FixBi: Bridging Domain Spaces for Unsupervised Domain Adaptation

Abstract

Unsupervised domain adaptation (UDA) methods for learning domain invariant representations have achieved remarkable progress. However, most of the studies were based on direct adaptation from the source domain to the target domain and have suffered from large domain discrepancies. In this paper, we propose a UDA method that effectively handles such large domain discrepancies. We introduce a fixed ratio-based mixup to augment multiple intermediate domains between the source and target domain. From the augmented-domains, we train the source-dominant model and the target-dominant model that have complementary characteristics. Using our confidencebased learning methodologies, e.g., bidirectional matching with high-confidence predictions and self-penalization using low-confidence predictions, the models can learn from each other or from its own results. Through our proposed methods, the models gradually transfer domain knowledge from the source to the target domain. Extensive experiments demonstrate the superiority of our proposed method on three public benchmarks: Office-31, Office-Home, and VisDA-2017. 1

Low-confidence Samples Matter for Domain Adaptation

“…To increase discriminability, more recent DA methods attempt to investigate the data structure in unlabeled target domain. Self-training as a typical approach generates target domain pseudo labels [12,26,29,36,39,43,47,49,58,89,92,96,97]. Another category is to construct the protoptypes [4,50,77,78,88,94] or cluster centers [10,24,66] across domains and then perform class-wise alignment.…”

“…Another category is to construct the protoptypes [4,50,77,78,88,94] or cluster centers [10,24,66] across domains and then perform class-wise alignment. Most of these approaches use a fixed probability threshold [12,36,50,58,78], a dynamic probability threshold [47,89], a fixed sample ratio [29,49,92], a dynamic sample ratio [58,96,97], or a threshold of other metrics [4,14,26,39] to choose trustworthy samples (i.e., highconfidence samples) and reject other low-confidence samples. To mitigate the harmful effect of noisy labels, it is reasonable to only utilize the reliable samples.…”

“…Some approaches try to add more constraints to lowconfidence samples. FixBi [47] proposes a self-penalization loss that enforces the maximum probability of the lowconfidence samples to be zero. Low-confidence samples are also taken into account by entropy minimization [14,43,55,72,80,91] and virtual adversarial training [8,61].…”

“…Mixup [87] and its variants [3,71,86] provide effective data augmentation strategies when paired with a crossentropy loss for supervised and semi-supervised learning. In domain adaptation, there are several methods [42,47,76,79,82,83] applying mixup. [76,79] combines the domainlevel mixup with domain adversarial training to learn a more continuous latent space across domains.…”

“…Consequently, the low-confidence samples are completely ignored in these methods. There are some approaches taking into account less reliable samples, such as self-penalization loss [47] and entropy minimization [14,43,55,80,91]. However, these methods only consider the probability or entropy of each individual sample and disregard the connections of these samples.…”