Implementation of High Speed Tangent Sigmoid Transfer Function Approximations for Artificial Neural Network Applications on FPGA

Abstract

Tangent Sigmoid (TanSig) Transfer Function (TSTF) is one of the nonlinear functions used in Artificial Neural Networks (ANNs). As TSTF includes exponential function operations, hardware-based implementation of this function is difficult. Thus, various methods have been proposed in the literature for the hardware implementation of TSTF. In this study, four different TSTF approaches on FPGA have been implemented using 32-bit IEEE 754–1985 floating point number standard, and their performance analyses and FPGA chip statistics are presented. The Van der Pol system ANN application was carried out using four different FPGA-based TSTF units presented. The Multilayer feed-forward neural network structure was used in the study. The FPGA chip statistics and sensitivity analyses were carried out by applying each TSTF structure to the exemplary ANN. The maximum operating frequency of ANNs designed on FPGA using the four different TSTF units varied between 184–362 MHz. The CORDIC-LUT-based ANN on FPGA was able to calculate 1 billion results in 3.284 s. According to the Van der Pol system ANN application carried out on FPGA, the CORDIC-LUT-based approach most closely reflected the reference ANN results. This study has a reference and key research for real-time artificial neural network applications used of tangent sigmoid one of the nonlinear transfer functions.

Keywords

• real-time systems
• field programmable gate arrays
• artificial neural networks
• approximation methods
• transfer functions