Page 204 - Deep Learning

The Growth of Competence 187

the learner’s current version of the strategy-to-be-learned. only a few opera-
tions, irregularly distributed across time, revise the rules. in one analysis, Kurt
Van Lehn identified 11 strategy changes in a problem-solving effort that lasted
27
20 minutes, a rate of 1 learning event per 2 minutes. A single learning event
might take between 1 second and 1 minute. Each learning event tweaks the
current version of the relevant skill, but the magnitude of the change is small.
observable effects on the time to solution, the probability of error and other
aspects of behavior are cumulative consequences of many minor changes in
the underlying strategy. The rate of change during a practice period is a func-
tion of the density of learning events, and the latter depends on the task envi-
ronment, the difficulty of the target task, the learner’s current level of mastery
and other factors.
The central theoretical task is to specify the number and nature of the
learning mechanisms. The growth of a cognitive skill is a complex type of
change and it is highly unlikely that it is due to a single mechanism. For exam-
ple, learning from instruction is quite different from learning in the course
of independent practice, and learning by adapting an already mastered skill
to a new task is obviously a different process from learning from error. A
successful skill acquisition theory needs to postulate a repertoire of distinct
learning mechanisms, or modes of plasticity in neuroscience terminology. For
ease of reference, i call this the Multiple Mechanisms Principle. The question is
how large a repertoire to expect. if not a single mechanism, then how many?
intuition rebels against the thought that there are a hundred distinct ways in
which practical knowledge can change. More than 1, but less than a 100; so 10
appears to be the right order of magnitude. As it turns out, most skill acquisi-
tion theories proposed over the last century are indeed multimechanism theo-
ries with less than 10 distinct mechanisms. The complexity of the hypothesized
mechanisms and the rigor with which they are specified have increased over
time.

A Century of progress
The beginning of systematic research on the acquisition of cognitive skills
can be specified with precision. William James’s comprehensive Principles of
Psychology, published in 1890, did not include a chapter on skill acquisition,
but the ph.d. thesis of his student E. L. Thorndike, begun in 1896 at Harvard
University but issued a few years later from teachers College at Columbia
University, reported experimental studies of how various species of animals
learned to escape from cages with nonobvious door-opening mechanisms. 28, 29