Page 396 - Deep Learning

Elements of a Unified Theory 379

alternative, incompatible representations are resolved by the cognitive system
on the basis of quantitative properties that reflect past usefulness as well as the
current context.
To summarize, the unified theory claims that non­monotonic change
occurs through a process that cycles back on itself: layered, selective and
capacity­limited feed­forward processes generate representations that are more
or less useful for the tasks a person undertakes. Because there is no mechanism
to maintain global coherence, the person might end up creating representations
that contradict each other in the sense of depicting materially incompatible
states of affairs or recommending mutually exclusive actions. A conflict of this
sort might remain latent for some time. it becomes manifest when the conflict­
ing representations happen to be activated at the same time. selection depends
on the relative activation levels among the relevant options. non­monotonic
changes occur when feedback tips the balance among the options available at
a particular point in the system. The resulting change is propagated forward
through the processing layers. if the propagation is amplified, it might trigger
changes of successively greater magnitude at each higher processing layer, until
the person finds himself maintaining a different and incompatible problem rep­
resentation, strategy or belief. The change may or may not take expression in
behavior or discourse. see Table 11.2 for brief summaries of the principles. The
three micro­theories stated in Chapters 4, 7 and 10 are three different instantia­
tions of these principles for three types of non­monotonic change.

necessary, sufficient, or Both?

The question arises whether all the abstract processing principles are necessary
conditions for non­monotonic change or whether a cognitive system could be
built that only satisfies some of them but nevertheless is capable of non­mono­
tonic change. There is no rigorous way to demonstrate either a positive or nega­
tive answer. But if we try to imagine a cognitive system without each of these
properties in turn, intuition favors necessity: Without layers of representation,
the system is stuck in a single possibility space: if there is no deeper layer to
reach into, there is no way to overcome the limitations on the current possibility
space. Without selective, feed­forward processing and without feedback links,
there is no way of altering the choices made at a lower level. Without the ability
to propagate a change through the system, a point change at a lower level might
not have any consequences. And so on. such thought experiments suggest that
a cognitive system that can learn from experience must exhibit all the properties
described by the abstract principles to be able to override that experience.