๐ŸŽ“How I Study AIHISA
๐Ÿ“–Read
๐Ÿ“„Papers๐Ÿ“ฐBlogs๐ŸŽฌCourses
๐Ÿ’กLearn
๐Ÿ›ค๏ธPaths๐Ÿ“šTopics๐Ÿ’กConcepts๐ŸŽดShorts
๐ŸŽฏPractice
๐ŸงฉProblems๐ŸŽฏPrompts๐Ÿง Review
Search
How I Study AI - Learn AI Papers & Lectures the Easy Way

Papers1

AllBeginnerIntermediateAdvanced
All SourcesarXiv
#Parameter Efficiency

KromHC: Manifold-Constrained Hyper-Connections with Kronecker-Product Residual Matrices

Intermediate
Wuyang Zhou, Yuxuan Gu et al.Jan 29arXiv

Hyper-Connections (HC) make the usual single shortcut in neural networks wider by creating several parallel streams and letting the model mix them, but this can become unstable when stacked deep.

#Hyper-Connections#Manifold-Constrained Hyper-Connections#Doubly Stochastic Matrix