Towards practical model fairness for trustworthy and safe AI

Abstract

As artificial intelligence (AI) has an increasing societal impact in our world, developing fair AI becomes important to avoid adapting or even amplifying social biases and discrimination. Although many techniques for training fair models have been proposed, most of them face significant limitations that make it challenging to apply them in practice. To address these challenges, this thesis provides fundamental solutions for 1) lowering the technical barriers of fair AI development and 2) achieving high fairness even when the training and test data contain errors or change over time, and thus lays the foundation for practical and trustworthy AI. Furthermore, we aim to extend our techniques to mitigate ethical concerns associated with foundation models, which are recently being used in many applications at an explosive pace, and suggest new opportunities for making foundational models fair and safe to use.