Machine Learning in Artificial Intelligence: Towards a Common Understanding

Conference Paper (PDF Available) · January 2019with 696 Reads
Conference: Hawaii International Conference on System Sciences (HICSS-52), At Grand Wailea, Maui, Hawaii
Cite this publication
Abstract
The application of “machine learning” and “artificial intelligence” has become popular within the last decade. Both terms are frequently used in science and media, sometimes interchangeably, sometimes with different meanings. In this work, we aim to clarify the relationship between these terms and, in particular, to specify the contribution of machine learning to artificial intelligence. We review relevant literature and present a conceptual framework which clarifies the role of machine learning to build (artificial) intelligent agents. Hence, we seek to provide more terminological clarity and a starting point for (inter- disciplinary) discussions and future research.
Machine Learning in Artificial Intelligence:
Towards a Common Understanding
Niklas Kühl
Karlsruhe Institute of
Technology
kuehl@kit.edu
Marc Goutier
Karlsruhe Institute of
Technology
marc.goutier@kit.edu
Robin Hirt
Karlsruhe Institute of
Technology
hirt@kit.edu
Gerhard Satzger
Karlsruhe Institute of
Technology
gerhard.satzger@kit.edu
Abstract
The application of “machine learning” and “arti-
ficial intelligence” has become popular within the last
decade. Both terms are frequently used in science and
media, sometimes interchangeably, sometimes with
different meanings. In this work, we aim to clarify the
relationship between these terms and, in particular, to
specify the contribution of machine learning to
artificial intelligence. We review relevant literature
and present a conceptual framework which clarifies
the role of machine learning to build (artificial)
intelligent agents. Hence, we seek to provide more
terminological clarity and a starting point for (inter-
disciplinary) discussions and future research.
1. Introduction
In his US senate hearing in April 2018, Mark
Zuckerberg stressed the necessary capabilities of
Facebook’s “AI tools (…) to (…) identify hate speech
(…)” or (…) terrorist propaganda” [1]. Researchers
would typically describe such tasks of identifying
specific instances within social media platforms as
classification tasks within the field of (supervised)
machine learning [2][4]. However, with rising
popularity of artificial intelligence (AI) [5], the term
AI is often used interchangeably with machine
learning–not only by Facebook’s CEO in the example
above or in other interviews [6], but also across
various theoretical and application-oriented contribu-
tions in recent literature [7][9]. Carner (2017) even
states that he still uses AI as a synonym for machine
learning although knowing this is not correct [10].
Such ambiguity, though, may lead to multiple
imprecisions both in research and practice when
conversing about methods, concepts, and results.
It seems surprising that despite of the frequent use
of the terms, there is hardly any helpful scientific
delineation. Thus, this paper aims to shed light on the
relation of the two terms machine learning and
artificial intelligence. We elaborate on the role of
machine learning within instantiations of artificial
intelligence, precisely within intelligent agents. To do
so, we take a machine learning perspective on the
capabilities of intelligent agents as well as the
corresponding implementation.
The contribution of our paper is threefold. First, we
expand the theoretical framework of Russel & Norvig
(2015) [11] by further detailing the “thinking” layer of
any intelligent agent by splitting it into separate
“learning” and “executing” sublayers. Second, we
show how this differentiation enables us to distinguish
different contributions of machine learning for intelli-
gent agents. Third, we draw on the implementations of
the execution and learning sublayers (“backend”) to
define a continuum between human involvement and
agent autonomy.
In the remainder of this paper, we first review
relevant literature in the fields of machine learning and
artificial intelligence. Next, we present and elaborate
our conceptual framework which highlights the con-
tribution of machine learning to artificial intelligence.
On that basis, we derive an agenda for future research
and conclude with a summary, current limitations, as
well as an outlook.
2. Related work
As a base for our conceptual work, we first review
the different notions, concepts, or definitions of
machine learning and artificial intelligence within
extant research. In addition, we elaborate in greater
detail on the theories which we draw upon in our
framework.
2.1. Terminology
Machine learning and artificial intelligence, as
well as the terms data mining, deep learning and
statistical learning are related, often present in the
same context and sometimes used interchangeably.
While the terms are common in different communities,
their particular usage and meaning varies widely.
Figure 1. General terminology used in this paper
For instance, in the field of statistics the focus is on
statistical learning, which is defined as a set of me-
thods and algorithms to gain knowledge, predict
outcomes, and make decisions by constructing models
from a data set [12]. From a statistics point of view,
machine learning can be regarded as an implemen-
tation of statistical learning [13].
Within the field of computer science, machine
learning has the focus of designing efficient
algorithms to solve problems with computational
resources [14]. While machine learning utilizes
approaches from statistics, it also includes methods
which are not entirely based on previous work of
statisticiansresulting in new and well-cited contri-
butions to the field [15], [16]. Especially the method
of deep learning raised increased interest within the
past years [17]. Deep learning models are composed
of multiple processing layers which are capable of
learning representations of data with multiple levels of
abstraction. Deep learning has drastically improved
the capabilities of machine learning, e.g. in speech
[18] or image recognition [19].
In demarcation to the previous terms, data mining
describes the process on how to apply quantitative
analytical methods, which help to solve real-world
problems, e.g. in business settings [20]. In the case of
machine learning, data mining is the process of
generating meaningful machine learning models. The
goal is not to develop further knowledge about
machine learning algorithms, but to apply them to data
in order to gain insights. Machine learning can
therefore be seen as a foundation for data mining [21].
