Making Robust Generalizers Less Rigid with Loss Concentration

Abstract

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 or difficult data points at test time. To achieve more robust and balanced generalization, methods applying sharpness-aware minimization to a subset of worst-case examples have proven successful for image classification tasks, but only using overparameterized neural networks under which the relative difference between "easy" and "hard" data points becomes negligible. In this work, we show how such a strategy can dramatically break down under simpler models where the difficulty gap becomes more extreme. As a more flexible alternative, instead of typical sharpness, we propose and evaluate a training criterion which penalizes poor loss concentration, which can be easily combined with loss transformations such exponential tilting, conditional value-at-risk (CVaR), or distributionally robust optimization (DRO) that control tail emphasis.