知识卡片
偏差治理需要两条路径
改进数据采集可减少已知偏差,但未知部署场景、稀缺样本与未察觉的偏差仍可能保留。后续路径一是增强或域随机化扩大训练变化,二是迁移学习或模型适配利用目标域信息。两者都应回到代表实际部署的评估,而不是把变换更多当作已经消除偏差。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
One way to reduce the issues of dataset bias and shift is to implement a rigorous data collection procedure that tries to minimize the biases present in the data. Although it is possible to prevent (or reduce) certain forms of bias, others will remain challenging because it is hard to collect the necessary data or because we are not aware of it. In many cases,
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第35章《数据偏差与分布漂移》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c035b.xhtml)
- 结论依据:摘录中关于偏差治理需要两条路径的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:One way to reduce the issues of dataset bias and shift is to implement a rigorous data collection procedure that tries to minimize the biases present in the data. Although it is possible to prevent (or reduce) certain forms of bias, others will remain challenging because it is hard to collect the necessary data or because we are not aware of it. In many cases,