In contrast, artificial intelligence applies
techniques like machine learning, statistical learning
or other techniques like descriptive statistics to mimic
intelligence in machines.
Figure 1 and the terms defined within this
paragraph lay the foundation of the remainder of this
work. However, the overall terminology and
relationships of the concepts is discussed
controversially [22]. Therefore, the focus of this paper
is to bring more insight to the terminology and more
precisely, to clarify the role of machine learning within
AI. To gain a broader understanding for the terms
machine learning and AI, we examine both in further
detail.
2.2. Machine learning
Machine learning describes a set of techniques that
are commonly used to solve a variety of real-world
problems with the help of computer systems which can
learn to solve a problem instead of being explicitly
programmed [23]. In general, we can differentiate
between unsupervised and supervised machine
learning. For the course of this work, we focus on the
latter, as the most-widely used methods are of
supervised nature [24]. With regard to supervised
machine learning, learning means that a series of
examples (“past experience”) is used to build
knowledge about a given task [25]. Although
statistical methods are used during the learning
process, a manual adjustment or programming of rules
or strategies to solve a problem is not required. In more
detail, (supervised) machine learning techniques
always aim to build a model by applying an algorithm
on a set of known data points to gain insight on an
unknown set of data [11], [26].
Statistical Learning
[Origin: Statistics]
Machine Learning
[Origin: Computer Science] Artifical Intelligence
applies
Others
Data Mining
Process Method set Instantiation
describes
application
process of
Others
Others
(e.g. Descriptive Statistics)
Implementation
Deep Learning
Thus, the processes of creationof a machine
learning model slightly vary in their definition of
phases but typically employ the three main phases of
model initiation, performance estimation and
deployment [27]: During the model initiation phase, a
human user defines a problem, prepares and processes
a data set and chooses a suitable machine learning
algorithm for the given task. Then, during the
performance estimation, various parameter
permutations describing the algorithm are validated
and a well-performing configuration is selected with
respect to its performance in solving a specific task.
Lastly, the model is deployed and put into practice to
solve the task on unseen data.
Learning in general depicts a key facet of a
human’s cognition which “refers to all processes by
which the sensory input is transformed, reduced,
elaborated, stored, recovered, and used” [28, p. 4].
Humans process a vast amount of information by
utilizing abstract knowledge that helps us to better
understand incoming input. Due to their adaptive
nature, machine learning models are able to mimic the
cognitive abilities of a human being in an isolated
manner.
However, machine learning solely represents a set
of methods that enable to learn patterns in existing
data, thus generating analytical models that can be
utilized inside larger IT artifacts.
2.3. Artificial intelligence
The topic of artificial intelligence (AI) is rooted in
different research disciplines, such as computer
science [18, 19], philosophy [20, 21], or futures
studies [22, 23]. In this work, we mainly focus on the
field of computer science, as it is the most relevant one
in identifying the contribution of machine learning to
AI and in differentiating both terms.
AI research can be separated into different research
streams [11]. These streams differ on the one hand as
to the objective of AI application (thinking vs. acting),
on the other hand as to the kind of decision making
(targeting a human-like decision vs. an ideal, rational
decision). This distinction leads to four research
currents which are depicted in Table 1.
According to the “Cognitive Modeling” (i.e. thinking
humanly) stream, an AI must be a machine with a
mind [34]. This also includes performing human
thinking [35], not only based on the same output as a
human when given the same input, but also on the
same reasoning steps which led to the very conclusion
[36].
1
In this case, the terms rational and intelligent are
used interchangeably in related work [11],[23]
The “Laws of Thought stream (i.e. thinking
rationally) requires an AI to arrive at the rational
decision despite what a human might answer.
Therefore, an AI must follow the laws of thought by
using computational models [37] which reflect logic.
The “Turing Test” (i.e. acting humanly) stream
implies that an AI must act intelligently when
interacting with humans. To accomplish these tasks,
an AI must perform human tasks at least as good as
humans [38]. These requirements can be tested by the
Turing Test [39].
Finally, the “Rational Agent” stream considers an
AI as a rational [11] or intelligent [40] agent
1
. This
agent does not only act autonomously but also with the
objective to achieve the rationally ideal outcome.
An alternative way to delineate AI is defining
intelligence in general and using the resulting insights
to create intelligent machines. Legg and Hutter [41]
use intelligence tests, theories of human intelligence
and psychological definitions to define a measurement
of intelligence. Based on their definition, they use an
agent-environment framework to describe intelligence
in general andin case the agent is a machine
artificial intelligence in particular. Their framework
exhibits many similarities to the “acting rationally”
stream.
Besides defining AI in general, the classification of
AI is another topic in the field of AI research. Searle
[42] suggests differentiating between weak and strong
AI. Whereas a weak AI only pretends to think, a strong
AI is a mind with mental states. Gubrud [43] however
categorizes AI by taking the type of task into account.
An artificial general intelligence (AGI) is an AI which
in general, i.e. in any domain, acts at least on the same
level as a human brain, however without requiring
Rationally
Thinking
“Laws of
thought “
Acting
Rational Agent
Application
to
Objective
Table 1. AI research streams based on
Russell & Norvig [11]
consciousness. In contrast, a narrow AI is an AI that
rivals or exceeds the human brain only in specific,
limited tasks [44].
