Remez algorithm based Sigmoid function hardware high-efficiency implementation method

Abstract

The invention discloses a remez algorithm based Sigmoid function hardware high-efficiency implementation method. The remez algorithm based Sigmoid function hardware high-efficiency implementation method is characterized in that: the Remez optimal consistency algorithm is firstly used to implement polynomial piecewise approximation on the Sigmoid function for four times to improve the fitting precision; and then the pipeline architecture optimization hardness is used to implement a module. The method effectively balances the contradiction between the resource consumption and the fitting precision, meets the real-time requirement, and meets different design indexes through the relatively strong flexibility.

Description

technical field [0001] The invention relates to the field of software and hardware collaborative programming for function realization, in particular to a high-efficiency hardware realization method of a Sigmoid function based on the Remez algorithm. Background technique [0002] According to the artificial neural network structure and mathematical model, the artificial neural network can be expressed as a quadruple [V, E, f, s], where V is the set of neuron nodes, and E is the set of connections between neurons , f is the activation function, and s is the learning algorithm. As the most mature and widely used artificial neural network, the most important feature of the error backpropagation (BP) artificial neural network is that it can simulate any given nonlinear function, and the nonlinear processing ability is mainly through the activation function of neurons. to realise. The nonlinear processing ability of artificial neural network largely depends on the characteristic...

AI Technical Summary

Problems solved by technology

[0012] In addition, there is the STAM algorithm, whose basic idea is to divide a lookup table in the original direct lookup table into multiple lookup tables according to predetermined rules, and then add and sum the outputs of multiple lookup tables to obtain the function value. Takes up much less storage than a direct lookup table, but is slower than a direct lookup table

[0013] Since the data format is related to precision and resource consumption, most of the existing technologies choose a custom data format to reduce logic resource consumption while ensuring fitting accuracy. In engineering applications, the 32-bit single-precision floating point of the IEEE754 standard is commonly used. Format, so that when the module designed by this method is operated with other modules, it needs to convert the data format, and the communication cost is relatively high; if the single-precision floating-point format is adopted, it is difficult to meet the approximation accuracy requirements by using the above existing solutions

Images