A fault injection approach for multiple-weight-fault tolerance of multi-layered neural networks

摘要

The methods for making multi-layered neural networks fault-tolerant by injecting intentionally the snapping of a link in the learning process have been studied in the literature. However, many of them considered the fault-tolerance only to the snapping of links. This paper is an additional report on fault-tolerance for multiple weight faults of links which Takanami et al. have already reported. A simple pattern recognition problem is used let a learning object. First, it is reconfirmed that fault-tolerance is obtained for the single and double weight faults. Next, it is shown that the fault tolerance for triple weight faults is also obtained. Next, the internal configuration of the network which has become fault-tolerant is analyzed by the covariance between inputs of output neurons. It is shown that the distribution of the covariances may become the measure of the degree of fault tolerance.