A Learning System with the Ability to Grasp Its Situation

Abstract

This paper reports a machine that is able to respond correctly to inputs having multiple meanings by grasping its situation. Unlike the conventional learning systems, the machine models the context relation between input patterns by a learning process. The machine can grasp its situation by the use of the modeled context relation, and hence it can respond correctly. To be concrete, the machine can correctly recognize patterns having multiple meanings such as 0 (``oh'' or ``zero'') as the situation may demand. The machine is composed of the backward coupling of the same kind of neuronlike elements called basic computational elements. The context relation between two input patterns is given by the coupling coefficient of the two basic computational elements correlated to the two patterns, respectively. Coupling coefficients make potentials of the elements correlated to the present inputs higher than the other elements. These higher potentials enable the machine to grasp the situation. Stability analysis is also made, and conditions for the machine to operate correctly are definitely shown under some assumptions. The machine was simulated on a computer, and several input patterns each of which had two meanings were recognized correctly.