Download Turing*s Legacy - Cognitive Science Department

Turing’s Legacy
Minds & Machines
“I believe that in about fifty years’ time it will
be possible to programme computers, with a
storage capacity of about 109, to make them play
the imitation game so well that an average
interrogator will not have more than 70 per cent
chance of making the right identification after
5 minutes of questioning”
-Alan Turing (1950)
Alan Turing
• Alan Turing was a British mathematician who was most
famous for his work in theoretical computer science;
• During World War II, Turing helped break German codes
using mechanical computers
• In 1952, he British government thought Turing’s
homosexuality was a crime, and forced him to go
through hormonal treatment.
• In 1954, age 41, Turing died from eating an apple laced
with cyanide; probably a suicide
• In 1999, Turing was listed as one on the top 100 most
important people of the 20th century
• On September 10, 2009, the British government
apologized for their treatment of Alan Turing.
Turing’s Legacy
• Turing’s legacy consists of 2 parts:
– Turing Machines (1936)
– Turing Test (1950)
Turing Test
• “I propose to consider the question, 'Can
machines think?' This should begin with
definitions of the meaning of the terms
'machine 'and 'think'. … [But] Instead of
attempting such a definition I shall replace
the question by another... The new form of
the problem can be described in terms of a
game which we call the 'imitation game'.“
-Alan Turing, “Computing Machinery and Intelligence”, 1950
The Imitation Game
Machine
Interrogator
Human
Some Initial Observations on the
Turing Test
• The Turing Test attributes intelligence purely on
verbal interactions. Is that ok?
• Well, physical characteristics (size, weight,
agility, etc) don’t seem to be relevant as far as
intelligence goes, so that seems right.
• However, shouldn’t we have to open up the
computer program and see how it works to make
this kind of determination?
• Then again, do we ever open up other human
beings to determine whether they are intelligent?
• Hmm, maybe Turing has a point.
The Turing Test:
Can Machines Think?
Premise 1: Machines can pass the Turing Test
Premise 2: Anything that passes the Turing Test
is intelligent
Conclusion: Machines can be intelligent
Can Machines pass the
Turing Test?
Computationalism
• Cognition can be defined in terms of informationprocessing:
–
–
–
–
–
–
Perception is taking in information
Memory/Beliefs/Knowledge is storing information
Reasoning is inferring new information
Learning is updating information
Planning is using information to make decisions
Etc.
• Information-processing can be done through
computations
• Therefore, cognition is computation.
Computationalism and the Brain
• Notice that the argument on the previous
slide is a purely conceptual one in that it is
not based on any empirical evidence.
• Indeed, it predicts the existence of some
kind of brain (computer) in any cognitive
being.
• So, the fact that we have a brain, which is
in many ways a computer, can be seen as
empirical confirmation of the view of
computationalism.
Computationalism and the Brain,
Part I
• The brain fits with computationalism:
– The brain is unlike any other organ; the heart, lungs,
liver, etc. all do something very much physical
(collect, filter, pump, etc.)
– The brain, however, is quite different: Its function
seems to be to take in signals, and send out signals,
in communication with the nervous system.
– Thus, the brain seems to be an informationprocessor: a computer of sorts.
– Indeed, we know that the nature of the mind changes
when the brain changes: thus, maybe:
• brain = ‘hardware’
• mind = ‘software’
Computers
• A ‘computer’ is something that computes, i.e.
something that performs a computation.
• Between the 17th and 20th century, a ‘computer’
was understood to be a human being; humans
who computed things!
• It was only by automating (mechanizing) this
process, that we obtained ‘computers’ as we
now think of them.
Computations
• A computation is a symbol-manipulation
algorithm.
– The symbols represent something
– Hence, the computation is about that
something: “we compute something”
Example: Long Division
Components for Computation
• In a famous 1936 paper, Turing argued
that all computations can be reduced to
the following basic components:
– One symbol string of arbitrary size
– An ability to move along this symbol string
– An ability to read and write symbols
• We now call this: a Turing-machine
Turing Machines Demo
Computable Functions
• We can use a Turing-machine to compute the
sum, and product, of any two numbers.
• These functions are therefore Turingcomputable
• Lots of other functions are Turing-computable
• E.g. all functions needed to run Microsoft Word
are Turing-computable (i.e. you can run
Microsoft Word on a Turing-machine)
The Church-Turing Thesis
• If a computer of type X can compute a
function f, we say that f is X-computable
• The Church-Turing Thesis:
– No matter what type of computer X you have:
All functions that are X-computable are
Turing-computable.
