Active Learning and Human-in-the-Loop

49
Lukas Biewald

Transcript of Active Learning and Human-in-the-Loop

Page 1: Active Learning and Human-in-the-Loop

Lukas Biewald

Page 2: Active Learning and Human-in-the-Loop

The Effect of Better Algorithms

Naïve Bayes Maximum Entropy SVM0%

5%

10%

15%

20%

25%

Classifier Error Rate

Active Semi-Supervised Learning for Improving Word Alignment(Vamshi ACL ’10)

Real World Data

2

Page 3: Active Learning and Human-in-the-Loop

The Effect of Better Features

Unigrams Bigrams Unigrams+Bigrams0%

5%

10%

15%

20%

25%

30%

Classifier Error Rate

3

Page 4: Active Learning and Human-in-the-Loop

The Effect of More Data

Active Semi-Supervised Learning for Improving Word Alignment(Vamshi ACL ’10)

Real World Data

N 2N 4N0%

2%

4%

6%

8%

10%

12%

14%

Classifier Error Rate

4

Page 5: Active Learning and Human-in-the-Loop

The Effect of Cleaner Data

90% Accurate Data 95% Accurate Data 100% Accurate Data0%

2%

4%

6%

8%

10%

12%

14%

Classifier Error Rate

5

Page 6: Active Learning and Human-in-the-Loop

6

Where Do Data Scientists Spend Their Time?

Source: CrowdFlower Data Science Report 2015

Page 7: Active Learning and Human-in-the-Loop

7

CrowdFlower Data Enrichment Platform

Page 8: Active Learning and Human-in-the-Loop

8

Color Data

Page 9: Active Learning and Human-in-the-Loop

9

Page 10: Active Learning and Human-in-the-Loop

10

Page 11: Active Learning and Human-in-the-Loop

11

Page 12: Active Learning and Human-in-the-Loop

12

Page 13: Active Learning and Human-in-the-Loop

13

Page 14: Active Learning and Human-in-the-Loop

14

Page 15: Active Learning and Human-in-the-Loop

Apple Watch

15

Page 16: Active Learning and Human-in-the-Loop

Apple Watch

16

Page 17: Active Learning and Human-in-the-Loop

Apple Watch

17

Page 18: Active Learning and Human-in-the-Loop

Apple Watch

18

Page 19: Active Learning and Human-in-the-Loop

Collecting the Same Data Over and Over

19

Page 20: Active Learning and Human-in-the-Loop

20

Open Data

Page 21: Active Learning and Human-in-the-Loop

21

Make Your Data Public Setting

Page 22: Active Learning and Human-in-the-Loop

22

Data for Everyone

Page 23: Active Learning and Human-in-the-Loop

23

Data For Everyone Library

Page 24: Active Learning and Human-in-the-Loop

24

Data for Everyone

Page 25: Active Learning and Human-in-the-Loop

25

Data For Everyone

Page 26: Active Learning and Human-in-the-Loop

Open Data API

26

Page 27: Active Learning and Human-in-the-Loop

URL Categorization

27

Page 28: Active Learning and Human-in-the-Loop

Categorize URLs

28

Page 29: Active Learning and Human-in-the-Loop

Record Data

29

Page 30: Active Learning and Human-in-the-Loop

Extracting Names and Titles

30

Page 31: Active Learning and Human-in-the-Loop

Summarization

31

Page 32: Active Learning and Human-in-the-Loop

Is an Image Funny?

32

Page 33: Active Learning and Human-in-the-Loop

Classifying Medical Images

33

Page 34: Active Learning and Human-in-the-Loop

Attributes of People

34

Page 35: Active Learning and Human-in-the-Loop

35

Page 36: Active Learning and Human-in-the-Loop

Kaggle accuracy

Baseline 12-May 13-May 14-May 15-May0%

10%

20%

30%

40%

50%

60%

70%

Accuracy of Best Performing Model

Accu

racy

36

Page 37: Active Learning and Human-in-the-Loop

Kaggle accuracy over time

12-M

ay

14-M

ay

16-M

ay

18-M

ay

20-M

ay

22-M

ay

24-M

ay

26-M

ay

28-M

ay

30-M

ay1-J

un3-J

un5-J

un7-J

un9-J

un

11-Ju

n

13-Ju

n

15-Ju

n

17-Ju

n

19-Ju

n

21-Ju

n

23-Ju

n

25-Ju

n

27-Ju

n

29-Ju

n1-J

ul3-J

ul5-J

ul0%

10%

20%

30%

40%

50%

60%

70%

80%

Accuracy of the Best Performing Model

Accu

racy

37

Page 38: Active Learning and Human-in-the-Loop

Kaggle Participation

12-M

ay

14-M

ay

16-M

ay

18-M

ay

20-M

ay

22-M

ay

24-M

ay

26-M

ay

28-M

ay

30-M

ay1-J

un3-J

un5-J

un7-J

un9-J

un

11-Ju

n

13-Ju

n

15-Ju

n

17-Ju

n

19-Ju

n

21-Ju

n

23-Ju

n

25-Ju

n

27-Ju

n

29-Ju

n0

200

400

600

800

1000

1200

1400

Number of Participating Teams

38

Page 39: Active Learning and Human-in-the-Loop

AIClassifier OutputConfident

Human in the Loop

39

Page 40: Active Learning and Human-in-the-Loop

Human in the Loop

Confident OutputAIClassifier Not Confident Human

Annotation

40

Page 41: Active Learning and Human-in-the-Loop

Human in the Loop

Confident OutputAIClassifier

Active Learning

Not Confident Human Annotation

41

Page 42: Active Learning and Human-in-the-Loop

Human in the Loop

Confident OutputAIClassifier

Active Learning

Not Confident Human Annotation

42

Page 43: Active Learning and Human-in-the-Loop

Active Learning

From hunch.net active learning tutorial ICML ‘0943

Page 44: Active Learning and Human-in-the-Loop

Active Learning Accuracy Improvement

44

Page 45: Active Learning and Human-in-the-Loop

Google Cars Miles Per Disengage

45

Page 46: Active Learning and Human-in-the-Loop

Adaptive Cruise Control

Image source: ExtremeTech

46

Page 47: Active Learning and Human-in-the-Loop

Advanced Chess

Image source: Computer Chess

47

Page 48: Active Learning and Human-in-the-Loop

AlphaGo

48

Page 49: Active Learning and Human-in-the-Loop

Lukas [email protected]@L2K

Thank You