[comp.ai.philosophy] Just Minds and Machines this time

cpshelley@violet.uwaterloo.ca (cameron shelley) (01/25/91)

In article <11656.9101241836@s4.sys.uea.ac.uk> jrk@information-systems.east-anglia.ac.uk (Richard Kennaway CMP RA) writes:
>

>--
>Richard Kennaway          SYS, University of East Anglia, Norwich, U.K.
>Internet:  jrk@sys.uea.ac.uk		uucp:  ...mcsun!ukc!uea-sys!jrk
>
>Relevance-to-newsgroup detector now registering approx. 0.0...
                                                         ^^^^

I concur!  Comparisons of pragmatic, existential, positivist, etc...
notions of truth are interesting but becoming both entrenched and
somewhat beside whatever it was that started this.  There are groups
such as sci.logic and talk.philosophy (?) which are better suited and
may arouse a larger volume of relevant comment.  Or maybe e-mail would
be sufficient.

While I'm here (:-), I wouldn't mind drawing some comment on a subject
I've been thinking about recently.  In reviewing some short papers on
some new neural net work, I was struck by the notion of 'error' being
employed.  One of the praises always sung of NN's is that they are
"robust", ie. no matter what input you give them, they won't simply
fail like symbolic programs, but will rather try to compensate and
produce a meaningful output.  In other words (IMHO), all errors are
treated as noise and an attempt is made to ignore them.  There is
essentially no notion of 'ill-formed' input as opposed to 'ill-transmitted'
input.  I argued (in the review) that this is epistemologically
inadequate, at least as a model of human cognition, since humans
show the ability to do recovery from both types of error (in different
fashion).  The cause, I believe, is that while the representation
is dynamically induced (the subject knowledge) the meta-knowledge
(or domain knowledge) is fixed by the structure of the NN so that
it cannot attempt more than one method of solving and therefore has
no redundancy -- a common error-handling technique in both machines
and people.

My intial suggestion was that a system should be created in which
two (or more) NN's with different structures be allowed to compete
for one output.  The question then is: is my analysis correct, and
if so, is the suggestion e-adequate both computationally and 
cognitively?

				Cam

Btw: I finished the review; you are not doing my work for me!

--
      Cameron Shelley        | "Absurdity, n.  A statement of belief
cpshelley@violet.waterloo.edu|  manifestly inconsistent with one's own
    Davis Centre Rm 2136     |  opinion."
 Phone (519) 885-1211 x3390  |				Ambrose Bierce

smoliar@isi.edu (Stephen Smoliar) (01/27/91)

In article <1991Jan25.022026.12999@watdragon.waterloo.edu>
cpshelley@violet.uwaterloo.ca (cameron shelley) writes:
>  In reviewing some short papers on
>some new neural net work, I was struck by the notion of 'error' being
>employed.  One of the praises always sung of NN's is that they are
>"robust", ie. no matter what input you give them, they won't simply
>fail like symbolic programs, but will rather try to compensate and
>produce a meaningful output.  In other words (IMHO), all errors are
>treated as noise and an attempt is made to ignore them.  There is
>essentially no notion of 'ill-formed' input as opposed to 'ill-transmitted'
>input.  I argued (in the review) that this is epistemologically
>inadequate, at least as a model of human cognition, since humans
>show the ability to do recovery from both types of error (in different
>fashion).

I want to pick up on this and perhaps try to assimilate it with Stuart
Hampshire's recent article on Wittgenstein (actually his review of the
new Monk biography) in THE NEW YORK REVIEW.  Hampshire offers what I feel
is an excellent summary of the essence of the TRACTATUS:

	According to the TRACTATUS the multiplicity of elements in
	a sentence ought to be a picture of the multiplicity of
	elements in a state of affairs.  The formal correspondence
	secures for us the reference to a particular point in reality.

