NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/bit/listserv/csgl/1389

Internet Message Format | 1992-11-06 | 4.0 KB

Path: sparky!uunet!know!hri.com!ukma!darwin.sura.net!paladin.american.edu!auvm!VAXF.COLORADO.EDU!POWERS_W
From: [e-mail] (William T. Powers)
Newsgroups: bit.listserv.csg-l
Subject: CONNECTIONISM AND hpct
Message-ID: <[e-mail]>
Date: 5 Nov 92 19:31:27 GMT
Sender: "Control Systems Group Network (CSGnet)" <[e-mail]>
Lines: 68
Comments: Gated by [e-mail]
X-Envelope-to: [e-mail]
X-VMS-To: @CSG
MIME-version: 1.0
Content-transfer-encoding: 7BIT


 [From Bill Powers (921104.2100)];

 Martin Taylor (921104.1340) --

 >Connectionism depends at root on the distribution of
 >representation, not on whether some elements have a continuous
 >rather than a discrete range of variation.  Your version of HPCT
 >isolates the responsibility for the control of particular percepts
 >to particular ECSs.  We would not do that, but would share the
 >responsibility through overlapping quasi-modular groups of ECSs.
 >Your version is to connectionist HPCT as classical AI is to
 >connectionist classifiers.  And it leads to related kinds of
 >problem.

 The difference you see depends on the dimension you're attending to.
 For me, the difference between AI and connectionism is that
 connectionist models let the signals representing variables be the
 important thing, where in AI all variables had to be converted first
 into symbols (words, mainly) before they could be operated upon, and
 then the operations were carried out by rule-driven algorithms for
 symbol manipulation instead of by computing devices that handle
 signals directly. To me, that's the difference between analogue and
 digital computing.

 To represent a system with fixed properties, it doesn't matter
 whether you use a distributed network or an equivalent set of
 individual funtions. BCP, p. 39:

 "It is convenient to think of the brain as a collection of localized
 functions, and of neural signals as occurring in definite pathways
 linking functions together. The model, however, will not be
 invalidated if these elements prove some day to be distributed over
 large volumes of the brain. The organizational properties of this
 model do not depend on its geometrical properties."

 If we want to account for the way these functions come into being,
 then the network representations will probably be necessary. During
 maturation, axons grow in ways that depend on what the system is
 doing, and even after maturation is complete, synapses appear, change
 their properties, and disappear.

 I think that such networks have limited scope; the size of a sensory
 or motor nucleus. On a larger scale there is clearly an architecture
 composed of separate modules and separate layers. The types of
 neurones are different in different modules of the brain; the brain
 is not just one huge network composed of identical elements operating
 by identical principles.

 The connectionist models I have seen do not impress me as much as
 they impress their inventors. A great deal of subjective
 interpretation is involved in saying that a network "classifies" its
 inputs, or even that it "recognizes" a form. What these networks
 actually do is a lot simpler than that: given a set of inputs, they
 will produce certain outputs over a range of the inputs. To label
 this process "classification" implies first that one already knows what
 classification is, and second that there is no other kind of
 perceptual operation of any importance. I reject both implications as
 unwarranted, the first because there has been no careful
 investigation of the elements of perception (of the kind I have tried
 to develop) and the second because it is obvious, at least to me,
 that a great deal more than classification goes on in perception. I
 think that connectionists are trying to accomplish in one jump what
 the real perceptual system does stage by stage. This may be possible,
 to some extent, but this sort of modeling will necessarily fail to
 account for perceptions of both lower and higher levels than
 classifications: perception of motion, for example, or perception of
 principles.