Page 392 - Deep Learning

Elements of a Unified Theory 375

preceding layers, so that the consequence of a choice at level N will be fed
back down to layers N­1, N­2, and so on. Both positive and negative feedback
can be passed back down through the processing layers. in the normal case,
complex representations send activation down through the layers to keep the
simpler representations consistent with themselves. These cases turn up as
context effects in psychological experiments: see a drawing as a face and the
probability increases that a line in the relevant place is interpreted as a mouth;
see a letter sequence as the word “WoRD” and the probability of seeing “o” as
the letter “o” rather than, say, the digit zero, goes up. Top­down propagation
of feedback is usually a stabilizing force.
However, because the choice in each processing unit is massively contingent
on the relative levels of activation of the competing options, feedback that propa­
gates down through the system has the potential to tip the balance among the
options in some unit. At such tipping points, the top­down feedback might have
a destabilizing effect. The perturbation propagates back up through the process­
ing layers and might result in a new representation at some higher level. if the
old and new representations are incompatible, then the switch from one to the
other looks to an outside observer like a qualitative change of mind. in short,
the cause of non­monotonic change is ultimately a change in the balance among
competing options at some low­level point in the processing system, caused by
the downward propagation of feedback derived from the outcomes of actions.
in the case of negative feedback, a shift in the balance among options
occurs because a previously dominant choice is suppressed. in such cases, the
key question in understanding the emergence of the new representation is not
how was this new representation accomplished? but rather, how and why were
the previously active representations inhibited? All that is needed for something
new to appear is that the standard operating procedure, the habitual way of
doing or thinking, is suppressed. As Edward Thorndike realized a century ago,
in a system that is continuously and spontaneously active, the suppression of
24
one option automatically leads to the execution of some alternative option.
A shift in the balance among options can likewise occur by positive feedback
that strengthens a previously dormant and hence unexplored option. in such
cases, the key question is, how did such and such an option benefit from the
positive feedback being propagated downward through the processing system? it
is possible that many cases of non­monotonic change involve the simultane­
ous inhibition of prior options and the activation of as yet unheeded ones.
The event that initiates non­monotonic learning is the propagation of
feedback – of one sort or the other – through the processing system, and that
feedback, in turn, originates in the environment; see Figure 11.2. Feedback is