一种神经网络的可配置激活函数模块设计

Implementation of Configurable Activation Function Module Based on Neural Network

神经网络是机器学习领域的一个重要分支,因其优秀的性能在很多基于嵌入式设备的应用中大放异彩。尽管已经有很多研究已经成功地将神经网络算法移植到小型化平台中,但是大多数的方法依然是基于某个固定模型的设计。本文提出了一种可配置的激活函数模块,比较了不同拟合方法的误差和耗费的资源,选择使用分段四阶多项式拟合激活函数的方法,拟合了6种常用激活函数,可以满足绝大多数神经网络的激活部分。这样在设计可配置神经网络的芯片时,通过配置该模块即可满足不同神经网络的需求。

Neural network is an important branch of machine learning,it is used in many embedded applications.Although many studies have successfully transplanted neural network algorithms into miniaturized platforms,most of them are still based on a fixed model.In the paper a configurable activation function module is proposed,compares the error and resource cost of different fitting methods,chooses the method of piecewise fourth-order polynomial fitting activation function,and fitting six commonly used activation functions.It can satisfy the activation part of most neural networks.In this way,when designing chips that can be configured with neural network,the modules can meet the requirements of different neural networks.