Programmable Analog Computer

The objective is to build and evaluate a novel neural computer architecture for tasks such as the real-time analysis of the output of large numbers of sensors tracking many objects simultaneously. This problem is analogous to the identification of aspecific event signature embedded in the vast quantity generated by a typical high-energy particle research facility, which represents one of the frontiers of neural network research. It is generally recognized that the power of neural networks increasesas the size of the network increases. Attempts to implement in very large (>103 nodes) practical hardware neural networks capable of very demanding pattern recognition functionality have been limited, and therefore these networks represent a largelyunknown opportunity. The scalability of the simplest implementation described therein, nothing more than a complex transistor interconnected by for example a matrix of polysilicon, lends itself to current large-scale integration techniques. Using today'sfabrication techniques it wouldbe possible to construct networks upwards of 106 nodes fully interconnected as a single chip, the chip comprising no more than a single layer of the actual junctions and topped by two layers of supply interconnections.As this represents a novel neural computing architecture, it should be emphasized that the ability to implement networks of current mirror elements in a highly interconnected fashion to a density exceeding 106, represents a significant technologicalbreakthrough in networks that rival those found living entities.Some of the projected applications include speech recognition, missile identification and other military applications involving complex pattern recognition, sensor interpretation using electromagnetic sensors arrays directly coupled to extensive,multi-layered networks, and image processing for recognizing potentially cancerous cell from ultrasound images, for example.