On 23 October 2026 at 16:00, Brent De Weerdt will defend their PhD entitled āInterpretable Transformer Networks for Compressed Sensing, Representation Learning, and Distributed Source Codingā.
Everybody is invited to attend the presentation in room LIC.0.04 or online via thisĀ link.
The current trend in artificial intelligence, and deep learning in particular, is to design increasingly large and complex models. Although these models achieve high accuracy, they require large datasets, high computational budgets, and often remain difficult to interpret. This thesis addresses these limitations by leveraging theoretical knowledge and data priors to design lightweight, performant, and inherently interpretable Transformer-based deep learning models.
The work first considers signal acquisition through compressed sensing. Using deep unfolding, the thesis designs neural networks for signal reconstruction from compressed measurements. By incorporating sparsity and temporal correlation, it develops interpretable Transformer models that reconstruct signals more accurately and efficiently than existing methods. The thesis then studies compact and privacy-preserving representations of human pose trajectories in video using contrastive learning and generative modelling, enabling anomaly detection through deviations from learned normal motion patterns. Finally, the thesis addresses distributed source coding, where correlated observations from multiple devices can be compressed more efficiently by exploiting inter-source dependencies. Building on LDPCA codes, it proposes a multi-rate Transformer decoder for Slepian-Wolf decoding that outperforms traditional decoders and reduces communication cost compared with alternative distributed compression methods.
Overall, the thesis develops deep learning architectures that incorporate signal processing, information theory, and data-specific prior knowledge, achieving state-of-the-art performance across several tasks while reducing the computational budget by at least an order of magnitude in most cases.
