Page 284 - Deep Learning

Error Correction in Context 267

to entire concertos; the fighter pilot masters takeoff and landing before he trains
the maneuvers of combat; the medical resident learns to take blood samples long
before he performs surgery; and so on. As the expert-to-be slides down each of
the learning curves for already initiated subskills, he also moves through the
curriculum. These two movements impact complexity in opposite ways. The
slide down the learning curve renders individual subskills streamlined and less
effortful to execute, while the broadening of the skill set renders the overall com-
petence more complex and more difficult to maintain, extend or restructure.
The relations between these opposing tendencies are depicted in Figure 8.5.
Time runs along the horizontal axis. Cognitive load, in terms of some mea-
sure that decreases with practice such as effort, number of errors or time to
task completion, is plotted on the vertical axis. The acquisition of each sub-
skill follows a negatively accelerated learning curve, but at any one time, mul-
tiple skills are practiced. Computational mechanisms like constraint-based
error correction can explain why the individual subskills improve according
to negatively accelerated learning curves but have little to say about the larger
pattern. At the longer time band, the details of the individual learning mecha-
nisms are irrelevant. The main change is the broadening of the skill set, a type
of change that is driven by the fact that trainees naturally move from simple

Cognitive load

Amount of training
Figure 8.5. The multiple-subskill view of long-term practice. Each subskill is learned
according to a power law, but the introduction of more subskills broadens the skill set.