Research on a self-built learning control mechanism in backpropagation training: convergence improvement.

概要

論文の詳細を見る
The main topic of this paper is the use of the self built learning control mechanism in backpropagation training, under the conditions encountered in processing remote sensing data. The approach based on the self-built learning control mechanism of controlling the learning rates and tuning of momentum term at the same time automatically in training of the feedforward neural network. An important feature of our novel method is that the initial learning parameters are not crucial to the success of the training, because the learning parameters are controlled that does not cause instability. Furthermore, the automatic initials learning rate selection and the optimal values of the learning control parameters were experimentally discussed. The proposed method was verified by a land cover study over Cianjur area, West-Java, Indonesia, with MOS-1 multi-spectral imagery. The simulations were presented to indicate the remarkable advantages of the proposed approach in both convergence rate and time saving.
社団法人日本写真測量学会の論文