| 摘要 |
PROBLEM TO BE SOLVED: To actualize superior versatility which has a small arithmetic quantity, enables fast learning, and is nearly free of an overlearning state wherein learning is carried on after convergence. SOLUTION: This device has a complete distribution type output form multi- layered neural network 14 which has M tutor signal output elements for the number Q of category classifications for M which is power of 2 and has M output layer units made to learn by using a binary tutor signal in complete distribution type output form obtained by assigning all different categories to all states and making learning input data sets correspond to tutor signals, a versatility rate processing means 40 which conducts a test for calculating the versatility rate of a test input data set when initial learning is converged and calculates the versatility rate of the test input data set, and an optimum intermediate-layer unit number decision means 41 which detects deterioration in versatility rate and decides whether or not variation in the number of intermediate layer output units is accepted and the number of optimum intermediate layer units.
|
