Page 224 - Deep Learning
P. 224
Error Correction: The Specialization Theory 207
to use the office key in the apartment door. Actions are only correct or incor-
rect in relation to some situation. To learn from an error is to distinguish the
situation in which the action is an error from those in which it is correct.
Adjusting the strength of an action or rule downward does not accomplish
this, because the adjustment lowers the probability of the relevant action in
every situation.
William James devoted four pages of his Principles of Psychology to an
explanation of how a child who is burned by a flame avoids being burned by
7
that flame again. According to James, “… we know that ‘the burnt child dreads
8
the fire,’ and that one experience usually protects the fingers forever.” The per-
ception of the flame, the reflexive action of reaching for it, the sensation of
pain and the equally reflexive retraction of the arm become associated in the
child’s mind, in that sequence. The next time the child sees the candle flame,
the action of reaching for it is initiated by the same lower, reflexive neural
pathways that made the child reach for the flame the first time, but it is inhib-
ited by the higher brain centers due to the chain of associations from the sight
of the flame to the retraction of the arm. This pseudo-neural mechanism does
not suppress the disposition to reach out and grab interesting objects, only
the tendency to perform this action with respect to the candle flame. James’s
explanation has the peculiar implication that errors are never truly corrected.
Error-producing response tendencies remain after learning. We commit fewer
erroneous actions over time because we spend more and more of our waking
time inhibiting those tendencies. A better theory of learning from error should
describe how the disposition that led to an erroneous action is altered so that
the action is no longer triggered in the situation in which it causes the unde-
sirable outcome, thus relieving the higher brain functions from the need to
monitor and inhibit.
James’s explanation is an attempt to capture the intuition that the cure for
an error is to avoid repeating the same action in the same situation in which it
caused trouble. But the phrase “in the same situation” overlooks the fact that
situations, once past, are history. If we perform action A in situation S and
experience a negative outcome, in which future situations should we avoid
doing A? It is not enough to say in S, because by the time the learner knows
the outcome of doing A in S, that situation is past and will never recur. We can
say, in situations that resemble S, but situations resemble each other to vary-
ing degrees along different dimensions. The better answer is, in situations that
resemble S in the relevant aspects, which immediately leads to the question of
how our brains know which aspects are relevant. If the candle hurts, what is
the relevant class of situations? The blue flame on a gas stove does not look like