Founding
Editor: F. de Hanika . Editor: Stephen Sokoloff . International Federation tor Systems Research
Schottengasse 3, A-1010 Wien, Austria·
Electronic Mail: K 323390 @ aearn, Bitnet
Artificial Intelligence is becoming more Brainlike
Charles François
Asociacion Argentina de Teoria General de Sistemas y Cibernetica.
(Reflections
on a lecture by Brother J. M. Ramlot O. P.)
Computers are to an increasing extent
taking on problems previously reserved for humans. Nowadays many of them can
deal not only with data but also with knowledge. The gap between artificial
and natural intelligence is gradually closing; therefore we have to constantly
reevaluate the differences between both forms.
Artificial intelligence can compete
successfully with natural intelligence 1) in resolving equations and 2) in
demonstrating and even in finding previously unknown demonstrations for mathematical
theorems. The algorithm ic and sequential capabilities
needed to perform these tasks can be found in both human brains and computers -
but the computer does this kind ofwork much faster.
Nevertheless the machine still needs an operator to put the data into
it, or at least to equip it with a collecting device. Besides, it generally
isn't able to discern whether the data is correct.
Also the operational algorithm, simple or
complex, must be introduced by a mind, and the machine cannot produce anything
otherthan what is contained in its programo Therefore artificial intelligence
is stilllargely dependent upon natural intelligence.
For some years now, however, we've seen
the emergence of expert systems, which handle not only data, but also
knowledge. This knowledge is kept apart from the rules governing its handling,
which constitute a second systems level, a sort of "metaknowledge"
in the form of algorithmic combinations.
E:xpert systems
have already produced spectacular results. A system of this type can include
the total knowledge of many human experts, and its algorithmic 'operation sometimes
leads to the discovery of things which all of the specialists- even the
designers of the system - had failed to notice. Of course the operations are
carried out much more rapidly and exhaustively in artificial than in natural intelligence
systems.
Iy creative and of inventing something which is not implicit in its
algorithm. The computer's Iimits are obvious:
- it doesn't "know" (isn't conscious of) what it
knows (what's in its memory and its algorithm). (At least, we think it doesn't
know.)
- it can't breakoutof itssequentiality. The parallel multisequentiality
of some systems is only a first approximation to the truly simultaneous
parallelism and the selfinterconnection of natural intelligence.
- it's a prisoner
of binary logic, which implies a particular type of reductionism.
However artificial intelligence may cross a
new threshold in the coming years, since we are now gaining a better understanding
of some properties of the brain. The following ones seem particularly
interesting:
- It has a fantastic combinatory capacity.
- It begins practically virgin, which is why we remember al-
most nothing about the first years of
Iife.
- Making use of varied mean so flearníng,
it must organize itself.
-It must acquire
perceptual and conceptual selectivity without having to actively register
everything in its surroundings.
-It is capable of forgetting.
These characteristics lead us to some
surprising conclusions when we take Ashby's ideas about variety and constrictions
into consideration:
- The brain receives and processes many
observations simultaneously.
- The brain needs
to form algorithms in order to be able to function usefully in its
super-complex and constantly changing environment.
- The "algorithmization" can be
transferred from brain to brain, and is called instruction/education.
-Learning leads to
the formation of mental algorithms by trial and error. Any algorithm can arise
as a result of the repeated strengthening of correct answers; it therefore
represents a set of organizing constrictions.
-Once established,
the algorithm replaces chance behavior with determined behavior. This
determination is, however, never absolute, probably beca use of the great potential
ofthe algorithms acquired in childhood and youth.
- The algorithmic
functions acquired by the brain tend to partly block creative capacity. This
phenomenon undergoes rapid enhancement during adolescence and youth - with few
exceptions, older people are no longer creative.
- The ability to
forget seems to be indispensible for relieving the brain of useless data: The
mechanism for forgetting is, however, by no means completely understood.
On the basis of these characteristics we
can conceive of a type of artificial intelligence which could go much further
than the most advanced machines currently available and come to much more
closely resemble natural intelligence. The prerequisites for this type of
system are:
-the ability to
"Iearn" not only the contents of data but also behavior patterns.
-the simultaneous functioning of many units.
-the formation of interconnections, at first by
chance, bet-
ween these units.
-the progressive
dynamic stabilization of certain of these i ntercon nections (ultrastabiIity).
-the maintenance of a great deal of variety,
partly by random utilization of the algorithms whichare formed.
-an ability to
selectively destroy certain portions of a memory.
The essentlal dlfferences between current
artificial and natural intelligence systems lie in the following characteristics
of the latter:
- Their multi-simultaneous mode of
operation prevents them from becoming absolutely determined.
- The entirealgorithm
acquired bythe brain isso large and complex that no human can possibly use all
its potential contents.
- The
"algorithmization" is never complete. A virgin reserve (potential
for variety) is always left over. Thus there is
always a margin of
imprevisibility as far as the future behavior of a natural intelligence system
is concerned.