News

Master thesis topics 2025-2026 available

PhD degree awarded

Samuel Wauthier successfully defended his dissertation on Model Reduction in Deep Active Inference.

In his dissertation, Samuel formulates an answer to the continuous growth in AI model sizes and computational demands in achieving Artificial General Intelligence. Specifically, he investigates the challenges of model complexity and proposes solutions for model reduction within the framework of deep active inference, an approach that integrates neuroscience and AI to create intelligent agents capable of adaptive behavior through learning from their environments.

Active inference refers to the process by which living organisms minimize surprisal (or variational free energy) to maintain homeostasis and effectively interact with their surroundings. In the field of AI, active inference has been extended to deep active inference, where neural networks are employed to model the beliefs that agents form about their environment. Samuel Wauthier addresses the issue of scaling deep active inference models and explores methods for reducing their complexity while minimally compromising performance.

Deep active inference relies on neural networks to learn generative models that predict the future states and observations of an agent based on past actions and observations. These models grow increasingly complex as they process large amounts of data and adjust their internal state spaces to reflect the environment. However, as the dimensionality of these state spaces grows, so do the computational and memory requirements, making it infeasible for use in real-time applications such as robotics, where resources are often limited. The primary focus of the PhD is to develop techniques aimed at reducing the complexity of these models. This involves identifying the smallest possible dimensionality of the model that still allows it to accurately predict states and observations within its environment, thereby reducing the computational burden while maintaining the effectiveness of the AI agent.

Nature Communications publication

FAIR Research Day

The Flanders AI Research Program supports cutting-edge AI research. Every year, all researchers involved in this program gather with industrial stakeholders at the FAIR Research Day event. In the 2024 edition, DECIDE had three active contributions.

Plenary pitch and demo: Privacy-aware camera-based ergonomy tracking

Repetitive tasks or improper posture can lead to long-term injuries of operators. While real-time posture tracking is possible with wearables, these can be uncomfortable to wear during long shifts and may disrupt workflow. In contrast, camera-based tracking eliminates the need for wearable. However, it often raises privacy concerns with workers. Within FAIR, we develop novel obfuscation algorithms that preprocess video data directly on the camera. These algorithms remove all privacy-sensitive information before the data is transmitted to a posture detection system.

Demo: Navigation in ambiguous environments

We proposed a novel computational model for navigation and mapping, integrating traditional cognitive mapping approaches with the cognitive principles of Active Inference. At the Research Day, visitors could see a demo of a robot navigating with this model in a simulated warehouse. Read more about this research.

Poster: Adaptive neural networks for real-time hyperspectral imaging

Hyperspectral cameras, capapble of capturing hundreds of spectral bands, allow robots to detect subtle changes in plant and fruit properties. However, they generate massive amounts of data, with hundreds of wavelengths measured for each pixel as opposed to three RGB values with traditional cameras.

Within FAIR, we conduct research on adaptive computation principles allowing real-time grape ripeness detection from hyperspectral data on board a mobile robot.

BELSPO FLOWS project kick-off

The FLOWS project aims to enhance flood crisis management by determining how and when earth observation data and derived products can best support it at three time steps: flood crisis, aftermath and reconstruction. DECIDE is responsible for developing real-time algorithms for flood mapping during data acquisition by UAVs.

On September 16th, all project partners gathered at ISSeP for the first consortium meeting.

Sony Depthsensing Solutions workshop

DECIDE professors Pieter Simoens and Sam Leroux are responsible for the course on Machine Learning in the Master of Science in Information Engineering Technology. This academic year, they teamed up with Sony Depthsensing Solutions to create an industry-relevant project. Guided by domain experts from Sony Depthsensing Solutions, students developed their own algorithm for hand gesture recognition from depth cameras.

It was a rewarding experience for all partners!