多激活函数的立体神经网络设计  

Stereo Neural Network Design with Multiple Activation Functions

作　　者：王春波 果福明 WANG Chunbo;GUO Fuming(Department of Financial Information Engineering,Heilongjiang University of Finance and Economics,Harbin 150025,Heilongjiang)

机构地区：[1]黑龙江财经学院财经信息工程学院,黑龙江哈尔滨150025

出　　处：《济源职业技术学院学报》2023年第3期62-67,共6页Journal of Jiyuan Vocational and Technical College

基　　金：黑龙江财经学院校级科研课题(XJZD202308)。

摘　　要：合适的激活函数和参数可大幅提高神经网络预测的准确率,因此,工作人员会耗费大量时间和精力对激活函数进行多轮对比测试,通过评价指标的表现做出最终选择。为减少这种无效科研工作时间,建立了激活函数池,将传统神经网络进行叠加,形成了立体神经网络。该网络可对激活函数池中的多个函数同时进行训练,自动调整参数,通过ROC和AUC对各个维度的网络进行评价,选出主网络对未知数据进行预测。该立体神经网络可以帮助工作人员节省大约25%的手工比对和调整参数时间,预测准确率与使用正确激活函数的传统神经网络持平。Appropriate activation functions and parameters can significantly improve the accuracy of neural network predictions.Therefore,researchers and practitioners spend a considerable amount of time and effort conducting multiple rounds of comparative testing on activation functions,evaluating their performance using various metrics,and making the final selection.To reduce the ineffective research workload,the activation function pool is established,and the traditional neural network is superimposed to form a multi-dimensional neural network.This network can simultaneously train multiple activation functions from the activation function pool,automatically adjusting parameters.Evaluation of the networks across various dimensions is performed using ROC and AUC,and the main network is selected to predict the unknown data.This multi-dimensional neural network can help researchers save approximately 25%of the time spent on manual comparison and parameter adjusting.Furthermore,its predictive accuracy is on par with traditional neural networks that use the correct activation functions.

关 键 词：人工神经网络 激活函数 评价指标 ROC AUC 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]