Abstract: Sparse regularization is essential for interpretable and efficient machine learning, but classical sparsity penalties are often non-smooth and difficult to combine with the gradient-based training procedures prevalent in deep learning. This talk discusses how overparameterization can address this challenge by replacing non-smooth objectives in the original parameters with smooth optimization in an expanded parameter space. The resulting methods recover classical and structured sparsity penalties while remaining compatible with modern training pipelines. Theoretical results characterize the induced penalties and connect the resulting optima and dynamics to sparse solutions. Empirically, the methods are effective in practice, reducing neural network sizes by a large factor.

Bio: David Rügamer is an Associate Professor of Statistics and Data Science at LMU Munich, where he heads the Munich Uncertainty Quantification AI Lab. He is a Principal Investigator at the Munich Center for Machine Learning, a Fellow of the Konrad Zuse School of Excellence in Reliable AI (relAI), an Ellis Member, and a Principal Investigator in the DFG Research Training Group METEOR. Before joining LMU Munich in his current role, he held interim professorships at TU Dortmund, RWTH Aachen University, and LMU Munich. He received his PhD in Statistics from LMU Munich, working on topics at the intersection of functional data analysis, boosting, and post-selection inference.

Beyond his research, he is actively involved in the academic community. He serves as an Associate Editor for TMLR and an Action Editor for the Journal of Statistical Software, and has held senior organizational roles including Position Paper Track Chair for NeurIPS 2026 and Program Chair for ProbML 2026. He has also served as (senior) area chair for major machine learning conferences including AISTATS, ICLR, ICML, NeurIPS, and UAI.