Active Learning - Carnegie Mellon School of Computer...
Transcript of Active Learning - Carnegie Mellon School of Computer...
![Page 1: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/1.jpg)
Active Learning
Yi Zhang10-701, Machine Learning, Spring 2011April 20th, 2011
Some pictures are from Burr and Aarti’s lectures
1
![Page 2: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/2.jpg)
Outline
Basic idea of active learning Supervised, semi-supervised, and active
learning Uncertainty sampling Version space reduction and query by
committee Expected error reduction Other active learning methods
2
![Page 3: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/3.jpg)
Why active learning?
Learning classifiers using labeled examples
But obtaining labels is◦ Time consuming, e.g., document classification◦ Expensive, e.g., medical decision (need doctors)◦ Sometimes dangerous, e.g., landmine detection
3
![Page 4: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/4.jpg)
Active learning
The learner actively chooses which examples to label !
Goal: reduce the number of labeled examples needed for learning
4
![Page 5: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/5.jpg)
Active learning
5
![Page 6: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/6.jpg)
Can active learning work?
Are all labeled examples equally important?
6
![Page 7: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/7.jpg)
Outline
Basic idea of active learning Supervised, semi-supervised, and
active learning Uncertainty sampling Version space reduction and query by
committee Expected error reduction Other active learning methods
7
![Page 8: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/8.jpg)
(Passive) supervised learning
8
![Page 9: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/9.jpg)
Semi-supervised learning
9
![Page 10: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/10.jpg)
Active learning
10
![Page 11: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/11.jpg)
Active learning vs. semi-supervised learning The same goal:◦ Attain good learning performance (e.g.,
classification accuracy) without demanding too many labeled examples
Different approaches◦ Semi-supervised learning: use unlabeled data◦ Active learning: choose labeled examples
11
![Page 12: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/12.jpg)
Outline
Basic idea of active learning Supervised, semi-supervised, and active
learning Uncertainty sampling Version space reduction and query by
committee Expected error reduction Other active learning methods
12
![Page 13: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/13.jpg)
Three active learning scenarios
Query synthesis◦ learner constructs examples for labeling
Selective sampling◦ Unlabeled data come as a stream◦ For each arrived point, learner decides to query
or discard
Pool-based active learning (*)◦ Given a pool of unlabeled data◦ Learner chooses from the pool for labeling
13
![Page 14: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/14.jpg)
Pool-based active learning
A pool of unlabeled samples A learned model (or a set of models) Choose: which sample to label next?
14
![Page 15: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/15.jpg)
Uncertainty sampling
Uncertainty sampling◦ Query the sample x that the learner is most
uncertain about◦ How to measure “uncertainty”?
15
![Page 16: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/16.jpg)
Uncertainty sampling
Uncertainty sampling◦ Maximum entropy
16
![Page 17: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/17.jpg)
Uncertainty sampling
Uncertainty sampling◦ Smallest margin (between most likely and
second most likely labels)
17
![Page 18: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/18.jpg)
Uncertainty sampling
Uncertainty sampling◦ Least confidence
18
![Page 19: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/19.jpg)
Uncertainty sampling
19
![Page 20: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/20.jpg)
Uncertainty sampling
20
![Page 21: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/21.jpg)
Outline
Basic idea of active learning Supervised, semi-supervised, and active
learning Uncertainty sampling Version space reduction and query by
committee Expected error reduction Other active learning methods
21
![Page 22: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/22.jpg)
Version space
The set of classifiers that are consistent with labeled examples
22
![Page 23: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/23.jpg)
Recall: uncertainty sampling
Uncertainty sampling
23
![Page 24: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/24.jpg)
Version space reduction
Version space reduction◦ Query the sample x that reduces the version most “Expected” reduction of version space “Worst case” reduction of version space “Best-case” reduction of version space
24
![Page 25: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/25.jpg)
A simplifying example
How to query? Binary search !
25
![Page 26: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/26.jpg)
A simplifying example Why binary search ?
◦ Recall: version space
◦ Maximize the “worst-case” reduction of version space
26
![Page 27: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/27.jpg)
Version space reduction for SVMs
27
![Page 28: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/28.jpg)
Query by committee
Query by committee◦ Keep a committee of classifiers◦ Query the instance that the committee members
disagree
28
![Page 29: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/29.jpg)
Query by committee
29
![Page 30: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/30.jpg)
Query by committee Query by committee◦ Keep a committee of classifiers◦ Query the instance that the committee members
disagree
QBC as version space reduction◦ Committee is an approximation to the version space
QBC as uncertainty sampling◦ Use committee members to measure the uncertainty
30
![Page 31: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/31.jpg)
Outline
Basic idea of active learning Supervised, semi-supervised, and active
learning Uncertainty sampling Version space reduction and query by
committee Expected error reduction Other active learning methods
31
![Page 32: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/32.jpg)
Expected error (uncertainty) reduction
32
![Page 33: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/33.jpg)
Outline
Basic idea of active learning Supervised, semi-supervised, and active
learning Uncertainty sampling Version space reduction and query by
committee Expected error reduction Other active learning methods
33
![Page 34: Active Learning - Carnegie Mellon School of Computer Sciencetom/10701_sp11/recitations/Recitation_13.pdf · Active Learning Yi Zhang 10-701, Machine Learning, Spring 2011 April 20th,](https://reader030.fdocuments.us/reader030/viewer/2022040712/5e1687ca7c4c0a27a755871c/html5/thumbnails/34.jpg)
Other active learning methods
Active learning is an active research area Cost sensitive active learning◦ Labeling costs of examples differ
Batch mode active learning◦ Query multiple instances at once
Multi-task active learning◦ Query labels for multiple learning tasks
See survey of Burr Settles ! [Settles, 2008]
34