Improving Spiking Neural Network With Frequency Adaptation for Image Classification

Spiking neural networks (SNNs) are promising in energy-efficient brain-inspired devices for their rich spatio-temporal dynamics, bio-plausible encoding, and event-driven information processing. However, the existing SNNs for image classification have fixed firing thresholds for the neurons and do not consider the adaptive properties of the neurons. In this article, we propose a high-performance SNN composed of neurons with spike frequency adaptation (SFA-SNN). We replace the fixed firing threshold with dynamic firing thresholds and incorporate them into the differential equation of neuron membrane potential, and then build an SNN on Pytorch. In addition, we introduce a new function to approximate the derivative of spike activity to solve its nondifferentiable problem, so that the SNNs can be trained in spatio-temporal domain using the error backpropagation algorithm. We verify the image classification performance of the proposed SFA-SNN on the static data set (including MNIST and Fashion-MNIST) and neuromorphic data set (including CIFAR10-DVS and DVS128-Gesture), and the accuracy results, including 99.52% on MNIST, 92.40% on Fashion-MNIST, 71.90% on CIFAR10-DVS, and 96.67% on DVS128-Gesture. We believe this work can help us better understand the intelligent information processing of the brain.