A learning strategy for multilayer neural network using discretized Sigmoidal function

In this paper a new approach for training the multilayer neural network (MNN), is proposed based on the scheme of discretization of the sigmoidal threshold activation function at regular intervals. The discretized values are stored in the form of a look up table (LUT). The weights are updated on a layer by layer basis using the values in the LUT. The proposed algorithm is applied to pattern classification problems and tested experimentally. It is found that the proposed algorithm takes lesser training time than the conventional backpropagation algorithm with continuous sigmoidal function. The proposed algorithm is more suitable for hardware implementation