ML@HZG episode 2: 'Don't Fear the Sphere'

0:00:01 Announcements

5 minute presentations
0:01:42 Julianna Carvalho
0:07:42 Tobias Finn
0:15:49 Lennart Marien

0:25:15 We cover "Spherical CNNs on Unstructured Grids," Jiang et al. 2019, arXiv.

The paper describes how convolution-like operations can be learned and applied when input data exist on an irregularly spaced 3D mesh instead of a flat plane. We discuss how this approach works, and why taking into account the sphere's curved geometry can be so important when designing an ML architecture. The discussion is mostly non-technical, but we mention a few equations and try to give intuitions as to what's going on.

Some links relevant to our discussion:

We cover additional approaches for 1-2 minutes each:

1:25:17 Adjusting the convolutional filters to compensate distortions