The thesis consists of two subjects. First, the mixed binary-to-full binary converter is designed using the neural-net analog-to-digital converters as basic computing elements. This converter operates asyncronously and its conversion time is faster than that of other conventional converters. Second, two different neural-net associative memory models are proposed and simulated. One is the expansion of the single layer bit-significance optimization model to the 2-layer model. By computer simulation, the 2-layer model as well as the single layer model has stronger error correction capability for highly correlated images than for the uncorrelated images. And the 2-layer model has better associative performance than the single layer model. The other is the 2-layer bidirectional associative memory model with hidden representation optimization. By optimizing the randomly assigned hidden representations, the associative performance is enhanced dramatically.
