can machines think? seminar by annervaz (07305063) jaideep (06305r01) k.v.m.v kiran (04005031) ...
TRANSCRIPT
CAN MACHINES THINK?
Seminar by Annervaz (07305063) Jaideep (06305R01) K.V.M.V Kiran (04005031) L. Srikanth (04005029) Vasudevan (06405004)
“I propose to consider the question, 'Can machines think?' ”
-Alan Turing
Motivation
Computers today capable of doing task which were previously thought to be exclusively in human domain
Any limitation to machine's capability?
Can machine do everything that human brain can do?
Machine “Intelligence” <=> Human Intelligence
Hot debate in AI community
Outline
Introduction
Can Machines Think?
Imitation game – Turing test
Proponents and opponents of TT
Introduction to interactive proof
TT as Interactive proof
Conclusion
“can” “machines” “think”
“can”
Theoretically possible?
Practically realizable?
“machines”
Every engineering technique permitted
Manner of operation need not be completely known
“think”
“thinking is as thinking does”
Intelligence necessary and sufficient for thinking
Intelligence (contd...)
“Intelligence” - Based on action:
Rational action: Act to achieve best (possible) outcome,
given what one knows
Human actions:Turing Test designed to measure this
Intelligence...(contd)
Based on thinking: Rational thinking:Irrefutable reasoning process; argument
structures that always yield correct conclusions given correct premises
Human thinking: studied in cognitive science which attempts to combine AI and experimental techniques from psychology
Artificial Intelligence: when an entity other than a natural life form possesses
“intelligence”
Outline
Introduction
Can Machines Think?
Imitation game – Turing test
Proponents and opponents of TT
Introduction to interactive proof
TT as Interactive proof
conclusion
Can Machines Think?Mathematical Objection
Gödel's Theorem
Gödel's Theorem
• Any sufficiently powerful consistent axiomatic system is necessarily incomplete in that there will always be statements that can neither be proved nor disproved from their axioms.
• Corollary:- No sufficiently powerful formal system is powerful enough to prove its own consistency
Gödel's Theorem – Proof Outline
The key idea is to construct a proposition P which asserts “P is not provable”
P is neither provable nor disprovable, as both leads to
contradiction
If P were false, P would be provable
If P were true, P is not provable
Implications of Gödel's Theorem
• Roger Penrose’s claim: “Human mathematicians are not using a knowably
sound algorithm in order to ascertain mathematical truth”.
If they were, it would constitute an algorithm which
can assert it’s own soundness which Gödel’s theorem proves is impossible.
Implications of Gödel's Theorem
• People seem to simply be able to “see” the truth of some statements (“intuition” and “insight”)
• Provability is a weaker notion than truth
• Lucas Argument
• “Some Gödel statements, the machine will be unable to produce as true, although a mind can see that it is true. And so the machine will not be an adequate model of the mind ”
Counter arguments
• Even humans fallible to Gödel's theorem
• “Lucas cannot prove this statement”
• Every human other than Lucas can see this statement is true
• But Lucas can never prove this, from just being ‘inside’ his ‘system’.
• Mind may be exhibiting ‘rational’ inconsistency,and thereby being a formal system consistent with Gödel's theorem
Can Machines Think?Consciousness Objection
Arguments
no mechanism can feel: Anger, grief, warmth, pleasure
Counter Arguments
What is feeling?
Machine can have its own set of feelings.
Aren't external actions enough?
Can Machines Think?Lady Lovelace Objection
Arguments
Machine is deterministic
Machine cannot originate anything
Counter Arguments Learning Machines
Low level determinism does not imply high level
predictable behavior
Can Machines Think?Continuity vs. discreteness
Arguments
Nervous system is continuous
Discrete state system like digital computer can't simulate
nervous system
Counter arguments:
Why can real thought be located only in continous-state
system?
Discrete-state system may still be intelligent
Church-Turing Hypothesis
Can Machines Think?Theological Objection
Arguments
Thinking is a part of Human soul
Man is made in gods own image
Intelligent machine can be a threat to humans
Counter Arguments
Why gods image cannot be passed on?
Why be so pessimistic and selfish?
Outline
Introduction
Can Machines think?
Imitation game – Turing test
Proponents and opponents of TT
Introduction to interactive proof
TT as Interactive proof
conclusion
Imitation game – Turing test
Objective: Interrogator determines which is man and
which is woman
B: tries to help C
A: tries and cause C to make the wrong
Imitation game – Turing test
Replace A with machine
Will the interrogator decide wrongly as often as earlier?
