Page 240 - Deep Learning

Error Correction: The Specialization Theory 223

about the exact nature and cause of the error. Merely designating an outcome
as undesirable or inappropriate does not provide much information about how
to correct it. The frequency of error is likewise void of useful information. The
information that resides in an erroneous action can only be extracted by pay-
ing attention to the qualitative nature of the error – that is, to what the learner
knows, what the constraints say and what the learner is seeing.

Rule Genealogies and Conflict Resolution
The mechanism described so far specifies the internal mechanics of a single
learning event, a local modification of a single rule in response to a self-de-
tected error. A rule typically needs to be revised multiple times on its path from
being overly general to being correct. In addition, a strategy of even modest
complexity consists of a large number of rules. For both reasons, acquiring or
adapting a skill necessarily requires a sequence of rule revisions. Three addi-
tional principles are needed to understand how a single learning event affects
an entire rule set.
When a rule R is revised, its two descendants Rʹ and Rʹʹ are added to the
stock of rules in the person’s memory but they do not replace R. Instead, all three
rules remain in memory. When either of the descendant rules is revised in turn,
it also produces two descendants, grandchildren, as it were, to R. In this way, the
specialization process generates a rule genealogy, a tree of descendants of R. The
original rule R is a root rule and the rules that have not yet generated any errors
and hence have no descendants are leaf rules; the versions between the root and
the leaves are intermediate rules. As the genealogy grows, it becomes deeper; that
is, the number of levels between the root rule and the leaf rules increases. The
tree also grows wider; that is, each level contains more rules than the one preced-
ing it. The root rule is the most general rule in a genealogy and the leaf rules are
the most specific. If we trace a path from the root node to some leaf node, the set
of situations to which each rule on the path applies – its domain of application –
gets smaller and smaller from level to level; see Figure 7.2.
If a situation falls within the domain of application of one of the leaf rules,
it is guaranteed to also fall within the domain of application of any of its ances-
tors, including the root rule. The question arises as to which version of the rule
should be entered into the conflict set to compete with alternative rules for
the control of action in that situation. The theory asserts that the most specific
rule within a genealogy is the only one entered into the conflict set. This might
turn out to be one of the leaf nodes, but it might also be one of the intermediate
rules or even the root rule.