Flanders AI Research Day 2025

Advancements in Explainable Graph-Based Deep Learning            De Krook - Blauwe Vogel

Prof Nikos Deligiannis (VUB, imec)

• This talk explores recent advancements in graph-based deep learning, with a particular focus on enhancing model interpretability and applying these methods to real-world challenges. We delve into the integration of physical knowledge with data-driven models to improve prediction accuracy in domains such as environmental monitoring, social media analysis, and image processing. The first part of the talk addresses how we combine graph convolutional networks (GCNs) with physical models to create more robust solutions for tasks like air quality forecasting. The second part focuses on the explainability aspect of these models, where we introduce novel techniques for interpreting complex graph-based deep learning architectures. We will present methods that utilize Layer-wise Relevance Propagation (LRP) to provide interpretable insights into model decisions, making them more transparent and trustworthy. Finally, the talk covers how these explainable graph models are applied to a variety of domains, from detecting fake news to segmenting images, highlighting their versatility and the critical role of interpretability in real-world applications.

Continual Learning at the Edge           Music Club

Prof Marian Verhelst, Chao Fang, Jonathan Swinnen, Stef Cuyckens, Karen Vanhalle (KULeuven)

• AI systems are increasingly being deployed on resource-constrained edge devices that operate in dynamic environments with changing data distributions. Traditional static models cannot adapt to these evolving conditions, leading to degraded performance over time. This necessitates adaptive learning systems that can continuously learn without catastrophic forgetting, which becomes crucial for real-world applications. However, the severe power limitations of edge devices make energy-efficient continual learning a critical challenge. This deep dive explores how AI systems can achieve energy-efficient continuous learning while operating under these resource constraints. The presentation offers three complementary perspectives. First, we introduce continual learning from an algorithmic standpoint, reviewing current benchmarks and methods while discussing the gap between laboratory settings and practical applications such as on-site fine-tuning and streaming data processing. Second, we examine artificial neural network-based hardware approaches for energy-efficient continual learning, introducing application-specific integrated circuit designs, reduced bit-width operations, and hardware-software co-optimization strategies that enable practical edge deployment while potentially reducing network complexity. Finally, we present neuromorphic computing as a brain-inspired approach for efficient continual learning. Drawing inspiration from the brain's remarkable energy efficiency and natural learning capabilities, we introduce spiking neural networks, neuromorphic chips, and bio-inspired continual learning algorithms as promising directions. This deep dive hence spans the path from algorithmic innovations to hardware implementations, demonstrating how energy-efficiency techniques across the entire computing stack can enable practical continual learning systems for edge applications.

Foundation Models for Interaction           Auditorium

Ruben Janssens, Prof. Tony Belpaeme, Eva Verhelst, Joachim De Baer (UGent)

• Foundation models such as Large Language Models (LLMs), speech models, and image-generating diffusion models, have radically transformed our ability to design interactive systems. For the first time, computers can engage in unconstrained, unscripted conversations with people, and generate infinitely personalizable creative content on-the-fly. This deep-dive session will show three projects that explore the capabilities of foundation models for interactive systems as part of the Flanders AI Research Program, with Prof. Tony Belpaeme (IDLab-AIRO, UGent-imec) as our guide through the session. We'll first look at how Vision-Language Models (VLMs) enable social robots to use visual context in social conversations with their users, with a talk given by Ruben Janssens (IDLab-AIRO, UGent-imec). Bringing social robots to language learning, Eva Verhelst (IDLab-AIRO, UGent-imec) will then show recent progress on using LLMs and diffusion models to co-create learning content with users. Finally, Joachim De Baer (IDLab-T2K, UGent-imec) will talk about chatbots in HR applications: will job interviews be automated soon?

Pitch: Beyond Words: Multimodal Approaches to Empathy in Conversational Agents
Sadegh Jafari (UGent)

• This presentation explores how empathy can be modeled in multimodal conversational agents by integrating text, speech, and visual signals. I will highlight current limitations in capturing and expressing empathy across modalities and propose a framework that uses large language models, emotional prompts, and reinforcement learning to address these gaps. The goal is to create more emotionally aware and engaging AI systems.

From Cancelling Variables to Building Polynomial Models           Collective A

Christof Vermeersch, Hans van Rooij, Sarthak De (KULeuven)

• This deep dive session will explore how elimination, essentially a linear technique, can be used to build polynomial models via realization theory. We want to demonstrate how elimination theory is not only useful for variable cancellation in multivariate polynomial systems, but also has a role in data-driven modeling. After a concise introduction on elimination and realization, two complementary talks will discuss and apply numerical elimination.
  Talk I: In algebraic geometry, elimination theory is used for reducing complexity by systematically cancelling some variables while preserving the algebraic relationships among the remaining ones. The first talk will introduce a numerical elimination approach that operates on the Macaulay matrix, whose (right) kernel encodes all algebraic dependencies between the monomials, using numerical linear algebra algorithms. We will show how elimination can be interpreted as finding the intersection between this kernel and an elimination space that encodes the variables to be removed.
  Talk II: After understanding the numerical machinery, we will present the model matrix, a parametric representation of a polynomial state-space model whose left kernel contains the difference equation that captures temporal correlation between consecutive outputs. This yields an orthogonal complementarity between the difference equation and the state-space model. We will repurpose our elimination procedure to reveal, rather than remove, the hidden state variables that transform the difference equation into a state-space representation. Exploiting the orthogonality, we can recover the vector space governing the model's parameters and deploy the above-mentioned intersection-based approach to realize the minimal state-space model. So, elimination helps building polynomial models.

