Transcript Powerpoint presentation 3

Artificial Intelligence
and
Computing Science
October 19, 2012
AI in the Seventies and Eighties
A belated realization: in most
situations, intelligent behaviour
requires knowledge
How many things do you know?
``I can’t think of a hundred things
that I know a thousand things
about’’
(Marvin Minsky, `Semantic Information Processing’, 1968)
AI in the Seventies
Winograd, SHRDLU
Expert Systems
Minsky, `Frames’
Rumelhart and McClellan, `Parallel
Distributed Programming’ (return of the
Perceptron)
SHRDLU’s World
Sample Dialogue
Winograd: What’s on the green block?
Shrdlu: A pyramid
Winograd: Put the red block on the green block
Shrdlu: OK
Winograd: Why did you move the purple pyramid?
Shrdlu: To clear off the green block
Winograd: Why did you put the red block on the green block?
Shrdlu: You told me to
Winograd: Pick up a big red block
Shrdlu: OK
Winograd: Grasp the pyramid
Shrdlu: I don’t understand which pyramid you mean
Winograd: Find a block which is taller than the one you are holding
and put it into the box
Shrdlu: By `it’, I assume you mean the block which is taller than the
one I am holding.
OK.
Winograd: What does the box contain?
Shrdlu: The blue pyramid and the blue block
Winograd: What is the pyramid supported by?
Shrdlu: The box
Winograd: Can the table pick up blocks?
Shrdlu: No
Things SHRDLU doesn’t know:
`red’ and `green’ are colours
SHRDLU’s world is flat
A table has legs but no arms
SHRDLU is a computer program
Expert Systems (rule-based)
Knowledge can be represented by a
number of `if…then’ rules plus an
inference engine.
E.g, ``IF temperature is high AND rash is present,
THEN patient has measles.’’
We can extract the rules from human experts via interviews.
This process is known as `knowledge engineering’:
`If an animal has fur, it is a mammal’
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal has scales, it is a fish’
`If an animal is a fish, it can swim’
`If an animal lays eggs and has fur, it is a duck-billed platypus’
This gives us a set of rules that an inference engine (or `expert system shell’)
can reason about.
Two popular modes of reasoning are forward chaining and backward chaining:
`If an animal has fur, it is a mammal’
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal is a bird, it lays eggs’
`If an animal has scales, it is a fish’
`If an animal is a fish, it can swim’
`If an animal lays eggs and has fur, it is a duck-billed platypus’
Forward chaining:
Given a new fact (`Tweety has feathers’), search for all matching conditionals,
draw all possible conclusions, and add them to the knowledge base:
:- Tweety is a bird
:- Tweety can fly
:- Tweety lays eggs
Potential problem: we run into the combinatorial explosion again
Backward chaining:
Given a query (`Does Tweety lay eggs?’),
search for all matching consequents
and see if the database satisfies the conditionals:
`If an animal has fur, it is a mammal’
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal is a bird, it lays eggs’
`If an animal has scales, it is a fish’
`If an animal is a fish, it can swim’
`If an animal lays eggs and has fur, it is a duck-billed platypus’
`Tweety has feathers’
Backward chaining:
`Does Tweety lay eggs?’
`If an animal has fur, it is a mammal’
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal is a bird, it lays eggs’
`If an animal has scales, it is a fish’
`If an animal is a fish, it can swim’
`If an animal lays eggs and has fur, it is a duck-billed platypus’
`Tweety has feathers’
Backward chaining:
`Does Tweety lay eggs?’
`Is Tweety a bird?’
`If an animal has fur, it is a mammal’
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal is a bird, it lays eggs’
`If an animal has scales, it is a fish’
`If an animal is a fish, it can swim’
`If an animal lays eggs and has fur, it is a duck-billed platypus’
`Tweety has feathers’
Backward chaining:
`Does Tweety lay eggs?’
`Is Tweety a bird?’
Does Tweety have feathers?’
