This is almost verbatim the text of the conference presented in plenary session PLO6 at the E.C.O. 2 symposium in Paris on april 27, 1989.) I would like to devote this talk to a tutorial review of the domain of optical computing as seen from the viewpoint of optical interconnects. We shall start with a short critical assessment of the state of optical computing, showing that optical interconnects are an important part of it. We shall then review the technologies and the families of components available and whose continuing development appears necessary in this context. The question of interconnects in microelectronics will then be examined, clearly supporting the assertion that optics has a part to play there. We shall close with a review of the fundamental advantages that can reasonably be expected from the introduction of optical interconnects in electronic computers. The general ideas will be illustrated whenever possible by recent examples from the litterature or from projects presently under investigation in our Institute in Orsay. To a large extent, this presentation will elaborate on work inspired by the pionneering 1984 article by Goodman, Leonberger, Kung and Athale.

We are investigating an opto-electronic implementation of a trainable pattern classification system based on a feed-forward neural network model. An architecture with two layers of interconnections is used to transform a large amount of scene information to a small feature space that is, in turn, transformed into classification data. By using two layers of interconnections the number of large inner products that must be calculated may be significantly reduced. Simulations have been performed on a digital computer that demonstrate the performance of a system for the rotation-invariant classification of printed characters. A possible optical implementation is outlined.

In this paper, we consider the use of optical neural networks to approach solutions of optimization problems. A computer neural simulation of the Traveling Salesman optimization Problem is presented. An optical system designed to achieve this task, using a synaptic matrix composed of NxN computer-generated holograms is described. We discuss how these systems could be extended for use in future machines.

Projection method formulations of the perceptron and Hopfield's associative content-addressable memory (ACAM) neural nets are presented. We show that the well-known single-layer perceptron learning algorithm can be formulated using the method of projections onto convex sets (POCS) and that its performance can be improved using this viewpoint. The operation of a modified, binary-valued Hopfield ACAM is shown to be equivalent to the method of generalized projections (GP). A direct extension of the binary-valued ACAM to the continuous-valued case lends itself to a POCS formu-lation.

In this paper, we present a hybrid optical neural network using a high resolution video monitor as a programmable associative memory. The superior resolution and the number of gray levels of the video monitor lead to the implementation of a larger number of neurons and a larger dynamic range. The system operates in a high speed asynchronous mode due to the parallel feedback loop. The programmability of the system permits the use of the Orthogonal Projection(OP) and the Multilevel Recognition (MR) algorithms to improve the error correction ability of the network. Using current integrated optics and electronic technology this optical neural system attains high learning and operational speed.

A multi-layer neural networks are proposed for the implementation of general logic operations. It is shown that such systems composed of simple identical neuronlike elements can perform any Boolean function of two, three or more variables. It has to be pointed out that the learning method and thresholding function for neuronlike element are identical with previously described for content addressable memory (CAM). Interesting aspect of doing logic operations on neural networks is the ability to accomplish precise logic operations such as and, or, xor etc. on highly corrupted data.

Optical systems with synthesized aperture may be considered as one of possible alternatives of large sized entrance aperture optical systems. This alternative of the problem solution determines general possibility of wave fronts optimum phasing within the synthesized aperture and wave fronts control under changing application conditions. However, in this case it is practically impossible to achieve complete filling of aperture. This results in necessity to determine dimensions of the synthesized aperture elements and their locations relatively common optical axis.

The spatial coordinates of detected photoevents in a given area convey information about the classical irradiance of the input scene. In this paper the effectiveness of photon-counting techniques for image recognition is discussed. A correlation signal is obtained by cross correlating a photon-limited input scene with a classical intensity reference function stored in computer memory. Laboratory experiments involving matched filtering, rotation-invariant image recognition, and image classification are reported. For many images it is found that only a sparse sampling of the input is required to obtain accurate recognition decisions, and the digital processing of the data is extremely efficient. Using available photon-counting detection systems, the total time required to detect, process, and make a recognition decision is typically on the order of tens of milliseconds. This work has obvious application in night vision, but it is also relevant to areas such as robot vision, vehicle guidance, radiological and nuclear imaging, and recognition of spectral signatures.

