For two years now, I have been exploring the notion that the most
important product of our brain - imagination - could well be an effort to
*neutralize* our sensory input, preventing it from penetrating the neural
networks of our brain. The results include a computer program that can
effectively construct such a network and a number of (in my opinion)
interesting observations that I will discuss in the introduction below.
Introducing this model is quite hard, since it has no loose ends: it
leaves us with no natural place to start discussing it. For this reason
the essence of this model does not lie in the beginning or end of its
description, but in the features of its workings.
1 How can a network learn to neutralize its sensory input?
By linking the neutralization level to global neural growth and
starvation, connections that increase neutralization can be collected over
time: Increasing neutralization strengthens all connections and grows new
(weak) ones. Decreasing neutralization weakens all connection, destroying
the weakest (new) ones. Note that a single *quantity* is monitored in
order to develop a *quality*.
2 What feeds our imagination if sensory input is completely neutralized?
The answer to this question is short: nothing. However, perfection is not
the issue here. Neutralizing (correctly predicting) ninety percent of our
input would be an excellent performance in the complex world we inhabit.
Note that the *gain* of the network has to increase in order to neutralize
ninety percent while receiving only ten percent of the actual input.
3 What prevents excessive feedback if reality takes an unexpected turn?
The increasing gain of the network proved to be a great obstacle for its
development. New connections were selected for their neutralization of
already sparse patterns (due to the effect of the network already in
place). However these connections proved to be far too sensitive to face
the original, complete input. In other words, the network could not simply
be switched on facing a familiar pattern: it had to 'replay' its evolution
very quickly in order to regain the original level of performance. The
neural correlate of this mechanism is suggested to be the modulating
effect of our *diffuse systems of arousal*: more recent connections are
suggested to be more sensitive to this modulation, resulting in the
required (partial) 'replay' of the network's development with each EEG
4 How could neutralization increase our chances survival?
Behavior, in this model, is interpreted as a means to increase the value
of our mental neutralizing capabilities. A child inevitably (since its
lack of control over its environment) has to learn to neutralize its
perceptions, while later on it will use the control it gains (by learning
to communicate and by developing senso-motorical skills) to preserve or
recreate the kind of environment it has already learned to neutralize. If
its initial environment has been a socially and physically healthy one,
this tendency provides an effective strategy for survival.
Feel very free to comment!
With kind regards, Mervyn
o This model has been presented in the Gronics'97 conference
article: "Feedback in Knowledge-Oriented Neural Networks"
Available at http://www.xs4all.nl/~mervyn/vankuyen.html
o Source code (Macintosh only) on request: mervyn at xs4all.nl