Mini-batch Dynamic Geometric Embedding for Unsupervised Domain Adaptation

Unsupervised domain adaptation has gotten a lot of attention due to its ability to improve learning performance in a target domain based on the knowledge extracted from a source domain. Recent studies show that graph-based models can accomplish good results for domain adaptation problems. However, most of these graph-based domain adaptation approaches cannot work in an end-to-end manner, leading to the limited scalable. To address this issue, we propose a learning method named Mini-batch Dynamic Geometric Embedding (MDGE), which seeks to find the relationship between batches source and target samples to learn discriminative representations. Specifically, to build a better graph representing sample relationship, we propose a class-specific sampling strategy to pick up samples which are then used as input of MDGE. Since the samples are effectively selected, we develop a method to dynamically build a subgraph that in turn supports the relationship update and helps the network backbone to extract more discriminative features. Comprehensive experiments on real-world visual datasets demonstrate the effectiveness of the proposed MDGE algorithm.