Posts by Collection

portfolio

publications

Clinical Features and Subgroup Patterns in Elderly and Super-Elderly TMD Patients

Yeon-Hee Lee, et al. (including Hee-Sung Kim)

Scientific Reports, 2025

Analysis of clinical features and subgroup patterns of temporomandibular disorders in elderly populations.

Deep Learning-Based Diagnosis of Temporomandibular Joint Osteoarthritis Using Whole-Body Bone Scans

Yeon-Hee Lee, Hee-Sung Kim, Seonggwang Jeon, Q-Schick Auh, Il Ki Hong, Sunju Choi, Fernando Guastaldi, Hyungsoon Im, Yung-Kyun Noh, Akhilanand Chaurasia

iScience, 2025

Deep learning models (CNN, VGG16-based) for automatic diagnosis of TMJ osteoarthritis from bone scintigraphy, achieving AUC > 0.90 on head-and-neck imaging.

Download Paper

Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data

Hee-Sung Kim, Hyeonseong Kim, Sungyoon Lee

ICML, 2026

We recover SAM’s flat-minima bias from unlabeled data alone by regularizing local inconsistency, enabling flat-minima methods for semi- and self-supervised learning.

Download Paper | arXiv | Poster | Slides |

Gradient Descent with Large Step Size Restores Symmetry in Deep Linear Networks with Multi-Pathway

Hee-Sung Kim, Sungyoon Lee

ICML, 2026

Large-step gradient descent at the Edge of Stability overrides the winner-takes-all architectural bias of multi-pathway deep linear networks, restoring shared representations across parallel pathways.

Download Paper | arXiv | Poster | Slides |

A Simple and Efficient Measure of Loss Landscape Curvature

Hee-Sung Kim, Sungyoon Lee

ICML Workshop on High-dimensional Learning Dynamics (HiLD), 2026

A scalable directional-curvature measure and forward-pass estimators for tracking Edge-of-Stability dynamics without Hessian-vector products.

Download Paper | Poster |

Hee-Sung Kim

Posts by Collection

portfolio

publications

Clinical Features and Subgroup Patterns in Elderly and Super-Elderly TMD Patients

Deep Learning-Based Diagnosis of Temporomandibular Joint Osteoarthritis Using Whole-Body Bone Scans

Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data

Gradient Descent with Large Step Size Restores Symmetry in Deep Linear Networks with Multi-Pathway

A Simple and Efficient Measure of Loss Landscape Curvature

talks

Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data

Gradient Descent with Large Step Size Restores Symmetry in Deep Linear Networks with Multi-Pathway

A Simple and Efficient Measure of Loss Landscape Curvature

teaching