Statistics Summer School at Warwick 2026

LMS Research Summer School in Robust Statistics and Reliable Learning Algorithms will be held from 29^th June to 3^rd July 2026 at the University of Warwick, Coventry, UK.

Abstracts (working)

Short courses

Algorithmic Robust Statistics (Ilias Diakonikolas)

The course will provide an overview of algorithmic techniques developed for various high-dimensional robust statistical problems, associated statistical-computational tradeoffs, and data analysis applications.

Topics in robustness and privacy (Po-Ling Loh)

The first part of the course will survey ideas from the rich literature in classical robust statistics, emphasizing connections to more modern work in statistics and machine learning. In the latter half of the course, we will shift our attention to differential privacy, introducing key concepts and discussing what privacy does (and does not) have to do with robustness.

Topics in conformal prediction and algorithmic stability (Rina Foygel Barber)

The first half of this course will be a tutorial that introduces the framework of conformal prediction, and will provide an overview of both theoretical foundations and practical methodologies in this field. We will cover methods including holdout set methods, full conformal prediction, cross-validation based methods, calibration procedures, and more, with emphasis on how these methods can be adapted to a range of settings to achieve robust uncertainty quantification without compromising on accuracy. The second half of the course will cover more advanced material on related topics, e.g., algorithmic stability, conformal inference for time series.

Plenary talks

Learn how to do from what you see: Causal Effect Estimation using Proxy Variables (Arthur Gretton)

Estimating the true effect of an action, whether a medical intervention or a business decision, is challenging when unmeasured factors influence both the treatment and the outcome. In this talk, we will explore a framework that addresses this hidden confounding problem using proxy variables: observable information that provides indirect clues about the unmeasured factors. Using machine learning tools such as kernel methods and deep neural networks, we can build mathematical "bridges" between these proxies to obtain the true causal effect. I will describe both the original "outcome bridge" procedure, a more recent "treatment bridge" procedure, and a doubly robust estimator that combines the two. I will further demonstrate how the proxy approach may be used in the setting of domain adaptation, without making the classical assumptions of covariate or label shift.

When One Person Changes the Graph: The Statistical Cost of Node-Private Community Detection (Olga Klopp)

Network data are powerful because they encode relationships, but this is also what makes them delicate: protecting a single individual may require protecting not only their presence in the graph, but also all of their incident edges. This talk studies the consequences of this stronger, node-level form of differential privacy for one of the canonical problems in network inference: community detection in stochastic block models.

The central question is simple to state: can we still recover the hidden communities of a sparse network while protecting each participant as an individual? A naïve sensitivity argument suggests a pessimistic answer, since changing one vertex may modify linearly many observations. The results behind this talk reveal a sharper picture. Exact recovery remains possible under pure node-level differential privacy by combining the exponential mechanism with a high-probability degree envelope and a Lipschitz extension argument. At the same time, privacy imposes an unavoidable logarithmic price: in the logarithmic-degree sparse regime, a privacy budget of order logn is both sufficient and necessary for polynomially strong recovery guarantees.

More broadly, the talk will explain how the recovery error decomposes into two forces: the classical non-private statistical signal and a second term governed by the privacy budget. This gives a rate-level map of the tradeoff between privacy and inference, and points toward a general methodology for reconciling sharp likelihood-based statistical procedures with strong, individual-level privacy constraints.

Applications and registration fees

[Applications and registration for the summer school are now closed.]

Please submit a CV, contact details of a referee, an abstract of your recent or ongoing work (if you are interested to present) and indicate whether you would like to present your work as a poster or as a 10-min short talk. Your application will then be confirmed via email. The deadline for applications is 10^th December 2025.

All accepted applicants will then be required to pay a registration fee to confirm their place. The fees are as follows:

• £150 for research students;
• £250 for early career researchers (within five years of completing their PhD, excluding career breaks).

Supported by