My Thai
University of Florida, FL·Computer Science
Machine learning works best with huge amounts of data, but much of the most useful data — medical scans, phone sensor readings, traffic records — is private and spread across many devices. Both projects here tackle the same underlying tension: how to build learning systems that are accurate, that can be trusted when attacked or given unfamiliar data, and that can explain their own decisions.
- Fields
Computer Science
Engineering
Mathematics & Statistics
- Topics
decentralized federated learning
bilevel optimization
stochastic optimization
- Grants
- 2 known grants, through Sep 2027
What they work on
- Design communication topologies and schedules for Decentralized Federated Learning, where devices train a shared model by exchanging updates peer-to-peer, using graph theory and network measurements to improve convergence and robustness.
- Develop stochastic algorithms for bilevel optimization — an optimization nested inside another, as in meta-learning — in decentralized settings, and establish theoretical convergence guarantees under periodic device-to-device communication.
- Characterize security threats unique to decentralized learning, arising from malicious participants and device-to-device links rather than a central server, and build principled defenses with privacy preservation.
- Quantify the trustworthiness of machine learning models through metrics such as adversarial robustness and semantic saliency, establishing theoretical bases and practical limits for these guarantees.
- Develop a proof-based explanation method that probes every hidden layer of a trained model to identify the critical layers and neurons behind an individual prediction, with a framework letting end users verify explanations.
- Identify retinal blood-vessel biomarkers of neurodegenerative disease using proof-based probing of biomarker graph networks, and link them to brain vasculature with cross-modality and spatio-temporal recurrent explainable models.
More about the research
One project attacks this at the level of how devices train a model together without pooling raw data; the other at the level of whether a trained model's predictions can be verified and understood well enough to use in medical diagnosis. Both combine mathematical theory with prototype systems and applications in Internet-of-Things networks, smart transportation, and medical imaging.
Why it matters
Privacy-preserving training across phones, vehicles and sensors could let machine learning be used on sensitive data that cannot legally or practically be centralized, while defenses against malicious participants make such systems safer to deploy in smart transportation and Internet-of-Things settings. On the medical side, explainable models of retinal blood vessels could support inexpensive, non-invasive early detection of neurodegenerative disease, with explanations and proofs that clinicians can check before acting.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Shuai Zhang
New Jersey Institute of Technology, NJ·Engineering
Develop topology sampling and graph neural network pruning methods for single-layer graph models, trimming network structure and model size so spatial-temporal AIoT data can be analyzed more efficiently.
Sanjay Purushotham
University of Maryland-Baltimore County, MD·Computer Science·Early-career
Develop multi-view federated learning algorithms that train shared models across hospitals holding different kinds of patient data, without moving the data between sites.
Joseph Camp
Southern Methodist University, TX·Computer Science
Develop an AI curriculum covering federated learning, explainable AI, homomorphic encryption and neural networks, taught through lectures, exercises and simulations and shared via the NAIRR Pilot.
The school
University of Florida: 65 labs on PI Pages, $1.3B a year on research, 795 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.