In the following, we will look into the Rational
Agent stream in some more detail as it is of
importance when regarding implementation of
machine learning within AI. We will come back to the
other three research streams in section 3 where we
show that they are compatible with our framework of
an agent-based AI.
According to the “Rational Agent stream, the
intelligence itself is manifested by the acting of agents.
These agents are characterized by five features,
namely theyoperate autonomously, perceive their
environment, persist over a prolonged time period,
adapt to change, and create and pursue goals” [11, p.
4]. An agent defines its action not for itself but with an
environment it interacts with. It recognizes the
environment by its sensors, has an agent program to
decide what to do with the input data, and performs an
action with its actuators. To become a rational agent,
the agent must also act to achieve the highest expected
outcome according to this performance measure
based on the current and past knowledge of the
environment and the possible actions.
When it comes to the general demarcation of
agents, according to Russel & Norvig, the agent
program can be segmented into four different agent
types [11]: A simple reflex agent reacts only based on
its sensor data whereas a model-based reflex agent also
considers an internal state of the agent. A goal-based
agent decides for the best decision to achieve its goals.
The fulfilment of a goal is a binary decision which
means it can either be fulfilled or not. On the contrast,
a utility-based agent has no binary goal but a whole
utility function which it tries to maximize. An agent
can become a learning agent by extending its program.
Such a learning agent then consists of a performance
element which selects an action based on the sensor
data and a learning element, which gets feedback from
the environment, generates own problems, and
improves the performance element if possible.
The agent-environment framework consists of
three components: an agent, an environment and a
goal. Intelligence is the measurement for the "agent’s
ability to achieve goals in a wide range of
environments” [41, p. 12]. The agent gets input by
perceptions generated from the environment. One type
of perceptions are observations of the environment,
while others are reward signals that indicate how well
the goals of the agent are achieved. Based on these
input signals, the agent decides to perform actions
which are sent back as signals to the environment.
3. A framework for understanding the role
of machine learning in artificial
intelligence
In order to understand the interplay of machine
learning and AI, we base our concept on the
framework of Russel & Norvig [11]. With their
differentiation between the two objectives of AI
application, acting and thinking, they lay an important
foundation.
3.1. Layers of agents
When trying to understand the role of machine
learning within AI, we need to take a perspective
which has a focus on the implementation of intelligent
agents. We require this perspective, as it allows us to
map the different tasks and components of machine
learning to the capabilities of intelligent agents. If we
regard the capabilities of thinking and acting of an
intelligent agent and translate this into the terms of
software design, we can reason that the acting
capabilities can be regarded as a frontend, while the
thinking part can be regarded as a backend. Software
engineers typically strictly separate form and function
to allow for more flexibility and independence as well
as to enable parallel development [45]. The frontend is
the interface the environment interacts with. It can take
many forms. In the case of intelligent agents it can be
a very abstract, machine-readable web interface [46],
a human-readable application [47] or even a humanoid
template with elaborated expression capabilities [48].
For the frontend to interact with the environment, it
requires two technical components; sensors and
actuators. Sensors detect events or changes in the
environment and forward the information via the
frontend to the backend. For instance, they can read
the temperature within an industrial production
machine [49] or read visuals of an interaction with a
human [50]. Actuators on the other hand are
components that are responsible for moving and
controlling a mechanism. While sensors just process
information, actuators act, for instance by
automatically buying stocks [51] or changing the
facial expressions of a humanoid [52]. One could
argue that the Turing test [39] takes place at the
interaction of the environment with the frontend, more
precisely the combination of sensors and actuators if
one wants to test the agent’s AI of acting humanly.
Despite every frontend having sensors and actuators,
it is not of importance for our work what the precise
frontend looks like; it is only relevant to note that a
backend-independent, encapsulated frontend exists.
Figure 2. Conceptual framework
The backend provides the necessary
functionalities, which depict the thinking capabilities
of an intelligent agent. Therefore, the agent needs to
learn and apply learned knowledge.
In consequence, machine learning is relevant in
this implementation layer. When regarding the case of
supervised machine learning, we need to further
differentiate between the process task that is building
(=training) adequate machine learning models [21]
and the process task that is executing the deployed
models [53]. Therefore, to further understand the role
of machine learning within intelligent agents, we
refine the thinking layer of agents into a learning
sublayer (model building) as well as an executing
sublayer (model execution)
2
. Hence, we regard the
necessary implementation for the learning sublayer as
the learning backend, while the executing sublayer is
denoted by the executing backend.
