Page 220 - Deep Learning

The Growth of Competence 203

Why is improvement gradual? The Raisins principle introduced in
Chapter 4 provides the answer: Each learning event, each application of a
learning mechanism like those described previously, produces a single muta-
tion in a single rule within a large rule set. The effect of a single application of
a learning mechanism to a single rule is necessarily minor. to master a com-
plex task requires many such mutations, multiple revisions of each one of a
large number of rules.
These answers cover the first four of the six questions about practice stated
at the beginning of this chapter. The fifth and sixth questions, about transfer of
training and the efficacy of instruction, are answered in Chapter 7.
The fact that the nine Modes Theory provides in-principle answers to the
fundamental questions about skill practice does not prove that it is the best possi-
ble theory, because some other theory might be able to provide even more satis-
factory answers. one way to improve the theory is to try to improve on the three
meta-theoretical principles that shape it: Multiple Mechanisms, information
Specificity and Maximal Effectiveness. But it is more plausible that there are mul-
tiple mechanisms than that all aspects of skill acquisition can be explained by a
single change mechanism. to question information Specificity is to claim that
learning mechanisms can operate on any type of information, but no such mech-
anisms have ever been proposed and it is not clear what this means. The poten-
tially weak point in the meta-theory is the assumption of Maximal Effectiveness.
The latter sounds too optimistic to be true. But which of the nine modes of learn-
ing could a critic plausibly deny? Even casual observation of everyday behavior
would quickly yield examples of each of the nine learning modes. All in all, to try
to improve the meta-theory is an unpromising line of attack.
A more plausible approach is to accept the meta-theory and to question
the particular mechanisms proposed for each type of information. The theory
might be improved by developing psychologically more plausible mechanisms
for one or the other mode of learning, or by increasing the empirical support
for one or the other mechanism. For example, a particularly weak component
of the theory is the pool of hypotheses for learning from errors, failures and
other types of negative feedback. For beings who constantly deal with novel
and changing task environments, errors and unexpected outcomes are facts
of life, so it is likely that we have evolved a more sophisticated mechanism for
making use of negative feedback than those proposed in past research.
Even if the mechanism for a particular mode of learning is well developed
and motivated, the question remains of whether it is the right mechanism –
that is, whether it is equivalent in its input and output to how the brain operates
when changing in the relevant mode. How well grounded is the mechanism