Page 408 - Deep Learning

The Recursion Curse 391

inductions, then the defense of our inductive inferences is circular: They are
justified with reference to a principle of inductive inference that is itself the
result of induction over past experiences.
There are differences between Hume’s formulation of the problem and
mine. The point made in this book is not epistemological but psychological.
The problem of generalization, of inducing new regularities from particulars,
is not as problematic for a psychologist as it was for Hume: It is an empiri-
cal fact that people construct generalizations, and the task of the psychologist
is primarily to describe how, by what processes, they arrive at those gener-
alizations and only secondarily to comment on their validity. My argument
presupposes that a person does in fact possess general knowledge structures
and that some of those structures do in fact apply to at least some of the situa-
tions in which he finds himself. The psychological question is not whether he
ought to apply them but how, by what processes, he decides whether to apply
them. Does he conclude that the situation at hand belongs to a tight context in
which regularities from past experience can safely be projected, or that there
is enough turbulence that the situation needs to be handled by drawing back
and leaping in a novel direction? The theory of deep learning is an attempt
to explain when and why people take one or the other route, not to evaluate
whether their choices tend to be correct.
Another difference between the argument of this book and Hume’s prob-
lem is that the present argument does not locate the source of the difficulty in
epistemic practices or in our cognitive system but in the world itself. Hume
did not doubt that there are valid regularities to be discovered. The difficulty,
as he saw it, is to identify the valid ones and to know with certainty that the
identification is correct. In his perspective, there is hope that a rigorous jus-
tification for inductive inferences will one day be found. We do not have one
now because we have not thought of it yet. Indeed, in Hume’s own words,
“… a man is guilty of unpardonable arrogance who concludes, because an argu-
4
ment has escaped his own investigation, that therefore it does not really exist.”
In the deep learning perspective, the matter stands differently. The world is a
turbulent, complex system in which change is the only constant and the laws
of change are themselves changing; reality has no rock bottom. The problem
of projectability arises out of the characteristic turbulence of reality, so no gen-
eral, once-and-for-all solution is possible.
The upshot is that as we go through life, circumstances will force us to
make judgments about whether to project prior experience or whether to
override it, but the decision cannot be reduced to a set of rules. Sometimes we
hit it right and sometimes we do not. The best we can do is to remember that