Page 233 - Deep Learning

216 Adaptation

Figure 7.1. The difficulty of inferring the appropriate rule revision, given a falsified
expectation. The top light is “off” although it is expected to be “on.” The required diag-
nostic inference that the two conditions indicator light is on and bottom light is on
should be added to the relevant rule.

environment can sometimes be altered or reversed through further action. This
is often what is meant in everyday contexts by correcting an error. An example
is to edit a misspelled word. The person reacts to the undesirable outcome – a
complaint from the spell-checking software – by taking further action. If that is
his only response, he might misspell that word again in the future. This mean-
ing of “error correction” is not central for the theory of learning from error.
An action is the overt expression of a (not necessarily conscious or delib-
erate) decision. If the action is incorrect, then the practical knowledge on
which the decision was based is erroneous. The overt error (the erroneous
action) must be distinguished from its cause, the flaw in the underlying prac-
tical knowledge. Since the seminal study of arithmetic errors by John Seely
Brown, Richard R. Burton and Kurt VanLehn, flaws in practical knowledge
are often called bugs in analogy with mistakes in computer code; science edu-
23
cation researchers use the term misconception; I prefer the term fault. To
learn from an action that generates an undesirable outcome is to eradicate
the fault (or faults) in the practical knowledge that recommended that action.
As a consequence, the learner has a lower probability of committing further
errors of the same type. To correct an error in this sense is to improve future
action, as opposed to revising the undesirable outcome of an action already
executed.
The difficulty of inferring the underlying fault (or faults) in a complex
system from observations of system behavior is well known among computer
24
programmers, electricians and physicians; see Figure 7.1 for an illustration.
The square symbolizes a device with two input buttons, an indicator light and,