Fault tolerance and robustness in neural networks

Abstract

Summary form only given. A framework by which the fault tolerance and robustness of neural networks can be assessed has been proposed. The possible effect on fault tolerance of various features of neural networks is discussed, as well as how to sensibly and realistically choose a method to assess the fault tolerance of a neural network. Advantages of systems employing neural networks with respect to error detection and recovery were considered. Also, the validity of applying conventional fault tolerance design methods is discussed. It is concluded that research should be aimed at abstract models rather titan at physical implementations. >