Manifold Mixup: Better Representations by Interpolating Hidden States

Tutkimustuotos: Artikkeli kirjassa/konferenssijulkaisussavertaisarvioitu

Tutkijat

  • Vikas Verma

  • Alex Lamb
  • Christopher Beckham
  • Amir Najafi
  • Ioannis Mitliagkas
  • David Lopez-Paz
  • Yoshua Bengio

Organisaatiot

  • Montreal Institute for Learning Algorithms
  • Sharif University of Technology
  • Facebook Artificial Intelligence Research

Kuvaus

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 Manifold Mixup, a simple regularizer that encourages neural networks to predict less confidently on interpolations of hidden representations. Manifold Mixup 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 Manifold Mixup learn class-representations with fewer directions of variance. We prove theory on why this flattening happens under ideal conditions, validate it on practical situations, and connect it to previous works on information theory and generalization. In spite of incurring no significant computation and being implemented in a few lines of code, Manifold Mixup improves strong baselines in supervised learning, robustness to single-step adversarial attacks, and test log-likelihood.

Yksityiskohdat

AlkuperäiskieliEnglanti
OtsikkoProceedings of the 36th International Conference on Machine Learning
TilaJulkaistu - 2019
OKM-julkaisutyyppiA4 Artikkeli konferenssijulkaisuussa
TapahtumaInternational Conference on Machine Learning - Long Beach, Yhdysvallat
Kesto: 9 kesäkuuta 201915 kesäkuuta 2019
Konferenssinumero: 36

Julkaisusarja

NimiProceedings of Machine Learning Research
KustantajaPMLR
Vuosikerta97
ISSN (elektroninen)6438-6447

Conference

ConferenceInternational Conference on Machine Learning
LyhennettäICML
MaaYhdysvallat
KaupunkiLong Beach
Ajanjakso09/06/201915/06/2019

ID: 38544178