We describe the combination of neural network training and volume holographic storage technologies using photorefractive crystals for real-time image processing. Experimental results on using the system for multi-channel distortion-invariant image recognition are presented.

A three-layered neural network (NN) for pattern recognition with feedback and complex states of neurons and interconnections is suggested. NN is based on adaptive resonance principles and consists of comparison, recognition and selective attention (vigilance) layers. Comparison is carried out in spectral domain, recognition and selective attention -- in image space. Parallel-sequential accessing to long-term memory is used. Adaptation is realized by creation of new recognition categories and by long-term memory variance when the input pattern is similar enough. Hybrid opto-electronic implementation of NN is used. The main optical part is a joint transform correlator with a dynamic holographic filter.

A hybrid optical-electronic system for pattern recognition based on a neural network model is under discussion. A structure of two-layered holographic neural network is proposed. In order to improve the iteration process an auxiliary sublayer with local connections is added to the first layer. A new class of optical high-performance processors -- intensity-linear holographic correlators to provide global connections in both layers of the neural network is proposed and discussed.

Optical neural network formed by placing the holographic correlator into the linear resonator is discussed. Variation of the attractor position by means of inhibitory optical interconnections to achieve new solution types is proposed. The experimental results are presented.

Basic models of neuron-like media and corresponding sets of spatio-temporal solutions are considered. The ways to use these models and solutions as basic units for designing hierarchic systems of decision making with the help of neural-like parallel complex image algorithms are considered.

This investigation is one of the preliminary steps to create an artificial system of visual image recognition and reconstruction implementing main principles of visual image processing in living biological systems. A previous study of neural net models with decreasing inner neurons' dynamics proved that net with lateral inhibition ties is able to convert image into compressed form and to pick out key elements of the picture. The present papers propose to use a reverse neural net with increasing inner neurons' dynamics with the same other parameters (the conjugate net) for image back reconstruction and image histograms' entropy measures as integral values for dynamic image processing monitoring and control.

Fourier-analysis and spatial frequency treatment of optical image formation by 2-D photodetector array are given. A spectral theory of the matrix photodetector accounting for its step-type behavior and crosstalk is developed. The theory gives analytically the expressions for the optical transfer function of matrix photodetectors.

A multichannel correlator based on photothermoplastic is presented in the paper. A one-layer neural network algorithm is realized for pattern recognition. Photothermoplastic holograms efficiency correction is investigated. Neural net learning process is realized using the filters efficiency adjustment. Recognition of the 5 images by 6 recorded feature patterns is shown.

Recently, a terminal attractor based associative memory (TABAM) with optical implementation techniques was published in Applied Optics (August 10, 1992). Herein perfect convergence and correct retrieval of the TABAM are demonstrated via computer simulation by adaptively setting the threshold values for the dynamic iteration for the unipolar binary neuron states using terminal-attractors and an inner-product approach. The simulations are completed by (1) exhaustive tests with all of the possible combinations of stored and test vectors in small- scale networks, and (2) Monte Carlo simulations with randomly generated stored and test vectors in large scale networks with a M/N ratio equals 4 (M: the number of stored vectors; and N: the number of neurons up to 256). The feasibility of optoelectronic implementation is discussed.

Object segmentation, recognition and localization are challenging because of the large amount of input data and because of the invariances required. We discuss strategies to overcome these problems, considering sensors, algorithms and architectures. Specifically, we address neural nets and Hough strategies. The ability of implicit learning makes neural nets interesting for industrial inspection: compared to classical methods they promise robustness against variations of the input data. Furthermore, no expert is necessary for supervision. The inherent parallelity simplifies the design of algorithms. However, the advantages are counterbalanced by a serious drawback: the high computational complexity -- if images are considered. The ability of optics, to help by its inherent parallelity is limited, because neural architectures are usually space variant and cannot simply be implemented optically. We discuss approaching these problems by feature extraction, by sparse algorithms and by space invariant architectures. A competitive strategy for object recognition and localization is based on probability tables, such as the Hough transform uses: a couple of weak but independent hypotheses can give a safe decision about the kind and the locus of an object. This method requires a learning phase prior to the working phase, as the neural strategy does. In that sense it is similar, however, the computational complexity can be much smaller. This makes it possible to segment, localize and recognize objects invariant against shift, rotation and scale.

Dynamic mechanisms of information processing by chemical light-sensitive media based on reactions of Belousov-Zhabotinsky type having nonlinear kinetics were discussed and implemented. The optical and digital system based on IBM PC AT 286 was used for input of visual information into media and control of data treatment. Investigated nonlinear dynamic media gave the opportunity to implement some image processing operations having rather high computational complexity (in particular, contour enhancement of closed image fragments, image segmentation and so on).

