知识卡片
反事实探针依赖属性控制质量
通过改变探针图像某一属性并观察输出,可研究模型对该变化是否敏感,补充受测试集相关性影响的观察比较。GAN潜空间方向与人工标注用于实现受控变换,但必须核实其它特征是否也改变、样本是否合理;不能自动视为真实世界因果干预。
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
- · 计算机视觉与社会
原书出处 · 1 处
- · 查看原书依据
modifying variables of a probe image and examining the algorithm's decision. GAN models, when coupled with human labeling to learn directions in the latent space corresponding to changes in the desired attributes, allow for such experimental intervention. Researchers [30] have shown that such counterfactual studies may lead to very different conclusions than observational studies alone, which can be influenced by biased correlations in the testing
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第04章《计算机视觉与社会》(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/c004.xhtml)
- 结论依据:摘录中关于反事实探针依赖属性控制质量的定义、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:modifying variables of a probe image and examining the algorithm's decision. GAN models, when coupled with human labeling to learn directions in the latent space corresponding to changes in the desired attributes, allow for such experimental intervention. Researchers [30] have shown that such counterfactual studies may lead to very different conclusions than observational studies alone, which can be influenced by biased correlations in the testing