3.2. Types of learning
The learning backend dictates first if the intelligent
agent is able to learn, and, second, how the agent is
able to learn, e.g., which precise algorithms it uses,
what type of data processing is applied, how concept
2
Russel & Norvig indicate a related relationship
by differentiating into learning elements and
performance elements [11].
drift [54] is handled, etc. Therefore, we pick up on the
terminology from Russel & Norvig [11] by regarding
two different types of intelligent agents: simple-reflex
agents as well as learning agents. This differentiation
especially holds for a machine learning perspective on
AI, as it considers whether the underlying models in
the thinking layer are once trained and never touched
again (simple-reflex)or continuously updated and
adaptive (learning). In recent literature, suitable
examples for both can be found. As an example for
simple-reflex agents, Oroszi and Ruhland build and
deploy an early warning system of pneumonia in
hospitals [55]: While building and testing the model
for the agent shows convincing results, the adaptive
learning of the system after deployment might be
critical. Other examples of agents with single-trained
models are common in different areas, for instance for
anaphora resolutions [56], prediction of pedestrians
[57] or object annotation [58]. On the other hand,
recent literature also gives examples for learning
agents. Mitchell et al. present the concept of “never-
ending learning” agents [59] which have a strong
focus on continuously building and updating models
within agents. An example for such an agent is shown
by Liebman et al., who build a self-learning agent for
music playlist recommendations [60]. Other cases are
for instance the regulation of heat pump thermostats
[61], an agent to acquire collective knowledge over
different tasks [62] or learning word meanings [63].
The choice on this feature in general (simple-reflex
vs. learning agent) influences the overall design of the
agent as well as the contribution of machine learning.
The overview of our resulting framework is depicted
in figure 2. In conclusion, in the case of a simple-reflex
agent, machine learning takes places as a once-trained
model in the execution sublayer. In contrast, it plays a
role in the learning sublayer of a learning agent to
continuously improve the model in the execution
sublayer. This improvement is based on knowledge
and feedback, which is derived from the environment
via the execution layer.
3.3. Continuum between human involvement
and machine involvement
When it comes to the executing backend and the
learning backend, it is not only of importance if and
how underlying machine learning models are
updatedbut how much automated the necessary
processes are. Every machine learning task involves
various process steps, including data source selection,
data collection, preprocessing, model building,
evaluating, deploying, executing and improving (e.g.
[21], [53], [64]). While a discussion of the individual
steps is beyond the scope of this paper, the autonomy
and the automation of these tasks as an
implementation within the agent is of particular
interest in each necessary task of the machine learning
lifecycle [27].
For instance, while the execution of a once-built
model can be fairly easily automated, the automated
identification of an adequate data source for a new
problem or retraining as well as a self-induced model
building are more difficult. Therefore, we need to view
the human involvement in the necessary machine
learning tasks of an intelligent agent, as depicted in
figure 3. While it is hard to draw a clear line between
all possible forms of human involvement in the
machine learning-relevant tasks of an intelligent agent,
we see this phenomenon rather as a continuum. The
continuum ranges between none or little agent
autonomy with full human involvement (e.g. [65]
[67]) on the one extreme as well as the full agent
autonomy and no or little human involvement for the
delivered task on the other (e.g. [68][70]). For
example, an intelligent agent with the task to
autonomously drive a car considering the traffic signs
already proves a high degree of agent autonomy.
However, if the agent is confronted with a new traffic
sign, the learning of this new circumstance might still
need human involvement as the agent might not be
able to “completely learn by itself” [71]. Therefore,
the necessary involvement of humans, especially in
the thinking layer (= executing backend and learning
backend), is of major interest when describing AI and
the underlying machine learning models. The degree
of autonomy for each step of machine learning can be
investigated and may help to characterize the
autonomy of an agent in terms of the related machine
learning tasks.
4. Research priorities for machine-
learning-enabled artificial intelligence
The presented framework of machine learning and
its role within intelligent agents is still on a conceptual
level. However, given the misunderstandings and
ambiguity of the two terms [69], we see potential for
further research with the aim both to clarify the
terminology and to map uncharted territory for
machine-learning enabled artificial intelligence.
First, empiric validation as well as continuous,
iterative development of the framework is necessary.
We need to identify various cases of intelligent agents
across different disciplines and to evaluate how well
the framework fits. It would be interesting to see how
practical and academic machine-learning-enabled
artificial intelligence projects map to the framework,
and, furthermore even quantify which share of such
projects works with learning agents and which with
non-learning agents. Additionally, such cases would
help us to gain a better understanding of the necessary
human involvement in state-of-the art intelligent
agentsand, therefore, determine the “degree” of
autonomy when regarding all aspects (acting,
executing, learning) of such agents.
Second, one aspect of interest would be to reduce
the necessary involvement of humans. As stated
before, we see this spectrum as a continuum between
human involvement and agent autonomy. Two
possibilities come immediately to mind. The methods
of transfer machine learning deal with possibilities on
how to transfer knowledge (i.e., models) from one
source environment to a target environment [72]. This
could indeed help to minimize human involvement, as
further research in this field could show possibilities
and application-oriented techniques to utilize transfer
Figure 3. Degree of agent autonomy and
human involvement
machine learning for automated adaption of novel or
modified tasks [73].
Additionally, regarding already deployed models
as part of the backend-layer, it is of interest not only
how the models are built initially, but how to deal with
changes in the environment. The so-called subfield of
concept drift holds many possibilities on how to detect
changes and adapt modelshowever, fields of
successful application remain rare [54], [74].
5. Conclusion
In this paper, we clarify the role of machine
learning within artificial intelligencein particular
intelligent agents. We present a framework, which
highlights the two cases of simple-reflex and learning
agents as well as the role machine learning can play in
each of them. In a nutshell, machine learning models
can be implemented as once-trained models within an
intelligent agentwithout the possibility to learn
additional insights from the environment (simple
reflex agent). Implementation-wise, we call this
sublayer of executing knowledge the executing
backend. In this case, the agent is able to utilize
(previously built) machine learning modelsbut not
build and update its own ones. If the agent, however,
is able to learn from its environment and is, therefore,
able to update the machine learning models within the
execution sublayer, it is a learning agent. Learning
agents have an additional sublayer, the learning
backend, which allows them to utilize machine
learning in terms of model building/training.
