Signal Processing - November 2017 - 125

data sets and a shared label space has remained constant.
Well-defined source and target data sets, discrete and shared
label spaces, and constraints on labeled and unlabeled data are
possibly some of the reasons for limiting the applicability of
domain adaptation to real-world settings. This section discusses a few problems with the way domain adaptation models are
developed and evaluated in the research community along with
some proposals for changes. Directions for future research in
domain adaptation are later outlined. Most of the proposed
solutions in domain adaptation are based on models developed
in the following environment:
1) Two different data sets are used to represent the source
and the target domains. Domain adaptation models trained
on specific data sets may perform exceedingly well, adapting between the two data sets. These models do not guarantee the same performance when adapting between
different data sets. Every new pair of data sets may require
training its own adaptation model. A more universal
approach would be to have two well-defined domains
(rather than data sets) to represent the source and the target. Algorithms developed to address a particular domain
shift can guarantee performance across applications that
encounter the same domain shift.
2) The source data set is labeled, and the target data set is
unlabeled in unsupervised domain adaptation. This ap--
pears to be a stringent and restrictive constraint because, in
a real-world setting, it is possible to have a few labeled
samples for the target data set. Most domain adaptation
models do not account for labeled target data. Optimal
model-parameters and model-design choices, which are
usually estimated using a labeled validation set, cannot be
applied for unsupervised domain adaptation because there
are no labeled target examples. There is no prescribed procedure to validate the model parameters of current domain
adaptation methods [32], [56], [98]. A few labeled target
examples are essential for validation purposes. Tzeng at al.
[39] outline a semisupervised adaptation model, where
some of the categories in the target have a few labeled data
points. Domain adaptation models that account for a few
labeled data points in the target domain can only outperform their unsupervised counterparts.
3) The label space of the source and target is exactly the same.
Domain adaptation models assume a shared label space
between the source and target domains. Most real-world
scenarios satisfy such a criteria. However, robust domain
adaptation models should account for a relaxed setting,
where there is no restriction on the label space of the domains
being exactly the same.
4) Current domain adaptation approaches are modeled to solve
only a specific subset of adaptation problems, which assume
closed-world representations with a fixed set of discrete
and disjoint labels. However, most real-world problems
have generic representations, and they cannot be limited to
discrete or disjoint label settings. Newer algorithms in
domain adaptation must remodel the basic problem setup and
evaluation protocol in step with real-world applications.

To solve real-world domain adaptation problems and eventually address the problem of artificial general intelligence, these
changes are suggested in domain adaptation research. Apart
from these changes to the basic approach in domain adaptation, the following subsections outline the specific directions
for future research in this area.

Modeling domain shift
The concept of a domain has been defined vaguely in computer
vision. Images from different data sets are viewed as belonging to different domains. Data sets have an inherent bias, and
images from a data set have certain properties that can help
identify the data set [99]. However, there has been limited
effort in understanding what creates this bias and on modeling the domain shift between data sets. The authors in [100]
attempt to identify the domainness-a measure for domain
specificity of an image.
The difficult problem of modeling domain shift in computer vision has been rarely addressed. There has been work
on identifying domains from a mixture of multiple data sets
and then studying transfer of knowledge between the domains
[101]. Although this does not necessarily model a domain, it
provides some direction toward identifying a domain through
analysis. The difficulties of modeling domain shift in computer vision mostly arise due to variations in representation and
not merely variations in the data being represented. The very
process of imaging (camera perspective and occlusion), storage
(resolution and size), and representation (color, brightness, and
contrast) can lead to variations. Image background (context)
is another cause for variation. Finally, the diversity in the real
data itself could also lead to variations in their images.
Most domain adaptation systems create adaptive models
that perform generic domain adaptation. The models are often
guided by the data sets that are used. On the other hand, it
might be beneficial to tailor the adaptation model to a specific
variation in the data. This would, however, need a comprehensive understanding of domain shift. It might also lead to
task-specific domain adaptation models based on the nature of
domain-shift, leading to increased adoption of domain adaptation in real-world applications.

New data sets
Current data sets for domain adaptation are not based on any
models of domain shift. They are merely data samples coming from different sources, all with the same categories. The
domain difference between these data sets is attributed to the
bias between the data sets, without a specific model characterizing the domain shift [99]. The domain adaptation procedures that are developed using these data sets can, therefore,
be considered very generic. There is no guarantee on the
performance of these procedures when applied to new problems. For example, if a domain adaptation approach were to
be developed using the digit data sets USPS and MNIST [102],
there is no guarantee that this procedure would work well for a
domain adaptation problem with medical images. On the other
hand, if a data set were to be created based on a domain shift

IEEE SIGNAL PROCESSING MAGAZINE

November 2017

125

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