MixMatch: A Holistic Approach to Semi-Supervised Learning

NeurIPS 2019  ·  David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, Colin Raffel ·

Semi-supervised learning has proven to be a powerful paradigm for leveraging unlabeled data to mitigate the reliance on large labeled datasets. In this work, we unify the current dominant approaches for semi-supervised learning to produce a new algorithm, MixMatch, that works by guessing low-entropy labels for data-augmented unlabeled examples and mixing labeled and unlabeled data using MixUp. We show that MixMatch obtains state-of-the-art results by a large margin across many datasets and labeled data amounts. For example, on CIFAR-10 with 250 labels, we reduce error rate by a factor of 4 (from 38% to 11%) and by a factor of 2 on STL-10. We also demonstrate how MixMatch can help achieve a dramatically better accuracy-privacy trade-off for differential privacy. Finally, we perform an ablation study to tease apart which components of MixMatch are most important for its success.

PDF Abstract NeurIPS 2019 PDF NeurIPS 2019 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data 		Result Benchmark
Image Classification CIFAR-10 MixMatch Percentage correct 95.05 # 131
Compare
Image Classification CIFAR-100 MixMatch Percentage correct 74.1 # 149
Compare
Semi-Supervised Image Classification CIFAR-10, 1000 Labels MixMatch Accuracy 92.25 # 1
Compare
Semi-Supervised Image Classification CIFAR-10, 2000 Labels MixMatch Accuracy 92.97 # 1
Compare
Semi-Supervised Image Classification CIFAR-10, 250 Labels MixMatch Percentage error 11.08 # 18
Compare
Semi-Supervised Image Classification CIFAR-10, 4000 Labels MixMatch Percentage error 6.24 # 26
Compare
Semi-Supervised Image Classification CIFAR-10, 500 Labels MixMatch Accuracy 91.35 # 1
Compare
Image Classification STL-10 IIC Percentage correct 88.80 # 43
Compare
Image Classification STL-10 CutOut Percentage correct 87.36 # 47
Compare
Image Classification STL-10 MixMatch Percentage correct 94.41 # 24
Compare
Semi-Supervised Image Classification STL-10, 1000 Labels MixMatch Accuracy 89.82 # 7
Compare
Semi-Supervised Image Classification STL-10, 5000 Labels MixMatch Accuracy 94.41 # 1
Compare
Image Classification SVHN MixMatch Percentage error 2.59 # 37
Compare
Semi-Supervised Image Classification SVHN, 1000 labels MixMatch Accuracy 96.73 # 7
Compare
Semi-Supervised Image Classification SVHN, 2000 Labels MixMatch Accuracy 96.96 # 1
Compare
Semi-Supervised Image Classification SVHN, 250 Labels MixMatch Accuracy 96.22 # 8
Compare
Semi-Supervised Image Classification SVHN, 4000 Labels MixMatch Accuracy 97.11 # 1
Compare
Semi-Supervised Image Classification SVHN, 500 Labels MixMatch Accuracy 96.36 # 2
Compare

Methods
Add Remove

No methods listed for this paper. Add relevant methods here
Contact us on: hello@paperswithcode.com . Papers With Code is a free resource with all data licensed under CC-BY-SA.
Terms Data policy Cookies policy