University of Florida — Graduate Research Assistant, NCR Lab
Advisor: Prof. Warren E. Dixon · Coadvisor: Prof. Dan P. Guralnik
Common thread: view learning, generation, and inference algorithms as dynamical systems, and use tools from nonlinear control, Lyapunov theory, and Bayesian estimation to understand and redesign them.
- Optimization as feedback control. Formulated optimization as a closed-loop controlled dynamical system on a Riemannian manifold coupling model parameters with internal optimizer states; a Lyapunov construction links objective reduction to regulation of the optimizer's internal dynamics, giving a common description of SGD, Adam, and Lion and a blueprint for designing optimizers around explicit stability objectives (NeurIPS 2026).
- Generation as controlled transport. Developed graph structured flow matching that transports the entire future skeleton trajectory through one conditional velocity field while preserving bone lengths by construction; a compact 4.47M-parameter model reaches near-best AMASS precision and transfers zero-shot to the unseen Human3.6M skeleton (preprint). Cast energy-based sampling as a control problem with a Kullback–Leibler Lyapunov function, yielding a finite-step stopping rule for Langevin sampling (preprint).
- Belief dynamics in inference and privacy. Built a Rao-Blackwellized particle filter that infers an agent's goal online from noisy trajectories, and control policies that regulate the observer's belief to bound goal-information disclosure via probabilistic control barrier functions (IFAC 2026; IEEE CDC 2026).
- Allocating model capacity and optimization effort. Given any importance-score vector, set an adaptive sparsity level via its effective sample size, with a provable loss-change bound; the method works with any importance criterion and architecture, validated on MLPs, CNNs, Transformers, LLaMA (7B–70B), and KANs, and even beyond networks on image pixels (ICML 2026 Spotlight). Reweighted each reward's advantage by batch-level saturation so multi-objective policy optimization stops spending effort on saturated objectives: +3.8 pts average over GDPO on five math benchmarks (preprint).