mixup: Beyond Empirical Risk Minimization

TLDR

Large deep neural networks are powerful, but they often memorize training data and are sensitive to adversarial examples. The authors propose mixup, a simple learning principle designed to alleviate these issues. Mixup trains a network on convex combinations of example pairs and their labels, encouraging linear behavior between training points. Experiments on ImageNet, CIFAR, Google Commands, and UCI datasets show that mixup improves generalization, reduces memorization of corrupt labels, increases adversarial robustness, and stabilizes GAN training.

Abstract

Large deep neural networks are powerful, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. In this work, we propose mixup, a simple learning principle to alleviate these issues. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples. Our experiments on the ImageNet-2012, CIFAR-10, CIFAR-100, Google commands and UCI datasets show that mixup improves the generalization of state-of-the-art neural network architectures. We also find that mixup reduces the memorization of corrupt labels, increases the robustness to adversarial examples, and stabilizes the training of generative adversarial networks.