Unified AL & OOD

📄 Transactions on Machine Learning Research (TMLR) 2025

Overview

Active learning and out-of-distribution (OOD) detection have long been studied as separate problems — but in practice they are deeply intertwined. When selecting data for labeling, you inevitably encounter OOD samples; when detecting OOD data, you benefit from a well-curated labeled set.

This work bridges the gap by proposing a unified framework that tackles both challenges simultaneously, providing a principled approach to select informative in-distribution samples while filtering out-of-distribution noise.

Key Contributions

• 🔀 Unified perspective — first framework jointly addressing active learning and OOD detection
• 🎯 Principled sample selection — simultaneously identifies the most informative samples and detects distributional shifts
• 💡 Data-efficient — leads to more robust learning pipelines with lower labeling budgets

Why It Matters

Real-world datasets are never clean. By unifying AL and OOD detection, this work enables machine learning pipelines that are both cost-effective and robust — a key requirement for deploying models in the wild.

(Schmidt et al., 2024)