FPGA-Tabanlı Yapay Sinir Ağı Uygulamaları İçin Yüksek Hızlı TanSig Aktivasyon Fonksiyonunun Gerçeklenmesi

Abstract

In recent years, Artificial Neural Networks (ANNs) are used from automotive sector to defense industry and from engineering to banking in many areas. Transfer/activation functions used in ANNS are divided into two sections as linear and non-linear. Tangent Sigmoid (TanSig) activation function is one of nonlinear activation functions. Hardware implementation of TanSig activation function because this function includes exponential operations is very difficult. Therefore, diverse approaches have been presented for hardware implementation of TanSig activation function. In this study, TanSig activation function has been modeled with Kwan approach as hardware on FPGA chips. In the design, 32-bit IEEE-754-1985 floating point number standard has been used and the design has been coded in VHDL. The maximum operation frequency of the designed FPGA-based Kwan TanSig AF unit is 397 MHz for VIRTEX-6 FPGA chip using Xilinx ISE 14.1 Project Navigator program. FPGA-based Kwan TanSig AF unit can produce approximately 397 million results in 1 second. In this study, Kwan TanSig activation function has been shown to be used in the FPGA-based embedded ANN applications. In the future studies, embedded system applications can be performed using FPGA-based Kwan TanSig activation function with ANNs.