Do Young Eun
North Carolina State University, NC·Computer Science
Training a machine-learning model means repeatedly checking how wrong the model is across a huge dataset. Doing that over every data point each time is too slow.
- Fields
Computer Science
Engineering
Mathematics & Statistics
- Topics
machine learning
decentralized learning
distributed optimization
- Grants
- 1 known grant, through Sep 2027
What they work on
- Design self-repellent random walks, multiple interacting walkers that avoid their collective history, with adaptive repellency across space and time to maximize sampling efficiency, and prove theoretical performance guarantees.
- Analyze stochastic approximation and decentralized learning algorithms driven by self-repellent random walks instead of MCMC, deriving finite-sample and long-run performance bounds that balance speed and efficiency.
- Develop low-complexity multi-timescale versions of stochastic approximation that always speed up learning in a given Markovian environment, and co-design them with nonlinear Markovian sampling for decentralized learning.
More about the research
Instead, algorithms check small samples, much like an opinion poll surveys a few people instead of everyone. Sampling cleverly, by favoring the most informative data, can make learning faster and more accurate. Standard sampling methods are built on Markov chain Monte Carlo (MCMC), a family of random-walk methods. These rest on the Markov assumption: each new sample depends only on the current one, not on the full history. The project asks whether sampling that deliberately uses past history can make sampling, optimization and decentralized learning more efficient, both in practice and in theory.
Why it matters
Many fields rely on MCMC and random-walk algorithms as the default tool. These include sampling from high-dimensional spaces with graph constraints, distributed inference using random walks on networks, and stochastic optimization with Markovian noise. More efficient sampling with proven guarantees could make large-scale and decentralized machine-learning training faster and more accurate.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Georg Oeltzschner
Johns Hopkins University, MD·Biomedical Engineering
Develop a multiverse analysis framework that fits brain 1H-MRS data with many candidate models, including those in existing MRS software, and quantifies how much metabolite estimates vary across them.
Evrydiki Nestoridi
Stony Brook University, NY·Mathematics & Statistics
Develop versatile techniques for bounding mixing times, the time a Markov chain takes to reach equilibrium, across particle systems, random walks on groups and card shuffles.
Hang Liu
Rutgers University-New Brunswick, NJ·Computer Science·Early-career
Design new ways to construct and update transition probabilities for Monte Carlo graph sampling and random walk methods so sampling on GPUs runs faster.
The school
North Carolina State University: 40 labs on PI Pages, $717M a year on research, 603 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.