Page 294 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

4. Question Versus Answer 287

the fastest species survive: the most adaptable species survive, where prediction
seems to be the key, and this lays down the necessary condition for consciousness.
In the second study [8], we analyzed publicly available brain EEG data collected
during awake, rapid eye movement (REM) sleep, and slow-wave sleep. Since awake
and vivid dreaming (REM sleep) are associated with consciousness and deep sleep
(slow-wave sleep) is associated with unconsciousness, we measured the predictabil-
ity in these EEG signal wave forms. We preprocessed the raw EEG signal, computed
the interpeak interval (IPI), the time distance between peaks in the EEG signal, and
measured how easy it is to predict the next IPI based on previous IPI data points. We
found that awake and REM EEG signals have higher IPI predictability than that of
slow-wave sleep, suggesting that IPI predictability and consciousness may be
correlated.
In this section, I discussed how synaptic plasticity mechanisms can be directly
linked to prediction, how delay in the nervous system may have led to predictive ca-
pabilities, and how predictive dynamics can serve as a precursor of consciousness. In
sum, prediction is a key function of the brain, and it should also be included as such
in artificial systems.

4. QUESTION VERSUS ANSWER
In both brain science and artificial intelligence, the general focus is to understand
how the brain solves problems relating to perceptual, cognitive, and motor tasks,
or how to make artificial intelligence algorithms solve problems in vision, natural
language processing, game playing, robot control, etc. That is, we are focused on
mechanisms that produce answers, and less on mechanisms that pose the questions.
Of course we know the importance of asking the right questions, and any researcher
is well aware of the importance of picking the right research question. Often times,
research involves finding new ways to conceive of the problem, rather than finding
new ways of solving problems as conceived [9], and this is especially essential when
the conceived problem itself is ill-formed so as to be unsolvable (e.g., how can we
prove Euclid’s fifth postulate [unsolvable] vs. can we prove Euclid’s fifth postulate
[solvable]).
In 2012, Mann and I discussed in Ref. [10] the need to start paying attention to
problem-posing, as opposed to problem-solving. It turns out that problem-posing has
been an active topic in the education literature (see Ref. [11] and many subsequent
publications). So, learning and problem-posing seem to be intricately related. How-
ever, this angle is not explored much in artificial intelligence, except for rare excep-
tions, and I strongly believe integrating learning and problem-posing can lead to a
much more robust and more general artificial intelligence. Some of those rare excep-
tions is Schmidhuber’s study, which explicitly addresses this issue. In his Powerplay
algorithm, both problems and solvers are parameterized and the algorithm seeks
specific problems that are solvable with the current capability of the solver, and
loop through this to train an increasingly general problem solver [12]. More recently,