Copy Link
Add to Bookmark
Report
AIList Digest Volume 4 Issue 205
AIList Digest Monday, 6 Oct 1986 Volume 4 : Issue 205
Today's Topics:
Humor - AI Limericks by Henry Kautz,
AI Tools - Turbo Prolog & Reference Counts vs Garbage Collection,
Philosophy - Emergent Consciousness & Perception
----------------------------------------------------------------------
Date: Fri, 3 Oct 86 14:58:37 CDT
From: Glenn Veach <veach%ukans.csnet@CSNET-RELAY.ARPA>
Subject: AI Limericks by Henry Kautz
gleaned from the pages of CANADIAN ARTIFICIAL INTELLIGENCE
September 1986 No. 9 page 6:
AI Limericks
by Henry Kautz
University of Rochester
*** ***
If you're dull as a napkin, don't sigh;
Make your name as a "deep" sort of guy.
You just have to crib, see
Any old book by Kripke
And publish in AAAI.
*** ***
A hacker who studied ontology
Was famed for his sense of frivolity.
When his program inferred
That Clyde is a bird
He blamed not his code but zoology.
*** ***
If your thesis is utterly vacuous
Use first-order predicate calculus.
With sufficient formality
The sheerist banality
Will be hailed by the critics: "Miraculous!"
If your thesis is quite indefensible
Reach for semantics intensional.
Your committee will stammer
Over Montague grammar
Not admitting it's incomprehensible.
------------------------------
Date: Fri, 26 Sep 86 11:40 PDT
From: JREECE%sc.intel.com@CSNET-RELAY.ARPA
Subject: Turbo Prolog - Yet Another Opinion
Although Turbo Prolog has been characterized by some wags as a "brain-dead
implementation" I think its mixture of strengths and weaknesses would be more
accurately described as those of an idiot savant. Some of the extensions,
such as the built-in string editor predicates, are positively serendipitous,
and you get most of the development time advantages of a fifth generation
language for a conventional application plus good runtime performance for
only $70. On the other hand, one tires quickly of writing NP-incomplete sets
of type declarations which are unnecessary in any other implementation....
If nothing else, for $70 you can prototype something that can be used to justify
spending $700 for a real PC Prolog compiler, or $18,000 for a VAX
implementation.
John Reece
Intel
------------------------------
Date: Fri, 26 Sep 86 18:52:26 CDT
From: neves@ai.wisc.edu (David M. Neves)
Reply-to: neves@ai.wisc.edu (David M. Neves)
Subject: Re: Xerox vs Symbolics -- Reference counts vs Garbage collection
When I was using MIT Lisp Machines (soon to become Symbolics) years
ago nobody used the garbage collector because it slowed down the
machine and was somewhat buggy. Instead people operated for hours/days
until they ran out of space and then rebooted the machine. The only
time I turned on the garbage collector was to compute 10000 factorial.
Do current Symbolics users use the garbage collector?
"However, it is apparent that reference counters will never
reclaim circular list structure."
This is a common complaint about reference counters. However I don't
believe there is very many circular data structures in real Lisp code.
Has anyone looked into this? Has any Xerox user run out of space
because of circular data structures in their environment?
--
David Neves, Computer Sciences Department, University of Wisconsin-Madison
Usenet: {allegra,heurikon,ihnp4,seismo}!uwvax!neves
Arpanet: neves@rsch.wisc.edu
------------------------------
Date: 26 Sep 86 15:35:00 GMT
From: princeton!siemens!steve@CAIP.RUTGERS.EDU
Subject: Garb Collect Symb vs Xerox
I received mail that apparently also went onto the net, from Dan Hoey
(hoey@nrl-aic.ARPA). He discussed garbage collection in response to
my unsupported allegation that, "S[ymbolics] talks about their garbage
collection more, but X[erox]'s is better." I am glad to see someone
taking up an informed discussion in this area.
First, I briefly recap his letter, eliding (well-put) flames:
+ In the language of computer
+ science, Xerox reclaims storage using a ``reference counter''
+ technique, rather than a ``garbage collector.''
+ If we are to believe Xerox, the reference counter
+ technique is fundamentally faster, and reclaims acceptable amounts of
+ storage. However, it is apparent that reference counters will never
+ reclaim circular list structure. As a frequent user of circular list
+ structure (doubly-linked lists, anyone?), I find the lack tantamount to
+ a failure to reclaim storage.
+ I have never understood why Xerox continues to neglect to write a
+ garbage collector. It is not necessary to stop using reference counts,
+ but simply to have a garbage collector available for those putatively
+ rare occasions when they run out of memory.