To some extent, the idea that it makes sense to talk about such "a particular
point in reality" at all was one of the major positions Wittgenstein chose to
contest in his later work;  but I find it interesting that much of
connectionism almost seems to have translated this idea of "a particular
point in reality" into a point in some multidimensional energy landscape.
What counts as robustness is nothing more than the observation that these
landscapes tend to be sufficiently smooth that perturbation of a starting
point will not severely disrupt its trajectory along this landscape.  Cam's
observation that this is but one way of viewing what "error" might be is well
taken;  and perhaps it obliges us to go back and think some more about
appropriate metaphors for these points in "reality."

Let us suppose that we still have some sort of multidimensional space as a
metaphor for reality;  but rather than filling it with an energy landscape,
suppose we instead insert a linkage structure, sort of like a linear undirected
graph, as a model of an agent's "knowledge" (whatever that may mean).  The
reason I wish to appear to this metaphor is as an alternative to modeling
reasoning in terms of the trajectory of a point in the space which seeks out
an energy sink.  Think, instead, of accommodating a point in space by asking
how that linkage structure might get flexed, or, perhaps expanded, in such a
way that it ultimately "meets" that point;  that resulting "meeting" might then
be regarded as an "interpretation" of that point.  Given sufficiently liberal
laws of what you could do with the linkages, any given point would obviously
be subject to multiple interpretations, which would mean it is not really a
fixed "point in reality" (which, as I assumed above, was one of the problems
Wittgenstein was trying to get away from).  From Cam's point of view, the
question of whether or not the point is "well-formed" ultimately boils down
to whether or not the linkages can be configured to "meet" it.

I realize this is all rather loose metaphor.  (I also remember what John
McCarthy recently had to say about metaphors.)  However, since the purpose
of this bulletin board is to kick around philosophical approaches, I regard this
as yet another pebble to toss in the pond.  Anyone who wishes to make waves
is certainly welcome to do so.
-- 
USPS:	Stephen Smoliar
	5000 Centinela Avenue  #129
	Los Angeles, California  90066
Internet:  smoliar@venera.isi.edu

cpshelley@violet.uwaterloo.ca (cameron shelley) (01/28/91)

In article <16510@venera.isi.edu> smoliar@venera.isi.edu (Stephen Smoliar) writes:
[...]
>
>I want to pick up on this and perhaps try to assimilate it with Stuart
>Hampshire's recent article on Wittgenstein (actually his review of the
>new Monk biography) in THE NEW YORK REVIEW.  Hampshire offers what I feel
>is an excellent summary of the essence of the TRACTATUS:
>
>	According to the TRACTATUS the multiplicity of elements in
>	a sentence ought to be a picture of the multiplicity of
>	elements in a state of affairs.  The formal correspondence
>	secures for us the reference to a particular point in reality.
>
>To some extent, the idea that it makes sense to talk about such "a particular
>point in reality" at all was one of the major positions Wittgenstein chose to
>contest in his later work;  but I find it interesting that much of
>connectionism almost seems to have translated this idea of "a particular
>point in reality" into a point in some multidimensional energy landscape.
>What counts as robustness is nothing more than the observation that these
>landscapes tend to be sufficiently smooth that perturbation of a starting
>point will not severely disrupt its trajectory along this landscape.  Cam's
>observation that this is but one way of viewing what "error" might be is well
>taken;  and perhaps it obliges us to go back and think some more about
>appropriate metaphors for these points in "reality."
>
I'd like to think so.  The predominant model for connectionist architectures
seems to be thermodynamics, at least in networks that use relaxation
methods such as "simulated annealing" in which training is thought of
as lowering the 'temperature' of a system so that its 'energy' becomes
minimized.  Of course, in classical physics, nature does not make
'mistakes' (ie. be "ill-formed"), whereas it can at least be 'deformed'
in modern physics.  

