Page 60 - Deep Learning

The Nature of the Enterprise 43

two previously unconnected ideas or concepts, is the oldest learning mecha-
nism of all. The notion of generalization (abstraction, induction) – the idea that
the mind extracts commonalities from sets of instances – has likewise been with
us since antiquity. Over time, psychologists have coined a wide variety of terms
to refer to what are ostensibly different types of cognitive change: association,
automatization, belief revision, categorization, chunking, conceptual change, con-
cept learning, conditioning, cognitive development, discrimination, equilibration,
generalization, habit formation, implicit learning, maturation, memorization,
perceptual learning, induction, knowledge acquisition, list learning, skill acquisi-
tion, schema extraction, stage transition, strategy change and theory change.
This plethora of technical terms implicitly claims that cognitive change
is a heterogeneous phenomenon and intuition provides some support. It is
certainly plausible that, for example, conceptual knowledge and skills are
acquired via different mechanisms. Furthermore, knowledge can change both
by becoming more abstract and by becoming more specific, two processes
that are each others’ opposites and hence difficult to explain with a single
mechanism. We should not expect a final theory of learning to pull together
all of cognitive change in a single law of learning to take its place next to
Isaac Newton’s law of gravitation as one of the triumphs of science. 44
On the other hand, the proliferation of terms is to some extent an arti-
fact of the sociology of research psychologists. Several of the terms label a dis-
tinct field of research, complete with its own pioneers, laboratory techniques,
phenomena, concepts, theories, controversies, authoritative reviews, heroes,
villains and in some cases its own conference and scientific journal. The prefer-
ence for developing one’s own concept is often rooted less in empirical evidence
than in dislike for the philosophical assumptions that lurk in the terminology
of others, and it is sometimes easiest to do cutting-edge research by moving that
edge closer to one’s own work with a deft terminological invention. In the race
for research grants, the appearance of doing something novel might be deci-
sive. It is also easy to confuse topic with approach and to regard, for example,
work by Gestalt psychologists on insight as a different field of research from
information-processing research on heuristic search, even though they study
the same phenomena. Another example is the work on cognitive consistency
45
from the 1950s and 1960s, which is all too easily dismissed as either old, mere
social psychology or both, even by those who study conceptual change, a field
that keeps itself too busy rediscovering what was learned back then to access
46
the original texts. To the extent that the proliferation of purported types of
change is driven by such sociological forces, it overestimates the heterogeneity
of cognitive change.