Georgios Leontidis

Embedding equivariance in hyperbolic networks can drastically accelerate convergence and reduce parameter redundancy, reshaping how we approach visual representation learning.

Confidence-based filtering can cut prediction error by over 65%, revealing a new approach to managing uncertainty in emissions monitoring for gas turbines.

Forget complex, multi-stage pruning pipelines: HiAP slashes Vision Transformer size with a single, end-to-end training pass that optimizes sparsity at multiple granularities.