| 摘要 |
PROBLEM TO BE SOLVED: To provide a method for obtaining the structure of a neural network which is high in generalization capacity and simple in configurations.SOLUTION: A neural network optimization device is configured to give an initial structure of a neural network, and to perform the update of a weight with this initial structure as a first input, and to delete a unit at random from the neural network subjected to the update of the weight to generate a neural network with a new structure, and to repeat processing of performing the learning of the neural network with the new structure, and to reduce the costs of the neural network to be calculated by using evaluation data. The neural network optimization device is configured to, when a state that the costs of the neural network before the deletion of the unit are larger than the costs of the neural network subjected to the update of the weight after the deletion of the unit continuously occurs by a predetermined number of times, end the learning of the neural network, and to output the neural network before the deletion of the unit as the optimal structure. |
