Explanatory Debiasing with Domain Experts for Representation Bias
A set of generic design guidelines for bringing domain experts (who lack AI expertise) into the process of reducing representation bias, by giving them explanations that let them understand where a dataset under-represents groups and steer the data-generation or augmentation algorithms that fix it. Students learn how human-in-the-loop, explanation-driven workflows can correct biased training data without sacrificing model accuracy, illustrated and evaluated in a healthcare application with a mixed-methods study of 35 experts.
R
Rosie
Text
Explanatory Debiasing: Involving Domain Experts in the Data Generation Process to Mitigate
Representation bias arises when training data under-represents certain groups, so an AI model performs poorly on them (e.g. a skin-condition classifier trained mostly on lighter-skinned images). Data…