Unverified50% confidenceFactExact time
A 2019 study by Yang et al. pointed out that BatchNorm may in fact introduce new gradient instability in the early stages of training.
1
Sources
50%
Confidence
Long-term
Relevance
8/22/2026
First Seen
Sources
ResNet残差连接为何有效？深层网络退化问题实验复现
Related Entities
Related Claims
UnverifiedDuring cross-domain fine-tuning, an excessively large learning rate will destroy pre-trained weights, while an excessively small one will result in almost no learning, requiring a careful trade-off.63% similarUnverified当前对齐方法主要依赖训练阶段的行为塑造，在面对精心设计的对抗性提示时安全边界可能被突破，这种现象称为'越狱'（jailbreaking）61% similarUnverified低精度引入的数值扰动会在RL的多轮迭代中通过策略梯度的乘积形式被放大，可能导致训练崩溃或模型性能退化60% similarVerifiedGANs have inherent drawbacks such as training instability and susceptibility to mode collapse, which prompted the research community to shift toward diffusion models that generate images through a gradual denoising Markov chain process.59% similarUnverifiedRAG technology addresses two fundamental shortcomings of LLMs: the timeliness limitations of training data and the hallucination problem.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/788587API
curl https://kongchang.com/api/v1/knowledge/claims/788587MCP
get_claim(id=788587)