Page 287 - Deep Learning

270 Adaptation

17
to the relevant domain. The cognitive advantage of an expert over a novice
disappears when he is confronted with problems in some other domain. The
evidence for this is particularly well established in chess: Chess masters per-
form orders of magnitude better on a variety of tasks designed to tap into their
chess knowledge, but the advantage disappears on analogous tasks with other
materials. For example, chess masters are much better than a chess novice at
remembering the locations of chess pieces after a brief exposure of a board
position, but they are not superior to the average person in remembering a list
of random digits; for an expert mental calculator, the relation is the opposite.
Similar indications of domain specificity have been observed by researchers in
a variety of domains.
The domain specificity of expertise contradicts common sense. The com-
mon view is that cognitive competence starts out highly concrete and task
specific in childhood, but over time evolves to become more abstract and for-
mal in character as a person matures; in short, children are concrete thinkers
while adults are abstract thinkers. Piaget’s theory of cognitive development
embraced this view. It is frequently expressed in the educational research lit-
erature as well as in the popular culture. One reason for this belief is perhaps
that “more abstract” tends to be equated with “more powerful.” In the realm
of declarative knowledge, there is truth in this equality. An abstract concept
or principle is powerful in part because it applies to many problems and sit-
uations. The matter stands differently with respect to practical knowledge.
Efficiency of performance depends on knowledge of exactly when, under
which circumstances, each relevant action is to be performed. A jet pilot can-
not fly a plane on the basis of the abstract principles of aerodynamics but has
to know exactly when to do what. In the course of extended practice, practical
knowledge moves from initial, general heuristics to highly detailed and spe-
cific procedures.
The domain specificity of the knowledge accumulated through long-term
practice is predicted by the constraint-based learning mechanism. The main
principle of this mechanism is that errors are unlearned by adding conditions
to goal-situation-action rules. The new conditions specialize and refine faulty
rules so that they apply only in the set of situations in which the action they
recommend is appropriate, correct or useful. The change that occurs through-
out practice thus moves practical knowledge from a small initial kernel of very
general dispositions to a large knowledge base of highly particular and specific
rules with detailed and precise conditions of applicability. This is the type of
change generated by constraint-based learning, and its end point is the type
of knowledge that researchers believe explains the observable performances