people
Brussels, Belgium
Ward Gauderis
Hello! I’m an FWO PhD Fellow studying compositionality as a mathematical foundation for deep learning. I use tensor networks and category theory to analyse neural representations as induced by the computations encoded in a model’s weight structure.
Instead of reading tea leaves in activation space, I treat the model’s weights as a formal compositional system, where local mechanisms compose into global properties. Emergent behaviour then becomes a direct function of the algebraic wiring, and the black box divides into parts small enough to conquer.
My work proposes tensor models to bridge neuro-symbolic AI and mechanistic interpretability, asking how models compose concepts to generalise… or not.
Antwerp, Belgium
Thomas Dooms
Hey! I’m a PhD student exploring weight-based interpretability. I’m interested in understanding how neural networks encode complex behaviours and knowledge.
Modern deep networks are astonishingly intricate and difficult to disentangle. I argue that we should embrace this complexity rather than shoehorn them into simplistic feature sets, flattening the very structures we hope to understand.
My work treats a model’s weights as a compositional system whose parts interact and recombine, much like phrases in a language. By developing flexible, formally grounded tools for compositional analysis, I aim to reveal how rich mechanisms and representations arise and how we can reason about them.