Abstract

Optical neural networks using a lenslet array interconnection have been reported recently,^1,2 in which two inexpensive liquid crystal televisions are used to generate the input pattern and the interconnection weight matrix (IWM). Each of the lenses images the entire IWM, but only one of the IWM-submatrices is used for interconnection. Thus the overall interconnection light efficiency of the systems is bounded by a factor of 1/N², which would limit the performance of a large-scale neural net operation. To improve the light efficiency of the optical neural network, we have replaced the lenslet array by a mirror array. N × N mirrors are affixed on a parabolic substrate. The position and orientation of the mirrors are finely tuned so that each mirror would reflect a corresponding IWM-submatrix onto the output detector. Compared with the lenslet array technique, the light efficiency of the system would be increased N² times.

© 1991 Optical Society of America