Thank you, John (Caulfield), Dr. Coffey, Dr. Potter, Ladies and Gentlemen. Having the opportunity to address the world's most innovative scientists in advanced computing is an honor, especially since the explosion in computing and data processing power has become modern folklore. Of all disciplines, computing has been blessed with the highest degree of innovation - and it's a good thing too because otherwise there might be building-sized slide rules with all sorts of crane-sized active control mechanisms manipulating the center bar. In fact, if not for innovation this could have very well been an SPIE Special Institute on the Structural Dynamics of Large Scale Calculating Machines. Sounds a little ridiculous, doesn't it? But then you don't suffer from the intellectual battle fatigue that forces most to hypothesize from the dark corners of "conventional wisdom." By definition, "conventional wisdom" does not spawn innovation and is inconsistent with the fact that major advances are frequently made by those who are exploring the challenge of their own far out ideas, rather than by those who have been directed to seek goals dictated by learned committees of "problem solvers." Most breakthroughs spring from the inspiration of a hunch rather than organized theoretical understanding and detailed systematic measurement programs. Sure - there will be failures but any innovator knows that success is built upon those failures.

Many future systems will need a computer incorporating good features of different architectures. Such a composite computer would divide a calculation into a number of subtasks to be executed on the most suitable subsystem. To date, relatively little experience has been gained in the design and use of such systems or even in interfacing different high performance computer architectures. It has been difficult enough to get one system working. Nevertheless complexities of modern problems force us to pull the advances made in signal processing, computer science, and large-scale scientific computing back together. We are assembling a production-oriented, Heterogeneous Element Supercomputer System (HESS) which incorporates high performance computers of different architectures to provide optimum overall performance on critical, computation-intensive problems facing the defense community. The system being assembled includes the Cray X-MP/12, the Laboratory for Computational Physics' Graphical and Array Processing System (GAPS), the NRL Network augmented by HUBNET, a multi-megabyte/sec fiber optics network connecting the main HESS participants, and additional highly parallel supercomputing elements to be installed for prototype use and evaluation.

A new class of computer architectures is emerging, driven mostly by the requirements of symbolic computing (artificial intelligence). They may be succinctly described as fine-grained, tightly-coupled multiprocessor systems. This means that the architectures are composed of a large number of relatively noncomplex processors that are coupled in such a manner that any given processor has direct communication with many other processors. Such machines are more readily adaptable to the data element structures associated with symbolic computing than are the sequential machines available today. It is important that the optical processing/computing community begin to investigate the role of optics in fine-grained, tightly-coupled multiprocessor systems.

Research efforts for optical computing devices, systems, and architectures in Japan are reviewed with an emphasis on recent developments. Activities pertaining to optical computing research in Japan are also reviewed.

The Fresnel transform provides a means to implement circuit switching networks by using optical techniques. Since photons can cross paths without interaction, non-blocking configurations are easy to achieve. Furthermore, any arbitrary interconnections between two N-port systems can be configured; the network can be rapidly reconfigured as needed due to the dynamic nature of acousto-optic cells.

Basic concepts of Fourier optics are presented, including the complex phasor representation of light waves, wave propagation in the Fresnel regime, coherent and incoherent imaging and spatial filtering, and schemes for the optical processing of 1-D signal information.

Correlation, convolution, and frequency filtering of images or signals may be achieved in noncoherent light using the optical transfer function (OTF) of an imaging system. Bipolar or complex-valued data and filter functions are represented in light intensity by the phase of a spatial frequency or temporal frequency carrier. This permits diffraction limited processing of complex-valued data which is input and output as patterns of light intensity.

Coherent processors, offer a myriad of complicated optical signal processings. However, coherent processing systems are usually plagued with coherent noise. These difficulties have prompted us to look at optical processing from a new standpoint. To consider whether it is necessary for all optical processing operations to be carried out by coherent light, we have found that there are several optical processing techniques which can be carried out by partially coherent light or white-light. One utilizes a spatially incoherent light to perform complex data processing, which has been proposed by Lohmann [1], Rhodes [2], and Stoner [3] and the other utilizes a white-light source, pursued by Leith and Roth [4], Yu [5], and Morris and George [6].

Analog optics is very fast but not very accurate. Digital electronics is much slower but much more accurate. Compromise (hybrid) systems appear to be intermediate in both speed and accuracy. As there are cases in which analog optics is too inaccurate and digital electronics is too slow; hybrid processors may have an important role to play.

