Abstract:

Aiming at the problem of low-precision and high-cost of the traditional convolutional neural network when the training set and test set have different distribution, convolutional neural network algorithm combined with transfer learning model is proposed. Firstly, using the principal component analysis algorithm to conduct unsupervised dimension reduction of the source data, meanwhile, the auto-encoder algorithm is combined to reduce the dimension of the target dataset in order to make the source domain and the target dataset have the similar distribution of features in the low dimension. Secondly, according to the characteristics of the feature extraction of convolutional neural network, JS divergence is used to determine whether convolution pool layer can be migrated and initialized hidden layer is used to complete the model of trCNN. Finally, a small number of the marked target datasets are used to train to finish the building of the classification model. The design experiment verifies that the classification model can accurately perform the classification work with a small amount of annotation data.

Key words: principal component analysis, auto-encoder, convolutional neural network, transfer learning

摘要：

针对传统的卷积神经网络算法在训练集与测试集分布不同时分类精度较低且标注成本较高的问题，提出结合迁移学习模型的卷积神经网络算法。使用主成分分析算法对源域数据进行无监督降维，同时结合自编码机算法对目标数据集降维，使源域和目标数据集在低维度下具有相似的特征分布；根据卷积神经网络特征提取的特点，利用JS散度来判别卷积池层能否迁移，并使用初始化的隐藏层补全trCNN模型；使用少量带标注的目标数据集进行训练，完成分类模型的构建。设计实验验证分类模型能够在使用少量标注数据情况下准确地完成分类工作。

关键词: 主成分分析, 自编码机, 卷积神经网络, 迁移学习