114
Robots Learning Like Babies Andrea Censi Laboratory for Information and Decision Systems Massachusetts Institute of Technology http://censi.mit.edu - [email protected] 時代基會暨省理學院MIT全球產研計劃台灣年會 2015 MIT ILP-EPOCH TAIWAN SYMPOSIUM Taipei, July 28, 2015 Meet the Future of Robotics and Machine Learning in Innovation Economy

Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Robots Learning Like Babies

Andrea Censi

Laboratory for Information and Decision SystemsMassachusetts Institute of Technology

http://censi.mit.edu - [email protected]

時代基⾦金會暨⿇麻省理⼯工學院MIT全球產研計劃台灣年會 2015 MIT ILP-EPOCH TAIWAN SYMPOSIUM

Taipei, July 28, 2015Meet the Future of Robotics and Machine

Learning in Innovation Economy

Page 2: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Outline

1. What is learning?

2. What is the value proposition for robotics?

3. What are the challenges in designing learning systems?

2

Page 3: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

What is learning?

‣ It’s about using data- …to make predictions- …to make decisions

3

Page 4: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

4

Page 5: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix

5

Page 6: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix

5

Page 7: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix

5

Page 8: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix

5

Page 9: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix suggests you a good movie to watch.

6

Page 10: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix suggests you a good movie to watch.

6

‣ Example of a “recommender” system- other classical example: shopping suggestions

Page 11: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Netflix suggests you a good movie to watch.

6

‣ Example of a “recommender” system- other classical example: shopping suggestions

‣ Netflix Challenge (2007-9): $1M for help in making the learning algorithm better.

Page 12: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

7

algorithminput output

Page 13: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

8

learning algorithm

dataa decision rule

Page 14: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

9

learning algorithm

data

Page 15: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

9

learning algorithm

dataJohn, “Godfather”,

Jane, “Godfather”,

Jane, “Terminator”,

...

Page 16: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

9

learning algorithm

dataJohn, “Godfather”,

Jane, “Godfather”,

Jane, “Terminator”,

...

you

rating?

a movie

Page 17: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

data

Page 18: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

parameters (high dimensional)

data

Page 19: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 20: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 21: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 22: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 23: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 24: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 25: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 26: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

parameters (high dimensional)

data

Page 27: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data

parameters (high dimensional)

data

Page 28: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10

parameters (high dimensional)

data

Page 29: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15

parameters (high dimensional)

data

Page 30: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.30

parameters (high dimensional)

data

Page 31: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.010.30

parameters (high dimensional)

data

Page 32: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.010.30 0.50

parameters (high dimensional)

data

Page 33: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.250.010.30 0.50

parameters (high dimensional)

data

Page 34: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.250.250.010.30 0.50

parameters (high dimensional)

data

Page 35: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.250.250.010.30 0.50

parameters (high dimensional)

data

Page 36: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.250.250.010.30 0.50

parameters (high dimensional)

data best* decision rule

Page 37: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.250.250.010.30 0.50

parameters (high dimensional)

data

* in the family considered

best* decision rule

Page 38: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

10

learning algorithm

candidates decision rules

error on test data 0.10 0.15 0.250.250.010.30 0.50

parameters (high dimensional)

data

* on the data available* in the family considered

best* decision rule

Page 39: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

11

Page 40: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

12

‣ Nest Learning Thermostat

Page 41: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Nest Learning Thermostat - acquired by Google ($2B)- see also: Ecobee in the Apple ecosystem

13

Page 42: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Nest Learning Thermostat - acquired by Google ($2B)- see also: Ecobee in the Apple ecosystem

13

~ NT 300

Page 43: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Nest Learning Thermostat - acquired by Google ($2B)- see also: Ecobee in the Apple ecosystem

13

+ learning =

~ NT 300

Page 44: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Nest Learning Thermostat - acquired by Google ($2B)- see also: Ecobee in the Apple ecosystem

13

+ learning =

~ NT 300 NT 8000

Page 45: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

14

March 2015 in Boston

Page 46: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

14

March 2015 in Boston

Boston, March 2015

Page 47: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

14

March 2015 in Boston

Boston, March 2015

“snow”fun only on the first day

Page 48: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

14

March 2015 in Boston

Boston, March 2015

“snow”fun only on the first day

Page 49: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

15

Page 50: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

16

Page 51: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

17

learning algorithm

data

Page 52: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

17

learning algorithm

datasensor data time series

Page 53: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

17

learning algorithm

data

“I’m cold”, “I’m hot”sensor data time series

Page 54: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

17

learning algorithm

data

turn on/off

temperature date

“I’m cold”, “I’m hot”sensor data time series

Page 55: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

18

learning algorithm

data

a control policy

turn on/off

temperature date

“I’m cold”, “I’m hot”sensor data time series

Page 56: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

18

learning algorithm

data

a control policy

turn on/off

temperature change

a model of the system

turn on/off

temperature date

“I’m cold”, “I’m hot”sensor data time series

Page 57: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

18

learning algorithm

data

a control policy

date

desired temperature

a model of the task

turn on/off

temperature change

a model of the system

turn on/off

temperature date

“I’m cold”, “I’m hot”sensor data time series

Page 58: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

19

Page 59: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

‣ What is a robot?

19

Page 60: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

‣ What is a robot?- has sensors

19

Page 61: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

‣ What is a robot?- has sensors- has actuators

19

Page 62: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

19

Page 63: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

20

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

‣ QUIZ: Which one is the robot?

Page 64: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

20

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

A

‣ QUIZ: Which one is the robot?

Page 65: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

20

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

A B

‣ QUIZ: Which one is the robot?

Page 66: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

20

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

A B C

‣ QUIZ: Which one is the robot?

Page 67: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

20

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

A B C

‣ QUIZ: Which one is the robot?

