Unverified85% confidenceFactTime unknown
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
1
Sources
85%
Confidence
Long-term
Relevance
5/31/2026
First Seen
Sources
Related Entities
Related Claims
Partially 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.74% similarUnverified分布偏移意味着训练集与测试集之间的KL散度不为零，即两个分布之间存在可度量的信息差距70% similarUnverified当生成模型反复在自身生成的数据上训练时，输出分布会逐渐退化，低概率事件被系统性遗忘68% similarUnverified研究团队证明学习分布与真实分布之间的总变差距离可分解为初始化误差、Score误差和求解器离散化误差三项67% similarUnverifiedThe training data of detection models perpetually lags behind the output distribution of the latest generation technology, a problem referred to in academia as "distribution shift."66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/9802API
curl https://kongchang.com/api/v1/knowledge/claims/9802MCP
get_claim(id=9802)