The program of this year's summer school will focus on Graph Neural Networks, covering both fundamental concepts and practical applications.
Participants can expect a mix of lectures and interactive sessions that provide insight into the mathematical foundations as well as real-world use cases of graph-based machine learning methods.
 
KCDS members as well as doctoral researchers from KIT and other universities/ research centers are welcome to join! There is no participation fee. Please note that KCDS doesn't cover travel and accomodation expenses.

Workshop: October 12+13, 2026 - places limited, registration necessary Poster session: October 12, 2026, 4.45 pm, KIT Campus South, Building 10.81, in front of the Theodor-Rehbock-Hörsaal (HS59). Posters will be presented by workshop participants. Attendance without presenting a poster is possible without registration. Keynote lecture: October 13, 2026 at 4.30pm, KIT Campus South, NTI Lecture Hall - no registration  
For more than two centuries, least-squares regression has been a cornerstone of statistical practice, while classical nonparametric smoothing methods have long served as standard tools for analysing complex data. In this workshop, we will revisit these methods from a modern perspective and ask: Are we making the best possible use of them? Recent work in statistical theory by Richard Samworth and others shows that these familiar methods can often be improved by incorporating additional structural information, such as shape constraints or properties of the underlying error distribution.
 
This two-day workshop, led by Richard Samworth, will explore recent developments in distributionally adaptive statistical methods. Richard Samworth is Professor of Statistical Science and Director of the Statistical Laboratory at the University of Cambridge. A Fellow of the Royal Society, he is the recipient of numerous distinctions, including the COPSS Presidents' Award, the David Cox Medal, and, in 2025, the Royal Statistical Society Guy Medal in Silver. His research has made fundamental contributions to nonparametric statistics, statistical learning theory, and high-dimensional methodology, particularly in shape-constrained estimation and adaptive nonparametric procedures. A defining feature of his research is the combination of rigorous theoretical guarantees with methods designed to be computationally efficient and practically applicable.
 
The first day will focus on linear regression and shape-constrained estimation. Starting from the classical least-squares framework, the workshop will examine how structural information, including monotonicity, can be used to improve estimation and inference. The second day will turn to nonparametric regression. It will begin with local polynomial methods and their theoretical foundations before introducing recent extensions, including Outrigger local polynomial regression, which adapts to the underlying error distribution while retaining strong theoretical guarantees.  
The theoretical lectures will be complemented by practical sessions in R and Python, allowing participants to apply the methods discussed during the workshop. A joint poster session, with a particular focus on early-career researchers, will provide an opportunity to present ongoing work, exchange ideas across disciplines, and receive feedback from other participants and senior researchers. The poster session on Monday and the plenary talk on Tuesday will also be open to researchers from the university and neighbouring institutions.
 
The workshop is primarily intended for doctoral candidates and postdoctoral researchers from the KCDS Graduate School and the Heidelberg Graduate School MathComp, as well as members of the Helmholtz Association and researchers in related fields who have a strong interest in modern mathematical statistics.
Participants will gain insight into current developments in adaptive statistical methodology and their connections to broader challenges in statistical learning and modern data analysis.
 
The workshop and keynote lecture are organized by the Institute of Statistics (STAT) in cooperation with MathSEE / KCDS and HGS MathComp at Heidelberg University. The workshop was made possible through Course Funding from HIDA, which supported its development and implementation.

The workshop for registered participants is complemented by a keynote by Richard Samworth to which everyone interested is invited, no registration necessary!
 
For more than two centuries, least-squares regression has been a cornerstone of statistical practice, while classical nonparametric smoothing methods have long served as standard tools for analysing complex data.  In this workshop, we will revisit these methods from a modern perspective and ask: Are we making the best possible use of them? Recent work in statistical theory by Richard Samworth and others shows that these familiar methods can often be improved by incorporating additional structural information, such as shape constraints or properties of the underlying error distribution.
 
Richard Samworth will explore recent developments in distributionally adaptive statistical methods. Richard Samworth is Professor of Statistical Science and Director of the Statistical Laboratory at the University of Cambridge. A Fellow of the Royal Society, he is the recipient of numerous distinctions, including the COPSS Presidents' Award, the David Cox Medal and the Royal Statistical Society Guy Medal in Silver in 2025. His research has made fundamental contributions to nonparametric statistics, statistical learning theory, and high-dimensional methodology, particularly in shape-constrained estimation and adaptive nonparametric procedures. A defining feature of his research is the combination of rigorous theoretical guarantees with methods designed to be computationally efficient and practically applicable.
 
Richard Samworth will present recent work with Elliot H. Young and Rajen D. Shah on "Outrigger local polynomial regression“. A preprint of the paper is available on arXiv https://arxiv.org/abs/2603.11282.