From Bob Clark (931119.1630 EST)

Hans Blom, (931117)

You refer to my post, Bob Clark (931113.1715 EST):

Instead of computing the needed predictions, the predictions could
be based on recorded observations of similar situations. Given a
modest library of such recordings, selection could begin with a
preliminary match, followed up with more accuracy as data
accumulate. "A first approximation, improved by iterating the

You comment:

Selection need not play a role. The process is more like building
up a "running average" that improves (whose standard deviation
becomes smaller) as more observations arrive.

Of course "selection need not play a role" -- if the individual has
the necessary skills for calculating a "running average." If only
minimal analytical skills are available, "selection" from recorded
observations appears an adequate procedure. "Selection" would be
based on the suitability of the remembered outcome.

You further comment:

                                              Another example is
the averaging process by which an auditory evoked potential
gradually becomes lifted above the amplitude of the noise when more
and more responses are recorded, despite the much greater amplitude
of the background EEG. Even though the signal-to-noise ratio is
terrible, the wanted information can be acquired because it repeats
and because we can observe many of those repeats. This is my basic
paradigm for learning, if I have to explain it in a hurry.

Such an accumulation of incoming data can certainly increase the
probability of a response. This concept seems the same as
"reinforcement," in "explaining" learning. To me, "description of
learning" would be better terminology. For such repetition to be
effective, the "repetitions" must be identical. Otherwise additional
noise is introduced. At the lowest level, the accumulation concept
appears workable. At higher levels, the occurrences of identical
events seems quite unlikely.

However, derivation of the entire concept of learning from your
"basic paradigm of learning," seems quite extreme. My example of
learning to toss a ball suggests additional considerations. Here the
recording includes not only the perceptions of the motions, but also
the accompanying series of reference signals producing the throwing
action. Ball throwing is rarely identically the same, so a group of
related recordings is needed. These could be averaged, or otherwise
combined to form a single result, but the ball player needs a
"vocabulary" of performances to fit changing needs. "Selection" from
among existing recordings seems close to a "minimum" description of
the process. More complex data reduction might be very useful for
the design of a mechanical system.

Learning a "performance" is more complex than learning to "recognize"
and identify. Each can be described in terms of "recordings" and
subsequent "selection" based on higher level requirements.
Additional manipulation, averaging, integrating, comparison with
other situations (also recorded) requires having learned these
procedures -- which can also be treated in terms of prior recordings,
and applied as directed by higher levels.

Regards, Bob Clark