Reliability Characteristics of Neural Networks Having Faulty Interconnections

An artificial neural network is a system composed of massively parallel, highly interconnected simple processing units. Its real promise for applications lies in specialized hardware. Currently VLSI and optoelectronics are the two prominent techniques being investigated for network implementations. In real situations, it is common to see a tremendous number of simple interconnections, along with the processing units, implemented on a large network chip. As the area of a single chip increases, the number of potential defects increases. Neural networks are known to have the capability of fault tolerance, but to what degree can a network tolerate defects? How can one estimate the results quantitatively in advance? How should the network specifications be updated by taking the interconnection faults into account? All of these questions will be answered in this dissertation for Hebbian-type associative memories. Equations for estimating the probabilities of direct one-step convergence of Hebbian -type associative memories having faulty interconnections are derived in this dissertation. Rigorous analysis is performed on open-circuited and short-circuited interconnection faults. With error bits in the probe vectors, both synchronous and asynchronous networks are explored. Our theoretical results are validated with extensive Monte Carlo simulations. The potential for using a three-layered network, trained with the backpropagation algorithm, as an associative memory is also examined. Advantages and disadvantages of this network, compared with Hebbian-type associative memories, are considered.