Page 111 - Deep Learning

94 Creativity

Problem Perception

Before someone can be said to be thinking about a problem, he must have
grasped the problem and formed at least a preliminary idea of what it means
to solve it. These two components are called the initial problem state and the
goal. For brevity, I focus on situations in which the problem solver encounters
the problem through visual perception, through language or through some
combination of the two. The theory of insight need not wait for a complete
theory of either visual perception or language comprehension, but it needs to
incorporate some of their principled features.

Visual perception
When a person’s eyes alight on an object, the percept appears to be constructed
instantaneously, without complex processing and without cognitive effort, and
what the person sees appears to him to be determined by the object itself: You
see a chair because there is a chair there. The speed, lack of subjective effort,
high veridicality and intersubjective validity of perception trick us into believ-
ing that perception is a passive recording of reality. Philosophers and psy-
chologists tirelessly repeat that this is not so. Perception is a constructive,
16
selective process of considerable complexity and a percept is an interpretation,
not a recording, of the physical situation in front of the eyes. The machinery
17
of interpretation has several properties that play key roles in insight.
Perception is layered. The layers are both anatomical – layers of neurons in
the visual cortex – and functional. Each layer consists of processing units that
18
receive input from the layer below, operate on what they receive and pass on
their outputs to the next layer. A processing unit corresponds to what we for
lack of a better term might call an encoding rule or a feature detector that reacts
to constellations of features and creates a higher-order feature that is passed
on to the next layer. The encoding rules can be conceptualized as having the
general form
1 {
ff … f } → g
,
,
,
i
2
where f , f , …, f are features identified in layer N, while g is a feature at layer
1
2
i
N+1. A processing unit can pass its result – the higher-order feature g – onto
multiple units in the next layer. To the extent that the processing unit trans-
mits its output selectively, along some of its outbound links but not others, it
constitutes a choice point.
The units in the bottom layer of the visual system detect very simple
perceptual features. The latter are elaborated by the intermediate layers. The