We are proposing a new transformation,based on computation of the sector area of a binary image. Transformation is insensitive to noise and also shift and rotation invariant. In this method lines of various positions and orientations intersect the area of interest. Each line divides the image into two portions (sectors), and the area of one of the sectors is assigned to that line. Digital realisation of the transform is time consuming - therefore we are proposing its optical realisation. Optical processor uses incoherent, white light and anamorphic optics.

We describe a training image correlation based pattern recognition system using a nonlinear processor to perform distortion invariant pattern recognition. The training image set is stored at the input plane of the optical processor side by side the input scene. The nonlinear processor operates by producing the joint power spectrum of the training image set and the input scene. A thresholding nonlinearity is applied to the joint power spectrum of the input scene and the reference images. For binary nonlinear optical processors, both the input signals and the joint power spectrum are binarized to only two values. Thus, a binary spatial light modulator can be used to implement the processor. The performance of the training image based nonlinear processor is compared to that of the conventional linear optical processors. For binary nonlinear optical processors, a single binary spatial light modulator can be used to read in sequentially the binarized input signals and the binarized joint power spectrum which results in a reduction in cost and complexity of the system.

We describe a multifunction nonlinear optical processor that can perform various types of operations such as image deconvolution and nonlinear correlation. This processor is joint power spectrum based and does not use any filters. In this technique both the input signal and the processing function are displayed side by side at the input plane of the processor. The joint power spectrum of the input signals is thresholded with varying nonlinearity to produce the specific operation. In image deconvolution the joint power spectrum is modified and hard clipped thresholded to remove the amplitude distortion effects and to restore the correct phase of the original image. An amplitude mask averaged over an ensemble of images is assigned to the thresholded power spectrum to provide the amplitude magnitude of the restored image. In optical correlation, the Fourier transform interference intensity is thresholded to provide higher correlation peak intensity and better defined correlation spot. We show that various types of correlation signals can be produced simply by varying the severity of the nonlinearity and without the need to synthesize the specific matched filter. For example, a phase-only correlation signal is produced by selecting the appropriate nonlinearity. An analysis of the nonlinear processor for image deconvolution is presented. Computer simulation of the proposed technique for a linearly smeared image is provided.

The use of holography in data processing for optical computing, pattern recognition, head up display, etc. induce a development of the methods for conceiving, realizing, and testing the holographic systems. The limitations of the mathematical models used up to now prevent any innovation in this field. As a first step to go beyond these limitations we propose : a) an alternative theory of diffraction by holograms, providing new methods for computing the grating parameters; b) applications to the conception of new devices for analysing or testing holograms, and of new holographic processors. The theory will be presented here only in the particular case of unslanted phase gratings without loss, but of course its domain of validity is larger. For the complete theory, see [1] (not yet published). Nevertheless, we do not restrict our study to the case of Bragg diffraction, as the classical coupled waves theories do, since our theory works independently of such an hypothesis.

A simple method for generating phase-only filters is proposed. The interference of the Fourier phase value with a tilted plane reference wave is done by computer calculation. This amplitude modulated transmittance function encodes the pure Fourier phase information of the object wave. It acts like an off-axis phase only filter. After the binarization step the filter function is recorded directly into the recording medium in final size using a laser scanner. This results in a binary point-orinted hologram. Compared with the conventional matched filter our binary filter can save the spatial bandwidth product. It further reduces the dynamic range of the Fourier spectrum, thus increasing the diffraction efficiency. We discuss the performces of these filters, present computer simulations and demonstrate experimental results.

Pure phase-only correlation method for pattern recognition is proposed. The digital analysis concerning the discrimination ability of this method is presented.

In pattern recognition and in optical metrology, optical transform systems have been widely applied. Their use is particularly appropriate when the object is detailed and the recognition depends upon features that can be coarsely sampled in the transform space. Now with the advent of neural-network software, it is shown that the major difficulty in applying this optoelectronic hybrid is overcome. Using the ring-wedge photodetector and neural-network software, we illustrate the classification technique using thumbprints. This is a problem of known difficulty to us. Instead of a 4 person-month effort to devise software for its solution, we find that a 4-hour effort is all that is required. Other experiments also discussed are the sorting of photographs of cats and dogs, particulate suspensions, and image quality of digital halftones. All of these are shown to be promising examples for the application of the ring-wedge detector and neural-network software.