• In short: Turing-machines can compute
anything that is computable.
Universal Turing Machines
Turing proved that there exists a Turing-machine
that can simulate any other Turing-machine
TM, I
UTM
Description of
machine TM
and input I
The Universal Turing Machine
TM(I)
The output that
Machine TM would
Give if I would be its
input
Computationalism and the Brain,
Part II
• There are reasons to believe that our brain is
‘programmable’ too:
– Neural connections get created and are being erased
in the brain
– More or less neural resources can be devoted to
certain tasks
– In case of neural defects, other areas of the brain can
take over
• In short, the brain is highly ‘plastic’, and it may
be as a result of this that we learn to become
more adept at tasks, or learn to do completely
new tasks.
0’s and 1’s
• Turing showed how all computation can be done
using a limited number of simple processes
manipulating a small number of symbols.
• In fact, it turns out you only need 2 symbols!
• You do need lots of these symbols, and you do
need to perform lots of these simple operations.
• But this is exactly how the modern ‘digital
computer’ does things. That is, at the ‘machine
level’, it’s all simple manipulations of 0’s and 1’s.
Physical Dichotomies
• The 0’s and 1’s are just abstractions
though; they need to be physically
implemented.
• Thus, you need some kind of physical
dichotomy, e.g. hole in punch card or not,
voltage high or low, quantum spin up or
down, penny on piece of toilet paper or
not, etc.
Computationalism and the Brain,
Part III
• Again, the brain fits with what we saw:
– Lots of simple devices, all organized together
to perform lots of simple operations
• Our brain has 1011 neurons, and 1014 neural
connections
• Early views on the brain supposed that neurons
firing or not would constitute 0’s and 1’s.
Causal Topology
• A physical system implements a computational
system if and only if that system implements a
certain causal topology.
• This topology is highly abstract. As long as you
retain the functionality of the parts, and the
connections between the parts, you can:
– Move parts
– Stretch parts
– Replace parts
• This is why there can be mechanical computers,
electronic computers, DNA computers, optical
computers, quantum computers, etc!
Computationalism and the Brain,
Part IV
• So are our brains organic, carbon-based,
‘meat-computers’?!
• Again, it seems to fit:
– Implements a complex causal topology
where, what the only thing that seems to
matter is how the neurons are connected.
Summary
• Two independent arguments for
computationalism:
– One conceptual: cognition is informationprocessing, and that’s exactly what computers
do
– One empirical: the mind seems dependent on
the brain, where the brain seems to be:
• an information-processing device
• that is ‘programmable’
• and is made of large numbers of simple devices
that are ‘wired’ to support complex informationprocessing capacities
Back to the the Turing Test:
Can Machines Think?
Premise 1: Machines can pass the Turing Test
Premise 2: Anything that passes the Turing Test
is intelligent
Conclusion: Machines can be intelligent
Cheap Tricks? Eliza
• A psychotherapist program developed by
Joseph Weizenbaum in 1966.
• Eliza used a number of simple strategies:
– Keywords and pre-canned responses
• “Perhaps I could learn to get along with my
mother”
-> “Can you tell me more about your family?”
– Parroting
• “My boyfriend made me come here”
-> “Your boyfriend made you come here?”
– Highly general questions
• “In what way?”
• “Can you give a specific example?”
Eliza and the Turing Test
• Many people conversing with Eliza had no idea
that they weren’t talking to a human.
• So did Eliza pass the Turing Test?
• (Or is it just easy being a psychotherapist?!)
• Eliza wasn’t really tested in the format that
Turing proposed.
• Still, it is interesting that humans were quick to
attribute human-level intelligence to such a
simple program.
• Maybe in a real Turing Test a relatively simple
computer program can ‘trick’ the interrogator as
well?
A Definition of Intelligence?
• Many people have similar criticisms with
regard to the Turing Test as a test or
definition of intelligence:
• In particular, the Turing Test is real sloppy:
– Who is the interrogator?
– How long is the conversation?
– What is the conversation about?
– How does the interrogator decide?
Not a Definition
• Turing himself clearly did not intend to propose a
definition of intelligence.
