Learning efficiency increasing method for neural net

Abstract

Abstract PURPOSE:To make efficient the learning of a neural net by altering an updated step width and the parameter value of a momentum so that an output error may obtain a minimum value in the repeating calculating process of the learning. CONSTITUTION:The particle of a present word shows a model to forecast the particle of a next word. A constitution is a multilayer network to be the object of back propagation algorithm, an input layer consists of 89 units, an intermediate consists of two layers of 16 units, an output layer consists of 89 units, and the number of times of sampling is 23 per one sentence. As parameters, respectively three types of 1/2 times, 1-time and 2-times of a last-time value are selected for a step width eta, and two types of 0 and 0.9 are selected for a momentum alpha. At the time of the initial learning or when the number of the samples is increased, alpha=0 is selected, and in the other cases, 0.9 is selected. Further, the eta decreases as the number of the samples increases. Thus, by automatically selecting the optimum parameter value for the quantity of the samples, and automatically regulating the parameters according to an advancing condition, the learning efficiency can be increased.

Inventors

• 👤 Masami Nakamura
• 👤 中村　雅己
• 👤 Kiyohiro Kano
• 👤 清宏 鹿野

Bibliographic Details