Making Robust Generalizers Less Rigid with Loss Concentration

Matthew J. Holland, Toma Hamada

While the traditional formulation of machine learning tasks is in terms of performance on average, in practice we are often interested in how well a trained model performs on rare…