Accelerating Neural Networks on FPGA via Tan-Sigmoid Approximation

Abstract

. This research examines the execution of feed-forward artificial neural networks on field-programmable gate arrays (FPGAs), addressing the challenge of limited hardware resources that restrict network size and hinder the direct execution of nonlinear activation functions. The proposed approach reduces resource usage through data normalization and restructuring, while approximating the tan-sigmoid function via a logarithmic method. Implemented on the Xilinx Spartan-3A (3sd3400afg676-4) platform, the design achieves a 3% optimization in slices and a 2% optimization in lookup tables (LUTs), demonstrating improved efficiency, enabling larger and more cost-effective neural networks on a single FPGA chip.