Fast Convolutional Neural Networks for Graph-Structured Data

Presented by: Xavier Bresson (Swiss Federal Institute of Technology) – Fast Convolutional Neural Networks for Graph-Structured Data

Why CNNs work?

Local stationary points.

CNN for Graph Structured Data

• Graph -> Euclidean Grid
• Graph coarse -> Downsampling(pooling)

Related work

Categories of graph CNNs

1. Spatial approach
2. Spectral(Fourier) approach

Convolution on Graph

• Graph Laplace
• Fourier transform on graph
• Localized Filters
  Fast Chebyshev Polynomial Kernels

Graph Coarsening

• Graph partitioning: Balance Cut/Graclus
• Fast Graph Pooling

Optimization

• Backpropagation
• Gradient Descent

Numeric Computation

• Tensorflow
• CUDA k40 (GPU x8 faster than CPU)

Result

• Euclidean CNNs
• Non-Euclidean CNNs

Future

• Social networks
• Gene networks
• etc.