Page 112 - Deep Learning

Creative Insight: The Redistribution Theory 95

top layer writes the final output of the perceptual system into working mem-
ory. For example, the perception of an object at the very least passes through
layers that (a) identify boundaries between light and shadow, (b) determine
the object’s shape and (c) assign the object category membership. Those tasks
are likely to be accomplished in that order (presumably with yet other layers
in between). The orderliness of the sequence of layers should not be overem-
phasized – the brain is a messy system – but it is a useful idealization.
A complex visual scene causes many parallel, simultaneous streams of
processing to travel upward through the successive layers. Color is not pro-
cessed in the same place in the brain or in the same way as either shape or spa-
19
tial location. Exactly how the parallel strands of information come together
into the subjective experience of seeing a single, integrated physical object of
a particular type, located in a particular place, is still unknown. However this
is accomplished, the implication is that mental representations are combinato-
rial structures with distinct constituents.
At each layer, different interpretations of the inputs from the preceding
layer are possible. The choice of which interpretation to pass on to the next
layer is assumed to happen via a version of the construction-integration process
20
hypothesized by Walter Kintsch for language comprehension. This type of
process requires horizontal links among the processing units within each layer.
Such links can be either excitatory or inhibitory. An excitatory link between
units A and B has the effect that if A is found in the stream of perceptual input,
the system is biased toward also finding B in that stream, and vice versa. The
two features “yellow” and “banana shape” are examples. The identification of
the color of an object as yellow increases the probability that the visual system
will conclude that the shape of the object is that of a banana and vice versa.
The identification of either feature causes the other feature to be activated. An
inhibitory link has the opposite effect. Seeing “blue” would bias the perceptual
system away from “banana shape” by lowering its activation level, because we
rarely see blue bananas. Horizontal links can be thought of as constraints on
processing. They limit the set of possible interpretations of the visual input
to the most plausible ones. The mutual excitatory and inhibitory interactions
continue within each layer until the activation levels settle and the process-
ing units that were stronger than their competitors pass on their processing
products and their activation to the next processing layer, where they enter
into new combinations.
The hypothesis of a sequence of layers should not be taken to imply that
processing is bottom-up. Information from layers further up in the stack is
21
fed back down to preceding layers. For the purpose of explaining insight, the