Exploring Invariance Matters for Domain Generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain generalization (DG) faces performance degradation with Out-Of-Distribution (O.O.D) data.
• Existing methods using CNNs struggle with spurious-invariant information and domain/class discrepancies.
• Causal mechanisms reveal limitations in current invariant feature extraction.

Purpose of the Study:

• To propose a self-weighted multi-view mining invariance domain generalization framework (SMIDG).
• To address limitations in traditional single-view feature extraction and mitigate spurious correlations.
• To enhance both coarse-grained and fine-grained generalization for robust O.O.D performance.

Main Methods:

• Employed a multi-view approach, including frequency domain features, to overcome single-view limitations.
• Utilized self-adaptive adversarial masks for eliminating spurious correlations and ensuring coarse-grained causal invariance.
• Implemented progressive contrastive learning and self-adaptive hard sample mining for fine-grained domain and feature invariance.

Main Results:

• The proposed SMIDG framework significantly improves performance on domain generalization tasks.
• Demonstrated superior results compared to state-of-the-art approaches across five benchmark datasets.
• Validated the effectiveness of multi-view learning and advanced techniques in mitigating O.O.D challenges.

Conclusions:

• The SMIDG framework offers a robust solution for domain generalization by addressing spurious correlations and discrepancies.
• The combination of multi-view mining, adversarial masks, and contrastive learning enhances model adaptability to O.O.D data.
• The study provides a significant advancement in achieving reliable performance across diverse data domains.