Talks

Machine learning is changing the way we simulate matter. In particular, Graph Neural Networks (GNNs), which describe a molecule as a network of interacting atoms, can now predict the energy of molecular systems with high accuracy, enabling fast and reliable simulations. This accuracy, however, comes at the price of transparency: unlike classical force fields, in which each term has a clear physical meaning, a trained GNN is a complex, high-dimensional function whose predictions are hard to trace back to the underlying physical interactions. Existing interpretability tools typically explain the model one configuration at a time, with a cost that grows quickly with the size of the system and with the number of configurations to be analyzed. In this contribution we present a different strategy: rather than analyzing the network on a given unique input, we use a kernel method (in essence, a linear regression performed after a nonlinear embedding of the data) to distill what the GNN has learned into a compact analytical potential. The resulting expression provides a transparent global approximation of the original model, in which the relevant interactions can be read directly, much as in a classical force field. Although illustrated here for molecular potentials, the approach is not tied to this specific application and can be extended to other neural architectures and learning tasks. In this way, a typically opaque object such as a GNN becomes physically inspectable, opening a very effective window into the so-called black box.

Graph Neural Networks accurately approximate the potential energy surfaces of molecular systems, but the learned potentials are hard to trace back to the underlying physical interactions. This talk presents an approach that builds interpretable parameters directly into the architecture, so that only the relevant interactions survive training, revealing which chemical species interact, at which spatial scale and with which many-body contributions, and finally distilling the network into a transparent analytical potential.

This talk explores the emerging field of topological signals, dynamical variables defined on higher-order networks such as simplicial or cell complexes. These are functions defined on simplicial complexes that generalize the idea of node-based dynamics to higher-order interactions, allowing signals to propagate across links, faces, and higher-dimensional simplices.

A collection of invited talks and seminars presented at various institutes and workshops, focusing on advancements in neural networks and machine learning.

A collection of invited talks and seminars presented at various institutes and workshops, focusing on advancements in neural networks and machine learning.

A collection of invited talks and seminars presented at various institutes and workshops, focusing on advancements in neural networks and machine learning.