Deep Adversarial Attention Alignment for Unsupervised Domain Adaptation: the Benefit of Target Expectation Maximization
share
In this paper we make two contributions to unsupervised domain adaptation in the convolutional neural network. First, our approach transfers knowledge in the deep side of neural networks for all convolutional layers. Previous methods usually do so by directly aligning higher-level representations, e.g., aligning the activations of fully-connected layers. In this case, although the convolutional layers can be modified through gradient back-propagation, but not significantly. Our approach takes advantage of the natural image correspondence built by CycleGAN. Departing from previous methods, we use every convolutional layer of the target network to uncover the knowledge shared by the source domain through an attention alignment mechanism. The discriminative part of an image is relatively insensitive to the change of image style, ensuring our attention alignment particularly suitable for robust knowledge adaptation. Second, we estimate the posterior label distribution of the unlabeled data to train the target network. Previous methods, which iteratively update the pseudo labels by the target network and refine the target network by the updated pseudo labels, are straightforward but vulnerable to noisy labels. Instead, our approach uses category distribution to calculate the cross-entropy loss for training, thereby ameliorating deviation accumulation. The two contributions make our approach outperform the state-of-theart methods by +2.6 six transfer tasks on Office- 31 on average. Notably, our approach yields +5.1 improvement for the challenging D → A task.
READ FULL TEXT