Critiques of the Turing Test

hesitantdoubtfulΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

55 εμφανίσεις


Prominent AI
Reseacher


Colleague of Alan Turing at Bletchley Park


1992 Paper:


Turing’s Test and Conscious Thought


Provides a critique of the test



Solipsism and the “Charmed Circle”



“…Turing underestimated the appeal of a more
subtle form of solipsism generalized to groups.”



The argument can be stated as: “the only way by
which one could be sure that a machine thinks is to
be a member of a charmed circle which has
accepted that machine into its ranks and can
collectively feel itself thinking.”


Subarticulate

Thought


“The test can only detect only those processes that
are susceptible to introspective verbal report.”


Many thought processes that cannot be articulated by
humans


A machine might be able to articulate them , even
when a human cannot.


Most highly developed mental skills are of the
verbally inaccessible kind (Hutchins)


“Expert Systems” famously failed in knowledge
extraction through dialog
-
acquisition.





Consciousness and Human
-
Computer
Interaction


What story is assigned to a sequence of events?


Cutaneous Rabbit


5 taps on the wrist


2 near the elbow


3 at the upper arm


Chinese Room


Consider a program that can appear intelligent in
conversation in Chinese


Suppose that someone who doesn’t speak Chinese
executes the program “by hand”


The non
-
Chinese speaker does not understand the
conversation, just as a computer does not
understand the conversation.


A successful Turing Test could be
accomplished through table lookup (given a
large enough memory)


Is this really intelligence?


Turing’s test might not be passed in the
foreseeable future, but that doesn’t really
matter.


Let machines make progress without the
requirement that they imitate people


Computers will provide their own
contributions without the need for imitation.


Weak AI


How the task is accomplished doesn’t matter


We can use a mechanism vastly different than what
humans do


Success is based strictly on performance


Strong AI


Tasks should use the same mechanisms used by
humans


We want to duplicate human intelligence


We want machines to be conscious of what they are
doing


Defined by a set of problems that are
generally considered to require intelligence in
humans


Knowledge Processing


Natural Language Understanding


Game Players


Diagnostic/Classification Problems


Machine Learning




“Rules of Thumb”


Methods that tend to work, but don’t guarantee
success.


Find a simpler problem you know how to solve and try
to generalize to the larger problem


Work backwards from the goal state



In the 1970’s and 1980’s many people
believed “expert systems” would replace
many if not most experts


“Knowledge Engineers” were tasked with
extracting and encoding knowledge from
experts.


It didn’t work very well, largely because much
if not most expertise is
subarticulate
.



Puzzle solving


Finding the best of a set of possible
permutations


Chess


Checkers


Go


Chinese Chess


Dots and Boxes


Given a set of facts, deduce “useful”
conclusions


Representation of facts


Method used for deduction


Identification of “useful” facts



If (some criteria) then some fact


If (some criteria) then perform some action


Expert Systems were often produced using
production rules.


Simplified model of basic building blocks of
the brain


Much smaller number of neurons


Much simpler model of how neurons work


Neural Networks are used in many pattern
matching/classification/generalization
problems.


Simulate evolution


Natural selection used as a form of search


Genetic Algorithms


A population of simulated genes evolves in an attempt
to solve a problem


Genetic Programming


A population of programs evolves in an attempt to
solve
a problem