Page 273 - Deep Learning

256 Adaptation

that are likely to execute over days, weeks or even months, that are learned over
several years and that are distributed over multiple operators. The gap between
the time scale of a single learning event, on the one hand, and the time scale
of expertise in the context of a complex system, on the other, raises the ques-
tions whether the constraint-based theory of learning from error applies at the
longer time band and the higher levels of complexity, and what implications, if
any, it has at the level of collectives. At the longest time band and highest level
of collectivity, the question is whether the fall of seemingly successful societies
can be understood in terms of identifiable errors.

SCALING ACROSS TIME AND COMPLEXITY

Simple skills can be acquired by practicing for a few minutes or at most an hour.
To learn how to tie a bowline knot, to master a new cell phone and to find one’s
way around an unfamiliar shopping center are examples. But how much prac-
tice is required before a person is competent to diagnose medical cases, pilot a
passenger jet or manage a multinational business? Thanks to empirical studies
by Anders Ericsson and other expertise researchers we know the answer: Top-
level performance in any field of skilled activity requires approximately 10 years
of practice, if we assume that the learner practices approximately four hours per
3
day. Unexpectedly, the order of magnitude of these numbers is relatively con-
stant across such otherwise different fields of activity as athletics, games, music
and the professions (e.g., business, engineering, medicine and the military).
The claim is not that this amount of practice automatically produces
expert performance, but that expert performance cannot be reached with less
than this amount of practice. The 10-year estimate should not be interpreted to
mean that learning stops after 10 years. On the contrary, the growth of compe-
tence is a lifelong affair, beginning in the 15-year period that we refer to as the
period of cognitive development, progressing through the adult years, which
for many include a decade or two of achieving expertise in a profession or
vocation, and continuing into late adulthood. It slows down only when further
improvements in cognitive performance are counteracted by cognitive aging.
The question of interest is how learning mechanisms that execute in a few
seconds or at most a couple of minutes relate to the growth of competence
at the time bands of cognitive development, expertise and life-span learning.
The constraint-based theory of learning from error presented in Chapter 7
is based on five principles: (a) Errors are caused by overly general (under-
specified) skills, or components of skills. (b) Errors are detected as deviations
from observed outcomes of actions from the outcomes expected on the basis