[論文レビュー] Unified Optimal Transport Framework for Universal Domain Adaptation

Wanxing Chang, Shi Ye+2 more|arXiv (Cornell University)|Oct 31, 2022

Domain Adaptation and Few-Shot Learning被引用数 24

ひとこと要約

UniOT は アンバランス OT と 標準 OT を 組み合わせて、Universal Domain Adaptation における 共通クラス検出と ターゲットプライベートクラス発見 を 共同で 取り扱い、適応的充填と OT ベースの ターゲット表現学習 を 導入し、新しい H3-score 指標を 用いて プライベートクラスの クラスタリングを 評価する。

ABSTRACT

Universal Domain Adaptation (UniDA) aims to transfer knowledge from a source domain to a target domain without any constraints on label sets. Since both domains may hold private classes, identifying target common samples for domain alignment is an essential issue in UniDA. Most existing methods require manually specified or hand-tuned threshold values to detect common samples thus they are hard to extend to more realistic UniDA because of the diverse ratios of common classes. Moreover, they cannot recognize different categories among target-private samples as these private samples are treated as a whole. In this paper, we propose to use Optimal Transport (OT) to handle these issues under a unified framework, namely UniOT. First, an OT-based partial alignment with adaptive filling is designed to detect common classes without any predefined threshold values for realistic UniDA. It can automatically discover the intrinsic difference between common and private classes based on the statistical information of the assignment matrix obtained from OT. Second, we propose an OT-based target representation learning that encourages both global discrimination and local consistency of samples to avoid the over-reliance on the source. Notably, UniOT is the first method with the capability to automatically discover and recognize private categories in the target domain for UniDA. Accordingly, we introduce a new metric H^3-score to evaluate the performance in terms of both accuracy of common samples and clustering performance of private ones. Extensive experiments clearly demonstrate the advantages of UniOT over a wide range of state-of-the-art methods in UniDA.

研究の動機と目的

• Motivate UniDA as a realistic setting where both domains have private classes and no prior label-set information.
• Develop a unified OT framework to address common class detection and private class discovery without predefined thresholds.
• Leverage partial/ unbalanced OT to detect common samples and OT-based representation learning to uncover target-private categories.
• Introduce a new evaluation metric (H3-score) that accounts for both common-class accuracy and target-private clustering quality.

提案手法

• Formulate common class detection as Unbalanced OT-based partial alignment between target samples and source prototypes, with adaptive filling to handle varying proportions of common classes.
• Use OT between target features and target prototypes for private class discovery, encouraging global discrimination and local sample consistency (L_global and L_local).
• Define L_CCD via pseudo-labels from the OT assignment for selected common samples and cross-entropy loss on those samples.
• Define L_PCD to jointly optimize prototype-based clustering with a memory-augmented OT solver, enforcing both global separation and local consistency.
• Integrate L_CCD and L_PCD with source supervision in an overall objective L_overall = L_cls + λ(L_CCD + L_PCD).
• Provide an automatic mechanism to adapt marginals (β) and to fill unbalanced positives/negatives to stabilize CCD.]
• research_questions:[

実験結果

リサーチクエスチョン

• RQ1 Can universal domain adaptation be achieved without threshold hand-tuning by formulating common class detection as unbalanced OT with adaptive filling?
• RQ2 Can target-private categories be effectively discovered and recognized within UniDA using an OT-based target representation learning framework?
• RQ3 Does combining partial inter-domain alignment with intra-domain OT clustering yield competitive performance across standard UniDA benchmarks?
• RQ4 Does the proposed H3-score provide a meaningful joint measure of common-class accuracy and private-class clustering quality?

主な発見

• UniOT achieves state-of-the-art H-scores across Office, Office-Home, VisDA, and DomainNet benchmarks compared to prior UniDA methods.
• The OT-based CCD with adaptive filling outperforms CCD without adaptive filling, highlighting the importance of balancing positive/negative detection.
• L_global contributes more than L_local in improving H3-scores, indicating the value of global cluster discrimination for private class discovery.
• The proposed H3-score, combining common-class accuracy, private-class discovery, and NMI-based clustering quality, demonstrates balanced performance improvements.
• UniOT exhibits robust performance under realistic category splits, without relying on predefined common/private thresholds.