Learning On The Hypersphere: The Multi-Geometric Neural Network

Link to Colab Notebook: colab.research...
Link to Nvidia Research Paper: arxiv.org/abs/...
Due to popular demand, I have started a Discord server! You can check it out here if you like: / discord
This video is about a new type of neural network that is being developed by researchers at the University of California, Berkeley. The network is called a multi-geometric neural network, and it is designed to learn from data that is distributed on a hypersphere. Hyperspheres are mathematical objects that are similar to spheres, but they have more dimensions. The new neural network is able to learn from data that is distributed on a hypersphere more effectively than traditional neural networks. This is because the new network is able to take advantage of the geometry of the hypersphere. The researchers believe that the new neural network could be used to solve a variety of problems, such as image recognition, natural language processing, and robotics.
The video also discusses the challenges of developing and training multi-geometric neural networks. The researchers are still working on improving the performance of the new network. However, they believe that the potential benefits of the new network are significant.
Here are some of the key points that are discussed in the video:
The challenges of developing and training multi-geometric neural networks.
The potential benefits of the new network.
The applications of multi-geometric neural networks.
Overall, the video is a good introduction to the topic of multi-geometric neural networks. It provides a clear overview of the challenges and potential benefits of the new network.

[@bt653]

If the "physics" of the representation are compatible with our actual universe, there could be huge potential for taking advantage of this to engineer computer hardware! The developments in photonic computing recently such as from Q.ANT ("Native Processing Unit") are quite interesting in what types of computations will be possible soon.

[@irbsurfer1585]

This appears to be a unique type of phase transition, distinct from standard grokking, but sharing some of its characteristics? A new kind of grokking specific to multi-geometry models, where spatial curvature acts as an implicit training signal? I think I really need to spend more time absorbing your ideas and brushing up on my hyperbolic geometry math. lol So just curvature differences, poincare disk model, and maybe just how trees and graphs fit with euculidean space? Or deeper still?

[@davidjohnston855]

What if you used your RL reasoning strategy to shape and optimize the layers? Or use your fibinacci strategy instead of backprop?

[@richardaragon8471]

I'll play around with this at some point in the future, I like it!

[@carson1391]

thats hilarious if they sent this to you. they be watchin lol

[@beo-w]

🐐