Unverified50% confidenceFactExact time
数据增强本质上是一种正则化技术，通过人为扩展训练分布来逼近真实数据生成分布，可以系统性地降低泛化误差
1
Sources
50%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
Unverified数据增强的核心原则是增强分布应当匹配真实测试分布，理想策略是让训练变换分布成为真实测试分布的一个超集79% similarUnverified数据增强从偏差-方差权衡视角看，本质是通过增加训练样本的有效多样性来降低模型方差76% similarUnverifiedBehavioral cloning suffers from distribution shift; improved algorithms such as DAgger address this by having the learning agent continuously interact with the expert to expand the training data.74% similarUnverified数据增强通过对训练样本施加随机变换（如旋转、翻转、缩放、裁剪、颜色抖动等）来人工扩大训练集规模，其理论基础是正则化73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/515875API
curl https://kongchang.com/api/v1/knowledge/claims/515875MCP
get_claim(id=515875)