Page 234 - Deep Learning

Error Correction: The Specialization Theory 217

on the other side, two output lights. The natural expectation is that the top but-
ton controls the top output light, while the lower button controls the bottom
light. This leads to a rule such as to turn on the top light, push the top button.
However, as the device is configured internally, a component is on only when
all its inputs are on, so both the indicator light and the lower output lamp have
to be on for the top button to switch on the upper output light. If the black box
component is on or off due to factors the user cannot observe, this conclu-
sion is difficult to draw. The diagnostic problem is no easier when the system
consists of knowledge and the error is an inappropriate, incorrect or unhelpful
response to some situation. How does the brain know which revision of the
underlying practical knowledge to choose out of the infinitely many logically
possible revisions? Something must give direction to the revision process if
the faulty strategy is to be revised in such a way that the fitness between action
and task environment improves. But what is the lesson that is embodied in an
error? How, by what process, can the brain use the information residing in a
constraint violation to infer the best, or at least a good or useful, revision of the
faulty rule? One approach to this problem is to consider the origin of errors.

The Origin of Errors

The commonplace observation that things can go wrong in different ways
suggests that there are distinct sources and types of errors. Different types of
errors might need to be corrected in different ways, by different processes. For
example, James Reason and Donald Norman have proposed a useful distinction
25
between slips and mistakes. Slips are errors that occur because the person does
not do what he intended to do. They are due to some glitch in the execution
of the person’s strategy – a momentary lapse of attention, for example – and
they are not informative with respect to the correctness of his relevant practical
knowledge. Mistakes, on the other hand, result from the (faithful) execution of
a faulty strategy. A child who intends to write “3” in an arithmetic problem but
writes “5” instead makes a slip; a child who consistently borrows without decre-
menting makes mistakes. Slips are not the focus of attention here.
Are there different types of mistakes? James Reason divided mistakes
26
into rule-based mistakes and knowledge-based mistakes, whereas Heinz
Heckhausen and Jürgen Beckmann distinguished implementation, initiation
and termination errors. In fact, so many taxonomies of errors have been pro-
posed that John W. Senders and Neville P. Moran felt compelled to propose a
taxonomy of such taxonomies! The proliferation of error taxonomies does not
inspire confidence in the proposed divisions of mistakes into types.