The document describes the mean teacher approach for semi-supervised learning. The mean teacher uses two neural networks, a student network and a teacher network, which are the same architecture but have different weights. The teacher's weights are an exponential moving average of the student's trained weights and provide better predictions to help train the student. This allows the student and teacher to improve each other in a virtuous cycle. Results shown demonstrate that the mean teacher approach achieves state-of-the-art results on image classification tasks using only a small percentage of labeled data.

1. SELF-ENSEMBLING
FOR VISUAL DOMAIN ADAPTATION
ICLR 2018

2. MEAN TEACHERS ARE
BETTERROLE MODELS
Weight-averaged consistency targets improve
semi-supervised deep learning results
ANTTI TARVAINEN & HARRI VALPOLA

3. HOW
THE MEAN TEACHER
WORKS

5. dog
dog
cat
horse horse
dog
cat
cat cat
horse dog
horse cat
horsecatdog

8. model’s
prediction
The truelabel pulls
predictions to
its direction.
SUPERVISED LEARNING
dog
dog
cat horse

9. The truelabel pulls
predictions to
its direction.
SUPERVISED LEARNING
dog
dog
cat horse
model’s
prediction

10. The truelabel pulls
predictions to
its direction.
SUPERVISED LEARNING
dog
dog
cat horse
model’s
prediction

11. The truelabel pulls
predictions to
its direction.
SUPERVISED LEARNING
dog
dog
dog
cat horse

12. The truelabel pulls
predictions to
its direction.
SUPERVISED LEARNING
dog
dog
cat horse

13. The truelabel pulls
predictions to
its direction.
SUPERVISED LEARNING
dog
dog
cat horse

14. WHAT ABOUT EXAMPLES
WITHUNKNOWN LABELS?
???
dog
cat horse

15. WHAT ABOUT EXAMPLES
WITHUNKNOWN LABELS?dog
How should we
adjust the prediction?
???
cat horse

16. dog
???
cat horse
We need two predictions:
a student and a teacher.

17. dog We need two predictions:
a student and a teacher.
Then, hopefully,
the student can learn
something from the
teacher.
???
cat horse

18. dog So how to do this?
Two ways.
???
cat horse

19. dog
We distort the student’s
input, making its task
more challenging.
1. MAKE THE
TASK HARDER.
cat horse

20. dog
1. MAKE THE
TASK HARDER.
cat horse
We distort the student’s
input, making its task
more challenging.
And then we train
the harder task to
predict the easier
tasks’ output.

21. dog
We maintain an exponential
moving average of weights
to create a better teacher.
2. MAKE THE
TEACHER BETTER.
???
cat horse

22. dog
We maintain an exponential
moving average of weights
to create a better teacher.
A mean teacher.
2. MAKE THE
TEACHER BETTER.
???
cat horse

23. dog
We maintain an exponential
moving average of weights
to create a better teacher.
A mean teacher.
Then we let the
student learn these
better predictions.
2. MAKE THE
TEACHER BETTER.
???
cat horse

24. dog
Combining these
two ways works
even better.
???
cat horse

25. dog
The student
and the teacher
improve each other
in a virtuous cycle.
???
cat horse

26. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

27. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

28. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

29. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

30. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

31. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

32. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

33. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

34. dog
???
cat horse
The student
and the teacher
improve each other
in a virtuous cycle.

35. HOW TO STARTUSING
THEMEAN TEACHER

36. θ
classification
cost
prediction
dog
Take a supervised model.
label
input
dog

37. θ θ’
classification
cost
prediction
dog
Make a copy of it.
label
input
dog
student teacher

38. θ θ’
exponentia
l moving
average
classification
cost
prediction
dog
Update teacher weights
after each training step.
label
input
dog
student teacher

39. θ θ’
exponentia
l moving
average
classification
cost
consistency
cost
prediction
dog
Add a cost between the two predictions.
label
input
dog
student teacher

40. θ θ’
exponentia
l moving
average
classification
cost
consistency
cost
prediction
noise noise
dog
Maybe add some noise.
label
input
dog
student teacher

41. θ θ’
exponentia
l moving
average
classification
cost
consistency
cost
prediction
noise noise
dog
Start using it for semi-supervised learning.
label
input
dog
student teacher

42. RESULTS

43. 50000
images
truck
4000
labeled

44. 0
2
4
12
10
8
6
testerrorrate
10,6 6,3
Virtual
Adversaria
l Training
Mean
Teacher
(ResNet)
using all labels
2,9
stateof theart
CIFAR-10 WITH4000 LABELS

45. IMAGENET 2012 WITH10%OF THELABELS
1280000
images
10%
labeled
brambling

46. 0
40
30
20
10
top-5validation
errorrate
using all the labels
Variational
Auto-Encoder
Mean Teacher
(ResNet-18)
3,8
stateof the art
35,2
19,8
IMAGENET 2012 WITH10%OF THELABELS

47. 0
10
20
30
40
top-5validation
errorrate
using all the labels
35,2
9,1
Variational
Auto-Encoder
Mean Teacher
(ResNet-152)
(theseresults included in the
Arxiv version of the paper)
3,8
stateof the art
IMAGENET 2012 WITH10%OF THELABELS

48. WHY TO USE
THEMEAN TEACHER

49. WHY MEAN TEACHER
1. It is easy to add to your model.
2. It can be adapted to different situations.
3. It gives good results.