No abstract

No abstract

Single-channel optoelectronic neurocoprocessor on the base of vector-matrix multiplier with column submatrices is designed. Two models with dynamic and constant threshold are proposed for implementation. Results of pattern recognition simulation showed fast and effective convergence to the memory patterns 1. 1NTRODUCI1ON Development of the neural network (NN) optical implementations is explained by principa1 parallelism of the neural algorithms. Most of the optical implementations can be divided into the two main classes: holographic correlators and vector-matrix multipliers (VMM). The second class bases on the description of many neural algorithms in terms of multiplication of an input vector by a weight matrix.Vector-matrix multiplication for the large scale input vector (N 256) can be realized by two methods with standard optoelectronic devices:1. Initial weight matrix is partitioned into N submatrices, where N is the dimension of the NN (number of the neurons). Each submatrix comprises all the weights of corresponding string of the mitial submatrix. In this case the product is defmed by the signals of the photodetector array and radiation inciding on the ith photodetector is collected by the ith lens from lenslet array situating behind the submatnces. However, using such system it is necessary to duplicate (multiplicate) input vector, which can be turned into the quasimatrix for compactness (Fig. 1). In other words, N quasimatrices of the input vector are needed to implement neural net with N neurons, that is at a loss economically in the first place. But optical part of the system is very simple in this case: array of the focusing lenses after the weight matrix and array of the projecting lenses before it. If we realize the input with LCD technique, it would be possible to design system without last one at all; but that would make capacity of the systern worse because of the large LCD time constants. Yl Yj IN 92 SPIE Vol. 2969 • O-8194-2375-01961$6.Oo * This paper combines two posters Fig.1 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/21/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

The comparative analysis of hologram and optoelectronic implementations of neural networks (NN) shows that at least for a small-format (N<128) and middle-format (N256 , . . 1024) optical NN the optoelectronic implementation is optimum. It is based on the original optical vector-matrix multiplier. This one contains quasi-rectangular LED-array for the N-length input vector load, projective monochrome PC's LCD for weight-matrix input, quasi-rectangular array of Si-photodiodes (or CCD), controlling PC (Notebook type), and additional electronics for linkage PC and optoelectronic neural processor. The peculiarities of such optical system assembled from the standard optoelectronic components and providing neural processing of the middle-format images (up to iO pixels) during one stroke of coprocessor (10. . . lOOns) were examined.

Mathemai.ica1 description and algebraic interpretation of recursive algorithm in optical binary pattern processing are carried out .The recursive optoelectronic processor configuration is described.