+ Dan Hoey
Xerox's system is designed for highly interactive use on a personal
workstation (sound familiar?). They spread the work of storage reclamation
evenly throughout the computation by keeping reference counts. Note that
they have many extra tricks such as "References from the stack are not
counted, but are handled separately at "sweep" time; thus the vast majority
of data manipulations do not cause updates to [the reference counts]"
(Interlisp-D Reference Manual, October, 1985). Even if this scheme were
to use a greater total amount of CPU time than typical garbage collection,
it would remain more acceptable for use on a personal, highly interactive
workstation. I have no idea how it can be compared to Symbolics for overall
performance, without comparing the entire Interlisp vs. Zetalisp systems.
Nevertheless, I can say that my experience is that Interlisp runs a "G.C."
every few seconds and it lasts, subjectively, an eyeblink. Occasionally
I get it to take longer, for example when I zero my pointers to 1500 arrays
in one fell swoop. I have some figures from one application, too. An
old, shoddy implementation ran 113 seconds CPU and 37.5 seconds GC (25% GC).
A decent implementation of the same program, running a similar problem twice,
got 145 seconds CPU, but 10.8 and 20.3 seconds GC (6.9% and 12% GC). (The
good implementation still doesn't have a good hashing function so it's still
slower.) I cannot claim that these figures are representative. I have
heard horror stories about other Lisps' GCs,
although I don't have any feel for Symbolics's "Ephemeral GC".
I have a strong feeling Xerox has other tricks besides the one about the
stack, which they don't want to tell anyone. I know they recently fixed
the reference counter from counting 16 or more references as "infinity"
(and thus never reclaimable) to an overflow scheme where the reference
count gets squirreled away somewhere else when it gets bigger.
Finally, normally the amount of unreclaimed garbage (e.g. circular lists)
grows much slower than memory fragments, so you have to rebuild your
world before unreclaimed garbage becomes a problem anyway.
Postfinally, Xerox makes a big deal that their scheme takes time proportional
to the number of objects reclaimed, while traditional systems take time
proportional to the number of objects allocated. I think Symbolics's
ephemeral scheme is a clever way to consider only subsets of the universe
of allocated objects, that are most likely to have garbage. I wish I knew
whether it is a band-aid or an advance in the state-of-the-art.
Absolutely ultimately, "traditional" GC I refer to, is known as "mark-
and-sweep".
Steve Clark {topaz or ihnp4}!princeton!siemens!steve
------------------------------
Date: Mon, 29 Sep 86 14:07:12 edt
From: segall@caip.rutgers.edu (Ed Segall)
Subject: Re: Emergent Consciousness
Why must we presume that the seat of consciousness must be in the form
of neural "circuits"? What's to prevent it from being a symbolic,
logical entity, rather than a physical entity? After all, the "center
of control" of most computers is some sort of kernal program, running
on the exact same hardware as the other programs. (Don't try to push
the analogy too far, you can probably find a hole in it.) Perhaps the
hierarchical system referred to is also not structural.
Might the brain operate even more like a conventional computer than we
realize, taking the role of an extremely sophisticated
(self-modifying) interpreter? The "program" that is interpreted is the
pattern of firings occurring at any given time. If this is so, then
moment-to-moment thought is almost completely in terms of the dynamic
information contained in neural signals, rather than the quasi-static
information contained in neural interconnections. The neurons simply
serve to "run" the thoughts. This seems obvious to me, since I am
assuming that neural firings can process information much faster than
structural changes in neurons.
I'd be interested to know about what rate neuron firings occur in the
brain, and if anyone has an intelligent guess as to how much
information can be stored at once in the "dynamic" form of firings
rather than the "static" form of interconnections.
I apologize in advance if what I suggest goes against well-understood
knowlege (not theory) of how the brain operates. My information is
from the perspective of a lay person, not a cognitive scientist.
------------------------------
Date: Mon, 29 Sep 86 09:34:01 EDT
From: "Col. G. L. Sicherman" <colonel%buffalo.csnet@CSNET-RELAY.ARPA>
Subject: Re: Consciousness as bureaucracy
Ken Laws' analogy between Bureaucracy and Man--more precisely, Man's
Mind--has been anticipated by Marvin Minsky. I do not have the reference;
I think it was a rather broad article in a general science journal.
As I recall, the theory that Minsky proposed lay somewhere between the
lay concept of self and the Zen concept. It seemed to suggest that
consciousness is an illusion to itself, but a genuine and observable
phenomenon to an outside observer, characterizable with the metaphor
of bureaucracy. Perhaps some Ailist reader can identify the article.
