Generative Inference Network for Imbalanced Domain Generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain generalization (DG) seeks to transfer knowledge from source domains to unseen target domains.
• Existing DG methods struggle with imbalanced data across domains and categories, hindering generalization.
• Imbalanced data negatively impacts the learning of robust classification models.

Purpose of the Study:

• To address the challenge of imbalanced data in domain generalization (IDG).
• To propose a novel method, Generative Inference Network (GINet), for improved domain generalization.
• To enhance the discriminative and generalization abilities of models in imbalanced scenarios.

Main Methods:

• Formulated a practical imbalanced domain generalization (IDG) scenario.
• Proposed Generative Inference Network (GINet) to augment reliable samples for minority domains/categories.
• Utilized cross-domain images to estimate common latent variables and discover domain-invariant knowledge.
• Generated novel samples using optimal transport constraints to enhance model robustness.

Main Results:

• GINet effectively augments minority domain/category samples.
• The method demonstrates improved robustness and generalization ability on benchmark datasets.
• Empirical analysis and ablation studies confirm the advantage of GINet over other DG methods.

Conclusions:

• The proposed GINet method offers a straightforward yet effective solution for imbalanced domain generalization.
• GINet enhances model generalization by addressing data imbalance issues.
• The approach shows significant advantages in elevating model generalization performance.