Page 364 - Deep Learning

Belief Revision: The Resubsumption Theory 347

the environment informs the person about how well the actions derived from the
theory work – that is, how good a tool the theory is for solving life’s cognitive
tasks. Environmental feedback takes different forms in different types of tasks.
For example, a poor decision, informed by an inadequate theory of the relevant
domain, might lead to an unnecessary waste of time, a type of outcome that can-
not be cast as a logical contradiction. If a theory works well – that is, if evoking it
enables the person to successfully solve the cognitive tasks he undertakes – then
the estimate of its utility increases; if the theory gets the person into trouble, then
the utility is decreased. Over time, the feedback generated by task performance
converges on a stable estimate of the cognitive utility of the relevant theory.
The contender theory Th(A+B) does not start at zero utility. That the-
ory emerged because it was needed to make sense of domain A and its utility
reflects its usefulness in that domain. Hence, the two competitors might start
out at comparable utility levels. They might be running neck to neck, as it
were, for some time, but if Th(B) generates negative feedback and Th(A+B)
generates positive feedback, the advantage of the former will decrease. The
person might switch back and forth between them with a gradually increasing
awareness that his thinking is incoherent and that a choice has to be made as
to what to believe. If the contender theory continues to accrue utility through
successful use, it will eventually surpass the resident theory.
If the contender theory is in fact more veridical than the resident theory, its
use will tend to increase its estimated utility. At first glance, this state of affairs
seems utterly implausible. Why would a theory that arises from an attempt to
understand domain B be less useful for dealing with that domain than a the-
ory formed in response to information about some other domain? When we
initially become familiar with a domain, we tend to perceive, categorize and
remember the domain in terms of perceptual features and simple relations
among its surface features. We spontaneously carve the world at the joints as
we see them. However, surface validity goes only so far, because surface features
are not, in fact, perfect, and sometimes not even good, predictors of underlying
mechanisms. Sharks and dolphins look similar but are quite different animals,
glass and diamond look similar but are quite different substances and so on.
The history of science is a long testimony to the fact that our perceptual systems
do not carve the world at its true joints. Effective and accurate theories require
that we stand back from the surface appearances and parse reality according to
features that might not attract our attention at first. Consequently, a theory that
began as an attempt to understand some other area of experience might hap-
pen to slice that domain in such a way that when its structure is superimposed
on another domain, it provides a better account than the theory initially and