Experiments that reveal the limitations of the small initial weights and the importance of the modified neural model

Training of a perceptron that consist of McCulloch-Pitts neural model with a semi-linear transducer function, with a gradient based algorithm such as delta rule or generalised delta rule may suffer from saturation both at initialization and while training is in progress, hence network paralysis. A modified neural model has been proposed to resolve saturation. This paper furnishes further experimental results of this model using small initial weights and demonstrates the effectiveness of the modified neural model