Informative pairs mining based adaptive metric learning for adversarial domain adaptation

Abstract

Adversarial domain adaptation has made remarkable in promoting feature transferability, while recent work reveals that there exists an unexpected degradation of feature discrimination during the procedure of learning transferable features. This paper proposes an informative pairs mining based adaptive metric learning (IPM-AML), where a novel two-triplet-sampling strategy is advanced to select informative positive pairs from the same classes and informative negative pairs from different classes, and a metric loss imposed with special weights is further utilized to adaptively pay more attention to those more informative pairs which can adaptively improve discrimination. Then, we incorporate IPM-AML into popular conditional domain adversarial network (CDAN) to learn feature representation that is transferable and discriminative desirably (IPM-AML-CDAN). To ensure the reliability of pseudo target labels in the whole training process, we select more confident target ones whose predicted scores are higher than a given threshold T, and also provide theoretical validation for this simple threshold strategy. Extensive experiment results on four cross-domain benchmarks validate that IPM-AML-CDAN can achieve competitive results compared with state-of-the-art approaches.