• In his paper Turing readily acknowledges that
one could have intelligent beings not being able
to pass the test simply by not having a humanlike intellect:
– “May not machines carry out something which ought
to be described as thinking but which is very different
from what a man does? This objection is a very
strong one, but at least we can say that if,
nevertheless, a machine can be constructed to play
the imitation game satisfactorily, we need not be
troubled by this objection”
A Sufficient Condition for
Intelligence?
• Many commentators interpret Turing’s statement
as saying that if a machine passes the Turing
Test, then it is intelligent, i.e. that passing the
Turing Test is a sufficient condition for
intelligence (since intelligence is a necessary
condition to pass it), but not a necessary one
(and hence it is not a definition).
• In logic:
– We have: P  I
– But not: I  P
Same Sloppiness …
And A Question
• As a sufficient condition for being
intelligent, the Turing Test suffers from
some of the same problems as before:
– such a criterion would still amount to a
subjective judgment based on imprecisely
defined behavioral criteria.
But why would Turing (not exactly known for
his sloppiness!) propose such a sloppy
test?
The Loebner Competition
• Modern day version of the Turing Test
• Multiple judges rank-order multiple humans and
multiple computer programs from ‘most likely to
be human’ to ‘least likely to be human’.
• Loebner has promised $100,000 for the first
computer program to be ‘indistinguishable from
a human’.
• Thus far, Loebner is still a rich man: occasionally
a judge will rank a program above a human, but
on the whole the judges systematically rank the
humans above the computer programs.
An OK Test After All?
• Apparently it is quite difficult to pass the test!
– When put to the real test, interrogators can see
through superficial trickery
• So it seems we could say that if something does
pass the test, then there is at least a good
chance for it to be intelligent.
• In fact, if we are turning this into an inductive
argument anyway, the sloppiness of the test isn’t
a huge concern either: we can now simply adjust
our confidence in our claim in accordance to the
nature of the conversation.
• So is this maybe what Turing was saying?
“Contrary Views”
• In his paper Turing goes over a list of
“Contrary Views on the Main Question”:
• Machines:
– can’t make mistakes
– can’t be creative
– can’t learn
– can’t do other than what they’re told
Machines Can’t Make Mistakes,
Be Creative, Learn
• “Can’t make mistakes”: A weird objection
to being intelligent …
• And machine do make mistakes!
– Not at the level of the program
– But at the level of interpretation
• And they are creative
• And do learn
Machines Can Only Do What
They’re told To Do
• The mistake in this objection is that while this statement
is true from the perspective of the underlying program, it
is not clear that a machine couldn’t do anything new or
creative when looked at from a higher level.
• Indeed, look at Deep Blue: Deep Blue beats every
human in chess, but that would be impossible if Deep
Blue couldn’t do any better than any of its programmers.
• Similarly, pending any mechanical glitches, machines
can’t fail but follow some program, and thus indeed not
make any errors from that point of view. However, it can
still make mistakes as the result of following that
program!
• (…. no quantum randomness needed!)
Turing’s Argument
• It looks like Turing makes the following
argument to support his view that
computers can pass the Turing Test:
– Passing the Turing Test requires nothing
more than some kind of information
processing ability
– Computers can have this information
processing ability
– Therefore, computers can pass the Turing
Test
The Earlier Argument for
Computationalism
• But compare this argument to the earlier
argument for Computationalism:
– Cognition requires nothing more than some
kind of information processing ability
– Computers can have this information
processing ability
– Therefore, computers can be cognitive
The Real Puzzle
• So why not change the argument into:
– Intelligence requires nothing more than some
kind of information processing ability,
– Computers can have this information
processing ability
– Therefore, computers can be intelligent
• So why didn’t Turing make this very argument?
Why bring in the game to make the case for
machine intelligence?
A Second Puzzle
• Also, why the strange set-up of the TuringTest? Why did Turing ‘pit’ a machine
against a human in some kind of contest?
Why not have the interrogator simply
interact with a machine and judge whether
or not the machine is intelligent based on
those interactions?
The Super-Simplified Turing Test
Interrogator
Machine
Answer: Bias
• The mere knowledge that we are dealing
with a machine will bias our judgment as
to whether that machine can think or not,
as we may bring certain preconceptions
about machines to the table.
• Moreover, knowing that we are dealing
with a machine will most likely lead us to
raise the bar for intelligence: it can’t write a
sonnet? Ha, I knew it!
