Page 270 - Deep Learning

Error Correction: The Specialization Theory 253

NORBERT WIENER’S INSIGHT

The acquisition of practical knowledge is shaped by the facts that we necessar-
ily operate on the basis of a small and unrepresentative sample of experiences,
that our material and social environments are turbulent and that we frequently
move to colonize unfamiliar task environments. Under those conditions, prior
experience is a limited guide to action, and errors and failures are unavoidable.
The error rate can only be reduced locally, in tight contexts and over short
periods of time.
According to the constraint-based specialization theory, errors are not
merely passively eliminated as a side effect of the acquisition of the correct
knowledge. Errors play an active part in their own elimination by providing
information that gives direction to the adaptive process. People have always
understood that errors are learning opportunities – witness proverbs like burnt
child dreads the fire – but they did not know how to conceptualize the infor-
mation that resides in errors. A negative outcome is not, by itself, informative,
and neither is the frequency with which a particular type of error occurs or
the type of situation in which it tends to occur. Comparisons between multiple
situations that produced positive and negative outcomes are informative but
they require psychologically implausible computations.
It was not until the cybernetics movement in the 1940s and 1950s that it
became clear that the information provided by errors resides in the deviations
between expected and observed outcomes. In his book Cybernetics, pub-
62
lished in 1948, the mathematician Norbert Wiener argued for an interdisci-
plinary science centered on the concept of feedback circles. Wiener’s genius
63
was to realize that intelligent, purposeful action, whether by animal, human
or machine, needs to be guided by the information provided by the devia-
tions between the intended (or expected) action outcomes and the observed
(or realized) outcomes. In engineering applications, outcomes are typically
values on quantitative variables, so the deviation between an intended and
an observed outcome can be expressed quantitatively. The desired correction
is achieved by feeding the magnitude of the deviation back into the system,
which then adjusts its operation so as to bring the actual outcome closer to
the intended one. For example, the operation of a thermostat is controlled by
the difference between the set temperature and the actual temperature in the
room.
The idea of a comparison between the projected and observed outcomes
of an action is more general than Wiener’s strictly quantitative definition of
negative feedback. The constraint-based specialization process described in