Extension of neuron machine neurocomputing architecture for spiking neural networks

The neuron machine (NM) is a synchronous neurocomputing architecture that can be used to design efficient large-scale neural network simulation systems. However the NM architecture has a limitation that it cannot support complex computations such as those for spiking neural network (SNN) models. In this paper, we review the NM architecture and propose an extension of it to support neural network models with complex synaptic and neuronal functions, by providing generalized memory structure and methods to design pipelined circuits for those functions. In addition, we discuss the designing of a neural network simulator that uses the proposed architecture and is implemented on a field-programmable gate array (FPGA) board. We show that the simulator implemented on a 200 MHz mid-range FPGA can run orders of magnitude faster than most existing board-level implementations. The proposed architecture has the additional advantages of simplicity, accuracy, and extensibility to the more biologically detailed neural and synaptic models compared with the existing event-driven approaches.