IEEE Computational Intelligence Magazine - February 2018 - 5

AlphaGo that beats the best human Go
player. These are re markable milestones in the history of machine intelligence. Yet, we are quite far from
designing true "intelligent systems" with
human-like reasoning ability. I firmly
believe that CI can help in this direction. In the recent past, deep neural
networks (DNNs) have made a big
news and they have reached an unbeatable performance limit in several applications! AlphaGo also uses DNNs along
with the state-of-the-art tree search to
reach an unbeatable performance level.
Do all these mean that we have realized
an intelligent system with human-like
cognitive and reasoning ability? In my
view, this is certainly not true. One of
the very plausible reasons is that such
systems do not perform brain style computing. Although such systems take
inspiration from the brain, neither the
architecture nor the low level processing
is close to that of the brain. Deep learning is a very promising area, but it
demands more research to understand
why it works so good for some problems! In spite of its superb success with
image classification, there are failures for
very simple cases [2],[3]. For example,
DNNs are found to recognize images
completely "unrecognizable to humans"
as human recognizable objects with
more than 99.99% confidence [2]; on
the other hand, with minor changes
imperceptible to human eyes, DNNs
can be made to misclassify [3]. These
raise several questions: What are the
causes for such successes? What are the
causes for such failures? What kind of
knowledge representation does it make?
Is there too much of redundancy in the
network? As a possible reason for the
success of DNNs, "the flattening of
manifold-shaped data in higher layers of
neural networks "[4] has been suggested
based on empirical studies. This is interesting, but we probably need more and
better understanding. The primary visual
cortex of a human uses sparse coding to
efficiently represent the visual world [5].
Thus, typically a small number of neurons become active to a specific property of the stimulus. Sparse encoding of
stimulus in the brain has led to the

A human being does not need thousands of examples
to learn an object. To design intelligent systems with
human-like reasoning skills, we may need to exploit
attributes of the human brain.
development of parsimonious representation of brain functions in terms of
multiple "Sparse Connectivity Patterns"
(SCPs) [6]. Each SCP encodes a system
level function and relates to a small set
of spatially distributed, functionally synchronous brain regions. The processing
in the brain is distributed. Thus, one of
the important attributes of brain is
believed to be sparse distributed representation (SDR) [7]. In an SDR, at a
given instant of time only a few neurons
will be active. For example, to represent
a "dog" we may use 2000 neurons with
only, say, 200 active neurons, where the
active neurons may represent different
attributes of a dog. Such a representation
is quite robust as even when a few neurons die, we shall still be able to recognize a dog. In an SDR, there may be
overlap between the sets of active neurons representing two different objects,
where the shared neurons may represent
some common attributes of the two
objects [7]. Another remarkable aspect of
human brain is that it can learn an
object or a concept with just a few
examples, it does not require hundreds
of instances. To design intelligent systems, we need to equip them with some
of these attributes of the human brain.
For example, we need to design neural
networks, which use SDR for knowledge representation. Next we discuss
another important aspect of computational intelligence.
We, the human beings, deal with
imprecise, uncertain and vague information in a routine manner to make
rational choices. Zadeh emphasized on
two remarkable human abilities: our
capability to converse, communicate
and reason with incomplete, imprecise
and uncertain information; and our
abilities to perform variety of tasks
(mental and physical) without any
numerical measurement or traditional

computation [8]. Human beings are
capable of computing with imprecise
information as well as with words, i.e.,
with non-numeric information. Naturally, an intelligent machine/system
should have these abilities. An intelligent system should be able to process
and reason with words as well as with
imprecise information. Here comes
another facet of CI, fuzzy sets, that
could be very useful. According to
Zadeh [8], "Computing with Words is a
methodology for reasoning, computing
and decision-making with information
described in natural language." Here the
objects of computations are not numbers but words. This is a definition with
a wide scope (I must mention, that
there are other views of CWW, for a
nice discussion, please see [8]). CWW
has progressed a lot, but not to the level
that one would be happy with. In my
view, CWW has not reached the level
of sophistication that is required to
develop "smart systems" with humanlike cognitive and reasoning abilities, at
least, to a limited extent. So far, CWW
has not got the level of attention that it
deserves. We need more research efforts
in this area. In fact, it is worth looking
at non-fuzzy modelling for CWW
because it may not be possible to use
fuzzy sets (Type 1 or Type 2) to represent every word needed for reasoning.
Moreover, a significant amount of work
is going on using tools like Long and
Short Term Memory (LSTM) and
DNNs for understanding text and
inferring from textual input. In my
view, this is also CWW, but without
using fuzzy sets. At this point, I want to
emphasize that it is not my intention to
claim or even suggest that neural networks and fuzzy logic are the necessary
and sufficient tools to reach goals of
designing intelligent systems with
human-like abilities. There are many

FEbruary 2018 | IEEE ComputatIonal IntEllIgEnCE magazInE

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2018