The Past and Future of Artificial Intelligence
Patrick Henry Winston
27 March 2017
Vision
To know where we are going,
we have to know
• where we are
• how we got here
• what is happing now
• what we can do next
What’s new
• Programs that explain their own work
• Self-aware systems
How we got here:
The Beginning
From Turing’s paper
I propose to answer the question, “Can
machines think?”
• Average interrogator
• Not better than 70% correct
• Five minutes
From Turing’s paper
Objections:
Arguments from:
1. Theological
5. Disabilities
2. Head in the sands
7. Continuity in the
nervous system
3. Mathematical
4. Lady Lovelace’s
8. Informality of
behavior
9. Extrasensory
perception
So why the Turing Test?
What we learned from Turing
• Serious people can think seriously
about computers thinking
• There is no reason to doubt that
computers will think someday
How we got here:
The Pioneers
At Stanford: Formal logic
At Carnegie Mellon:
Human problem solving
At MIT:
Many methods working together
AI History
The First Wave
James Slagle 1961
Analogical reasoning
Anytime algorithms
Bayesian techniques
Case-based reasoning
Computational linguistics
Connectionism
Constraint satisfaction
Cross-modal clustering
Cyc
Explanation based learning
Frames
Fuzzy sets
Genetic algorithms
Image understanding
Information retrieval
Intermediate features
K-lines
Knowledge engineering Question answering
Layered reasoning
Reasoning hierarchies
Legged locomotion
Regularity based learning
Lisp, Smalltalk, and Logo Reinforcement learning
Logic systems
Representation
Machine learning
Rule based systems
Means-ends
Search
Memory based control Self-organizing maps
Model based systems
Soar
Multiple methods
Speech understanding
Multiple resolution
Streams and counter streams
Neural nets
Structure mapping
Open mind
Subsumption
Planning and scheduling Symbolic mathematics
Probabilistic reasoning Visual and motor routines
Problem reduction
Wait-and-see algorithms
Propagator architecture
Winograd
Winston
Waltz Sussman
If you get the representation right
Then you are almost done.
What we learned from early AI
Architectures that deploy
Methods that exploit
Constraints exposed by
Representations that support
Models of
Language, Vision,
Problem solving,
and Learning
AI History
The Second Wave
Mycin Program: medical diagnosis
If x’s type is primary bacteremia
and x’s suspected portal is
gastrointestinal
and the site of the culture of x is sterile
then there is evidence that x is
bacteroides
Analogical reasoning
Anytime algorithms
Bayesian techniques
Case-based reasoning
Computational linguistics
Connectionism
Constraint satisfaction
Cross-modal clustering
Cyc
Explanation based learning
Frames
Fuzzy sets
Genetic algorithms
Image understanding
Information retrieval
Intermediate features
K-lines
Knowledge engineering Question answering
Layered reasoning
Reasoning hierarchies
Legged locomotion
Regularity based learning
Lisp, Smalltalk, and Logo Reinforcement learning
Logic systems
Representation
Machine learning
Rule based systems
Means-ends
Search
Memory based control Self-organizing maps
Model based systems
Soar
Multiple methods
Speech understanding
Multiple resolution
Streams and counter streams
Neural nets
Structure mapping
Open mind
Subsumption
Planning and scheduling Symbolic mathematics
Probabilistic reasoning Visual and motor routines
Problem reduction
Wait-and-see algorithms
Propagator architecture
Artificial Intelligence
Cognition
Machine learning
Natural language
Vision
Robotics
Artificial Intelligence
Robotics
Cognition
Machine
learning
Natural
language
Vision
…
1985-2010
AI History
Third Wave
1996
2010
2011
A group of young people playing a game of frisbee
Analogical reasoning
Anytime algorithms
Bayesian techniques
Case-based reasoning
Computational linguistics
Connectionism
Constraint satisfaction
Cross-modal clustering
Cyc
Explanation based learning
Frames
Fuzzy sets
Genetic algorithms
Image understanding
Information retrieval
Intermediate features
K-lines
Knowledge engineering Question answering
Layered reasoning
Reasoning hierarchies
Legged locomotion
Regularity based learning
Lisp, Smalltalk, and Logo Reinforcement learning
Logic systems
Representation
Machine learning
Rule based systems
Means-ends
Search
Memory based control Self-organizing maps
Model based systems
Soar
Multiple methods
Speech understanding
Multiple resolution
Streams and counter streams
Neural nets
Structure mapping
Open mind
Subsumption
Planning and scheduling Symbolic mathematics
Probabilistic reasoning Visual and motor routines
Problem reduction
Wait-and-see algorithms
Propagator architecture
A refrigerator filled with lots of food and drink
Lesson Learned
Deep nets are amazing, but they do not
• See like us
• Think like us
• An important part of the story,
but not the whole story
Third wave
Worried people
2014
Musk: With Artificial
Intelligence, we are
summoning the demon.
