Download PHILOSOPHY OF ARTIFICIAL INTELLIGENCE Artificial intelligence

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Computer vision wikipedia , lookup

Human-Computer Interaction Institute wikipedia , lookup

AI winter wikipedia , lookup

Wizard of Oz experiment wikipedia , lookup

Computer Go wikipedia , lookup

Intelligence explosion wikipedia , lookup

Turing test wikipedia , lookup

Embodied cognitive science wikipedia , lookup

Human–computer interaction wikipedia , lookup

Visual Turing Test wikipedia , lookup

Existential risk from artificial general intelligence wikipedia , lookup

Chinese room wikipedia , lookup

Ethics of artificial intelligence wikipedia , lookup

History of artificial intelligence wikipedia , lookup

Philosophy of artificial intelligence wikipedia , lookup

Transcript
Artificial intelligence
• “Artificial Intelligence is the science of
making machines do things that would
require intelligence if done by men”
PHILOSOPHY OF ARTIFICIAL
INTELLIGENCE
Marvin Minsky
Prof.Dr John-Jules Meyer
– Weak AI thesis
– Strong AI thesis
Dr Menno Lievers
2
Weak AI Thesis
Strong AI Thesis
• The computer is (only) a powerful aiding
tool for the study of the human mind
• An adequately programmed computer has a
cognitive state - computer programs explain
human cognition
• It is possible to devise machines that behave
like people and possess human capabilities,
such as the ability to think, reason, ..., play
chess, walk, ..., have emotions, pain, ...
• It is possible to construct machines that
perform useful “intelligent”
tasks
assisting human users
– possible??
– desirable?!
– Difficult enough?!
3
4
Can a machine think?
The Turing Test
• Try to first answer the question ‘in principle’,
independent of available technology
• Is consciousness necessary for thinking?
• You may replace ‘thinking’ or ‘being
intelligent’ by ’displaying cognitive activity’
• A human A communicates by email
with a human B and a computer C
• A poses questions to both B and C to
discover which is the human
• If A doesn’t succeed to distinguish B
and C, the computer C passes the
Turing Test
5
6
– Human mental processes are often non-conscious
• 'sleeping problem solver'
• 'blindsight'
1
Has the Turing Test been passed
already?
The Turing Test Set-Up
C
• Turing test: based on link between ’thinking' en
'conversation'
• Two famous ‘conversation’ programs:
B
– ELIZA
– PARRY
•
ELIZA and PARRY are based on relatively simple
pattern matching algoritms: this is not thinking…?!!
A
7
8
Objections against the Turing
Test
Objections against the Turing
Test
1. Chimpanzee objection: chimpanzees, dolphins, ...
will not pass the Turing Test, while they are
obviously intelligent and able to think! So a negative
result does not say anything about being able to
think / being intelligent.
2. sensory versus verbal communication: the TT only
concerns verbal communication: no test of the
computer’s ability to relate words to things in the
world.
9
Conclusion?!
simulation objection: simulated X ≠ X. This objection
says that thinking cannot ever be simulated
perfectly
4. Black Box objection: the external behaviours are
equal does not imply that the processes are
themselves equal!
3.

