Optimizing event-driven spiking neural network with regularization and cutoff

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