Budget-Agnostic Bayesian Active Learning with Kernel-Weighted Evidence

Speaker: Daphna Weinshall (Hebrew University of Jerusalem)

Date: 2/3/26

Abstract: Modern deep learning often relies on massive labeled datasets, yet in many domains (such as medicine) labels require expensive expert effort, limiting feasibility. Active Learning (AL) addresses this problem by selectively curating informative examples within a fixed budget. Yet most existing AL paradigms and methods lean on prediction uncertainty to some extent, which cannot be reliably estimated in the small-budget regime. I will begin by describing our work on very low-budget AL, where we recast query selection as a k-set coverage problem: we select points whose fixed-radius neighborhoods in representation space cover as much probability mass as possible. I will then present a Bayesian extension that supports multiple noisy annotators with different costs and reliabilities. To reconcile conflicting annotations, we replace binary coverage with kernel-weighted “soft evidence,” and maintain Bayesian posteriors over labels via Dirichlet–Categorical updates with discounted (kernel-weighted) counts. I will then conclude with open problems and next steps.