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

OOD Detection, which determines whether data belongs to the existing training distribution (In-Distribution, ID) or to a new distribution (Out-of-Distribution, OOD), is an essential technology in various fields such as autonomous driving and medical diagnosis.

The existing Outlier Exposure (OE) approach improves performance by leveraging auxiliary OOD datasets, but it has the limitation that securing sophisticated OOD data in real-world environments requires enormous cost and labor. To address this, research on GAN- or Diffusion-based synthetic data generation has been actively conducted recently.

In this research, LG AI Research's DI Lab proposes the generation of Semantic-Discrepant (SD) outliers with two key properties.

• Proximity to ID: The generated data must be sufficiently similar to the existing ID data
• Semantic discrepancy: While incidental information (Nuisance) such as background should be preserved, the core semantic information must be clearly distinguished from ID

The research team obtained pseudo-labels through SCAN, a state-of-the-art self-supervised learning methodology, and used a Classifier-Free Guidance (CFG) Diffusion model to generate photorealistic and semantically distinguishable outliers, thereby innovatively improving OOD detection performance.