The increasing availability of data together with the high computational power of modern architectures lead the success of Deep Learning techniques in recent years. Convolutional Neural Networks and Recurrent Neural Networks represent in this sense a de facto standard today. Unfortunately, despite the great success CNNs and RNNs achieved on images, audio signals, videos and sequences of data (e.g. text), none of the aforementioned techniques is actually able to deal with non-Euclidean structured data (e.g. graphs and manifolds). Motivated by the lack of techniques capable to process such structures and by the increasing availability of network (e.g. social networks, regulatory networks, citation networks) and 3D data (e.g. shapes, point clouds), Geometric Deep Learning approaches appeared on the scene in the last years. Geometric Deep Learning is an umbrella term that refers to all the techniques which attempt to generalize classic Deep Learning approaches to non-Euclidean structured data. In this talk we’ll introduce some of the major GDL architectures that have been introduced for learning on graphs, together with some possible applications of these. We’ll present in particular novel solutions for solving Matrix Completion for Recommender systems, Neutrino detection via IceCube signals and Fake News detection on Social Media.
Copyright. All Rights Reserved.