Interpolation Consistency Training for Semi-Supervised Learning
Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, David Lopez-Paz
Code Available — Be the first to reproduce this paper.
ReproduceCode
- github.com/vikasverma1077/ICTOfficialIn paperpytorch★ 0
- github.com/kevinghst/ICTpytorch★ 0
- github.com/yanzhicong/mixup_sslpytorch★ 0
- github.com/perrying/realistic-ssl-evaluation-pytorchpytorch★ 0
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
| Dataset | Model | Metric | Claimed | Verified | Status |
|---|---|---|---|---|---|
| CIFAR-10, 1000 Labels | ICT (CNN-13) | Accuracy | 84.52 | — | Unverified |
| CIFAR-10, 2000 Labels | ICT (CNN-13) | Accuracy | 90.74 | — | Unverified |
| CIFAR-10, 4000 Labels | ICT (CNN-13) | Percentage error | 7.29 | — | Unverified |
| CIFAR-10, 4000 Labels | ICT (WRN-28-2) | Percentage error | 7.66 | — | Unverified |
| SVHN, 1000 labels | ICT (WRN-28-2) | Accuracy | 96.47 | — | Unverified |
| SVHN, 1000 labels | ICT | Accuracy | 96.11 | — | Unverified |