification. The model combines a pretrained ResNet-34 semantic branch with a learnable grouped CCNN branch. During training, it computes a directional finite-time Lyapunov growth rate and regularizes it through perturbation invariance, class separation, and stable-range constraints.

Critical strengths

(1) The model combines learnable dynamics with pretrained semantic features and clearly separates the training and inference paths, so the stability regularization does not introduce additional FTLE-specific computation at inference time. (2) The paper provides augmentation-only and feature-invariance controls with matched backbones, preprocessing, and training protocols. It also evaluates the three Lyapunov loss terms separately, which helps identify the sources of the observed gains. (3) The appendix provides detailed descriptions of the state update, tangent-vector propagation, loss coefficients, and training schedule.

Critical weaknesses

(1) Appendix A.2 correctly states that Eq. 42 measures the growth rate along a single propagated direction, and Eq. 43 also clearly shows that this quantity does not exceed the maximum finite-time growth rate. However, the perturbation bound in Appendix A.3 uses maximum growth rates over different time intervals and also includes an input-sensitivity term and perturbations of the encoded features. Constraining the growth rate along one fixed initial direction does not directly imply control over these quantities. (2) The periodic condition in Eq. 56 of Appendix B does not directly imply Eq. 57, because the final update generally depends on the threshold state at the previous time step rather than on the initial state. Eq. 65 gives a nonzero internal state at time zero, which conflicts with the zero initialization stated earlier, and its denominator is inconsistent with Eqs. 70 and 71. Eq. 69 uses the derivative sign of the equilibrium-condition equation to determine the stability of the discrete system, but the required Jacobian analysis of the state update is not provided. (3) In Table 3, LRDR improves the mean perturbed accuracy over Aug-only and Feature Inv by 0.38 and 0.31 percentage points, respectively. However, its relative performance drop is 2.08%, which is higher than the 1.81% and 1.92% of these two controls. The paper therefore needs to distinguish more clearly between improving clean and perturbed accuracy and reducing sensitivity to perturbations. (4) Section 4.4 states that LRDR improves the mean perturbed accuracy over ACDF by 1.79 percentage points, but the values 83.96 and 81.17 in Table 3 differ by 2.79 percentage points. The text should be checked.

Questions for authors

(1) What is the empirical relationship between the growth rate along a fixed tangent-vector direction and the maximum growth rate used in the perturbation bound in the appendix? Can the paper compare these two quantities on representative test samples, together with the actual trajectory changes caused by input perturbations? (2) Please check Eqs. 57, 65, and 69 in Appendix B and clarify the parameter conditions under which the periodic and chaotic behavior claims hold. (3) How exactly does the rate perturbation modify the timestamps and observation window? If the events are rebinned according to each sample's own duration, does the perturbation still change the model input? Do the other perturbations involve timestamp clipping or reordering? (4) How does a temporal aggregation method perform when it uses the same paired augmentation and feature regularization but does not use nonlinear dynamics?

Reviewer_J8ABOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LRDR, a representation learning framework for event camera classification. The method encodes event streams into polarity-separated time series,extracts semantic features using an ImageNet-pretrained ResNet-34, and models temporal evolution via a grouped residual dynamic module. The authors further characterize the stability of each group's dynamics using the Finite-Time Lyapunov Exponent (FTLE), and train the model with stability invariance, inter-class separation, and boundedness losses.Classification accuracy is evaluated on N-MNIST, N-Caltech101, N-CARS and ASL-DVS, and several synthetic event perturbations are tested on N-Caltech101. The paper reports competitive classification performance, around 1.2GMACs and an inference latency of 8.5 ms. Ablation studies are conducted to analyze the contribution of each loss term.

Critical strengths

S1.The design of the method incorporates the temporal structure of events into representation learning. Unlike approaches that simply convert events into images or time surfaces, LRDR explicitly preserves polarity-separated time series and imposes constraints on dynamic features during training. This design aligns with the temporal properties of event streams. S2.The experiments span multiple classification benchmarks,including ablation studies and complexity analysis. The authors report not only classification accuracy but also comparisons across different modules and loss terms,alongside MACs and latency metrics, enabling readers to preliminarily evaluate the method's performance and computational cost.

Critical weaknesses

W1. The paper adopts FTLE for training regularization and achieves promising results under synthetic perturbations. However, it has not yet demonstrated that this constraint reliably reflects the model's robustness to real sensor variations or natural distribution shifts. W2. The baselines in Table 1 may follow different experimental protocols, and the results reported herein are obtained with a fixed data split. Consequently, it remains unclear whether the marginal performance gain is statistically significant and consistently reproducible.

Questions for authors

Q1.Were all baselines in Table 1 re-run under exactly the same data split, preprocessing and evaluation protocol as LRDR? If not, could you provide baseline results under this unified protocol, together with statistics over multiple independent splits or paired tests, to verify whether the marginal performance gain is consistent?

Reviewer_dKPPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Inspired by the primary visual cortex, this work presents Lyapunov-Regularized Dynamical Representation Learning (LRDR). It encodes event streams into a learnable nonlinear dynamical system and uses finite-time Lyapunov exponents to regularize stability, preserving consistency for perturbed samples while boosting inter-class separability. LRDR reaches SOTA accuracy of 85.7% on N-Caltech101 and 99.9% on ASL-DVS. It validates that finite-time stability can be directly optimized as a representational geometry to yield robust temporal representations for event-based vision.

Critical strengths

The idea of Lyapunov-Regularized Dynamical Representation Learning seems interesting for the event-based vision.

Critical weaknesses
  1. This paper bears obvious traces of large-language-model generation. Worse still, many of its illustrations and result plots appear to be AI-generated, which is unacceptable. I also cast doubt on the rigor of its other experimental results.

  2. The benchmark datasets used, like N-Caltech101 and ASL-DVS, are small-scale, which is not enough to validate the effectiveness. Recent large-scale benchmark datasets are suggested, such as HARDVS, CeleX-HAR.

  3. The related works mentioned in Section 2 are relatively old, the works published in year 2025 and 2026 are scarce.

  4. The compared methods in Table 1 are old, without comparing with SOTA.

  5. The temporal fusion seems a simple combination of attention, mean and last, which is not novel.

Questions for authors
  1. Please add more experiments on large-scale benchmark datasets, like HARDVS, CeleX-HAR.

  2. More SOTA models need to be introduced for comparison.

  3. All the figures need to be adjusted without using the large model.

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

Related papers

Discussion (0)

Sign in to comment.