Neural-Network Model of Induction Processes in the Way of Thinking by Using Random Excitation Patterns of Neural Elements
Masahiro Agu, Kazuo Yamanaka · Japanese Journal of Applied Physics · 1990
Based on the viewpoint that inductive prediction is regarded as an evolution process of information in the brain, the process of induction is modeled as a random selection procedure of plausible hypothesis under given data, by using random excitation patterns of probabilistic neural elements. The joint probability P(H, D) of hypothesis H and datum D is realized as the statistical distribution of the fluctuating excitation patterns of the neural network. The posterior probability P(H|D) can be estimated by selecting the fluctuations consistent with the observed datum D among those excited spontaneousely in the network. The estimation of a prototype pattern from its perturbed or incomplete patterns is treated as a typical induction problem, whose simple computer simulatin is given.