Signal Processing - November 2017 - 9

interpretable deep-learning models with the aim of creat-
ing new algorithms and methods that can overcome current
limitations of machine-learning systems in their lack of
ability to explain the actions, decision, and prediction out-
comes to human users while promising to perceive, learn,
decide, and act on their own. This new class of machinelearning systems will allow users to understand and thus trust
the system's outputs and to foresee and predict future system
behaviors. To this end, neural networks and symbolic systems
need to be integrated, enabling the machine-learning systems
themselves to construct models that will explain how the
world works. That is, they will discover by themselves the un-
derlying causes or logical rules that shape their prediction and
decision-making processes interpretable to human users in
symbolic and natural language forms. An initial work in this
direction makes use of an integrated neural-symbolic repre-
sentation called tensor-product neural memory cells, which
can be decoded back to symbolic form without loss of infor-
mation after extensive learning in the neural-tensor domain.
Another potential breakthrough in machine-learning
research is in new algorithms for reinforcement and unsu-
pervised deep learning, which make use of weak or even no
training signals paired to inputs to guide the learning. Effec-
tive reinforcement-learning algorithms will allow machinelearning systems to learn via interactions with possibly
adversarial environments and with themselves.
The most challenging problem, however, is unsupervised
learning, for which no satisfactory machine-learning algo-
rithms have been devised so far in practical applications.
The development of unsupervised learning techniques is
significantly behind that of supervised and reinforcement
deep learning. The most recent development in unsuper-
vised learning exploits sequential output structure and ad-
vanced optimization methods to alleviate the need for using
labels in training prediction systems [12].
Future advances in unsupervised learning include taking
into account new sources of learning signals such as the struc-
ture of input data and building conditional generative models.
In this context, the recent popular topic of generative adver-
sarial networks [2] is a highly promising direction exploiting
the long-standing concept of analysis by synthesis. A closely
related open problem is multimodal deep learning with crossdomain information as low-cost supervision. Standard speech
recognition, image recognition, and text classification meth-
ods make use of supervision labels within each of the speech,
image, and text modalities separately. This is far from how
children learn to recognize speech and classify text. For ex-
ample, children often get a distant "supervision" signal for
speech sounds by an adult pointing to an image scene or text.
A final future direction for tackling open problems in
machine learning is the paradigm of learning-to-learn or
metalearning; i.e., how to design a machine-learning system
that improves or automatically discovers a learning algo-
rithm. Learning-to-learn is a powerful emerging paradigm
and a fertile research direction expected to impact a wide
range of real-world applications.

Holcombe Department of Electrical and Computer Engineering
Faculty Search in Computer Engineering and
Electrical Engineering
The Holcombe Department of Electrical and Computer Engineering
at Clemson University is seeking applicants for multiple computer
engineering and electrical engineering tenure-track or tenured
faculty positions at the rank of assistant professor or associate
professor. The Department has a particular interest in applicants
in the following technical areas: (1) machine learning, computer
vision, artificial intelligence, signal processing, with collaborations in
biomedical engineering, health science, or automotive engineering; (2)
embedded computing, sensors, wearables; (3) high-performance
computing with an emphasis on big data, high-performance
networking, or accelerated computing architectures; and (4) cyber
security and cyber-physical system security. Outstanding assistantprofessor candidates will be considered for the Warren Owens
Assistant Professorship.
The Holcombe Department of ECE is one of the largest and
most active departments in Clemson University, with 32 primary
and 14 affiliated full-time faculty members, approximately 550
undergraduates and 190 graduate students. Annual research
expenditures exceed $8.6 million. Many members of the faculty
are known internationally; they include eight IEEE Fellows, three
endowed chairs, and four named professors. Annual funded
research expenditures exceed $8.6 million. The Department and
Clemson have highly successful computing-focused research
programs in high-performance computing and networking; privacy,
communications security, and secure control systems; and mobile
health devices.
Clemson University is the largest land-grant institution in South
Carolina, enrolling 18,600 undergraduates and 4,800 graduate
students. Seven colleges house strong programs in architecture,
engineering, science, agriculture, business, social sciences, arts
and education. A faculty of 1,500 and staff of 3,700 support 84
undergraduate degree offerings, 73 master's degree programs and
40 Ph.D. programs. An annual operating budget of approximately
$956 million and an endowment of $621 million fund programs
and operations. The University has externally funded research
expenditures of $100 million per year. Research and economic
development activities are enhanced by public-private partnerships
at 4 innovation campuses and 6 research and education centers
located throughout South Carolina. Clemson University is ranked
23th among national public universities by U.S. News & World Report.

Applicants must have an earned doctorate in electrical
engineering, computer engineering, or a closely related field.
Applicants should submit a current curriculum vitae, statements of
research and teaching strategy, and a minimum of five references
with full contact information. Application material should be
submitted electronically at the following Web link:
http://apply.interfolio.com/39731
To ensure full consideration, applicants must apply by December
1, 2017; however, the search will remain open until the position
is filled.
Clemson University is an AA/EEO employer and does not discriminate
against any person or group on the basis of age, color, disability,
gender, pregnancy, national origin, race, religion, sexual orientation,
veteran status or genetic information. Clemson University is building
a culturally diverse faculty committed to working in a multicultural
environment and encourages applications from minorities and women.

http://apply.interfolio.com/39731

Table of Contents for the Digital Edition of Signal Processing - November 2017