Investigation of out-of-distribution detection across various models and training methodologies

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning algorithms excel in many areas but falter with data distributions different from training sets.
• Out-of-distribution (OOD) detection is crucial for reliable AI, yet current methods often fail to improve OOD performance alongside in-distribution accuracy.

Purpose of the Study:

• To comprehensively study OOD detection performance across various models and training techniques.
• To verify the phenomenon where OOD performance does not consistently increase with in-distribution classification accuracy.

Main Methods:

• Evaluated multiple pre-trained computer vision models using existing and novel OOD detection methods.
• Introduced Trimmed Rank with Inverse softMax probability (TRIM), a new method focusing on model weights and training strategies.

Main Results:

• Observed significant performance disparities among current OOD detection methods.
• TRIM demonstrated promising results, showing high compatibility between OOD performance and in-distribution accuracy.

Conclusions:

• Existing OOD detection methods exhibit performance inconsistencies.
• TRIM offers a simple, effective approach to enhance OOD detection, potentially bridging the gap between in-distribution accuracy and OOD data identification.