In Section 7.2 I explained the structurally associative memory by analogy to Quillian
networks. But, as hinted there, the Quillian network has several undesirable properties not shared
by the STRAM. Some of these are relatively technical, such the fact that the Quillian network
has no connection with the details of analogical reasoning. But there are also more philosophical
differences. In particular, I will argue that there are absolutely crucial phenomena which the
Quillian network cannot even begin to explain.
What is at issue here is what Israel Rosenfeld (1988, p.3) has called
a myth that has probably dominated human thought ever since human beings began to write
about themselves: namely, that we can accurately remember people, places and things because
images of them have been imprinted and permanently stored in our brains; and that, though we
maynot be conscious of them, these images are the basis of recognition and hence of thought and
action.
In Rosenfeld’s opinion, a careful examination of classical neurological experiments shows that
"the fundamental assumption that memories exist in our brains as fixed traces, carefully filed and
stored, may be wrong" (p.5).
For instance, consider the well-known series of experiments carried out by Wilder Penfield,
beginning in the 1930’s. He stimulated various areas of the brain in conscious patients and noted
that this appeared to elicit recollections of "forgotten memories." At first sight this would seem
to speak against Rosenfeld’s point — the natural interpretation is that Penfield was touching the
areas of the brain in which those memories were stored. But actually things are not so clear.
Recent experiments show that these forgotten memories are actually "fragmentary impressions,
like pieces of a dream, containing elements that are not part of the patient’s past experiences" (p.
7).
Also, these forgotten memories occur only when the brain stimulation is simultaneous activity
with the limbic system. Since the limbic system is the seat of emotion, this is evidence in favor
of Freud’s observation that memory without emotion would be unrecognizable. As Gloor et al
(1982) put it, describing their observations of epileptic patients:
[W]hatever we experience with our senses… even after it has been elaborated as a percept in
the temporal neocortex, must ultimately be transmitted to limbic structures in order to assume
experiential immediacy. This may… imply that all consciously perceived events must assume
kind sort of affective dimension, if only ever so slight.
Rosenfeld proposes that, rather than storing traces, memory stores procedures. According to
him, what happened in Penfield’s experiments was that certain processes were activated, which
then constructed the so-called forgotten memories on the spur of the moment, based partly on
emotional factors and partly on information somehow stored in nearby parts of the cortex.
Further support for this point of view is given by the work of Mahl et al (1964), who observed
these forgotten memories depend significantly upon "the patient’s mental content at the time of
stimulation." Sometimes they may not be memories at all, but merely rearrangements of the
ideas which occupied the patient’s mind just prior to stimulation.
No one denies that part of memory consists of procedures. For instance, every time we form a
spoken word from its syllables, we are applying certain phonological procedures. However, most
contemporary psychologists would agree with Broca, who argued in 1861 that there is a crucial
structural difference between the image-based memory responsible for the storage of words and
their meanings, and the procedural "memory for the movements necessary for articulating
words."
Against such arguments, Rosenfeld summons an impressive variety of evidence. In each case,
a phenomenon which at first appears to depend onimage-based memory is seen to require
procedural memory. For instance, he refers to David Marr’s demonstration that shapes can be
recognized as shapes without any reference to previous knowledge, merely by executing certain
procedures on the appropriate visual stimuli. This shows that shape recognition probably does
not depend upon searching a memory store of shapes until a match is found — it may, rather, be a
matter of summoning appropriate procedures from the memory. But if shape recognition does
not require a store of shapes, then why should memory contain such a store at all?
Rosenfeld admits that "Marr did not totally abandon the idea of fixed memories, since
ultimately the naming of a shape required, in his scheme, a memory search" (p. 113). To
Rosenfeld, this is the major limitation of Marr’s approach. However, it seems to me that this
shows exactly where Rosenfeld goes too far. Clearly he is right that the brain does not hold,
somewhere in its memory, little pictures of circles or ellipses or squares. Rather, it stores certain
procedures or functions which characterize these shapes. But this fact does not imply all that he
says it does.
For the purpose of illustration, let us consider a highly oversimplified example. Let y denote
the function which, from P and r, generates the circle with radius r and center P; and let z=(P,r).
Then, roughly speaking, we may say that a certain collection of stimuli x is "representable as a
circle" to the extent that (y,z) is a pattern in x. For each shape, there will be one or more such
characterizing patterns. I do not mean to imply that the mind stores shapes by fitting them to
their standard mathematical equations, but only that it characterizes shapes by certain
"symmetries" or "computational shortcuts" which manifest themselves as patterns. Algebraic
equations are one kind of specifying pattern, but not the only kind. For example, one of the
patterns characterizing the shape "square" might be "invariance under reflection and ninety-
degree rotation".
