Alexander Jung

Associate Professor

Aalto University, Finland

Machine learning (ML), the driving force behind most AI systems, trains models by minimizing prediction errors on a given dataset. Federated learning (FL) extends this paradigm to networks of distributed ML tasks, where each task involves a separate model and dataset. Just as empirical risk minimization is a foundational principle in ML, total variation (TV) minimization can provide a unifying framework for FL. This talk explores the mathematical structure of TV minimization and its role in designing trustworthy AI. We demonstrate how carefully chosen components of TV minimization lead to AI services that are robust, privacy-preserving, and explainable.