SOM NEURAL NETWORK LEARNING CONTROL DEVICE

摘要

PPROBLEM TO BE SOLVED: To enable quick convergence of a SOM (Self-Organizing Maps) model by suppressing inconsistencies and, further by resolving inconsistencies in an early stage. PSOLUTION: A SOM neural network learning control device for causing a grid-like learning elements to learn from input data according to weights employs an asymmetric neighborhood function for deciding the weights. The direction of asymmetry of the asymmetric neighborhood function is reversed appropriately. Further, the degree of asymmetry is reduced while the asymmetry direction is reversed. Even if inconsistencies appear, convergence is obtained in learning cycles proportional to the square of system size (number of nodes). PCOPYRIGHT: (C)2007,JPO&INPIT