Summary

In this chapter, we dove into the concept of bias in ML, and especially explored angles in vision and language. We opened with a general discussion of human bias and introduced a few ways these empirically manifest in technology systems, frequently without intention. We introduced the concept of “intersectional bias”, and how commonly your first job in detecting bias is listing a few common types of intersections you want to be wary of, including gender or race and employment, for example. We demonstrated how this can easily creep into large vision and language models trained on datasets crawled from the internet. We also explored methods to mitigate bias in ML models. In language, we presented counterfactual data augmentation along with fair loss functions. In vision, we learned about the problem of correlational dependencies, and how you can use open source tools to analyze your vision dataset and solve sampling problems.

Finally, we learned about monitoring...