Page 226 - Deep Learning

Error Correction: The Specialization Theory 209

The phrase the left turn was an error must therefore be understood to mean
turning left instead of right at that intersection, approaching from the south with
destination X in mind, was a less effective way to get to X than one of the alter-
native turns available at that intersection. An action token is a mistake when
it is performed in a situation in which it does not serve the goal of the person
performing it, at least not as well as some other action. An error is an action
that does not lie on the (shortest) path to the goal.
Like common sense, this view defines errors as deviations from a norma-
tive standard (the best or correct solution). This concept of error belongs to
an omniscient observer who can access the entire situation tree and evaluate
every action in relation to its alternatives. Although the objective view is a
useful analytical tool, it does not explain how people detect their own errors.
The concept of a deviation from a standard might capture the essence of error-
hood, as it were, but it does not provide someone who has not yet acquired that
standard with a cognitive mechanism for detecting errors. To be informative
during learning, errors must be recognizable by the learner himself. A theory
of error detection requires a shift from an objective to a subjective view of
errors.
Situations produced by erroneous actions typically contain features that
reveal that an error has been committed. In everyday task environments the
features can be perceptually intrusive, as when the learner smells something
burning while cooking or hears grinding noises from the gearbox while learn-
ing to drive with a standard transmission. In technical task domains, error
detection might be less direct. A chemistry student might detect an error in a
laboratory procedure by comparing the weight of the reaction products with
the weight of the reactants and observe that they differ. In symbolic task envi-
ronments, error signals tend to be subtler. A statistics student might suspect
that the magnitude of a quantity is unreasonable, and a programmer might
recognize unbalanced parentheses as a syntax error. 10
The recognition of such error signals typically requires task-specific knowl-
edge. For example, a driver trying to find a location north of his or her current
location who encounters a road sign saying Route 60 south knows that he or
she has made a wrong turn. To recognize this traffic sign as an error signal
one must know that roads are labeled according to their directions and that
south and north are opposites. In other situations, the knowledge required to
recognize an error signal might be more extensive. To recognize the answer
to a statistics problem as unreasonable one must have some knowledge about
the range of reasonable answers. To recognize syntax errors in computer code
one has to know something about the correct syntax. To check a laboratory