VLSI Realization of Artificial Neural Networks with Improved Fault Tolerance

The feed forward neural network which is a model of the cerebral neural network has in-built fault tolerance. The conventional back-propagation algorithm reduces errors between the learning examples and the output of a multilayer neural network (MNN). However, it is not assured that the MNN behaves in the same manner when faults occur. For these reasons the study of fault tolerance in artificial neural networks (ANN) is valuable. The method proposed here improves the fault tolerance of the feedforward network to stuck-at-faults of weights by manipulating the activation function. Using this technique the average recognition rate is found to be 80%. This technique has the advantage that no extra hardware is required and that the complexity of the network is not increased. Also the computation time and learning cycles are reduced as there are no weight relevance calculations. This fault tolerance technique can also be suitably used for hardware implementation of artificial neural networks along with other redundancy methods. Keywords—Fault tolerance, Artificial Neural Network neural network, sigmoid function, stuck-at-faults, VLSI realization.