Page 221 - Deep Learning

204 Adaptation

in empirical observations? on this dimension, each of the nine modes does as
well as any hypothesis in psychology. All nine modes are supported by system-
atic empirical studies as well as by common sense. Some mechanisms, such
as learning by analogy and learning from solved examples, have been sup-
ported in dozens of studies. From this perspective, the nine Modes Theory is
as empirically grounded as any theory in cognitive psychology.
The methodological implication of multiple mechanisms is that observ-
able changes in performance are not due to any one mechanism but to the
interactions among the different change mechanisms. We should not expect to
derive empirical phenomena from any one mechanism by itself, even though
the majority of research studies attempt precisely such single-mechanism
tests. The more telling test would compare data with predictions derived from
a model that learns in all nine ways. The question is what to use as a baseline.
What counts as a good versus unsatisfactory fit to the data? We do not have
several nine-mechanism theories to compare; as long as we accept that differ-
ent implementations of a mechanism cannot be reliably differentiated with
the kinds of behavioral data currently available, then we have only a single
ninemode theory. Estimates of goodness of fit are then meaningless. one par-
tial solution is to compare the full nine-mode theory with models that lack one
or the other mechanism. The issue is whether each of the nine mechanisms
adds predictive power to the model, that is, whether it does a better job of
accounting for data than a model without that mechanism. if this is true for all
nine mechanisms, then the complete theory is supported. James J. Jewett and
i made a modest move in this direction by building a simulation model that
could learn from either positive or negative feedback or both simultaneously.
68
Simulation runs with the model did indeed show that having both mecha-
nisms operating at the same time produced better fit to human learning curves
than either mechanism by itself. A serious test of the full nine Modes Theory
requires a simulation model that can learn in all nine ways. developing such a
model is a task for the future. The first century of skill acquisition research pre-
pared the ground and put the main conceptual support in place, but finalizing
the theory might take a second century. A reasonable first step is to improve
the theory of learning from negative feedback. This is the task of Chapter 7.