Page 242 - Deep Learning

Error Correction: The Specialization Theory 225

again but in a different way, generating a different pair of descendants and
hence starting a new genealogy from the same root.
Because A is known to be an error in some of the situations in the third
subset, it is also necessary to consider actions other than A in those situations.
This will occur because rules from other rule genealogies, recommending
other actions, will also be entered into the conflict set. Within the conflict set,
the same conflict resolution scheme applies: A more specific rule is preferred
over a less specific rule. If a leaf rule R is the result of multiple revisions and
2
it recommends action B, it will be preferred over the root rule R that recom-
1
mends action A. The rule that has been revised a greater number of times is
guaranteed not to cause certain types of errors, so it makes sense to prefer it
over any general rule. If the situation falls within the domain of application of
one of R ’s descendants, then that descendant will enter the conflict set instead
1
of R , and there will be competition between two specific rules recommend-
1
ing different actions. The conflict resolution scheme has additional features as
well, but the details are not important for present purposes.
In short, when a rule triggers a constraint violation, the relevant rule set is
augmented with two specialized versions of that rule, but the rule itself is also
kept in memory. In each operational cycle, the most specialized version of a
rule that matches the current situation enters the conflict set. Within the con-
flict set, more specific rules are also preferred over less specific ones. The effect
of this conflict resolution scheme is to consider action A in situations in which
it is known to avoid certain types of errors, and at the same time give alternative
actions greater priority in situations in which A’s correctness is still unknown.

Relation to Alternative Mechanisms
According to the specialization theory, the acquisition and adaptation of a
cognitive skill begins with general problem-solving methods. The rules that
implement such methods generate errors precisely because they are gen-
eral; that is, they put minimal restrictions on the actions they recommend.
Consequently, those actions are sometimes executed in situations in which
they are not appropriate, correct or useful. To detect such an error, the learner
has to have sufficient knowledge about his task environment to evaluate the
outcomes of his actions. An error is recognized as a constraint violation. To
correct the fault in the responsible production rule is to specialize that rule by
adding conditions that restrict its application to situations in which the rele-
vant constraint is not violated. During conflict resolution, more specialized
rules are preferred over less specialized ones.