Sigmoid Function Implementation Using the Unequal Segmentation of Differential Lookup Table and Second Order Nonlinear Function

Abstract

This paper discusses the artificial neural network (ANN) implementation into a field programmable gate array (FPGA). One of the most difficult problem encounters is the complex equation of the activation function namely sigmoid function. The sigmoid function is used as learning function to train the neural network while its derivative is used as a network activation function for specifying the point at which the network should switch to a true state. In order to overcome this problem, two-steps approach which combined the unequal segmentation of the differential look-up table (USdLUT) and the second order nonlinear function (SONF) is proposed. Based on the analysis done, the deviation achieved using the proposed method is 95%. The result obtained is much better than the previous implementation that uses equal segmentation of differential look-up table.