Abstract

We describe a new generic approach for realizing neural network models. Schematically the basic system consists of a special purpose electronic integrated circuit—the neural processor (NP) and a 2-D spatial light modulator (SLM). The synaptic efficacies matrix is stored on the SLM and loaded in parallel into the NP. This is done by imaging the contents of the SLM onto an array detector which acts as the input unit of the NP. The NP processes the state of the network by using the synaptic efficacies supplied by the SLM. Several different architectures for the NP are described which enable the construction of networks with both binary and analog neurons, analog programmable synapses, and either synchronous or asynchronous updating. The expected properties of these systems are discussed based on the state of the art of microelectronic signal processing and SLM technology. It is shown that networks with 10²–10³ neurons can be built with update rates of typically 10 kHz for the complete networks. Finally, preliminary experimental results are described.

© 1988 Optical Society of America