Systems, Man & Cybernetics - July 2015 - 10

a general-purpose expert system. With added domain
knowledge, other expert systems were derived from
Dendral, such as MYCIN and MOLGEN. These systems
achieved success rates comparable with those of human
experts in specific domains. Expert systems that played
checkers (Arthur Samuel in the 1960s) and chess (Chase
and Simon in the 1970s [3]) started defeating top players.
In 1967, Minsky predicted that "within a generation, the
problem of creating artificial intelligence will be substantially solved" [4].
Rules, Yes, But Not Everything
Criticism on the AI dreams started coming as early as
1965, when Dreyfus published his article "Alchemy and
Artificial Intelligence" [5]. Subsequently, his book What
Computers Can't Do: A Critique of Artificial Reason
was published in 1972. By the end of the 1970s, doubts
abounded. Feigenbaum wrote that the "expert himself
doesn't always know exactly what it is he knows about
his domain" and "knowledge threatens to become ten
thousand special cases" [6, p. 6]. In 1982, Minsky admitted, "The AI problem is one of the hardest science has ever
undertaken" [7].
Learning from Examples
In 1943, McCulloch and Pitts published the paper "A Logical Calculus of the Ideas Immanent in Nervous Activity."
This encouraged a new connectionist model that learned
from known patterns. Its development started with the
perceptron in 1958, went through a hibernation phase due
to Minsky's 1969 book Perceptrons, revived after the nonlinear multilayer perceptron learning algorithm by Werbos
and the Parallel Distributed Processing Group appeared
in the 1980s, and, finally, culminated in support vector
machines, which were introduced by Vapnik.
Although the generalization problem was mitigated by
support vector machines, the problem of feature selection
remained. The Neocognitron, proposed by Fukushima
in 1980, could reveal important features from training
images and then classify patterns. With a similar basic
idea, an improved model, the convolution neural network,
was introduced by LeCun in 1998. In addition to image
recognition and video analysis, recently, it was successfully applied to natural language processing and decision
making in games like Go.
Awareness-A Path from Data to Features,
Concepts, and Knowledge
In this article, we consider CA as a bridge connecting raw
data (e.g., image, audio, and so on) to perception and cognition. CA is also a bridge connecting soft computing and
AI. It is possible to derive intelligence step by step, ranging from sensory-level to human-level awareness [8]. In
general, awareness is a state directed to an object, and it
provides a mechanism for detecting some events [9], [10].
In particular, the detected events themselves may not give
10

IEEE Systems, Man, & Cybernetics Magazine July 2015

rise to awareness immediately, but they can provide information and a knowledge base for the awareness of more
complex events. Creating different levels of awareness in
a computing machine will lead to intelligence step by step,
from low levels to higher levels. This is the main approach
proposed by CA.
CA has been studied for more than two decades in the
context of computer-supported cooperative work, ubiquitous computing, and social networks. So far, CA has been
considered by researchers as a process for acquiring and
distributing context information related to what is happening, what has happened, and what is going to happen
in a particular environment. The main purpose of CA is
to provide context information in a timely manner so that
a human (or machine) can take action or make decisions
proactively before something happens.
A direct way to realize awareness in a computing
machine is to model and emulate human awareness.
A human has his or her own goals and views of the
world. He or she senses various facts in the environment, and, based on what he or she senses or feels,
the sensory information is incorporated into his or
her memory to form a knowledge base. In the same
way, CA can be realized by a f low of data sensing,
information processing, and experience accumulation. Note that CA not only emulates human awareness
but also improves human awareness. That is, we can
develop CA by observing human awareness, and, ultimately, the developed CA can sense minute, vague, and
vast amounts of experiences that cannot be sensed by
human awareness due to the physical limitations of the
human body. With CA support, we humans can be more
aware of our environment.
The Purpose of This Study
So far, CA has been typically studied from a high-level,
macroscopic, or social perspective. To derive intelligence
based on awareness, it is necessary to study the microscopic and mesoscopic aspects of awareness. To this
aim, we need to understand the mechanism that converts
low-level awareness to high-level one inside human or
animal brains and to build specific mathematical and
computational models for emulating awareness. This article is a joint work of the four cochairs of the Technical
Committee on Awareness Computing of the IEEE SMCS.
Awareness: Sometimes an Enigma
To realize awareness in a computing machine, it is necessary to review the meaning of awareness and reach some
consensus about the goal of CA research. The use of the
word awareness is broadly categorized into two meanings:
explicit awareness and tacit awareness. "I am aware that
the next Monday is a holiday," "I am aware of the traffic
rules in New Zealand," and "The network server is aware
of the computing platform I am using to write this article"
are examples of explicit awareness, where awareness

Table of Contents for the Digital Edition of Systems, Man & Cybernetics - July 2015