Page 245 - Deep Learning

228 Adaptation

the discrimination process is triggered? When it is triggered, the learner
must retrieve all applications of the relevant rule from memory and compare
them with respect to shared and unshared features to identify the poten-
tially discriminating ones. Finally, the presence or absence of a particular
feature does not by itself prove causation. The comparison process is unable
to determine whether the presence of a particular feature is accidental or
related to the outcome. For example, if a person rents three red cars of brand
X and all three break down, while three yellow rental cars of brand Y do not,
then the discrimination mechanism is as prone to form the rule “do not rent
red cars” as “do not rent cars of brand X.” These problems reduce the psycho-
logical plausibility of discrimination mechanisms that are based on the idea
of comparing rule applications with positive and negative outcomes.
In contrast, the process of constraint-based specialization – apply the con-
straints, notice constraint violations and specialize the relevant rule according to
the algorithm specified previously – does not require extensive memory storage
or the retrieval of episodic information, it is computationally cheap and it does
not require decisions that are necessarily based on insufficient information. The
reason is that it operates on a richer type of information. Thorndike’s weakening
mechanism operates on the action that produced the negative outcome, while
rule discrimination operates on the situations that appear in the execution his-
tory of the relevant rule. In contrast, constraint-based specialization operates
on the information that resides in the relation between the observed and the
expected outcomes of an action. A single constraint violation contains sufficient
information to uniquely determine the appropriate revision of the relevant rule.
At a higher level of abstraction, the constraint-based specialization theory
bears a family resemblance to a wide range of theories centered around the
34
action-conflict-change principle. The latter has been articulated in many dif-
ferent ways by different theorists. Jean Piaget’s disequilibrium, Leon Festinger’s
cognitive dissonance, Roger Schank’s expectation failures and Karl Popper’s
falsified predictions are different conceptualizations of cognitive conflict.
Different types of conflicts imply different types of processes for overcoming
them. For example, cognitive dissonance supposedly triggers changes that
restore consistency to the learner’s belief system (see Chapter 10), while expec-
tation failures might trigger so-called tweaking of the explanation patterns
that generate failed expectations. Although the theory of constraint-based
specialization proposes novel conceptualizations of both conflict and change,
it is similar to other conflict-driven theories in that it builds on the principle
that intelligent action must be guided by the information residing in the rela-
tion between expected and observed outcomes of actions.