Incremental Dictionary Learning for Unsupervised Domain Adaptation

Domain adaptation (DA) methods attempt to solve the domain mismatch problem between source and target data. In this paper, we propose an incremental dictionary learning method where some target data called supportive samples are selected to assist adaptation. The idea is partially inspired by the bootstrapping-based methods [1, 3], which choose from the target domain some samples and add them into source domain for retraining the classifier. However, the suitable sample selection and stopping criteria for DA setting is a tricky problem. For the sample selection criteria, we choose supportive samples that are close to the source domain, so that they act as a bridge to connect the two domains and reduce the domain mismatch. More sepecifically, given the source dictionary D, we select the target samples that minimize the reconstruction error when represented by D. Then we augment the source domain by adding supportive samples and retrain the dictionary. For the stopping criteria, we guarantee that the domain mismatch decreases monotonically during adaptation. This is realized by checking whether adding new supportive samples will reduce the domain dissimilarity after each iteration. The proposed approach is shown in Fig. 1. Supportive Samples Selection: We select the supportive samples using W (k+1) by solving the following optimization problem: