Page 405 - Deep Learning

388 Conclusion

“adaptability,” “creativity” or “open­mindedness.” The situation is the other way
around: it is the occurrence of non­monotonic changes (with positive outcomes)
in a person’s life trajectory that causes us, as observers, to use such labels. if a
person’s life trajectory exhibits a sustained streak of non­monotonic changes
with positive outcomes, we characterize him in one way; if his life develops in a
different way, we describe him differently. The attribution of differences among
individuals to differences in unanalyzed powers like “adaptability,” “creativity”
or “mindfulness” is a post facto characterization, not a cause and hence not an
explanation, for differences in the disposition for non­monotonic change.
When we list the factors that affect the probability of non­monotonic
change, its low probability of occurrence ceases to be mysterious: Minds are
programmed by evolution to push forward before drawing back to leap, the
relevant cognitive processes are not under voluntary control and the occur­
rence of a change requires the simultaneous occurrence of multiple conditions.
The latter may be hindered or facilitated by cognitive parameters like working
memory capacity and retrieval thresholds. Even with favorable values on the
architectural parameters, non­monotonic changes are unlikely unless a person
undertakes projects that require such changes, works hard at them, engages
in extensive preparations, subjects himself to varied experiences and exposes
himself to negative feedback. Taken together, these factors explain why non­
monotonic change is difficult even in the presence of cognitive mechanisms
that make such changes possible. Alternative theories of deep learning are
no doubt possible, but any satisfactory theory must explicate this dialectic
between difficulty and possibility.