Thesis Proposal - Eric Sturzinger
— 3:30pm
Location:
In Person
-
Reddy Conference Room, Gates Hillman 4405
Speaker:
ERIC STURZINGER
,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://www.linkedin.com/in/eric-sturzinger-6203b265
Survival-Critical Machine Learning
Foundational Machine Learning (ML) models are trained on massive datasets collected and curated over months and years and largely deployed in hyperscale datacenters. As a result, such collection and training are strongly decoupled in which learning is not urgent. This is optimal when the cost of an incorrect response to a query is not catastrophic. However, ML systems that operate in adversarial or hostile physical environments must learn in a unique, edge computing paradigm on limited a priori training data.
This proposal introduces Survival-Critical Machine Learning (SCML), a new ML framework that defines optimal system design and operation that maximizes an ML system’s survivability in adversarial environments. In such scenarios, threats can evolve and morph continuously and an incorrect inference can potentially be catastrophic. In classical ML, the key performance indicator (KPI) of a system is some measure of output accuracy or quality with respect to the input. In SCML, the KPI is defined by the time delay between the initial arrival of new data and when the derived knowledge from it is embedded in a newly deployed model, when learning is urgent. In order to maximize survivability by adapting to evolving threats, SCML systems leverage Live Learning. Live Learning tightly couples data collection, inference, transmission, labeling, and training. It is feasible to implement SCML, which is a new model of continuous learning, for improved survival of edge-based systems in adversarial environments.
This work defines critical enhancements to Live Learning and its integration into a scalable and distributed SCML system. It explores the optimization of both traditional system-level design and properties as well as ML-focused improvements. Combined, these minimize an SCML system’s exposure to threats, thus maximizing its survivability in adversarial or hostile environments. We seek to inform and influence future SCML system design by modeling and evaluating the behaviors of a team of distributed SCML systems that perform Live Learning.
Thesis Committee:
Mahadev Satyanarayanan (Chair)
Rashmi Vinayak
Jeff Schneider
Babu Pillai (Intel Labs)