Fairness in Deep Learning: A Computational Perspective

Nevertheless, fairness in machine learning remains a problem. Machine learning algorithms have the risk of amplifying societal stereotypes by over associating protected attributes, e.g., race and gender, with the main prediction task [33]. Eventually they are capable of exhibiting discriminatory behaviors against certain subgroups. For example, a recruiting tool believes that men are more qualified and shows bias against women [26], facial recognition performs extremely poorly for darker skin females [5], recognition accuracy is very low for subgroup of people in pedestrian detection of self-driving cars [33]. The fairness problem might cause adverse impacts on individuals and society. It not only limits a person's opportunities which he/she is qualified, but also might further exacerbate social inequity. Among different machine learning models, the fairness problem of deep learning models has especially attracted attention from academia and industry recently. First, deep learning models have outperformed conventional machinePermission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.