Research Seminar - Marco Mondelli

The Institute for Statistics and Mathematics is pleased to invite you to the next research seminar, which will take place on campus:

Marco Mondelli, Institute of Science and Technology Austria (ISTA))
Learning from Data of All Sorts (Synthetic, Private, Surrogate, and Mixture): Lessons from High-Dimensional Regression
Wednesday, October 14, 2026, at 5:15 p.m., in Building D4, Room D4.0.127.

Abstract: Machine learning models are routinely trained on all sorts of data: synthetic data obtained via generative models; private data held by external providers; surrogate data labeled by pretrained teacher models; and mixture data coming from heterogeneous domains. While each of these settings has its own distinctive features, in this talk I will outline a common methodology based on asymptotic characterizations of the risk, as well as non-asymptotic deterministic equivalents. Our approach provides precise predictions and establishes rigorous scaling laws for the error (under classical source-capacity conditions) for a broad class of high-dimensional ridge regression models. These include model shift, covariance shift, heterogeneous noise levels, dataset sizes growing at different rates, and the addition of noise to enforce differential privacy. In turn, the theoretical results provide insights into common questions in machine learning practice: How should synthetic data be selected? When should one commit to purchasing privatized data without explicitly assessing its performance? To what extent can surrogate data labeled by an imperfect teacher improve performance? When does data mixing produce synergies, significantly outperforming the use of data from each domain in isolation? We will provide quantitative answers to these questions and, going beyond linear models, we will validate our findings using practical datasets.

For further information and the seminar schedule, please see:
www.wu.ac.at/en/statmath/research/resseminar