Continuous learning: a design methodology for fault-tolerant neural networks

Abstract

Fault tolerance in artificial neural networks is an important feature, in particular when the application is critical or when maintenance is difficult. This paper presents a general design methodology for designing fault-tolerant architectures, starting from the behavioral description of the nominal network and from the nominal algorithm. The behavioral level is considered to detect errors due to hardware faults, while system survival is guaranteed by the reactivation of learning mechanisms of the nominal network. An example of the use of this methodology is presented and evaluated.