Page 252 - Deep Learning

Error Correction: The Specialization Theory 235

that the information contained in positive outcomes is richer or more useful
than the information contained in negative outcomes.
The successful acquisition of problem-solving skills at the level of high
school or college science supports the sufficiency of the constraint-based spe-
cialization mechanism. Applications to other task domains provide evidence
for robustness as well as answers to two of the main problems to be solved
by a theory of skill acquisition: the problem of transfer and the problem of
instruction.

THE PROBLEM OF TRANSFER
Competence grows in two seemingly distinct scenarios. On the one hand, the
learner sometimes appears to be faced with an entirely unfamiliar task and
has to construct a brand-new skill. Parents and teachers are constantly intro-
ducing the young of the species to novel tasks (today we are going to start with
algebra). Adults also experience such cases. The members of the post–World
War II generation had to learn, among many other skills, how to use Internet
search engines, a task that bears little resemblance to any prior task. In this
acquisition scenario, the learner appears to pick up a piece of mental chalk and
write the code for a new cognitive strategy on his own blank slate. On the other
hand, a learner sometimes masters a skill only to see the task environment
change due to its own internal causal dynamics or due to externalities, forcing
him to adapt his strategy to the changed circumstances. The switch from land-
lines to cell phones is an example; some features of phone use remained the
same, others changed. This adaptation scenario highlights the fact that practi-
cal knowledge often needs to be used in situations other than the one in which
it was acquired.
These two scenarios – acquiring a brand-new skill and adapting an exist-
ing skill to a new task – are discussed separately in the cognitive research
literature under the two labels “skill acquisition” and “transfer of training.”
Research studies are typically conceptualized as being about one topic or the
other. The implicit assumption is that acquisition is the fundamental case. A
strategy has to be acquired before it can be transferred, or so it seems. In this
view, transfer must be accomplished by some additional cognitive mechanism,
over and above those needed to acquire a cognitive strategy in the first place.
Identifying the mechanism of transfer is a long-standing research problem in
cognitive psychology. 40
There is no doubt that people can transfer what they learn in one situation
to another. I prove this every time I eat in a restaurant where I have not eaten