D

Page 68: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

20

‣ What is a robot?- has sensors- has actuators- interacts with the physical world

A B C E

‣ QUIZ: Which one is the robot?

D

Page 69: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

21

learning algorithm

data

Page 70: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

21

learning algorithm

datasensorimotor experience

Page 71: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

21

learning algorithm

data

task examples (given by user )

sensorimotor experience

Page 72: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

21

learning algorithm

data

a control policy

command

observations desired state

task examples (given by user )

sensorimotor experience

Page 73: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

21

learning algorithm

data

commands

change in state

a model of the system

a control policy

command

observations desired state

task examples (given by user )

sensorimotor experience

Page 74: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

21

learning algorithm

data

observations

desired state

a model of the task

commands

change in state

a model of the system

a control policy

command

observations desired state

task examples (given by user )

sensorimotor experience

Page 75: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics)

22

Page 76: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics)

22

Page 77: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics)

22

Page 78: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics) - Learning allows high performance

with very cheap components

23

Page 79: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics) - Learning allows high performance

with very cheap components

23

learning algorithm

data

Page 80: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics) - Learning allows high performance

with very cheap components

23

recorded commands

recorded observationslearning

algorithm

data

Page 81: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics) - Learning allows high performance

with very cheap components

23

recorded commands

recorded observationslearning

algorithm

datacommands

change in state

a model of the system

Page 82: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Kiva Systems (now AmazonRobotics) - Learning allows high performance

with very cheap components

24

recorded commands

recorded observationslearning

algorithm

data

a model of the system

Page 83: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

‣ Learning makes systems more robust.

25

interferencesabotage faults / wear

Page 84: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

‣ Learning makes systems more robust.

25

interferencesabotage faults / wear

Page 85: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Dyson 360eye- learns a 3D map of your house

26

Page 86: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ Dyson 360eye- learns a 3D map of your house

26

Page 87: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ “Learning by demonstration”

27

Page 88: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ “Learning by demonstration”

27

[Muelling, Peters]

Page 89: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ “Learning by demonstration”

27

[Muelling, Peters]

learning algorithm

data

Page 90: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ “Learning by demonstration”

27

[Muelling, Peters]

user demonstrations

robot’s own experience

learning algorithm

data

Page 91: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Examples of learning

‣ “Learning by demonstration”

27

[Muelling, Peters]

user demonstrations

robot’s own experience

learning algorithm

data

control policy

commands

current state

Page 92: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Learning

‣ It’s about using data- …to make predictions- …to make decisions

28

Page 93: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Learning

‣ It’s about using data- …to make predictions- …to make decisions

28

‣ Value proposition for robotics:- add value: more complex functionality- reduce design cost- reduce building cost- reduce operating cost

Page 94: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Learning

‣ It’s about using data- …to make predictions- …to make decisions

28

‣ Value proposition for robotics:- add value: more complex functionality- reduce design cost- reduce building cost- reduce operating cost

but…

Page 95: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

29

Page 96: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

30

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

Page 97: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

30

more adaptive less adaptive

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

Page 98: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

30

more adaptive less adaptive

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

Human

Page 99: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

30

more adaptive less adaptive

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

Fruit flyHuman

Page 100: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

30

Dyson 360eye

more adaptive less adaptive

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

Fruit flyHuman

Page 101: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

30

Dyson 360eye

more adaptive less adaptive

Downsides of Learning

‣ Adaptivity comes at a (computational) cost.

Fruit flyHuman iRobot’s Roomba

Page 102: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ It might be harder to understand your design.

31

Page 103: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ It might be harder to understand your design.

31

‣ Just like “genetic algorithms”…

Page 104: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ It might be harder to understand your design.

31

2006 NASA ST5 spacecraft antenna.

‣ Just like “genetic algorithms”…

Page 105: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ It might be harder to understand your design.

32

Page 106: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ It might be harder to understand your design.

32

Page 107: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

33

Downsides of Learning

(image, label)(image, label)(image, label)(image, label)(image, “gorilla”)(image, label)

learning algorithm

data

‣ It might be harder to understand your design.

Page 108: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

33

Downsides of Learning

(image, label)(image, label)(image, label)(image, label)(image, “gorilla”)(image, label)

learning algorithm

data

‣ It might be harder to understand your design.

image of person

“gorilla”

Page 109: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

33

Downsides of Learning

(image, label)(image, label)(image, label)(image, label)(image, “gorilla”)(image, label)

learning algorithm

data

‣ It might be harder to understand your design.

image of person

“gorilla”

evidence: + eyes + hair - clothes

Page 110: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ Machine learning systems are hard to maintain.

34

Page 111: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Downsides of Learning

‣ Machine learning systems are hard to maintain.

34

Page 112: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Learning

‣ It’s about using data- …to make predictions- …to make decisions

35

‣ Value proposition for robotics- add value: more complex functionality- reduce design cost- reduce building cost- reduce operating cost

Page 113: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Learning

‣ It’s about using data- …to make predictions- …to make decisions

35

‣ Value proposition for robotics- add value: more complex functionality- reduce design cost- reduce building cost- reduce operating cost

‣ But:- systems need more computational resources- systems are hard to trust- systems are less maintainable

Page 114: Robots Learning Like Babies - Andrea Censi · ‣ Netflix suggests you a good movie to watch. 6 ‣ Example of a “recommender” system-other classical example: shopping suggestions

Learning

‣ It’s about using data- …to make predictions- …to make decisions

35

‣ Value proposition for robotics- add value: more complex functionality- reduce design cost- reduce building cost- reduce operating cost

‣ But:- systems need more computational resources- systems are hard to trust- systems are less maintainable

‣ Learn about learning: “The Master Algorithm” by Domingos (Sep’15)