AI is our biggest
existential threat
1970
Once the computers
get control, we might
never get it back. We
would survive at their
sufferance. If we’re
lucky, they might keep
us as pets.
~1970
Newell and Simon:
The SOAR reasoning
system is the answer;
Alert the people
~1980
Feigenbaum:
Rule based expert systems
are the answer
~1990
Brooks:
Subsumption is the answer
Present
Hinton:
Deep neural nets
are the answer
What we learned from Minsky
We need it all: representations, methods,
agents, all working together
AlphaGo vs Lee Sedol, 2016
Messages for worried people
• New technology has unexpected dangers
• Ordinary computing is changing the world
• AI is not our top threat
• We are our biggest threat
Third wave:
What’s next
• Computational statistics,
• machine learning,
• deep nets
enabled by
• massive computing,
• massive training data
Companies protecting data,
not programs
What next after next:
The fourth wave
Why I am optimistic
• Massive, free computing
• Excited people
• Emerging round table
• Accumulated progress
• Better questions
Merge → an inner language
∞
We, and only we,
build complex,
deeply nested,
symbolic descriptions
of situations and
events
The Fourth Wave
AI
Artificial
Perception
The Fourth Wave
AI
Artificial
Perception
Reasoning
The Fourth Wave
AI
Recipe following
Artificial
Perception
Reasoning
The Fourth Wave
AI
Story understanding
Recipe following
Artificial
Perception
Reasoning
The Strong Story Hypothesis
The mechanisms that enable us
humans to tell, understand, and
recombine stories separate our
intelligence from that of other
primates.
Story Understanding
Symbolic Description
Merge
Perception
Sequencing
Fairy and folk tales
Religious parables
Ethnic narratives
History
Literature
Experience
News
…
Law
Business
Medicine
Defense
Diplomacy
Science
Engineering
…
What is read
Macbeth is a thane and Macduff is a thane. Lady Macbeth is evil and greedy. Duncan is the
king, and Macbeth is Duncan's successor. Duncan is an enemy of Cawdor. Macduff is an
enemy of Cawdor. Duncan is Macduff's friend. Macbeth defeated Cawdor. Duncan becomes
happy because Macbeth defeated Cawdor. The witches danced and had visions. Macbeth talks
with the witches. The witches predicted that Macbeth will become king. The witches astonish
Macbeth. Duncan executes Cawdor. Macbeth becomes Thane of Cawdor. Duncan rewarded
Macbeth because Duncan became happy. Macbeth wants to become king because Lady
Macbeth persuaded Macbeth to want to become the king. Macbeth invites Duncan to dinner.
Duncan complements Macbeth. Duncan goes to bed. Duncan's guards become drunk and sleep.
In order to murder Duncan, Macbeth murders the guards and Macbeth stabs Duncan. Mabeth
becomes king. Malcolm and Donalbain flee. Macbeth's murdering Duncan leads to Macduff's
fleeing to England. In order to flee to England, Macduff rides to the coast and Macduff sails on
a ship. Then, Macduff's fleeing to England leads to Macbeth's murdering Lady Macduff.
Macbeth hallucinates at a dinner. Lady Macbeth says he hallucinates often. Everyone leaves
because Lady Macbeth tells everyone to leave. Macbeth's murdering Duncan leads to Lady
Macbeth's becoming distraught. Lady Macbeth has bad dreams. Lady Macbeth thinks she has
blood on her hands. Lady Macbeth kills herself. Birham Wood is a forest. Burnham Wood
goes to Dunsinane. Macduff's army attacks Dunsinane. Macduff curses Macbeth. Macbeth
refuses to surrender. Macduff kills Macbeth. The end.
Is Duncan dead?
Why did Macduff kill Macbeth?
Is there revenge? Pyrrhic victory?
Deduction rules
If x kills y, then y becomes dead.
Macbeth murders Duncan.
Macbeth murders Duncan
Duncan becomes dead
Explanation rules
If x angers y, then y may kill x.
Macbeth angers Macduff….
Macduff kills Macbeth.
Macbeth angers Macduff
Macduff kills Macbeth
The Doctrine of Necessity
Introduce only mechanisms that are
needed to model an observed
behavior.
Guess reason: abduction rule
If x murders y, then x must be insane.
Macbeth murders Duncan.
Macbeth is insane
Macbeth murders Duncan
Censor
If y becomes dead, y cannot become
unhappy.
Duncan becomes dead.
Duncan becomes dead
Duncan becomes unhappy
Concept patterns
X's harming Y
leads to Y's harming X.
Situational explanation, Chinese
Dispositional explanation, American
Duncan is a person. Lady Macbeth is a person. Macduff is
a person. Macbeth is a person. A thane is a noble.
Macbeth is a thane. Macduff is a thane. Lady Macbeth is
greedy. Macbeth defeats a rebel. Appear is a success.
