An evolvable NoC-based spiking neural network architecture

Abstract

Nature employs bio-inspired concepts such as evolution and learning to develop complex and intelligent organisms, capable of adaptation and fault tolerance. Brain-inspired paradigms such as Spiking Neural Networks (SNNs) offer the potential of elegant, low-power and robust methods of performing computing. Previous work by the authors reports a reconfigurable mixed signal Network on Chip (NoC)-based SNN architecture, with reconfigurable analogue neuron cell and digital NoC The SNN architecture includes an array of neural tiles, each incorporating a NoC router for packet-based neuron interconnect. This paper presents a Genetic Algorithm (GA) based evolution framework which interacts with the SNN architecture to evolve SNN-based solutions to problems. Simulation results are presented which verify the adaptability of the reconfigurable NoC-based SNN architecture in evolving a solution for the XOR benchmark problem. Results on the synthesised neural tile area utilisation for FPGAs are also presented. This work contributes to the realisation of a large scale reconfigurable mixed signal hardware platform for SNNs.