Page 86 - Deep Learning

The Production of Novelty 69

The ways in which these theories provide direction by taming the com-
binatorial explosion leave something to be desired. The first issue is how the
mind selects the set of concepts, ideas or mental elements that should be con-
sidered as potential ingredients in a new combination. Poincaré’s appeal to the
preparatory phase of deliberate problem analysis as a source of constraints on
the unconscious generator is more problematic than helpful: The claim that
conscious work on a problem tends to activate a set of relevant concepts is
plausible, but it poses the puzzle of how we ever come to apply a concept that
we did not anticipate to be useful for the problem at hand. This hypothesis
does not help explain how the unconscious idea generator can produce a com-
bination that goes against the expectations created by prior experience.
The alternative is to place no restrictions on the cognitive elements that
can participate in new combinations, but this aggravates the problem of direc-
tion. How, given some set of initial elements, are new combinations generated?
terms like “blindly,” “randomly” and “by chance” claim that the combination
process proceeds without direction. This hypothesis is tempting because it frees
the combination process from prior experience and hence promises to help
explain how people go beyond experience. it also frees the mind – and hence
the theorist – from having to solve the unsolvable problem of how to com-
pute exactly which idea combination will turn out to be the solution to any one
problem. But without any restriction on the set of components, a random com-
bination process provides no protection against the combinatorial explosion.
The particular evaluation processes proposed by these combinatorial theo-
ries are even more problematic. Poincaré’s idea of selection by beauty is difficult
to generalize to other areas than mathematics. The invention of the computer
certainly ranks as a creative project, but it is difficult to see beauty in those early
roomfuls of vacuum tubes and crisscrossing wires. a deeper problem is how,
by what process, the unconscious measures mathematical beauty. The sophis-
tication of this process makes its attribution to the unconscious implausible.
There are no reasons to believe that there exists an evaluation function that
unerringly selects the right constellation of ideas for any problem, and none
to believe that evolution has endowed our brains with the capacity to compute
such a function. Similar concerns apply to evaluation based on stability and
interestingness. Perfect evaluation filters contradict the fact that human beings
are fallible; sometimes the evaluation process picks a promising new combi-
nation that turns out not to be the solution. a closely related weakness is that
generate-and-test mechanisms do not explain how the direction of a creative
project is impacted when an idea that is not in fact useful nevertheless slips
through the successive envelopes of selection and leads to failure.