geometric deep learning
play

Geometric Deep Learning going beyond Euclidean data Michael - PowerPoint PPT Presentation

Geometric Deep Learning going beyond Euclidean data Michael Bronstein Imperial College London / Twitter Perceptron ( " ! " ( # ! # + * = sign / 0 1 ( $ ! $ & = ( $)" 1 simplest neural network Rosenblatt 1957


  1. Anisotropic filters on meshes Anisotropic spectral filters on meshes Boscaini et B 2016

  2. Spectral graph convolution, take 2 ! & ' % * # ! " # = % ⋱ ! & ) • Matrix function applied to " • Interpret eigenvalues & ' , … , & ) as frequencies and ! & as spectral transfer function • Make ! & parametric with - 1 parameters • Make ! & expressible in terms of simple matrix operations, to avoid explicit computation of % • Possible to guarantee stability under graph perturbations • Possible to guarantee localization Levie et B 2019; Levie, Monti et B 2018

  3. Polynomial filter (ChebNet) ! " = $ % + $ ' " ' + ⋯ + $ ) " ) • Number of learnable parameters * 1 Defferard et al. 2016

  4. Polynomial filter (ChebNet) ! " = $ % & + $ ( " ( + ⋯ + $ * " * • Number of learnable parameters + 1 • Efficient computation + ℰ ~+ / avoiding graph FT altogether 1 2 3 6 4 7 5 8 9 11 10 + ℰ non-zeros Defferard et al. 2016

  5. Polynomial filter (ChebNet) ! " = $ % & + $ ( " ( + ⋯ + $ * " * • Number of learnable parameters + 1 • Efficient computation + ℰ ~+ / avoiding graph FT altogether • Localization to p -hops since " * is localized to p -hops 1 2 3 6 4 7 5 8 9 11 10 Defferard et al. 2016

  6. Polynomial filter (ChebNet) ! " = $ % & + $ ( " ( + ⋯ + $ * " * • Number of learnable parameters + 1 • Efficient computation + ℰ ~+ / avoiding graph FT altogether • Localization to p -hops since " * is localized to p -hops • Generalization across graphs (stability under graph perturbation) • Can be used with other operators, e.g. 0 • Can be used with directed graphs Defferard et al. 2016

  7. Spatial graph convolution

  8. 1D grid = ring graph 1 2 3 … $ adjacency matrix = Shift operator

  9. Convolution, revisited + = + % * ' * % ) ' ) % & ' & !(#) , ' = % & - + % ) ' ) + % * ' * Sandryhaila, Moura 2013

  10. Graph Convolutional Networks (GCN) 1 2 3 6 4 Node-wise 7 5 transform. 8 9 11 10 Node-wise features Kipf, Welling 2016

  11. Graph Convolutional Networks (GCN) 1 2 3 6 4 Node-wise 7 5 transform. 8 9 11 10 Node-wise Graph diffusion features + = ReLU 123 Kipf, Welling 2016

  12. Graph Convolutional Networks (GCN) 1 2 3 6 4 Node-wise 7 5 transform. 8 9 11 10 Node-wise Graph diffusion features + = softmax 4ReLU 49: ; : < Kipf, Welling 2016

  13. Graph Attention Networks (GAT) * + = , 0 +- 1 - -∈/ + " ! attention s core exp ξ 1 + (, 1 - ( ) 0 +- = ∑ 7∈/ + exp ξ 1 + (, 1 7 ( ) # = % &, (, ) & Monti et B 2017; Veličković et al. 2018

  14. Message Passing Neural Network (MPNN) General aggregation function * + = , - 1 2 + , 2 . , 3 +. , ) " .∈0 + ! # = % &, (, ) Gilmer et al. 2017 (MPNN); Wang et B 2018 (EdgeConv)

  15. Images Graphs § Convolution § Message passing § Local operations (window) § Local operations (1-hop) § Constant number of neighbours § Different number of neighbours § Fixed ordering of neighbours § No ordering of neighbours § Shift equivariance § Permutation invariance § O( n ) complexity § O( n ) complexity

  16. Pooling

  17. Graph coarsening

  18. Graph coarsening ! (#) = ! ! (&) • Sequence of coarser graphs with adjacency matrices ! (#) , ! (&) , …

  19. Graph coarsening ! (#) = ! ! (*) ! (&) • Sequence of coarser graphs with adjacency matrices ! (#) , ! (&) , … • Pooling of features on collapsed vertices ( (#) , ( (&) , … • Interleave convolutional / pooling layers • Learnable pooling Ying et al. 2018 (DiffPool)

  20. “…we might be witnessing a new field being born.”

  21. ICLR 2020 submissions keyword statistics ∆ between 2020 and 2019 in % Plot: Pau Rodríguez López

  22. “We expect the following years to bring exciting new methods and applications”

  23. Vosoghi et el. 2018

  24. Did 'Muslim Migrants' Attack a Catholic Church? A video of pro-migrant protesters being removed from the Basilica of Saint-Denis in France was shared with the inflammatory and incorrect claim that it shows Muslim immigrants attacking a church. Monti et B 2019

  25. Acquired by Twitter in 2019

  26. Recommender systems and link prediction decoder encoder Graph Graph Node embedding Reconstructed Input graph graph

  27. High-energy physics Jet image: LCH

  28. LHC: stop pair production GNN architecture for event graph classification Abdughani et al. 2018

  29. LHC: Particle reconstruction Ju et al. 2019

  30. IceCube: neutrino detection

  31. IceCube: neutrino detection Light deposition for a high-energy ROC curve comparing different methods for single muon in IceCube detector neutrino detection Choma et B 2018

  32. LHAASO CR experiment Graph-structured LHAASO-KM2A detectors ROC of classifying light component from the activated by a 500-TeV EAS event background (red=EM & blue=muon detectors) Jin, Chen, He 2019

  33. Astrophysics: Redshift regression Predicted galaxy redshift from photometric observations using MoNet-style GNN vs groundtruth spectroscopic measurement Beck et al 2019

  34. Neutrino detection Light deposition for a high-energy ROC curve comparing different methods for single muon in IceCube detector neutrino detection Choma et B 2018

  35. Computational chemistry and drug design 10 60 Synthesizable molecules Computational cost 10 12 10 -2 Graph NN #Candidates 10 3 10 9 DFT 10 5 10 6 Schrödinger 10 3 Experiment “Computational funnel” Stokes et al. 2020 Duvenaud et al. 2015; Gilmer et al. 2017; Jin et al. 2020

  36. Hyperfoods Veselkov et B 2019

  37. Hyperfoods Veselkov et B 2019

  38. Hyperfoods Veselkov et B 2019

Recommend


More recommend