Page 293 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

286 CHAPTER 14 Meaning Versus Information, Prediction Versus Memory

spike-timing-dependent plasticity can help explain more complex phenomena such
as orientation flash lag effect (see Ref. [5] and references within).
Second, we will consider predictive properties in brain dynamics and how it can
be related to higher level phenomena such as consciousness. As we discussed above,
prediction seems to be a key function of the brain. How can it also be used to gain
insights into phenomena such as consciousness? In consciousness studies, the neural
correlate is highly sought after, where neural correlates of consciousness refer to the
“. neural events and structures . sufficient for conscious percept or conscious
memory” [6]. This view is somewhat static (of course it depends on the definition
of “event”), and its dependence on sufficient condition can lead to issues relating
to the hard problem of consciousnessdhow and why it “feels” like it. In our
view, it would be better to first consider the necessary condition of consciousness,
and this led us to the realization that the property of brain dynamics, not just isolated
“events,” need to be considered. We also found that predictive property in brain dy-
namics has an important role to play in consciousness, and this is how the discussion
of consciousness comes into picture in this section [7].
Let us consider necessary conditions of consciousness. We begin by considering
consciousness and its subject. There cannot be a consciousness without a subject,
since consciousness, being a subjective phenomenon, cannot be subjective without
a subject. Next, consider the property of the subject (or let us say self). Self is the
author of its own actions, and there is a very peculiar property about these actions
authored by the selfdthat it is 100% predictable. When I say “I will clap my hands
in 5 seconds,” I will make sure that happens, so that my behavior in such a case is
100% predictable. This is quite unlike most phenomena in the world that is not so
much the case. In order to support such prediction, some part of the brain has to
have a dynamic pattern that has a predictable property. That is, based on past data
points in the neural dynamic trajectory, it needs to be possible to predict the current
data point. This, we believe, is an important necessary condition of consciousness
(see Ref. [7] for details). Through computational simulations and secondary analysis
of public EEG data, we showed that predictive dynamics can emerge and have
fitness advantage in synthetic evolution [7], and conscious states such as awake con-
dition and REM sleep condition exhibit more predictive dynamics than unconscious
states (slow-wave sleep) [8].
For the first study [7], we evolved simple recurrent neural network controllers to
tackle the 2D pole-balancing task, and found that successful individuals have a vary-
ing degree of predictability in their internal dynamics (how the hidden unit activities
change over time). This is discouraging, since if individuals with internal dynamics
with high or low predictability are equally good in behavioral performance, predic-
tive dynamics may not evolve to dominate. However, a slight change in the environ-
ment made individuals with high predictive dynamics to proliferate. The only
change necessary was to make the task a little harder (make the initial random tilt
of the pole to be more). This suggests that predictive internal dynamics has a fitness
value when the environment changes over time, and this happens to be how the
nature is, thus predictive dynamics will become dominant. Not the strongest or