Transcript Will Androids Dream of Electric Sheep? A Glimpse of Current and

Artificial Intelligence:
Prospects for the 21st Century
Henry Kautz
Department of Computer Science
University of Rochester
What is Artificial Intelligence?
• Study of principles for understanding
and building intelligent agents
–
–
–
–
–
–
Human, animal, or mechanical
How to perceive the world
How to reason and make decisions
How to learn
How to act (motion, speech)
How to cooperate with other agents
Can’t Win Definition of AI
• AI = making a computer solve a
problem that requires human
intelligence
– By definition, any problem solved by AI
no longer requires human intelligence
– So, AI never succeeds!
• Useful idea: study tasks people
perform in order to understand
intelligence
Outline
• Approaches to AI
– Task based (“Classical AI”)
– Neural networks
• Which Way Will Achieve AI?
– Criticisms
– Ray Kurzweil’s Perspective
– A Middle Ground
Classical AI
• The principles of intelligence are
separate from the hardware (or
“wetware”)
• Look for these principles by studying
how to perform individual tasks that
require intelligence
Success Story: Medical Expert
Systems
• 1980: First expert level performance
– diagnosis of blood infections
• Today: 1,000’s of systems
– Often outperform doctors
Success Story:
Chess
I could feel – I
could smell – a new
kind of intelligence
across the table
- Garry Kasparov
(1997)
•Examines 5 billion positions /
second
•Intelligent behavior emerges
from brute-force search
Success Story: Robotics (1)
Rendezvoused with an asteroid, 1998-2000
Capable of autonomous diagnosis & repair
Success Story: Robotics (2)
• DARPA Grand Challenges, 2004-2007
– Races in desert and urban environments by
fully autonomous vehicles
– Succeeded with “off the shelf” AI
technology!
Success Story: Text to Speech
• Kurzweil Reading Machines, 1978-2006
Neural Networks
• Develop computational models of the
brain at the neural level
– McCulloch & Pitts model (1943): very
simple, but a pretty good approximation
of most real neurons
Success Story: Face
Recognition
• Programming a
neural net
that learns to
recognize
faces can now
be done as
homework
problem!
Success Story: Brain-Computer
Interfaces
Miguel
Nicolelis
(2003),
Duke
University
Success Story: MRI Imaging of
Specific Thoughts
Tools
Buildings
• Tom Mitchell (CMU) 2006
Food
Which Approach Will Achieve
AI?
• Criticism of Classical AI:
– Successes so far are in all narrow
domains
– We can never explicitly program enough
“commonsense” into a AI system to make
it a true general intelligence
– The human brain has a completely
different architecture than a modern
computer
Which Approach Will Achieve
AI?
• Criticism of Neural Networks:
– Successes so far are in all narrow
domains
– Building an AI by studying neural
processes is like trying to reverseengineer Windows Vista by watching bits
– “Summation and threshold” is just
another kind of logic gate!
A Middle Ground
• Most AI researchers (including me)
believe that AI will be accomplished by
a combination of ideas from both camps
– Studying tasks tells us what needs to be
computed
– Studying brains tells us what classes of
algorithms are possible
– We can implement those algorithms in many
ways
A Middle Ground
• Neural nets are not necessary the best
way to implement all the thing the brain
does!
– Evolution rarely produces optimal solutions!
• Machine learning is compatible with both
the classical and neural net approaches
– Learning from text on the Internet will
solve the problem of getting enough
“commonsense” information
Pioneering an Emerging Area
• Assisted Cognition
– Computer systems that enhance the
abilities, independence, and safety of
persons with cognitive disabilities
• Aging and age-related diseases
• Brain injury
• Developmental disabilities
– Computer caregivers
Examples
• Maintaining a daily schedule
– Compensating for memory problems
– Compensating for lowered self-initiative
– Step-by-step task prompting
• Navigation
– Indoors and outdoors
• Safety and health
– Need for immediate help
– Long term health trends
Activity of Daily
Living Monitoring
• Goal: Accurate,
automated ADL logs
– Changes in routine
often precursor to
illness, accidents
– Human monitoring
intrusive &
inaccurate
Object-Based Activity Recognition
• Activities of daily living involve the
manipulation of many physical objects
– Kitchen: stove, pans, dishes, …
– Bathroom: toothbrush, shampoo, towel, …
– Bedroom: linen, dresser, clock, clothing, …
• We can recognize activities from a
time-sequence of object touches
Sensing Object Manipulation
• RFID: Radiofrequency
identification tags
–
–
–
–
Small
Semi-passive
Durable
Cheap
• Near future: use
products’ own tags
Wearable RFID Reader
• Bracelet reads tags near hand, transmits information
wirelessly to monitoring system
• Soon will be built into a wristwatch
Interpreting the Sensor Data:
Machine Learning
• Machine learning algorithms automatically create
the recognition system from training examples
• Can handle sensor noise and user errors
Using Commonsense Knowledge
• Can further improve the system by
adding “commonsense knowledge”
• Example: a travel mug is like a cup
Results: Detecting ADLs
Activity
Prior
Work
SHARP
RFID
Legend
Personal Appearance
92/92
Oral Hygiene
70/78
Toileting
73/73
Washing up
100/33
Appliance Use
100/75
Use of Heating
84/78
Care of clothes and linen
100/73
Making a snack
100/78
Making a drink
75/60
Use of phone
64/64
Leisure Activity
100/79
Infant Care
100/58
Medication Taking
100/93
Housework
100/82
General solution
Point solution
Inferring ADLs from
Interactions with
Objects
Philipose, Fishkin,
Perkowitz, Patterson,
Hähnel, Fox, and
Kautz
IEEE Pervasive
Computing, 4(3), 2004
All Watched Over by Machines of Loving Grace,
by Richard Brautigan
I like to think
(it has to be!)
of a cybernetic ecology
where we are free of our labors
and joined back to nature,
returned to our mammal
brothers and sisters,
and all watched over
by machines of loving grace.