Massively Parallel Neural Network Simulation Paul Fox * * University of Cambridge Computer Laboratory, JJ Thompson Avenue, Cambridge, CB3 0FD, UK ABSTRACT A system for the simulation of neural networks in real time using FPGAs. KEYWORDS : Neural Network; Simulation; FPGA 1 Introduction The structure of an individual neuron is well known, as way in which the brain is divided into functional units. However, the way in which neurons are networked together to form such functional units is less well understood. One way that neuroscientists use to gain a greater understanding of this is by computer simulation of neural networks, whose param- eters and connectivity can be altered and the resulting behaviour observed. 2 Modelling Neural Networks Mathematically Many attempts have been made to represent the behaviour of neurons using mathematical equations, hence allowing them to be simulated by a computer system. One of the most well known is by Hodgkin and Huxley[HH52]. This is believed to be very accurate, but is also complex, requiring ten equations to simulate a single neuron. This project is using a much simpler model proposed by Izhikevich [Izh03]. This uses two equations per neuron, with one being quadratic and the other linear. This approach is less accurate, but the relative simplicity of these equations makes this acceptable, particularly if the pattern and timing of neural spikes is believed to be of more interest than the exact behaviour of the interior of a neuron. 1 E-mail: paul.fox@cl.cam.ac.uk