Unified Optimal Transport Framework for Universal Domain Adaptation (2210.17067v2)

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.

• The paper introduces UniOT, a unified Optimal Transport (OT) framework for Universal Domain Adaptation (UniDA) that addresses both common class detection and private class discovery.
• UniOT uses OT-based partial alignment for automatic common class detection and OT-based target representation learning for private class discovery, evaluated with a new $ ext{H}^3$-score.
• Extensive experiments show UniOT outperforms existing UniDA methods on several benchmarks, demonstrating its robustness and reduced reliance on manual threshold tuning.

Unified Optimal Transport Framework for Universal Domain Adaptation

The paper "Unified Optimal Transport Framework for Universal Domain Adaptation" presents a novel approach to Universal Domain Adaptation (UniDA) using a unified framework based on Optimal Transport (OT). UniDA is a complex task where knowledge is transferred from a labeled source domain to an unlabeled target domain without constraints on label sets, making the detection and alignment of common samples critical, along with the recognition of private target-domain categories.

Summary of Key Contributions

The authors propose UniOT—a unified OT framework addressing two critical components in UniDA:

1. Common Class Detection (CCD):
   • Utilizes OT-based partial alignment to detect common classes without requiring predefined thresholds, which is vital given the varied ratios of common categories.
   • By analyzing the assignment matrix from OT, UniOT can automatically discern differences between common and private classes.
2. Private Class Discovery (PCD):
   • Introduces OT-based target representation learning to foster global discrimination and local consistency, avoiding over-reliance on source data.
   • Implements a new metric, the $\text{H}^3$-score, which measures accuracy and clustering performance, providing a robust evaluation of target representations.

The implementation includes adaptive techniques for filling to handle unbalanced class proportions, and a dynamic update mechanism for marginal probability vectors, ensuring effective representation and alignment without relying on exhaustive prior configurations.

Numerical Results and Implications

Extensive experimental validation demonstrates UniOT's superiority across several benchmarks, including Office, Office-Home, VisDA, and DomainNet with substantial improvements over existing UniDA methods. The strong numerical results validate the efficacy of using OT in detecting common classes and discovering private categories automatically.

The precise components of the UniOT framework contribute to significant advancements in domain adaptation scenarios, highlighting the method's robustness in diverse and challenging real-world settings without dependency on prior domain knowledge.

Practical and Theoretical Implications

Practically, UniOT reduces the reliance on manual interventions for threshold settings, enhancing adaptability across datasets with varying class distributions. It paves the way for more realistic implementations in environments with unbounded category sets.

Theoretically, integrating OT as a central mechanism for domain adaptation advances the field by providing a structured approach in tackling the inherent challenges of category and domain gaps. This conceptual shift offers insights into utilizing transport-based optimization for efficient domain alignments and establishing new norms for target representation evaluations using the $\text{H}^3$-score.

Future Directions

The research prompts further investigation into refining memory models for more efficient and scalable feature filling strategies within the OT framework. Additionally, exploring extensions of UniOT to other machine learning tasks beyond domain adaptation remains a promising avenue.

In conclusion, the paper contributes significantly to UniDA by leveraging the mathematical optimization properties of OT, setting a new benchmark for future research in domain transfer learning methodologies.