Human-Aware Technology for Optimization           De Krook - Twistappel

Tias Guns (KULeuven)

• Combinatorial optimisation is widely used to solve scheduling, sequencing, rostering, routing and other assignment problems. The traditional paradigm is that of model+solve, where an expert user expresses their problem in constraints and a highly efficient solver searches for the optimal solution.
  To make this approach more human-aware, we will revisit this into the model+solve+explain paradigm. We'll cover recent work on LLM-based translation from natural language to formal combinatorial problem specifications; on tools to compare CP/ILP/PB/SAT solvers from a single specification; and on explainable constraint solving techniques that allow explaining why there is no (better) solution, or how to best deal with disruptions.
  We end the talk by showing how we are further developing these techniques on Flanders Make's multi-agent high-mix low-volume assembly scheduling use case, where human and robot workers collaborate in specialized product assembly.

The potential of predictive modelling for next generation clinical decision support in dermatology therapy optimisation, with case examples of psoriasis           Expo2

Bram Steenwinckel, Sofie Van Hoecke, Femke Ongenae, Ayko Chevaillier, Jo Lambert (UGent), Rani Soenen and Anke Eylenbosch (UZGent)

• This deep dive session explores how artificial intelligence is advancing precision medicine in psoriasis by enabling predictive, personalized, and value-based care. Topics include, among others, AI models that distinguish responders from non-responders by integrating clinical, psychosocial, and economic factors to optimize treatment selection and improve long-term outcomes. We also highlight the prediction of next-visit PASI scores, showcasing how machine learning can support dermatologists in forecasting disease severity based on patient history and contextual information. In addition, we present the application of Model-Informed Precision Dosing (MIPD), which leverages population pharmacokinetic models to tailor biologic therapies beyond traditional therapeutic drug monitoring (TDM).

Pitch: From Data to Personalized Decision for Psoriasis Therapy Optimization
Ayko Chevaillier (UGent)

• Psoriasis is a chronic skin condition that affects patients’ quality of life. Current treatment outcomes are highly variable and many patients do not respond optimally to biologic therapies. This leads not only to prolonged suffering and reduced well-being but also to significant healthcare costs due to ineffective treatment cycles and disease complications.
  My research addresses this challenge by applying machine learning to model disease progression and treatment response. By predicting the Psoriasis Area and Severity Index (PASI) at future visits and define for which patients a particular treatment is effective, my models can help to create more proactive, personalised treatment strategies. Early results with models such as Random Forests, XGBoost, and CatBoost, where I not only incorporate clinical features but also psychosocial factors, show strong predictive performance. This highlights the potential to reduce disease worsening, enhance quality of life, and lower healthcare costs by optimising therapy selection and minimising ineffective treatments.
  By enabling more precise forecasting of clinical scores, classifying responders and more advanced tasks such as survival analysis of time-to-clearance, my work has the potential to improve patients’ lives and drive evidence-based updates to clinical guidelines.

Tools for responsible AI innovation           De Krook - Laatkomer

Pieter Duysburgh (VUB), Rob Heyman, Neena Singh (Knowledge Centre Data & Society, imec-SMIT, VUB)

• Technology never exists in social isolation. There are societal views and expectations, power dynamics and ethical considerations that can determine whether an innovation thrives, stalls, or fails entirely.
  This interactive session by the Knowledge Centre Data & Society combines a theoretical perspective on the relation between technology and ethics, with practical tools and insights on how to make ethical and social considerations part of your AI innovation trajectory.
  Some frameworks that will be presented include the shortened ALTAI (Assessment List for Trustworthy AI), the guidance ethics approach, and ALLY (building blocks for developing a organisational strategy for responsible AI). We will give you a hands-on experience with these tools, providing the first step in how they can enhance your innovation process and heighten the success of your AI project.

Towards Fairer Language Models           Bar Bougie Left

Pieter Delobelle

• This talk will explore how biases in AI language models can lead to unfair outcomes in real-world applications. We will examine methods for measuring and mitigating these biases, particularly across different languages. We will also address broader safety concerns and technical approaches to creating more fair and responsible AI systems.

Tutorial: Data-Efficient Modeling and Optimization           Bar Bougie Right

Sebastian Rojas Gonzalez, Prof. Ivo Couckuyt, Prof. Tom D'haene (UGent, imec)

• An increasing number of scientific and engineering fields are confronted with expensive experiments, such as costly physical lab experiments or computer simulations, to study complex, real-world phenomena or solve challenging design problems. When dealing with such expensive experiments or process measurement data, surrogate modeling methods are firmly established as facilitators for design space exploration, sensitivity analysis, uncertainty quantification, optimization, visualization, etc.
  For the sub-class of low-dimensional black-box optimization problems that are expensive to evaluate, Bayesian learning and optimization has become a very popular data-efficient alternative, due to its flexibility, tractability and built-in uncertainty quantification. This tutorial will introduce the standard (Bayesian) framework for expensive black-box optimization, and will then delve into advanced topics like handling noisy data and dealing with multiple design objectives. Finally, we will demonstrate the practical value of these methods on several real-life use cases. We will do live hands-on examples with Python Notebooks.
  Participants of this tutorial will: 1) Learn or review the basics of Bayesian (active) Learning; 2) Understand the standard Bayesian Optimization framework; 3) Learn how to quantify and exploit the uncertainty in such methods; 4) Learn the state-of-the-art on how to handle noisy data and multiple design objectives; 5) See an algorithm performing an actual expensive optimization task.

