Out-of-Distribution Generalization and Detection With Limited Source Data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Out-of-distribution (OOD) generalization and detection are crucial for robust AI.
• Current methods often require extensive source data for generalization and unknown category detection.
• Addressing both covariate and category shifts simultaneously with limited data remains a challenge.

Purpose of the Study:

• To propose a novel method for simultaneous OOD generalization and detection.
• To overcome the limitation of requiring large source datasets.
• To effectively handle both covariate and category shifts under data constraints.

Main Methods:

• A novel distribution estimation method using pretrained classifier weights and an auxiliary dataset.
• Generation of estimated in-distribution (ID) feature representations.
• Constrained optimization on estimated wild data (ID, covariate-shifted OOD, category-shifted OOD) for generalization and detection.

Main Results:

• The proposed method achieves competitive results with limited source data.
• Demonstrates effectiveness in handling both covariate and category shifts.
• Outperforms baseline methods specializing in only OOD generalization or detection.

Conclusions:

• The novel method successfully addresses OOD generalization and detection with limited data.
• It provides a viable solution for scenarios with scarce training resources.
• The approach offers a unified framework for tackling diverse OOD challenges.