Lunes, 4 de abril de 2022, 12:00-13:00 

Enlace:   https://us06web.zoom.us/j/89903156301?pwd=eHRRS0hwVEFXWmRZaDh0WHowMHM0QT09

Ponente:  Jim Portegies (Eindhoven University of Technology)

Título: PDE-based Group Equivariant Convolutional Neural Networks

Abstract: In recent work, we have introduced PDE-based Group Equivariant Convolutional Neural Networks (PDE-G-CNNs). They fit into image-analysis pipelines, where images get represented as (or rather lifted to) functions on homogeneous spaces. I'll explain what PDE-G-CNNs are, how we make them computationally feasible, and what some of their main advantages are. The key message is that by restricting the architecture of the neural networks, we enforce particular input-output relationships (equivariance) so that we can achieve both higher data-efficiency and similar performance with fewer trainable parameters.
This is joint work with Bart Smets, Erik Bekkers and Remco Duits.