A new architecture for all-optical neural network computing

Abstract

The main features of artificial neural networks are a large number of nonlinear processing elements and massively parallel interconnections among themselves. Many researchers have studied hardware of such neural artificial networks and software for highly parallel computing. In terms of the hardware, two different approaches, VLSI techniques and optical neural networks.have been proposed. Basic neural operations in a simple artificial neural network model are based on a spatial weight sum operation, including arithmetic operation and addition, and a nonlinear operation. In each neuron, the synaptic weights and the input signals form other neurons are multiplied, and their sum is subjected by a nonlinear operation to obtain an output. In a general neural network model, arithmetic operations in a neuron include subtraction and negative multiplication, because of bipolar weights corresponding to excitatory weights and inhibitory weights.