>Let us suppose that we still have some sort of multidimensional space as a
>metaphor for reality;  but rather than filling it with an energy landscape,
>suppose we instead insert a linkage structure, sort of like a linear undirected
>graph, as a model of an agent's "knowledge" (whatever that may mean).  The
>reason I wish to appear to this metaphor is as an alternative to modeling
>reasoning in terms of the trajectory of a point in the space which seeks out
>an energy sink.  Think, instead, of accommodating a point in space by asking
>how that linkage structure might get flexed, or, perhaps expanded, in such a
>way that it ultimately "meets" that point;  that resulting "meeting" might then
>be regarded as an "interpretation" of that point.  Given sufficiently liberal
>laws of what you could do with the linkages, any given point would obviously
>be subject to multiple interpretations, which would mean it is not really a
>fixed "point in reality" (which, as I assumed above, was one of the problems
>Wittgenstein was trying to get away from).  From Cam's point of view, the
>question of whether or not the point is "well-formed" ultimately boils down
>to whether or not the linkages can be configured to "meet" it.
>
Hmmm.  An interesting notion!  Could you elabourate on "meet"?  If I
understand correctly, you mean the structure can be deformed by some
composition of adjustments applied to its parts.  This seems plausible
with the proviso that the entire structure need not be completely
connected (or am I missing something?).  The notion "ill-formed" would
be the requirement of an 'illegal' translation of the structure, while
"ill-transmitted" would be uncertainty about the point being met.
Would we have to know all possible translations of the structure in 
order to separate 'legal' from 'illegal'?  Complexity could then be
measured by the number of elementary translations and deformations
necessary to meet the point in question.  By then, the default (or
acquisition of) the structure becomes an issue...

				Cam

--
      Cameron Shelley        | "Absurdity, n.  A statement of belief
cpshelley@violet.waterloo.edu|  manifestly inconsistent with one's own
    Davis Centre Rm 2136     |  opinion."
 Phone (519) 885-1211 x3390  |				Ambrose Bierce

smoliar@isi.edu (Stephen Smoliar) (01/29/91)

In article <1991Jan27.185935.18038@watdragon.waterloo.edu>
cpshelley@violet.uwaterloo.ca (cameron shelley) writes:
>In article <16510@venera.isi.edu> smoliar@venera.isi.edu (Stephen Smoliar)
writes:
>
>>Let us suppose that we still have some sort of multidimensional space as a
>>metaphor for reality;  but rather than filling it with an energy landscape,
>>suppose we instead insert a linkage structure, sort of like a linear
>>undirected
>>graph, as a model of an agent's "knowledge" (whatever that may mean).  The
>>reason I wish to appeal to this metaphor is as an alternative to modeling
>>reasoning in terms of the trajectory of a point in the space which seeks out
>>an energy sink.  Think, instead, of accommodating a point in space by asking
>>how that linkage structure might get flexed, or, perhaps expanded, in such a
>>way that it ultimately "meets" that point;  that resulting "meeting" might
>>then
>>be regarded as an "interpretation" of that point.  Given sufficiently liberal
>>laws of what you could do with the linkages, any given point would obviously
>>be subject to multiple interpretations, which would mean it is not really a
>>fixed "point in reality" (which, as I assumed above, was one of the problems
>>Wittgenstein was trying to get away from).  From Cam's point of view, the
>>question of whether or not the point is "well-formed" ultimately boils down
>>to whether or not the linkages can be configured to "meet" it.
>>
>Hmmm.  An interesting notion!  Could you elabourate on "meet"?  If I
>understand correctly, you mean the structure can be deformed by some
>composition of adjustments applied to its parts.  This seems plausible
>with the proviso that the entire structure need not be completely
>connected (or am I missing something?).  The notion "ill-formed" would
>be the requirement of an 'illegal' translation of the structure, while
>"ill-transmitted" would be uncertainty about the point being met.
>Would we have to know all possible translations of the structure in 
>order to separate 'legal' from 'illegal'?  Complexity could then be
>measured by the number of elementary translations and deformations
>necessary to meet the point in question.  By then, the default (or
>acquisition of) the structure becomes an issue...
>
I think we're basically on the same channel here.  I was using "meet" in the
set-theoretic sense of intersection.  The more I think about it, I think there
are (at least) two implications to this approach:

	1.  There is the issue of whether or not, and how, the linkages can
		be manipulated in order to achieve a meeting.

	2.  There is the issue of WHERE on the linkages this meeting actually
		takes place.

In a very loose sense the first issue has to do with whether or not that
particular point in space is well-formed, in the notion that Cam wishes
to pursue.  Under the assumption that it IS well-formed, the second issue
then takes on the matter of HOW it will be interpreted.  This is what I was
trying to get at in saying that different ways of manipulating the linkages
might lead to different interpretations:  Depending on the specific
manipulations, the point may meet different locations on the linkage,
itself.

