Learning under change: what can we trust, and what to do when we cannot?

Speaker: Nicola Gnecco, Imperial College London

Abstract: We train machine learning models on the data we can collect, but we often deploy them under different conditions. For instance, environments change, policy interventions occur, or inputs fall outside the collected data. In these situations, collecting more data from the same regime is not enough to provide guarantees on the deployed setting. To address this, we need assumptions about the relationship between the collected and deployed data (via invariances, structured shifts, or regularities in extreme events), and diagnostics when these assumptions fail. In this talk, I will present a research program for robust learning under changing conditions, illustrating the main ideas across the following examples: (i) distribution generalization across different environments, where we learn targets that are stable under shifts; (ii) causal discovery in heavy-tailed systems, where we use the signature in the tails to learn how large shocks propagate through a system; and (iii) a disagreement-based test flagging when certain inputs are likely outside the raining distribution.

Biography: Nicola Gnecco is an Assistant Professor in Statistics in the Department of Mathematics at Imperial College London. His research focuses on reliable machine learning under changing conditions, drawing on ideas from causal inference and extreme value theory. Previously, he was a postdoctoral researcher at the Copenhagen Causality Lab (University of Copenhagen) and a visiting postdoctoral researcher at UCL and UC Berkeley, supported by an SNSF Postdoc.Mobility grant (210976). He received his PhD in Statistics from the University of Geneva, supervised by Sebastian Engelke, and completed his MSc thesis at ETH Zurich, supervised by Nicolai Meinshausen.

Host: Prof. Ernst Wit