Research
Discretized Diffusion Dynamics
From stochastic optimization to generative AI — a unifying, non-asymptotic theory of the stochastic systems behind modern machine learning.
Diffusion processes offer a powerful mathematical framework for understanding and improving machine learning algorithms. My work develops a non-asymptotic theory of discretized diffusion dynamics and applies it across three domains — stochastic gradient descent, reinforcement learning, and generative AI. By characterizing finite-time behavior precisely rather than only in the limit, this theory turns asymptotic intuition into actionable guidance for algorithm design, uncertainty quantification, and decision-making in stochastic systems.
Sampling and Noise Schedules for Diffusion Models
Generative diffusion models sample from complex distributions by discretizing a stochastic process, and the choice of noise schedule governs sample quality. Existing guarantees of order-optimal sampling error held only for specific piecewise-constant schedules, leaving the schedules used in practice unexplained. I establish a bound on the exponential integrator’s discretization error that is amenable to analytic optimization; minimizing it via the calculus of variations shows that a broad class of schedules attains order-optimal sampling, and yields closed-form optimal schedules under mild regularity conditions.
Seo Taek Kong, Weina Wang, R. Srikant
Optimizes the discretization error of the exponential integrator to derive order-optimal, closed-form noise schedules for diffusion sampling.
Finite-Time Analysis of Stochastic Approximation
Stochastic approximation (SA) underlies both stochastic gradient descent and reinforcement learning.
I derive rigorous uncertainty estimates of the error incurred by SA algorithms by establishing non-asymptotic central limit theorems.
By doing so, I am able to unify various analytical insights including central limit theorems, moment bounds, and concentration inequalities, which had been developed separately.
Seo Taek Kong, R. Srikant
Derives finite-time Wasserstein-p convergence rates for nonlinear SA under Markovian, martingale, and i.i.d. noise.
Seo Taek Kong, Sihan Zeng, Thinh T. Doan, R. Srikant
Establishes a non-asymptotic central limit theorem for Polyak–Ruppert averaging in linear two-time-scale stochastic approximation, with insights into the optimal time-scale separation.
Human-in-the-Loop ML
Applied work on making the most of human labeling effort, along two complementary directions.
Active learning selects the most informative data to label, reducing the annotation needed to train accurate models.
Label aggregation produces accurate labels from a collection of noisy labels by estimating the reliability of labelers.
Saptarshi Mandal*, Seo Taek Kong*, Dimitrios Katselis, R. Srikant
Seo Taek Kong, Soomin Jeon, Dongbin Na, Jaewon Lee, Hong-Seok Lee, Kyu-Hwan Jung
Introduces an active learning algorithm that selects the most informative samples to label, reducing annotation complexity.