Semantic enhanced for out-of-distribution detection.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Existing out-of-distribution (OOD) detection methods often sacrifice performance on same-manifold OOD (SMOOD) data while improving on general OOD benchmarks.
• This performance gap stems from a failure to capture comprehensive in-distribution (ID) semantic features.

Purpose of the Study:

• To develop a novel approach for improving OOD detection by enhancing the learning of robust ID semantic features.
• To address the limitations of current methods that compromise SMOOD data performance.
• To propose a method that generalizes effectively to various OOD scenarios.

Main Methods:

• Utilizing features from multiple "semantic perspectives" to create a comprehensive semantic representation of ID samples.
• Perturbing batch sample mean and variance during inference to increase model sensitivity to OOD data.
• The proposed method avoids using OOD samples during training and does not require pre-trained models or inference-time pre-processing.

Main Results:

• The method successfully enhances the semantic representation of ID data.
• Achieved state-of-the-art results on standard OOD benchmark datasets.
• Demonstrated significant improvements in detecting SMOOD data, addressing a key limitation of prior work.

Conclusions:

• The proposed strategies effectively improve the generalization ability of models to OOD data.
• This approach offers a robust solution for OOD detection without relying on OOD samples or complex training procedures.
• The findings suggest a new direction for developing more reliable and comprehensive OOD detection systems.