Abstract

Monolithically integrated optoelectronic circuits have emerged as a viable solution for implementing nonlinear operations and providing gain in optical neural networks. However, before they can be incorporated into a practical system, these devices must be in the form of large arrays and have sufficient optical gain to compensate for the optical losses due to the low efficiency of the interconnection medium. Furthermore, they much dissipate little electrical power so that the performance will not be limited by heat dissipation and at the same time exhibit high input sensitivity to accommodate the low input power received by the neurons. In this paper, several integration schemes utilizing GaAs light emitting diodes (LEDs), heterojunction bipolar phototransistors. metal-semiconductor field-effect transistors (MESFETs) and optical field-effect transistors (OPFETs) are presented. Typical results in optoelectronic neurons that incorporate phototransistors as the detectors show a differential optical gain of 10-40, which the neurons that incorporate OPFETs as detectors show a gain of approximately 80. The electrical power dissipations in these neurons are less than 2 mW/neuron and the switching energy is measured to be between 10 and 40 pJ. While these results generate optimism, issues like scalability, processing limitations and compatibility, device integrability, input-output isolation, and tradeoff between power dissipation and optical gain still need to be addressed before high density arrays can be practically incorporated in neural networks.

© 1991 Optical Society of America