When it comes to the implementation of these two
sublayers, it is of importance to capture the degree of
autonomy that the machine learning within the agent
requires. This aspect focusses on the human
involvement in the necessary machine learning tasks,
e.g. the data collection or the choice of an algorithm.
The research at hand is still in a conceptual state
and has certain limitations. First, while the proposed
framework allows to deepen the understanding of
machine learning within AI, empirical studies are still
required to see how well existing machine-learning-
enabled AI applications fit into this scheme. Expert
interviews with AI designers could validate the model
and complete and evaluate the level of detail.
Furthermore, we need to find ways to quantify the
human involvement in machine-learning related tasks
within AI to gain better understanding of the degree of
autonomy of state-of-the-art agents.
Although at an early stage, our framework should
allow scientists and practitioners to be more precise
when referring to machine learning and AI. It
highlights the importance of not using the terms
interchangeably but making clear which role machine
learning plays within a specific agent implementation.
References
[1] “The Washington Post,” “Transcript of Mark
Zuckerberg’s Senate hearing,” 2018. [Online].
Available:
https://www.washingtonpost.com/news/the-
switch/wp/2018/04/10/transcript-of-mark-
zuckerbergs-senate-
hearing/?utm_term=.4720e7f10b41. [Accessed:
15-Jun-2018].
[2] Z. Waseem and D. Hovy, “Hateful Symbols or
Hateful People? Predictive Features for Hate
Speech Detection on Twitter,” in Proceedings of
the NAACL Student Research Workshop, 2016, pp.
8893.
[3] W. Warner and J. Hirschberg, “Detecting hate
speech on the world wide web,” Proceeding LSM
’12 Proc. Second Work. Lang. Soc. Media, no. Lsm,
pp. 1926, 2012.
[4] H. Chen, W. Chung, J. Qin, E. Reid, M. Sageman,
and G. Weimann, “Uncovering the dark web a case
study of jihad on the web,” Int. Rev. Res. Open
Distance Learn., vol. 14, no. 4, pp. 90103, 2013.
[5] H. Fujii and S. Managi, “Trends and priority shifts
in artificial intelligence technology invention: A
global patent analysis,” Econ. Anal. Policy, vol. 58,
pp. 6069, 2018.
[6] “University of Wisconsin,” “Mark Zuckerberg :
How to Build the Future,” Interview Transcript,
2016. .
[7] Information Commissioner’s Office, “Big Data,
artificial intelligence, machine learning and data
protection,” Data Protection Act and General Data
Protection Regulation, 2017. [Online]. Available:
https://ico.org.uk/media/for-
organisations/documents/2013559/big-data-ai-ml-
and-data-protection.pdf. [Accessed: 15-Jun-2018].
[8] J. A. Brink, “Big Data Management, Access, and
Protection,” Journal of the American College of
Radiology, vol. 14, no. 5, pp. 579580, 2017.
[9] T. Nawrocki, P. D. Maldjian, S. E. Slasky, and S.
G. Contractor, “Artificial Intelligence and
Radiology: Have Rumors of the Radiologist’s
Demise Been Greatly Exaggerated?,” Academic
Radiology, 2018.
[10] C. F. Camerer, “Artificial intelligence and
behavioral economics,” in Economics of Artificial
Intelligence, University of Chicago Press, 2017.
[11] S. J. Russell and P. Norvig, Artificial Intelligence:
A Modern Approach, 3rd ed. 2015.
[12] T. Hastie, R. Tibshirani, J. Friedman, and J.
Franklin, “The elements of statistical learning: data
mining, inference and prediction,Math. Intell.,
vol. 27, no. 2, pp. 8385, 2005.
[13] O. Bousquet, U. von Luxburg, and G. Rätsch,
Advanced Lectures on Machine Learning: ML
Summer Schools 2003, Canberra, Australia,
February 2-14, 2003, Tübingen, Germany, August
4-16, 2003, Revised Lectures, vol. 3176. Springer,
2011.
[14] M. Mohri, A. Rostamizadeh, and A. Talwalkar,
Foundations of machine learning. MIT press, 2012.
[15] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme
learning machine: a new learning scheme of
feedforward neural networks,” in Neural Networks,
2004. Proceedings. 2004 IEEE International Joint
Conference on, 2004, vol. 2, pp. 985990.
[16] F. Sebastiani, “Machine learning in automated text
categorization,” ACM Comput. Surv., vol. 34, no. 1,
pp. 147, 2002.
[17] Y. A. LeCun, Y. Bengio, and G. E. Hinton, “Deep
learning,” Nature, 2015.
[18] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.
Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P.
Nguyen, T. N. Sainath, and B. Kingsbury, “Deep
Neural Networks for Acoustic Modeling in Speech
Recognition,” IEEE Signal Process. Mag., 2012.
[19] K. He, X. Zhang, S. Ren, and J. Sun, “Deep
Residual Learning for Image Recognition,” in 2016
IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2016.
