Page 274 - Deep Learning

Error Correction in Context 257

of declarative knowledge about the task environment. (c) Declarative knowl-
edge is normative rather than descriptive in nature; it consists of constraints
rather than propositions. (d) Errors are unlearned by incorporating violated
constraints into the applicability conditions of faulty knowledge elements.
(e) The conditions to be added are identified through the particular computa-
tions described in Chapter 7.
What kinds of long-term effects should we observe if the mechanism of
constraint-based specialization is executed many times over? How does a skill
change as it is mastered? What are the cumulative effects? Do the five key fea-
tures of constraint-based specialization punch through to higher system levels
and influence changes that unfold over 10 years or more, a stretch across five or
six orders of magnitude? If so, which empirical regularities at the higher levels
of time and complexity can be explained with their help and which are driven
by other factors?

The Patterns of Skill Acquisition
The acquisition or adaptation of a skill through practice is necessarily gradual.
The learner’s initial representation of the target skill consists of rules – goal-
situation-action associations – that are incomplete and possibly incorrect. To
learn a skill is to revise the initial rules by incorporating more information
about the task environment into them. A rule undergoes only a minor change
in the course of any one learning event, so every rule requires multiple revi-
sions before it ceases to generate errors. Complex skills consist of large col-
lections of rules, so the target skill as a whole is necessarily reached through
a long sequence of learning events. What little empirical evidence is available
suggests that learning events occur relatively infrequently, perhaps one event
per 5 or 10 minutes, on average. Hence, a skill that requires 20 learning events
to master cannot be learned with less than 1.5–3 hours of practice. If change is
necessarily gradual, what pattern or patterns does it exhibit?

The short­term learning curve
Improvement due to practice is not only gradual but it also follows a well-
documented temporal pattern: If performance, measured in terms of the time
to complete the target task, is plotted as a function of the amount of experience
with the task, usually measured by the number of training trials (attempts to
4
perform the task), the result is a negatively accelerated curve. That is, the rate of
improvement is fastest in the beginning and gradually diminishes as mastery is
approached. The curve eventually approaches a horizontal line, the asymptote,