Computing Machinery and Intelligence

boorishadamantΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 3 χρόνια και 5 μήνες)

52 εμφανίσεις

Computing Machinery and
Intelligence

By Alan M. Turing

Computing Machinery and
Intelligence


Published in “Mind:
A Quarterly Review
of Psychology and
Philosophy”, in
1950.


“I propose to
consider the
question, ‘Can
machines think?’”

About the paper


Describes the imitation game, now called
the Turing Test.


Possibly one of most important and disputed
topics in AI, philosophy of mind, cognitive
science.


The foundation of AI, and its ultimate goal?


Useless and even harmful?


A key paper regardless.

Which is machine,
and which is
woman???

The Imitation Game

Which is man,
and which is
woman???

Let’s try it…computer of poet?

At six I cannot pray:

Pray for lovers,

through narrow streets

And pray to fly

But the Virgin in their dark wintry bed

Let’s try it…computer of poet?

What seas what shores what granite
islands toward my timbers

And woodthrush calling through the fog

My daughter.

Let’s try it…computer of poet?

Men with picked voices chant the names

of cities in a huge gallery: promises

that pull through descending stairways

to a deep rumbling.

Let’s try it…computer of poet?

Where were thou, sad Hour, selected from
whose race is

guiding me,

Lured by the love of Autumn's being,

Thou, from heaven is gone, where was lorn
Urania

When rocked to fly with thee in her clarion o'er
the arms of death.

A Brief History of AI, pre
-
Turing Test


Greek mythology: Hephaestus, idea of intelligent
robots.


13
th

century: talking heads, supposedly owned by
Robert Bacon, Albert the Great


15
th

century: da Vinci drafted robot design


16
th

century: the Maharal of Prague’s Golem


17
th

century: Descartes


“animals are complex
machines”


19
th

century: Charles Babbage’s Analytical Engine


1940’s: Isaac Asimov


“Three Laws of Robotics”


1943: McCulloch and Pitts, model neurons with
algorithms?

Turing’s contemporaries, and
subsequent related work in AI


Claude Shannon, 1950: algorithm for playing Chess.


Alan Newell and Herbert Simon, 1956: one of first expert
systems, Logic Theorist.


Noam Chomsky, 1957: analyze language mathematically,
Syntactic Structures.


Friedberg, 1958: genetic algorithms.


Joseph Weizenbaum, 1966: writes computer program ELIZA,
with some success at imitation game


Computerized human psychologist


Minsky and Papert, 1968: wrote book
Perceptrons
, showing
some limitations of neural nets. Slowed research in area.


Kurzweil Reading Machine, 1976: read printed text.


MYCIN, 1979: expert system that diagnosed some diseases.

Proponents and Opponents of
AI


Lots of debate about potential success and
limitations of AI.


Herbert Simon, 1958: “within ten years a digital
computer will be the world’s chess champion.”


Hubert Dreyfus, 1972:
What Computers Can’t
Do


Human intelligence is more than manipulation of
symbols.


John Searle, 1980: Opposed idea of strong AI,
that machines can think, with “Chinese Room”
thought experiment.

The Paper


‘Can machines think?’


Not a meaningful question, definitional issues


Instead, suggests imitation game


Description of machines, and universality of
digital computers


Possible objections to the question and the
test, with responses:


Theological, mathematical, arguments from
consciousness, originality, etc.


Learning machines

Digital Computers


‘Are there imaginable digital computers
which would do well in the imitation
game?’


Manchester Mark 1

Predictions


In 1950, Turing predicted that 50 years later it will
be possible to program a computer with ~100 Mb
memory to pass TT 70% of the time, with 5 minute
conversations.


It will be natural to speak of computers ‘thinking’.


“[The machine] may be used to help in making up
its own programmes, or to predict the effect of
alterations in its own structure.”


“We may hope that machines will eventually
compete with men in all purely intellectual fields.”

Some Objections


Theological objection: Thinking is part of
humans’ souls, and so animals/machines
can’t think.


Head
-
in
-
the
-
sand objection: Consequences
of thinking machines are dreadful, so let’s
hope it’s not possible.


Futuristic movies and books build upon this fear.


Machines will never be able to do X.


X = {be kind, friendly, have sense of humor, fall
in love, etc.}

Mathematical Objection


Gödel’s Incompleteness Theorem: in any consistent
logical system that includes number theory, there
are statements that can’t be proved or disproved


The halting problem: no machine can determine
whether another machine will halt on a given input


These show limitations to discrete
-
state machines


But humans are not infallible


Judge of the imitation game will not know if
incorrect response is because of limitation or
human error.

Consciousness


“Not until a machine can write a
sonnet…because of thoughts and emotions
felt…could we agree that machine equals
brain…”


Professor Jefferson


Really just attack on TT, but TT does not
test whether computer thinks or feels.


Solipsism: the only way to
really
know if a
machine is thinking is to
be

the machine.

Lady Lovelace’s Objection


Wrote about Babbage’s Analytical Engine.


Machine can not
originate

anything, and only does
what it is programmed to do.


But what about learning machines?


Maybe machines can’t surprise?


But then again, humans often are surprised by
machines.


In addition, what is surprise? Theorems may not be
surprising after they are proven, but is there no
virtue in proving them?

Continuity of the Nervous
System


Nervous system is not a discrete
-
state
machine, so can’t mimic by computer.


