Adaptive Domain Generalization Via Online Disagreement Minimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks exhibit performance degradation due to distribution shifts between training and deployment.
• Existing Domain Generalization (DG) methods often fail to outperform simple Empirical Risk Minimization (ERM).

Purpose of the Study:

• To propose a novel framework, AdaODM, that adaptively modifies source models at test time for improved generalization to unseen domains.
• To enhance existing DG algorithms by providing a consistent performance improvement.

Main Methods:

• AdaODM employs a shared domain-generic feature extractor and multiple domain-specific classifiers trained adversarially.
• The feature extractor maps samples to a domain-invariant space, while classifiers capture domain-specific decision boundaries.
• At test time, AdaODM fine-tunes source models to minimize prediction disagreement among classifiers, aligning target features.

Main Results:

• AdaODM consistently improves the generalization capacity of DG methods like ERM and CORAL.
• The framework achieved state-of-the-art performance across four DG benchmarks: VLCS, PACS, OfficeHome, and TerraIncognita.
• Adaptive modification at test time effectively addresses distribution differences between target and source domains.

Conclusions:

• AdaODM offers a general and orthogonal approach to enhance existing DG algorithms.
• The adaptive test-time modification strategy is crucial for robustly transferring models to unseen domains.
• This framework represents a significant advancement in achieving reliable performance under distribution shifts.