Complexity Class Analysis with Machine Learning; Privacy, Prejudice, and Pixels: A Journey Through the Implications of Machine Learning

Author:
Hong, Mitchell, School of Engineering and Applied Science, University of Virginia

Advisors:
Wylie, Caitlin, University of Virginia
Vrugtman, Rosanne, EN-Comp Science Dept, University of Virginia

Abstract:

Machine Learning (ML) has emerged as one of the most revolutionary technologies in the 21st
century. Consequently, the questions of when, where, and by whom it should be used arise. The
collective societal response to these questions is 'whenever, wherever, and by whoever,' which is
completely justifiable. ML technology has enabled coastal countries to predict flood patterns,
law enforcement to detect faces with high precision and accuracy, and automobiles to drive
without any human input. The prowess of ML is unquestionable, but efficacy cannot be simply
evaluated on performance – social and ethical factors must be considered. There exists a delicate
balance between harnessing the transformative potential of ML and mitigating its adverse social
impacts. A holistic approach is necessary in advancing the application of ML technology
throughout society. My research covers two topics. My technical paper highlights a powerful use
case of ML. My research paper explores the implications of ML on data privacy and bias within
society.
My technical paper investigates the cutting-edge application of ML in the analysis of complexity
classes. The study of complexity classes or the “hardness” of problems has been a leading field
in computer science research. Efficiency of software is bounded to the complexity class of the
underlying problem the software solves. There are many solutions to the same problem; some
more effective than others. This is determined by decomposing a solution into its constituent
parts and analyzing them to determine the spatial and computational upper bounds. Based off of
these analyses, various problems have been sorted into their respective complexity classes over
the last few decades. My paper proposes the introduction of ML into complexity class analysis.
Leveraging a description of a problem along with its associated complexity class, a machine
learning model can be trained to predict an arbitrary problem’s complexity class. With this,
computer scientists can automate the categorization of problem difficulty and potentially find
discrepancies between their intuition and the output of the model. ML has the potential to assist
in determining whether P = NP, one of the most debated fields in Computer Science research. If
P = NP, then the class of problems solvable in polynomial time (P) would be equivalent to the
class of problems whose solutions can be verified in polynomial time (NP), meaning that every
problem whose solution can be quickly verified could also be quickly solved. This would alter
the computational landscape in an enormous manner.
My research paper examines the implications of ML technology on society. ML, while
innovative and powerful, has the potential to harm its stakeholders. Models are inherently limited
by the quality of data they train on – they utilize a vast amount of data, some of which is
gathered without sufficient ethical and moral considerations. When trained with sensitive data,
models have the potential to leak information, exposing critical information that was once
secure. If a model is trained on biased data, the model will output biased results. In critical
environments like healthcare and law enforcement, biased results are detrimental and will
perpetuate bias in a rapid manner. Due to the complexity of ML technology, matters of data
privacy and bias are difficult to resolve. My research examines these complications and
challenges associated with legal regulations.
My research is confined to discussions within academia and my own analytical insights, rather
than to practical field applications. While it covers the broad, overarching problems of ML
technology, much more attention needs to be focused on the minute details of ML in application-specific contexts. In more detail, ML will be applied differently across various sectors, making it
crucial to distinguish its uses, implementation methods, and the individuals it will impact. Future
research should address these topics, using a more technical and experimental approach – instead
of an overview, build a model and demonstrate infringements on data privacy and bias.