Again, interrogator can’t tell difference

Extra
-
Sensory Perception


Seems to acknowledge overwhelming
statistical evidence for telepathy.


Imitation game fails with ESP, since human
can communicate with interrogator via ESP.


Telepathic human is better at guessing games
(i.e. which hand is coin in?)


To solve this, Turing suggests putting
subjects in ‘telepathy
-
proof room’.

Learning Machines


“Presumably the child brain is something like a
notebook as one buys it from the stationer’s. Rather
little mechanism, and lots of blank sheets.”


Replicate child brain, and then feed it information.


Gives estimate of amount of storage in human
brain: 10
9

decimal digits.


Much less than currently believed.


Believes that once memory is available,
constructing a computer with a human
-
like mind is
“mainly one of programming”.


Even discusses ways of “teaching” computer.

Later debate on the TT


Stuart Shieber’s analogy:


Deniers: intelligence is like bad cold.


There is a germ, a hidden cause.


Can’t “fake it”.


Approvers: intelligence is like fluency in Italian.


Talk to someone in Italian for an hour.


Can’t say, “he doesn’t really know Italian, he’s just faking it.”


Now say someone gets good grades, does well on
Psychometry


Can you say, “He’s not really intelligent, he’s just faking it to
get into a good University”?

Searle’s “Chinese Room”


Thought experiment, 1980.


Refined consciousness objection.


There is a room, with a man who only speaks English.


Man has book, with instructions: given some scribble
in Chinese, output this scribble.


A man fluent in Chinese sends messages (in Chinese)
into room, and gets responses (also in Chinese).


He can’t distinguish b/w man in room and fluent Chinese
speaker.


But does this mean the room knows Chinese??


Conclusion: TT only tests for “weak” AI, not “strong”
AI.

Psychologism and
Behaviorism


Ned Block, 1981.


Intelligence can’t be based
only

on behavior


TT does not demonstrate general capacity of
machine for producing reasonable responses


Even a
mindless

machine can pass TT:


Have all possible conversations of given length in memory


Machine just looks up correct response


Clearly not intelligent


For intelligence, need capacity and
compactness
:


No exponential blow
-
up in storage

TT Variations


Harnad’s Total Turing Test: same as TT, but
machine has to respond to
all
inputs, not just
verbal.


Needs robot with sensorimotor capabilities


Watts’s Inverted Turing Test: roles reversed.
Computer shouldn’t be able to distinguish its own
outputs from those of a human.


Schweizer’s Truly Total Turing Test: machines
shouldn’t just be able to converse or play chess, but
develop

language and
invent
chess.


Subject Matter Expert Turing Test: test only in some
field.

TT as Interactive Proof


Shieber’s argument in favor of TT, against Block:


Block: “Intelligence is capacity to produce sensible verbal
responses to verbal stimuli without exponential storage”


Shieber: TT does test for this!


Conventional proof: prover P sends proof of
assertion to verifier V, who verifies correctness.


IP adds interaction and randomness


Interaction: many rounds of message
-
passing


Randomness: V may use random bits in message


Also, V approves, but possibly with some small chance of
error

Interactive Proof of

Capacity


TT as proof of capacity


“capacity to produce sensible responses to stimuli”


Consider space of sequences of verbal stimuli


A machine/person has capacity if it answers
correctly on, say, 50% of space


Now run k tests, and say subject passes 75% of
time


Chernoff: subject has capacity with probability exp(O(k))

Interactive Proof of

Compactness


TT as proof of compactness


“without exponential blow
-
up in storage”


How can TT test for storage of machine?


Assumption: machine can’t store more than amount of
information in world.


Upper bound:


Store bit at finest granularity: 10
-
35

meters


Volume of universe: 10
79

meters


Total storage capacity of universe: < 10
200


So computer needs capacity to pass a TT of log 10
200

= 670 bits = 140 words < 1 minute of conversation!

Current TT Achievements


No computer has passed TT to date.


AOLiza: modern spin
-
off of ELIZA


IRC chatbots: people have conversations online with
machines, can’t distinguish


Loebner Prize: annual TT competition, prizes for passing TT
and for “most human conversation”.


No prize awarded for former yet.


Longest Bets Foundation has $10,000 bet b/w Mitch Kapor
and Ray Kurzweil:


Will a computer pass a TT by 2029?


CAPTCHA’s:
c
ompletely
a
utomated
p
ublic
T
uring

test to tell
c
omputers and
h
umans
a
part

Conversations with a
Computer

PROGRAM: So what do you do for a living?

JUDGE: I think. I do research. I write. I love. I
live.

PROGRAM: I wonder whether you think
sometimes! I am feeling like the judge here!

JUDGE: Huh?

Conversations with a
Computer

PROGRAM: Is love the most important thing to
you? How important is that to you?

JUDGE: essential cure to my existential
schizophrenia.

PROGRAM: A striking example of the essential
identity of the two tongues. Existential
schizophrenia, that’s amusing.

So do computers think?


Interview with Gary Kasparov’s advisor
after loss to Deep Blue:


Q: “Did Gary Kasparov think the
computer was thinking?”


A: “Not thinking but that it was showing
intelligent behavior…it understands
strategy…”

What now?

“We can only see a short
distance ahead, but we
can see plenty there
needs to be done.”

-
Alan M. Turing