Matrix algebra provides a mathematical language into which different classes of problems can be formulated in a consistent manner. These problems include those encountered in signal and image processing and numeric as well as symbolic computing. The fundamental operations of matrix algebra involve the arithmetic operations of multiplication and addition/subtraction along with global interconnections between one- or two-dimensional arrays of numbers. Both of these characteristics match the advantages provided by an optical system. A hierarchical view of the matrix operations is given and different optical architectures for implementing the basic operations of matrix algebra are surveyed.

We sometimes wish to undo what has been done to optical data using repeatedly identical nonlinear optical processors, such as non-linear imaging devices, matrix-vector multipliers with threshold logic, etc. This undoing is mathematically, equivalent to finding the inverse operator D-1, if the direct operator D represents the nonlinear optical processor in repeated usage.

A novel optical device for nonlinear signal processing is described based upon the following observations: (a) A phase space for signal processing is identified with a time-frequency joint representation (TFJR) that appears almost everywhere naturally, for example in bats, in music, etc. (b) A sudden slow down mechanism is responsible for the transition from a phase coherent-to-incoherent wavefront and provides us the sharpest tone transduction from a Bekesy traveling wave in a model of the inner ear. The cause of the slowdown is physically identified to be due to three forces. This has been used to derive a cubic deceleration polynomial responsible for a cusp bifurcation phenomenon which occur for every tone transducted along the nonuniform elastic membrane. The liquid-filled inner ear cochlea channel is divided by the membrane into an upper duct that has hair cells for the forward sound-generated flow and the lower duct for the backward balance-return flow.

We review the physical principles underlying logic decision making and communication for both optics and electronics. Research directions towards demonstration machines are indicated using the experimentally proven properties of bistable memories, logic gates and optical transphasors. The demonstration of restoring optical logic, a 'lock and clock' architecture and of a simple finite state machine is described.

Nonlinear optics can contribute decisions to optical signal processing and computing. Etalons permit massive parallelism and global interconnectivity. GaAs etalons appear more attractive for real systems, but ZnS interference filters are convenient for simple demonstrations of logic operations and pattern recognition.

Spatial filtering is one of the main assets of optics for information processing. We use spatial filtering for performing many logic operations simultaneously. The input data are arranged in matrix form. The type of operation is usually homogeneous across the matrix. The input is characterized as diffracting, as scattering or as birefringent structure. Specific issues to be addressed here are: incoherent inputs; inhomogeneous programming; simultaneous performance of several operations on all data, for example XOR and AND as parts of a half-adder; utilization of polarization.

A general approach is described for optically implementing massively parallel logic. Cellular array logic and the cellular automaton theory are emphasized as a guiding principle in the design of optical parallel computers. These theoretical approaches are important clues to elucidate the general characteristics and the limitation of optical parallel logic. We propose the use of Minnick's cellular logic array, which is well known as a simple two-dimensional array of processor elements to synthesis arbitrary logic functions. With reference to the optical cellular logic architecture, a space-variant logic is proposed, that is, different logical operations are performed in parallel. This space-variant logic gate technique is related to the multiple instruction-stream multiple data-stream (MIMD) logic operation technique. The new method is the MIMD extension of Tanida and Ichioka's optical shadow-casting logic, which is based on a space-invariant or a SIMD (single instruction-stream multiple data-stream) logic gate array. Simple examples are demonstrated. An application of this method to the Minnick cellular architecture is discussed. Finally we will discuss the possibility of the Minnick cellular array to reconfigurable architectures and self-organizing architectures, which are implemented by changing the interconnection networks.

The potential of optics-based technology for performing the fundamental decision and interconnection operations required in any computing system is reviewed. Examples in which only interconnection operations are performed optically and in which both interconnection and decision operations are performed optically are discussed.

Future developments in molecular electronicsi-b not only offer the possibility of high density archival memories, 1015 to 1018 gates/cc, but also new routes to fabrication of high levels of parallel processors (> 106) and hence to new computer architectures. A central theme of molecular electronics is that information can be stored as conformational changes in chemical moieties or functional groups. Further, these functional units are chosen or designed so that their structure facilitates the storage of information via reversible conformational changes, either in bond distances or in bond angles, or both. In exploring possible switching and information storage mechanisms at the molecular-size level, it has become apparent that there are many analogues or alternatives possible for any logical function which might be desired. It is even more exciting to realize that some structural chemical units or configurations offer completely new functional or logical capabilities. The example offered below is the molecular analogue of the CASE statement in PASCAL (proposed by an NRL summer student employee7). As suggested in the title, one of the purposes of this article is to enhance the appreciation of the universality of the 'chemical' or 'molecular' systems to express logical functions. The literature on molecular electronic concepts is growing and some reviews are available1-4. Two Molecular Electronic Device (MED) workshops5-6 have been held in Washington, D.C. (1981 and 1983) and an International Symposium on Bioelectric and Molecular Electronic Devices 8 was held in Tokyo, 20-21 November 1985. Beyond the strong interest current in Japan9, interest is also developing in England and Soviet block11.