Tutorial: Open-source tools for visual exploration of complex data           Collective B

Jefrey Lijffijt (UGent), Edith Heiter (UGent), Fuyin Lai (UGent), Laura Hajzoková (UGent, VIB)

• In this tutorial we will teach you how to explore, interpret, and discover new patterns from complex data using data visualisation techniques. We will use the tools developed by the AIDA & DAMBI groups at UGent/VIB under the FAIR program, such as TRACE and other open-source code. We developed these tools for the analysis of data in the domain of single-cell biology, but they are applicable to any domain. To run the tools hands-on, you need a laptop where you can create custom Python environments. If you have own data available (in tabular form), it may be feasible to explore your own data during the session. Participants of this tutorial will gain state-of-the-art skills in exploratory data analysis.

Our Use Cases / Demo's

• Health

  • Fitness on the edge

  • Intensive Care Units

  • AI for Sports

  • Sharpen Tumor Insights with HSI

  • Epidemic Decision Making

  • Personalized Therapy Dermatology

  • Cardio Digital Twin

  • Single Cell Technologies

  • Medical Imaging

  • Real-World Evidence

  • Wearable Physiological Monitoring @ Home

• Planet & Energy

  • Living Environment

  • Renewable Energy Production

  • Smart Grid - Decision Support for The Low Voltage Distribution Grid

• Society

  • Education & Training

  • Multi-modal conversational explainable systems

  • Digital Humanities

  • Public Employment Services

  • Collaborative Learning

  • Mobility: intelligent intersections

• Industry

  • Smart Machines - Monitoring & Control

  • Robotics for Agrifood

  • Prognostics and Health Management for Assets

  • Human-Machine Flexible Production

  • Autonomous Systems in Manufacturing Floor

  • Straight-Through-Digitalization

  • Production Optimization

  • Monitoring & Control of Production & Machine

Our posters

• Resurfacing the Instance-only Dependent Label Noise Model through Loss Correction
  Mustafa Enes Aydın (KU Leuven)

• o3 (mini) Thinks Harder, Not Longer
  Marthe Ballon (Vrije Universiteit Brussel)

• PHYSICS-INFORMED MODEL-BASED REINFORCEMENT LEARNING FOR PANDEMIC MITIGATION
  Veronica Biancacci (Vrije Universiteit Brussel)

• CSD-AFNet: Computationally Efficient Atrial Fibrillation Classification from ECGs using 2D Causal Strided Dilated Convolutions
  Lennert Bontinck (University of Ghent)

• Structured querying on tabular embeddings - an experimental comparison
  SEBASTIAN BUGEDO BUGEDO (UHasselt)

• SubTSMD: discovering subspace motifs with temporal variations in high-dimensional time series
  Louis Carpentier (KU Leuven)

• Cash or Comfort? How LLMs Value Your Inconvenience
  Mateusz Cedro (University of Antwerp)

• Improving Tree-Structured Parzen Estimators for Bayesian Optimization
  Vipul Chalotra (UHasselt)

• ReferGPT: Towards Zero-Shot Referring Multi-Object Tracking
  Tzoulio Chamiti (Vrije Universiteit Brussel)

• Virtual staining of pathology slides
  Thomas De Kerf (University of Antwerp)

• Relational neurosymbolic Markov models
  Lennert De Smet (KU Leuven)

• Interpretable Concept-Based Memory Reasoning
  David Debot (KU Leuven)

• Programmatic Reinforcement Learning
  Senne Deproost (Vrije Universiteit Brussel)

• Rashomon Sets: A Path Toward Fairer AI
  Gilles Eerlings (UHasselt)

• Structured Light + Machine Learning: Real-Time 3D Quality Inspection
  Rhys Evans (KU Leuven)

• Decision Support for High Mix Low volume Scheduling
  Marco Foschini (KU Leuven)

• Graph Connectionist Temporal Classification for Phoneme Recognition
  Henry Grafé (KU Leuven)

• Uncertainty Calibration via Proper Scores for Trustworthy AI
  Sebastian Gruber (KU Leuven)

• Oxytrees: Model Trees for Bipartite Learning
  Pedro Ilídio (KU Leuven)

• Towards Reliable Automatic Evaluation of Emotional Text Generation
  Sadegh Jafari (University of Ghent)

• Evaluating Visual Context in Social Human-Robot Dialogue
  Ruben Janssens (University of Ghent)

• A calibration test for evaluating set-based epistemic uncertainty representations
  Mira Juergens (University of Ghent)

• Leveraging Expert Knowledge in Bayesian Optimization for Engineering Design
  Wouter Koppen (University of Ghent)

• Bayesian Optimization for Microwave Design
  Yens Lindemans (University of Ghent)

• RNNPerf: Unified Prediction of Performance and Energy for Distributed DNNs
  CHENGJIN LYU (University of Ghent)

• Embeddings as Equivalence Relations in Logic Programming
  Jaron Maene (KU Leuven)

• Probabilistic Predictive Control with Exploratory Uncertainty Quantification: An Autonomous Vehicle Usecase
  Mohammad Mahmoudi Filabadi (University of Ghent)

• Exploration Guidance using Unlabeled and Incomplete Data in Reinforcement Learning
  Elias Malomgré (University of Ghent)

• A Method for Optimizing Connections in Differentiable Logic Gate Networks
  Wout Mommen (imec)

• A Verification Framework for Bounding Decision Quality in Predict-and-Optimise Models
  Anna Münz (Aachen University)

• An evidence-based neuro-symbolic framework for ambiguous image scene classification
  Giulia Murtas (Sirris)

• Koopman Theory vs. Physics-Informed Neural Networks for Cell-Signaling Dynamic Inference.
  Sabrina Negroni (VIB)

