Aproximação da função tangente hiperbólica em hardware

Abstract

The ANN (artificial neural networks) are widely used in various applications in engineering, especially for troubleshooting nonlinear nature, classification and clustering, in pattern recognition, function approximation, among others. The implementation of RNA in reconfigurable devices is a major challenge, since many factors such as floating point precision, the tangent hiperbolic activation function and the area used in the FPGA are involved. This paper presents the implementation of various types of tangent hiperbolic activation function based LUT (Look-Up Tables) in order to assist in choosing the best approximation of nonlinear functions in hardware. It presents also a comparison of the results obtained with the literature currently available, and various implementation characteristics were analyzed as the area used in FPGA, and the error rate in memory access time by statistical tools CAD.

Index Terms - FPGA, sigmoid activation function, ANN