Page 394 - Deep Learning

Elements of a Unified Theory 377

Elementary
representations
Feedback
propagation
Combinatorial
generation
Higher-order
representations
Higher-order
representations
Outcome
perception
Decision
making
Environment
Physical Motor
causation action
Figure 11.2. The deep learning circle.

There might not be any visible effects at the lowest layers; the change
becomes noticeable in overt behavior only when it has reached a certain mag­
nitude and begins to affect what the person says and what he does. To the
outside observer, this looks mysterious; where did that action or that utterance
come from? indeed, the person himself might not have conscious access to the
layers of representation at which change originated and so might be as sur­
prised as an observer by the effect on his thinking or discourse. The downward
propagation of feedback and the upward propagation of a point change work
as a team; in conjunction they explicate the phrase drawing back to leap into a
specific cognitive mechanism.
Whether forward re­propagation from a changed processing unit is
dampened or amplified is contingent on the exact distribution of activation
levels across the processing system. The construction of a complex mental
representation, be it a problem space, a rule set or a belief system, requires a
large number of choices, and the choices are made on the basis of the relative
activation levels of all the relevant units and links. Activation is in turn an
aggregate function of past experience, inputs from lower layers, inputs from
excitatory and inhibitory links within the same layer and positive and negative
feedback propagating downward from the higher layers. other factors include
the person’s goal and the exact distribution of information arriving through
perception at each moment in time. The total set of choices behind a particular
representation is massively contingent on the exact distribution of action over
the entire cognitive system, moment to moment. it is not possible to posses