Fault models for artificial neural networks

The author describes a method by which fault models can be developed for neural networks visualized at the abstract level, thus allowing their inherent fault tolerance to be probed. The derivation of such fault models has two stages: the location of where faults can occur, and the definition of the faults´ characteristics. As an example, a fault model for the multilayer perceptron neural network model is developed for each stage. The abstract nature of such fault models increases the possibility of their being generic in nature due to the independence of implementation. Also, they will allow the inherent fault tolerance of a neural network to be constructively and realistically investigated