Human-Centered Artificial Intelligence

Transcription

1 Human-Centered Artificial Intelligence Mark

2 Alien intelligences 2

3 Alien intelligences Artificial intelligences are inscrutable to most humans 2

4 Alien intelligences Artificial intelligences are inscrutable to most humans Humans are inscrutable to artificial intelligences 2

5 Human-centered artificial intelligence 3

6 Human-centered artificial intelligence Understanding humans 3

7 Human-centered artificial intelligence Understanding humans Helping humans understand them 3

8 Human-centered artificial intelligence Computational creativity Understanding humans Helping humans understand them 3

9 Human-centered artificial intelligence Understanding humans Helping humans understand them 3

10 Challenges & opportunities Understanding humans Helping humans understand them 3

11 Challenges & opportunities Understanding humans 3

12 Specifying goals 4

13 Specifying goals 4

14

15

16 Commonsense goal failure Do what I want 6

17 Commonsense goal failure Do what I want the way I would do it! 6

18 Commonsense goal failure Do what I want the way I would do it! Knowledge bases? Lots of sensors? Demonstration? 6

19 Learning from stories If computers could comprehend stories then humans can transfer commonsense procedural knowledge to computers by telling stories 7

20 Machine enculturation Human cultural values are implicitly encoded in stories told by members of a culture Allegorical tales Fables Contemporary fictional literature, TV, & movies 8 Riedl. CHI Workshop on Human-Centered Machine Learning, 2016.

21 Natural language Natural language processing is not a solved problem Humans are noisy (variable) Humans shouldn t need to know autonomous system capabilities or execution environment 9

22 Quixote Reinforcement learning: AI devises a program for operating in an environment through trial and error Intuition: Reward the agent for performing actions that mimic the stories that it has been told 10 Harrison & Riedl. AIIDE Conference, 2016.

23 Quixote Exemplar stories Model learning A model Trajectory tree creation A trajectory tree Reinforcement learning Reward assignment A policy mapping states to actions A trajectory tree with events assigned reward values 11 Environment

24 Quixote Exemplar stories Model learning A model Trajectory tree creation A trajectory tree Reinforcement learning Reward assignment A policy mapping states to actions A trajectory tree with events assigned reward values 11 Environment

25 choose restaurant Fast food restaurant drive to restaurant walk/go into restaurant read menu drive to drive-thru wait in line choose menu item take out wallet place order pay for food drive to window wait for food get food find table sit down eat food clear trash leave restaurant drive home 12

26 arrive at theatre go to ticket booth wait for ticket choose movie buy tickets go to concession stand order popcorn / soda show tickets buy popcorn enter theatre turn off cellphone find seats sit down Going on a date to the movies eat popcorn watch movie use bathroom talk about movie discard trash hold hands kiss leave movie 13 drive home

27 Quixote Exemplar stories Model learning A model Trajectory tree creation A trajectory tree Reinforcement learning Reward assignment A policy mapping states to actions A trajectory tree with events assigned reward values 14 Environment

28 Reinforcement learning Fill gaps between events Leave House Go to bank Go to hospital Go to doctor Withdraw money Get prescription hospital Get prescription doctor Don't get prescription hospital Don't get prescription doctor Go to pharmacy Buy strong drugs Buy weak drugs Go home World state space 15 Harrison & Riedl. AIIDE Conference, 2016.

29 Reinforcement learning Fill gaps between events Leave House Go to bank Go to hospital Go to doctor Withdraw money Get prescription hospital Get prescription doctor Don't get prescription hospital Don't get prescription doctor Go to pharmacy Buy strong drugs Buy weak drugs Go home World state space 15 Harrison & Riedl. AIIDE Conference, 2016.

30 Reinforcement learning Fill gaps between events Leave House leave house Go to bank Go to hospital Go to doctor Withdraw money Get prescription hospital Get prescription doctor Don't get prescription hospital Don't get prescription doctor Go to pharmacy Buy strong drugs Buy weak drugs Go home World state space 15 Harrison & Riedl. AIIDE Conference, 2016.

31 Reinforcement learning Fill gaps between events Leave House Go to bank Go to hospital Go to doctor leave house go doctor Withdraw money Get prescription hospital Get prescription doctor Don't get prescription hospital Don't get prescription doctor Go to pharmacy go bank Buy strong drugs Go home Buy weak drugs go hospital World state space 15 Harrison & Riedl. AIIDE Conference, 2016.

