Page 391 - Deep Learning

374 Conclusion

and routine belief formation processes might create contradictory beliefs that
apply to one and the same domain of experience. Because it is impossible to
maintain global coherence, a person’s knowledge base is likely to contain mul­
tiple conflicts, and such conflicts can remain latent for a long time.

Feedback and Point Changes

non­monotonic learning requires a mechanism for changing the balance
between competing and incompatible processing options. if one option beats
its competitors at some moment in time, why does it not continue to do so? if
it did, we would be doomed to try and try again the same path that we have
tried before, regardless of how unsuccessful it turns out to be. if feed­forward
propagation were the sole modus operandi of the cognitive system, as it might
be in other mammals and might have been in pre­humans, then we would only
be capable of mastering tight contexts that approximate clockwork worlds.
only change begets change, so a change at one point in the cognitive sys­
tem must be due to a change elsewhere in the system. What gives creativity,
adaptation and conversion their aura of mystery is, in part, that this explana­
tory principle is circular: if we cannot explain a particular change without
referring to a change elsewhere in the system, then we must apparently explain
the origin of the latter change in the same way. Consistent with this obser­
vation, explanations for non­monotonic changes typically move the change
backward one step through the causal links and leave it there: A person breaks
out of an impasse by restructuring the perception of the problem – but what
caused the restructuring to happen when it did and not earlier or later? What
changed? A new idea comes about through a remote connection between
two distinct concepts – but why was the person suddenly making a remote
connection not made before? What changed? A person forms a new belief in
response to anomalous information – but anomalies are always accumulating,
so why did this change not happen earlier or later? What changed? Attributing
a change to some other change invites the question of how or why the latter
change occurred, and this is a problem of the same type and form as the origi­
nal one and hence has to be dealt with in the same way. There is no obvious
way to terminate this regress.
in the present theory, a non­monotonic change at one point in the cog­
nitive system is caused by an altered balance among options at some lower
layer. The origin of the latter change lies in the information generated by the
activation, articulation and execution of the higher­level knowledge represen­
tations. Feedback links carry activation and information from high layers to