深度神经网络模型构建及优化策略

Model construction and optimization strategies of deep neural networks

针对深度学习构建网络模型以及确定模型参数的问题,在分析神经网络基本结构和线性模型局限性的基础上,研究了深度神经网络设计的关键因素和优化策略。结合手写数字识别问题,对优化策略、动态衰减学习率、隐藏层节点数、隐藏层数等情形下的识别正确率进行了实验。结果表明,不同神经网络模型对最终正确率有质的影响,相同优化策略在不同参数取值时对最终正确率有很大影响,并进一步探究了具体选取优化策略和参数的方法。

Aiming at the issues of building the network model and determining the model parameters in deep learning,on the basis of analyzing the basic structure of neural networks and the limitations of the linear model,the key factors and optimization strategies of designing deep learning neural networks are studied.Combined with the handwritten numeral recognition problem,a large number of experiments are carried out on the recognition accuracy under the conditions of optimization strategy,dynamic attenuation learning rate,number of hidden layer nodes and number of hidden layers.The results show that different neural network models have a qualitative impact on the final accuracy rate,and the same optimization strategy has a great impact on the final accuracy rate when different parameters are selected.Furthermore,the specific selection method of optimization strategy and parameters is explored.