Learning Discrete Structures for Graph Neural Networks Luca - - PowerPoint PPT Presentation
Learning Discrete Structures for Graph Neural Networks Luca Franceschi , Mathias Niepert, Massimilano Potil, Xiao He Poster later: Pacific Ballroom # 177 Introduction & Motivations Aim: apply Graph Neural Networks (GNN) to settings in which
Poster later: Pacific Ballroom # 177
1
Aτ~Pθ
wt+1= Φ(wt,A1) = wt - γ∇Lt(wt,A1) wt+τ= wt+τ-1 - γ∇Lt+τ-1(wt+τ-1,Aτ)
Data points Initialize parameters Sample graphs Compute hypergradients and update θ of graph generator GCN: Graph generator:
A1~Pθ
Compute gradients of and update GCN parameters
wt+τ wt+τ-1 wt ...
Validation
∇θ 𝔽[F(wθ,τ , θ)]
nodes See Franceschi et al. Forward and Reverse Gradient-based Hyperparameter Optimization, ICML 2017
2
3
3
3
3
3
3
3
3
4