A Fast BP Learning Algorithm via A Hybrid Approach

A hybrid approach based on changing the activation function and using a robust error estimator for improving the learning speed of BP network is suggested in this paper.The activation function used in this paper is a non-differentiable piecewise linear function and its derivative function is re-defined as a new continuous one.The network learning is proceeding under new error function.Significant improvement is observed in the simulations of the XOR and encoder/decode examples.