SUPERPARRY: program containing all conversations of
length ≤100 words: is finite in principle and programmable;
will pass the Turing test; however, does not think !?!
10
Can we improve the Turing Test ?
• In any case we need the following
criteria:
– Output criterion: competition between two
‘agents’
– Design criterion: it is not about the humanlike way of thinking, think also of
hypothetical aliens (or animals…)
11
12
2
What is thinking / intelligence?
Symbol System Hypothesis
• thinking is an intentional notion, it has goal/actiondirected; it has to do with explaining and predicting of
behaviour −−−> planning, being flexible, adaptable
• Generalise this notion: it is about being 'massively
adaptable' → this notion is applicable to nontraditional
matters
such
as
extraterrestrial
intelligence, animals, computers / machines (artificial
intelligence)
∴ "robots are able to think" may then be a sensible
statement
• thinking = 'being massively adaptable'
• Is this achievable using digital computers?
– I.o.w. if we can make machines ‘think’, is a digital computer
the right kind of machine?
• symbol system hypothesis (SSH): yes!:
– a universal symbol system (= general-purpose storedprogram computer): symbol manipulator operating by
executing fundamental operations, such as branch, delete,
output, input, compare, shift, write, copy is a 'massively
adaptable' machine
13
Intelligent systems
• An intelligent ('massively adaptable’) system
should be able to:
–
–
–
–
–
–
–
–
14
GOFAI recipe for an IS
(IS)
1. Use a sufficiently expressive, inductively defined,
compositional language to represent 'real-world'
objects, events, actions, relations, etc.
2. Construct an adequate representation of the
world and the processes in it in a universal
symbol system (USS) : extensive Knowledge
Base (KB)
3. Use suitable input devices to obtain symbolic
representation of environmental stimuli
15
16
Generate plans
Analyze situations
Deliberate decisions
Reason and revise 'beliefs'
Use analogies
Weigh conflicts of interest, preferences
Decide rationally on the basis of imperfect information
Learn, categorize
GOFAI recipe for an IS
Employ complex sequences of the fundamental
operations of the USS to be applied to the
symbol structures of the inputs and the KB,
yielding new symbol structures (some of these
are designated as output)
5. This output is a symbolic representation of
response to the input. A suitable robot body can
be used to ‘translate’ the symbols into real
behaviour / action
4.
17
• The SSH says:
• In this way a thinking (= massively adaptable)
machine is obtained!
18
3
Doubts about the SSH
Status SSH
• the SSH is an interesting conjecture, that may
appear strange, but may be true after all
(there are more strange things that are held
to be true: e.g. relativity theory, quantum
mechanics...); however:
• How can such a machine really understand?
• Or wonder whether a sentence is true?
• or desire something?
– Is there any evidence by the state of the art in AI?:
• Not (yet): all AI at the moment is rather limited; the
original GPS project has more or less failed, and modern
AI is not yet sufficiently convincing(?!)
• ... Etc.
– Philosophical (analytical) considerations (Searle)
19
20
Strong Symbol System
Hypothesis (SSSH)
Philosophical objections against
Strong AI & SSH: Searle
• SSH:
computers
(i.e..
univ.
symbol
manipulators) can think
• SSSH: ONLY computers (univ. symbol
manipulators) can think, i.e. the only things
capable of thinking are univ. symb. manip.;
ergo, the human mind is a univ. symb.
manip., a computer!!!
• Is the question whether a computer is suitable device
for thinking an empirical one?
• Searle: the question whether a symbol manipulating
device can think is not empirical, but analytical, and
can be answered negatively :
– The SSSH is even more controversial than the
SSH.
– a universal symbol manipulator (USS) operates purely
syntactically and is not able to really understand what it is
doing!
– syntax is insufficient for dealing with semantics (=
"understanding of what symbols actually mean")
21
Searle’s Gedankenexperiment
22
The Chinese room
• John Searle tries to argue by means of a
Gedankenexperiment that a computer
cannot think, or more precisely, cannot
perform an intelligent task, such as e.g.
answer questions in Chinese about a
Chinese text, and really understand what
it is doing.
ダソまめキずそぜゑわボ
Text with
questions
in Chinese
Sam
Answers in
Chinese
Suppose we have a computer program Sam
capable to answer questions in Chinese about
Chinese texts
23
24
4
The Chinese room
The Chinese room
• Chinese room argument:
– Joe in the room executing the computer program
Sam manually, does not understand the story nor
the questions, nor the answers: only manipulation
of meaningless symbols: "Sam 'run' on a human
computer"
– Executing the program does not enable Joe to
