A theoretical ML paper came out recently that directly ties the notion of bias/fairness in machine learning to the tuning of the bias parameter in deep neural n

A theoretical ML paper came out recently that directly ties the notion of bias/fairness in machine learning to the tuning of the bias parameter in deep neural network architectures. The authors used stochastic backpropagation to classify an unbalanced dataset (85% class 0, 15% class 1) and show, through a series of topological cross-validation techniques, that very small amounts of perturbation in the bias parameter results in significant increase in the overall bias of the model towards the minority class (class 1).