Does the structure have to be connected?  I do not see that as necessary.  I am
more interested in the extent to which this metaphor can be useful interpreted
into a set of viable rules for manipulation the structure.  There is no reason
to eliminate rules which would pull the structure apart unless it could be
demonstrated that they led to undesirable consequences (such as, perhaps,
an inability to recover information about where the meeting took place when
it was finally achieved).  My own sense of aesthetics seems more inclined to
allowing the "limbs" of the structure to be stretched or shrunken, rather than
allowing the structure to be pulled apart;  but, as I said, I think it is more
important to start thinking about reasonable ways to talk about the
manipulation, itself.

Another possibility might be that the manipulations are sufficiently powerful
that the structure can always access any point in the space.  In this case
there would no longer be an issue of ill-formed points.  This might be a way
in which the manipulation of the structure reflects an adjustment in
interpretation to accommodate what might have otherwise been regarded
as an ill-formed point.

Perhaps I can try to illustrate what I am trying to say here with a concrete
example.  Peter Todd recently published some of his work in trying to use
connectionism for musical composition.  The basic approach was to "train"
a network with some examples of melodies and then, through control of some
inputs, allow a trained network to synthesize new melodies.  During training,
Todd has some very strict rules about what constitutes well-formed input.
However, when he leaves the network to its own devices, so to speak, the
results it yields do not respect those rules.  As an outside observer, he
interprets them in a way which basically "makes sense" according to the
original INTENT of his rules.  To return to my metaphor, in the strictest
sense Todd's network converges to points in space which are actually
ill-formed.  However, his human intelligence controls this metaphorical
system of linkages in such a way that he can still assign interpretations
to those points which are consistent with what he originally had in mind.
Robustness is thus a matter of finding the best interpretation for a given
situation (for some metric for "best") rather than trying to decide, in any
absolute sense, whether or not that situation is "well-formed."
-- 
USPS:	Stephen Smoliar
	5000 Centinela Avenue  #129
	Los Angeles, California  90066
Internet:  smoliar@venera.isi.edu

cpshelley@violet.uwaterloo.ca (cameron shelley) (01/30/91)

In article <16537@venera.isi.edu> smoliar@venera.isi.edu (Stephen Smoliar) writes:
[...]
>I think we're basically on the same channel here.  I was using "meet" in the
>set-theoretic sense of intersection.  The more I think about it, I think there
>are (at least) two implications to this approach:
>
>	1.  There is the issue of whether or not, and how, the linkages can
>		be manipulated in order to achieve a meeting.
>
>	2.  There is the issue of WHERE on the linkages this meeting actually
>		takes place.
>
>In a very loose sense the first issue has to do with whether or not that
>particular point in space is well-formed, in the notion that Cam wishes
>to pursue.  Under the assumption that it IS well-formed, the second issue
>then takes on the matter of HOW it will be interpreted.  This is what I was
>trying to get at in saying that different ways of manipulating the linkages
>might lead to different interpretations:  Depending on the specific
>manipulations, the point may meet different locations on the linkage,
>itself.
>

I don't think that's too controversial.  The existence of multiple
interpretations is a time-honoured way of describing ambiguity: in
this case, there is more than one set of translations that meet the
desired point.  I would then go on to distinguish "ill-transmitted",
in which the desired point has been underdefined resulting in 
multiple points which satisfy the constraints of the transmission
(and each point may be ambiguous in its own right), from "ill-formed"
in which the transmitted constraints themselves cannot be interpreted
(by some sense of adequacy) so that no points at all have been
defined.  Note that the last notion can be quite independant of errors
in transmission.  The concern in my original post was that connectionist
architectures treat "ill-formed" as "ill-transmitted" when I don't
think this does the subject justice.  No doubt this is for the 
convenience of being able to use results from information theory
in all cases.  In dealing with interpretation in general, I think
such (0,1,plural) boundary cases are most profitably treated as
separate until evidence indicates otherwise.

