To provide a method and a device for discovering important weight of neural net and a storage medium with stored important weight discovery program for neural net, which can discover a numerical rule with a clear important weight even when plural unwanted weights are included in a neural net.
The weight of the neural net and a penalty coefficient for every weight are initialized, a weight for minimizing a learning target function with square value penalty term is found while using a secondary learning method, a weight for verifying a crossing and updating the penalty coefficient is found while using the secondary learning method, a search direction for updating the penalty coefficient is analytically found from the provided weight on the basis of the theorem of a shadow function, a search width for updating the penalty coefficient is found on the base of a Gauss-Newton method, and the penalty coefficient is updated while using the secondary learning method.