NormAUG: Normalization-Guided Augmentation for Domain Generalization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Supervised deep learning models struggle with domain shift, leading to performance degradation on unseen test data.
• Domain generalization aims to create models robust to variations across different data distributions.
• Data augmentation is key for enhancing training data diversity in domain generalization.

Purpose of the Study:

• To propose NormAUG, a novel normalization-guided augmentation method for improved domain generalization.
• To enhance model robustness and generalization capabilities by introducing feature-level diversity.
• To theoretically and empirically validate the effectiveness of the proposed method.

Main Methods:

• NormAUG employs a two-path architecture: a main path and an auxiliary (augmented) path.
• The auxiliary path utilizes batch normalization with diverse domain statistics (single or combined) during training.
• An ensemble strategy is applied to the auxiliary path's predictions during testing.

Main Results:

• NormAUG effectively improves the generalization of deep learning models to unseen domains.
• The method introduces diverse information at the feature level, enhancing robustness.
• Experiments on benchmark datasets demonstrate significant performance gains.

Conclusions:

• NormAUG offers a simple yet effective approach to address domain shift challenges in deep learning.
• The proposed method enhances model generalization by leveraging normalization-guided augmentation.
• NormAUG provides a promising direction for developing more robust and adaptable AI systems.