The Role of Dialog in Augmented Intelligence

17 views

Published on

Murray Campbell, IBM Research, presention "The Role of Dialog in Augmented Intelligence" as part of the Cognitive Systems Institute Speaker Series

Published in: Technology
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
17
On SlideShare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

The Role of Dialog in Augmented Intelligence

  1. 1. The Role of Dialog in Augmented Intelligence Murray Campbell Distinguished Research Staff Member IBM T.J. Watson Research Center
  2. 2. Outline • Augmented Intelligence? • Augmented Intelligence + Communication • Communication = Dialog (sometimes) • Augmented Intelligence + Dialog = Interesting Research Challenges © 2017 International Business Machines Corporation
  3. 3. Augmented Intelligence • Humans + Computers can achieve better performance than either alone • When might this be true? – They are complementary • Gaps in computer capabilities are covered by human capabilities • Gaps in human capabilities are covered by computer capabilities – The benefits of collaboration outweigh the costs of communication © 2017 International Business Machines Corporation Computer Capabilities Gaps: Creativity, goal setting, commonsense reasoning, rare cases, etc. Human Capabilities Gaps: Memory, attention, large scale search, stamina, etc. Augmented Intelligence Gaps are covered
  4. 4. Augmented Intelligence Examples • Aircraft autopilot • Legal discovery • Healthcare – IBM Watson and University of North Carolina Cancer Center study – 1000 cancer patients • 99% match on recommended treatment • 30% had valid alternative treatments missed by physicians (current clinical trial, recently approved, etc.) • Computer chess – The gap is narrowing – But it is taking a long time © 2017 International Business Machines Corporation
  5. 5. When can dialog be useful for Augmented Intelligence? • Sufficient time to make a decision • Sufficient complexity to justify an interaction • Sufficient importance to justify an interaction • Explainability is important • Complex goals/preferences of human agent • Gaps and strengths of agents are at least somewhat understood • Non-expert human users • Hands free • To build trust © 2017 International Business Machines Corporation
  6. 6. Dialog Challenges • Scripted approaches to dialog are not easily scaled – Learning approaches (Deep Learning, (Deep) RL, Adversarial Learning) are emerging • Evaluation/measurement/annotation • Detecting and expressing emotion • Multi-agent • Scalable interaction with domain knowledge • Reasoning © 2017 International Business Machines Corporation
  7. 7. Dialog Challenges • Scripted approaches to dialog are not easily scaled – Learning approaches (Deep Learning, (Deep) RL, Adversarial Learning) are emerging • Evaluation/measurement/annotation • Detecting and expressing emotion • Multi-agent • Scalable interaction with domain knowledge • Reasoning © 2017 International Business Machines Corporation Customer: I bought the wingtip shoes in a size 9 but they’re too small, can I exchange for another size? Agent: Sure, the only size we have left in stock is 8.5, will that be ok? Customer: I want to add my wife to our insurance policy. Agent: Ok, I can help you with that. Is she an immediate family member?
  8. 8. Learning Simple Reasoning for Dialog © 2017 International Business Machines Corporation 1. Emma is in the garden. 2. Hannah is in the garden. 3. Hannah picks up the gift. 4. Hannah goes to the kitchen Question: Where is the gift? Agent: The gift is in the kitchen. 1. Emma is in the garden. 2. Hannah is in the garden. 3. Hannah picks up the gift. 4. Someone goes to the kitchen. Question: Where is the gift? Agent: In statement 4, who went to the kitchen? Response: Emma went to the kitchen. Agent: Then the gift is in the garden. bAbI Task (Facebook) Blabby task • Given examples of short stories with ambiguities, along with question/answers (toy domain) • Learns to interact, reason and answer questions about stories it has never seen before • Can answer almost perfectly on easier (5-6 sentence) stories • Can answer more than 80% correct on harder 15-20 sentence stories openreview.net/pdf?id=rJ0-tY5xe
  9. 9. Conclusion • Augmented Intelligence systems are likely to be important for the foreseeable future • In some cases, dialog may be the best communication mechanism for Augmented Intelligence systems • There are some significant (and interesting) research challenges in using dialog in Augmented Intelligence systems © 2017 International Business Machines Corporation
  10. 10. Thank You

×