[Paper Review] Joint Semantic Domain Alignment and Target Classifier Learning for Unsupervised Domain Adaptation

Dongdong Chen, Yisen Wang+3 more|arXiv (Cornell University)|Jun 10, 2019

Domain Adaptation and Few-Shot Learning59 references4 citations

TL;DR

This paper proposes SDA-TCL, a novel unsupervised domain adaptation method that jointly optimizes semantic domain alignment and target classifier learning in a unified feature space using shared class centers. By integrating discriminative center loss and shared class centers across domains, the method achieves state-of-the-art performance on Office-31 and VisDA benchmarks, demonstrating faster convergence and improved domain gap reduction.

ABSTRACT

Unsupervised domain adaptation aims to transfer the classifier learned from the source domain to the target domain in an unsupervised manner. With the help of target pseudo-labels, aligning class-level distributions and learning the classifier in the target domain are two widely used objectives. Existing methods often separately optimize these two individual objectives, which makes them suffer from the neglect of the other. However, optimizing these two aspects together is not trivial. To alleviate the above issues, we propose a novel method that jointly optimizes semantic domain alignment and target classifier learning in a holistic way. The joint optimization mechanism can not only eliminate their weaknesses but also complement their strengths. The theoretical analysis also verifies the favor of the joint optimization mechanism. Extensive experiments on benchmark datasets show that the proposed method yields the best performance in comparison with the state-of-the-art unsupervised domain adaptation methods.

Motivation & Objective

• Address the limitation of existing unsupervised domain adaptation methods that separately optimize domain alignment and classifier learning, leading to suboptimal performance.
• Overcome the challenge of jointly optimizing feature-level alignment (semantic domain alignment) and label-level discrimination (target classifier learning), which operate in different spaces.
• Develop a holistic optimization mechanism that leverages both intra-class compactness and inter-class separation to improve generalization on target domains.
• Theoretical and empirical validation of the joint optimization mechanism to demonstrate its superiority over decoupled approaches.

Proposed method

• Proposes a unified framework, SDA-TCL, that jointly optimizes semantic domain alignment and target classifier learning in the feature space using shared class centers across domains.
• Introduces a discriminative center loss that pulls samples toward their predicted pseudo-label class centers and pushes them away from other class centers in the target domain.
• Enforces semantic domain alignment by sharing class centers between source and target domains for the same class, promoting inter-domain class-level compactness.
• Uses a two-stage training strategy where target pseudo-labels are gradually introduced via a hyperparameter $I_s$, improving training stability.
• Employs a balanced loss function with a hyperparameter $K$ to control the trade-off between domain alignment and classifier learning, ensuring robustness to hyperparameter choice.
• The joint optimization operates entirely in the feature space, avoiding the complexity of separate optimization in label and feature spaces.

Experimental results

Research questions

• RQ1 How can semantic domain alignment and target classifier learning be effectively combined in a unified optimization framework for unsupervised domain adaptation?
• RQ2 What is the impact of jointly optimizing intra-class compactness and inter-class separation on domain generalization and convergence speed?
• RQ3 How does the joint optimization mechanism compare to decoupled approaches in terms of reducing domain discrepancy and improving accuracy?
• RQ4 What is the optimal timing and weighting strategy for incorporating target pseudo-labels during training to maximize performance?
• RQ5 Can shared class centers across domains serve as an effective bridge to unify feature-level and label-level optimization objectives?

Key findings

• SDA-TCL achieves state-of-the-art accuracy on the Office-31 dataset with a 85.6% average accuracy, outperforming previous SOTA methods such as TCL-ours (84.3%) and SDA-ours (85.0%).
• On the VisDA benchmark, SDA-TCL achieves 81.9% accuracy in the synthetic-to-real transfer setting, surpassing the next best method (81.5%) by a clear margin.
• The method shows faster and more stable convergence than RevGrad, with error rates decreasing more rapidly during training on the A→W task.
• The domain gap, measured by $ ext{dist}_{ar{A}}$, is significantly reduced by SDA-TCL, indicating effective domain alignment compared to baseline methods.
• The ablation study confirms that SDA-TCL outperforms both SDA-ours and TCL-ours, demonstrating the superiority of joint optimization over individual components.
• The method is robust to hyperparameter choices: performance remains strong across a wide range of $K$ values (1 to 9), indicating stability in the weighting of the two objectives.