Page 407 - Deep Learning

390 Conclusion

prior experience and consequently experience a higher rate of non-monotonic
change. But the paradoxical implication of turbulence is that a higher rate
of change will not necessarily be associated with a higher level of success. A
stronger disposition to change might not by itself be useful unless we can also
learn something about the conditions under which we are better off projecting
prior experience and the conditions under which we are better off drawing
back to leap. The ideal is not to change with high frequency, but to project
prior experience when the situation at hand is like prior situations and hence
does in fact lend itself to be handled on the basis of experience, and to draw
back and leap when the situation at hand is, in fact, essentially different and
requires a novel response.
It is tempting to believe that we can learn to differentiate one class of situ-
ations from the other, and hence become more accurate in choosing between
extrapolating and leaping. After all, a person’s lifetime contains many examples
of both types of situations, and the study of history extends that database over
generations. If we analyze this accumulated experience from the deep learning
point of view, we might be able identify conditions that reliably characterize
situations in which we should project rather than leap and vice versa, and so
improve the accuracy of our choices.
A moment of reflection reveals the fallacy hiding in this temptation. Our
knowledge of the situations in which projection worked well and the situa-
tions in which drawing back to leap worked better is itself a strand in our past
experience. It is subject to the same turbulence as all experience: We cannot
know whether our past experience of when past experience is useful, is itself
useful in any future situation. That is, we cannot project whatever experience
we have of projecting past experience onto the next situation. Are we better
off overriding that experience and making choices as to when to project and
when to override in some novel way? The argument against the projectability
of past experience applies recursively to our past experience of projecting past
experience; similar with our experience of drawing back to leap. In program-
ming terminology, turbulence is recursive; that is, it applies to itself. Hence, we
cannot, even in principle, learn from experience when to project prior experi-
ence and when to override it.
Philosophers recognize this as a version of “Hume’s problem” or “the
problem of induction.” David Hume asked an epistemological question: How
3
can we know that a regularity induced from past examples is true of the entire
set that the examples were drawn from? The problem, said Hume, is that if
we apply a principle of inductive inference, what guarantees the validity of
that principle? If the principle is itself a generalization over successful past