"There is no royal road to geometry.'' --- Euclid
The Geometric Learning And Discovery (GLAD) group is mainly interested in the following directions:
1. Geometric Deep Learning (GDL): Model images, videos, spatio-temporal data, social networks, and scientific data as geometric objects with certain symmetry, such as sets, manifolds, graphs, etc, and then developing specific machine learning models to represent and process these data. Particularly for graph data processing via GNNs, we research the expressivity of GNNs, how to train GNNs efficiently, how to perform self-supervised GNN learning.
2. GDL for Robot Sensing and Decision-Making: Towards tactile data representation, video analysis, sound signal processing, multi-modal data sensing and fusion; model-based reinforcement learning, imitation learning with imperfect supervision.
3. Science-Informed GDL: Embed the knowledge from Geometry, Quantum Mechanics, Computational Chemistry, Computational Biology, and other science domains into existing GML models, with the application to multi-body problem, molecular dynamics simulation, drug design, protein docking, antibody design, crystal material design, and so on.