Page 236 - Deep Learning

Error Correction: The Specialization Theory 219

Table 7.1. Examples of overly general rules in four task domains. The underlined parts
of the correct rules are the conditions that have to be learned.

Past tense of verbs
Overly general rule: If the root of a verb is V, then its past tense is V+ed.
Example symptom: “goed,” “runned,” “speaked,” etc.
Correct rule: If the verb is V and V is not one of {go, run, speak, … },
then its past tense is V+ed.
Multiplication
Overly general rule: OP(X + Y) = OP(X) + OP(Y), where “OP” is any
operation.
Example symptom: √34 = √25 + 9 = √25 + √9 = 5 + 3 = 8
Correct rule: If “OP” is multiplication, then OP(X + Y) = OP(X) +
OP(Y).
Highway lane changes
Overly general rule: If you want to switch into the left lane, then turn left.
Example symptom: Cutting in front of a car in the left lane.
Correct rule: If you want to switch into the left lane and the left lane is
empty, then turn left.
Cell phone volume control a
Overly general rule: If you want to increase the sound volume, push the
volume button.
Example symptom: Pushing the button before making a call.
Correct rule: If you want to increase the sound volume and a call is in
progress, then push the volume button.
a T his example applies to a particular cell phone brand for which the volume control button works
only when a call is in progress.

verbs (e.g., comed and runned). A large corpus of such events can be explained
by the hypothesis that children have a single general rule for all regular verbs
and many specific rules, one for each irregular verb. Overregularization occurs
when the rule for regular verbs is applied to an irregular verb. That is, over-
regularization errors are a type of displacement error, caused by an overly gen-
eral past tense rule. Table 7.1 shows additional examples from three other task
domains.
The Displacement Hypothesis explains incorrect actions in terms of
the normal functioning of the cognitive system. It is common for practical
knowledge to be overly general. Only when we perform exactly the same task
over and over again do we develop rules that are specific to that very task.
This view of error contrasts with the assumption that performance errors
are caused by some malfunction in the cognitive architecture such as lack