Adaptive Mixing of Auxiliary Losses in Supervised Learning