AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-19
Towards optimal selection schedules for online active learning
This paper studies the optimal query policy for online linear classification, where labeling samples is costly. Under Gaussian design in the high-dimensional regime, it shows that greedy myopic policies select samples by margin intervals that shrink over time. The authors numerically find optimal threshold schedules, design a simple policy, and compare it with other heuristics on synthetic and real medical image data.