[20] C. Schommer, “An Unified Definition of Data
Mining,” CoRR, vol. abs/0809.2696, 2008.
[21] I. H. Witten, E. Frank, and M. a. Hall, Data Mining:
Practical Machine Learning Tools and Techniques,
Third Edition, vol. 54, no. 2. 2011.
[22] Cross Validated, “What is the difference between
data mining, statistics, machine learning and AI?”
2014.
[23] J. R. Koza, F. H. Bennett, D. Andre, and M. A.
Keane, “Automated Design of Both the Topology
and Sizing of Analog Electrical Circuits Using
Genetic Programming,” in Artificial Intelligence in
Design ’96, 1996.
[24] M. I. Jordan and T. M. Mitchell, “Machine
learning: Trends, perspectives, and prospects,”
Science. 2015.
[25] T. M. Mitchell, Machine Learning, no. 1. 1997.
[26] T. Hastie, R. Tibshirani, and J. Friedman, The
elements of statistical learning: data mining,
inference and prediction, vol. 9. Springer, 2017.
[27] R. Hirt, N. Kühl, and G. Satzger, “An end-to-end
process model for supervised machine learning
classification: from problem to deployment in
information systems,” in Proceedings of the
DESRIST 2017 Research-in-Progress, 2017.
[28] U. Neisser, Cognitive Psychology. 1967.
[29] N. J. Nilsson, Artificial Intelligence: A New
Synthesis, vol. 125, no. 12. 1998.
[30] K. Segerberg, J.-J. Meyer, and M. Kracht, “The
Logic of Action,” The Stanford Encyclopedia of
Philosophy, 2016. [Online]. Available:
https://plato.stanford.edu/archives/win2016/entries
/logic-action/. [Accessed: 15-Jun-2018].
[31] R. Thomason, “Logic and Artificial Intelligence,”
The Stanford Encyclopedia of Philosophy, 2016.
[Online]. Available:
https://plato.stanford.edu/archives/win2016/entries
/logic-ai/. [Accessed: 15-Jun-2018].
[32] Committee on Technology National Science and
Technology Council and Penny Hill Press,
“Preparing for the future of Artificial Intelligence,”
Committee on Technology National Science and
Technology Council and Penny Hill Press, vol. 58.
CreateSpace Independent Publishing Platform,
2016.
[33] P. Stone, R. Brooks, E. Brynjolfsson, R. Calo, O.
Etzioni, G. Hager, J. Hirschberg, S.
Kalyanakrishnan, E. Kamar, S. Kraus, K. Leyton-
Brown, D. Parkes, W. Press, A. Saxenian, J. Shah,
M. Tambe, and A. Teller, “Artificial Intelligence
and Life in 2030,” One Hundred Year Study Artif.
Intell. Rep. 2015-2016 Study Panel, p. 52, 2016.
[34] J. Haugeland, Artificial Intelligence: The Very Idea.
MIT Press, 1989.
[35] R. Bellman, An Introduction to Artificial
Intelligence: Can Computers Think? Boyd &
Fraser, 1978.
[36] A. Newell and H. A. Simon, “GPS, a program that
simulates human thought,” 1961.
[37] D. McDermott and E. Charniak, “Introduction to
artificial intelligence,” Int. J. Adapt. Control Signal
Process., vol. 2, no. 2, pp. 148149, 1985.
[38] E. Rich and K. Knight, “Artificial intelligence,”
McGraw-Hill, New, 1991.
[39] A. M. Turing, “Computing Machine and
Intelligence,” MIND, vol. LIX, no. 236, pp. 433
460, 1950.
[40] D. L. Poole, A. Mackworth, and R. G. Goebel,
“Computational Intelligence and Knowledge,”
Comput. Intell. A Log. Approach, no. Ci, pp. 122,
1998.
[41] S. Legg and M. Hutter, “Universal intelligence: A
definition of machine intelligence,” Minds Mach.,
vol. 17, no. 4, pp. 391444, 2007.
[42] J. R. Searle, “Minds , Brains , and Programs,” vol.
3, pp. 119, 1980.
[43] M. A. Gubrud, “Nanotechnology and international
security,” Fifth Foresight Conference on Molecular
Nanotechnology, vol. 1. 1997.
[44] R. Kurzweil, “The Singularity Is Near: When
Humans Transcend Biology,” Book, vol. 2011. p.
652, 2005.
[45] Laureate Online Education, “Model View
Controller Design Pattern,” Online, vol. 3, pp.
20002007, 2007.
[46] N. Kühl, M. Mühlthaler, and M. Goutier,
“Automatically Quantifying Customer Need
Tweets : Towards a Supervised Machine Learning
Approach BT - Hawaii International Conference
on System Sciences (HICSS-51), Waikoloa
Village, Hawaii, United States, 3rd - 6th January
2018,” 2018.
[47] C.-W. You, M. Montes-de-Oca, T. J. Bao, N. D.
Lane, H. Lu, G. Cardone, L. Torresani, and A. T.
Campbell, “CarSafe: a driver safety app that detects
dangerous driving behavior using dual-cameras on
smartphones,” Proc. 2012 ACM Conf. Ubiquitous
Comput. - UbiComp ’12, pp. 671672, 2012.