32 Reinforcement learning Fill gaps between events Leave House Go to bank Go to hospital Go to doctor leave house go doctor Withdraw money Get prescription hospital Get prescription doctor Don't get prescription hospital Don't get prescription doctor Stairs Go to pharmacy Buy strong drugs Buy weak drugs Go home Drive Main St. go hospital go bank World state space 15 Harrison & Riedl. AIIDE Conference, 2016.

33

34

35 Machine enculturation Social conventions prevent conflict Robots that follow the rules of society will be safer 17 Riedl. CHI Workshop on Human-Centered Machine Learning, 2016.

36 Challenges & opportunities Understanding humans Helping humans understand them 18

37 Challenges & opportunities Helping humans understand them 18

38 Autonomous system failures 19

39 Possible solution: open the black box

40 AI rationalization 21 Harrison, Ehsan, Riedl. arxiv , 2017.

41 AI rationalization Creating an explanation comparable to what a human would say if he or she were controlling the robot in the same situation 21 Harrison, Ehsan, Riedl. arxiv , 2017.

42 AI rationalization Creating an explanation comparable to what a human would say if he or she were controlling the robot in the same situation Takes inspiration from what humans do 21 Harrison, Ehsan, Riedl. arxiv , 2017.

43 AI rationalization Creating an explanation comparable to what a human would say if he or she were controlling the robot in the same situation Takes inspiration from what humans do Human understandable 21 Harrison, Ehsan, Riedl. arxiv , 2017.

44 AI rationalization Creating an explanation comparable to what a human would say if he or she were controlling the robot in the same situation Takes inspiration from what humans do Human understandable Helps build trust; useful in time-critical situations 21 Harrison, Ehsan, Riedl. arxiv , 2017.

45

46

47 Neural machine translation 23 Harrison, Ehsan, Riedl. arxiv , 2017.

48 Neural machine translation Woah! Car beside me and a gap above. Fortune favors the brave. 23 Harrison, Ehsan, Riedl. arxiv , 2017.

49 Neural machine translation 24 Harrison, Ehsan, Riedl. arxiv , 2017.

50 Neural machine translation Woah! Car beside me and a gap above. Fortune favors the brave Harrison, Ehsan, Riedl. arxiv , 2017.

51 AI Rationalization 25

52 AI Rationalization Target users are those without technical backgrounds 25

53 AI Rationalization Target users are those without technical backgrounds Meant to convey fast, approximate explanations 25

54 AI Rationalization Target users are those without technical backgrounds Meant to convey fast, approximate explanations Meant to foster rapport and trust 25

55 AI Rationalization Target users are those without technical backgrounds Meant to convey fast, approximate explanations Meant to foster rapport and trust Coupled with more thorough explanations & visualizations Work by Alex Endert, Georgia Tech 25

56 Challenges & opportunities Understanding humans Helping humans understand them 26

57 Challenges & opportunities Understanding humans Helping humans understand them 26

58 Understanding helps AI Woah! Car beside me and a gap above. Fortune favors the brave

59 Understanding helps AI Woah! Car beside me and a gap above. Fortune favors the brave. 27

60 Punchline Average reward Training iterations (x100) 28

61 Punchline Average reward Standard Q-learning Training iterations (x100) 28

62 Punchline Average reward Learning from demonstration Standard Q-learning Training iterations (x100) 28

63 Punchline Language-based guidance Average reward Learning from demonstration Standard Q-learning Training iterations (x100) 28

64 Human-centered artificial intelligence Computational creativity Understanding humans Helping humans understand them 29

65 Human-centered artificial intelligence Computational creativity 29

66 Computational creativity 30

67 Computational creativity 30

68 Computational creativity 30

69 Computational creativity 30

70 Computational creativity 31

71 Computational creativity Most computational creativity is learning a pattern from data and trying to make new inputs fit the pattern 31

72 Computational creativity Most computational creativity is learning a pattern from data and trying to make new inputs fit the pattern AI can t reach human-level creativity without making intuitive leaps 31

73 Computational creativity Most computational creativity is learning a pattern from data and trying to make new inputs fit the pattern AI can t reach human-level creativity without making intuitive leaps AI can t augment human creativity if AI can t keep up with human collaborator s intuitive leaps 31

74 Computational creativity Most computational creativity is learning a pattern from data and trying to make new inputs fit the pattern AI can t reach human-level creativity without making intuitive leaps AI can t augment human creativity if AI can t keep up with human collaborator s intuitive leaps Computational creativity is about making AI gracefully handle novel situations it was never trained for 31

75 + =? 32

76 Concluding thoughts AI appears less alien Maybe safer? Computational creativity to handle contingencies very different from input Human-centered AI is an essential mix of capabilities for robots in the human world 33