Page 253 - Deep Learning

236 Adaptation

before: I smoothly transfer the relevant parts of a typical restaurant script (be
seated, order, eat, pay), and I do not get confused if the tablecloth has a novel
color scheme or the items on the menu are different from any I have seen
before. Everyday life suggests that transfer of prior skills to altered or novel
situations is ubiquitous and nearly automatic. Empirical evidence for powerful
transfer mechanisms should therefore be plentiful.
But when cognitive psychologists conduct experiments to measure trans-
41
fer, they tend to find less transfer than they expect. Even when experimental-
ists set up a sequence of problem-solving experiences that is ideal for transfer,
their experimental subjects can appear peculiarly obtuse in their inability to
apply what they learned in the beginning of the sequence to the challenges at
its end. Teachers and educators report frustrations that mirror those of the
psychologist. Students are forever disappointing their teachers by not applying
the knowledge they have learned – or seem to have learned – to novel prob-
lems for which that knowledge is relevant. Educators speak in despair about
inert knowledge and some have proposed that knowledge acquired in one situ-
ation applies to future situations only if it is processed at the time of learning
so as to prepare for those very situations, a rather pessimistic view of human
flexibility. 42, 43 In short, transfer of practical knowledge works smoothly and
successfully in everyday life, but when we try to put it under the microscope
or enhance it, it turns fickle.
A theory of skill acquisition should resolve this puzzle, but the Specialization
Principle that is at the heart of the theory of learning from error appears to
make transfer more rather than less difficult to understand. If skills are morph-
ing from higher levels of generality to greater specificity in the course of prac-
tice, the end result should be even less transferable than the starting point.
High performance depends on a close fit between skill and task, but better
fit to one task necessarily lowers fit to another, different task. To be effective,
44
practical knowledge has to be specific; to transfer, it has to be abstract. We
cannot have it both ways, or so it seems.

The Constraint-Based Theory of Transfer
The rule genealogies and conflict resolution scheme introduced previously
provide a resolution to this dilemma. Recall that when new, more special-
ized versions of rules are added to memory, the previous versions are kept
in memory but downgraded in priority. A task will be processed by the most
specific rule (or rules) that matches the situation at hand. In the case of a
highly familiar task, the most specific matching rules will be the leaf nodes in