ai for social good: the essential role of human machine partnership · 2018. 1. 4. · the...

AI for Social Good:

The Essential Role of Human Machine Partnership

MILIND TAMBE

Founding Co-director, Center for Artificial Intelligence in Society (CAIS)

University of Southern California

tambe@usc.edu

Co-Founder, Avata Intelligence

USC Center for Artificial Intelligence in Society (CAIS.USC.EDU)

7/20/2017 3

Mission Statement: Advancing AI research driven by…

Grand Challenges of Social Work

▪ Ensure healthy development for all youth

▪ Close the health gap

▪ Stop family violence

▪ Advance long and productive lives

▪ End homelessness

▪ Achieve equal opportunity and justice

…

7/20/2017 4

Overview of CAIS Project Areas

AI Assistants for Low Resource Communities

▪ Social networks: Spread HIV information, influence maximization

▪ Real-world pilot tests: Big improvements

7/20/2017 5

Overview of CAIS Project Areas

7/20/2017 6

▪ Machine learning/planning: Predicting poaching spots, patrols

▪ Real-world: Uganda, South Asia…

AI Assistants for Conservation

▪ Game theory: security resource optimization

▪ Real-world: US Coast Guard, US Federal Air Marshals Service…

Overview of CAIS Project Areas

7/20/2017 7

AI Assistants for Public Safety and Security

Lesson Learned:Essential Nature of Human-Machine Partnership

▪ Building decision aids/assistants: Humans focus on their expertise

7/20/2017 8

▪ Low Resource Communities

▪ Public Safety and Security

▪ Wildlife Conservation

▪ Partnership at individual and organization level:

AI Assistant: HEALER

7/20/2017 9

AI Assistant for HIV Prevention:Using Social Networks to Spread HIV Information

7/20/2017 10

Challenge: Adaptive selection in Uncertain Network

K = 5

1st time step

7/20/2017 11

127/20/2017

137/20/2017

Challenge: Adaptive selection in Uncertain Network

K = 5

2nd time step

7/20/2017 14

Challenge 3 : Adaptive selection

K = 5

3rd time step

7/20/2017 15

NO LONGER A SINGLE SHOT

DECISION PROBLEM

7/20/2017 16

Creating an Adaptive Policy: “POMDP” [2015]

▪ Homeless shelters – sequentially select nodes under uncertainty

➢ Policy driven by observations about edges

Action

Choose youth for

intervention

Observation: Which edges exist?

Adaptive Policy

POMDP POLICY

Generation

Actual social

Network in the

real world:

Real Networks – Simulation Results [2016-2017]

7/20/2017 17

0

10

20

30

40

50

60

70

80

90

Venice Hollywood

Ind

irec

t In

flu

ence

Degree CR PSINET HEALER-1 HEALER-2

Pilot Tests

with 170 Homeless Youth [2017]

Recruited youths:

Preliminary network —> HEALER

Bring 4 youth for training, get edge data —> HEALER

Bring 4 youth for training, get edge data —> HEALER

Bring 4 youth for training

HEALER HEALER++ DEGREE CENTRALITY

62 56 55

7/20/2017 18

Results: Pilot Studies

7/20/2017 19

0

20

40

60

80

100

HEALER HEALER++ Degree

Percent of non-Peer Leaders

Informed Not Informed

0

20

40

60

80

100

HEALER HEALER++ Degree

Informed Non-Peer Leaders Who Started Testing for HIV

Testing Non-Testing

Analysis: Pilot Studies

7/20/2017 20

0

5

10

15

20

25

HEALER HEALER++ Degree

% of edges between peer leaders

0

20

40

60

80

100

120

HEALER HEALER++ Degree

% Coverage of communities in 1st stage

AI Assistant: HEALER

7/20/2017 21

Next Steps: AI Assistant for HIV Prevention

▪ 900 youth study begun at three locations in Los Angeles

➢ 300 enrolled in HEALER/HEALER++

➢ 300 enrolled in no condition

➢ 300 in Degree centrality

7/20/2017 22

“Picking youth as peer leaders was

changing their self esteem and the

sense of confidence….”

Eric Rice

Overview of CAIS Project Areas

AI Assistants for Low Resource Communities

▪ Substance abuse, suicide prevention…

▪ Modeling gang violence, matching homeless and homes…

7/20/2017 23

AI Assistant for Protecting Wildlife in Uganda

7/20/2017 24

Data from Queen Elizabeth National Park, Uganda

AI Assistant for Poacher behavior prediction

Number of poaching attacks over 12 years: ~1000

How likely is an

attack on

a grid Square

Ranger patrol

frequency

Animal density

Distance to

rivers / roads

Area habitat

Area slope

…

7/20/2017 25

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

L&L

Uniform Random SVM CAPTURE Decision Tree Our Best Model

Poacher Behavior Prediction

Poacher Attack Prediction

Results from 2015

7/20/2017 26

7/20/2017 27

Real-world Deployment: (1 month)

Real-world Deployment: Results

▪ Two 9 sq KM patrol areas: Predicted hot spots with infrequent patrols

▪ Poached Animals: Poached elephant

▪ Snaring: 1 elephant snare roll

▪ Snaring: 10 Antelope snares

Historical Base Hit

RateOur Hit Rate

Average: 0.73 3

7/20/2017 28

7/20/2017 29

0

0.02

0.04

0.06

0.08

0.1

0.12

High (1) Low (2)

Cat

ch p

er

Un

it E

ffo

rt

Experiment Group

Real-world Deployment: Field Test 2 (6 months)

▪ Catch Per Unit Effort (CPUE)

➢Unit Effort = km walked

➢Our high CPUE: 0.11

➢Our low CPUE: 0.01

Historical CPUE: 0.04

7/20/2017 30

Field Test Side Effects:Queen Elizabeth National Park

▪ Rangers followed poachers’ trail; ambushed camp

➢ Arrested one (of 7) poachers

➢ Confiscated 10 wire snares, cooking pot, hippo

meat, timber harvesting tools.

▪ Pursuit of poachers

▪ Signs of road building, fires, illegal fishing

Building AI Assistants:Human Machine Partnership in “AI for Social Good”

▪ Humans focus on their expertise, AI assistants on theirs:

➢ E.g., social workers interact with youth, AI assistants on selecting youth

▪ Lessons in Building Assistants:

➢ Right task division for humans vs machines

➢ Right adjustable autonomy; human may be wrongly biased

➢ Explanation of output

▪ Implementation requires organization level partnership:

➢ Immersion opens up our eyes to complexity; builds up trust over time

➢ Need a champion on the inside

7/20/2017 31

AI for Social Good

7/20/2017 32

THANK YOU

CAIS.USC.EDU