用于无监督域适应的深度对抗重构分类网络

Deep Adversarial-Reconstruction-Classification Networks for Unsupervised Domain Adaptation

最近迁移学习的新方法对抗域适应,将生成对抗网络(GAN)的思想添加到深度网络中,能够学习数据的可迁移表示形式进行域适应。虽然通过GAN的思想能够很好地提取出源域数据和目标域数据的共同特征,有效地进行不同域之间的知识迁移,但现有的对抗域适应算法不能有效地保留目标域数据的局部特征,而目标域数据的某些特征可能会对分类精度有显著的提升。为了避免原始数据的局部特征因对抗性学习遭到破坏,利用多任务神经网络来保留目标域数据的局部特征。提出了一个深度对抗重构分类网络的模型(DARCN)。DARCN受到自动编码器的启发,在对抗域适应的基础上,添加了自动编码器的解码部分,这样能够有效地从低维特征重建原始数据。该模型学习了以下任务的共享编码表示:带标签的源域数据的监督分类;不带标签的目标域数据的无监督重构;源域和目标域的不可区分性。最后,最小化标签分类器的分类损失和解码器的重构损失,同时最大化域判别器的分类损失,通过梯度下降法能够有效地解决此类优化问题。实验结果证明了目标域局部特征的保留对领域自适应任务是十分关键的。

Recently,a new method of transfer learning called adversarial domain adaptation,embeds the idea of the generative adversarial networks(GAN)into the deep networks.It can learn the transferable representation of data for domain adaptation by the thought of the GAN.Although this method can extract the common features of the source domain data and target domain data,and effectively transfer knowledge between different domains,the existing adversarial domain adaptation algorithms cannot effectively retain the local features of the target domain.However,some features of the target domain data may significantly improve the classification accuracy.In order to avoid the destruction of the local features of the original data due to adversarial learning,a multi-task neural network is used to retain the local features of the target domain data.A model of deep adversarial-reconstruction-classification networks(DARCN)is proposed.DARCN is inspired by the auto-encoder.On the basis of adversarial domain adaptation,the decoding part of the auto-encoder is added,which can effectively reconstruct the original data from low dimensional features.The model learns shared coding representations for the following tasks:supervised classification of labeled source domain data,unsupervised reconstruction of unlabeled target domain data and indistinguishability of source domain and target domain.Finally,the classification loss of the label classifier and the reconstruction loss of the decoder are minimized,and the classification loss of the domain discriminator is maximized.The gradient descent method can effectively solve such optimization problems.The experimental results prove that the preservation of local features of target domain is critical for domain adaptation tasks.