Rethinking the Flow-Based Gradual Domain Adaptation: A Semi-Dual Optimal Transport Perspective

Technical Summary: Rethinking the Flow-Based Gradual Domain Adaptation: A Semi-Dual Optimal Transport Perspective

Problem Statement
Gradual Domain Adaptation (GDA) aims to bridge the distributional gap between a labeled source domain and an unlabeled target domain by progressively adapting models through a sequence of intermediate domains. While flow-based approaches have gained attention for synthesizing these intermediate domains by interpolating between source and target distributions, they face a critical limitation: they typically rely on explicit estimation of the target domain's probability density function (PDF). Existing methods often estimate the target PDF (e.g., via score matching) to define the driving force (velocity field) for the flow. However, estimating PDFs from samples alone is an ill-posed problem; inaccuracies in this estimation can lead to samples being driven into low-density regions, resulting in a mismatch between generated and true target samples, thereby degrading downstream task performance.

Methodology: E-SUOT Framework
To address the reliance on explicit PDF estimation and the instability of existing dual formulations, the authors propose Entropy-regularized Semi-dual Unbalanced Optimal Transport (E-SUOT). The methodology proceeds as follows:

1. Reformulation as Optimization: The authors recast the flow-based GDA problem, traditionally viewed as a gradient flow minimizing an $f$-divergence, as a recursive optimization problem regularized by the squared 2-Wasserstein distance (a Jordan-Kinderlehrer-Otto scheme). This formulation avoids the need for explicit PDF evaluation.
2. Semi-Dual Formulation: They derive a semi-dual formulation of this optimization problem. Unlike primal formulations that require density estimation, the semi-dual objective depends only on expectations over the source and target samples. This allows the use of Monte Carlo methods to approximate integrals without accessing the target PDF.
3. Addressing Instability: The standard semi-dual formulation involves a composite "sup–inf" structure (adversarial training), which the authors prove can lead to non-unique optimal solutions and training instability. To resolve this, they introduce an entropy regularization term to the primal objective.
4. Stable Sequential Optimization: The entropy regularization transforms the adversarial min–max problem into a stable sequential optimization procedure. Theoretically, this ensures the uniqueness of the optimal solution and guarantees convergence.
5. Training Workflow: The framework trains two neural networks in an alternating manner:
   • A potential function $w_\phi$ is optimized to maximize the semi-dual objective (which now includes the entropy term).
   • A transport map $T_\theta$ is optimized to minimize the transport cost conditioned on the learned potential $w_\phi$.
   • Once the sequence of transport maps is learned, the classifier is fine-tuned stage-by-stage along the generated intermediate domains.

Key Contributions

1. Semi-Dual Formulation: The paper develops a semi-dual formulation for intermediate domain generation in flow-based GDA that eliminates the need for explicit estimation of the target domain PDF, thereby avoiding the errors associated with density estimation.
2. Entropy-Regularized Stability: The authors introduce an entropy regularization term to address the inherent instability and non-uniqueness of the semi-dual formulation. This results in the novel E-SUOT framework, which guarantees a unique optimal solution and a stable training process.
3. Theoretical Analysis: The paper provides theoretical proofs regarding the convergence of the E-SUOT framework and establishes an upper bound on the generalization error of the adapted model, linking the error to the cumulative transportation cost and label continuity along the adaptation path.
4. Empirical Validation: Extensive experiments on GDA tasks (Portraits, MNIST with rotation) and Unsupervised Domain Adaptation (UDA) tasks (Office-Home) demonstrate that E-SUOT achieves superior performance compared to state-of-the-art methods, including self-training, gradient flow-based methods (GGF, CNF), and optimal transport-based approaches (MMDSW, SMMD).

Experimental Results

• GDA Tasks: On the Portraits, MNIST $45^\circ$, and MNIST $60^\circ$ datasets, E-SUOT consistently outperformed existing baselines. For instance, it improved accuracy by 21.5%, 23.4%, and 38.6% over the source-only baseline, respectively, and surpassed the second-best methods (e.g., STDW, GOAT) across all datasets.
• UDA Tasks: On the Office-Home dataset, E-SUOT achieved the best average accuracy among compared methods. While not always the top performer on every specific transfer direction, it demonstrated high stability and competitiveness, often securing second-best results with small gaps to the best methods.
• Ablation Studies:
  • Removing entropy regularization (reverting to adversarial training) caused significant performance drops, confirming the necessity of the regularization for stability.
  • Using alternative $f$-divergences (e.g., $\chi^2$, Softplus) or barycentric projection strategies resulted in inferior performance compared to the proposed KL-divergence-based semi-dual approach.
  • The method showed robustness to variations in UMAP embedding dimensions and hyperparameters like step size and batch size.

Significance and Claims
The paper claims that E-SUOT advances the field of GDA by unifying flow-based methods and optimal transport within a semi-dual formulation. It identifies and resolves fundamental issues of stability and generalization that have limited previous approaches, specifically the reliance on inaccurate PDF estimation and the instability of adversarial training. By providing a theoretically grounded, entropy-regularized framework, the authors argue that their approach offers a more robust and reliable mechanism for constructing intermediate domains. The work suggests that avoiding explicit density estimation while ensuring stable optimization is key to effective gradual adaptation, particularly in scenarios where intermediate domains are unavailable or difficult to synthesize. The authors maintain a modest stance, acknowledging limitations such as the assumption of label invariance along the transport path and the current reliance on multi-stage offline training, suggesting future directions for incorporating label information and improving parameter efficiency.