UFCOD: Few-Shot Cross-Domain OOD Detection via Diffusion Geometry
A new framework called UFCOD has been introduced by researchers to address few-shot cross-domain out-of-distribution (OOD) detection. Traditional OOD detectors necessitate training on a particular in-distribution (ID) dataset. In contrast, UFCOD allows for the detection of various new ID-OOD task combinations using only a limited number of ID samples during inference, without requiring extra training. This approach utilizes information-geometric analysis of diffusion trajectories to derive two energy features: Path Energy, which measures integrated score magnitude, and Dynamics Energy, which assesses score smoothness. Together, these features create a discrete So.
Key facts
- UFCOD performs few-shot cross-domain OOD detection.
- It uses a single pre-trained model with no additional training.
- Only a handful of ID samples are needed at inference time.
- Key insight: diffusion noise predictions are score functions.
- Two energy features extracted: Path Energy and Dynamics Energy.
- Method is based on information-geometric analysis of diffusion trajectories.
- Standard OOD detectors are trained on a specific ID dataset.
- UFCOD works on arbitrary new ID-OOD task pairs.
Entities
—