BROWSE
Detail
Conference Paper
Low-Likelihood EBM Samples for Out-of-Distribution Detection
Cited - time in 
Share 
Share
- Authors
- Assefa Seyoum Wahd, Marcella Astrid, Seung-Ik Lee
- Issue Date
- 2023-02
- Citation
- 영상처리 및 이해에 관한 워크샵 (IPIU) 2023, pp.1-6
- Publisher
- 한국정보과학회
- Language
- English
- Type
- Conference Paper
- Abstract
- Out-of-distribution detection (OOD) is an essential step in ensuring the safe deployment of machine learning models. Several approaches have been proposed to use real-world data as a training OOD data. However, OOD data may not be accessible during training, for example in case of rare events, or it may requires huge of amount of data collection. In this paper, we present EBMOOD, an out-of-distribution detection approach that utilizes energy-based models (EBMs) to generate low-likelihood data and the generated pseudo-OOD data is used to train an OOD detector. Our approach achieves competitive results compared to methods that utilize external OOD data and up to 26.24% decrease in FPR95 over the baseline.
- KSP Keywords
- Data Collection, Energy-based models, Rare events, Real-world data, Safe deployment, machine learning models