DeepSeek's Manifold Constrained Hyper Connections (mHC) and the evolution of ResNets
DeepSeek's new mHC: Manifold-Constrained Hyper-Connections is one of the top trending papers in the ML space currently - and this video visually breaks down everything you need to know about it. We talk about Residual connections, their role in Deep Learning, the idea of Hyper-Connections, and build intuition about the new Manifold-Constraint idea that is introduced in this paper. I have tried to use as many visualizations and illustrations to break down these concepts - feel free to pause, rewind, and rewatch sections that require extra attention! This is a very simple optimization on the sur…
Watch on YouTube ↗
(saves to browser)
Chapters (7)
Intro
1:51
Language Models
4:29
Residual Connections
5:51
Identity Mapping
8:47
Hyper Connections
14:43
Signal Gain
16:30
Manifold Constraints and the Birkhoff Polytope
DeepCamp AI