Real-world Stories and Long-term Risks and Opportunities.
Tom Dietterich, Ph.D.
Technical and Business Perspectives on the Current and Future Impact of Machine Learning - MLVLC
October 20, 2015
3. What is AI?
Smart Software
vision, speech, touch
choosing actions to achieve goals
learning
understanding and predicting
behavior
Credit: Andrej Karpathy, Li Fei-Fei
4. Exciting Progress: Perception
2013 2014 2015
23% Word Error
8%
Google Speech Recognition
Credit: Fernando Pereira & Matthew
Firestone, Google
Google Translate from Images
Credit: www.bbc.com
“a black and white cat is
sitting on a chair."
Credit: Jeff Donahue, Trevor Darrell
12. Tool AI: What’s Next?
Deeper understanding of video:
• What type of play?
• Who carried the ball?
• Was the pass complete?
• Which players made mistakes?
• Which players achieved their
goals?
Credit: Alan Fern
13. Tool AI: What’s Next?
Deeper understanding of text:
• Is Yoadimnadji dead? Yes
• Is he in Paris? Yes
• Is he married? Yes
• Where is his wife? In Paris
• Where will she be in the future? In
Chad
“Pascal Yoadimnadji has been
evacuated to France on
Wednesday after falling ill and
slipping into a coma in Chad.
His wife, who accompanied
him to Paris, will repatriate his
body to Chad.”
14. Tool AI: What’s Next?
Linking Big Data to Medicine:
• Web search logs can detect adverse
drug interaction events better than
FDA’s existing Adverse Drug
Interaction Reporting Service
White, Harpaz, Shah, DuMouchel, Horvitz, 2014
15. Tool AI: What’s Next?
Improved Personal Assistants that
combine
• Knowledge of recipes (lasagna
ingredients)
• Wine pairing recommendations
(cabernet or pinot noir)
• Brother’s home address
• Routes to brother’s home address
• Stores along that route that have
cheap cabernet or pinot in stock
to produce a plan
Source: Wired August 12, 2014
17. Autonomous AI Systems: What’s Next?
• AI Hedge Funds
• High Speed Trading
• Self-Driving Cars
• Automated Surgical Assistants
• Smart Power Grid
• Autonomous Weapons
18. The Motivations The Dangers
High-Stakes Autonomy
Advances in AI enable exciting
applications
There is a great potential to save
lives
There is a need to act at lightning
speed
Bugs
Cyber Attacks
Mixed Autonomy
Misunderstanding User
Commands
19. Software Quality
Many AI methods give only probabilistic
guarantees
Research Question: How can we ensure
safe performance of AI-based
autonomous systems?
"a young boy is holding a
baseball bat."
Credit: Andrej Karpathy, Li Fei-Fei
21. Cyber Attacks on AI Systems
Training Set Poisoning:
Make yourself invisible to a computer
vision system
Make yourself look “normal” to an
anomaly detection system
Bias the behavior of the system
Bid slightly higher on certain stocks
Prefer to show certain advertisements
Credit: Katherine Hannah
22. Mixed Autonomy
Auto-pilot unexpected hand-off to
pilots
Pilots lack situational awareness
and make poor decisions
Question: How can we make
imperfect autonomous systems
safe?
AF447 Pilots Final Conversation
Pilot 1: “What the...how is it we
are going down like this?”
Pilot 2: “See what you can do with
the commands up there, the
primaries and so on...Climb climb,
climb, climb
Credit: www.aviationlawmonitor.com
24. Trustable AI for Autonomy
Many research teams are at work
Verification and Self-Monitoring
Knowledge of Desirable and Acceptable Behavior
Improved User Interaction
26. 1980 1990 2000 2010
40
80
120
160
200
240
280
320
US Industrial Production
US Manufacturing Employment
Index:01.1972=100
Similar trends are seen worldwide
Source: Andrew McAfee / US Federal Reserve
US is Producing More with Fewer
Employees
27. As AI advances, many existing jobs have
the potential to be automated
Source: Frey & Osborne, 2014: “The future of employment: How susceptible are jobs to computerisation?”
28. Three Reasons to be Optimistic
AI often complements human intelligence rather than
replacing it
Increases in productivity increases in wealth
increases in consumption demand for new
goods and experiences new kinds of jobs
A focus on current jobs, underestimates the creativity
of people to develop new products and services, new
industries, etc.