Optical pattern recognition is an activity readily describable in information or entropy terms. In that respect, for example, target discrimination can be done with rather low energy if we had the wisdom to design precisely the proper measurements. We show that the traditional Fourier optical pattern recognition apparatus is capable of achieving such low energy operation automatically. Viewed in another way, we can say that the energy per calculation in this (and all other coherent optical processors) can be significantly less than the "thermal limit" of kT. While digital computer theorists have shown that they can also beat the kT limit, digital computers now in use operate at about 108 kT. A comparison of coherent optical versus digital low energy schemes will show why coherent optics works so well.

Investigation of using phase conjugate technique to remove phase distortion in a joint transform correlator (JTC) is presented. For improving the accuracy of detection, a phase encoding method based on the nonlinearity of phase conjugation is also proposed. Experimental results and computer simulations are provided.

A method for the computer holographic codification of composite real filters is presented. This is especially useful for the optical implementation of the correlation in pattern recognition using the joint-transform scheme. Numerical and experimental results of the performances of some of these filters are shown for simple binary objects and for more realistic grey tone ones.

A wavelength-multichannel correlation technique based on matched filtering is proposed to identify a colour object with a non uniform colour distribution. The filters are matched to the object which is illuminated with different wavelengths. This correlation gives the information of colour distribution in addition to other information usually obtained from image recognition. The theoretical study of this technique, numerical simulations and discussion are presented.

Two types of optical correlators have been built to investigate real-time pattern recognition. The first employs one-dimensional devices to perform the two dimensional correlation in real time. This architecture uses an array of light emitting diodes (LED's) to input an electronically stored reference image into the processor in parallel. The input scene data is introduced into the processor one line at a time using an acousto-optic device (AOD). Multichannel time integrating correlations are performed in the row direction using the AOD and in the column direction using a charge coupled device (CCD) operating in the time delay and integrate mode. A processor has been built using this technology which correlates a 64 x 44 pixel binary reference image with a 256 x 232 input scene at video rates. The second correlator is a space integrating Fourier transform based correlator. A magneto optic-device (MOD) is used at the Fourier transform plane to rapidly change filter functions. The binary nature of the MOD device necessitates using either a binary phase or binary amplitude representation of the desired complex filter function. For this reason, several types of Binary Phase-Only Filter (BPOF) representations have been analyzed and experimentally investigated. Experimental correlation results have been obtained using both the Hartley BPOF and a newly developed class of complex binary filters, called Quad-Phase-Only Filters (QPOF). The performance of the two systems will be compared on the basis of processing speed, space bandwidth product, processor size and light efficiency. The inherent differences between incoherent and coherent processing and their implications for filter design will also be discussed. Finally, estimates of future performance will be presented.

We describe the architectural design of a compact optical processor implementing the symbolic substitution algorithm of hexagonal lattice gas automata. A ping pong architecture for iterative algorithm is given.

The multiwavelength laser transillumination is about to reveal itself as a very convenient mammography method to detect and localize suspicious areas in the breast. Primary experiments had clearly shown that the position of an heterogeneous zone relative to the detector greatly influences the response signal. To minimize its influence, we have combined a laser transillumination technique with an image reconstruction method. We did not scan point by point the studied medium, but directly transilluminated the medium along an axis. Then, the reconstruction is achieved from a set of projections, defined by 128 points, which also yields 128 x 128 pictures. Since photonic signals are used, it appears to be judicious to include an optical spatial filter in the Fourier plane. Because of the high scattering caracteristic of biological media, an incoherent to coherent convertion is necessary to obtain an optical Fourier transform of the projections. This is achieved by a liquid crystal light valve. The setup can be considered as an experimental approach of an 'optical scanner' when used in the field of biological imaging.

