AI Research Seminar: Tom Lefebvre, Manfred Opper, Tolga Birdal, Jonas Degrave

AI Research Seminar: Tom Lefebvre (UGent), Manfred Opper (TU Berlin), Tolga Birdal (Imperial College London), Jonas Degrave (Deepmind)

We are pleased to invite you to a research seminar on a broad range of topics by top AI researchers. After the talks, there will be time for discussions and networking during the coffee break.

Location: Auditorium 1, first floor, iGent Tower, Technologiepark-Zwijnaarde 126, 9052 Gent.

Schedule:
14:00 - 14:30: Tom Lefebvre (UGent)
14:30 - 15:00: Manfred Opper (TU Berlin)
15:00 - 15:30: Tolga Birdal (Imperial College London)
15:30 - 16:00: Jonas Degrave (Deepmind)
16:00 - 17:00: Coffee break

17:00 - 18:30: PhD defense of Rembert Daems (separate registration)

Denoising Entropy Regularization in Optimal Control
Tom Lefebvre, Ghent University

Entropy regularized optimal control problems have become common practice in the robotics and reinforcement learning community. One of the major merits of entropy regularization is the increased tractability of the resulting optimization problem which is happily exploited by the many algorithms that derive from the paradigm. However, the fault is in the merit, since one does no longer solve the original problem, spawning the question what problem one does solve? In my talk I will navigate the landscape of entropy regularization in optimal control, illustrate exact relations with traditional optimal control problems, Maximum Likelihood Estimation problems, the theory of Linearly Solvable Optimal Control and, ultimately, the ubiquitous Model Predictive Path Integral control method.

Tom Lefebvre received the M.Sc. degree in control engineering and automation from Ghent University, Belgium, in 2015, and the Ph.D. degree in electromechanical engineering in 2019. Since 2019, he has been a post-doctoral research assistant. His work is positioned at the intersection of estimation, control, and, machine learning. His current research focuses on combining adaptive control architectures with active learning strategies.

Bayesian Inference for Point Process models - A latent variable approach
Manfred Opper, TU Berlin and Centre for Systems Modelling and Quantitative Biomedicine, University of Birmingham

In my talk I consider the nonparametric Bayesian estimation of intensity functions for spatio (- temporal) point processes. As an example I discuss an inhomogeneous Poisson process and its generalisation to a model of earthquake occurrences. To model the spatial variability of intensity functions, a popular assumption is a Gaussian process prior probability distribution together with a nonlinear link function. This model leads to a non-Gaussian and infinite dimensional inference problem. I will show that this problem can be made tractable by augmenting the model with various extra latent random variables. This will allow for efficient inference, either by variational approximations or by Monte Carlo sampling.

Manfred Opper received his Ph.D. in Physics and his Habilitation in Theoretical Physics from the University of Giessen, Germany. He has held posts at Aston University and the University of Southampton in the UK. From 2006 until his retirement in 2021 he has been Professor of Methods in Artificial Intelligence at TU Berlin. Since 2020 he is a part time Professor of Machine Learning at the University of Birmingham (UK). From 2022- 2025 he was visiting professor at the University of Potsdam. He is interested in the development and theoretical analysis of methods for probabilistic inference in machine learning using techniques from statistical mechanics and statistics. In recent years, he has worked particularly in the area of data assimilation, where he has developed approximations for inference in stochastic differential equations, point processes, and dynamical processes on large networks. He also works on the relationship between inference and optimal stochastic control, and on the application of random matrix theory to machine learning algorithms.

Topological Deep Learning and Its Applications in Drug Design
Tolga Birdal, Imperial College London, UK

Topological deep learning is a rapidly growing field that pertains to the development of deep learning models for data supported on possibly higher-order topological domains such as simplicial complexes, cell complexes, and hypergraphs, which generalize many data modalities encountered in scientific computations and drug design. In this talk, Tolga will present this unifying deep learning framework and its components, while introducing a new domain called combinatorial complexes. Next, building this rich combinatorial and algebraic structure, Tolga will develop a general class of message-passing neural networks as well as transformers. Tolga will also address a bunch of applications from CAD design to molecular modeling, which paves the way to a new way to drug design and shows great promise for AI4Science.

Tolga Birdal is an assistant professor in the Department of Computing of Imperial College London. Previously, he was a senior Postdoctoral Research Fellow at Stanford University within the Geometric Computing Group of Prof. Leonidas Guibas. Tolga has defended his masters and Ph.D. theses at the Computer Vision Group under Chair for Computer Aided Medical Procedures, Technical University of Munich led by Prof. Nassir Navab. He was also a Doktorand at Siemens AG under supervision of Dr. Slobodan Ilic working on “Geometric Methods for 3D Reconstruction from Large Point Clouds”. His current foci of interest involve topological / geometric machine learning and 3D computer vision. More theoretical work is aimed at investigating and interrogating limits in geometric computing and non-Euclidean inference as well as principles of deep learning. Tolga has several publications at the well-respected venues such as NeurIPS, CVPR, ICCV, ECCV, ICLR, T-PAMI, ICRA, IROS, ICASSP and 3DV. Aside from his academic life, Tolga has co-founded multiple companies including Befunky, a widely used web-based image editing platform.

Science While Falling into the Singularity
Jonas Degrave, Deepmind, UK

Jonas Degrave works at DeepMind since 2017. There he developed algorithms to control nuclear fusion reactors with reinforcement learning, with work published in Nature and featured in international media such as the New York Times and the BBC. Today he works on Project Astra, where he researches the interface between humans and AI. He obtained his Ph.D. at Ghent University in 2018 with the title “Incorporating a-priori knowledge into deep neural networks for controllers of robots with legs.” He also exhibited new media art at Speculum Artium, with positive reviews of the work in the New Scientist, The Guardian and Der Spiegel. Jonas is an advocate of developing the singularity as quickly as possible, but argues that society needs to prepare more quickly and is often naive about what a future along with AGI might look like.