Learning Generalizable Models via Disentangling Spurious and Enhancing Potential Correlations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain generalization (DG) aims to develop models that perform well on unseen target domains after training on multiple source domains.
• Acquiring domain-invariant representations is crucial for mitigating domain shift and improving model generalization.
• Existing methods often focus on either sample or feature perspectives, but not both simultaneously.

Purpose of the Study:

• To enhance model generalization ability in domain generalization settings.
• To develop a method that learns domain-invariant representations from both sample and feature perspectives.
• To disentangle spurious correlations and enhance potential correlations within the data.

Main Methods:

• A frequency restriction module was developed to guide models towards relevant object feature-label correlations from the sample perspective, disentangling spurious correlations.
• A Tail Interaction module was introduced to implicitly enhance potential correlations among all samples across source domains from the feature perspective.
• These modules were integrated with Convolutional Neural Networks (CNNs) and Multi-Layer Perceptrons (MLPs) as strong baselines.

Main Results:

• The proposed approach significantly improved generalization performance.
• Models equipped with the frequency restriction and Tail Interaction modules achieved superior results compared to baseline methods.
• An average accuracy of 92.30% was reported on the Digits-DG dataset.

Conclusions:

• The combined sample and feature perspective approach effectively learns domain-invariant representations.
• The developed modules successfully disentangle spurious correlations and enhance potential correlations, leading to better generalization.
• The method offers a promising direction for advancing domain generalization techniques in machine learning.