Mask-Shift-Inference: A novel paradigm for domain generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain Generalization (DG) aims to create models that perform well on unseen data distributions (Out-Of-Distribution or OOD generalization).
• A key challenge in DG is learning domain-invariant representations due to shifts between source and target domains.
• Existing methods often rely on complex constraints or assumptions for optimization.

Purpose of the Study:

• To propose a novel paradigm, Mask-Shift-Inference (MSI), for robust Domain Generalization using Convolutional Neural Networks (CNNs).
• To shift the focus of domain-invariant representation learning to feature channels in the latent space.
• To enhance the cognitive ability of models for improved generalization.

Main Methods:

• A two-branch architecture with a main module and domain-specific sub-modules is introduced.
• Unstable feature channels are progressively masked during forward propagation to remove domain-specific spurious information.
• Domain style is shifted to the closest source domain using Fourier transform for style matching, minimizing the domain gap without additional model updates.

Main Results:

• The MSI paradigm effectively discards unstable channels, mitigating domain shifts and spurious correlations.
• Domain style matching and shifting reduces the domain gap, improving model performance on unseen target domains.
• Extensive experiments on PACS, VLCS, Office-Home, and DomainNet datasets demonstrate the superiority and effectiveness of MSI.

Conclusions:

• MSI offers a logically progressive approach to Domain Generalization, intuitively excluding confounding domain-specific information.
• The method implicitly learns semantically invariant representations, achieving robust Out-Of-Distribution generalization.
• MSI enhances model generalization capabilities by focusing on feature channel adaptation and domain style manipulation.