Statistical learning from high-dimensional and distributed data

Speaker: Faicel Chamroukhi (irt-SystemX) Abstract Modern learning algorithms must tackle real-world problems involving complex, heterogeneous, often unlabeled, high-dimensional, and potentially large-scale and/or distributed data. In this talk, the speaker will present a statistical learning approach focused on the design of latent variable models, with approximation capabilities and learning guarantees. He will begin by introducing mixture-of-experts models designed for heterogeneous data and high-dimensional functional predictors, which may be noisy, and their training via regularization methods, enabling sparse and interpretable representations. Finally, when data are inherently distributed and/or constrained by confidentiality requirements, the speaker will present federated learning and aggregation strategy for (statistical or neural) models trained in parallel.