Fakultät Informatik


Kolloq. Prof. Bronstein, topic: Deep learning on graphs and its applications to matrix completion

Friday, 23rd of June 2017, 2:00 pm FMI HS 3 (MI-Building, Campus Garching)

The rise of graph-structured data such as social networks, regulatory networks, citation graphs, and functional brain networks, in combination with resounding success of deep learning in various applications, has brought the interest in generalizing deep learning models to non-Euclidean domains.
In the first part of this talk, I will introduce a new spectral domain convolutional architecture (CNN) for deep learning on graphs. The core ingredient of our model is a new class of parametric rational complex functions (Cayley polynomials) allowing to efficiently compute localized regular filters on graphs that specialize on frequency bands of interest. Our model requires a constant number of parameters and has complexity that scales linearly with the size of the input data for sparsely-connected graphs.
In the second part of the talk, I will extend graph convolutional neural networks to multiple graphs and show their application to matrix completion problems, which are among the most common formulations of recommender systems. Assuming to be given pairwise relationships between users/items, we formulate the matrix completion problem as deep learning on the user/item graphs. Our matrix completion architecture combines graph convolutional neural networks and recurrent neural networks acting together as a learnable non-linear diffusion process on the score matrix. This neural network requires a constant number of parameters and has linear complexity in the matrix size. I will conclude with state-of-the-art results on some standard benchmarks showing the promise of graph learning approaches.
Based on joint works with Federico Monti (USI), Ron Levie (TAU) and Xavier Bresson (NTU).

Michael Bronstein is an associate professor of Informatics at USI Lugano in Switzerland, associate professor of Applied Mathematics at Tel Aviv University in Israel, and a Principal Engineer at the Intel Perceptual Computing. Michael got his Ph.D. with distinction in Computer Science from the Technion in 2007. He has previously held visiting appointments at Politecnico di Milano, Stanford, INRIA, Technion, and University of Verona. He is a Senior Member of the IEEE, alumnus of the Technion Excellence Program and the Academy of Achievement, ACM Distinguished Speaker, and a member of the Young Academy of Europe. His research appeared in the international media such as CNN and was recognized by numerous prestigious awards, including several best paper awards, three ERC grants (Starting Grant 2012, Proof of Concept Grant 2016, and Consolidator Grant 2016), Google Faculty Research Award (2016), Radcliffe Fellowship from the Institute for Advanced Study at Harvard University (2017), and Rudolf Diesel Industrial Fellowship from TU Munich (2017). In 2014, he was invited as a Young Scientist to the World Economic Forum, an honor bestowed on forty world's leading scientists under the age of forty. He was a guest speaker at the World Economic Forum meeting in Dalian, China in 2015. Michael is the author of the first book on deformable 3D shape analysis, editor of four books, over 100 papers in top scientific journals and conferences, and inventor of over 25 granted patents. He has chaired over a dozen of conferences and workshops in his field, and has served as area chair at ECCV 2016 and ICCV 2017 and as associate editor of the Computer Vision and Image Understanding journal. Besides academic work, Michael is actively involved in the industry. He has co-founded and served in leading technical and management positions at several startup companies, including Invision, an Israeli startup developing 3D sensing technology acquired by Intel in 2012.

contact person:
Daniel Cremers
Phone: +
Email: cremers(at)tum.de



« December - 2017 »