Page 401 - Deep Learning

384 Conclusion

The issue of how easy or difficult it is to override past experience is, in part,
an issue of balance and priority between monotonic and non­monotonic
responses to situations and tasks. our brains are still designed to prioritize the
projection of prior experience, because this cognitive strategy imposes little
cognitive load and works well in tight contexts. it is possible that the increase
in cognitive ability may have given the human species sufficient control over
its survival rate to slow down its own evolution. The capabilities that sup­
port routine processing and monotonic learning enabled agriculture, techno­
logical innovation, new forms of social interaction, organized protection for
children and, eventually, health practices. These factors must have increased
the proportion of each generation of humans that lived to reproductive age.
Perhaps we are weak non­monotonic learners because evolution toward yet
greater capacity for this type of change slowed as the advantages of the already
achieved gains in cognitive ability increased the survival rate and hence low­
ered the selective pressure. Even without this factor, it is likely that the ideal
balance between monotonic and non­monotonic change was different for
hunter­gatherers who lived short lives under restricted circumstances than it
is for people who live through future shock after future shock as the pace of
26
change speeds up. These are speculative but plausible hypotheses. They imply
that our cognitive system is biased to prioritize extrapolation over restructur­
ing, projecting past experience over leaping in a new direction.
When we deliberately tackle a project that requires non­monotonic
change, we attempt to override this part of our evolutionary programming.
But non­monotonic cognitive change is something that happens to a person
rather than something he does. it is no easier to control this aspect of brain
function with an act of will than it is to exert voluntary control over the opera­
tion of the kidneys or the liver. This type of change happens when certain
triggering conditions hold and not otherwise, and the probability of non­
monotonic change is, in part, determined by the fact that there are multiple
such conditions. The relevant cognitive raw materials, the knowledge struc­
tures that represent the relevant options, have to have been created by prior
monotonic learning and the relevant feedback has to be sufficient to alter the
balance among options in one or more choice points. in addition, the distribu­
tion of strengths and activation levels across the processing system has to be
such that a change at some particular point can propagate through the cogni­
tive system in an amplified way. The probability that all these conditions are
at hand simultaneously is low. in short, the built­in preference for monotonic
over non­monotonic change, our lack of conscious control over the relevant
processes and the dependence of non­monotonic change on the simultaneous