Learning machine with multi-input single output circuits connected in hierarchical structure.

摘要

In a learning machine with multi-input single-output circuits (18, 19, 20) connected in a hierarchical structure which sums up weighted input signals and subjects the sum thus obtained to a non-linear processing to provide an output signals, each weight is changed (16, 65) to provide a desired value of each output signal in learning. Using, as a direction of changing the weight, a conjugate gradient direction is used in place of the most sudden drop direction which has been conventionally adopted can reduce the time required for the learning. Also dynamically setting, as a learning parameter, an optimum value in place of a fixed value which has been conventionally adopted can reduce the time required for learning. Moreover, in calculating errors for plural learning parameters relative to a certain weight changing direction, a product-sum of the input signals and weights in a secret layer (22) and a product-sum of the input signals and the weight changing direction in the secret layer are stored (77) so that the number of operations of acquiring outputs from the secret layer thereby to greatly reduce the time required for learning. Furthermore, if it is detected (92) that the learning has been fallen into an unsuitable state where further advancing the learning cannot effectively reduce the errors, the weight are adjusted to resume the learning thereby reducing the time required for learning.