Abstract

The potential application of optical systems to perform high speed, low cost signal processing with large parallelism has attracted the attention of researchers for many years. General optical processors have been developed that compute matrix-vector multiplications and other linear algebraic operations using incoherent light. One example is the Optical Matrix- Vector Multiplier (OMVM), which calculates the discrete operation of a matrix-vector product, rather than the continuous correlation and convolution more commonly associated with optical processing [1]. The OMVM can be used to compute discrete Fourier transforms (DFT’s), and for performing linear algebraic operations, including matrix-matrix multiplications. It has been suggested as a method for implementing associative memory [3-5] and optical crossbars [4]. The first OMVM had several disadvantages, including low accuracy, low speed, and a nonprogrammable matrix mask. Recent implementations use real-time spatial light modulators (SLM) [5-7] and acousto-optic cells [8]. The two-dimensional spatial light modulators used in many of these optical processors operate at millisecond speeds, are expensive and have low resolution [5, 7]. One-dimensional modulators such as acousto-optic cells are faster, but the major drawback of computing matrix-matrix operating using one-dimensional devices is that to calculate two-dimensional matrix-matrix operations, data from the rows and columns of matrices must be loaded serially. The cycle time through the processors increases with the order of the matrix, and the natural parallelism of optics is lost.

© 1987 Optical Society of America