This paper discusses application of the electrically modulated and unusually strong Raman emitted light produced by an adsorbed monolayer of phthalocyanine molecules on silver electrode or silver bromide substrates and on neural membranes. The analysis of electronic energy levels in semiconducting silver bromide and the adsorbed phthalocyanine molecules suggests a lasing mechanism as a possible origin of the high enhancement factor in surface enhanced Raman scattering. Electrically modulated Raman scattering may be used as a carrier of information which is drawn fran the fast intramolecular electron transfer aN,the multiplicity of quantum wells in phthalocyanine molecules. Fast switching times on the order of 10-13 seconds have been measured at room temperature. Multilevel and multioutput optical signals have also been obtained fran such an electrically modulated adsorbed monolayer of phthalocyanine molecules which can be precisely addressed and interrogated. This may be of practical use to develop Nlecular electronic devices with high density memory and fast parallel processing systems with a typical 1020 gate Hz/cm2 capacity at room temperature for use in optical computers. The paper also discusses the electrooptical modulation of Raman signals obtained from adsorbed bio-compatible phthalocyanine molecules on nerve membranes. This optical probe of neural systems can be used in studies of complex information processing in neural nets and provides a possible method for interfacing natural and man-made information processing devices.

This article describes models of associative pattern learning, adaptive pattern recognition, and parallel decision-making by neural networks. It is shown that a small set of real-time non-linear neural equations within a larger set of specialized neural circuits can be used to study a wide variety of such problems. Models of energy minimization, cooperative-competitive decision making, competitive learning, adaptive resonance, interactive activation, and back propagation are discussed and compared.

Information processing in future computers as well as in higher animals must refer to a complicated knowledge base which is somewhat vaguely called memory. Especially if one is dealing with natural data such as images and sounds, one has to realize the two aspects to be discussed: 1. The internal representations of sensory information in the computing networks. 2. The memory mechanism itself. Most of the experimental and theoretical works have concentrated on the latter problem, which might be named the "back-end" problem of memory. This paper contains some new results which show that both of the above functions, viz. formation of the internal representations and their storage, can be implemented simultaneously by an adaptive, massively parallel, self-organizing network.

We describe a parallel, distributed, associative model based on Hebbian modification of connection strengths between simple elements. Some extensions to the simple model are described and interpretations of the model as a gradient descent algorithm and as maximizing a probability density function are given. The model is applied to concept formation, since most versions of this modelling approach form equivalence classes of inputs that act like much like psychological concepts. Some computer simulations of concept-like behavior are described. Some kinds of computation can be performed effectively with these techniques.

The performance limits of optical, electro-optical, and electronic artificial neural systems (ANS) processors (also known as neurocomputers) are discussed. After a brief introduction, an overview is provided of the recently revived field of ANS. Next, ANS performance measures are defined and a neurocomputer taxonomy is presented. Finally, the designs and performance limits of the various types of neurocomputers are discussed.

A brief overview of background and developments in the emerging field of neural net models for optical computing is presented. At this early stage the field is offering a new and intellectually stimulating approach to signal processing that dove-tails with and compliments the capabilities of optics.

A single homogeneous layer of neural network is reviewed. For optical computing, a vector outer product model of neural network is fully explored and is characterized to be quasi-linear (QL). The relationships among the hetero-associative memory [AM], the ill-posed inverse association (solved by annealing algorithm Boltzmann machine (BM)), and the symmetric interconnect [T] of Hopfield's model E(N) are found by applying Wiener's criterion to the output feature f and setting [EQUATION].

In introduction, we have Sam Horvitz who is in charge of this panel and who will produce the report. Our purpose is to review the state of the art in neural network computing as well as the connection with optical computing in the future. I'm sure there is a lot of controversy and interesting projections of the future that will come out of this meeting. Here is Mr. Horvitz.

Neural processors are self-organizing Hamiltonian systems with adaptive energy functions, and are commonly referred to as "Neural Networks." They are characterized by a set of differential or difference equations, and can process information by means of their state response to initial or continuous input. Such processors consist of a large number of mutually interconnected nonlinear devices, appropriately called Processing elements or more prosaically referred to as "neurons." See references 1-3 for elaboration.

