A generalist-specialist paradigm for multilayer neural networks

A generalist-specialist paradigm for building neural networks is presented. The underlying principle is derived from an analogy with medicine. In most cases, a general practitioner or generalist is able to make a sure diagnosis: however, the generalist sometimes hesitates between several diagnoses and chooses to send his patient to see one or several specialists. This neural network architecture offers many advantages: improvement of the network performances, enhancement of the network interpretation, and selection of the critical frontiers between attraction valleys. The design of specialists is strongly dependent on the ambiguity threshold. A small threshold will lead to a rather simple network, whereas a larger value ensures a better robustness for the global network. This design methodology allows one to adjust the architecture´s complexity to the overall problem´s complexity. The authors present the definition of ambiguities matrices for the generalist networks, discuss methods for constructing specialists, and examine the functioning of the overall network