Invited speakers

About the talk

The allure and the challenges of deep reinforcement learning

In a nutshell, reinforcement learning (RL) means making good decisions by learning from experience. The generality of this problem formulation, and the prospect of minimal supervision, make RL a plausible candidate for solving tomorrow’s problems. RL has been demonstrated to work at scale, in particular when combined with deep learning. In this talk I’ll expand on the key challenges that remain, and illustrate them using the case-study of AlphaStar, a deep RL agent that learned to play the game of StarCraft II at grandmaster level.

About the talk

Machine learning-based research strategies – a game changer for science?

Research is increasingly driven by machine learning strategies for data analysis, but also discovering new hypotheses and findings. Material science, chemistry, astrophysics, meteorology and many more disciplines are experimenting with machine learning settings of various types for various purposes. The keynote lecture will present examples from ongoing research projects based on machine learning strategies and will discuss the impact of AI on scientific research in general.

About the talk

Challenges in real-time 3D imaging, and how machine learning comes to the rescue.

Tomography is a powerful technique for visualizing the interior of an object in 3D from a series of its projections, acquired at different angles during a tomographic scan. At the heart of the technique is a mathematical inverse problem (known as “reconstruction”). At present, the steps of image acquisition, reconstruction, and analysis are usually carried out sequentially, often analyzing the data after the scan has long finished.
In this lecture I will present the research of my team, along with many collaborators, to develop algorithms, and software that make the entire tomographic imaging pipeline work in “real-time”, while the scan is taking place. Machine learning can address many of the algorithmic challenges involved, but certainly not in a plug-and-play manner. When scanning objects in 3D, we typically do not have access to a large database of training examples, the circumstances of each scan are different, and so are the internal properties of the objects. In addition, training times and memory requirements tend to increase drastically when scaling up from 2D to fully 3D imaging. I will go over each of these problems and present various strategies for making machine learning feasible and efficient in real-time 3D imaging environments. I will illustrate the results by providing examples from scientific research (real-time synchrotron tomography and real-time electron tomography), industry (real-time quality control) and cultural heritage (interactive technical art investigation).

About the talk

We aren’t doing AI research

There is a large diversity of excellent research in the BENELUX addressing different aspects of artificial intelligence. The Dutch AI Manifesto distinguishes seven of these aspects, called AI foundations. The creation of artificial intelligence is a multidisciplinary challenge that requires integration of several of these AI foundations. The Dutch AI Manifesto mentions three multidisciplinary challenges: Socially-Aware AI, Explainable AI, and Responsible AI. Although these are important challenges, they do not address the creation of artificial intelligence. If we claim to do AI research instead of only creating smart solutions, then we should also address the multidisciplinary challenge of creating artificial intelligence. In the talk, I will address several aspect of this multidisciplinary challenge.

About the talk

AI for industrial decision-making

Operational decision-making problems are everywhere, from train maintenance scheduling, ambulance dispatching, to order batching in warehousing. These problems are typically studied in the
Operations Research community, where expert and domain knowledge is used to find good solutions. In this talk, I will introduce new research directions my team have been working on these years that
investigate how data can augment traditional decision making methods. I will use examples from practice to illustrate our approaches and the emerging research challenges.

About the talk

Neuro-Symbolic = Neural + Logical + Probabilistic

The overall goal of neuro-symbolic computation is to integrate high-level reasoning with low-level perception, or phrased differently System I and System II. We argue 1) that neuro-symbolic computation should integrate neural networks with the two most prominent methods for reasoning, that is, logic and probability, and 2) that neuro-symbolic integrated methods should have the pure neural, logical and probabilistic methods as special cases. We make some observations w.r.t. the state-of-the-art with regard to these claims.

This is joint work with Robin Manhaeve, Sebastijan Dumancic, Giuseppe Marra, Thomas Demeester and Angelika Kimmig.