`If an animal has fur, it is a mammal’
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal is a bird, it lays eggs’
`If an animal has scales, it is a fish’
`If an animal is a fish, it can swim’
`If an animal lays eggs and has fur, it is a duck-billed platypus’
`Tweety has feathers’
Backward chaining:
Conclusion: Yes, Tweety does lay eggs
This method is used by Prolog, for example
`If an animal has feathers, it is a bird’
`If an animal is a bird, it can fly’
`If an animal is a bird, it lays eggs’
Potential problem: A lot of rules have exceptions.
Frames
(Marvin Minsky, 1974)
A frame allows us to fill in default
knowledge about a situation from a
partial description. For example,
``Sam was hungry. He went into a
Mcdonalds and ordered a hamburger.
Later he went to a movie.’’
Did Sam eat the hamburger?
So we can economically
represent knowledge by
defining properties at the
most general level, then
letting specific cases
inherit those properties…
Event
Transaction
Buying something
Buying a hamburger
Return of the perceptron
(now called a `neural net’)
Changes since 1969:
Hidden layers
Non-linear activation function
Back-propagation allows learning
Rumelhart and McClelland
`Parallel Distributed
Processing’
Use neural nets to represent knowledge by
the strengths of associations between
different concepts, rather than as lists of
facts, yielding programs that can learn
from example.
Conventional Computer Memory
Register One
01100110
Register Two
11100110
Register Three
00101101
....
AI: 1979-2000
Douglas Lenat, `CYC’,
Douglas Hofstadter, `Fluid Analogies’
Brian Hayes, `Naïve Physics’
CYC’s data are written in CycL, which is a
descendant of Frege’s predicate calculus
(via Lisp).
For example,
(#$isa #$BarackObama #$UnitedStatesPresident)
or
(#$genls #$Mammal #$Animal)
The same language gives rules for deducing
new knowledge:
(#$implies
(#$and
(#$isa ?OBJ ?SUBSET)
(#$genls ?SUBSET ?SUPERSET))
(#$isa ?OBJ ?SUPERSET))
What CYCcorp says CYC knows about
`intangible things’.
Intangible Things are things that are not physical
-- are not made of, or encoded in, matter. These
include events, like going to work, eating dinner, or
shopping online. They also include ideas, like
those expressed in a book or on a website. Not the
physical books themselves, but the ideas
expressed in those books. It is useful for a
software application to know that something is
intangible, so that it can avoid commonsense
errors; like, for example, asking a user the color of
next Tuesday's meeting.
Questions CYC couldn’t answer
in 1994
What colour is the sky?
What shape is the Earth?
If it’s 20 km from Vancouver to Victoria,
and 20 km from Victoria to Sydney, can
Sydney be 400 km from Vancouver?
How old are you?
(Prof. Vaughan Pratt)
Hofstadter: Fluid Analogies
Human beings can understand similes, such as
``Mr Pickwick is like Christmas’’
Example:
Who is the Michelle Obama of Canada?
Michaelle Jean, Governor-General
Head of government
Spouse
Spouse
Head of State
Spouse
Spouse
One of Hofstadter’s approaches to solving these
problems is `Copycat’, a collection of independent
competing agents.
If efg becomes efw, what does ghi become?
If aabc becomes aabd, what does ijkk become?
Inside Copycat:
ij(ll)
ij(kk)
ijkk
(ijk)l
(ijk)k
aabd:jjkk
aabc:ijkk
aabd:hjkk
If efg becomes efw, what does ghi become?
COPYCAT suggests whi and ghw
If aabc becomes aabd, what does ijkk become?
COPYCAT suggests ijll and ijkl and jjkk and hjkk
Hofstadter:
``What happens in the first 500 milliseconds?”
