A global model for fault tolerance of feedforward neural networks

It is commonly assumed that neural networks have a built in fault tolerance property mainly due to their parallel structures. The international community of Neural Networks discussed these properties only until 1994 and afterwards the Subject has been mostly ignored. Recently the subject was again brought to discussion due to the possibility of using neural networks in nano-electronic systems where fault tolerance and graceful degradation properties would be very important. In spite of these two periods of work there is still need for a large discussion around the fault model for artificial neural networks that should be used. One of the most used models is based on the stuck at model but applied to the weights. This model does not cover all possible faults and a more general model should be found. The present paper proposes a model for the faults in hardware implementations of feedforward neural networks that is independent of the implementation chosen and covers more faults than all the models proposed before in the literature.