Optimising Event-Driven Spiking Neural Network with Regularisation and Cutoff

TL;DR

This work targets efficient, adaptive inference in Spiking Neural Networks (SNNs) by introducing a principled cutoff mechanism that can terminate inference at any timestep, paired with a Regularising Cosine Similarity (RCS) penalty to optimise training for cutoff. The approach is applicable to both ANN-to-SNN conversion and direct SNN training, and is designed to minimize inference timesteps while preserving accuracy. Empirical results show 1.76–2.76× faster timesteps for CIFAR-10 and 1.64–1.95× for event-based datasets, with near-zero accuracy loss, and notable improvements in optimal cutoff metrics (OCT) especially under event-based conditions. The work demonstrates robust compatibility with existing conversion and training methods and offers practical pathways for energy-efficient, latency-aware SNN deployment, along with a public codebase.

Abstract

Spiking neural network (SNN), as the next generation of artificial neural network (ANN), offer a closer mimicry of natural neural networks and hold promise for significant improvements in computational efficiency. However, the current SNN is trained to infer over a fixed duration, overlooking the potential of dynamic inference in SNN. In this paper, we strengthen the marriage between SNN and event-driven processing with a proposal to consider a cutoff in SNN, which can terminate SNN anytime during inference to achieve efficient inference. Two novel optimisation techniques are presented to achieve inference efficient SNN: a Top-K cutoff and a regularisation.The proposed regularisation influences the training process, optimising SNN for the cutoff, while the Top-K cutoff technique optimises the inference phase. We conduct an extensive set of experiments on multiple benchmark frame-based datasets, such asCIFAR10/100, Tiny-ImageNet, and event-based datasets, including CIFAR10-DVS, N-Caltech101 and DVS128 Gesture. The experimental results demonstrate the effectiveness of our techniques in both ANN-to-SNN conversion and direct training, enabling SNNs to require 1.76 to 2.76x fewer timesteps for CIFAR-10, while achieving 1.64 to 1.95x fewer timesteps across all event-based datasets, with near-zero accuracy loss. These findings affirms the compatibility and potential benefits of our techniques in enhancing accuracy and reducing inference latency when integrated with existing methods. Code available: https://github.com/Dengyu-Wu/SNNCutoff

Figures (6)

• Figure 1: An illustrative diagram showing the regularisation for optimising SNN and the cutoff mechanism for reducing latency on CIFAR10-DVS dataset. Cutoff is triggered when $Y_{\text{gap}}$ is greater than $\beta$, a value dynamically determined by a confidence rate as introduced in Section \ref{['sec:cutoffmech']}.
• Figure 2: Illustration of inference process in a LIF neuron within the hidden layer. Input spikes charge the membrane potential $V^l_i(t)$ through weighted and bias currents learned during the training stage. When $V^l_i(t)$ reaches the threshold $V^l_{\text{thr}}$, the neuron will generate a spike and then reset the $V^l_i(t)$. In our study, the Integrate-and-Fire (IF) neuron with soft reset is a special case of LIF model when $\tau=1$.
• Figure 3: Evaluation of the Top-K cutoff on CIFAR10-DVS, using a directly trained SNN model with $T = 10$ (as detailed in Section \ref{['sec:exp']}): (a) The increase of $\beta$ limits the number of samples eligible for cutoff, while concurrently enhancing the confidence in the cutoff decision. (b) To enhance the readability, the inference timestep of 16 samples from the test dataset under varied trigger conditions.
• Figure 4: Comparison of SNN with and without Top-K cutoff on CIFAR10 (left) and CIFAR10-DVS (right) across various training methods: (a) The Top-K cutoff is determined by $\epsilon$ values ranging from 0.00 to 0.50 in increments of 0.05. (b) The statistic data is extracted from testing samples under $\epsilon=0.02$ for CIFAR10 $\epsilon=0.0$ for CIFAR10-DVS.
• Figure 5: Comparison of Top-K cutoff accuracy before (dashed lines) and after (solid lines) regularisation, across a range of $\epsilon$ values from 0.00 to 0.50, increasing in steps of 0.05. The accuracy for full-length input is detailed in the legend.