29. What will be the new jobs?
Jobs involving hard-to-automate skills
high levels of social skills
deep understanding of human experience
and emotion
high levels of creativity
Jobs where human + machine is better
than either alone
Augmented Cognition
combining strategic thinking with detailed
tactics: “Centaur Chess”
combining physical dexterity with
information access: augmented reality
vehicle maintenance
Source: Tartajubow.blogspot.com
Source: metaio / designboom.com
30. Historical and cultural artifacts become
more valuable
AI cannot produce another Roman
Empire, Greek Civilization, or Al Andaluz
AI + Augmented Reality may make
tourism even more compelling and
meaningful
Credit: http://www.hellovisitspain.com
31. Life-Long Personal AI Assistant?
Human-machine pairs work best when they
know each other very well
strengths and weaknesses
easy communication
preferred ways of sharing tasks
A vision:
Student enrolls in technical college
Student is given an AI personal assistant
The student and the assistant train together
Learn how to solve problems jointly and efficiently
Employer hires the pair together
source: starwars.wikia.com
33. Some people are very afraid
December 2, 2014
October 27, 2014
34. AI Misconceptions
Intelligence is Not a Threshold Phenomenon
Progress in AI is the accumulation of thousands of incremental
improvements
Robots will not “wake up” one day and be “truly intelligent” or
“superintelligent” or “conscious” or “sentient”
“Tool AI” systems are already smarter than people along many
dimensions
35. AI Misconceptions
Autonomy Will Not Happen Spontaneously
There is no threshold above which AI systems suddenly have free will
Systems need to be designed and built to be autonomous
They must be given access to resources (money, power, materials,
generalized task markets, communications with people)
36. The danger of
“Autonomous AI” is not “AI”
but “Autonomy”
An autonomous system can be dangerous for many reasons:
It could consume vast resources
It could injure or kill people
It could apply AI to help it do these things
37. How Can We Maintain Control of
Autonomous Systems?
Case 1: Decisions are made at human time scales
Examples:
Aircraft autopilot
Most factory robots
Surgical robots
Humans can monitor and respond to problems
Similar to supervision in a human organization
38. Case 2: Very High Speed Decision Making
Examples:
stock market
power grid
self-driving cars
≥18,520 market crashes and spikes
have been detected 2006-2011
below the 950ms level
Flash Crash (10.5.2010)
39. Automated Monitoring
Signals might predict or
detect flash crashes
Example: Inventory of High-
Speed Traders
Normal market Flash crash
Source: Vuorenmaa, Tommi; Wang, Liang, 2013
40. When Monitoring Detects A Problem:
Then What?
Stock market:
Halt trading
Unwind transactions
Self-driving car:
Slow down and pull over in a safe spot
Smart Electric Grid:
??
41. We Should Never Create a
Fully Independent Autonomous System
By definition, a fully independent
autonomous system is a system over
which we have no control
42. Summary
AI has been making steady progress
Exciting applications of Tool AI are coming soon
Potential applications of Autonomous AI are being proposed
Many of these involve high-stakes decision making
There are many risks that must be addressed before it will be safe to field
such systems
Information technology—including AI—may be contributing to
unemployment but the picture is far from clear
Tool AI will not spontaneously become autonomous
We must find ways to control autonomous AI systems
Sources:
Fernando Pereira & Matthew Firestone, Google
2. http://www.bbc.com/news/technology-30824033
3. http://cs.stanford.edu/people/karpathy/deepimagesent/
Sources: Darrell UC Berkeley
From https://ece.uwaterloo.ca/~vganesh/talks/SATSMT-Dagstuhl-Aug8-12-2011-part1.pdf
Sources:
1. http://www.forbes.com/sites/thomasbrewster/2015/07/24/chrysler-recall-exploit/
2. http://blogs.wsj.com/digits/2015/08/06/hackers-take-control-of-a-tesla-sort-of/
3. http://www.wired.com/wp-content/uploads/2015/07/2015_0724_GK_RifleHack072-1024x682.jpg
As these systems become more capable, it is more and more important that we
Data poisoning for an ML algorithm? Adversarial computer vision examples.
Can you get confidence?
https://flic.kr/p/agnZGL
photo credit: http://www.aviationlawmonitor.com/tags/air-france-flight-447/
Air France 447
Conair 3407