A statistically resilient method of weight initialization for SFANN

Proper weight initialization is one of the important requirements for faster training in feedforward artificial neural networks. Conventionally, these weights are initialized to small uniformly distributed random values so as to break the symmetry of weights during training, that is allow the weights to acquire different values. In this work, we have proposed a new weight initialization technique (NWIT) for sigmoidal feedforward artificial neural networks. The proposed method NWIT ensures that the output of neurons are in the active region and the range of activation function is fully utilized. The proposed routine is compared with random weight initialization method for 11 function approximation task. The proposed method NWIT is as good as if not better when compared to random weight initialization technique (RWIT).