TRAINING METHOD FOR FAULT TOLERANT FEEDFORWARF NEURAL NETWORK

概要

論文の詳細を見る
Feedforward neural network, trained with the backpropagation algorithm is widely used. The algorithm is based on the minimization of the mean squared error. Although it was thought that neural networks are fault tolerant as they have many parallel processing elements, recently extensive research has proved that neural networks are not intrinsically fault tolerant, And the fault tolerance ability have to be enhanced by adequate scheme. A connection link is called weak link if the stuck-at-0 fault of its weight causes an error at the network's output. In this paper it is assumed that the network's output is a logic value (0 or 1)and the output is faulty if the difference between the theoretical and the practical value is greater than a specified limit. The proposed training method here attempts to improve the weak links so that the network becomes fault tolerant for a stuck-at-0 of a single weight. The method to improve the weak links is presented in section 2. In section 3, the results obtained for character recognition applicatlon and XOR logic function are evaluated.
社団法人電子情報通信学会の論文
1995-09-05