Continuous Disentangled Joint Space Learning for Domain Generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain generalization (DG) aims to develop models that perform well on unseen data from new domains.
• Existing DG methods often focus on domain-invariant features, potentially neglecting valuable domain-specific semantic information.

Purpose of the Study:

• To propose a novel DG method, continuous disentangled joint space learning (CJSL), that leverages both domain-invariant and domain-specific information.
• To improve the robustness and effectiveness of DG models by simulating domain-specific representations for test samples.

Main Methods:

• CJSL formulates and learns a continuous joint space (CJS) for domain-specific representations through iterative feature disentanglement.
• During inference, CJSL simulates domain-specific representations for test samples by sampling from the learned CJS using Monte Carlo methods.
• This approach enables the use of multiple domain-specific classifiers for robust predictions.

Main Results:

• CJSL demonstrated superior performance compared to 19 state-of-the-art (SOTA) methods.
• The proposed method achieved significant improvements across seven benchmark datasets.
• Empirical results validate the effectiveness of leveraging both domain-invariant and simulated domain-specific information.

Conclusions:

• CJSL offers a novel and effective approach to domain generalization by integrating domain-invariant and domain-specific feature learning.
• The method's ability to simulate domain-specific representations enhances model generalization to unseen domains.
• The proposed technique represents a significant advancement in the field of domain generalization.