Let us suppose, then, that in the mind’s "language" each shape is a certain collection of
characterizing procedures. Then what is wrong with calling this collection a "label" or "trace" of
the shape? It seems clear that, in general, a mind would do best to store such collections in
proximity to each other. After all, they will very often be used to recognize the same shapes.
Rosenfeld thinks Marr is wrong to speak of a "memory search". But does he believe that a
mind always immediately selects the most appropriate procedures? If a mind recognizes a shape
by recognizing certain symmetries and other patterns in it, then what could possibly be wrong
with the hypothesis that the mind has to search a little to determine the appropriate patterns?
A careful study of Rosenfeld’s book reveals that the structurally associative memory accounts,
schematically at least, not only for Marr’s work but for all the phenomena which Rosenfeld
adduces against the image-based model of memory. From the fact that most memory relies on
procedures, one cannot conclude that these procedures are not organized and accessed according
to a network structure. Returning to Quillian networks, I agree with Rosenfeld that "chair" is stored in the memory as a collections of procedures for determiningwhat is a chair. But I still
maintain that the Quillian network can be a useful approximation to the actual network of
interrelations between the procedures associated with various entities.
INFORMATION AND RELATION
Extending this point of view, I concur with Erlich (1979, p. 200) that each item stored in
memory should be "considered capable, by rights, of performing two different functions: the
informative function and the relational function." That is: each item in memory is acted on by
other items in memory, and also acts on other items in memory. The Quillian approach
emphasizes the static, "acted-upon" aspect of memory; whereas the Rosenfeld approach stresses
the dynamic, procedural, "acting-on" aspect, and considers actions on external stimuli as well as
other memory items.
For instance, "chair", "house", "meal" and so forth are collections of procedures which act
very little to transform other entities in memory — mainly they act on external stimuli. But
logical and mathematical structures, as well as words such as "in", "on" and "besides", are
primarily relational: they are collections of procedures which serve primarily to act on other
entities stored in memory.
More precisely, what I mean here by "A acts on B" is simply: "A is a function which takes B
as arguments." Those entities which are patterns between other entities in memory will thus "act
on" many other entities. This definition is validated by the fact that such entities will very often
be invoked by analogical searches proceeding from B to A; and by the fact that if A acts on B,
recognition of B as a pattern in an entity will often be followed by recognition of A.
In sum: the STRAM is a memory model which 1) accounts for the procedural nature of
memory, 2) recognizes the approximative value of static semantic networks, 3) explains the self-
organizing, generative nature of memory, and 4) acknowledges the intricate structural
interdependence of memory with cognition.
SPARSE DISTRIBUTED MEMORY
Another way to look at this dichotomy is to observe that the STRAM is superficially similar
to Kanerva’s (1988) "sparse distributed memory", in which entities are coded as binary sequences
and stored, each in several places, near other sequences that are "similar". Kanerva measures
similarity by the Hamming distance — the fewer places in which two sequences differ, the more
similar they are. This is ideal for a memory storing images of physical objects, which may be
conveniently characterized by a list of binary qualities. For instance, one could associate with a
physical object A a binary sequence a1…an as follows: a1 is 1 if and only if the object is red, a2 is
1 if and only if the objectis green, a77 is 1 if and only if the object is dirty, etc. Given any such
assignation of sequences to objects, the similarity of two objects A and B could be plausibly
measured by the number of places in which the sequences a1…an and b1…bn differed.
But for a memory storing relations, procedures and so forth, it is much more direct to measure
similarity structurally, as is done in the STRAM. It would certainly be possible to encode complete information about a complex procedure in a binary sequence. This possibility lies at the
core of algorithmic information theory. But this is not a natural way to relate procedures. To
make the Hamming distance mesh smoothly with analogical reasoning, one would have to
enumerate all possible patterns w1,…,wm, and assign each entity A a binary sequence a1,…,an
based on the formula: for some fixed C, and 0%r<k, amk+r=1 if and only if wm is a pattern in A
with intensity greater than rC but less than (r+1)C.
Given this setup, Kanerva’s sparse distributed memory would be extremely similar to the
STRAM. But it would completely ignore the fact that the procedures being stored were
procedures and not merely images of physical objects. The emphasis would be informational.
The STRAM gives equal balance to the informational and relational functions of the entities
stored in memory.
Kaynak: A New Mathematical Model of Mind
belgesi-952