• Zone-Adaptive Bayesian Active Learning for Feasible Region Identification
  Ioana Nikova (University of Ghent)

• Robust 3D localization with visible light positioning
  Arman Nikraftar (University of Ghent)

• Data efficient robust 3D visible light positioning
  Arman Nikraftar Khiabani (University of Ghent)

• Raw2Event: Converting Raw Frame Camera into Event Camera
  Zijie Ning (KU Leuven)

• When and how to transform the target variable?
  Loren Nuyts (KU Leuven)

• LLM-Based Schema Matching for Real-World Data
  Marcel Parciak (UHasselt)

• Brain scanning a CNN - Visualization and optimization of a convolutional neural network through image morphing
  Christopher Patzanovsky (UHasselt)

• Using Weak Supervision for Schema Matching Tasks
  Dante Pinto (UHasselt)

• Enhancing Speech Understanding in LLMs via Phoneme-Aware Training and Multi-Encoder Fusion
  Jakob Poncelet (KU Leuven)

• Constraint-Guided PINNs: A Constrained Optimization Approach
  Wout Rombouts (KU Leuven)

• SigmaNicheNet: unravelling metabolite mediated cell-cell communication by inferring metabolite activity from downstream gene expression profiles
  Ruth Seurinck (VIB)

• Poking creativity in an LLM by seeding it with realtime open-ended interaction
  Marnix Van Soom (Vrije Universiteit Brussel)

• Clarity in complexity: a dashboard for navigating algorithmic fairness in HR
  Anna Verheyden (KU Leuven)

• Patch-based Querying Identifies Structures of Interest in Electron Microscopy Data
  Niels Vyncke (University of Ghent)

• Evaluating Epidemic Interventions with Hypergraphs
  Nick Wils (University of Antwerp)

• HERMES: A Unified Open-Source Framework for Realtime Multimodal Physiological Sensing, Edge AI, and Intervention in Closed-Loop Smart Healthcare Applications
  Maxim Yudayev (KU Leuven)

A Compositional Approach to Interpretability           Auditorium

Prof. José Oramas (UAntwerpen), Prof. Geraint Wiggins (VUB), Ward Gauderis (VUB), Thomas Dooms (UAntwerpen)

• Interpretability aims to gain insights into a neural network’s behaviour by extracting salient structure from its internal representations and computations. This offers researchers a targeted lens to make generalisable statements about a model and uncover learned knowledge.
  Current methods find structure through common properties such as low-rankness, sparsity constraints, clustering or hierarchical structures. These algorithms or proxy models rely on the assumption that such abstractions are meaningful to predicting model behaviour, which is unfortunately often not the case due to non-linearities in the network. Because each model component operates on a different manner their isolated interpretation is possible but their composition is destabilized; any joint behaviour is determined by the input, not by the composition of the parts. As a result, even perfectly disentangled components combine in an entirely input-dependent manner.
  In this deep dive, we argue that neural networks learn inherently entangled concepts and mechanisms that should be highlighted, rather than ignored. However, the current syntax used to define and analyse them is meaningful exclusively either at the level of individual neurons or of entire layers and therefore fails to support such intermediate mechanisms.  We will present a compositional tensor-based formulation that can provide the means to decompose the elements that define these architectures and analyse, with high faithfulness, the representations and computations therein.

AI Innovations for Environmental Modeling           De Krook - Twistappel

Jorge Sousa (VITO), Tanja Vanachteren (VITO) and Nikola Dukic (KULeuven).

• The “Living Environment” use case integrates advanced artificial intelligence methods into environmental modeling to improve decision support. This presentation combines three key innovations: self-supervised learning for urban representation, AI-driven remote sensing, and surrogate modeling for air quality applications.
  We apply self-supervised learning / few shot learning to very high-resolution aerial imagery, producing dense, object-centric urban representations that surpass traditional patch-based methods in capturing buildings, vegetation, and land cover. These methods will be illustrated through a hands-on demo, showcasing their potential for more scalable, transferable environmental analysis from remote sensing data.
  To complement this, we leverage AI for remote sensing by combining aerial data with satellite observations (e.g., Terrascope). This fusion enhances spatial and temporal monitoring of land use, environmental dynamics, and water quality, while AI-driven data augmentation addresses sensor gaps to deliver continuous time-series.
  Finally, we develop AI surrogate models trained on computationally intensive physics-based air quality models. These surrogates retain high fidelity while enabling rapid scenario analysis and operational forecasting. Complementary approaches, including recurrent and graph neural networks, are benchmarked to further advance forecasting capabilities.
  Together, these innovations demonstrate how AI can transform environmental modeling into scalable, adaptive, and decision-oriented tools for air quality and living environment applications.

Discovering Machine Learning with Knowledge Graphs           De Krook - Blauwe Vogel

Ioannis Dasoulas (KULeuven, Flanders Make)

• The sheer number of available machine learning (ML) artefacts including models, datasets, papers, and hyper-parameters is increasing in such a rate that it is often very difficult to navigate and discover optimal solutions. While there are many ML repositories (e.g., Kaggle, OpenML, Papers with Code) and conferences that publish ML artefacts, there is often lack of interoperability between them, lack of important metadata as well as inefficient recommendation techniques for users to find what they need. This work focuses on modelling ML information as inter-connected graphs to improve discoverability and recommendation of ML artefacts. We develop ontologies and taxonomies for modelling different aspects of the ML lifecycle (e.g., datasets, model training and evaluation) and their metadata and integrate ML information from diverse repositories to knowledge graphs, using this unified schema. In addition, we leverage external knowledge bases and LLMs to investigate enriching existing metadata and, hence, the relationships between ML artefacts. Finally, we leverage knowledge graph embedding techniques, to represent ML artefacts in the latent space and investigate using this representation for improving the recommendation of ML datasets and pipelines and improving meta-learning models. Our results demonstrate how complex knowledge graph querying across different sources can facilitate the discovery of ML information, as well as that embedding ML information can improve the recommendation of ML artefacts and meta-learning techniques.

