階層型ニューラルネットワークの初期値設定問題に関する仮説検定および写像の観点からの検討

概要

論文の詳細を見る
For the error back-propagation learning rule in the multi-layered neural networks, the usual way to set the initial values of the unit parameters such as weights and biases has been to give uniform random numbers of a certain range. The setting, however, results frequentry in local minimum to stop the learning and usually in slow convergency. To alleviate these problems, we propose a systematic initialization which sets biases to be zero and weight vectors of a hidden layer so as to be equal in length and uniform in distribution. In this sense the proposed initialization has a structual property, which is theoretically derived from hypothesis tests under no priori infomations By Monte Carlo simulations, the proposed initialization was confirmed to be valid for the convergency
近畿大学の論文
1991-12-01