An introduction and comparative overview to the state of the art of two-dimensional spatial light modulator technology is provided, touching on the basic operation and performance of most of the more promising electronically- and optically- addressed device technologies. The fundamental functional capabilities and potential applications of these light control devices are also discussed, and some projections are offered on the future directions of spatial light modulator technology.

Electroabsorption and electrorefraction properties of bulk and multiple quantum well structures made fr III-V compound semiconductor materials are discussed in this paper. The large electroabsorption that has already been observed is suitable for spatial light modulation. The potential advantages for using these structures for spatial light modulation include high speed of response and monolithic integration with detectors and electronic devices on the same chip. Variations of the electro-optical properties such as absorption wavelength birefringence. nonlinearity and speed of response as a function of structural

The photorefractive effect in nonlinear crystals can be utilized to perform such elementary optical computing operations as image amplification, 2D optical logic, phase conjugation, spatial correlation and convolutions. The potential applications of these elementary optical computing operations to analog optical computing (e.g., matrix inversion, solving matrix eigenvector/eigenvalue problems), digital optical processing, pattern recognition and image processing are described.

A pipelined processor is characterized by the use of multiple processors, each programmed to perform a specific part of an overall computing task. The task is typically broken down into segments that are conveniently performed as a unit, but which must be completed before the next segment is started. Such computational units, placed sequentially so as to perform the computing task, comprise the pipeline. Data and input parameters are provided to each computing unit as needed, externally from the system memory or internally from the preceding unit. To help control the flow of data and results, the pipeline is often run synchronously, with one output from each unit occurring each i time units. Units which complete their tasks in less time must wait for the other units to finish. This favors the use of systolic pipelines, as proposed by Kungl.

A variety of integrated optic Bragg modulator modules that incoorporate an array of channel waveguides, a linear microlens array, an array of acoustooptic or electrooptic diffraction gratings, and an integrating lens in LiNb03 single-mode Channel-Planar composite waveguides with a substrate size of 0.2 x 1.0 x 1.8 cm3 have been realized. These integrated optic device modules were utilized successfully to perform a variety of experiments in computing, RF signal processing, and communications. These device modules together with various other device architectures should serve to facilitate the very large channel capacity that is inherent with the diode laser, the optical fiber, and the photodetector arrays for future multichannel optical computing as well as RF signal processing and communication systems.

There is a large potential payoff for successful application of rapid laser beamsteering techniques to optical media access. Optical disc storage provides a huge amount of memory; can this memory be utilized effectively for AI applications? The ultimate capability of an AI expert may be related to the amount of memory available. This paper describes the considerations which are likely to be important for implementation of rapid memory access techniques. Two issues are likely to be important. First, rapid access to the storage medium will have a strong influence on real-time performance, and second, automated learning and organization of knowledge will be important to get large amounts of information into the expert system. A review of laser beamsteering techniques suitable for random access to optical memory media is presented, grouped according to basic principles of operation. Photorefractive beamsteering, which may be particularly useful for optical computing applications, is discussed in detail. In the second part of the paper, methods for creation, organization, and utilization of very large rule bases are considered, and preliminary experiments in this area are presented.

Recent optical data processing (ODP) research at Carnegie-Mellon University (CMU) is reviewed. This includes pattern recognition work on feature extraction and correlation, optical linear algebra processing, and most recently optical symbolic, associative, and adaptive artificial intelligence (Al) optical processors.

A review of optical processing research performed in the Optical Systems Laboratory at Texas Tech University is presented. The major topic areas covered are (1) space-variant optical processing; (2) real-time electro-optical signal processing and (3) optical experiments for undergraduate engineering laboratories.

All known methods of computed tomography (CT) for image reconstruction from parallel projections have been derived from Fourier transform (the central slice theorem) and inverse Fourier transform, and the present Hankel transform tomography being an angular slice theorem is of no exception. Such a simplified viewpoint and unified theory is expected to serve the reader of optical computing and to pool interdisciplinary knowledge from a broader scientific community. The community can apply CT to optical computing or innovate novel approaches to solve the problem of real time and safe dosage computed tomography.

The alignment problem among an object o (7), the OFT lens and a phase hologram exp(iφ(r)) is investigated for holographic coordinate transformations, both theoretically and numerically, using FFT replacing OFT for two interesting examples in human visual systems. The optimum alignment is found to be at a specific saddle point r = r0 of the phase hologram defined by Δφ(r0)= k = 0 and Δ2φ = 0 where the OFT lens axis k = 0 (D.C.) and the object origin r = 0 must be overlayed. The result of such a D.C.-saddle point alignment can be understood in terms of the minimization of high order (cubic and above) contributions of the Fourier phase integral from a general viewpoint of a stationary phase approximation.