适用于具有多分类器的卷积神经网络训练方法

Method for Training Convolution Neural Network With Multiple Classifiers

摘要: 为了提升视条件而定的深度卷积网络（conditional deep learning network，CDLN）的分类准确率，提出一种多分类器联合训练的方法.在训练网络时将多个分类器的输出误差同时进行反向传播，以校正网络权重.以LeNet-5、AlexNet为基础构造神经网络CDLN-L和CDLN-A，以MNIST、CIFAR-100和Pascal Voc数据集为实验样本进行实验，网络的分类准确率均得到提升，最高提升了4.39%.实验表明，提出的联合训练方法能够提升CDLN的分类准确率.

Abstract: To improve the classification accuracy of convolution neural network similar to conditional deep learning network (CDLN), a method of joint training with multiple classifiers was proposed in this paper. When training the network, all the error signals of the classifiers were applied to update weights by Back Propagation. In the experiments, CDLN-L and CDLN-A based on LeNet-5 and AlexNet were studied on the MINIST, CIFAR-100 and Pascal Voc databases, and an increase of 4.39% in classification accuracy was achieved. The experiments demonstrate that the proposed method can improve the accuracy of the network similar to CDLN.