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, 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 with examples in (i) distribution generalization across different environments, (ii) causal discovery in heavy-tailed systems, and (iii) a disagreement-based test flagging when inputs are likely outside the training distribution.

Host: Prof. Ernst Wit