A simplified retinal neural network (RNN) model has been considered. The main properties of this model are as follows: (1) primary transform of input raster simulates a decrease of resolution from the fovea to the retinal periphery; (2) the RNN consists of two layers, i.e., excitatory and inhibitory ones, each of them being formed by elements with identical properties excluding input transform; (3) each element of the excitatory layer is inhibited by the retinotopically corresponding element of the inhibitory layer; and (4) receptive field size and time constant of inhibitory neurons are more than those of excitatory ones. Two versions of the RNN differing in several aspects from each other were developed. In the first model the Gauss transform was used as a primary transform of the input raster. In addition, a wide range of the RNN and visual stimulus parameters was tested by computer simulation. The primary transform in the second model was performed by brightness averaging on neuron receptive fields. In the last case, qualitative behavior of the RNN was studied analytically. It was shown that neuron dynamics in response to moving stimuli and the preferable velocity of motion depended on neuron position in the RNN. In particular, foveal neurons were tuned to lower velocity as compared with peripheral ones.

Multilayer neural networks are widely applied in fields of pattern recognition, speech processing, optimization problems, non-linear identification, non-linear adaptive control and other applications. They are trained usually by the error back-propagation algorithm. The main calculation problem of the algorithm is the goal function gradient searching implemented successively backward from the output layer. Two-layer neural networks can solve the approximation problem for a complicated non-linear function of many variables, as well as be effectively applied to automatic control problems, namely for the non-linear dynamic object identification. Calculation of the goal function gradient can be performed directly for two- layer neural networks, omitting the error back-propagation procedure, while a large number of calculations on each step remain. A training procedure simplified from the calculation point of view aimed at hardware implementation is suggested below for two-layer neural networks.

The problem of neural network synthesis with a predetermined sequence of activity pattern is considered. An algorithm for NN interconnection matrix obtainment, giving an approximate but robust decision to the problem, is proposed and studied at length. The limits of satisfactory functioning of the proposed mechanisms are determined. Different types of NN activity dynamics visualization are compared. The applications of the results obtained for the information processing devices and for interpreting of the neurophysiological data are discussed.

We consider an opportunity of utilizing massively parallel architecture similar to that of the well known connection machine (CM) in neurocomputers. It is shown that use of optical means in such a system permits considerable increase in its productivity.

A closed multilayer dynamic neural network, unlike a nonclosed one, can form infinite pattern sequences, both periodic and aperiodic. The closing enhances interactions between time pattern sequences and formation of new sequences. Methods of recording time sequences are considered; prospects of use are discussed.

A new multiport associative memory based on the optoelectronic principle of data processing is suggested. This memory enables M users to execute simultaneously and independently a parallel content-based keys search and data retrieval into a common memory of N stored words by M search arguments, as well as a random-access writing of keys and data. The main parameters of such an associative memory are evaluated and its hardware implementation is discussed.

An artificial neural net system with decreasing response of elements corresponding to models of some economic and biological systems is investigated. Some special modes of neural dynamics are considered. The existence of two-dimensional chaos is shown. The conditions arising of various modes of neural dynamics are analyzed.

The neurocomputer based on parallel mathematical logical unit and random access memory is produced. Test experiments of neurocomputer operation are made. The main application of the neurocomputer is pattern recognition. There is a possibility to apply the hardware developed for evolutionary optimization.

The problems of using simple neural networks for the spatially invariant recognition of contours is discussed in this paper. An optoelectronic circuit for primary information processing in recognition problems is offered.

The results of research into possibilities of making an apparatus for automatic recognition (vowels and syllables) in continuous speech are given. Methods of the preliminary processing of the electric analogue of the speech signal with the aim to decrease its excessiveness and to obtain the invariance with respect to acoustic phonetic categories are studied. A programmed model of a neural network for identification of signal fractions corresponding to acoustophonetic pattern is implemented. The experimental results of research of the computer model of an apparatus of automatic recognition are given.

Architectures of optoelectronic neural network modules based on digital 3-D holographic memory are evaluated in terms of holographic memory design, module scalability and learning capability. A feed-forward module with 900 inputs and 9 outputs with 12-bit interconnection accuracy and perceptron learning are demonstrated experimentally.

The paper presents the obtained results of learning of the two-layer (64 X 8) neural network (TLNN), the results of tolerant noising of weight matrixes and results of hardware implementation. The imperfection of optics satisfies the margin requirements of the TLNN model.

We study how it is possible to use media containing oscillator elements for parallel image processing. Regimes of self-sustained oscillations and damped quasi-harmonic oscillations under resonant excitation are discussed. The oscillators may be isolated or interacted. It is shown that oscillator media are very convenient for extraction of images from noise, amplification of images, removing small scale distortions from patterns, halftoning, parallel extraction of regions of equal intensity or color, determination of common and distinct parts of images, moving elements, points of extrema, contours, etc.

A method to construct a unipolar interpattern association (IPA) interconnection weight matrix (IWM) is presented. By searching the redundant interconnection links, a method that removes all negative links is introduced. Computer simulation as well as experimental results have shown that the unipolar IWM IPA neural network performs better than that of the bipolar IWM IPA model.

