Research

Discretized Diffusion Dynamics

My research develops non-asymptotic theory to characterize the effects of noise in stochastic systems and guide algorithm design, drawing on applied probability and control. Treating algorithms as discretized diffusions, I study the effects of algorithmic parameters that are obscured in the asymptotic limit. I apply this framework to understand and improve diffusion models for generative AI, stochastic gradient descent, and reinforcement learning.

Generative AI

Diffusion visualization

Diffusion models generate realistic data using a denoising process. The rate at which noise is injected to or removed from data determines the quality of generated data, and its effect is captured by non-asymptotic theory.

  1. Preprint2026
  2. Preprint2025

Theoretical Foundations: Stochastic Approximation

  1. Preprint2026
  2. Revised and resubmitted to IEEE TAC2025

  3. When to Use Polyak-Ruppert Averaging in Stochastic Approximation

    S. T. Kong, R. Srikant

    Work in progress2026

Crowdsourcing, Reliability, and Active Learning

  1. TMLR2025
  2. AAAI2023

  3. A Neural Pre-Conditioning Active Learning Algorithm to Reduce Label Complexity

    S. T. Kong, S. Jeon, D. Na, J. Lee, H. S. Lee, K. H. Jung

    NeurIPS2022

Machine Learning for Clinical Decision-Making

Chest radiograph with a radiologist-annotated pulmonary nodule in the left upper lung.
model confidence0.92
Radiology Report
Findings
  • Solitary subcentimeter nodule in the left upper lung zone. AI: 0.92
  • No prior imaging available for comparison.
Impression
  • Indeterminate subcentimeter left upper zone nodule.
  • Recommend non-contrast chest CT for further characterization.
Clinical decision-making: a model flags a nodule, and a radiologist confirms and issues the report and recommendation.
  1. Translational Vision Science & Technology2024

  2. Self-Accumulative Vision Transformer for Bone Age Assessment Using the Sauvegrain Method

    H. J. Choi, D. Na, K. Cho, B. Bae, S. T. Kong, H. Ahn, S. Choi, J. Kim

    ECCV Workshop2024

  3. Volume Is All You Need: Improving Multi-Task Multiple Instance Learning for WMH Segmentation and Severity Estimation

    W. Jung, C. H. Suh, W. H. Shim, J. Kim, D. Lee, C. Park, S. T. Kong, K. H. Jung, H. Heo, S. J. Kim

    MICCAI Workshop (ML in Clinical Neuroimaging)2022

  4. Augmenting Magnetic Resonance Imaging with Tabular Features for Enhanced and Interpretable Medial Temporal Lobe Atrophy Prediction

    D. Lee, C. H. Suh, J. Kim, W. Jung, C. Park, K. H. Jung, S. T. Kong, W. H. Shim, H. Heo, S. J. Kim

    MICCAI Workshop (ML in Clinical Neuroimaging)2022
  5. MICCAI Workshop (Limited & Noisy Data)2022
  6. Journal of Digital Imaging2022
  7. IEEE EMBC2021

  8. Aggregation of Cohorts for Histopathological Diagnosis with Deep Morphological Analysis

    J. Park, Y. R. Chung, S. T. Kong, Y. W. Kim, H. Park, K. Kim, D. I. Kim, K. H. Jung

    Scientific Reports2021
  9. Applied Sciences2021
  10. MICCAI2020

* Equal contribution.