Page 398 - Deep Learning

Elements of a Unified Theory 381

feedback again and again robs an option of activation, some other, incompat­
ible option must eventually become stronger, given that activity is the natural
state of the cognitive system.
The unified theory does not compete with theories of the cognitive archi­
tecture. it does not propose an alternative set of basic processes to compete
with the processes proposed in various theories of the cognitive architecture.
instead, it places constraints on the latter: if the abstract processing princi­
ples in Table 11.2 are indeed both necessary and sufficient for non­monotonic
change to occur, then any model of the cognitive architecture, whatever its
nature, has to implement those properties, lest it be unable to explain creativ­
ity, adaptation or conversion. To the extent that dolphins and chimpanzees are
capable of non­monotonic cognitive change, their minds should turn out to
instantiate these principles as well. Artificial intelligence systems and robots,
to the extent that we want them to mimic the human capability for non­
monotonic change, will have to be built in accordance with the deep learning
principles. When space aliens finally land, we should expect their cognitive
architecture to instantiate the deep learning principles, unless they borrowed
their spaceships from some other species.

EVOLUTIONARY ORIGINS
An explanation for non­monotonic learning, or for any other cognitive func­
tion, is stronger if it suggests how that function came into being. How did we
acquire the capabilities of creating novelty, adapting to unfamiliar environments
and revising our beliefs? in the absence of a hypothesis about origin or source, a
computational explanation for a cognitive function is like a rabbit pulled out of
a hat. The act of pulling the rabbit out is easy; the trick is to get the rabbit into
the hat in the first place. likewise, a cognitive theory is more plausible if it sug­
gests particular explanations for the origin of the mechanisms it postulates.
A noteworthy feature of the three non­monotonic change mechanisms
proposed in this book is that the components of those mechanisms are not
themselves non­monotonic processes. That is, the three micro­theories do not
hide a homunculus, a black box that carries most of the explanatory burden.
There is no “idea generation” box in the theory of insight in Chapter 4, no “flex­
ibility” box in the theory of adaptation from Chapter 7 and no “evidence eval­
uation module” in the theory of belief revision in Chapter 10. The component
processes are either routine processes or monotonic learning processes. They
are not exotic and they are not specially designed for the purpose of explaining
non­monotonic change. They have typically been proposed to explain a wide