LATEX IEEE TEMPLATE TUTORIAL 1
Literature Review
Student name
Abstract—The abstract goes here.
Index Terms—CMPE185, LATEX Tutorial, IEEEtran, journal, LATEX, paper, template.
✦
1 INTRODUCTION
A RTIFICIAL intelligence (AI) is becomingincreasingly prevalent in society, with
the potential to revolutionize numerous
fields, from healthcare and transportation
to education and finance. However, as AI
systems become more integrated into our
daily lives, it is essential to recognize and
mitigate the potential for bias in their design
and implementation. Bias can be defined as
the systematic ways in which AI systems can
be designed, trained, or used in a manner
that unfairly favors or disadvantages certain
groups of people or outcomes. The impact of
bias in AI can be significant, as AI systems are
increasingly being used in decision-making
processes that affect people’s lives, such as
hiring, lending, and criminal justice. Bias in
AI systems can result in unfair outcomes and
reinforce existing inequalities.
There are three key articles that was pro-
fessional journal articles that dig deep into this
topic: ”Mitigating Unwanted Biases with Ad-
versarial Learning” by Zhang, Lemoine, and
Mitchell, ”Man is to Computer Programmer as
Woman is to Homemaker? Debiasing Word Em-
beddings” by Bolukbasi, and ”Fairness and Ab-
straction in Sociotechnical Systems” by Craw-
ford and Calo. These articles provide valuable
insights into the nature of bias in AI and pro-
pose solutions to address the issue. By examin-
ing these articles in-depth, this review aims to
provide a comprehensive understanding of the
current state of research on bias in AI and the
ways in which it can be addressed.
2 MITIGATING UNWANTED BIASES
WITH ADVERSARIAL LEARNING
Mitigating Unwanted Biases with
Adversarial Learning, written by Zhang,
Lemoine, and Mitchell, addresses the issue of
bias in machine learning models and proposes
a method to mitigate it. The authors note
that machine learning models often reflect the
biases that exist in the data they are trained on.
This can lead to unfair outcomes, especially
when the models are used in high-stakes
decision-making processes such as hiring,
lending, and criminal justice. The authors
propose a method called adversarial learning
to mitigate unwanted biases in machine
learning models. Adversarial learning is a
two-part process that involves training two
models simultaneously: a primary model and
an adversarial model. The primary model is
trained to predict the outcome of interest, while
the adversarial model is trained to predict the
sensitive attribute(s) that the primary model
may use to make decisions, such as gender or
race.
During training, the adversarial model tries
to predict the sensitive attribute(s) from the out-
put of the primary model. At the same time, the
primary model tries to make accurate predic-
tions while minimizing the ability of the adver-
sarial model to predict the sensitive attribute(s).
0000–0000/00$00.00 © 2007 IEEE
LATEX IEEE TEMPLATE TUTORIAL 2
By training the two models simultaneously,
the primary model learns to make predictions
that are accurate and fair, while the adversarial
model learns to identify and reduce the effects
of unwanted biases. The authors tested their
method on several datasets and found that it
was effective in mitigating unwanted biases
in machine learning models. The method was
able to significantly reduce the disparities in
prediction accuracy between different groups
and to produce fairer outcomes. The proposed
method is a promising approach to reducing
unwanted biases in machine learning models,
and could have significant real-world impact by
making decision-making processes fairer and
more equitable. However, the authors note that
their method is not a silver bullet and further
research is needed to fully understand its limi-
tations and potential drawbacks.
3 MAN IS TO COMPUTER PROGRAM-
MER AS WOMAN IS TO HOMEMAKER?
DEBIASING WORD EMBEDDINGS
Bolukbasi’s article provides a comprehensive
survey of bias in machine learning, with
a focus on the different types of bias that
can arise during the various stages of the
machine learning process. The author discusses
several types of bias, including sampling
bias, measurement bias, and algorithmic bias.
Additionally, the article provides a detailed
review of the various techniques that have been
proposed to mitigate bias in machine learning,
including pre-processing data, post-processing
decisions, and optimizing algorithmic fairness.
One of the key strengths of Bolukbasi’s
survey is its clear and accessible explanation
of the different types of bias that can arise
during the machine learning process. The
author provides several examples to illustrate
each type of bias and discusses the potential
consequences of each. For instance, the author
notes that measurement bias can arise when
certain groups of people are underrepresented
in the training data, leading to inaccurate
predictions for those groups.
