Papers of the day   All papers

FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence


Ian Goodfellow: The quiet semisupervised revolution continues

8 replies, 1109 likes

David Berthelot: FixMatch: focusing on simplicity for semi-supervised learning and improving state of the art (CIFAR 94.9% with 250 labels, 88.6% with 40). Collaboration with Kihyuk Sohn, @chunliang_tw @ZizhaoZhang Nicholas Carlini @ekindogus @Han_Zhang_ @colinraffel

5 replies, 943 likes

Alexey Kurakin: Fixmatch: code for training on Imagenet dataset is released and available here: by Kihyuk Sohn @D_Berthelot_ML @chunliang_tw @ZizhaoZhang Nicholas Carlini @ekindogus @alexey2004 @Han_Zhang_ @colinraffel

0 replies, 264 likes

hardmaru: Happy to see SOTA benchmarks moving from CIFAR-10, to CIFAR-10 with only 40 training labels!

1 replies, 139 likes

Yann N. Dauphin: Ten years ago, the mcRBM really crushed it with an amazing 71% accuracy on CIFAR-10. Today, Fixmatch reaches 88.6% with 1000x fewer labelled examples. @D_Berthelot_ML #ML10YearChallenge

0 replies, 95 likes

Quoc Le: This work continues our efforts on semi-supervised learning such as UDA: MixMatch: FixMatch: Noisy Student: etc. Joint work with @hieupham789 @QizheXie @ZihangDai

1 replies, 64 likes

Colin Raffel: As a spurious, n=1 datapoint, the FixMatch paper ( is on its *third* submission whereas ReMixMatch and MixMatch were both unanimously accepted on first submission. FixMatch is simpler and works better and should supplant the other *ixMatches completely.

1 replies, 32 likes

Aakash Kumar Nain: Paper for the day!

0 replies, 16 likes

mat kelcey: Awesome example of how to stitch together a semi supervised learning pipeline! Makes me wonder how these operators; pseudo label, dataset union, augment, etc; could be represented and "learnt" through evolutionary approaches or framed as RL... @D_Berthelot_ML doable?

1 replies, 11 likes

Aakash Kumar Nain: Why I liked this paper so much? The "ablation study" section is so clearly written! Thanks @D_Berthelot_ML @colinraffel et al

1 replies, 10 likes

Daisuke Okanohara: FixMatch achieves new SOTA performance on semi-supervised learning tasks. 1) generate pseudo-labels with weak-augmentation and keep high confidence samples 2) train the model using generated training data with strong augmentation.

0 replies, 8 likes

phalanx: FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence paper: code:

0 replies, 8 likes

Greg Linden: @colinraffel Here's a direct link to the FixMatch paper, in case that is useful to others too:

0 replies, 4 likes

Alexey Romanov: FixMatch gets even better 88.61% accuracy in the same situation #MachineLearning

0 replies, 1 likes

彡Sαι彡: Thanks @CShorten30 @labs_henry My top -5 are, 1. Fixmatch - SSL @GoogleAI 2. Scaling laws of Neural language models @OpenAI 3. Squinting at VQA models @MSFTResearch 4. Jax&Swift

1 replies, 0 likes

Wei-Ning Hsu: Local Prior Matching: a simple yet effective semi-supervised ASR objective, showing SOTA WER-recovery (82.22%/91.39%) on LibriSpeech, effective with 360-60K hrs of speech Collaboration with Ann Lee, @syhw @awnihannun paper: code:

1 replies, 0 likes


Found on Jan 22 2020 at

PDF content of a computer science paper: FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence