Bias in Image Understanding Algorithms
Fourth-Year Image Understanding
One 2-hour class
Fri 03 March 2023
In this module, students examine technical and ethical aspects of bias in image understanding algorithms. The module begins with a discovery activity led by the computer science instructor in which students look through the Celeb-A image dataset. They consider the uses of the dataset as well as the demographics that are missing from the dataset. The computer science instructor then discusses the different sources of bias in datasets, including those described in studies by Timnit Gebru, as well as some strategies for mitigating them. Next, the ethics instructor explores how to approach situations in which it is not possible to fully mitigate bias without cost. First, students create simple linear classifiers that illustrate the trade-offs between overall accuracy and subgroup accuracy. In a discussion with the ethics instructor, students then explore the different reasons that might lead them to prioritize subgroup accuracy against overall accuracy, and how those reasons might apply to different demographic subgroups.
This module was developed by Steven Coyne, Babak Taati, and David Lindell. Diane Horton, David Liu, and Sheila McIlraith provided feedback on this module.
Module materials coming soon.