Development of an optical neural network module

1. Introduction The main features of artificial neural networks are the large number of nonlinear processing elements and the massively parallel interconnections among them. Many researchers have studied the hardware required for artificial neural networks and the software for such highly parallel computations. In terms of the hardware, two different approaches have been proposed: VLSI-based neural networks and optical neural networks. The VLSI techniques, in which remarkable progress has been made, are adopted to many artificial neural networks. Compared with the complex 3-D interconnections of the human brain, the VLSI techniques have an inherent 2-D architecture with a poor density of wiring between 2-D surfaces. Optical signals, on the other hand, can flow in the 3-D space. Many advantages of applying optics to artificial neural networks have been discussed, and many optical and optoelectronic neural networks have been proposed. The optics, which has inherent parallelism and high speed features , offers high potential interconnections in terms of density, capacity, and flexibility. Optical techniques lead to huge parallel operations and interconnections, and provide useful hardware for artificial neural networks. Optical neu-rochips, which integrate the sophisticated VLSI fabrication techniques with optical interconnection techniques, have been studied. In order to develop a large scale optical neural network computing system, a smaller module is developing at the first step. We have two approaches considered, that is, the reversal input superpositioning technique (RIST) 1 in which all neural operations are done in the positive range, and the 2-D architecture with a multiple imaging system. 2