Active Sites Implementation for the B-Matrix Neural Network

Abstract

This thesis is concerned with the problem of memory recall for the feedback neural network called the B-Matrix network. A new model -- the Active Sites model -- is proposed to reduce the computational complexity of the B-Matrix approach. We also develop a new delta learning rule that increases the memory retrieval capacity of the Active Sites model. These techniques are extended to a multi-level, non-binary neural network, which has a much higher capacity than the binary network. Through simulations and analysis, it is demonstrated that to retrieve a memory from a neural network, one does not need to have the whole memory and, as in real life, a fragment of that memory may be sufficient to recall it.