Ziv Scully
Cornell University, NY·Computer Science·Early-career
Training and running artificial intelligence (AI) models uses huge amounts of data center time and energy. Much of that cost comes from scheduling decisions, such as which candidate model to try training next or which incoming task to serve first.
- Fields
Computer Science
Mathematics & Statistics
Engineering
- Topics
Gittins index
scheduling
queueing theory
- Grants
- 1 known grant, through Mar 2031
What they work on
- Develop Gittins index acquisition functions for exotic Bayesian optimization problems, so that searches over possible AI models find higher-quality models with less training time and energy.
- Design Gittins index schedulers that order incoming AI inference tasks in queues to minimize tail latency, the waiting time of the slowest requests, while keeping throughput high.
- Prove that the new Gittins index algorithms are optimal or near-optimal under stated assumptions, and test them through open-source prototypes and pilot studies.
More about the research
This project asks whether a classical but rarely used mathematical tool, the Gittins index, can make those decisions far more efficient. The Gittins index acts as a "universal prioritizer." It compresses everything known about a complicated task into a single priority number.
Why it matters
More efficient training and scheduling could cut the energy use of AI data centers and make AI services faster. The exploration methods could also help with engineering design searches such as drug discovery and fusion reactor design. The scheduling methods could reduce waiting times in other computer systems and in everyday services like food delivery and urgent care clinics.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Jia Liu
Ohio State University, OH·Engineering·Early-career
Develop communication-efficient distributed optimization methods at the wireless physical layer, so machine learning training and inference across edge devices use less communication.
Seyed Iravani
Northwestern University, IL·Engineering
Develop behavioral and analytical models of how customers react to delay announcements, combining stochastic dynamic programming and queueing theory with prospect theory and econometrics to decide what wait information to give and when.
Somayeh Sojoudi
University of California-Berkeley, CA·Engineering·Early-career
Analyze when key non-convex machine-learning problems have spurious solutions (points that look locally best but are not globally best), and how the amount of data and problem structure remove them.
The school
Cornell University: 37 labs on PI Pages, $1.6B a year on research, 537 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.