Timezone: »

Manifold Mixup: Better Representations by Interpolating Hidden States
Vikas Verma · Alex Lamb · Christopher Beckham · Amir Najafi · Ioannis Mitliagkas · David Lopez-Paz · Yoshua Bengio

Tue Jun 11 11:20 AM -- 11:25 AM (PDT) @ Hall A

Deep neural networks excel at learning the training data, but often provide incorrect and confident predictions when evaluated on slightly different test examples. This includes distribution shifts, outliers, and adversarial examples. To address these issues, we propose \manifoldmixup{}, a simple regularizer that encourages neural networks to predict less confidently on interpolations of hidden representations. \manifoldmixup{} leverages semantic interpolations as additional training signal, obtaining neural networks with smoother decision boundaries at multiple levels of representation. As a result, neural networks trained with \manifoldmixup{} learn flatter class-representations, that is, with fewer directions of variance. We prove theory on why this flattening happens under ideal conditions, validate it empirically on practical situations, and connect it to the previous works on information theory and generalization. In spite of incurring no significant computation and being implemented in a few lines of code, \manifoldmixup{} improves strong baselines in supervised learning, robustness to single-step adversarial attacks, and test log-likelihood.

Author Information

Vikas Verma (Aalto University)
Alex Lamb (Universite de Montreal)
Christopher Beckham (Ecole Polytechnique de Montreal)
Amir Najafi (Sharif University of Technology)
Ioannis Mitliagkas (University of Montreal)
David Lopez-Paz (Facebook AI Research)
Yoshua Bengio (Mila / U. Montreal)

Yoshua Bengio is recognized as one of the world’s leading experts in artificial intelligence and a pioneer in deep learning. Since 1993, he has been a professor in the Department of Computer Science and Operational Research at the Université de Montréal. He is the founder and scientific director of Mila, the Quebec Institute of Artificial Intelligence, the world’s largest university-based research group in deep learning. He is a member of the NeurIPS board and co-founder and general chair for the ICLR conference, as well as program director of the CIFAR program on Learning in Machines and Brains and is Fellow of the same institution. In 2018, Yoshua Bengio ranked as the computer scientist with the most new citations, worldwide, thanks to his many publications. In 2019, he received the ACM A.M. Turing Award, “the Nobel Prize of Computing”, jointly with Geoffrey Hinton and Yann LeCun for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing. In 2020 he was nominated Fellow of the Royal Society of London.

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors