Programme
The scientific programme of the StatML Summer School brings together internationally recognised researchers whose work is shaping the future of statistics and machine learning. Through a series of in-depth lecture courses, participants will explore contemporary research challenges, gain insight into emerging methodologies, and engage with ideas at the forefront of the field.
The programme will begin on Monday 12 July and end on Friday 16 July.
Professor David Dunson, Duke University
Learning Latent Structure: Bayesian Models and Spectral Methods
Latent variable models provide a powerful framework for learning structure from complex, high-dimensional data. These lectures develop Bayesian and spectral approaches to latent structure learning, emphasizing scalable computation, identifiability, and uncertainty quantification. Starting with factor models, we examine how increasing ambient dimension can improve recovery of unobserved factors and enable efficient inference on covariance matrices and latent probability distributions. We then consider adaptive shrinkage, scientifically informed representations, and shared structure across multiple data views and studies. Finally, we introduce identifiable discrete and multilayer latent models for categorical and network data. Throughout, recent methodological developments are connected to theoretical guarantees and applications, highlighting when simple computational procedures can retain statistical accuracy and when richer joint modeling remains necessary.
Professor Yi Yu, University of Warwick
Differential privacy and related topics
This short course will introduce some of the statistical foundations of trustworthy learning, with differential privacy as a central theme. We will develop selected tools from modern statistical decision theory and minimax analysis that are useful for understanding the fundamental limits imposed by privacy and related constraints. A range of examples will illustrate how privacy changes both the design of statistical procedures and the achievable rates of estimation, testing, and learning. We will also discuss connections with other notions of trustworthy inference, potentially including robustness, fairness, and learning from distributed or heterogeneous data. The emphasis throughout will be on general principles, mathematical ideas, and techniques that transfer across different statistical models and learning problems.