Demographic Bias Mitigation at Test-Time Using Uncertainty Estimation and Human-Machine Partnership
Abstract
Facial attribute classification algorithms frequently manifest demographic biases by obtaining differential performance across gender and racial groups. Existing bias mitigation techniques are mostly in-processing techniques, i.e., implemented during the classifier’s training stage, that often lack generalizability, require demographically annotated training sets, and exhibit a trade-off between fairness and classification accuracy. In this paper, we propose a technique to mitigate bias at the test time i.e., during the deployment stage, by harnessing prediction uncertainty and human-machine partnership. To this front, we propose to utilize those lowest percentages of test data samples identified as outliers with high prediction uncertainty. These identified uncertain samples at test-time are labeled by human analysts for decision rendering and for subsequently re-training the deep neural network in a continual learning framework. With minimal human involvement and through iterative refinement of the network with human guidance at test-time, we seek to enhance the accuracy as well as the fairness of the already deployed facial attribute classification algorithms. Extensive experiments are conducted on gender and smile attribute classification tasks using four publicly available datasets and with gender and race as the protected attributes. The obtained outcomes consistently demonstrate improved accuracy by up to 2% and 5% for the gender and smile attribute classification tasks, respectively, using our proposed approaches. Further, the demographic bias was significantly reduced, outperforming the State-of-the-Art (SOTA) bias mitigation and baseline techniques by up to 55% for both classification tasks.