understand the story, questions, etc., ergo
executing the program does not enable the
computer to understand the story, questions etc. !
ダソまめキずそぜゑわボ
Text with
questions
in Chinese
Answers in
Chinese
Joe
Replace computer program Sam by
human Joe executing the program instructions
25
26
Chinese room: Searle’s
conclusion
But …?!?
• running a program does not lead to
understanding, believing, intending, thinking
…!
• But… cannot we ‘prove’ in the same way
that humans (i.e. our brains) cannot
think …?!?
– Let the global population (5 billion people) simulate a brain B
with its 100 billion neurons: then each person controls some
20 neurons
– No person knows what B is thinking…
– So, neither do(es) (the neurons in) brain B.
• "merely manipulating symbols will not enable
the manipulating device to understand X,
believe Y, think Z..."
∴ the SSH is FALSE !
27
28
The “Systems Reply”
Counter-objection
• 'The systems reply': Not only the symbol
manipulator Joe is concerned but the system
as a whole: it could be possible that the whole
system does understand!
•
Searle contra de systems reply:
1. Joe does not understand, but Joe + paper +
pencil would understand ?!? (cynically)
2. Let Joe learn all rules of the program by heart;
then there is no ‘bigger’ system any more of
which Joe is part; in fact everything is part of Joe
in that case!
29
30
5
The Chinese room revisited
And the debate goes on…
• Searle:
− SSH ⇒ 'toilet paper' machine (= TM)
thinks as well ?!?!
− biological objection to the SSH and AI
ダソまめキずそぜゑわボ
Text with
questions
In Chinese
Answers in
Chinese
• Copeland:
Joe
− although Joe may say of himself that he
does not understand, an external observer
may still say that Joe does understand!!!
31
The Great Debates in AI
32
Can computers think?
• Can computers think?
• Can the Turing Test determine whether computers
can think?
• Can physical symbol systems think?
• Can Chinese Rooms think?
• Can connectionist networks think? Can computers
think in images?
• Do computers have to be conscious to think?
• Are thinking computers mathematically possible?
• Is the brain a computer?
• Can computers have free will?
• Can computers have emotions?
• Can computers be creative?
• Should we pretend computers will
never be able to think?
35
36
Can the TT determine whether
computers can think?
Can computers think?
• If a simulated intelligence passes,
is it intelligent?
• Does the imitation game determine
whether computers can think?
• Is passing / failing the test decisive?
• Have any machines passed the test?
• Is the test a legitimate intelligence
test?
• Does God prohibit computers from
thinking?
• Can computers understand arithmetic?
• Can computers draw analogies?
• Are computers inherently disabled?
• Can computers reason scientifically?
• Can computers be persons?
37
38
6
Can Physical Symbol Systems
Think?
Can Physical Symbol Systems
Think?
• Can the elements of thinking be
represented in symbolic form?
• Can physical symbol systems learn as
humans do?
• Do humans use rules as physical symbol
systems do?
• Can a symbolic knowledge base represent
human understanding?
• Can symbolic representations account for
human thought?
• Does thinking require a body?
• Can physical symbol systems think
dialectically?
• Is the relation between hardware and
software similar to that between human
brains and minds?
• Does mental processing rely on heuristic
search?
• Do physical symbol systems play chess as
humans do?
39
40
Can Connectionist Networks
Think?
Can Chinese Rooms Think?
• Can the Chinese Room, considered as a
total system, think?
• Can an internalized Chinese Room
think?
• Can brain simulators think?
• Can robots think?
• Do Chinese Rooms instantiate programs?
• Can computers cross the syntax-semantics
barrier?
• Are connectionist networks
vulnerable to the arguments
against physical symbol systems?
• Do connectionist networks follow
rules?
• Does the subsymbolic account offer
a valid account of connectionism?
41
Can Computers Think in Images?
• Can images be realistically represented
in computer arrays?
• Can computers recognize Gestalts?
• Are images less fundamental than
propositions?
• Is image psychology a valid approach to
mental processing?
• Can computers represent the analogue
properties of images?
42
Do Computers Have to Be
Conscious to Think?
• Can computers be conscious?
• Is consciousness necessary for thought?
• Is the consciousness requirement
solipsistic?
• Can functional states generate
consciousness?
• Can higher-order representations
produce consciousness?
43
44
7
Are Thinking Computers
Mathematically Possible?
• Can automata think?
• Does Gödel’s theorem show that machines
can’t think / can’t be conscious?
• Does Gödel’s theorem show that
mathematical insight is nonalgorithmic?
• Do mathematical theorems like Gödel’s
show that computers are intrinsically
limited?
45
8