Unlocking Deep Learning for Graphs скачать в хорошем качестве

Unlocking Deep Learning for Graphs

Dominique Beaini, PhD, Research Unit Lead at Valence Discovery PNA -- Principal Neighbourhood Aggregation for Graph Nets (NeurIPS2020) https://arxiv.org/abs/2004.05718 DGN -- Directional Graph Networks (ICML2021) https://arxiv.org/abs/2010.02863 SAN -- Rethinking Graph Transformers with Spectral Attention (NeurIPS2021) https://arxiv.org/abs/2106.03893 Reading group organized by students from Mila (Quebec's AI institute) Abstract: Deep learning has successfully transformed how we tackle different problems, especially in the fields of image/text recognition and generation. However, most methods remained constrained to these grid-like data structures and struggle to translate to more complex geometries such as meshes and graphs. Having powerful deep learning on graphs will unlock an unprecedented number of applications in different fields, such as social networks, road navigation, and drug discovery. In this talk, I will discuss the main challenges of applying deep learning on graphs, why early methods have struggled, and how spectral theory is one of the main keys in unlocking graph deep learning. Specifically, I will discuss some of our recent work including “Directional Graph Networks”, which offers the first generalization of convolutional neural networks to graphs, and “Rethinking Graph Transformers with Spectral Attention”, which offers the first generalization of Transformers to graphs. Twitter: @dom_beaini, @valence_ai Linkedin: dbeaini, valence Valence website: www.valencediscovery.com Personal website: https://mila.quebec/en/person/dominiq... ----------------------------------------------------------------------------------- 00:00 Start 00:45 Overview and problematic 01:48 Graph Neural Networks overview 06:28 PNA -- Principal Neighbourhood Aggregation 13:58 DGN -- Directional Graph Networks 30:12 SAN -- Spectral Attention Network 53:24 Overview and Conclusion 55:31 Questions & Answers