Self-training with Noisy Student improves ImageNet classification

Qizhe XieMinh-Thang LuongEduard HovyQuoc V. Le

   Papers with code   Abstract  PDF

We present a simple self-training method that achieves 88.4% top-1 accuracy on ImageNet, which is 2.0% better than the state-of-the-art model that requires 3.5B weakly labeled Instagram images. On robustness test sets, it improves ImageNet-A top-1 accuracy from 61.0% to 83.7%, reduces ImageNet-C mean corruption error from 45.7 to 28.3, and reduces ImageNet-P mean flip rate from 27.8 to 12.2... (read more)

Benchmarked Models

RANK
MODEL
REPO
CODE RESULT
PAPER RESULT
ε-REPRODUCED
BUILD
1
EfficientNet-L2
(NoisyStudent)
88.4%
--
2
EfficientNet-L2 475
(NoisyStudent)
88.2%
--
3
EfficientNet-B7
(NoisyStudent)
86.8%
--
4
EfficientNet-B6
(NoisyStudent)
86.5%
--
5
EfficientNet-B5
(NoisyStudent)
86.1%
--
6
EfficientNet-B4
(NoisyStudent)
85.2%
--
7
EfficientNet-B3
(NoisyStudent)
84.1%
--
8
EfficientNet-B2
(NoisyStudent)
82.4%
--
9
EfficientNet-B1
(NoisyStudent)
81.4%
--
10
EfficientNet-B0
(NoisyStudent)
78.7%
--