Page 217 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

208 CHAPTER 10 Computers Versus Brains: Game Is Over or More to Come?

2. AI APPROACHES
Dominated by the proliferation of digital computers, symbolic approaches have
been prominent in AI for the decades 1960s to 1980s [10,11]. As the computational
power of computers expanded, the sophistication of symbol manipulations has
increased tremendously. A leading AI paradigm explicitly declared that advanced
manipulation of a physical symbol system has all the sufficient and necessary
means of intelligent behavior in man and machine [12,13].Accordingtothis
view, external events and perceptions are transformed into inner symbols to repre-
sent the state of the world. There is a set of rules describing possible deductions
and actions through logical inferences using these rules. The rules are part of an
inference engine; together with a knowledge base and corresponding data they
constitute an expert system. The knowledge base sometimes can be huge to ac-
count for the intricate specifics of the problem domain. In expert systems, problem
solving takes the form of a search through the space of symbols using inferences
[14,15].
Symbolic AI approaches produced many successful implementations. However,
symbolic models are often inflexible, rigid, and unable to accommodate too many
changes inevitable in real-life situations. Dreyfus’ situated intelligence approach
is a prominent example of a philosophical alternative to symbolism [16]. Dreyfus
ascertains that intelligence is defined in the context of the environment, and a fixed
symbol system cannot grasp the essence of intelligence [16,17]. Situated intelligence
finds its successful applications in embodied intelligence and robotics; see for
example, Ref. [18]. Deep learning and neural networks implement an alternative to
symbolic approaches. They employ massive data to build AI classifiers without the
need for preset symbolic rule base. Neural networks can be viewed as bottom-up
methods starting from data and moving toward abstractions at higher levels,
while symbolic AI implements top-down processing by employing a broad existing
knowledge base to make inferences applicable to the specific problem at hand. Neural
networks emphasize parallel-distributed processing over adaptive structures and they
are robust and flexible. Knowledge-based neural networks are devoted to combine the
advantages of symbolic and neural approaches [19].
Artificial neural networks appeared first in the 1940s in McCulloch and Pitts’
pioneering work [21] and in Wiener’s cybernetics [22]. Mathematical guarantees of
their functional approximation properties were provided by Kolmogorov and Arnold
[23,24], with many developments in following decades. Werbos’ backpropagation
presented the first constructive algorithm to build such approximation [25], but it
was not until the late 1980s that neural networks became mainstream through parallel
and distributed processing [26e28]. There is no way to properly reference the massive
work on neural networks in the past decades; we just mention LeCun’s LeNet convo-
lutional networks [29] and long short-term memories [30] as two examples of the
fundamental work that lead to the deep learning revolution of AI after the turn of
the century [31]. Fig. 10.2 illustrates the relationship between symbolic AI and
nonsymbolic neural networks approaches. As we introduce next, brains teach us