Improving generalizing capability of connectionist model through emergent dynamic behavior

An attempt is made to couple a connectionist model with a nonlinear dynamic system. To construct a closed system with two multilayer perceptrons, three types of cognitive mechanisms are built into it: supervised learning, the DREAM phase, and the EXPERIENCE phase. The EXPERIENCE phrase is the same as that of the self-supervised perceptron. In the DREAM phrase, the stimulus as the combination of taught conceptions comes from the interior of the net. For a cognitive system, it is considered that any one of these mechanisms is not complete. In the authors´ test, the generalizing capability is improved dramatically. The recognition of untaught samples is shown to be exact, the evolution process of the net exhibits complex behavior, and the internal representation plays an important role