[AIRG] AIRG 12/10: Graph Convolutions

[DAVID MERRELL <dmerrell@xxxxxxxxxxx>]

Hi AIRG folk,

At this week's AIRG meeting, we will discuss generalizations of Convolutional Neural Networks.

In a nutshell: 

• CNNs work great for images because they exploit the mesh-like relationships between pixels. CNNs impose a useful inductive bias in that setting.
• However, there are many kinds of data where the variables are related, but those relationships are messy -- they are described by some arbitrary graph, rather than a rectangular mesh.
• Is there a way to retain some of the benefits of CNNs in this generalized setting? How do we generalize a convolution to domains where "sliding windows" and "strides" don't make sense?

I recommend a couple of papers on this subject:

• Broad survey of convolutions for non-euclidean data:
  https://arxiv.org/pdf/1611.08097.pdf
  
• Focus on Spectral Networks: 
  https://arxiv.org/pdf/1312.6203.pdf
  
  

I plan to spend our time walking through some of the math, building intuition about how these things work.

See you on Wednesday!

-David Merrell