Thanks for the #Flanders AI Research Program (FAIR) for supporting this collaboration and to the involved persons for the fruitful collaboration.

#Neuromorphic #hyperdimensional #interpretability #Explainability #VSA #xai #AI #ML #FAIR #UAntwerp #imec

#HDC models aim to be an energy-efficient alternative to current #AI systems and thanks to the efforts of our collaborators, their decision-making process is now more interpretable.

#Neuromorphic #hyperdimensional #interpretability #Explainability #VSA #xai #AI #ML #FAIR #UAntwerp #imec

Thanks to our collaborators from the #VUB, Ward Gauderis and Geraint Wiggins; as well as #sqIRL members Thomas Dooms and José Oramas for the nice collaboration. #UAntwerp #FlandersAI #FAIR

Benjamin Vandersmissen (25/04 evening) will show the effects that using a twin network has on learning processes and share insights on how TNA leads to superior predictive performance in a number of tasks for several architectures. #deeplearning #ML #ICLR2025 #sqIRL
openreview.net/forum?id=TEm...

Thomas Dooms will show how to bilinear MLPs can server as more transparent component that provides a better lens to study the relationships between inputs, outputs and the weights that define the models. #mechinterp #interpretability #ML #AI #XAI #ICLR2025 #sqIRL
openreview.net/forum?id=gI0...

Bilinear MLPs Enable Weight-based Mechanistic Interpretability
M. Pearce, T. Dooms, A. Rigg, J. Oramas, L. Sharkey

We show that bilinear layers can serve as an interpretable replacement for current activation functions; enabling weight-based interpretability.

preprint: arxiv.org/abs/2410.08417

Improving Neural Network Accuracy by Concurrently Training with a Twin Network
B. Vandersmissen, L. Deckers, J. Oramas

We show the effectiveness of TNA lies on a better exploration of the parameter space and the learning of more robust and diverse features

preprint: openreview.net/forum?id=TEm...

Neural networks are not black boxes. They're the opposite of black boxes: we have extensive access to their internals.

I think people have accepted this framing so innately that they've forgotten it's not true and it even warps how they do experiments.