Vavoso, Greg, School of Engineering and Applied Science, University of Virginia
Advisors
Ferguson, Sean, University of Virginia
Powell, Harry, EN-Elec/Computer Engr Dept, University of Virginia
Abstract
The technical project in this portfolio is a temperature checkpoint system. This is an
automated health screening door connecting temperature measurements to a door lock. When a
prospective entrant approaches a door outfitted with this system, they will place their wrist under
a mounted box, triggering a motion detector and awakening a non-contact temperature sensor.
The temperature sensor will then transmit a temperature to a microcontroller and decide whether
or not to unlock the door. Data is also sent to an external server for data analytics and further
computation. Access to this external server is available through a locally hosted website. The
STS portion of research was also inspired by the recent COVID-19 pandemic. In the early onset
of the pandemic, rapid shifts in human behavior caused various machine learning systems to fail
in unpredictable ways. The past data that these machine learning systems were trained on did not
reflect the current behavior as the pandemic hit, thus the systems could not predict outside of
their trained range. The paper aims to understand the harm in depending on machine learning
systems that could potentially fail catastrophically. The paper synthesizes various sources on
technological dependence in order to form a set of precautions: 1) aim for complete transparency
during design, 2) have a robust backup system prepared, and 3) design with detection
mechanisms in place for immediate transition to the backup system upon failure. These two
projects are coupled in that the door locking mechanism is an example of a system that could be
harmfully depended on if various backup systems are not in place. Some of these precautions
were used in the design of this system, such as various backup mechanisms in the case of failure.
For example, if the system loses power, the mechanism unlocks rather than locking to prevent
trapping individuals. Some of the hands-on design of the technical system inspired and led to the
development of the set of precautions outlined in the STS portion of this portfolio.
Vavoso, Greg. Mitigating Harmful Machine Learning Dependencies. University of Virginia, School of Engineering and Applied Science, BS (Bachelor of Science), 2021-05-14, https://doi.org/10.18130/2er1-hz43.
