Page 61 - Deep Learning

44 Introduction

A second observation that warns against overestimating the heterogeneity
of cognitive change is the fact that the brain is a relatively homogeneous com-
putational medium. All parts of the brain contain neurons, and all neurons
receive input signals from a large number of other neurons, integrate those in
some complicated and as yet only partially understood manner and respond
by sending a signal forward to another large set of neurons. Neurons in dif-
ferent parts of the brain vary in size, the number of connections and in other
ways, but they function in fundamentally the same way. This observation sug-
gests that there might only be a small number of types of changes at the neural
level, or modes of plasticity in the terminology preferred by neuroscientists.
In short, it is implausible that all cognitive change can be explained by the
operation of a single learning mechanism, be it analogy, association, chunking,
equilibration, or restructuring. Instead, theory construction should aim for a rep-
ertoire of diverse mechanisms that jointly suffice to explain the behavioral phe-
nomena associated with cognitive change. In this respect, a cognitive psychology
of learning will be more like geology than like chemistry. On the other hand,
intuition rebels against the idea that the mind employs hundreds of qualitatively
distinct change mechanisms. Steering clear of the extremes, the investigations
reported in this book assume that a unified theory of cognitive change will list a
handful of learning mechanisms. This is an imprecise guess, but the best guess.

Triggering Conditions

To describe a change in a knowledge representation in terms of a combination of
basic processes is clarifying, but for an account to be truly explanatory, it has to be
augmented with a specification of the relevant triggering conditions: When, under
which mental circumstances, does the mind apply or execute any one learning
mechanism? For example, when does the mind create a new link between two
nodes in a conceptual network? Classical accounts of association claim that links
are created when the referents of the relevant nodes are either very similar, occur
adjacent to each other in space or time (contiguity) or stand in a cause-effect rela-
tion to each other, but there might be other conditions as well. 47
The question of triggering condition is relevant for every change mecha-
nism. When, under which mental circumstances, does the mind invent a new
subgoal, revise a strategy, restrict the application of a concept or schema or
lower or raise the activation level of a representation? Robert M. Gagné iden-
tified the question of triggering condition as central by putting it in the title
48
of his 1965 book The Conditions of Learning and it remains central. Unless
the relevant triggering conditions are specified with some precision, the claim