The Dynamics of Learning in some Digital Networks

This thesis is concerned with a study of learning in feedback networks of adaptive logic circuits. Random networks have been studied by various researchers, but previous work has not considered the adaptation mechanisms in dynamic logic networks which result from exposure to a non-random environment. Starting with a consideration of some limitations of a single-layer static network, the concept of a dynamic net (i.e. one with feedback connections) is introduced. The behaviour of the system is described in terms of its cycling activity in state space, and the effect of training on the state structure is considered. Subsequent experimental investigations consider unsupervised learning in the net where early evidence of a clustering effect is seen. This effect is found to be more pronounced when constraints are applied to the system in the sense that controlling gates are included in the feedback path. The nature and definition of memory and perception in such nets, and the response of the net to sequences of inputs is also presented and discussed. In conclusion, a simple probabilistic analysis is developed so as to provide a basis for a general understanding of dynamic networks of this kind.