Computer Vision Foundations · 免费预览
测量与修复算法偏差
面部分析性能如何随人口统计群体而变化，数据集偏差如何产生这种变化，以及生成模型和反事实探针如何攻击它。
1. Vision systems in society
Vision systems in society
2. Facial analysis across demographic groups
Facial analysis across demographic groups
3. Where the bias comes from
Where the bias comes from
4. Generating the missing data
Generating the missing data
5. Why report both error types per group
A face recognition vendor reports 99% top-line accuracy for its system. Why does this single number fail as evidence of fairness across demographic groups?
6. The factor between groups
In a vendor test, one demographic group shows a false positive rate of for a given face recognition system. Another group shows a false positive rate that is a factor of 100 higher. Write the higher false positive rate as a decimal.