Page 254 - Deep Learning

Error Correction: The Specialization Theory 237

the relevant genealogies. A completely unfamiliar task will only match against
the root rules (i.e., the rules that encode general methods). Between these
two extremes lies the succession of rules of intermediate generality that were
created in the learning events that lead from the root node to the current leaf
node.
Consider what happens when a person encounters a task that bears some
relation to an already mastered task without being identical to it. By hypoth-
esis, the leaf rules in the relevant rule genealogies will not match. However, less
specific rules in those genealogies might nevertheless match, precisely because
they are less specific. Conflict resolution by specificity causes the novel task to
be filtered through the rule genealogies until it encounters intermediate rules
that do match. Those rules constitute the initial rule set for the transfer task.
Conflict resolution by specificity will automatically pull the most specific prior
rules that match out of the total set of rules available. Prior practical knowl-
edge will hence be re-used precisely to the extent that it is applicable to an
altered or novel task.
How useful the prior rules will be depends on the relation between the
set of prior tasks and the transfer task. If the two are utterly different, only the
root rules might match the transfer task and the learner is then indeed in a
blank slate situation. If the task is not entirely unfamiliar, the matching rules
will be some distance from the root rule and hence embody some prior learn-
ing, but they might require more or less additional revisions before they stop
violating constraints in the novel task. If the intermediate rules are close to the
leaf nodes, then the learner almost knows the transfer task already and he will
master it with little cognitive work, corresponding to the ease of adaptation
that characterizes everyday life. If this transfer mechanism works, it should
enable the HS model to exhibit transfer.

Simulating Successful Transfer
Ernest Rees and I explored the problem of transfer in the context of the task of
counting a set of objects, a task that preschool children master so early and so
well in the face of haphazard adult instruction that developmental psycholo-
gists suspect that the relevant constraints – called counting principles by Rochel
Gelman, C. R. Gallistel and their co-workers – might be, if not innate, at least
supported by some preparatory neural structures that make their acquisition
45
highly probable in the presence of relevant stimuli. An example of a counting
principle is that each object in the set to be counted is assigned exactly one
number (One-One Mapping Principle). The case of counting thus provides a