Page 385 - Deep Learning

368 Conclusion

basic processes, as in the cognitive architecture approach, i have addressed
each type of non­monotonic change on its own terms, so to speak, and pro­
posed a cognitive mechanism that is specifically designed to resolve the ques­
tions and issues that pertain to that type of change. in the case of creative
insight, the initial problem representation is abandoned in favor of a new
one that might break the impasse; in the case of adaptation, the initial, overly
general solution method is constrained to apply only to situations in which
it does not cause errors; and in the case of belief revision, truth values are
converted when the relevant area of experience is subsumed under another,
incompatible but more effective belief system. lined up side by side, these
three mechanisms appear unrelated in that they are composed of qualitatively
different basic processes.
But a list of unrelated micro­theories does not serve the purpose of uni­
fication. it leaves questions unanswered: How does each mechanism relate to
the others? What do the three cases of non­monotonic change have in com­
mon and how do they differ? The question arises whether deep learning –
non­monotonic cognitive change – is a label of convenience, a verbal handle
on an arbitrary bundle of research topics shaped primarily by the investi­
gator’s interests and personal history. or is deep learning a natural kind, a
type of cognitive change that can be characterized in an abstract, principled
way, analogous to, for example, speciation in biology, combustion in chem­
istry or wave propagation in physics? if deep learning is a natural kind, the
explanations for particular instances ought to share certain properties, and
unification can be achieved by capturing those properties in a set of abstract
principles.

PRINCIPLES OF DEEP LEARNING

The principles proposed in this section specify properties that hold across the
processing mechanisms postulated in the three micro­theories of creativity,
adaptation and conversion. in conjunction, they constitute a first draft of a
unified theory of deep learning; see Figure 11.1. The question whether these
properties are necessary, sufficient or both for a cognitive system to be able to
override prior experience is discussed after the principles have been stated.

spontaneous Activity
A human brain is never at rest. At any moment in time, millions of brain cells
propagate their signals downstream to other brain cells. Activity is the natural