Signal Processing - November 2017 - 120

A few domain adaptation methods have labeled target data sets
but use the source data to augment the number of labeled samples.
These methods are also referred to as supervised domain adaptation and are used to train classifiers for unseen target data [35],
[37], [38]. Tzeng et al. [39] propose a new paradigm for supervised
domain adaptation where a few labels are present for only a few of
the categories in the target domain. Their reasonable assumption
is that, as is often encountered in a real-world setting, it is possible
to get a few labeled target samples. Although the popular setting
is two-domain adaptation, there are also examples of multisource
domain adaptation as in [40]-[42]. In this article, we consider the
following definitions:
■■ supervised domain adaptation: when the target domain has a
few labeled samples for all of the categories
■■ unsupervised domain adaptation: when the target domain
does not have any labeled samples.
A majority of the domain adaptation literature cited in this article
is unsupervised.

Relevance of domain adaptation
Human intelligence is a competitive benchmark that machine
intelligence is seeking to emulate and eventually outperform.
One of the hallmarks of human intelligence is the ability to adapt
and transfer knowledge across multiple domains. For example,
if humans are familiar with a language, they can easily understand almost anyone speaking it, even if they were to hear it for
the first time; or, if a person has learned to drive a car, he or she
can easily adapt to driving a truck by adapting some previously
learned knowledge to the new setting. To enhance machine intelligence to the level of human intelligence and beyond, machinelearning models will have to model knowledge transfer. The
ability to transfer knowledge will provide tools to process the
vast amounts of unlabeled data available in the form of online
video, audio, images, and text. These advances in artificial intelligence and machine-learning will greatly benefit a wide range
of signal processing applications, including communication systems, financial markets, medical imaging, robotics, and digital
video processing.
The state-of-the-art algorithms for domain adaptation are
dominated by deep-learning-based approaches. Deep-learning
methods are outperforming standard nondeep-learning techniques for domain adaptive image classification. The success of
deep-learning methods has led to a rapid growth in domain adaptation research. It is necessary to categorize the myriad approaches and organize them to get a better understanding of the current
research in domain adaptation. This article provides a classification of deep-learning approaches for domain adaptation. It also
highlights the drawbacks with current approaches and outlines
directions for future research.

Shallow domain adaptation: Survey
Prior to the introduction of deep neural networks for vision
(AlexNet [43]), computer vision researchers relied on handcrafted features like scale invariant feature transform (SIFT) [44],
histogram of oriented gradients (HoG) [45], etc. to create a bagof-words-based vector representation for images and videos [46].
120

Domain adaptation techniques developed and studied using these
features are called shallow methods (as opposed to deep-learning
methods). It is important to understand some of these approaches,
as they form the basis of our understanding of domain adaptation. In addition, the first batch of deep-learning methods developed for domain adaptation are based on a few of these shallow
domain adaptation techniques.
There are many nondeep-learning (shallow) approaches that
address the problem of domain-shift in unsupervised domain
adaptation. All of these procedures work at the level of features,
i.e., the images are represented as feature vectors, and the domain
adaptation algorithm attempts to reduce the domain disparity
between the feature vectors of the source and the target. Since
the goal is to classify the target data, one straightforward technique is to modify a support vector machine (SVM) classifier
trained for the source data and adapt it to classify target data.
In [10], Bruzzone and Marconcini introduce the domain adaptive SVM (DASVM), where the source SVM decision boundary is iteratively modified and adapted to classify target data. In
[47], Aytar and Zisserman develop a projective model transfer
SVM (PMT-SVM), where a transformation matrix is learned
to adapt SVM decision boundaries across domains. Hoffman
et al. [48], develop the max-margin domain transfer (MMDT)
approach, where a linear SVM decision boundary for a source
is transformed to classify target data. Their work is an extension
of the seminal work by Saenko et al. [49], where a transformation
matrix is learned to cluster the source and target data based on
category. Both of these methods consider a few labeled samples
in the target domain. Other linear procedures project the source
and the target data to a common subspace, where domain alignment is bettered. In [50] and [51] the authors estimate a common
subspace to align the principal axes of the source and target feature spaces.
When linear feature-based approaches like linear transformations are inadequate in overcoming domain disparity, nonlinear
techniques are applied to ameliorate the domain discrepancy
between the source and the target. Nonlinear transformations
project the data points to a high-dimensional space and align
the domains in that space. Reducing domain disparity through
nonlinear alignment of data has been made possible with the
maximum mean discrepancy (MMD), which is a nonparametric
distance estimate designed by embedding the data into a reproducing kernel Hilbert space (RKHS). The data are mapped to a
high-dimensional (possibly infinite-dimensional) space defined
by U (X) = [z (x 1), f, z (x n)]. z : R d " H defines a mapping
function, and H is a RKHS. When two data sets belong to the
same distribution, their MMD is zero. Gretton et al. in [52], introduced the MMD to estimate the distance between the source and
the target data sets, which is given by
MMD = 1 / z ^ x si h - 1
ns i =1
nt
ns

j =1

t
z ^ x jh

. (2)
H

The distance between the two distributions is the distance
between their means in an RKHS. When the RKHS is universal, the MMD measure approaches zero only when the

IEEE SIGNAL PROCESSING MAGAZINE

November 2017

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