融合随机傅里叶特征的混合神经网络模型

Hybrid Neural Network Model Based on Random Fourier Features

深度神经网络模型作为目前最成功的学习模型之一,在模型训练时涉及大量的参数学习,特别对于较大规模的样本存在特征处理困难、模型结构复杂等问题.线性或亚线性时间复杂度的核近似算法具有计算开销小、易于扩展到大规模数据的优点.因此,利用线性核近似算法可以进行特征处理并降低深度神经网络模型的结构复杂程度.本文提出一种基于傅里叶特征空间的混合神经网络模型优化方法,并由此构建一个混合深度神经网络模型.该模型将基于随机傅里叶特征变换的浅层网络与卷积神经网络相结合,傅里叶层进行数据特征处理与提取的同时,降低混合网络模型的结构复杂程度,优化深度神经网络模型结构.实验结果表明本文提出的模型拥有较少的参数量和较低的浮点计算量,同时模型可保持较高的测试准确率以及更快的收敛效率.

As one of the most successful learning models at present,the deep neural network model involves a lot of parameter learning in model training,especially in large-scale samples with difficulties in feature processing and complex model structure.Kernel approximation algorithms with linear or sublinear time complexity have the advantages of low computational overhead and easy expansion to large-scale data.Therefore,linear kernel approximation algorithms can be used for feature processing and reduce the structure complexity of deep neural network models.This paper presents a hybrid neural network model optimization method based on Fourier feature space,from which a hybrid depth neural network model is constructed.The model combines shallow network based on random Fourier feature transformation with convolution neural network.The Fourier layer performs data feature processing and extraction while reducing the structure complexity of the hybrid network model and optimizing the structure of the deep neural network model.The experimental results show that the proposed model has fewer parameters and lower floating-point computation,while maintaining a high test accuracy and faster convergence efficiency.