Page 116 - Deep Learning

Creative Insight: The Redistribution Theory 99

them are links. Each node is associated with a level of activation that fluctuates
over time. At each moment in time, a small subset of the nodes have activa-
tion levels above a threshold. Those elements are immediately available for
processing; they form the current content of working memory. Activation is
passed along the links from elements that are currently above threshold to
other, related but not yet active nodes. If a knowledge element receives enough
activation to rise above threshold, it “comes to mind” as we say. “Retrieval” is
a label for the event that occurs when the activation of a knowledge element
rises above threshold. As activation spreads from a source node N, it is passed
along its outbound links. A certain amount of activation is lost in each step of
the spreading process, so the amount that is spread from a given source node
N decreases gradually with increased distance from N. There are several vari-
ants of this theory that differ in the quantitative details of the spreading pro-
cess, but those details need not concern us here.
Memory retrieval is selective. A person can keep only a small amount of
information in an active state at any one time – working memory has a limited
capacity – but the knowledge store is vast, so the retrieval process necessar-
ily makes choices, however implicit and unconscious, about what to retrieve.
Retrieving an element X constrains which other elements can also be retrieved
at the same time. In a problem situation, activation initially spreads from the
problem representation and the goal. The initial encounter with the problem
thus determines the knowledge elements that are initially marshaled to solve
it, and those elements in turn become sources from which activation spreads.
Retrieval is a cyclic, iterative search through memory for those knowledge ele-
ments that are most likely to be relevant for the problem at hand.
Memory links are acquired in the course of experience, so the structure
of the memory network mirrors the structure of experience. Reaction time
studies reveal, to no one’s surprise, that the two concepts of chair and table are
28
more closely linked in people’s heads than chair and, for example, parachute.
The strengths of memory links are influenced by a variety of factors, including
frequency of use, recency of use, estimated capacity demands and past use-
fulness. The structure of the knowledge network – the set of links – and the
strength of the links serve to project the accumulated prior experience onto
the current situation. In conjunction, the relative strengths determine what
knowledge is retrieved on the initial encounter with a problem, and what is
retrieved constitutes the memory system’s best guess as to what knowledge is
relevant for the problem at hand. 29
Psychological research on memory has focused on information com-
municated via pictures, texts or observations of events. For the purpose of