Projection-wise Disentangling for Fair and Interpretable Representation Learning: Application to 3D Facial Shape Analysis.