Another area in which (to my knowledge) the '0' case is not treated is
theoretical phonology.  Although the 'null' symbol turns up in phonological
rules (deletion), it is never defined.  This also means that the theory
assumes there is some certain method of distinguishing the properties of
speech sound from those of other sound before phonology is invoked.
But I digress... (sorry!)

>Does the structure have to be connected?  I do not see that as necessary.  I am
>more interested in the extent to which this metaphor can be useful interpreted
>into a set of viable rules for manipulation the structure.  There is no reason
>to eliminate rules which would pull the structure apart unless it could be
>demonstrated that they led to undesirable consequences (such as, perhaps,
>an inability to recover information about where the meeting took place when
>it was finally achieved).  My own sense of aesthetics seems more inclined to
>allowing the "limbs" of the structure to be stretched or shrunken, rather than
>allowing the structure to be pulled apart;  but, as I said, I think it is more
>important to start thinking about reasonable ways to talk about the
>manipulation, itself.
>

Well, empirically, it is possible to lose information, and I don't see
any advantage in requiring the entire 'knowledge' structure to be
connected.  However, this may depend on how 'knowledge' is defined.

>Another possibility might be that the manipulations are sufficiently powerful
>that the structure can always access any point in the space.  In this case
>there would no longer be an issue of ill-formed points.  This might be a way
>in which the manipulation of the structure reflects an adjustment in
>interpretation to accommodate what might have otherwise been regarded
>as an ill-formed point.
>

Again, I see this as ignoring a potential problem.  Under-specification
may be treated in this fashion, but I still see 'non-specification' as
something that must be dealt with.  Perhaps even a further distinction
is necessary:  a system shouldn't waste time interpreting random noise
('non-transmission'), but should simply baulk at nonsense ('ill-formed'
transmission), and deal as mentioned above with noise in a sensible
transmission.  For example, the system should deal differently with:
"%^%#*%^&%^%^%%^$#@!#$#" (noise), "After being run *^@r by the truck,
the man said 'Ouch'!" (where "over" can be interpolated), "flying
airplanes make people ill" (ambiguity), and "pick peek poke pack puck pink"
(nonsense).  Should I interpret the last as a description of a
psychodelic hockey game, or just respond "What?"  I think both should
be options, but NN's (as they exist currently) do not have a choice
in such a case.

>Perhaps I can try to illustrate what I am trying to say here with a concrete
>example.  Peter Todd recently published some of his work in trying to use
>connectionism for musical composition.  The basic approach was to "train"
>a network with some examples of melodies and then, through control of some
>inputs, allow a trained network to synthesize new melodies.  During training,
>Todd has some very strict rules about what constitutes well-formed input.
>However, when he leaves the network to its own devices, so to speak, the
>results it yields do not respect those rules.  As an outside observer, he
>interprets them in a way which basically "makes sense" according to the
>original INTENT of his rules.  To return to my metaphor, in the strictest
>sense Todd's network converges to points in space which are actually
>ill-formed.  However, his human intelligence controls this metaphorical
>system of linkages in such a way that he can still assign interpretations
>to those points which are consistent with what he originally had in mind.
>Robustness is thus a matter of finding the best interpretation for a given
>situation (for some metric for "best") rather than trying to decide, in any
>absolute sense, whether or not that situation is "well-formed."

