[NeurIPS 2023] OOD Detection via Synthetic Outlier Generation (Part 1) - LG AI Research BLOG

Out-of-Distribution (OOD) Detection is a technology that determines whether new data belongs to an existing data distribution (In-Distribution), and it is essential in fields such as autonomous driving and medical diagnosis. Recently, the Outlier Exposure (OE) approach, which utilizes auxiliary datasets, has been drawing attention, but there is a problem in that securing high-quality outlier data in real-world environments requires enormous cost.

To address this, LG AI Research introduced Semantic-Discrepant (SD) Outlier Generation technology based on a Diffusion model. Unlike existing generative models that had the limitation of producing blurry images, this methodology generates photo-realistic outliers.

The core of the proposed SD Outlier generation is to satisfy two properties.

• ID similarity: The generated data must be sufficiently close to the existing data distribution
• Semantic discrepancy: While auxiliary information (Nuisance) such as background is maintained, the core semantic must be clearly distinguished

The research team combined Pseudo-labels obtained through the SCAN methodology with a Classifier-Free Guidance (CFG) model to build a new OOD detection framework that works effectively even in unlabeled environments.