FPGA-based RBF neural network activation function implementation method

Abstract

The invention discloses an FPGA-based RBF neural network activation function implementation method in the technical field of machine learning and intelligent control, and aims to solve the technical problems that in the prior art, a table look-up method for implementing an activation function in an FPGA consumes more memory resources, and a piecewise function approximation method is not high in calculation precision. The method comprises the following steps: dividing an independent variable domain of an activation function into a core interval and an edge interval by taking a function inflection point as a demarcation point according to geometrical characteristics of the activation function; subdividing the edge interval into at least two sub-intervals, and approximating the activation function of each sub-interval by adopting a piecewise function to obtain a fitting function of each sub-interval; performing hardware language description on the calculation execution process of the activation function on the FPGA device; based on an FPGA device subjected to hardware language description, using a hyperbolic coordinate rotation algorithm to calculate an activation function of a core interval, and using a fitting function to calculate an activation function of an edge interval.

Description

technical field [0001] The invention relates to a method for realizing an activation function of an FPGA-based RBF neural network, belonging to the technical field of machine learning and intelligent control. Background technique [0002] With the rapid development of artificial intelligence technology, intelligent information processing and intelligent real-time control systems represented by fault prediction, fault diagnosis, and fault prevention are more and more applied to actual projects. In order to obtain good control performance and meet engineering real-time control and high-performance operation, nonlinear systems also need the support of intelligent algorithms such as artificial neural network and fuzzy control. Most of the intelligent algorithms contain complex exponential functions, which must not only meet the requirements for data processing speed, but also ensure the accuracy of calculations. FPGA (Field Programmable Gate Array) is a product further develope...

AI Technical Summary

Problems solved by technology

[0004] For the deficiencies in the prior art, the purpose of the present invention is to provide a method for implementing an FPGA-based RBF neural network activation function, to solve the problem of utilizing the table look-up method in the prior art to realize the activation function in the FPGA that consumes many memory resources and utilizes segmentation The function approximation method realizes the technical problem that the calculation accuracy of the activation function is not high in FPGA

Images