Progressive Invariant Causal Feature Learning for Single Domain Generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Single domain generalization (SDG) faces challenges in transferring models to unseen domains due to unknown domain shifts.
• Learning domain-invariant features is crucial for mitigating the impact of domain shifts in SDG.

Purpose of the Study:

• To develop a novel method for accurately capturing domain-invariant features in SDG.
• To improve the generalization ability of models across multiple unseen target domains.

Main Methods:

• Proposes Progressive Invariant Causal Feature Learning (PICF) using a causal perspective and front-door adjustment.
• Introduces a foreground feature filter to extract object-related causal features by removing irrelevant confounders.
• Enhances causal feature invariance through training with augmented features combined with randomly-sampled styles.

Main Results:

• The PICF method effectively bridges the gap between seen and unseen domains by capturing invariant causal features.
• Significant performance improvements were observed across multiple datasets when PICF was integrated with state-of-the-art methods.
• Achieved a notable 4.7% accuracy improvement on the PACS dataset.

Conclusions:

• PICF demonstrates superior performance in enhancing model generalization for single domain generalization tasks.
• The causal approach offers a robust framework for learning domain-invariant features, addressing key challenges in SDG.
• The method's plug-and-play nature allows for integration with existing SDG techniques, highlighting its versatility.