Page 125 - Deep Learning

108 Creativity

instrumental in producing it. If a unit P is involved in generating a result or
outcome R through option O, and R is evaluated negatively, then the activation
of O is decreased. If alternative options Oʹ, Oʹʹ, …, are available in P, and if the
person continues to attend to the problem, the balance among the options is
altered. O will eventually lose out to some competing option even though O
was preferred initially.
The change in P propagates upward through the layers of processing,
potentially causing a qualitative change in the problem representation in
working memory. This in turn alters the distribution of activation over mem-
ory, which might cause potentially useful but hitherto unheeded knowledge
elements to be retrieved. The ultimate consequence is that the heuristic search
process moves through a different solution space.
The core of this redistribution theory of insight is a specification of the
exact conditions under which the process of responding to negative feedback
with gradually lowered activation levels will produce a qualitative change in the
problem representation. Consider a choice point (processing unit) in the visual
system with, for example, three different outbound options. The argument
proceeds through six hypotheses that jointly constitute sufficient conditions
for insight; that is, for the involuntary and unexpected passage of a previously
unheeded option into consciousness after a period without progress.
First, each choice point is associated with a certain amount of activation.
To a first approximation, the amount of activation can be thought of as the
sum of the activations contributed by the inputs to the unit. This assumption
is familiar from network theories of all kinds.
Second, the activation associated with a choice point is distributed across
its outbound links (options) in proportion to their relative strengths. The
strength of an option measures past experience; more precisely, it is propor-
tional to the frequency with which that option has been executed, how often it
has been associated with success or failure, the expected payoff if it is executed,
the estimated cost or effort of execution and perhaps other variables as well.
The strength determines the proportion of activation allocated to an option
when the choice point is active.
Third, there is a threshold such that when the activation associated with a
choice point is distributed over the outbound options, some options might rise
above threshold, while others remain below. Options that are above thresh-
old pass activation and information – the result of whatever computation the
processing unit performs – onto the next processing layer. Options that receive
a level of activation less than the threshold remain dormant or unheeded.
That is, neither activation nor information is passed along those links, nor is