Macbeth has a success. Witches talk with Macbeth.
Witches have visions. Duncan rewards Macbeth because
Duncan becomes happy. Macbeth wants to become king
because Lady Macbeth persuades Macbeth to want to
become king. Macbeth murders Duncan.
Duncan becomes dead because
Person x becomes dead whenever
Person y kills person x.
Macbeth is a thane and Macduff is a thane. Lady Macbeth
is evil and greedy. Duncan is the king, and Macbeth is
Duncan's successor. Duncan is an enemy of Cawdor.
Macduff is an enemy of Cawdor. Duncan is Macduff's
friend. Macbeth defeated Cawdor. Duncan becomes happy
because Macbeth defeated Cawdor. The witches danced
and had visions. Macbeth talks with the witches. The
witches predicted that Macbeth will become king. The
witches astonish Macbeth. Duncan executes Cawdor.
Macbeth becomes Thane of Cawdor. Duncan rewarded
Macbeth because Duncan became happy. Macbeth wants
to become king because Lady Macbeth persuaded Macbeth
to want to become the king. Macbeth invites Duncan to
dinner. Duncan complements Macbeth. Duncan goes to
bed. Duncan's guards become drunk and sleep. In order to
murder Duncan, Macbeth murders the guards and Macbeth
stabs Duncan. Mabeth becomes king. Malcolm and
Donalbain flee. Macbeth's murdering Duncan leads to
Macduff's fleeing to England. In order to flee to England,
Macduff rides to the coast and Macduff sails on a ship.
Then, Macduff's fleeing to England leads to Macbeth's
murdering Lady Macduff. Macbeth hallucinates at a
dinner. Lady Macbeth says he hallucinates often. Everyone
leaves because Lady Macbeth tells everyone to leave.
Macbeth's murdering Duncan leads to Lady Macbeth's
becoming distraught. Lady Macbeth has bad dreams. Lady
Macbeth thinks she has blood on her hands. Lady Macbeth
kills herself. Birham Wood is a forest. Burnham Wood
goes to Dunsinane. Macduff's army attacks Dunsinane.
Macduff curses Macbeth. Macbeth refuses to surrender.
Macduff kills Macbeth. The end.
The story is about Pyrrhic
victory. Lady Macduff was
Macduff's wife. Macbeth
wanted to become king
because Lady Macbeth
persuaded Macbeth to want
to become king. Macbeth
murdered Duncan, probably
because Duncan was a king
and Macbeth was Duncan's
successor. Macduff fled to
England. Macbeth killed
Lady Macduff. Macduff
killed Macbeth, probably
because Macbeth angered
Macduff.
On a common-sense level, the Healthy
reader thinks Samantha grabs an
umbrella, probably because Samantha
wants to reach the handle, and Samantha
fails to reach the handle.
The Schizophrenic reader thinks
Samantha grabs an umbrella because
Samantha wants to go into the rain.
What we learned about stories
Simple mechanisms support many abilities:
• Instruction with student model
• Summary via concepts
• Story composition
• Mental illness explanation
• Hypothetical reasoning
• Self-aware processing
Can a machine be self aware?
Suitcase Words
Intelligence
Creativity
Emotion
Self awareness
…
Lu is a student. Shan is a student. Lu inhabits America.
Lu fails his dissertation defense. Goertz is Lu's advisor.
Goertz and Lu are not friends. Lu is Chinese. Lu went
to US to do phD. Lu had highest entrance exam score.
Lu is a bachelor. Lu is lonely. Lu owns a gun. Lu
practices shooting. Lu passes his second dissertation
defense. Lu becomes a lab assistant because Lu does
not find a job. Shan is younger than Lu. Shan graduates
with Lu. Shan received national award. Faculty
rejected Lu's appeal. Goertz angers Lu. Shan comes
from a small Chinese village. Shan is married. Shan is
social. Shan has friends. Shan is successful. Shan
outperforms Lu. In order to kill Shan, Lu shoots Shan.
Then, Lu shoots himself.
Did Lu kill Shan because America
is individualistic?
Can you explain why you think so?
Did Lu murder Shan?
Why did Lu murder Shan?
How did Lu murder Shan?
Was Lu lonely?
Was Lu lonely before he murdered Shan?
What happened after Lu murdered Shan?
Do you believe America is individualistic?
What is the story about?
Story Processor
Neural net
Neural net
Neural net
Probabilistic
program
Probabilistic
program
Rule-based
system
Subsumption
system
Planning
system
What you need to know
• Part of the answer = computational statistics.
• That part enabled by massive computing,
massive data.
• Tomorrow’s answer is humanlike thinking
• Tomorrow’s answer is on the horizon.
Expected contributions
• Smart systems on another level.
• Smart systems with explanations.
• Smart systems that understand us.
• A better understanding of ourselves
and each other and that will make the
world a better place