Page 38 - Deep Learning

The Need to Override Experience 21

complexity and turbulence, change is the only constant. Furthermore, change
is thoroughgoing, and the rules that control change are themselves changing.
In this kind of world, prior experience is guaranteed to be misleading most
of the time, although it might provide a good enough approximation in local
contexts or over short periods of time. Learning in this kind of world requires
cognitive capabilities other than those implied by empirical inductivism.

DEEP LEARNING

If prior experience is a seriously fallible guide, learning cannot consist solely or
even primarily of accumulating experiences, finding regularities therein and
projecting those regularities onto the future. to successfully deal with thor-
oughgoing change, human beings need the ability to override the imperatives
of experience and consider actions other than those suggested by the projec-
tion of that experience onto the situation at hand. Given the turbulent charac-
ter of reality, the evolutionary strategy of relying primarily on learned rather
than innate behaviors drove the human species to evolve cognitive mecha-
nisms that override prior experience. This is the main theme of this book, so it
deserves a label and an explicit statement:

The deep Learning Hypothesis
In the course of shifting the basis for action from innate structures to
acquired knowledge and skills, human beings evolved cognitive processes
and mechanisms that enable them to suppress their experience and over-
ride its imperatives for action.
The deep Learning Hypothesis does not deny the existence of cognitive
mechanisms that operate on the basis of experience. Inductive learning works
in tight contexts and people obviously do possess the processes for encoding
episodic information into memory, inductive reasoning, projection and plan-
ning that are described in cognitive psychology textbooks. The type of learn-
ing supported by those processes generates new knowledge that is consistent
with what was known before. Borrowing a useful term from logicians, I refer
to such additive cognitive growth as monotonic. 52
The claim of the deep Learning Hypothesis is that monotonic learning
is at most half the story. The other half describes how we abandon, override,
reject, retract or suppress knowledge that we had previously accepted as valid
in order to track a constantly shifting and fundamentally unpredictable envi-
ronment and thereby indirectly create mental space for alternative or even
contradictory concepts, beliefs, ideas and strategies. A complete theory of