Back    Zoom +    Zoom -
Liang Wenfeng, Others Publish Signed Paper Claiming DeepSeek Opens New Chapter in Architecture
Recommend
2
Positive
4
Negative
1
Chinese AI startup DeepSeek released a new paper on New Year's Day, proposing a new architecture called mHC (Manifold Constrained Hyperconnection), aimed at addressing the instability issues of traditional hyperconnections in large-scale model training while maintaining significant performance gains.

DeepSeek's mHC extends the traditional Transformer single residual stream into a multi-stream parallel architecture, utilizing the Sinkhorn-Knopp algorithm to constrain the connection matrix on a doubly stochastic matrix manifold, successfully resolving the numerical instability and signal explosion issues caused by the disruption of identity mapping properties in hyperconnections (HC) during large-scale training.

The first authors of the paper included Zhenda Xie, Yixuan Wei, and Huanqi Cao. DeepSeek founder Liang Wenfeng was also listed among the authors.
AASTOCKS Financial News
Website: www.aastocks.com