[48] E. Guizzo, “How Aldebaran Robotics Built Its
Friendly Humanoid Robot, Pepper, “IEEE
Spectrum,” 2014. [Online]. Available:
https://spectrum.ieee.org/robotics/home-
robots/how-aldebaran-robotics-built-its-friendly-
humanoid-robot-pepper. [Accessed: 15-Jun-2018].
[49] K. Woo, S. Meninger, T. Xanthopoulos, E. Crain,
D. Ha, and D. Ham, “Dual-DLL-based CMOS all-
digital temperature sensor for microprocessor
thermal monitoring,” in Digest of Technical Papers
- IEEE International Solid-State Circuits
Conference, 2009.
[50] T. Geller, “How do you feel? Your computer
knows,” Commun. ACM, vol. 6, no. 8, pp. 2426,
2014.
[51] L. A. Teixeira and A. L. I. De Oliveira, “A method
for automatic stock trading combining technical
analysis and nearest neighbor classification,”
Expert Syst. Appl., vol. 37, no. 10, pp. 68856890,
2010.
[52] K. Berns and J. Hirth, “Control of facial
expressions of the humanoid robot head ROMAN,”
in IEEE International Conference on Intelligent
Robots and Systems, 2006, pp. 31193124.
[53] P. Chapman, J. Clinton, R. Kerber, T. Khabaza, T.
Reinartz, C. Shearer, and R. Wirth, “Crisp-Dm
1.0,” Cris. Consort., p. 76, 2000.
[54] J. Gama, I. Žliobaitė, A. Bifet, M. Pechenizkiy, and
A. Bouchachia, “A survey on concept drift
adaptation,” ACM Comput. Surv., vol. 46, no. 4, pp.
1–37, 2014.
[55] F. Oroszi and J. Ruhland, “An early warning system
for hospital acquired,” in 18th European
Conference on Information Systems (ECIS), 2010.
[56] X. Yang, J. Su, and C. L. Tan, “A twin-candidate
model for learning-based anaphora resolution,”
Comput. Linguist., vol. 34, no. 3, pp. 327356,
2008.
[57] Z. Zheng, L. Zheng, and Y. Yang, “Pedestrian
alignment network for large-scale person re-
identification,” arXiv Prepr. arXiv1707.00408,
2017.
[58] A. M. Jorge, J. P. Leal, S. S. Anand, and H. Dias,
“A study of machine learning methods for detecting
user interest during web sessions,” in Proceedings
of the 18th International Database Engineering &
Applications Symposium on - IDEAS ’14, 2014, pp.
149157.
[59] T. M. Mitchell, W. Cohen, E. Hruschka, P.
Talukdar, J. Betteridge, A. Carlson, B. D. Mishra,
M. Gardner, B. Kisiel, J. Krishnamurthy, N. Lao,
K. Mazaitis, T. Mohamed, N. Nakashole, E. A.
Platanios, A. Ritter, M. Samadi, B. Settles, R.
Wang, D. Wijaya, A. Gupta, X. Chen, A. Saparov,
M. Greaves, and J. Welling, “Never-Ending
Learning,” AAAI Conf. Artif. Intell., pp. 2302
2310, 2015.
[60] E. Liebman, M. Saar-Tsechansky, and P. Stone,
“Dj-mc: A reinforcement-learning agent for music
playlist recommendation,” in Proceedings of the
2015 International Conference on Autonomous
Agents and Multiagent Systems, 2015, pp. 591599.
[61] F. Ruelens, S. Iacovella, B. J. Claessens, and R.
Belmans, “Learning agent for a heat-pump
thermostat with a set-back strategy using model-
free reinforcement learning,” Energies, vol. 8, no.
8, pp. 83008318, 2015.
[62] M. Rostami, S. Kolouri, K. Kim, and E. Eaton,
“Multi-Agent Distributed Lifelong Learning for
Collective Knowledge Acquisition,” arXiv Prepr.
arXiv1709.05412, 2017.
[63] Y. Yu, A. Eshghi, and O. Lemon, “VOILA : An
Optimised Dialogue System for Interactively
Learning Visually-Grounded Word Meanings
(Demonstration System),” in Proceedings of the
SIGDIAL 2017 Conference, 2017, pp. 197200.
[64] U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth,
“The KDD process for extracting useful knowledge
from volumes of data,” Commun. ACM, vol. 39, no.
11, pp. 27–34, 1996.
[65] Y. Nagar and T. W. Malone, “Making business
predictions by combining human and machine
intelligence in prediction markets,” Int. Conf. Inf.
Syst. ICIS 2011, pp. 116, 2011.
[66] O. Russakovsky, L. J. Li, and L. Fei-Fei, “Best of
both worlds: Human-machine collaboration for
object annotation,” in Proceedings of the IEEE
Computer Society Conference on Computer Vision
and Pattern Recognition, 2015, vol. 0712June,
pp. 21212131.
[67] A. Holzinger, “Interactive machine learning for
health informatics: when do we need the human-in-
the-loop?,” Brain Informatics, vol. 3, no. 2, pp.
119131, 2016.
[68] T. Hata, M. Suganuma, and T. Nagao, “Controlling
an Autonomous Agent for Exploring Unknown
Environments Using Switching Prelearned
Modules,” Electron. Commun. Japan, vol. 101, no.
5, pp. 8493, 2018.
[69] A. Rosenfeld, N. Agmon, O. Maksimov, and S.
