Optical neurochip based on a three-layered feed-forward model

Perceptrons with Hebbian learning based on wave ensembles in spatially patterned potentials

A general scheme to realize a perceptron for hardware neural networks is presented, where multiple interconnections are achieved by a superposition of Schrödinger waves. Spatially patterned potentials process information by coupling different points of reciprocal space. The necessary potential shape is obtained from the Hebbian learning rule, either through exact calculation or construction from a superposition of known optical inputs. This allows implementation in a wide range of compact optical systems, including (1) any nonlinear optical system, (2) optical systems patterned by optical lithography, and (3) exciton-polariton systems with phonon or nuclear spin interactions.

Optical learning neurochip with internal analog memory

An optical neurochip with learning capability and a memory function is reported for the first time, to our knowledge. The neurochip is a three-dimensional optoelectronic integrated circuit consisting of a light-emitting diode array and a variable-sensitivity photodetector array. The principle of operation and the fundamental characteristics of the neurochip are described in detail. The synaptic weights, which, are acquired through learning, are stored in the variable-sensitivity photodetector as the detection sensitivity with analog values. Both the positive and the negative synaptic weights are memorized with one variable-sensitivity photodetector element by changing the polarity of the electric signal applied to the photodetector. A storage time of ~20 min. was obtained. With a fabricated optical neurochip that had 32 neurons and 32 x 32 synapses, experiments of on-chip learning for pattern classification were performed successfully.

Dynamic optical neurochip using variable-sensitivity photodiodes

A novel type of a dynamic optical neurochip that uses sensitivity-variable photodiodes (VSPD's) as variable interconnection weight elements is proposed and analyzed. The chip consists of the line-shaped light-emitting diode array and the VSPD matrix array. A monolithic integration of these arrays is presented. This dynamic chip has advantages such as less optical cross talk, higher processing speed, and realization of analog synaptic weights. Also, experimental results of the VSPD array using metal-semiconductor-metal structure are shown. The results of the computer simulation using this chip as a learning element show that the unwanted effect of the optical cross talk on the recognition rate can be alleviated by the learning capability of the chip. It is also found that the theoretical maximum density is ~2000 neurons/cm(2) for the backpropagation model.