Page 400 - Deep Learning

Elements of a Unified Theory 383

For example, layered processing – the presence of both excitatory and
inhibitory links within a layer and the presence of downward feedback links –
might have emerged because they improve the accuracy and utility of vision,
independent of any need to override prior experience and re­represent unfa­
miliar situations. But once these features were in place, they provided the pos­
sibility of re­perceiving the environment. likewise, the capability of growing
cognitive skills by specializing them might have evolved as a component of the
ability to acquire new skills. But once specialization and conflict­resolution
mechanisms that are sensitive to specificity were in place, they provided the
ability to override an already learned skill as well as creating a new one. Finally,
the ability to retrieve declarative representations from memory is useful in a
wide range of knowledge­using scenarios, but once this ability was in place,
remote associations – bisociations – became possible. For each micro­theory,
it is possible to hypothesize selective pressures toward greater capacity that
are independent of the need for non­monotonic change but which eventually
endowed the relevant cognitive mechanisms with the properties that make
such changes possible. This two­phase scenario is speculative but consistent
with how complex adaptations evolve.

ON THE DIFFICULTY OF NON­MONOTONIC CHANGE
scientific theories have many functions. Conceptual advances sometimes sup­
port new technologies, but another, equally important function is to help us
understand puzzling aspects of life. The most interesting puzzle regarding non­
monotonic change is its low probability of occurring. Why are we so eager to
push forward but so reluctant to draw back to leap in a new direction? if we are
capable, in the sense of possessing the requisite cognitive processes, to see, think,
speak and act in novel ways, why do we so often fail to do so when the circum­
stances call for it? We are stuck on unfamiliar problems – slow to adapt old hab­
its to novel tasks but quick to sweep belief­contradicting information under the
cognitive rug. When goal attainment is blocked, our first response is to expand
more effort, to execute the tried and true strategies with greater energy, greater
attention to detail and higher fidelity to the intended course of action. We push
forward before we draw back to leap. Why is this the case, if we evolved a capa­
bility for non­monotonic change during the hunter­gatherer era?
The first part of the answer is that the two­stage evolutionary process envi­
sioned in the previous section did not replace monotonic with non­monotonic
learning, just as the evolution of vision did not replace the prior senses of smell
and touch. The new function was superimposed on top of prior functionality.