This suggests that the notion of "ill-formed" is context-dependant.  My
response in this case would be that Peter's experiment has defined away
the notion of ill-formed at the start, in the sense that the structure
of the net has been controlled so that it can only produce results which
have possible 'melodic' interpretations.  In any case, this seems to be
bringing the mountain to Mohammed; simply deforming the topology of
the knowledge structure to a point doesn't appear to me to be much
different than moving a point across the topology.  But your point is
well taken, some notion of 'systematic accomodation' seems useful.

				Cam

--
      Cameron Shelley        | "Absurdity, n.  A statement of belief
cpshelley@violet.waterloo.edu|  manifestly inconsistent with one's own
    Davis Centre Rm 2136     |  opinion."
 Phone (519) 885-1211 x3390  |				Ambrose Bierce

smoliar@isi.edu (Stephen Smoliar) (01/31/91)

In article <1991Jan29.165646.17764@watdragon.waterloo.edu>
cpshelley@violet.uwaterloo.ca (cameron shelley) writes:
>  For example, the system should deal differently with:
>"%^%#*%^&%^%^%%^$#@!#$#" (noise), "After being run *^@r by the truck,
>the man said 'Ouch'!" (where "over" can be interpolated), "flying
>airplanes make people ill" (ambiguity), and "pick peek poke pack puck pink"
>(nonsense).  Should I interpret the last as a description of a
>psychodelic hockey game, or just respond "What?"  I think both should
>be options, but NN's (as they exist currently) do not have a choice
>in such a case.
>
Cam, you interpreted my account of Peter Todd's work as suggesting "that the
notion of 'ill-formed' is context-dependent."  I would say that the above
nonsense example supports a similar conclusion:  How you choose to interpret
depends upon the context in which you received it.  I would even say the same
of your noise example.  If you saw those symbols in some avant-garde poetry
magazine, you might very well try to kick in SOME attempt at interpretation,
rather than just writing them off as noise (even if they had been scrupulously
generated by a truly random source).

Nevertheless, I think we are basically in agreement.  Todd's system "works"
because its architecture assumes a single context.  I have not seen any
convincing demonstration of a neural net which would be capable of maintaining
multiple contexts, such as your nonsense example requires, let alone having a
level of control for DECIDING which context is appropriate for a given
interpretation task.
-- 
USPS:	Stephen Smoliar
	5000 Centinela Avenue  #129
	Los Angeles, California  90066
Internet:  smoliar@venera.isi.edu

cpshelley@violet.uwaterloo.ca (cameron shelley) (01/31/91)

In article <16558@venera.isi.edu> smoliar@venera.isi.edu (Stephen Smoliar) writes:
>In article <1991Jan29.165646.17764@watdragon.waterloo.edu>
>cpshelley@violet.uwaterloo.ca (cameron shelley) writes:
>>  For example, the system should deal differently with:
>>"%^%#*%^&%^%^%%^$#@!#$#" (noise), "After being run *^@r by the truck,
>>the man said 'Ouch'!" (where "over" can be interpolated), "flying
>>airplanes make people ill" (ambiguity), and "pick peek poke pack puck pink"
>>(nonsense).  Should I interpret the last as a description of a
>>psychodelic hockey game, or just respond "What?"  I think both should
>>be options, but NN's (as they exist currently) do not have a choice
>>in such a case.
>>
>Cam, you interpreted my account of Peter Todd's work as suggesting "that the
>notion of 'ill-formed' is context-dependent."  I would say that the above
>nonsense example supports a similar conclusion:  How you choose to interpret
>depends upon the context in which you received it.  I would even say the same
>of your noise example.  If you saw those symbols in some avant-garde poetry
>magazine, you might very well try to kick in SOME attempt at interpretation,
>rather than just writing them off as noise (even if they had been scrupulously
>generated by a truly random source).
>

Actually, I was wondering about that after sending that post.  It seems to
me that you're correct.  If I equate 'context' with 'interpretation function'
being applied,  then I suppose you could have a crack at anything.  I am
reminded that in Montague semantics (if I recall correctly), the truth
assignment of a function has a superscript identifying which of the
multiple 'worlds' it is applying to.  The situation seems similar here,
with 'world' playing the role of 'context'.