Emergent consciousness has always been a hope of A.I. I side with
those who suggest that consciousness depends on contact with the world
... even though I know some professors who seem to be counter-examples!
:-)
------------------------------
Date: 2 Oct 86 17:14:00 EDT
From: "FENG, THEO-DRIC" <theo@ari-hq1.ARPA>
Reply-to: "FENG, THEO-DRIC" <theo@ari-hq1.ARPA>
Subject: Perception
I just ran across the following report and thought it might contribute some-
thing to the discussion on the "perception" of reality. (I'll try to summarize
the report where I can.)
according to
Thomas Saum in the
German Research Service,
Special Press Reports,
Vol. II, No. 7/86
A group of biologists in Bremen University has been furthering the theory devel-
oped by Maturana and Varela (both from Chile) in the late '70's, that the brain
neither reflects nor reproduces reality. They suggest that the brain creates its
own reality.
Gerhard Roth, a prof. of behavioral physiology at Bremen (with doctorates in
philosophy and biology), has written several essays on the subject. In one, he
...writes that in the past the "aesthesio-psychological perspective"
of the psychomatic problem was commonly held by both laypersons
and scientists. This train of thought claims that the sensory organs
reporduce the world at least partially and convey this image to the
brain, where it is then reassembled ("reconstructed") in a uniform
concept. In other words, this theory maintains that the sense organs
are the brain's gateway to the world.
In order to illustrate clearly the incorrectness of this view,
Roth suggests that the perspectives be exchanged: if one looks at
the problem of perception from the brain's angle, instead of the
sense organs, the brain merely receives uniform and basically homo-
geneous bioelectric signals from the nerve tracks. It is capable of
determining the intensity of the sensory agitation by the frequency
of these signals, but this is all it can do. The signals provide no
information on the quality of the stimulation, for instance, on whe-
ther an object is red or green. Indeed, they do not even say any-
thing about the modality of the stimulus, i.e. whether it is an
optical, acoustical, or chemical stimulation.
The constructivists [as these new theoreticians are labeled],
believe that the brain is a self-contained system. Its only access
to the world consists of the uniform code of the nerve signals which
have nothing in common with the original stimuli. Since the brain
has no original image, it cannot possibly "reporduce" reality; it
has to create it itself. "It (the brain) has to reconstruct the di-
versity of the outside owrld from the uniform language of the neu-
rons", Roth claims. The brain accomplishses this task by "interpret-
ing itself", i.e. by construing what is going on inside itself.
Thus, the brain "draws conclusions" from the degree to which it is
agitated by the modality of the original stimulus: all neuronal im-
pulses reaching the occipital cortex, for example, are visual im-
pressions.
This isolated nature of the brain and its reality, however, are
by no means a blunder on the part of nature; indeed, they are not
even a necessary evil, Roth explains. On the contrary, it is an
adaptive advantage acquired by more highlly developed creatures dur-
ing the course of their phylogenic development. If the brain had di-
rect access to the environment, Roth argues, then one and the same
stimulus would necessarily always result in one and the same reac-
tion by the organizsm. Since, however, the human brain has retained
a certain amount of creative scope for its reconstruction of reality,
it is in a position to master complicated stiuations and adapt itself
to unforeseen circumstances.
Only in this way is it possible to recognize an object in differ-
ent light intensities, from a new angle of vision, or at a distance.
Even experiments with "reversal spectacles" demonstrate man's powers
of adaptation in interpreting reality: after a little while, test
persons, who see the world upside down with special glasses, simply
turn their environment around again in their "mind". When, after a
few days, they remove the spectacles, the "real" world suddenly seems
to be standing on its head.
This mobility and adaptability on the part of our perceptive fa-
culties were obviously much more important for the evolution of more
highly developed vertebrates than was a further intensification of
the signal input by the sense organs. The million fibers in man's
optic nerve are only double the number of a frog's; the human brain,
on the other hand, has one hundred thousand times more nerve cells
than a frog brain. But first and foremost, the "reality workshop",
i.e., the cerebral area not tied to specific sense, has expanded
during the evolution of man's brain, apparently to the benefit of
our species.
Contact: Prof. Dr. Gerhard Roth, Forschungsschwerpunkt Biosystemforschung,
Universitat [note: umlaut the 'a'] Bremen, Postfach 330 440,
D-2800 Bremen 33, West Germany.
[conveyed by Theo@ARI]
------------------------------
End of AIList Digest
********************
