Fault tolerant design using error correcting code for multilayer neural networks

A new fault tolerant multilayer neural network (NN) which can correct an error caused by a fault in the output layer neuron is proposed. The principle of the design is that an error correcting code is used for the output space of NN, and NN learns this code in the training phase. Simulation experiments for some examples in relatively small pattern recognition models are examined. As a result, it can be concluded that SEC-DED codes together with some training method are sufficient for the examples, and the proposed design can be adapted effectively to practical applications