The combination of digital parallel computing and optical neural computing give us a new type of hybrid computing. A temporal coding technique is employed for preprocessing digital computing. This data is subjected by a three-layer optical neural computing system to make character recognition.

For the model of coupled phase oscillators, problems of associative memory are considered. It is demonstrated that the networks of oscillators possess some necessary properties for associative memory design. In the case of networks with a tree-like structure of connections, only one pattern can be placed at any prescribed point in the phase space. An approach to the analysis of the number and possible locations of memory images is suggested. It is based on the fact of the existence of invariant sets of critical points.

The system of coupled oscillators, interacting via arbitrary symmetric matrix of connections, is studied from a viewpoint of associative memory modelling. A self-consistent field approach which consists in operating with a finite number of macrovariables (appropriate inner products which can be regarded as order parameters) is used. A system of dynamic equations of oscillatory network being rewritten in terms of macrovariables has a form of independent equations. Being completed by functional equations for order parameters, this system provides a self-consistent description of the oscillatory network. In particular, the approach can be used as an instrument for studying the dependence of the number of network phase locked states on the frequency distribution and the architecture of connections. The abilities of this approach are demonstrated in the case of a network with all-to-all uniform connections.

The Boltzmann machine (BM) for the travelling salesman problem is considered. The BM modification having only the distance connections between neurons and using the `column replacement' rule is proposed. Computer simulation results are presented. The optoelectronic hardware of this BM is discussed.

This paper presents two smart VLSI/FELC spatial light modulators which are designed as a somatic plane and a synaptic plane for optoelectronic implementations of artificial neural networks.

A spatio-temporal memory model (STM) based on 1-D point mapping is considered. An optoelectronic device for high-effective and fast point mapping is proposed.

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 (N equals 256...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 PCs LCD for weight-matrix input, quasi-rectangular array of Si-photodiodes (or CCD), controlling PC (notebook type), and additional electronics for a linkage PC and optoelectronic neural processor. The peculiarities of such an optical system assembled from the standard optoelectronic components and providing neural processing of the middle-format images (up to 10³ pixels) during one stroke of a coprocessor (10...100 ns) were examined.

The operation of multiplication of a vector by a multidimensional matrix is known to provide the basis for neural network algorithms. As a consequence the processor using such algorithms should repeatedly multiply the input data by matrix elements kept in the memory. The making of such a processor using traditional microelectronics technologies runs into the problem of realizing a parallel system of independent interconnects with specified weight coefficients. At the same time optical holographic methods are quite suitable by their nature for solving this problem because photons do not interact in propagation and the holographic means of representing information are associative, i.e., the information is not localized in the memory cells but distributed in the medium. The use of holographic optical elements (HOE) provides a means for solving this problem. HOEs were first proposed in and are designed for transforming wavefronts according to a predetermined law. In this paper the technology of manufacturing computer-generated HOEs by using electron-beam lithography is presented. HOEs made this way allow one to realize neural network algorithms with high accuracy.

A mathematical model of an optoelectronic matrix-vector multiplier which can be efficiently applied to the pattern recognition problems and neural networks is described with consideration for disturbing factors which are present in a real optoelectronic system, such as noise and errors of light sources and photodetectors, ADC nonlinearity and point spread function of the optical system. Analytic formulae relating processor accuracy to the value of error probability and accuracy parameters of its processor elements as well as optical system quality are delivered. The obtained theoretical results are closely approximated with modern experimental data.

The paper is devoted to investigations of an optoelectronic generator (OEG) which under certain conditions may be considered as a multistable device. Operations regimes of the device and optimum conditions of generation are investigated. A possibility to regulate the operation states of the OEG by means of short radio pulses is examined.

An acousto-optic system with a feedback which controls frequency of acoustic waves excited in an acousto-optic cell is investigated theoretically and experimentally. In the system an amplitude transparent situated in front of a photodetector defines a nonlinear dependence of light intensity registered by the photodetector on the ultrasound frequency. Conditions for appearance of multistable states, which differ by amplitude, frequency and diffracted beam direction, are found. Applications of the system for information channels switch in optical communication and for stabilization of laser beam direction are discussed as well.

A model of reversible recording medium based on photochromic molecules dispersed in a glassy polymeric matrix is considered. It is shown that electron energy migration and quenching of electron-excited photochroms can lead to bistability of photoconversion extent and absorption dichroism under linearly polarized excitation. Applications of the model to optical information processing are analyzed.

Three primary colors (red -- 620 nm, green -- 540 nm, blue -- 455 nm) are obtained by means of A2B6 e-beam pumped semiconductor lasers. Semiconductor plates 5 cm in diameter are used. Light output in white is 3000 lm (9 W). High resolution (up to the 2000 pix/line) gives us the possibility of using such an L-CRT projector in HDTV. The main characteristics of basic elements are given. It is possible to reach the L-CRT light output of 10⁴ lm.