Abstract
Optical processors for neural networks are based on vector-matrix multiplication machines that can, potentially, compete with serial computers, because of their parallelism and ability to facilitate densely connected networks. However, in most systems proposed, the multiplication supports only two quadrants, disallowing the possibility of using both positive and negative neuron outputs to increase network capacity and learning rate. We propose and demonstrate a fast four quadrant optoelectronic multiplying machine, for a bipolar vector and a continuous matrix containing both positive and negative values, which can be used during both the recall and learning sessions of feedforward neural networks. The machine is based on two adjacent liquid crystal television displays (LCTVs), which enable it to perform an exclusive-OR (XOR) operation between the patterns presented on the displays, thus implementing a complete signed multiplication. The absolute values of the matrix elements are presented on a monitor while the sign is applied to a LCTV to achieve the said four quadrant multiplication. The summation of the vector-matrix multiplication is done either by cylindrical optics or by lenslet arrays. Experimental results, obtained by using common commercial components, present a novel useful and reliable approach for a four quadant electro-oprical matrix-vector multiplication in general and for feedforward neural network training and recall stages in particular.
© 1992 Optical Society of America
PDF ArticleMore Like This
Wenlu Wang, Shutian Liu, Jie Wu, and Chunfei Li
CMH6 Conference on Lasers and Electro-Optics (CLEO:S&I) 1992
Jonathan George, Rubab Amin, Armin Mehrabian, Jacob Khurgin, Tarek El-Ghazawi, Paul R. Prucnal, and Volker J. Sorger
SpW4G.3 Signal Processing in Photonic Communications (SPPCom) 2018
S. D. Pethel, C. M. Bowden, and C. C. Sung
FB4 Nonlinear Dynamics in Optical Systems (NLDOS) 1992