LG AI Research 555

Existing Out-of-Distribution (OOD) Detection is fundamentally about detecting data not seen during training. Existing Outlier Exposure techniques incur costs for collecting external samples, and have limitations in distinguishing Near-OOD that is similar to the training data.

SONA (Semantic Outlier Generation via Nuisance Awareness), proposed by LG AI Research, is a framework that leverages Diffusion models to directly use ID images to generate sophisticated Outliers. This approach distinguishes between an image's Semantic information and Nuisance (unnecessary information) to enable effective Outlier synthesis.

The core process of SONA is as follows:

• Masking: Uses the difference in Conditional-Unconditional Noise Estimation to distinguish and mask Semantic regions from Nuisance regions.
• Guidance: Removes the existing Semantic information while preserving the Nuisance, and mixes a new Random OOD Semantic with the existing ID Semantic.
• Loss Optimization: Trains the model to focus on semantic differences through three Loss Terms that perform ID classification, OOD distinction, and minimization of Mutual Information between Semantic information.

Through this approach, SONA resolves the instability of existing Diffusion-based generation methods and can accurately detect a wide range from Near-OOD to Far-OOD.