Unverified50% confidenceFactExact time
低秩参数化将完整协方差矩阵分解为Σ = WW^T + D的形式，使参数量从O(d²)降至O(dr)
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMLA通过低秩分解将KV Cache显存开销从O(n·d_model)压缩到O(n·d_latent)，其中d_latent约为d_model的1/870% similarUnverifiedL2算法相对于基向量系数位长β达到O(n⁴·β·(n+β))的位复杂度，在β=O(n)时相当于O(n⁶)，而经典LLL约为O(n⁹)67% similarUnverifiedLoRA的原理是将微调中权重变化量ΔW分解为两个低秩矩阵的乘积（ΔW = A×B）65% similarUnverified在鲁棒回归中将似然函数从高斯替换为t分布，等价于在损失函数中降低离群点的权重，能提升模型对异常值的鲁棒性65% similarUnverified哈达玛矩阵满足正交性条件 H_n · H_n^T = n · I_n，使哈达玛变换可逆且不丢失信息64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/894739API
curl https://kongchang.com/api/v1/knowledge/claims/894739MCP
get_claim(id=894739)