Inherent fault tolerance analysis for a class of multi-layer neural networks with weight deviations

The general formula of computing the deviation of the output of a multilayer neural network (MLNN) with respect to the deviations of its input and of its weights is presented. The upper bound of the deviation propagation from level to level is well estimated with certain probability. Based on this, one can analyze the relation between the topological structure of an MLNN and its fault tolerance property, which can be used to correctly design fault tolerant MLNNs.<
>