This replaces 'Can Machines think?'
Practical version: “Will an average interrogator have
more than, say 70% chance, of making the right
identification after 5min of questioning?”
Turing Test
Test of adequacy of an agent's verbal behavior
TT is based on the idea that ability to produce sensible
verbal response is intelligence
Tests the “human action” part of intelligence
Can be theoretically formulated as below-
Turing Test(contd..)
Premise 1:If an agent passes a TT, then it produces a sensible sequence of verbal responses to a sequence of verbal stimuli.
Premise 2: If an agent produces a sensible sequence
of verbal response to a sequence of verbal stimuli, then
it is intelligent.
Conclusion:Therefore, if an agent passes TT then it is
intelligent.
Outline
Introduction
Machines can't think
Counter arguments - 'machines can think !'
Imitation game – Turing test
Proponents and opponents of TT
Introduction to interactive proof
TT as Interactive proof
conclusion
Opponents of Turing Test
It is widely agreed that TT is not a necessary condition for intelligence.
Any machine would require sensory organs and
sociological training to pass TT: very difficult task even
for an “intelligent” machine.
TT is of little value in guiding actual AI research
Total test: should involve responding to all inputs, not
just verbal
Block's Argument
TT merely tests behavior
TT is silent about internal working.
Memorizing machine: machine that stores sensible
responses to all possible sequences of verbal inputs
Practically infeasible but possible in principle
Such a machine does not fit into our concept of “intelligence”
So TT is not a sufficient condition for intelligence
Some extra-conditions on the working of the machine are
required
Proponents-Stuart's argument
TT a sufficient condition for intelligence!
logically necessary for intelligence
Extra conditions can be revealed by a TT
“Slight weakening” of proof criterion required
Weakening: Statistical Proof instead of a logical one
Weakening makes no conceivable difference from a
practical standpoint
TT Rephrased
Premise 1:
If an agent passes k rounds of a TT of at least one
minute in length, then (with a prob. of error exponentially
small in k) it has a capacity to produce a sensible
sequence of verbal responses to a sequence of verbal
stimuli that is logarithmic in the storage capacity of the
agent, whatever they may be.
TT Rephrased (continued...)
Premise 2:
If an agent has the capacity to produce a sensible
sequence of verbal responses to a sequence of verbal
stimuli that is logarithmic in the storage capacity of the
agent, whatever they may be, then it is intelligent.
Conclusion:
If an agent passes k rounds of a TT of at least 1 minute
in length (with probability of error exponentially small in k)
it is intelligent
Outline
Introduction
Machines can't think
Counter arguments - 'machines can think !'
Imitation game – Turing test
Proponents and opponents of TT
Introduction to interactive proof
TT as Interactive proof
conclusion
Interactive proof System
An interactive proof system is an abstract machine that models computation as the exchange of messages between two parties
Prover, P having unlimited computation power
Verifier, V with polynomially bounded computation power
Assertion 's'
Randomization and Interaction
Multiple rounds of message-passing
Interactive Proof Examplegraph non-isomorphism
Graph Isomorphism – if there is any edge between any two vertices in first graph, then there should be an edge in between corresponding vertices in other graph
Given graphs G0 and G
1
P's assertion s
G0 and G
1 are NOT isomorphic
Interactive Proof example(s:Graphs G
0&G
1 are Not Isomorphic)
P(bit – b')[infinite resources]
b'=0 If Isomorphic(G0,G') b'=1 otherwise Send b' to V
V(bit – b)[limited resources]
b = rand(0,1) [Prob=0.5] G' = random permute(G
b)
Send G' to P
s: proved, if b==b'
Interactive Proof Example(continued...)
V selects Gb randomly, b=0 or 1
V does a random permutation G' of Gb
V sends G' to P
P checks if G' isomorphic to G0.
If so,sends back b' = “0”, else b' = “1”
V checks if b=b'. If so, V accepts proof- assertion proved
else V rejects proof
Interactive Proof ExampleGraph Non-Isomorphism
If G0 and G
1 are isomorphic, then clue provides no help in guessing
the number Prover guesses randomly, being wrong about half the time Probability of false positive after k rounds is 1 in 2k.
