Understanding Failures in Out-of-Distribution Detection with Deep Generative Models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep generative models (DGMs) are theoretically suited for out-of-distribution (OOD) detection.
• Empirical studies show DGMs often assign higher probabilities to OOD data than in-distribution data.
• This counterintuitive behavior is frequently attributed to a misalignment between likelihood-based OOD detection and the nature of OOD data.

Purpose of the Study:

• To investigate the reasons behind DGMs' failure in OOD detection.
• To analyze the validity of the typical set hypothesis for OOD detection.
• To identify the primary cause of OOD detection failures in DGMs.

Main Methods:

• Theoretical analysis to prove performance limitations without assumptions on out-distributions.
• Interrogation of the typical set hypothesis and its implications for OOD detection.
• Analysis of consequences arising from assumed support overlap between in- and out-distributions.
• Illustration of OOD detection failures caused by estimation errors.

Main Results:

• No OOD detection method can guarantee performance beyond random chance without specific assumptions about relevant out-distributions.
• The typical set hypothesis is found to be arbitrary for OOD detection, and support overlap assumptions are problematic.
• Model misestimation, rather than a fundamental misalignment, is identified as the primary cause of OOD detection failures.
• Even minimal estimation errors in DGMs can lead to significant OOD detection failures.

Conclusions:

• Model misestimation is the key factor hindering effective OOD detection in DGMs.
• Future research should focus on mitigating estimation errors for improved OOD detection capabilities.
• The findings have significant implications for the development of robust deep generative models and OOD detection techniques.