In terms of techniques for mitigating bias,
Bolukbasi provides a detailed review of the var-
ious approaches that have been proposed. For
example, the author discusses pre-processing
techniques, such as oversampling or undersam-
pling, which can be used to balance the rep-
resentation of different groups in the training
data. Additionally, the author discusses post-
processing techniques, such as threshold cali-
bration and reject option classification, which
can be used to adjust the decision-making pro-
cess to ensure fairness. Finally, the author dis-
cusses several techniques for optimizing algo-
rithmic fairness, such as constraint optimiza-
tion and regularization. Bolukbasi’s survey pro-
vides a comprehensive and accessible overview
of the different types of bias that can arise
in machine learning, as well as the various
techniques that have been proposed to miti-
gate these biases. By synthesizing existing re-
search, the article provides a useful resource
for researchers and practitioners interested in
understanding and addressing bias in machine
learning systems.
4 FAIRNESS AND ABSTRACTION IN SO-
CIOTECHNICAL SYSTEMS
The third article is ”Fairness and Abstraction
in Sociotechnical Systems” by Crawford and
Calo. This article takes a different approach
to understanding and mitigating bias in
AI by exploring the concept of fairness in
sociotechnical systems. The authors argue that
fairness cannot be understood solely through
the lens of individual algorithms or models,
but must be examined in the context of larger
systems and social structures. The article
begins by discussing the limitations of current
approaches to fairness in AI, which often focus
on individual algorithms or models and fail
to account for the larger social, political, and
economic structures that shape the use and
impact of AI. The authors argue that these
structures can create and perpetuate biases in
AI systems, even when individual algorithms
or models are designed to be fair.
LATEX IEEE TEMPLATE TUTORIAL 3
To address this issue, the authors propose
a new framework for understanding fairness
in sociotechnical systems, which they call
”fairness through abstraction.” This framework
involves abstracting away from individual
algorithms or models and instead examining
the larger social, political, and economic
structures that shape the use and impact of AI.
By focusing on these structures, the authors
argue, we can identify and address biases in AI
systems at a more fundamental level.
The article provides several examples of
how this framework can be applied in practice.
One example is the use of predictive policing
algorithms, which have been shown to dis-
proportionately target communities of color.
Rather than focusing solely on the algorithm
itself, the authors argue that we must examine
the larger social and political structures that
give rise to over-policing in these communities.
By addressing these structures, we can work to
reduce the biases that are built into predictive
policing algorithms. Another example the ar-
ticle provides is the use of hiring algorithms,
which have been shown to perpetuate gender
and racial biases. Again, the authors argue that
we must examine the larger social and eco-
nomic structures that give rise to these biases,
such as gender and racial discrimination in the
workplace. By addressing these structures, we
can work to reduce the biases that are built into
hiring algorithms.
5 CONCLUSION
In conclusion, bias in artificial intelligence is a
critical issue that requires attention and miti-
gation efforts. The reviewed articles highlight
the various forms of bias that can arise in AI
systems and the potential impacts on differ-
ent domains, such as healthcare, finance, and
criminal justice. Moreover, they emphasize the
importance of using diverse and representative
datasets, fair algorithms, and proactive mon-
itoring and testing to mitigate unwanted bi-
ases. Adversarial learning and abstraction are
promising approaches to identify and mitigate
bias, but further research is needed to explore
their effectiveness and applicability in different
contexts. The ethical and social implications of
bias in AI should also be addressed through
collaboration among researchers, policymakers,
and other stakeholders. It is essential to ensure
that AI technologies are designed and deployed
in a fair and inclusive manner that benefits all
individuals and communities.
6 REFERENCE
[1] Zhang, B., Lemoine, B., & Mitchell, M.
(2018). Mitigating unwanted biases with adver-
sarial learning. arXiv:1801.07593
[2] Bolukbasi, T., Chang, K. W., Zou, J.
Y., Saligrama, V., & Kalai, A. T. (2016). Man
is to Computer Programmer as Woman is
to Homemaker? Debiasing Word Embeddings.
arXiv:1607.06520
[3] Crawford, K., & Calo, R. (2016). Fairness
and abstraction in sociotechnical systems.
https://dl.acm.org/doi/10.1145/3287560.3287598