Unified Optimal Transport Framework for Cold-Start Active Learning

A recent preprint on arXiv (2608.03249) introduces a cohesive optimal transport framework tailored for Cold-Start Active Learning (CSAL), aimed at efficiently choosing valuable subsets from unlabeled data without prior knowledge or human input. The authors highlight that current CSAL techniques, which depend on various heuristics like typicality, coverage, or diversity, each possess unique inductive biases, leading to inconsistent performance across different tasks. They emphasize that the key challenge lies not in creating another heuristic but in enabling CSAL to adapt to specific data and tasks automatically. To tackle this, they re-evaluate CSAL using optimal transport concepts, presenting a generalized transport selection framework that encapsulates the allocation structures of existing methods. Furthermore, they conduct a theoretical analysis of the trade-offs governed by entropic regularization and establish a task-agnostic minimax bound for cold-start selection, laying a solid foundation for adaptive CSAL. This research is significant for machine learning and artificial intelligence, especially in contexts where labeled data is limited and efficient selection is essential.