Unverified50% confidenceTradeoffExact time
In kernel density estimation (KDE), when the Gaussian kernel bandwidth parameter σ is too small the estimate degenerates into a collection of discrete points, while when σ is too large it over-smooths and loses detailed information, reflecting the bias-variance tradeoff.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified该测试采用KL散度（KLD）作为客观量化指标，衡量量化后输出分布与BF16参考分布的偏离程度，KLD越低说明量化造成的信息损失越小70% similarUnverified从KL散度角度看，每一代合成数据与真实分布之间的散度都在单调递增，且误差会在迭代中累积69% similarPartially VerifiedKL divergence D(p,q) describes how many extra bits are wasted when using an incorrect distribution q to encode data that actually follows distribution p, compared to using the correct distribution p.65% similarUnverifiedCWD 通过 Softmax 对每个通道的空间激活值归一化转化为概率分布，再用 KL 散度计算师生模型通道分布差异作为蒸馏损失64% similarUnverifiedThe exponential distribution Exp(λ) describes the waiting time between successive events in a Poisson process and possesses the memoryless property; the Gamma distribution is the distribution of the sum of k independent exponential random variables and degenerates to the exponential distribution when k=1.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/831939API
curl https://kongchang.com/api/v1/knowledge/claims/831939MCP
get_claim(id=831939)