Fade: fairness-aware deep ensemble for quantifying uncertainty

This paper brings together two important aspects of trustworthy ML that go beyond mere accuracy that is calibration and fairness. A model is well calibrated if predictions that have a confidence of 0.6 are true 60% of the time. For fairness, we require calibration-by-groups and no underestimation, i.e., that predictions are consistent with respect to the actual distribution, across sensitive feature values. The deep ensemble method provides uncertainty estimates, as a form of transparency, by retraining the same neural network architecture with different weight initializations to produce a distribution of predictions. This method provides a compelling approach to approximating Bayesian predictive distributions, resulting in well-calibrated predictions. Through a series of experiments, we demonstrate that deep ensemble models are often well calibrated in terms of posterior probabilities but less so when a sensitive attribute is involved, resulting in biased predictions. This should not come as a surprise because the distributions of weights in the deep ensemble are usually optimized to maximize generalization accuracy without explicit consideration of fairness. To address this issue, we propose Fairness-Aware Deep Ensemble (FADE), a multi-objective optimization strategy to optimize the deep ensemble model on accuracy and fairness. We empirically evaluate our framework on two synthetic and twelve real-world datasets. We find that FADE can obtain fairer models while still attaining adequate overall generalization accuracy and calibrated probability estimates.