S4L: Self-Supervised Semi-Supervised Learning
Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, Lucas Beyer
Code Available — Be the first to reproduce this paper.
ReproduceCode
Abstract
This work tackles the problem of semi-supervised learning of image classifiers. Our main insight is that the field of semi-supervised learning can benefit from the quickly advancing field of self-supervised visual representation learning. Unifying these two approaches, we propose the framework of self-supervised semi-supervised learning and use it to derive two novel semi-supervised image classification methods. We demonstrate the effectiveness of these methods in comparison to both carefully tuned baselines, and existing semi-supervised learning methods. We then show that our approach and existing semi-supervised methods can be jointly trained, yielding a new state-of-the-art result on semi-supervised ILSVRC-2012 with 10% of labels.
Tasks
Benchmark Results
| Dataset | Model | Metric | Claimed | Verified | Status |
|---|---|---|---|---|---|
| ImageNet - 10% labeled data | VAT + Entropy Minimization | Top 5 Accuracy | 83.39 | — | Unverified |
| ImageNet - 10% labeled data | S4L-MOAM (ResNet-50 4×) | Top 1 Accuracy | 73.21 | — | Unverified |
| ImageNet - 10% labeled data | Rotation + VAT + Ent. Min. | Top 5 Accuracy | 91.23 | — | Unverified |
| ImageNet - 10% labeled data | S4L-Rotation (ResNet-50) | Top 5 Accuracy | 83.82 | — | Unverified |
| ImageNet - 10% labeled data | S4L-Exemplar (ResNet-50) | Top 5 Accuracy | 83.72 | — | Unverified |
| ImageNet - 10% labeled data | Exemplar (joint training) | Top 5 Accuracy | 83.72 | — | Unverified |
| ImageNet - 10% labeled data | VAT + Entropy Minimization (ResNet-50) | Top 5 Accuracy | 83.39 | — | Unverified |
| ImageNet - 10% labeled data | VAT (ResNet-50) | Top 5 Accuracy | 82.78 | — | Unverified |
| ImageNet - 10% labeled data | VAT | Top 5 Accuracy | 82.78 | — | Unverified |
| ImageNet - 10% labeled data | Pseudolabeling (ResNet-50) | Top 5 Accuracy | 82.41 | — | Unverified |
| ImageNet - 10% labeled data | Pseudolabeling | Top 5 Accuracy | 82.41 | — | Unverified |
| ImageNet - 10% labeled data | Exemplar Fine-tuned (ResNet-50) | Top 5 Accuracy | 81.01 | — | Unverified |
| ImageNet - 10% labeled data | Exemplar | Top 5 Accuracy | 81.01 | — | Unverified |
| ImageNet - 10% labeled data | Rotation Fine-tuned (ResNet-50) | Top 5 Accuracy | 78.53 | — | Unverified |
| ImageNet - 10% labeled data | Rotation | Top 5 Accuracy | 78.53 | — | Unverified |
| ImageNet - 1% labeled data | Rotation (joint training) | Top 5 Accuracy | 53.37 | — | Unverified |
| ImageNet - 1% labeled data | Pseudolabeling | Top 5 Accuracy | 51.56 | — | Unverified |
| ImageNet - 1% labeled data | Exemplar (joint training) | Top 5 Accuracy | 47.02 | — | Unverified |
| ImageNet - 1% labeled data | VAT + Entropy Minimization | Top 5 Accuracy | 46.96 | — | Unverified |
| ImageNet - 1% labeled data | Rotation | Top 5 Accuracy | 45.11 | — | Unverified |
| ImageNet - 1% labeled data | Exemplar | Top 5 Accuracy | 44.9 | — | Unverified |
| ImageNet - 1% labeled data | VAT | Top 5 Accuracy | 44.05 | — | Unverified |