Research on SNN Learning Algorithms and Networks Based on Biological Plausibility

Abstract

Spiking Neural Networks, inspired by the brain’s neuronal information processing mechanisms, utilize sparse, event-based spike signals to emulate biological computation. These networks aim to replicate the spike-driven behavior of biological neurons, incorporating temporal dynamics alongside spatial inputs, potentially offering enhanced spatiotemporal processing capabilities compared to conventional artificial neural networks. Due to their event-driven nature and binary spike information representation, they hold promise for ultra-low power consumption, making them suitable for energy-constrained applications. Recent advancements suggest these networks are approaching, and sometimes exceeding, the performance of traditional deep learning methods in domains like image classification and speech recognition, particularly concerning energy efficiency. This paper provides a comprehensive review focusing on the biological plausibility of spiking neural networks. It begins by discussing the concept of biological plausibility and outlining fundamental principles and development approaches for spiking neuron models. It then details various temporal encoding strategies used to convert input data into spikes. The review subsequently explores the evolution of learning algorithms, covering both biologically inspired unsupervised methods, such as Spike-Timing-Dependent Plasticity, and supervised learning approaches adapted primarily from backpropagation techniques using surrogate gradients. Finally, the paper surveys recent advancements in deep network architectures specifically designed for spiking neural networks. Challenges and future research directions are also discussed.