Hyperdimensional Computing (HDC) for Edge Intelligence           De Krook - Laatkomer

Werner Van Leekwijck (imec), Laura Smets (UAntwerpen), Hans Cappelle (imec), Ryan Antonio (KULeuven), Prof Marian Verhelst (KULeuven)

• Hyperdimensional computing (HDC) has been introduced as a brain-inspired, lightweight, and energy-efficient computing paradigm suitable for wearable internet of things (IoT) devices, near-sensor artificial intelligence (AI) applications, and on-device processing. It originated at the intersection of symbolic AI and connectionism, with the potential of combining the advantages of both, i.e., the transparency of symbolic representations, and flexibility and robustness of connectionist representations. 
  This session will explain the basics and current usages of HDC for edge applications, algorithmic advances that have been developed in the FAIR program, performance insights and benchmarks on COTS platforms, and a glimpse at accelerators for this paradigm. The session is organized as follows: 1) HDC basics and state-of-the-art (Werner Van Leekwijck - imec); 2) Algorithmic Advances (Laura Smets - UA); 3) Benchmarking HDC on COTS platforms (Hans Cappelle - imec); 4) HDC acceleration (Ryan Antonio - KULeuven).

Neutrality and bias in (generative) AI           Music Club

Prof. Tijl De Bie (UGent)

• Neutrality and bias in classical AI are elusive concepts that can be formalized in a multitude of ways. In generative AI, the conceptual and methodological challenges are far greater yet. In this talk, I will investigate the notions of `neutrality' and `bias' from the perspective of ideological diversity, content moderation, and censorship, and I will touch upon the implications for regulatory initiatives that aim to mitigate the harmful impact of generative AI on public discourse.

Selected pitches from Flanders AI Research Program PhD students           Bar Bougie Left

• Christopher Patzanovsky (UHasselt), Visualization and Pruning Techniques for Decision Making Pathways in Neural Networks

• Hicham Azmani (VUB), Learning to Interact: Multi-Agent RL for Energy Systems

• Xuemin Duan (KULeuven), SHACL as a Unified Data Profile for Consistency Validation in Knowledge Graphs

• Adrick Tench (UGent), LLM-assisted causal discovery

• Anna Verheyden (KULeuven), Clarity in complexity: a dashboard for navigating algorithmic fairness in HR

• Apurva Shah (VUB), Adaptive Multi-Objective algorithms for Fair Collective Decision-Making in Dynamic Environments

• David Debot (KULeuven), Towards Accurate and Explainable AI using Concept-based Reasoning ModelsVarious PhD students

(Team) Sports Analytics           Collective A

Pieter Robberechts (KULeuven)

• In this deep dive, we intent to give an accessible overview of existing work on the use of machine learning in sports. The field of sports analytics is vast, though, which is why we narrow the scope to team sports, for which much work exists and in which the authors have the most expertise. We will touch on the following topics: (1) the event data and spatial-temporal data that is available for team sports; (2) which particular machine learning techniques have been used in the literature; (3) how to assess the value and quality of actions performed by players; and (4) ways of identifying and describing team tactics, and how to play out alternative paths. The overall goal is to leave participants with a general idea of the state-of-the-art and challenges in sports analytics. We hope this will inspire participants to apply their research to use cases in sports.

Trustworthy AI in Medical Image Analysis           Expo2

Prof. Matthew Blaschko (KULeuven), Siri Willems (imec)

• Medical image analysis is a crucial technology for improving healthcare outcomes.  This session discusses two key points in deploying trustworthy AI in this area: (i) the incorporation of domain specific information in the form of hyperspectral imagery, and (ii) modeling uncertainties to give statistically valid estimates of imaging based biomarkers.  Contributions to these domains from KU Leuven and Imec within the context of the Flanders AI program will be discussed.

Tutorial: An introduction to data analysis and compression via tensor decomposition using pyTensorlab           Bar Bougie Right

Nico Vervliet, Raphaël Widdershoven, Prof. Lieven De Lathauwer (KULeuven)

• Storing and analyzing data both effectively and efficiently is key in various applications in domains such as signal processing, machine learning, scientific computing, bioinformatics, and audio and image processing. If a dataset has more than two dimensions and can be represented as a multiway array of numbers, or a tensor, using advanced tensor techniques leads to better and faster ways to compress data, recover and interpret hidden signals, learn patterns, and so on. In this tutorial, we showcase two key tensor decompositions: the canonical polyadic decomposition (CPD), which writes a tensor as a minimal sum of simple rank-1 terms, and the low-multilinear-rank approximation (LMLRA)---also known as the Tucker decomposition---which relates the different subspaces in each mode through a core tensor. We illustrate how these techniques can be used to extract information and compress data through various examples. To allow participants to focus on the application at hand, the recently released Python version of Tensorlab, pyTensorlab, will be used, and we will demonstrate its core features. Laptop with Python 3.10 or higher installed is required.

