Deep learning technology has been widely deployed in increasingly critical decision-making tasks, such as mortgage approval, credit card assessment, college admissions, employee selection, and recidivism prediction. However, these areas are observed to be subject to long-standing systematic discrimination against certain people on the basis of diverse background traits, including race, gender, nationality, age, and religion. Unfortunately, the introduction of intelligent algorithms into these areas has failed to relieve the discrimination conundrum because people with minority backgrounds are institutionally underrepresented in the historical data that fuel the algorithmic systems. Thus, the unfairness residing in biased historical data is inherited and sometimes intensified by a machine learning (ML) model that is trained on such data. The consequent fairness concerns are particularly severe due to the black-box nature of ML algorithms. Therefore, mitigation of the fairness issues arising in ML applications is both urgent and important.