Reproduce “Joint Out-of-Distribution Filtering and Data Discovery Active Learning” on OpenResearch

About this paper

Sebastian Schmidt, Leonard Schenk, Leo Schwinn, Stephan Günnemann

This paper from Technical University of Munich and BMW Group introduces Joda, a method addressing Open-Set Discovery Active Learning (OSDAL) by simultaneously filtering out irrelevant out-of-distribution data while discovering novel, task-relevant categories. Joda achieved superior accuracy and the fastest class discovery across diverse datasets and conditions compared to existing active learning approaches.

Key insights

• The Open-Set Discovery Active Learning (OSDAL) scenario formalizes a real-world problem where unlabeled data includes in-distribution, discoverable novel, and irrelevant out-of-distribution samples.
• A single model can effectively perform integrated OOD filtering, in-distribution sampling, and novel category discovery, removing the need for auxiliary models.
• Both Outlier Exposure loss during training and energy-based filtering are critical for robust OOD discrimination, while a self-balancing selection metric facilitates balanced discovery.

Summary by alphaXiv.