Generalization Beyond Feature Alignment: Concept Activation-Guided Contrastive Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Contrastive learning achieves state-of-the-art results in domain generalization (DG).
• The core feature alignment strategy in contrastive learning can paradoxically hinder model generalization.
• Neuron interpretability reveals that indiscriminate alignment minimizes feature diversity.

Purpose of the Study:

• To investigate the negative impact of feature alignment on model generalization in DG.
• To propose a novel approach, Concept Contrast (CoCo), to enhance feature diversity and generalization.
• To demonstrate the effectiveness and versatility of CoCo across various contrastive DG methods.

Main Methods:

• Characterizing the problem of feature alignment using neuron interpretability and activation views.
• Introducing Concept Contrast (CoCo), a plug-and-play module that contrasts high-level concepts encoded in neurons.
• Integrating CoCo into four canonical contrastive methods for domain generalization.

Main Results:

• CoCo effectively relaxes element-wise feature alignments by focusing on concept-level contrasts.
• CoCo consistently improves the generalization capability of various contrastive DG methods.
• Neuron coverage analysis indicates CoCo potentially invokes more meaningful neurons during training.

Conclusions:

• Feature alignment in contrastive learning can limit diversity and harm domain generalization.
• Concept Contrast (CoCo) offers a simple yet effective solution to enhance feature diversity and improve DG.
• CoCo's ability to leverage high-level concepts and activate more meaningful neurons presents a promising direction for future research in DG.