Tutorial: Constraint guided deep learning with Congrads           Collective B

Prof. Peter Karsmakers, Quinten Van Baelen, Wout Rombouts (KULeuven)

• Neuro-symbolic AI (NeSy) methods aim to integrate deep learning with knowledge-based systems. One such method is Constraint Guided Gradient Descent (CGGD), which enables the inclusion of domain knowledge in the form of inequality constraints, monotonicity constraints, and partial differential equations. The key innovation of CGGD lies in its ability to handle hard constraints involving continuous variables, which sets it apart from many existing approaches. By incorporating domain knowledge into the learning process, the reliance on large amounts of data for training deep learning models can be reduced.
  The Congrads toolbox provides a practical implementation of CGGD, allowing users to define domain knowledge as constraints that, in conjunction with data, guide the learning process. Additionally, users can define custom constraint types, which are seamlessly integrated into the learning framework.
  In this tutorial, we first introduce the types of problems where Congrads is applicable. Next, through a series of hands-on use cases, participants will learn how to apply Congrads to new applications. This workshop requires a laptop.

Advancements in autonomous robots & drones           Music Club

Yuri Durodie (VUB), Carlos Tiana Gomez (Flanders Make) 

• Anchorless localization for autonomous robots in GPS-denied environments
  As autonomous mobile robots gain traction across logistics, agriculture, inspection, and manufacturing, reliable localization remains a critical challenge—especially in GPS-denied environments like greenhouses, warehouses, and indoor industrial sites. Traditional systems using UWB anchors, QR codes, or cameras are effective but costly, making them impractical for smaller-scale deployments. This experiment explores a new frontier: anchorless localization using only onboard sensors. By evaluating real-world performance in complex environments—such as metal-rich structures and dense crops—we assess both the technological viability and economic impact of this approach. The goal: enable safe, collaborative robot operation without the need for expensive infrastructure. 
  This innovation opens the door to scalable, cost-effective deployment of autonomous platforms in both small and large environments, with fewer robots and lower setup costs. Ideal for drone developers, robotics manufacturers, and integrators looking to expand into constrained or dynamic spaces.

• AI-powered mobile robotic assistants: bridging virtual intelligence with real-world scenario's
  While recent AI breakthroughs have revolutionized virtual tasks like text generation, real-world challenges—such as reducing staff workload in hospitals—require a new kind of intelligence. This session explores a novel training framework for robotic generalist assistants that combines AI-generated strategies with real-world interaction, enabling robots to operate safely and effectively in dynamic environments.
  We’ll focus on hospital use cases, particularly automating logistical tasks in clinical trials—like delivering supplies, guiding patients, and supporting emergency workflows—to free up qualified staff for higher-value care. The same framework applies to hospitality and warehousing, where robots can enhance guest experiences or navigate complex indoor spaces without costly infrastructure.

AI applied in agriculture           De Krook - Laatkomer

Hiep Luong (University of Ghent, imec), Gaoyuan Liu (VUB), Pieter Jan Delvaux (Flanders Make)

• Digital twins of orchards: Realistic 3D reconstruction in outdoor environments
  Accurate 3D modeling of orchards is a game-changer for agricultural automation. In this session, we share insights from AI experiment, where we developed a high-resolution 3D reconstruction using sensor fusion—combining dual RGB cameras with LiDAR scanning—to create realistic digital twins of complex orchard environments. These detailed models are not just visualizations—they actively guide robotic pruning systems. By integrating 3D data into decision-making and motion control, robots can prune with precision, avoid damaging branches, and adapt to the natural variability of fruit trees. The technology shows strong potential for broader orchard tasks and inspection robotics. Whether you're a fruit grower, agri-tech manufacturer, or innovator in inspection systems, this session offers a glimpse into how digital twins are transforming automation in the field.

• Smarter motion planning for Agricultural Robotics: From pruning to inspection
  Precision tasks like pruning, harvesting, and inspection in orchards demand robotic systems that can navigate cluttered, unstructured environments with high reliability. While perception in agricultural robotics has advanced, motion planning remains a bottleneck—especially for complex tasks like pruning fruit trees, where rigid structures and dense foliage pose unique challenges. This session introduces a novel behavior planning framework designed to boost the adaptability and efficiency of robotic manipulators in agriculture. Featuring a dual-modal motion planner that blends goal-oriented planning with reactive servoing, the system leverages both vision and tactile sensing to operate effectively even in occluded areas. Demonstrated on a robotic pruning system for fruit trees, the approach significantly improves planning speed and success rate. Beyond pruning, the technology scales to a wide range of orchard tasks—from frost prevention to harvesting—and is adaptable to other crops like pears and vegetables. The control algorithms also show promise in inspection robotics and warehouse automation, where high-DoF manipulators must operate safely in constrained environments.
  Join us to explore how advanced planning can unlock new levels of autonomy and scalability in agricultural and inspection robotics.

• Explainable (vision) AI for more sustainable 'Crop Management'
  Vision AI is transforming agriculture by enabling precise, real-time weed detection using high-resolution cameras, drones, and sensors. But to truly empower farmers, AI must be interpretable. This session introduces a cutting-edge system that combines Vision AI with Explainable AI and Visual-Language Models (VLMs), allowing farmers to understand and trust AI decisions through image-based explanations and interactive question answering.
  Built on a multi-stage pipeline and supported by a Retrieval-Augmented Generation (RAG) system, the technology integrates visual data with domain-specific knowledge for context-aware insights. The result: targeted spraying, reduced environmental impact, and optimized resource use—adaptable across crops and farming contexts.
  Join us to explore how this innovation supports smarter farming, sustainable practices, and broader applications in anomaly detection and dynamic system control.