>Nevertheless, I think we are basically in agreement.  Todd's system "works"
>because its architecture assumes a single context.  I have not seen any
>convincing demonstration of a neural net which would be capable of maintaining
>multiple contexts, such as your nonsense example requires, let alone having a
>level of control for DECIDING which context is appropriate for a given
>interpretation task.

Ok, let me try this out.  To summarize the information distinctions
I made previously (about how well transmitted constraints can be 
satisfied in general), within a fixed context, let @ = noise, 0 = under-
defined point, 1 = uniquely-defined point, * = multiply-defined points.

[Opps!  I should mention that by 'point', I mean a coordinate in a
solution space...]

Thus the options (under a fixed context) that a complete classification
system should have are (@,0,1,*).  I know of no NN which has a node
dedicated to '@', ie. to explicitly recognizing over-noisy input;  I am
also unaware of any which explicitly recognizes '0' (which I am also
calling "ill-formed"), ie. units which are individually recognizable
but form no interpretable whole (regardless of noise); on the other
hand, NN's are quite capable of producing 1 or * (more) interpretations
by activating a single output mainly, or several equally.  This 
implies that current NN design could well be augmented with nodes
on the output indicating '@' and '0' which would be mutually 
inhibitory.  I noticed in the Dec CACM, that Kevin Knight suggested
putting noise in normal training pairs, ie. ([a,p,^,l,e],apple);  what
I'm suggesting is ([^,%,#,&,&],@).  This also leads me to suggest
pairs like ([s,t,k,a],0), bearing in mind that I'm fixing the context
on english words.  You may take issue with the example I've chosen,
but I think the idea is at least sound.

As far as varying the context goes, there are examples of I know of
in which entire sub-structures of a NN are mutually exclusive but
deal with the same input 'item'.  My original mention of two
differently structured nets independantly competing for the same
'output' seems similar.  On the other hand, I have no idea how a
net might be allowed to vary its context freely (or creatively!)
I would be curious to know how the linkage structure you proposed
might account for this.

--
      Cameron Shelley        | "Absurdity, n.  A statement of belief
cpshelley@violet.waterloo.edu|  manifestly inconsistent with one's own
    Davis Centre Rm 2136     |  opinion."
 Phone (519) 885-1211 x3390  |				Ambrose Bierce

smoliar@isi.edu (Stephen Smoliar) (02/01/91)

In article <1991Jan31.040637.15353@watdragon.waterloo.edu>
cpshelley@violet.uwaterloo.ca (cameron shelley) writes:
>
>As far as varying the context goes, there are examples of I know of
>in which entire sub-structures of a NN are mutually exclusive but
>deal with the same input 'item'.  My original mention of two
>differently structured nets independantly competing for the same
>'output' seems similar.  On the other hand, I have no idea how a
>net might be allowed to vary its context freely (or creatively!)
>I would be curious to know how the linkage structure you proposed
>might account for this.
>
So would I!  I seem to have started with a casual metaphor, and now I have to
account for it!  Bearing in mind that I still do not know how seriously I want
to take this metaphor (primarily because I am not sure I want to buy into
viewing knowledge states in terms of points in a space of some unknown
dimension and a topology which is probably not Euclidean), I would say
the way to approach context is in terms of the current configuration of
the linkage system.  As I said before, I tend to view interpretation in
terms of both how the linkages will move to meet a given point and where
that meeting takes place.  How the linkages will be disposed to move (and
I use that word intentionally acknowledging the influence of Minsky's original
K-lines paper) will depend not only on where the point is but also on how the
linkages are currently configured.  That configuration should, in at least a
remote sense, be a representation of previous (mental) activity, which is to
say a context established by what the linkages were last doing.  Varying
context may then be a matter of shaking the linkages around before trying
to approach the point (which is beginning to sound a bit like raising the
temperature before cooling down to anneal).
-- 
USPS:	Stephen Smoliar
	5000 Centinela Avenue  #129
	Los Angeles, California  90066
Internet:  smoliar@venera.isi.edu