摘要
针对传统多层感知机(Multi-Layer Perceptron,MLP)模型在手写数字识别方面识别精度不高、识别效率较低的问题,提出改进的多层感知机模型,引入Dropout解决过拟合问题,Adagrad优化参数调试过程,ReLU解决梯度弥散问题,并在TensorFlow软件平台上实现该改进模型。实验表明,该改进的MLP模型能够有效地进行特征学习,在提高识别性能上表现优秀。与传统MLP算法模型相比,识别准确率提高了将近7.0%,识别效率提高了27.3s。
Aiming at the problem that the traditional MLP(Multi-Layer Perceptron)model has low recognition accuracy and low recognition efficiency in handwritten digit recognition,a modified multilayer perceptron model is proposed.And Dropout is introduced to solve the over-fitting problem,and by using Adagrad to optimize the parameter debugging process and ReLU to solve the gradient dispersion problem,the modified model is implemented on the TensorFlow software platform.Experiments show that the modified MLP model could effectively perform feature learning and perform better in improving recognition performance.Compared with the traditional MLP algorithm model,the recognition accuracy is greatly improved nearly by 7.0%higher;and the recognition efficiency improved by 27.3s.
作者
何平
刘紫燕
HE Ping;LIU Zi-yan(College of Big Data and Information Engineering,Guizhou University,Guiyang Guizhou 550025,China)
出处
《通信技术》
2018年第9期2075-2080,共6页
Communications Technology
基金
贵州省科学技术基金项目(黔科合基础[2016]1054)
贵州省联合资金项目(黔科合LH字[2017]7226号)
贵州大学2017年度学术新苗培养及创新探索专项(黔科合平台人才[2017]5788)~~
