dynamically switching between synergistic workflows for crowdsourcing christopher h. lin, mausam,...

53
Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto, ON, Canada 1

Upload: emil-bruce

Post on 24-Dec-2015

223 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

1

Dynamically Switching between Synergistic Workflows for Crowdsourcing

Christopher H. Lin, Mausam, Daniel S. Weld

University of Washington

AAAI-2012, Toronto, ON, Canada

Page 2: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

2

Page 3: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

3

Translate AB

Page 4: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

4

Translate AB

HumanTrAB

ChooseHumanTr

AB

Page 5: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

5

Translate AB

MachineTrAB ProofRd

Page 6: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

6

Translate AB

TranslateAC

TranslateCB

Page 7: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

7

www.crowdsourcing.com

Is this a face?

Yes No

Page 8: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

8

Congressman Tony Coelho landed a million - dollar job on Wall Street after resigning over a controversial junk - bond investment last summer.

NER Tagging

Page 9: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

9

Congressman Tony Coelho landed a million - dollar job on Wall Street after resigning over a controversial junk - bond investment last summer.

NER Tagging

location

Page 10: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

10

www.crowdsourcing.com

location

film

Congressman Tony Coelho landed a million - dollar job on Wall Street after resigning over a controversial junk - bond investment last summer.

Page 11: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

11

www.crowdsourcing.com

Wall Streethttp://en.wikipedia.org/wiki/Wall_StreetWall Street is the financial district of New York City, named after and centered on the eight-block-long street running from Broadway to South Street.…

Wall Street (1987 film)http://en.wikipedia.org/wiki/Wall_Street _(1987_film)Wall Street is a 1987 American drama film released by 20th Century Fox. It was directed by Oliver Stone and stars Michael Douglas, Charlie Sheen, and Daryl Hannah…

Congressman Tony Coelho landed a million - dollar job on Wall Street after resigning over a controversial junk - bond investment last summer.

Page 12: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

12

70%

90%

A/B Testing Accuracy

Page 13: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

13

70%

90%

A/B Testing Accuracy

Page 14: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

14

www.crowdsourcing.com

us_countylocationcitytown

location

Washington StateWashington D.C

Page 15: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

15

www.crowdsourcing.com

us_countylocationcitytown

location

Washington StateWashington D.C

Page 16: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

16

www.crowdsourcing.com

us_countylocationcitytown

location

Washington StateWashington D.C

Page 17: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

17

N Create job using best workflow

[Dai et al, 2010]

Page 18: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

18

N Create job using best workflow

[Dai et al, 2010]

SUBOPTIMAL

Page 19: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

19

N Create job using best workflow

Page 20: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

20

www.crowdsourcing.com

us_countylocationcitytown

location

Washington StateWashington D.C

Page 21: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

21

www.crowdsourcing.com

Washingtonhttp://en.wikipedia.org/wiki/WashingtonWashington, D.C., formally the District of Columbia and commonly referred to as Washington, “the District,” or simply D.C., is the capital of the United States….

Washington(state)http://en.wikipedia.org/wiki/Washington_(state)Washington () is a state in the Pacific Northwest region of the United States located north of Oregon, west of Idaho, and south of the Canadian province of British Columbia, on the coast of the Pacific Ocean….

Washington StateWashington D.C

Page 22: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

22

www.crowdsourcing.com

Washingtonhttp://en.wikipedia.org/wiki/WashingtonWashington, D.C., formally the District of Columbia and commonly referred to as Washington, “the District,” or simply D.C., is the capital of the United States….

Washington(state)http://en.wikipedia.org/wiki/Washington_(state)Washington () is a state in the Pacific Northwest region of the United States located north of Oregon, west of Idaho, and south of the Canadian province of British Columbia, on the coast of the Pacific Ocean….

Washington StateWashington D.C

Page 23: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

23

www.crowdsourcing.com

Washingtonhttp://en.wikipedia.org/wiki/WashingtonWashington, D.C., formally the District of Columbia and commonly referred to as Washington, “the District,” or simply D.C., is the capital of the United States….

Washington(state)http://en.wikipedia.org/wiki/Washington_(state)Washington () is a state in the Pacific Northwest region of the United States located north of Oregon, west of Idaho, and south of the Canadian province of British Columbia, on the coast of the Pacific Ocean….

Washington StateWashington D.C

Page 24: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

24

N Create job using best workflow

[Dai et al, 2010]

Before

Page 25: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

25

AGENTHUNT

Page 26: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

26

AGENTHUNT

Dynamically

Page 27: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

27

POMDPStates

Actions

Transition Function

Observation Probabilities

Page 28: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

28

POMDPStates

Actions

Transition Function

Observation Probabilities

Page 29: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

29

Page 30: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

30

2 + 2 = ?

www.crowdsourcing.com

Page 31: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

31

2 + 2 = ?

www.crowdsourcing.com

EASY

d = 0

Page 32: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

32

www.crowdsourcing.com

?=

Page 33: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

33

www.crowdsourcing.com

?=HA

RD

d = 1

Page 34: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

34

Correct Answer

Workflow Difficulties

Page 35: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

35

POMDPStates

Actions

Transition Function

Observation Probabilities

Page 36: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

36

K+1 actions

Page 37: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

37

POMDPStates

Actions

Transition Function

Observation Probabilities

Page 38: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

38

POMDPStates

Actions

Transition Function

Observation Probabilities

Page 39: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

39

Page 40: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

40

Difficulty

Page 41: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

41

Worker Ability

Difficulty

Page 42: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

42Source:Wikipedia

Page 43: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

43[Figure courtesy Peng Dai]

4.0

2.0

1.0

0.5

0.25

Page 44: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

44

Page 45: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

45

LEARN NEW PARAMETERS

Page 46: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

46

[Dai et al, 2010]

AgentHunt beats SingleBestWorkflow

Page 47: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

47

[Dai et al, 2010]

AgentHunt beats SingleBestWorkflow

Page 48: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

48

AgentHunt SingleBestWorkflow

AverageAccuracy (%)

92.45 85.85

AverageNet Utility

-13.36 -18.35

AgentHunt beats SingleBestWorkflow

Page 49: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

49

Contributions

• A probabilistic model for multiple workflows

• A decision-theoretic agent that dynamically switches between workflows

• Real experiments that show you should not throw away your workflows

Page 50: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

50

INTERESTED?

http://cs.washington.edu/homes/chrislin

Page 51: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

51

Repeated Labeling

Romney et al, 1986; Sheng et al, 2008; Raykar et al, 2010; Whitehill et al, 2009; Dai et al, 2010, 2011; Kamar et al, 2012; Parameswaran et al, 2010; Karger et al, 2011; Snow et al, 2008;

Complicated Workflows

Little et al, 2009; Bernstein et al, 2010; Liem et al, 2011; Noronha et al, 2011; Lasecki et al, 2011; Kulkarni et al, 2012;

AI Control of Workflows

Dai et al, 2010,2011; Shahaf et al, 2010; Weld et al, 2011;Kamar et al, 2012;

Page 52: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

52

How many people are there in this military base?

www.crowdsourcing.com

Source:Wikipedia

Page 53: Dynamically Switching between Synergistic Workflows for Crowdsourcing Christopher H. Lin, Mausam, Daniel S. Weld University of Washington AAAI-2012, Toronto,

53

Christopher H. Lin, Mausam, and Daniel S. Weld. “Crowdsourcing Control: Moving Beyond Multiple Choice.”

In UAI 2012