Neuro-anatomical models of brain function have proved fertile ground in the development of efficient computational structures in recent times. Keynotes of these structures include a high degree of parallelism, distributed storage of information, robustness and fault- tolerance, and the usage of very simple processing elements individually performing tasks of low computational complexity. These characteristics of neural networks lend themselves to a plethora of engineering problems. Applications include (i) image and signal processing, and particularly inthe areas of robot and computer vision, and automatic speech generation and recognition, (ii) associative and content-addressable memories with applications to a variety of areas including expert systems utilizing large data bases, automated directory assistance, and coding and decoding systems, and (iii) adaptive systems capable of self-organization and "learning." These applications utilize fully the high degree of parallelism and the distributed format of inormation storage in these networks. Rapid convergence results typically, as a consequence of the highly parallel operation. In addition, the distributed storage of informationresults in a robust, fault-tolerant system. This is a particularly appealing feature for systems in hostile environments, and for applications where constant monitoring is unfeasible as in space-borne systems. Another feature useful in practice is the utilization of very simple computational elements as the basic component of these networks. This augurs well for the possibility of low cost, mass production of these networks for specific applications. The area has a particularly rich inter-disciplinary flavour. Engineering analysis, architecture design, and development of materials for these networks exist symbiotecally with investigations in artificial intelligence and parallel computing in computer science, spin glasses in statistical physics, and modelling of neural phenomena in neurobiology.