Pitch: Hyperspectral cameras in robotic sensing
Ciem Cornelissen (UGent)

• While robots are increasingly equipped with vision, their perception is often limited to what the human eye can see. This restricts their ability to perform complex tasks that require understanding material composition or identifying subtle, invisible properties. My research tackles this challenge by integrating hyperspectral cameras into robotic sensing systems.
  Unlike standard cameras that capture three colors (RGB), hyperspectral imagers analyze hundreds of light bands, allowing the robot to identify the unique "spectral fingerprint" of any material it observes. This technology unlocks powerful applications, from a robot in agriculture determining the precise ripeness of fruit or detecting early-stage plant disease, to an industrial robot sorting visually identical plastics in a recycling facility.
  My pitch will demonstrate how this fusion of robotics and advanced sensing paves the way for a new generation of intelligent automation, creating robots that can make nuanced decisions far beyond current capabilities.

AI driven smart design engineering           De Krook - Twistappel

Prof. Ivo Couckuyt, Prof. Dirk Deschrijver (University of Ghent, imec)

• Accelerating engineering with artificial intelligence: Driving faster design and optimization with limited data
  In industries like engineering, manufacturing, and healthcare, access to large, labeled datasets is often limited - making it challenging to apply traditional Machine Learning (ML) methods. However, the data that is available tends to be of high quality, e.g., coming from time-consuming physics-based simulations or costly lab-controlled experiments. Data scarcity has emerged as a major challenge in recent years, and going from the traditional "big data" to “small data” requires a shift in approach. In this talk, we introduce data-efficient ML (or surrogate modeling), which enables practitioners to build and deploy effective models trained on just 10 to 100 examples. Data-efficient ML can accelerate tasks such as design space exploration, optimization, sensitivity analysis, generative design, and the development of cost-effective digital twins. We explore several real-world examples from industry, from creating trustworthy digital twins, to the design of ventilation systems and optimizing the manufacturing process. Data-efficient approaches unlock many new use cases in which only small (labeled) datasets are available. This minimizes the time, engineering effort, and data required to go live and create practical value from artificial intelligence.

• Accelerating  engineering product design with physics-informed AI
  Designing high-performance engineering devices—whether mechanical, electronic, or photonic—remains a complex, time-intensive process. Lengthy cycles of simulation, measurement, and optimization often slow innovation and inflate costs, especially when the gap between available technology and desired functionality continues to grow. 
  This session explores how physics-inspired models enriched with advanced AI techniques can transform the design process. By combining the predictive power of AI with the reliability of physical principles, engineers can automate and accelerate device development—making it faster, more accurate, and cost-effective. 
  From photonic components in telecom to mechanical systems in aerospace and electronics for renewable energy, this approach empowers design teams to push boundaries, reduce time-to-market, and deliver smarter, high-performance products.

Industry 5.0: unlocking hidden machine performance           Bar Bougie Left

Denis Steckelmacher (VUB), Tom Lefebvre, (UGent), Dmitry Kravchenko (Umicore), Xavier Chesterman (VUB), Leander Schietgat (VUB)

• Better control and more knowledge on complex multi-physics systems with Machine Learning

  The MPRL2Real project aims to bridge the gap between lab-validated PoC and industrial use cases by developing a systematic approach to derive a simulation environment from historical data collected on large multi-physics industrial plants. This simulation environment can subsequently be used to safely validate ML-enhanced controllers to increase the confidence level and maturity level of these controllers, to assess their suitability before real-life implementation. We will zoom into 2 methods for achieving this: (i) Reinforcement learning (RL) for automatically learning a control policy from interacting with a system (real or simulation). RL is known to sometimes produce quite creative policies, that would not have been envisioned by people (ii) Bayesian Optimization (BO) which enhances the performance of the real system by finding the optimal parameters of a low-dimensional parameterized controller during automated commissioning or tuning.

  This session presents a method to leverage ML to improve the control of complex multi-physics systems in a safe way. We will detail how to develop a simulator for such a system, at low cost, such that an RL agent or BO can then interact with it and produce original policies and optimized control actions. We will then explain how to evaluate these policies, assert their usefulness in the real world, and produce real human knowledge by analyzing the policy and its decisions. This human knowledge can then be applied with confidence.

  Whether you're an industrial manufacturer or an AI solution provider, join us to explore how Machine Learning can think outside the box and uncover new control strategies!

• Scaling failure diagnosis across several industrial machines
  Industrial machine data is notoriously inconsistent—varying across configurations, sensors, and signal naming conventions—making cross-machine analysis and failure diagnosis complex and time-consuming. As data volumes grow, manual pattern detection becomes impractical, demanding scalable automation.
  This session introduces a robust, scalable toolbox that tackles these challenges head-on. It automatically builds machine ontologies using expert knowledge and pattern mining, and applies knowledge graph reasoning to enable transparent, cross-type failure diagnosis. Developed through FAIR research and validated with real-world fleet data from geared wind turbines, the toolbox is designed for realistic industrial scenarios with noisy, heterogeneous data.
  Whether you're in manufacturing, energy, automotive, or electronics, this approach offers a powerful way to streamline diagnostics, reduce downtime, and future-proof your operations.

