DeepSeek改造何恺明残差连接！梁文峰亲自署名，十年首次重大升级

Core Viewpoint - The article discusses the evolution and enhancement of the residual connection, a fundamental component in deep learning introduced by He Kaiming in ResNet, and presents a new approach called Hyper-Connections (HC) that aims to improve performance while addressing potential issues related to signal amplification and stability in deep learning architectures [2][7][11]. Group 1: Residual Connections and Their Evolution - Residual connections have been a cornerstone of deep learning since the introduction of ResNet in 2016, allowing signals to pass directly from shallow to deep layers without modification [7][9]. - The rise of Transformer architectures has made residual connections a standard feature in large language models like GPT and LLaMA [10]. - Hyper-Connections (HC) expand the residual flow width from C dimensions to n×C dimensions, introducing three learnable mapping matrices to manage information flow [11]. Group 2: Performance and Stability Challenges - Experiments by the DeepSeek team indicate that the Hres matrix, responsible for internal information exchange in HC, significantly enhances performance [12]. - However, when HC is extended to multiple layers, the composite mapping loses its identity property, leading to potential issues such as sudden loss spikes and gradient fluctuations during training [14]. - The peak amplification factor of signals in HC can reach 3000, which poses risks of signal distortion during inter-layer propagation [16]. Group 3: Theoretical Framework and Constraints - The core idea of the DeepSeek paper is to constrain the residual mapping matrix to a specific manifold formed by double stochastic matrices, which ensures three key theoretical properties: norm preservation, combinatorial closure, and geometric interpretation [17][19]. - The Sinkhorn-Knopp algorithm is employed to project any matrix onto this manifold, effectively reducing the signal amplification issue observed in HC [21]. Group 4: Engineering Optimizations - The paper details the memory access costs associated with expanding the residual flow width, highlighting significant increases in read and write operations for HC compared to standard residual connections [24]. - To mitigate these costs, the team developed infrastructure optimizations, including the TileLang framework for merging operations and specialized kernels for the Sinkhorn-Knopp algorithm [25][26]. - The paper also discusses pipeline parallelism enhancements to overlap computation and communication, improving overall efficiency [27]. Group 5: Experimental Validation - The paper validates the proposed methods on MoE models of sizes 3B, 9B, and 27B, with an expansion rate of n set to 4 [30]. - In the 27B MoE model, the modified HC (mHC) demonstrated a stable training curve, achieving a loss reduction of 0.021 compared to the baseline while maintaining gradient stability [31]. - Performance improvements were noted in downstream tasks, with mHC outperforming both the baseline and HC in various benchmarks [32][35].