A generalized ABFT technique using a fault tolerant neural network

In this paper we first show that standard bp algorithm cannot yeild to a uniforminformation distribution over the neural network architecture. a measure of sensitivity isdefined to evaluate fault tolerance of neural network and then we show that the sensitivityof a link is closely related to the amount of information passes through it. based on thisassumption, we prove that the distribution of output error caused by s-a-0 (stuck at 0) faultsin a mlp network has a gaussian distribution function. udbp (uniformly distributedback propagation) algorithm is then introduced to minimize mean and variance of theoutput error. simulation results show that udbp has the least sensitivity and the highestfault tolerance among other algorithms such as wrta, n-ftbp and adp. then a mlpneural network trained with udbp, contributes in an algorithm based fault tolerant(abft) scheme to protect a nonlinear data process block. the neural network is trained toproduce an all zero syndrome sequence in the absence of any faults. a systematic realconvolution code guarantees that faults representing errors in the processed data will resultin notable nonzero values in syndrome sequence. a majority logic decoder can easily detectand correct single faults by observing the syndrome sequence. simulation resultsdemonstrating the error detection and correction behavior against random s-a-0 faults arepresented too.