Unverified50% confidenceFactExact time
Gaussians with fixed covariance are a canonical example of NGD being exactly optimal, as their mixture geodesics coincide completely with Fisher–Rao geodesics, with R reaching machine-precision-level 1 in experiments.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen the optimization objective involves quadratic potential function families such as fixed-covariance Gaussians, NGD can be used with confidence, as it is theoretically already optimal.74% similarUnverifiedOU过程的平稳分布为高斯分布，其转移核为高斯分布，使任意时刻的条件分布可以精确计算62% similarUnverifiedWhen the prior is a uniform distribution, MAP degenerates into MLE; a Gaussian prior corresponds to Ridge regression, and a Laplace prior corresponds to Lasso regression60% similarUnverifiedMulti-model distributed output (an independent large model for each stage) delivers the best quality but is the slowest and most token-intensive; direct output is the fastest and is suitable for early-stage validation; single-model step-by-step output is the balanced choice.60% similarUnverifiedQLoRA论文提出的NF4针对预训练权重服从正态分布的特性设计了最优量化格点59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/983526API
curl https://kongchang.com/api/v1/knowledge/claims/983526MCP
get_claim(id=983526)