We present a time and space integrating optical architecture for multi-layer finite impulse response neural networks (FIRNN). The proposed architecture is capable of forward propagation and on-line learning in the form of backward propagation. FIRNNs are first presented and analyzed. From the analysis it is observed that the implementation of FIRNNs requires the calculation of temporal convolutions, which inspire the use of time-integrating and space- integrating optical architectures. A novel device is proposed for the space-integrating architecture, based on the use of a rotating volume hologram. Initially, two single-layer architectures based on the space integrating and time integrating architectures are presented, leading to the multi-layer architecture, which uses a combination of both architectures, folding them together in such a way that all the operations of the order O(N³) are performed optically and only the less computationally intensive operations are performed electronically.

Date Published: 1 October 1999
PDF: 10 pages
Proc. SPIE 3804, Algorithms, Devices, and Systems for Optical Information Processing III, (1 October 1999); doi: 10.1117/12.363977

Published in SPIE Proceedings Vol. 3804:
Algorithms, Devices, and Systems for Optical Information Processing III
Bahram Javidi; Demetri Psaltis, Editor(s)