When you look around you, you see the world as objects interacting with one another. When scientists study nature, they define galaxies, stars, lakes, droplets of water, molecules, and sub-atomic particles in complex configurations. Structured models--based on object- and relation-centric representations and computations--are crucial to human intelligence, and this talk will describe how structured models are being developed and used in artificial intelligence (AI). A key method we'll explore is the "graph neural network" (GNN), a class of deep learning architectures which can be trained to support various prediction, inference, and decision-making functions. I'll present several lines of work which use GNNs and their relatives to learn to simulate physics, model complex objects, and plan compositional behaviors. The goal of the talk is to provide a survey of structured models in AI, and to help people who would like to explore these methods get started.
The seminar will be done remote only, using ZOOM for this event.
M. Girone, M. Elsing, L. Moneta, M. Pierini
Event co-organised with the IML workshop