7.1 Quillian Networks

   It is rather difficult to study the structure of human long-term memory, since we cannot look
into a person’s brain or mind to determine the processes by which their memory organizes itself.
The only aspect of memory which is open to study is memory access. In particular, a great deal
of attention has been devoted to the time required for memory access.
   One frequently studied phenomenon is "priming": an experimenter shows the subject one
letter, then shows two other letters simultaneously, and measures how long it takes for the
subject to determine if the two simultaneously presented letters are the same or different. The
answer comes considerably faster when the one preliminary letter is the same as one of the two
simultaneous letters (Posner and Snyder, 1975). This shows that when something is summoned
from memory it is somehow "at the top" of the memory for a period afterwards, somehow more
easily accessible.
   According to the idea of structurally associative memory, if x is more easily accessible than y,
those things which are similar to x should in general be more easily accessible than those things
which are similar to y. This has been shown by many different experimenters, e.g. Rips et al
   This is an essential result; however, it is clear that this simple feature would be predicted by all
sorts of different memory models. Using similar techniques, psychologists have attempted to
determine the structure of memory in further detail. For instance, Collins and Quillian (1969)
performed several experiments to test Quillian’s network theory of memory (Fig. 3), according to
which concepts are stored as nodes in a digraph. For instance, chair, couch and table would be
stored on nodes emanating from the furniture node; and coffee table and dinner table would be
stored on nodes emanating from the table node. In their experiments, subjects were asked to
verify statements of the form "an X is a Y" — say, "a couch is a table", or "a couch is furniture".
Collins and Quillian predicted that the time required to verify the sentence would be a linear
function of the number of links between the concepts in the memory digraph.
   This hypothesis at first appeared to be correct; but further experiments showed that the model
has difficulty dealing with negative responses. Therefore Rips et al proposed an alternate model
of memory in which the similarity of twoentities is defined as the amount by which their
"semantic features" overlap. According to their experiments, this sort of similarity is a far better
predictor of reaction time than Quillian’s hierarchical distance.
   Collins and Loftus (1975) responded with an improvement of the Quillian model, according
which concepts are stored in the nodes of a network, and each link of the network is assigned a
weight corresponding to the degree of association between the concepts that it connects. Memory
access then involves two stages: 1) "spreading activation", in which an activated node spreads its
activation to neighboring nodes, and 2) evaluation of the most active portion of the network. This
accounts for the data because it incorporates the "feature overlap" of Rips et al into the network
structure. Ratcliff (1989) has criticized the model for not adequately explaining the process of
evaluation; but this seems to me to be beside the point.
   The model of Collins and Loftus is somewhat similar to the structurally associative memory;
the biggest difference is that in the Collins and Loftus model, "similarity" is imposed a priori
from the outside. The model does not explain how the mind assigns these degrees of similarity.
Psychologically, this is unsatisfactory. However, it must be noted that "Quillian networks" of the
sort studied by Collins and Loftus have since become quite common in AI programs. The a priori
nature of similarity is no problem here, since the writer or user of the program can specify the
pertinent degrees of similarity. Quillian networks are a simple and effective way of representing
   A unique Quillian network may be derived from any network of emergence by a very simple
process. To explain this we shall require one preliminary concept.
Definition 7.1: The contextual distance between x and y, relative to the set V, is the sum over
all v in V of d#[St(x%v)-St(x),St(y%v)-St(v)]. It will be denoted d##(x,y).
This measures the distance between x and y relative to the set V: not only direct structural
similarities between x and y are counted, but also similarities in the ways x and y relate to
elements of V.
    Then, to form the Quillian network corresponding to a given network of emergence, one must
simply create a new link between each two nodes, and weight it with the contextual distance
between the entities stored at the nodes, relative to the other entities in the memory.
    The structurally associative memory network refers to a deeper level than the Quillian
network, but it is fully compatible with the Quillian network. It therefore seems reasonable to
hypothesize that anything which the Quillian network explains, the structurally associative
memory can also explain. However, there are certain phenomena which require a deeper
analysis, and hence the full power of the structurally associative memory.
    For example, it will be seen below that the details of the relationship between memory and
analogy fall into this category. The Quillian network supports avery rough form of analogy based
on a priori "similarity", but to explain the full subtlety of analogical reasoning, the structurally
associative memory is required. In general, it would seem that the Quillian network is not
really suited to be a large-scale, adaptive, self-organizing memory. The structurally associative
memory is based on pattern recognition and hence it can easily be modified on the basis of new
pattern recognitions. And the STRAM stores more refined data: it stores information about the
type of relation between two entities.
    The remainder of this chapter will be devoted to showing that the STRAM is capable of
functioning as a large-scale, adaptive, self-organizing memory; and showing how its structure
relates to analogical reasoning.
Kaynak: A New Mathematical Model of Mind

Belgeci , 2280 belge yazmış

Cevap Gönderin