Page 289 - Deep Learning

272 Adaptation

knowledge will hence be re-used precisely to the extent that it is applicable to
an altered or novel task. The amount of cognitive work needed to transfer the
prior skill is a function of how many learning events are required before the
rules have been adapted to the new task and stop generating errors.
This explanation for expert flexibility works best at the longest time
scale. In the course of 10 years of practice, an expert creates, according to
the constraint-based theory, a vast knowledge base of rule genealogies, each
containing many rules of varying levels of specificity. When a new situation
is encountered, the most specific rules that match are activated. Given widely
varied prior experience and 10 years of training, the person’s knowledge base
will contain tens of thousands of such intermediate rules. Because situations
and problems within a domain resemble each other to some degree, the best-
fitting subset of prior rules might be almost able to handle the new situation.
The versions that fit are likely to be overly general, and hence require some
revisions. The issue is how much cognitive work those revisions represent. In
situations that exhibit only small differences to previously encountered situ-
ations, the amount of cognitive work required to specialize the existing rules
to fit the new situation might be minor. If the work required to specialize
the rules is a small fraction of the total effort required to perform the task,
it will look to an observer as if the expert already knew how to carry out the
novel task correctly. A large knowledge base of rule genealogies provides a
resource that enables experts to re-specialize their previously learned rules
to novel circumstances with a minimum amount of cognitive work, provid-
ing the illusion of having mastered the unfamiliar circumstances before they
were encountered.
In summary, the constraint-based specialization theory, although orig-
inally developed to explain learning from errors in single skills, is compat-
ible with three key features of expertise: The size of the knowledge base is
explained by the fact that the repertoire of subskills keeps growing, each new
subskill requiring a period of specialization. Although any one subskill can be
mastered in a few hours of training, the number of subskills is large, so mas-
tery of the entire skill set continues over a long time. The domain specificity of
expertise is predicted by the mode of change implied by the constraint-based
theory. If rules start out general or incomplete and become more and more
specific, then the ultimate knowledge base consists of highly specific rules.
The flexibility of such a knowledge base is explained by the constraint-based
transfer mechanism introduced in Chapter 7. The key features of constraint-
based error correction are compatible with the cumulative effects of practice
over a long time.