In recent years, artificial neural networks a.k.a. deep learning have significantly improved the fields of computer vision, speech recognition, and natural language processing. The success relies on the availability of large-scale datasets, the developments of affordable high computational power, and basic deep learning operations that are sound and fast as they assume that data lie on Euclidean grids. However, not all data live on regular lattices. 3D shapes in computer graphics represent Riemannian manifolds. In neuroscience, brain activity (fMRI) is encoded on the structural connectivity network (sMRI). In genomics, the human body functionality is expressed through DNA, RNA, and proteins that form the gene regulatory network (GRN). In social sciences, people interact through networks. Eventually, data in communication networks are structured by graphs like the Internet or road traffic networks.
Deep learning that has originally been developed for computer vision cannot be directly applied to these highly irregular domains, and new classes of deep learning techniques must be designed. This is highly challenging as most standard data analysis tools cannot be used on heterogonous data domains. The workshop will bring together experts in mathematics (statistics, harmonic analysis, optimization, graph theory, sparsity, topology), machine learning (deep learning, supervised & unsupervised learning, metric learning) and specific applicative domains (neuroscience, genetics, social science, computer vision) to establish the current state of these emerging techniques and discuss the next directions.
This workshop will include a poster session; a request for posters will be sent to registered participants in advance of the workshop.
(Nanyang Technological University, Singapore)
Michael Bronstein (USI Lugano, Switzerland)
Joan Bruna (New York University)
Yann LeCun (New York University)
Stanley Osher (University of California, Los Angeles (UCLA))
Arthur Szlam (Facebook)