Unverified50% confidenceSolutionExact time
In practice, Muon uses approximately 5 steps of Newton-Schulz iteration to approximate the matrix sign function, reducing the computational overhead of exact SVD.
1
Sources
50%
Confidence
Long-term
Relevance
9/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified完整雅可比矩阵的计算代价是O(m×n)次前向传播，实际应用中研究者通常采用随机化SVD或Lanczos迭代等算法近似计算其低秩分解65% similarUnverifiedLoRA leverages the low-rank approximation principle of SVD to decompose the incremental updates of large weight matrices into the product of two low-rank matrices, thereby compressing the number of trainable parameters from billions to the millions range.64% similarUnverified传统梯度下降方法（如SGD、Adam）要求目标函数可微且连续，在面对离散、非凸、多模态问题时容易陷入局部最优63% similarUnverified微积分（特别是多元微积分和链式法则）直接驱动梯度下降这一核心优化算法60% similarUnverifiedWhen ε ≈ 1, the gradient second-moment term in the denominator is diluted, adaptive scaling has almost no effect, and the optimizer behaves close to pure SGD.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/937868API
curl https://kongchang.com/api/v1/knowledge/claims/937868MCP
get_claim(id=937868)