Curriculum Vitae

PhD Researcher

IDLab, Ghent University - imec

Research on self-supervised representation learning and label-efficient latent world models for autonomous perception and control, spanning multi-modal sensor fusion, action-label-free pretraining, and raw-action planning.

• Developed LaFreeWM, a label-efficient JEPA world model that infers discrete latent actions during action-label-free pretraining and uses 1% or 10% labelled trajectories for raw-action planning
• Identified a prediction-planning mismatch: removing quantization reduced held-out prediction MSE by about 60-fold but lowered long-horizon planning success from 60.7% to 18.0%
• Developed Le MuMo JEPA, a multi-modal self-supervised framework for unified RGB-LiDAR-thermal representations, accepted at the CVPR 2026 URVIS Workshop
• Designed learnable fusion tokens as an efficient latent bottleneck and evaluated frozen representations on Waymo, nuScenes, and FLIR for depth, segmentation, and 3D detection
• Developed LISA for illumination-invariant hyperspectral learning and OHSLIC for efficient UAV-based phenotype segmentation
• Research includes the LaFreeWM preprint and publications in CVPR Workshops, IEEE Internet of Things Journal, WACV Workshops, and Neural Computing and Applications

PhD

Ghent University - imec

Computer Science Engineering

• Multi-Modal Self-Supervised Learning
• Predictive Representation Learning
• Latent World Models
• Autonomous Perception

Advanced Master

KU Leuven

Artificial Intelligence

• Deep Learning
• Computer Vision
• Data Science
• Machine Learning

Master of Science

Ghent University

Physics

• Complex Systems
• Statistical Mechanics
• Artificial Intelligence Electives

Bachelor of Science

Ghent University

Physics

• Theoretical Physics
• Mathematics
• Computational Methods

LaFree World Model: Label-Free Pretraining for Action-Conditioned JEPA World Models

Preprint

A label-efficient JEPA world model that learns discrete latent actions from unlabelled observation sequences, then grounds its predictor with limited labelled data for raw-action planning. The study also exposes a prediction-planning mismatch in which lower prediction error does not necessarily produce better control.

Le MuMo JEPA: Multi-Modal Self-Supervised Representation Learning with Learnable Fusion Tokens

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops - URVIS

A self-supervised framework that learns unified representations from RGB images and aligned companion modalities. Extends LeJEPA to the multi-modal setting by learning fusion tokens that act as a latent bottleneck between modality-specific patch stems inside a shared transformer.

In-Field Mapping of Grape Yield and Quality with Illumination-Invariant Deep Learning

IEEE Internet of Things Journal

Development of the LISA framework for illumination-invariant hyperspectral imaging analysis in precision viticulture. An end-to-end IoT-enabled robotic system for non-destructive, real-time mapping of grape yield and quality.

Computational Fairness in Adaptive Neural Networks

Neural Computing and Applications

A novel investigation into the fairness of adaptive neural networks, analyzing disparities in computational resource allocation across demographic subgroups. Introduces computational resource allocation as a new dimension of AI fairness.

Adaptive Clustering for Efficient Phenotype Segmentation of UAV Hyperspectral Data

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops

Introduction of the OHSLIC algorithm for computationally efficient, real-time phenotype segmentation from hyperspectral imaging data. Enables on-device processing on UAVs.