摘要 |
PROBLEM TO BE SOLVED: To achieve precise and high speed learning of a neural network.SOLUTION: Processing in which, a path having the highest fidelity between a propagating signal and coupling load value, or the path and the paths disposed in the vicinity of the path, are selected, out of paths connected to respective neurons 51-54 of respective SOMs 41-43 forming respective layers 31-33 of the hierarchical neural network, and residual paths are set as sparses, is performed for the unit of the SOMs 41-43. An input signal inputted to the hierarchical neural network propagates in an order direction, an output signal is acquired, then, the acquired output signal and a target signal which is a set with the input signal are compared. Then, according to the matching degree, the fidelity between the signal propagating the selected path and coupling load value, is increased or decreased. |