Real-time spatial domain image correlation and convolution were first demonstrated by White and Yariv using four-wave mixing in photorefractive nonlinear materials, but the full three channel input capability has not been exploited. We report experimental studies showing the application of the third input channel for composite pattern recognition and real-time spatial filtering of correlations or phase conjugate images. Potential applications are discussed for programmable image comparison, optical transformations to solve rotation and scale variance, and an all-optical cellular automaton computer.

A real and positive filter for pattern recognition is presented. The filter, based on the circular harmonic (CH) expansion of a real function, is partially rotation invariant. As it is real and positive, the filter can be recorded on a transparency as an amplitude filter. Computer simulations of character recognition show a partial rotation invariance of about 40°. Optical experiments agree with these results and with acceptable discrimination between different characters. Nevertheless, due to experimental difficulties, the method is onerous for use in general pattern recognition problems.

Through a slight modification of conventional monochromatic correlation, we formulate a color correlation which is capable of spectral and spatial recognition. Expressing RGB (red, green and blue) images, which are commonly used, in a vector form, the color correlation is decomposed of 9 terms, 3 intraband and 6 interband. To reduce the calculation time, we introduce preprocessings; projection onto appropriate plane and vector in RGB color space. The former one enables us to recognize an object with taking into account its both shape and color, and the latter one improves the sharpness of the cross-correlation peak. We also show the effectiveness of the color correlation applied to a quantitative measurement of stomach's surface from stereo endoscopic pictures.

The capability of Optics to carry out operations in parallel on large masses of data has fueled the current interest in many areas of Optics. However many of the Optical methods proposed have so far only been implemented in digital simulations. Some new or recently implemented methods for invariant pattern recognition will be reviewed, and the problems associated with implementing them with Optics will be discussed. Some of the methods discussed will include pattern recognition with advanced invariant matched filters designed to enhance classification performance and reduce sidelobes and the effects of noise and clutter; neural network content-addressable memories and symbolic substitution systems; and neural logic modules and their combinations with other neural networks to carry out fuzzy logic operations. The optical implementation of the above exemples and of some others will be discussed. In particular, a new representation of 2-D objects that represents any object by a small number of coefficients with invariant (rotation and scale) properties will be introduced, a method to realize it optically will be described, and its combination with a neural network to accomplish invariant pattern classification will be proposed.

This paper presents an efficient method for obtaining phase distribution of an object spectrum by using a fringe scanning phase conjugate interferometer. Detailed analysis of this proposed technique is provided and experimental demonstrations for verifying this technique are also given. The major advantages of this proposed technique are its potential high accuracy of the phase measurement and low space-bandwidth product requirement for the detection system.

The goal of cost effective, high performance optical systems capable of pattern matching tasks has fuelled research for over two decades. It is only recently with the introduction of novel optical/electronic architectures that such systems are becoming viable. Although conceptually elegant, the single output from an optical correlator is of restricted application due to its sensitivity to image distortions. A multichannel hybrid optical/electronic architecture has therefore been devised, exhibiting greatly enhanced flexibility and performance. This hybrid processor is being used in conjunction with a robotic handling system as part of an ESPRIT programme. The video signal from a robot's TV camera is relayed to an electrically addressed spatial light modulator in the processor. This input image is then replicated optically and fed into each of the system's parallel channels where it is cross-correlated with a set of differing reference patterns. The multiple correlation signals are detected and compiled into an electronic descriptor code. A 'Euclidean' search through a look-up table of codes then extracts object identification and rotation. Scale information and out-of-class recognition is also possible since the system can be taught to memorise new codes in real time.

It has been shown theoretically that optical correlation systems which use phase only filters are capable of producing sharp correlation peaks and good discrimination between different input objects, whilst at the same time having high optical efficiency. In addition to this, the resolution required in the filter plane is very much lower than that required for an equivalent conventional Van der Lugt correlator using a matched filter which has been made holographically. Here, we describe a flexible phase only filter implemented using a special-purpose liquid crystal spatial light modulator, and present results which demonstrate how it can be applied in optical correlation systems.

Specific optical architectures for performing logic and numeric functions using symbolic substitution are detailed. Optical laboratory results of the operations are presented and advanced considerations concerning these operations are discussed.