Unverified50% confidenceFactExact time
MGD 对学生模型的特征图随机遮盖一定比例（通常75%）的空间区域，训练轻量级生成模块以未遮盖区域重建教师模型对应位置的完整特征
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
Unverified领域自适应作为迁移学习子领域，常用方法包括对抗训练（DANN）、最大均值差异最小化（MMD）以及基于元学习框架（如MAML）的少样本适应60% similarUnverifiedFine-tuning typically requires only one percent or less of the original training cost to achieve domain performance close to that of a dedicated model.57% similarUnverified在过参数化区间存在无数个能完美拟合训练数据的解，优化算法如SGD倾向于选择范数最小的解，这种隐式正则化使模型获得更好泛化性能57% similarUnverifiedModels trained with SFT (Supervised Fine-Tuning) have an 'imitation ceiling,' where their performance upper bound is constrained by the quality and diversity of the labeled data.57% similarUnverified微调VLA模型时通常采用分层学习率策略：视觉编码器使用极小学习率或冻结，语言模型层使用中等学习率，动作专家使用较大学习率57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/541882API
curl https://kongchang.com/api/v1/knowledge/claims/541882MCP
get_claim(id=541882)