Bases of a cast system for formal neural nets

Abstract

Formal neural nets considered here are McCulloch-Pitts type with interact ions of afferents, that is formal neurons capable of computing any logical function of the inputs. The main problem considered by the CAST system is that of network synthesis from the state transit ion matrix of a net. The system consists of three blocks: an input block, which accepts state transition matrix specification and provides the number of neurons and the logical functions to be performed by them. A second block provides, from the function of each neuron, the corresponding neural diagram, including the interaction of afferents. This interaction can be optimized to maximum reliability against changes in threshold. The third block generates the whole net including the effect of presynaptic and postsynaptic excitations and inhibitions. The system allows for quick investigation on the effect of synaptic changes upon global behaviour (i.e. state transition matrix) when, for example, the network is subjected to learning changes in pre- and postsynapses. Efferent interaction, which are necessary for arbitrary probabilistic nets, are not considered in the system, for the moment.