Page 192 - Deep Learning

The Growth of Competence 175

100

Time to Task Completion (secs) 60
80

40

20

0
0 5 10 15 20
Trial
Figure 6.1. Mean task completion times for 106 students practicing an unfamiliar
symbol manipulation task as a function of the number of trials.

less need to backtrack; the performer hesitates less and completes the task
faster; and he can achieve more finely specified outcomes (come to a full stop
five feet before the stop sign). As he performs the task over and over again, his
actions gradually shape themselves to fit the contingencies and causal laws
of the task environment (when the roadway is wet, it takes longer to stop the
car, so start braking earlier).
improvement is not only gradual but follows a particular form. When
performance is plotted as a function of amount of practice, the result is a well-
11
documented temporal pattern called the learning curve. Learning curves are
invariably negatively accelerated. That is, the rate of improvement is high ini-
tially but gradually diminishes, producing a concave curve. Figure 6.1 shows
a learning curve for a group of college students practicing a simple symbol
12
manipulating exercise. The curve drops sharply in the beginning, but the
rate of change decreases smoothly. After 20 practice trials, improvements
have almost ceased and the curve approaches a stable level, referred to as
the asymptote. our familiarity with such practice effects veils a conceptual
puzzle: Why is skill acquisition gradual? Why is a skill not fully mastered at
the moment the target task has been completed correctly for the first time?
Why does the rate of improvement follow a decreasing (negatively acceler-
ated) curve?
practicing in a training environment would be pointless unless the skills
acquired there can be executed in the real task environment. We expect a