Page 383 - Deep Learning

366 Conclusion

constant across contexts and tasks, shared by all individuals (variations along
a few quantitative parameters aside) and stable over time, at least in relation
to the duration of individual cognitive processes. The architecture is presum­
ably innate. in this view, variability in task performance across tasks, across
individuals and over time is accounted for by variations in the knowledge that
is applied in any one performance. unification is served – and this is newell’s
fundamental innovation – by the fact that the same set of basic processes is used
to explain every cognitive phenomenon. simplicity is served because the basic
processes and their modes of combination can be stated relatively concisely.
The complexity that is squeezed out of the basic machinery bulges out in the
specification of the knowledge that the cognitive architecture applies to gen­
erate behavior. The soar theory by newell, J. laird and P. s. Rosenbloom and
the ACT­R theory proposed by Anderson and co­workers are two attempts to
propose specific theories of the form that newell envisioned, but other theo­
14
rists have proposed models of the cognitive architecture as well. no proposal
has won wide acceptance, but the search for the right specification of the cog­
nitive architecture is the only theoretical programme in cognitive psychology
that provides a clear vision of what a final theory of human cognition might
look like.
The research strategy followed in this book departs from the cognitive
architecture programme. The subject of this book is not human cognition.
i believe that the latter is too vast and ill­defined a target for theory forma­
tion. To illustrate, consider whether dreaming is a cognitive phenomenon.
on what grounds would it be excluded? Dreaming appears to be a process
of manipulating mental representations, the very stuff of cognition. But if
dreaming is included among bona fide cognitive phenomena, then no contem­
porary theory of the cognitive architecture stands any chance of being even
approximately correct. We are too far from understanding the alterations in
consciousness during the wake­sleep cycle, let alone the function of dreaming.
similarly, a complete theory of cognition should explain the role of emotions
in cognition, the distinction between meaning and reference in language and
the digital­to­analogue conversion that presumably occurs somewhere on the
path from decision to movement. These and other research problems are so far
from being solved that proposing theories of human cognition is premature.
The work reported in this book follows the alternative strategy of focusing
on a single, particularly interesting cognitive phenomenon: non­monotonic
cognitive change. People demonstrably undergo such changes and the goal
of explaining them poses theoretical challenges that are significantly different
from those of explaining stable performance or monotonic learning. The first