Safe and Ethical Implementation of Intelligent Systems
Name
dai-zhengdai-phd-eecs-2024-thesis.pdf
Description
Thesis PDF
Size
31.54 MB
Format
Adobe PDF
Checksum (MD5)
c10239985849ec5a37de5136aa8afb1d
Author(s)
Dai, Zheng
Advisor(s)
Gifford, David K.
Date Issued
September 2024
Publisher
Massachusetts Institute of Technology
Abstract
In the year 2024, the prospect of solving human level tasks using intelligent systems is no longer the subject of science fiction. As these systems play an increasingly critical role in our day-to-day lives, it becomes ever more important to consider the safety and ethics surrounding their implementation. This is a multifaceted challenge spanning multiple disciplines, involving questions at the regulatory, engineering, and theoretical levels. This thesis discusses three projects that span these levels. We first explore the problem of tracing causal influence from training data to outputs of generative models. In our exploration we encounter the phenomenon of unattributability, and consider its scientific and regulatory implications. We next tackle the challenge of designing a high diversity library of therapeutics that is depleted of dangerous off-target binders using intelligent systems, developing a suite of inference and optimization tools along the way. Finally, we derive universal bounds for the robustness of image classifiers that inform us of how safe these intelligent systems can be in theory. Together, these projects present a multilevel overview of the safe and ethical implementation of intelligent systems.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
