DeepSeek releases new paper by Liang Wenfeng: Proposes mHC new architecture to improve large model training stability

PANews, January 1st – According to Golden Ten Data, DeepSeek has released a new paper proposing a novel architecture called Manifold-Constrained Hyperconnection (mHC). This architecture aims to address issues in Hyperconnection Network (HC) technology, such as training instability and limited scalability caused by the disruption of the identity mapping property. The mHC architecture restores the identity mapping property by mapping the residual connection space of HC to a specific manifold, while also incorporating rigorous infrastructure optimization to ensure efficiency. This results in significant performance improvements and superior scalability. DeepSeek anticipates that mHC, as a flexible and practical extension of HC, will contribute to a deeper understanding of topological architecture design and point to promising directions for the evolution of foundational models. The paper lists Zhenda Xie, Yixuan Wei, and Huanqi Cao as co-first authors, with Wenfeng Liang also included among the authors.