Making industry knowledge accessible with AI           Auditorium

Willy Carlos Tchuitcheu (VUB)

• Towards Next-Generation of Intelligent Document Understanding: Open Questions and Challenges
  This session presents a research agenda for advancing intelligent document understanding in the era of Large Language Models (LLMs). As LLMs increasingly rely on intermediate reasoning traces, such as Chain-of-Thought (CoT), the need for task-specific, compatible frameworks becomes critical in document-centric applications. These applications span summarization—requiring coherent and synthesized narratives—to fact-checking, which demands rigorous hypothesis validation grounded in document content, and visual question answering, which involves multimodal reasoning. While Retrieval-Augmented Generation (RAG) has shown promise in improving context retrieval, it faces practical limitations in scalability and adaptability due to the heterogeneous structures of real-world documents. 
  Beyond the presentation, this session will feature an interactive Q&A to gather audience insights on their information processing workflows and expectations. Key discussion points include: Would a custom AI chatbot improve their document workflows? What challenges do they foresee in adopting LLMs (e.g., cost, data privacy)? How large and dynamic is their internal knowledge base, and how does it scale over time? Is there a need for in-house, lightweight, task-specific LLMs through knowledge distillation? 
  Your feedback will help shape future research directions and practical solutions for next-generation document understanding.

Next-gen AI in industry: Explainability, Privacy & Performance           De Krook - Blauwe Vogel

Joran Michiels (KULeuven), Emilio Gamba (Flander Make)

• Beyond explainability: Guiding AI models through their explanations
  The growing use of machine learning in critical domains has increased the demand for models that are not only accurate but also understandable and adaptable. While post-hoc explanations are widely used to shed light on model predictions, they are still mostly treated as passive tools. This limits their usefulness: explanations may help users interpret a model, but they rarely provide a mechanism to improve it. This presentation explores a more active role for explanations through the concept of interactive local post-hoc explanations. Instead of being the endpoint of analysis, explanations become an interface for guiding models. We present and evaluate two complementary approaches: training set augmentation, where user corrections to explanations are transformed into new labeled data to retrain the model, and loss augmentation, where explanations directly influence the model’s optimization objective. Both strategies allow end-users to iteratively refine models, improving not only performance but also the faithfulness of the explanations themselves. Our experiments on benchmark datasets demonstrate that interactive explanations can deliver a clear win-win: explanations become more meaningful to users, while models achieve higher accuracy and efficiency. Beyond benchmarks, these results highlight strong potential for practical use in industry settings, where interactive explanations could align models more closely with expert knowledge.

• Privacy-preserving Vision AI: Balancing safety, accuracy, and anonymity in industrial environments
  How can industries benefit from camera-based AI—improving safety, ergonomics, and productivity—while protecting worker privacy and company-sensitive information? Many manufacturers see the potential of vision AI, yet its adoption is often slowed by concerns from employees, unions, and management about surveillance and data misuse. Conventional anonymization techniques like blurring or pixelation rarely offer the right balance: either too much useful information is lost, or sensitive details remain visible. 
  In PRIVAI, Flanders Make and IDLab Ghent worked with three industrial partners (Ardis, ExoVibe and PixelVision) to validate and improve privacy-aware monitoring in real manufacturing environments. The project investigated how to filter human features (such as skin color or tattoos) and IP-sensitive details data essential for AI applications like operator activity tracking, forklift-pedestrian detection, and ergonomic risk assessment.
  In this presentation, we (1) address the technical challenges of implementing privacy-preserving video processing in industrial environments, (2) examine barriers to adoption based on a user-acceptance study with workers, unions, and management, and (3) provide insights and guidelines for responsible deployment, together with directions for future research into scalable, trustworthy privacy-aware vision AI.

The AI Act and conformity assessment           Expo2

Prof. Koen Vranckaert (KULeuven)

• In this session, we will discuss the Artificial Intelligence Act's requirements for high-risk AI systems and focus in particular on the relationship between the Artificial Intelligence Act's legal requirements and their expression in harmonized standards, as well as the conformity assessment procedures that make these requirements "come to life". Participants are expected to leave this session with an increased understanding of the AI Act's impact in their sector, not only in theory, but also in practice as harmonized standards and conformity assessment bodies will be designated in the future. Thus, participants are expected to leave a bit more "ready" for when the AI Act becomes operational. 

The Flanders AI Research Program           Bar Bougie Right

Sabine Demey (FAIR)

• This session will explore the FAIR program’s progress in addressing its grant challenges and its potential future impact. We’ll highlight key enrichment projects and use cases. Depending on the audience, we may dive deeper into specific deliverables or address targeted questions.

Towards Zero-Manual Reconfiguration           Collective A

Bart Vandoninck (Flanders Make)

• Traditional production systems are often engineered for the assembly of a specific product. In high-mix, low-volume manufacturing, this leads to significant reprogramming effort whenever configurations change. Autonomous systems—such as mobile robots, AGVs, autonomous forklifts, drones, robotic arms, and inspection systems—are increasingly deployed on production floors to improve efficiency and adapt to variability in dynamic manufacturing environments. In this deep-dive session we demonstrate how AI upgrades Industry 4.0 agents with new intelligence can achieve zero-manual reconfiguration. Intelligent perception enables robust understanding of the environment through multi-modal sensor fusion. Multi-agent collaboration allows autonomous systems to organize themselves for complex assembly and logistics tasks with minimal human intervention. Integration with digital twins ensures real-time tracking and mapping of assets, while also generating hybrid datasets from both physical and simulated setups for training, validation, and testing. Enhancing these systems with AI capabilities ultimately enables zero-manual reconfiguration, where production resources adapt automatically to variable demands, advancing toward true manufacturing autonomy.

Demo tours           Arena

Our challenge managers will take you on a guided tour towards a selection of our use cases / demo's. 

• Demo tour 1: Selected Use Cases from the different FAIR Program domains (Health, Industry, Planet & Energy, Society)

• Demo tour 2: Industry Use Cases from FAIR Program 

Please note that all the demo's are freely accessible throughout the day.