Are there any active communities working in the space of hyperbolic deep learning? I’m specifically interested in Flux-compatible libraries which define the basic DL operations for either the Poincare or Lorentz hyperbolic spaces: addition, multiplication, softmax, etc. Bonus points for hyperbolic versions of Adam, etc, along with docs/examples.
If these don’t exist and I was to build them, is there an existing org where the package could/should live?
Not sure, this is the first time I hear of this interesting thing:
http://hyperbolicdeeplearning.com/hyperbolic-neural-networks/
Again it uses Python/PyTouch, do you have good reason to think Julia would be better? Why is Python, not Julia, still used for most state-of-the-art AI research? - #4 by Palli
The Adam optimizer?
It’s amazing to me how advanced math (e.g. softmax and tanh) has anything to do with even natural language processing (in ordinary ANNs), and now more advanced geometry than Euclidean: https://arxiv.org/pdf/1006.5169.pdf
Hyperbolic Geometry of Complex Networks
[…] SNLI task and dataset. We evaluate our method on two tasks. The first is natural language inference, or textual entailment
I should have said that I know about the “Manifold aware” library http://github.com/geoopt/geoopt which is based on pytorch (which I actively use). I could always use that in python, or call it using PyCall.
Julia has better potential for non-vanilla use cases where there is substantial computation occurring outside of the kernel functions. I’m attempting to evaluate the cost of using Julia for this work, having a good prior sense of the benefit.
I just discovered the JuliaManifolds organization which seems to be what I was looking for. 
@tbreloff Did you ever get a hyperbolic deep learning model trained? I wanted to prototype something and wasn’t sure if I needed to fall back to PyTorch.