IEEE Signal Processing Magazine - January 2018 - 173

PERSPECTIVES

(continued from page 180)

and the vertical (i.e., individual AI application areas) perspectives.

The first rise of AI
In the first rising wave of AI, starting in
the 1960s and based on expert knowledge
engineering, domain experts devised computer programs according to the knowledge about the (very narrow) application
domains they have [3], [4]. The experts
designed these programs using symbolic
logical rules grounded on careful representation and engineering of such knowledge.
These knowledge-based AI systems tend
to be effective in solving narrow-domain
problems by examining the "head" or most
important parameters and reaching a solution about the appropriate action to take
in each specific situation. These "head"
parameters are identified in advance by
human experts, leaving the "tail" parameters and cases untouched. Since they lack
learning capability, they have difficulty in
generalizing the solutions to new situations
and domains.
The typical approach during this wave
is exemplified by the expert system, a
computer system that emulates the decision-making ability of a human expert.
Such systems are designed to solve complex problems by reasoning about knowledge [3]. The first expert systems were
created in the 1970s and then proliferated
in the 1980s. The main "algorithm"
used was the inference rules in the form of
"if-then-else" [5].
The main strength of these first-generation AI systems is its transparency
and interpretability in their (limited) capability in performing logical reasoning.
They use handcrafted expert knowledge
that is often effective in narrowly defined problems, although the reasoning
cannot handle uncertainty that is ubiquitous in practical applications. Due to this
strength, the first-generation AI systems
are still in use today. Examples are narrow-domain dialogue systems and chatbots, chess-playing programs, traffic
light controllers, optimization software
for logistics of good deliveries, etc.
The early research and system design
of speech recognition, a long-standing
AI challenge in machine perception,

were based on the AI paradigm of expert
knowledge engineering during the first
rising wave of AI. During the 1970s and
early 1980s, the expert-system approach
to speech recognition was quite popular,
driven largely by Carnegie Mellon University and Massachusetts Institute of
Technology speech researchers; e.g., the
spectrogram reading method as elaborated in [6]. However, the lack of general
abilities to learn algorithmically from
data and to handle uncertainty in reasoning in the knowledge-based approach was
acutely recognized by researchers, along
with the second rise of AI in the 1980s.
The author has been part of the speech
research community since the late 1980s
and contributed to the transition from
knowledge-based speech recognition to
a data-driven one powered by statistical
machine-learning methods [7]-[13].

The second rise of AI
The second rising wave of AI for speech
arrived in the 1980s (and somewhat later
for other AI areas) after clear evidence that
learning and perception capabilities are
crucial for complex AI systems but missing in knowledge-based expert systems.
This is not just for speech recognition but
also for vision and other AI systems. For
example, when the Defense Advanced
Research Projects Agency opened its first
Grand Challenge for autonomous driving,
most vehicles then relied on the knowledge-based paradigm. Much like speech
recognition, the autonomous driving and
vision researchers quickly realized the
limitation of the first-generation AI paradigm due to the need for automatic learning equipped with uncertainty handling
and generalization capabilities.
This second-generation AI paradigm
was based on machine learning, which
we now call shallow due to the lack of
abstractions constructed by many-layer
or "deep" representations of data which
would come in the third rise of AI. In such
shallow machine learning, engineers do
not need to be concerned with constructing precise and exact rules as required for
the first-generation AI systems. Rather,
they focus on statistical models [14],
[15] or simple neural networks [16] as an
IEEE SIGNAL PROCESSING MAGAZINE

January 2018

underlying engine and then automatically
learn or "tune" the parameters of the engine using the training data to make them
handle uncertainty and generalize well
from one condition to another and from
one domain to another. The key algorithms and methods for machine learning include expectation-maximization
(EM), Bayesian networks, support vector
machines, decision trees, and, for neural
networks, the backpropagation algorithm.
Generally, the machine-learning-based
AI systems perform much better than the
earlier, knowledge-based counterparts.
Successful examples include almost all
AI tasks in machine perception-speech
recognition [10], [17]-[20], face recognition [21], visual object recognition [22],
handwriting recognition [23], and machine
translation [24].
In speech recognition, for more than
20 years from the 1980s to 2010, the
paradigm was completely switched to
and dominated by the (shallow) machinelearning paradigm using a statistical generative model called the hidden Markov
model (HMM) integrated with Gaussian
mixture models, along with various versions of its generalization [25]-[28]. The
main algorithms and methods include
the Viterbi algorithm, Baum-Welch algorithm (which is a special case of EM
when applied to the HMM), and extended
Baum-Welch (which includes learning
algorithms for maximizing mutual information, minimizing classification errors,
and minimizing phone errors) [29].
Among many versions of the generalized HMMs were statistical and neuralnet-based hidden dynamic models (see
[30]-[32] and [66]). The former adopted
EM and extended Kalman filter algorithms for learning model parameters
[33], [34] and the latter used backpropagation [35]. Both of them made extensive
use of multiple latent layers of representations for the generative process of speech
waveforms following the long-standing
framework of analysis by synthesis in
human speech perception. More significantly, inverting this deep generative
process to its counterpart of an end-toend discriminative process gave rise to
the very first industrial success of deep
173

Table of Contents for the Digital Edition of IEEE Signal Processing Magazine - January 2018