Kraus, “Intelligent agent supporting humanmulti-
robot team collaboration,” Artif. Intell., vol. 252,
pp. 211231, 2017.
[70] H. O. Al-sakran, “Intelligent Traffic Information
System Based on Integration of Internet of Things
and Agent Technology,” Int. J. Adv. Comput. Sci.
Apl., vol. 6, no. 2, pp. 3743, 2015.
[71] A. Gudigar, S. Chokkadi, and R. U, “A review on
automatic detection and recognition of traffic sign,”
Multimed. Tools Appl., vol. 75, no. 1, pp. 333364,
2016.
[72] J. Lu, V. Behbood, P. Hao, H. Zuo, S. Xue, and G.
Zhang, “Transfer learning using computational
intelligence: A survey,” Knowledge-Based Syst.,
vol. 80, pp. 1423, 2015.
[73] K. Weiss, T. M. Khoshgoftaar, and D. D. Wang, “A
survey of transfer learning,” J. Big Data, vol. 3, no.
1, 2016.
[74] L. Baier, N. Kühl, and G. Satzger, “How to Cope
with Change? Preserving Validity of Predictive
Services over Time,” in Hawaii International
Conference on System Sciences (HICSS-52), 2019.
  • Conference Paper
    Full-text available
    Entwicklungen der Informationstechnologie stellen zunehmend Möglich-keiten bereit, den menschlichen „Wissensarbeiter“ durch kognitive Assistenzsysteme auf Basis Künstlicher Intelligenz (KI) in seinen Entscheidungen oder Aktionen zu unterstützen. Im vorliegenden Beitrag wollen wir beleuchten, welche Anforderungen an die zugrundeliegenden maschi-nellen Lernverfahren zu stellen sind, um die individuelle sowie gesellschaftliche Akzeptanz solcher „Augmented Intelligence“ zu gewährleisten.
  • Article
    Artificial intelligence is a rapidly evolving computerized technology affecting multiple aspects of our lives. It is predicted that artificial intelligence will lead to a fundamental change in practice of many professional fields, including medicine. One of the most significant advances in artificial intelligence involves digital imaging and image recognition. Consequently, radiologists, who work in the most digitalized field of medicine, need to be familiar with this rapidly progressing technology. "Artificial intelligence," "machine learning," and "deep learning" are terms that tend to be used interchangeably in terms of advanced computer algorithms, but each has a different meaning. Objectives for this article are to demystify these terms for radiologists and to establish a basic understanding of this topic for the reader. We also discuss the impact that artificial intelligence might have on the field of radiology in the foreseeable future. Although artificial intelligence is unlikely to replace radiologists any time soon (if ever), we explore how this technology could be beneficial to radiologists.
  • Article
    Full-text available
    This study is the first to apply a decomposition framework to clarify the determinants of AI technology invention. Consisting of 13,567 AI technology patents for the 2000-2016 period, our worldwide dataset includes patent publication data from the U.S., Japan, China, Europe, and the Patent Cooperation Treaty (PCT). We find that priority has shifted from biological and knowledge-based models to specific mathematical models and other AI technologies, particularly in the U.S. and Japan. Our technology type and country comparison shows that the characteristics of AI technology patent publication differ among companies and countries. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Full paper is available for download. https://ideas.repec.org/p/eti/dpaper/17066.html
  • Article
    Machine learning addresses the question of how to build computers that improve automatically through experience. It is one of today’s most rapidly growing technical fields, lying at the intersection of computer science and statistics, and at the core of artificial intelligence and data science. Recent progress in machine learning has been driven both by the development of new learning algorithms and theory and by the ongoing explosion in the availability of online data and low-cost computation. The adoption of data-intensive machine-learning methods can be found throughout science, technology and commerce, leading to more evidence-based decision-making across many walks of life, including health care, manufacturing, education, financial modeling, policing, and marketing.
  • Article
    Most current speech recognition systems use hidden Markov models (HMMs) to deal with the temporal variability of speech and Gaussian mixture models to determine how well each state of each HMM fits a frame or a short window of frames of coefficients that represents the acoustic input. An alternative way to evaluate the fit is to use a feed-forward neural network that takes several frames of coefficients as input and produces posterior probabilities over HMM states as output. Deep neural networks with many hidden layers, that are trained using new methods have been shown to outperform Gaussian mixture models on a variety of speech recognition benchmarks, sometimes by a large margin. This paper provides an overview of this progress and represents the shared views of four research groups who have had recent successes in using deep neural networks for acoustic modeling in speech recognition.
  • Conference Paper
    Full-text available
    We present an approach to detecting hate speech in online text, where hate speech is defined as abusive speech targeting specific group characteristics, such as ethnic origin, religion, gender, or sexual orientation. While hate speech against any group may exhibit some common characteristics, we have observed that hatred against each different group is typically characterized by the use of a small set of high frequency stereotypical words; however, such words may be used in either a positive or a negative sense, making our task similar to that of words sense disambiguation. In this paper we describe our definition of hate speech, the collection and annotation of our hate speech corpus, and a mechanism for detecting some commonly used methods of evading common "dirty word" filters. We describe pilot classification experiments in which we classify anti-semitic speech reaching an accuracy 94%, precision of 68% and recall at 60%, for an F1 measure of. 6375.