Temporal Reversal Regularization for Spiking Neural Networks: Hybrid Spatio-Temporal Invariance for Generalization

TL;DR

This work tackles overfitting in spiking neural networks by introducing Temporal Reversal Regularization (TRR), which uses input/feature temporal reversal and a Hadamard-based spike firing-rate hybridization to learn perturbation-invariant representations. The method yields a multi-head training regime with a consistency objective and a PAC-Bayesian justification showing tightened generalization bounds. Empirically, TRR improves generalization across static, neuromorphic, and 3D tasks, reduces average spike firing rate for energy efficiency, and enhances adversarial robustness, across multiple architectures at low latency. The approach is architecture- and task-agnostic and supports extensions with neuromorphic data augmentation, making TRR a practical strategy for deploying low-power, high-performance SNNs in real-world settings.

Abstract

Spiking neural networks (SNNs) have received widespread attention as an ultra-low power computing paradigm. Recent studies have shown that SNNs suffer from severe overfitting, which limits their generalization performance. In this paper, we propose a simple yet effective Temporal Reversal Regularization (TRR) to mitigate overfitting during training and facilitate generalization of SNNs. We exploit the inherent temporal properties of SNNs to perform input/feature temporal reversal perturbations, prompting the SNN to produce original-reversed consistent outputs and learn perturbation-invariant representations. To further enhance generalization, we utilize the lightweight ``star operation" (Hadamard product) for feature hybridization of original and temporally reversed spike firing rates, which expands the implicit dimensionality and acts as a spatio-temporal regularizer. We show theoretically that our method is able to tighten the upper bound of the generalization error, and extensive experiments on static/neuromorphic recognition as well as 3D point cloud classification tasks demonstrate its effectiveness, versatility, and adversarial robustness. In particular, our regularization significantly improves the recognition accuracy of low-latency SNN for neuromorphic objects, contributing to the real-world deployment of neuromorphic computational software-hardware integration.

Figures (8)

• Figure 1: Overview of the proposed method. TRR perturbs temporal and static data by (a) input and (b) spike feature temporal reversal, allowing the SNN to produce original and temporally reversed outputs, and (c) encouraging both outputs to be as similar as possible to learn generalized perturbation-invariant representations. In addition, TRR hybridizes the original and temporally reversed spike firing rates and expands the implicit dimensionality with a lightweight "star operation”. This serves as a spatio-temporal regularizer that further facilitates the generalization of SNNs.
• Figure 2: Influence of the balance coefficient $\alpha$ on the performance. A larger $\alpha$ indicates stronger perturbation regularization. As a whole, our method is insensitive to $\alpha$ and consistently outperforms the baseline.
• Figure 3: Visualization of ASFR. Our method has a lower ASFR than the baseline, favoring low-power deployment.
• Figure 4: Visualization of original, temporally reversed, and hybride spike firing rates and perturbations after 1 epochs of TRR training. Shown here are the first 16 channels of the penultimate layer of the VGG-9 network on the CIFAR10-DVS, where the input is the example in Fig. \ref{['visfull']}. The results show that the "star" operation hybridization caused a significant negative perturbation (blue area in the rightmost subfigure).
• Figure 5: Visualization of original, temporally reversed, and hybride spike firing rates and perturbations after 100 epochs of TRR training. Shown here are the first 16 channels of the penultimate layer of the VGG-9 network on the CIFAR10-DVS, where the input is the example in Fig. \ref{['visfull']}. The "star" operation hybridization induced fewer negative perturbations than at the beginning of training, indicating that the model learned perturbation-insensitive generalized representations.