HTorch - Hyperbolic Learning with MCTensor

Last updated on Oct, 2023
Code
Table 3 in the MCTensor paper.

We want to build all hyperbolic learning models (with manifolds as Halfspace, Lorentz, Poincare, etc.) on top of MCTensor algorithms. This might address the well-known NaN problems in hyperbolic learning, especially for manifolds that are known to be numerically unrobust (Poincare). We have released a library called HTorch and we will continue to develop and maintain such library. We are looking forward to writing a paper on HTorch.

Hyperbolic learning Efficient high-precision floating-point arithmetic
Wentao Guo
Wentao Guo
CS PhD student @ Princeton

I am a second-year CS PhD student at Princeton University advised by Prof. Tri Dao. I am obsessed by the sweet intersection of hardware efficiency and ML model architecture design, and I want to contextualize the expressiveness of ML models among the principles of hardware efficiency.