Many learning tasks, such as model alignment and fine-tuning, rely on costly human feedback, making it essential to minimize such interventions. In this talk, we formalize this challenge in the setting of online classification tasks, motivated by question-answering problems, and introduce the framework of online classification with expert guidance. For this problem, we propose algorithms tailored to both low- and high-budget regimes. We establish regret upper bounds for these algorithms by combining concentration-of-measure techniques with tools from convex geometry. Our approach is further validated through experiments, including evaluations on real-world question-answering datasets using embeddings derived from state-of-the-art large language models.
Alexandre Proutiere is Professor in the Decision and Control Systems at KTH Royal Institute of Technology, Sweden. Before joining KTH in 2011, he was a researcher at Microsoft Research, and a research engineer at France Telecom R&D. He was an invited lecturer and researcher at the computer science department ENS Paris. His current research interests include probability, optimization, and machine learning. Education and awards: PhD in Applied Mathematics from Ecole Polytechnique, Graduated in Mathematiques from Ecole Normale Superieure, Engineering degree from Telecom Paris, Engineer from Corps des Mines. Awards: ACM Sigmetrics rising star award in 2009, ACM best papers awards at Sigmetrics 2004 and 2010, and Mobihoc 2009. ACM Sigmetrics test of time award in 2025. Recipient of an ERC consolidator grant 2012-2017.