Implementation of event-driven spiking neural networks in field programmable gate array (FPGA)

Abstract

Spiking Neuron Networks (SNNs) are a fascinating new field of artificial intelligence and computational neuroscience that is directly inspired by the complex work of biological brain systems. Unlike traditional feedforward neural networks (FNN) and recurrent neural networks (RNN), which rely on continuous data streams and feedback for learning, SNN uses an innovative information processing and transmission method. They use discrete synaptic events called spikes to communicate between neurons, which is very similar to pulse-based communication in the human brain and nervous system.

In this dissertation, based on human neuron system and basic principles of SNN. We develop a simulation system for SNN data processing from input to output, simultaneously validating the network structure and neuron parameters through software-level training optimization. Through our work, we find that SNN has practical usage in the field of image recognition. So we continue to try to implement SNN in hardware-level, aiming to leverage the unique characteristics of hardware architectures to unlock new capabilities and efficiencies.