Fault tolerance in neural networks
Hamed Elsimary, S. Mashally, S. Shahine · 2003
In the hardware realization of neural networks, some physical faults arise during the fabrication that can cause the network to fail. The implementation method will determine the type of fault that will occur. This assertion is examined using a backpropagation (BP) network model for pattern recognition of a subset of the English alphabet, simulated faults in the neurons, and synaptic connections between pairs of neurons. Observing the effects on the system performance when retrieving patterns, it was found that neural networks are tolerant against noise in input pattern and neural networks are tolerant against damage or faults in synaptic links or even neurons. The simulation results showed that, as the probability of damage increases, the network still produces correct results up to a certain limit. This BP model and the Hopfield model were compared for the same application (pattern recognition), which showed that the BP model is better in the sense that the number of nodes is not varied with the number of example patterns as in the case of the Hopfield model.>