A round-up of topology-based papers at ICML 2020

The spheres are an example that we generated to showcase the benefits. It is arguably a little bit contrived, but we think that preserving this sort of relationship can be useful. The experiments in this paper indicate that this is also a suitable perspective to assume for more complex data sets, such as image manifolds.

There's also the option to go for a more local perspective, but in general I would say that preserving connectivity information is important because it aids reconstruction.

I guess that when working with 3D orientation data, this could be useful?