Page 285 - Deep Learning

268 Adaptation

to complex undertakings. The complexity of the total skill set – symbolized in
the figure by the height of the vertically stacked single-skill learning curves –
increases over time. Because the two movements – streamlining individual
subskills and broadening the repertoire – push cognitive complexity in oppo-
site directions, it is impossible to derive specific predictions about the time
course of complexity in the long-term growth of competence.
In short, competence grows along three dimensions, each with its own
shape. The mastery of each individual skill follows a negatively accelerated
curve of improvement that is driven both by the elimination of errors and
other unnecessary steps and by the optimization and speedup of the remain-
ing steps. The discovery of new, more powerful strategies for individual tasks
superimposes the multiple overlapping waves pattern on the learning curves
for individual skills. Progress follows the learning curve for each strategy, but
the strategies that apply to the same task also succeed each other over time,
resulting in the inverted U-shaped pattern in the frequency of use of any one
strategy. Finally, the changes in performance on each subtask are embedded
within the widening repertoire of distinct but interrelated skills. At the 10-year
time band, the growth of competence is not well described as a big learning
curve, except in certain special cases. Instead, there is a growth of complexity
as more subskills are integrated into the overall competence. The three move-
ments depicted in Figures 8.1, 8.4 and 8.5 do not combine to produce any over-
all pattern or regularity, as far as we know. The internal mechanics of learning
from error impacts the shape of the learning curve for individual strategies
but do not punch through to the overlapping waves pattern or the growth of
complexity, both of which are due to other factors.

The Anatomy of Competence

The end point of extended practice is expertise. Empirical studies of experts
as well as efforts to implement expert systems in Artificial Intelligence have
converged on three characteristics of expertise: the large size of the expert’s
knowledge base, the domain specificity of expert knowledge and the ability of
experts to behave flexibly.

Size of knowledge base
Why does it take so long to become expert? The answer is that there is a lot
to learn. Simple skills like tying a bowline knot, mastering a new cell phone
and finding one’s way around an unfamiliar shopping center can be captured
in one or two dozen goal-situation-action rules. Not so expertise. We have