4TU+.AMI Community Event 2026

Title: Statistical Models for Dimension Reduction of Temporal Multi-Omics Data

Abstract: Dimension reduction is a longstanding topic in statistics. However, modern biological studies increasingly involve complex, high-dimensional multi-omics data collected over time. This setting introduces challenges such as temporal dependence and high dimensionality, which are not well addressed by standard approaches like linear mixed models.

In this talk, I present an extension of the PO2PLS framework, which simultaneously performs dimension reduction and models temporal correlations through random effects. Parameters are estimated via maximum likelihood using the EM algorithm, with inference based on likelihood ratio tests. The method yields subject-specific scores for detecting outliers and enables ranking of metabolites based on explained variance and temporal dynamics. Moreover, the method is computationally efficient.

Application to two metabolomics datasets from the TwinsUK study provides biologically interpretable results and new insights. For a subset of variables, the approach recovers the same metabolite rankings as standard longitudinal models, while scaling effectively to higher-dimensional data.

About the speaker: Jeanine Duistermaat is a Professor of Applied Statistics at Eindhoven University of Technology. She has worked on statistical genetics, statistical bioinformatics, and modeling data along a continuum, such as weather data and images. Since 2024, she has been a member of the Royal Netherlands Academy of Arts and Sciences (KNAW).

Title: Mathematics running through our veins

Abstract: This presentation focuses on methods and algorithms for personalizing mathematical models of the cardiovascular system using clinical data. These models are based on fluid and solid mechanics and span multiple scales. Therefore, not only is their robust numerical solution challenging, but they also need to be integrated with data in cost-effective data assimilation frameworks.

About the speaker: Cristóbal Bertoglio is an Associate Professor of Computational Mathematics at the University of Groningen. His scientific interests include numerical fluid mechanics, inverse problems in fluid-solid interaction, magnetic resonance imaging, and the clinical translation of cardiovascular modeling. His group develop mathematical models and numerical methods for studying the cardiovascular system, with a focus on integrating data and models at different levels of complexity. 

Title: Bias-Optimal Bounds for SGD: A Computer-Aided Lyapunov Analysis

Abstract: Stochastic Gradient Descent (SGD) has become a cornerstone algorithm in optimization, popular for training a wide range of machine learning models. The typical non-asymptotic analysis of SGD requires assumptions to be made on the problem, which are not always easy or possible to verify in practice. In this work, we provide an analysis that does not make such unrealistic assumptions. We extend previous studies to incorporate a larger range of step-sizes, and obtain bounds which are, in some sense, optimal. The proof design is inspired by the Performance Estimation Problem framework, which both guides our proof and numerically demonstrates the optimality of our bounds.

The talk is based on the following paper: https://arxiv.org/abs/2505.17965

About the speaker: Daniel Cortild is a DPhil (PhD) student at the Mathematical Institute of the University of Oxford. His research interests include convex optimization, fixed-point iterations, computer-aided proofs, stochastic optimization, and hierarchical optimization.

Pitches from  Sara Giarrusso (Eindhoven), Gregor Gantner (Twente) and Maike Meier (Groningen).

Sara Giarrusso is an Assistant Professor at the department of Mathematics and Computer Science at TU/e. Her research lies at the interface between physics, chemistry, and mathematics, and aims at developing reliable and practical methods to describe electronic structure from a density-functional-theory standpoint.

Gregor Gantner is an Associate Professor in the Mathematics of Computational Science group at the University of Twente. He is interested in the adaptive numerical methods for partial differential equations, in particular with finite element methods, boundary element methods, and space-time methods.

Maike Meier is an Assistant Professor in the Computational and Numerical Mathematics group at the University of Groningen. Her research interests lie in numerical linear algebra and inverse problems, specifically using approximate computing techniques such as randomization and mixed-precision arithmetic to speed-up computations for large linear systems. 

Title: Structure-preserving model reduction: From the formulation on manifolds to data-driven realizations

Abstract: Capturing and preserving physical properties, e.g., system energy, stability and passivity, using data-driven methods is currently a highly-researched topic in surrogate modeling. To ensure that the desired physical properties are retained, structure-preserving projection techniques are used in the field of model order reduction (MOR).  In this talk, we present structure-preserving MOR with nonlinear projections, which are needed for problems with slowly decaying Kolmogorov n-widths. To precisely define and highlight the quantities that we would like to retain, we start with a formulation of initial value problems on manifolds, which we consider as the full-order model (FOM). Already at this level, we define what we mean by adding structure to the FOM and how this can be detailed geometrically. This formalism allows to introduce a novel projection technique, the generalized manifold Galerkin (GMG). By adapting the underlying non-degenerate tensor field, this GMG projection can be used for a structure-preserving reduction of various initial value problems that give rise to interesting physical properties, which include, but are not restricted to, Lagrangian and (port-)Hamiltonian systems. Once that we have derived the geometric formulation, we focus on data-driven ansatzes to realize the presented reduction methods. 

About the speaker: Silke Glas is an Assistant Professor in the Department of Applied Mathematics at University of Twente, where her research focus is on model reduction, especially structure-preserving model reduction and model reduction on manifolds.

Title: Convergence of energy-based learning in resistive circuits 

Abstract: Energy-based learning is a biologically plausible alternative to the popular back-propagation method for training artificial neural networks. It deals with models that are governed by a parameterized energy function. Learning is achieved by shaping the energy function such that its minima coincide with given data. The energy-based perspective is a promising avenue for training analog circuits in an energy-efficient manner, especially because training can be performed using relatively simple, local parameter updates. The purpose of this talk is to make some steps towards a theoretical understanding of energy-based learning, applied to nonlinear resistive networks. For these networks, we propose an energy-based learning algorithm and we establish conditions under which this algorithm converges to a suitable vector of parameters, explaining the data.

About the speaker: Henk van Waarde is an Assistant Professor in the Systems, Control and Optimization group at the University of Groningen. His research interests include direct data-driven control, system identification, kernel-based modeling of (physical) dynamical systems, experiment design, and pplications to networked systems and neuromorphic computing. He aims to develop a systems and control theory that is grounded on measured data.

There will be a dinner at the end of the evening. Participants of the dinner will be responsible for their own expenses.