SOTAVerified

Interpolation Consistency Training for Semi-Supervised Learning

2019-03-09Code Available0· sign in to hype

Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, David Lopez-Paz

Code Available — Be the first to reproduce this paper.

Reproduce

Code

Abstract

We introduce Interpolation Consistency Training (ICT), a simple and computation efficient algorithm for training Deep Neural Networks in the semi-supervised learning paradigm. ICT encourages the prediction at an interpolation of unlabeled points to be consistent with the interpolation of the predictions at those points. In classification problems, ICT moves the decision boundary to low-density regions of the data distribution. Our experiments show that ICT achieves state-of-the-art performance when applied to standard neural network architectures on the CIFAR-10 and SVHN benchmark datasets. Our theoretical analysis shows that ICT corresponds to a certain type of data-adaptive regularization with unlabeled points which reduces overfitting to labeled points under high confidence values.

Tasks

Benchmark Results

DatasetModelMetricClaimedVerifiedStatus
CIFAR-10, 1000 LabelsICT (CNN-13)Accuracy84.52Unverified
CIFAR-10, 2000 LabelsICT (CNN-13)Accuracy90.74Unverified
CIFAR-10, 4000 LabelsICT (CNN-13)Percentage error7.29Unverified
CIFAR-10, 4000 LabelsICT (WRN-28-2)Percentage error7.66Unverified
SVHN, 1000 labelsICT (WRN-28-2)Accuracy96.47Unverified
SVHN, 1000 labelsICTAccuracy96.11Unverified

Reproductions