• By shielding the interrogator from the
interrogated, such a bias and bar-raising is
eliminated in the Turing-Test.
The Simplified Turing Test
Interrogator
Machine or Human
Level the Playing Field
• Since we know we might be dealing with a
machine, we still raise the bar for the entity
on the other side being intelligent.
• Through his set-up of the test, Turing
made sure that the bar for being intelligent
wouldn’t be raised any higher for
machines than we do for fellow humans.
• Still, we are left with the earlier puzzle:
why bring up any kind of test (and a sloppy
one at that!) at all?
My Answer
• I propose that the convoluted set-up wasn’t
merely a practical consideration to eliminate bias
in some strange game, but rather the point of his
article, which is that if we put a label ‘intelligent
being’ on other human beings based on their
behavior then, just to be fair, we should do the
same for machines, whether we are correct or
precise in any such attributions or not.
• In other words, Turing’s point was that we don’t
have a precise definition of ‘intelligence’, but that
we do have a fuzzy concept of it, and that our
use of slapping this label onto things (human or
otherwise) should at least be consistent.
‘Imitation Game’ vs ‘Turing Test’
• In other words, I think it is likely that Turing
never intended to propose any kind of test
for machine intelligence (let alone propose
a definition!).
• Interesting fact: In his original article
Turing uses the word ‘pass’ or ‘passing’ 0
times, ‘test’ 4 times, and ‘game’ 37 times.
The Turing ‘Test’ as Harmful!
• Moreover, I believe that seeing Turing’s contribution as
laying out a test is harmful.
• The harm is that we have been thinking about the goal of
AI in these terms, and that has been, and still is,
detrimental to the field of AI.
• E.g. In “Essentials of Artificial Intelligence”, Ginsberg
defines AI as “the enterprise of constructing a physical
symbol system that can reliably pass the Turing Test”
• But trying to pass the test encourages building cheap
tricks to convince the interrogator, which is exactly what
we have seen with Eliza, Parry, and pretty much any
entry in the Loebner competition.
• This kind of work has advanced the field of AI, and our
understanding of intelligence … exactly zilch!
• So, I think we really should no longer refer to the Turing
Test as the Turing ‘Test’!!
Grand Challenges
• Maybe the Turing Test (and the Loebner
competition) is a kind of Grand Challenge?
– Landing people on moon
– Chess (Deep Blue)
– Urban Challenge
– Jeopardy (Watson)
• But at this point in time, I feel that trying to
create human-level intelligence in a
computer is a ridiculously-grand challenge,
and hence a ridiculous Grand Challenge
How to Read Turing’s Paper
• So what did Turing really mean? Ultimately, this is an
issue of history, and not an issue we should be
concerned about.
• Better questions to ask are: What, if anything, can we
learn from Turing’s paper? What would be a fruitful
interpretation of his paper?
• Well, there are many interesting parts of the paper,
especially in Turing’s responses to the ‘Contrary Views’.
• I also believe that seeing Turing’s paper as laying out a
genuine test is harmful, not helpful.
• Instead, I believe a fruitful reading of his paper is to see
the Turing ‘Test’ as a statement about the use of the
word ‘intelligence’.
Pluto and Planets
• Asking how many planets there are in our solar
system seems to be a factual matter:
– We believe there is a straightforward fact of the
matter to this issue.
• If I say: “There are X planets in our solar system” then this
statement is either true or false.
– How many planets there are is an empirical issue:
observations will tell us how many there are
• However, as the case of Pluto demonstrated,
things aren’t that easy. This issue isn’t just an
empirical issue, but also one of interpretation.
• Maybe the same is true for machine intelligence!
Artificial Flight and
Artificial Intelligence
• Imagine going back 100 years when the Wright
Brothers had their first flight.
• We can imagine people say: “Well, but that’s not
real flight. There is no flapping of the wings!”
• But over time, we realized that it is, from the
standpoint of using concepts that help us think
and make sense of the world around us, a good
idea to consider airplanes as really flying.
• Again, maybe the same is true for intelligence!
The original question, “Can machines think?”, I believe
to be too meaningless to deserve discussion. Nevertheless
I believe that at the end of the century the use of words and
general educated opinion will have altered so much that one
will be able to speak of machines thinking without expecting
to be contradicted.
-Alan Turing (1950)