Find the O
XXXXXXXXXXX
XXXXXXXXXXX
XXXXXXXXXXX
XXXXXXXXOXX
XXXXXXXXXXX
Find the X
XXXXXXXXXXX
XXXXXXXXXXX
XXXXXXXXXXX
XXXXXXXXXXX
XXXXXXXXXXX
Find the O
XXOXXXOXXOX
XXXXXOXXOXX
XXXOXXOXOXX
XXOXXOXXOXX
OXXXXXOXXOX
What eye sees
What I see
The Cutaneous Rabbit
Naive model of perception:
World
Vision
Awareness
Better model of perception:
World
Vision and Knowledge
Awareness
We recognise all
these as instances
of the letter `A’.
No computer can
do this.
Hofstadter’s program
`Letter Spirit’ attempts
to design a font.
Naïve Physics
Hayes, ‘Naïve Physics Manifesto’, 1978
“About an order of magnitude more work than any
previous AI project …’’
Hayes, `Second Naïve Physics
Manifesto’, 1985
“About two or three orders of magnitude more
work than any previous AI project…”
One sub-project of naïve physics:
Write down what an
intelligent 10-year-old
knows about fluids
Part of this is knowing how we talk about fluids:
For example:
Suppose Lake Chad
dries up in the dry
season and comes
back in the wet
season.
Is it the same lake
when it comes back?
Suppose I buy a cup of
coffee, drink it, then
get a free refill.
Is it the same cup of
coffee after the refill?
2011: IBM’s Watson Wins Jeopardy
Inside Watson:
4 Terabytes disk storage: 200 million pages
(including all of Wikipedia)
16 Terabytes of RAM
90 3.5-GHz eight-core processors
One of the components of Watson is a Google-like search
algorithm.
For example, a typical Jeopardy question in the category
`American Presidents’ might be
``The father of his country, he didn’t really chop down a
cherry tree’’
Try typing `father country cherry tree’ into Google
The first hit is `George Washington – Wikipedia’
But Watson also needs to know how confident it should
be in its answers
Conspicuous Failures, Invisible Successes
In 2012, we have nothing remotely
comparable to 2001’s HAL.
On the other hand, some complex
tasks, such as attaching a printer to a
computer, have become trivially easy
A different approach: robot intelligence
Grey Walter’s machina speculatrix, 1948
BEAM robotics,
Queen Ant, a light-seeking hexapod, 2009
AI Now: Robot Intelligence
Rodney Brooks, `Cambrian Intelligence’
-Complex behaviour can arise when a
simple system interacts with a complex
world.
-Intelligent behaviour does not require a
symbolic representation of the world.
SPIRIT: Two years on
Mars and still going.
Brook’s approach invites us to reconsider
our definition of intelligence:
…is it the quality that distinguishes
Albert from Homer?
…or the quality
that
distinguishes
Albert and
Homer
from a rock?
`Chess is the touchstone of intellect’
-- Goethe
…but perhaps we are most impressed by
just those of our mental processes that
move slowly enough for us to notice
them…
Strong AI:
``We can build a machine that will have a
mind.’’
Weak AI:
``We can build a machine that acts like it
has a mind.’’
Strong AI (restatement):
``We can build a machine that, solely by
virtue of its manipulation of formal
symbols, will have a mind.’’
Hans Moravec:
``We will have humanlike competence
in a $1,000 machine in about forty
years.’’
---- 1998
Hubert Dreyfus:
``No computer can ever pass the Turing
Test, or do any of the following things
[long list, including `play master-level
Chess’].’’
1965; MacHack beat
Dreyfus in 1967
…and if a program did pass the
Turing test, what then?
John Searle:
``Even if a computer did pass the Turing
test, it would not be intelligent, as we
can see from the Chinese Room
argument’’
John Bird:
``A computer is a deterministic system,
and hence can have neither free-will,
responsibility or intelligence -- whether
it passes the Turing test or not.’’
``This is an AND gate:
A
C
B
A
B
C
0
0
1
1
0
1
0
1
0
0
0
1
Given A and B, does the computer have any choice
about the value of C?
… but a computer is just a collection of AND gates and
similar components. If none of these components can
make a free choice, the computer cannot make a free choice.’’
The Brazen Head