Page 235 - Deep Learning

218 Adaptation

A plausible approach is to divide errors into types based on their origins.
The concept of general problem-solving methods suggests a particular hypoth-
27
esis about the origin and hence the correction of errors. At the first encounter
with the target task, a learner is not so much practicing as solving the problem
of how to perform the practice task. In the earliest stage of practicing – the
very first training trials – he cannot yet have a strategy that is adapted to the
specific task (or else the task is not unfamiliar). But if he is acting, he is, by defi-
nition, executing some strategy or another. General methods are not adapted
to the particular structure of any one task environment. They are widely appli-
cable, but inefficient. Their function is not to produce solutions but to enable
the learner to generate task-relevant behaviors vis-à-vis an unfamiliar task.
Like other strategies, general strategies can be described as collections of
rules. The latter serve as templates for task-specific rules. For example, the gen-
eral heuristic, if you want to switch on an electrical device, then push its power
button, when applied to a particular device such as a computer projector, might
generate the specific rule, if you want to turn on a computer projector of brand
X, then push the red button to the right. The conditions on such rules (i.e., the
part between if and then) are likely to be incomplete because the general rules
do not contain any knowledge about the specific situation to which they are
applied. Consequently, a rule has some probability of matching situations in
which the action it recommends is not, in fact, appropriate, correct or useful.
To continue the projector example, perhaps the projector works only if it is
switched on after being hooked up to the computer, so a more complete rule
is, if you want to turn on a computer projector of brand X, and the projector has
been hooked up to the computer, then push the red button to the right. Without
this additional condition, the rule will recommend pushing the power button
even in situations in which the projector and the computer are not yet con-
nected, perhaps causing a communication problem between the two devices.
I refer to errors caused by overly general rules as displacement errors because
they arise when an action was recommended even though it should not have
been and thereby displaced whatever action would otherwise have been per-
28
formed in that situation. The claim that a significant proportion of errors in
real situations is due to displacement is the Displacement Hypothesis.
Evidence that errors are generated because rules and strategies are overly
general at the outset of learning has been found in multiple task domains; see
29
Table 7.1. Overregularization in language acquisition is a classical example.
Children at a certain age apply linguistic rules to cases that constitute excep-
tions to those rules. An example is the overregularization of past tense forma-
tion. That is, the tendency to use the suffix “-ed” even in the case of irregular