B Truth value of assertion B' Conclusion about assertion
0 True(not isomorphic) 0 True(not isomorphic)
1 True(not isomorphic) 1 True(not isomorphic)
0 False(isomorphic) 0 True(not isomorphic)
1 False(isomorphic) 0 False(isomorphic)
Outline
Introduction
Machines can't think
Counter arguments - 'machines can think !'
Imitation game – Turing test
Proponents and opponents of TT
Introduction to interactive proof
TT as Interactive proof
conclusion
TT As Interactive Proof
capacity conception If an agent has the capacity to produce a sensible
sequence of verbal responses to a sequence of verbal stimuli, whatever they may be, then it is intelligent
generalizability compactness conception
If an agent has the capacity to produce a sensible sequence of verbal responses to a sequence of verbal stimuli, whatever they may be, without requiring storage exponential in the length of the sequence,then it is intelligent
logarithmic storage
TT As Interactive Proofcapacity
P: Computer
V: Interrogator
assertion s: “P has the capacity to produce sensible
sequence of verbal responses to a sequence of verbal
stimuli, whatever they may be”
TT As Interactive Proofcapacity
space: all possible verbal stimuli sequences
tp: fraction of space for which P can perform correctly
tl: lower bound of tp for acceptance
if tp>tl then P has general capacity
TT As Interactive Proofcapacity
select sample (size K) uniformly
t: fraction of sample for which P can perform correctly
ts: lower bound of t for passing
TT As Interactive Proofcapacity
false positive t
p<t
l and t>t
s
Pr[t>ts] < e-ck, using Chernoff bounds
Pr(false positive) decrease exponentially with k
Choice of ts,t
l does not change basic natue of argument
Similarly, Pr(false negative) decreases exponentially with
k
TT As Interactive Proofcompactness
length of sequence is greater than logarithm of storage capacity
By Quantum theory and bounded volume of universe,
information capacity of universe estimated to be 10185
Turing Test of less than 1 min enough to judge
Above “Critical TT length” too short? Counter-intuitive?
TT As Interactive Proofcompactness
Reasons for short TT length requirement: TT unrestricted; all queries of any sort on any topic
allowed. Machine we wish to unmask is of a particular sort- one
that has memorized answers to every possible such query
In IP samples are independent. Here, judge free to use knowledge from previous responses Only reduces probability of error
Properties of IP and TT
non transferability
provide proof only to verifier
lack of closure under composition
fails under composition
What we Proved
Premise 1:
If an agent passes k rounds of a TT of at least one
minute in length, then (with a prob. of error exponentially
small in k) it has a capacity to produce a sensible
sequence of verbal responses to a sequence of verbal
stimuli that is logarithmic in the storage capacity of the
agent, whatever they may be.
What we proved
Premise 2(modified compactness conception-based):
If an agent has the capacity to produce a sensible
sequence of verbal responses to a sequence of verbal
stimuli that is logarithmic in the storage capacity of the
agent, whatever they may be, then it is intelligent.
Conclusion: If an agent passes k rounds of a TT of at least 1 minute
in length (with probability of error exponentially small in k) it is intelligent
Conclusions
Varied definitions of intelligence, AI Not much consensus among experts on how to
determine or measure it Numerous attempts made to define and measure-
scientists,philosophers, engineers One such attempt, TT quite popular and proved with
slight weakening to be sufficient for intelligence. AI has influenced and has been influenced by various
other fields. The debate has both inspired as well as distracted from
research of practical value
References
Block, N. 1981. Psychologism and behaviorism. Philosophical Review XC(1):5-43.
Dennett, D. C. (1985) Can machines think? In: How we know, ed. M.
Shafto, Harper and Row.
Harnad, Stevan. (2006) The Annotation Game: On Turing (1950) on
Computing, Machinery, and Intelligence.
Hofstadter, Douglas R. (1999), Gödel, Escher, Bach, Basic Books.
James H. Moor. An analysis of the Turing test. Philosophical Studies,
30:249-257, 1976.
References
Stalker, D. (1978), ‘Why Machines Can’t Think: A Reply to
James Moor’, Philosophical Studies 34, pp. 317-320.pp. 317–320.
Shieber, S. M. To appear. The Turing test as interactive proof,
Nous.
Stuart M. Shieber 2006. Does the Turing Test Demonstrate
Intelligence or Not? In Proceedings of the Twenty-First National
Conference on Artificial Intelligence (AAAI-2006), Boston,
MA,16-20 July.
Turing, A. (1950), ‘Computing Machinery and Intelligence’, Mind
59(236), pp. 433–460.