Ph.D. Candidate · Industrial & Systems Engineering · KAIST
Hi! I am a Ph.D. candidate at KAIST, advised by Professor Heeyoung Kim at the Industrial Statistics Lab. My research focuses on efficient uncertainty quantification (UQ) for modern deep learning — designing UQ methods that are scalable, practical, and reliable under distribution shift, class imbalance, and noisy labels.
I'm especially interested in single-pass UQ methods such as evidential deep learning, efficient Bayesian deep learning, and conformal prediction. More broadly, I'm drawn to robust AI systems — including time-series anomaly detection and learning under incomplete or imbalanced data. My recent work includes F‑EDL (NeurIPS 2025), DAEDL (ICML 2024), and the upcoming Courtroom Analogy framework (ICML 2026).
Previously, I was a research intern at Samsung Advanced Institute of Technology and SK Hynix, where I worked on uncertainty-based anomaly detection and quality testing for semiconductor manufacturing.
Currently open to research collaborations and post-doc / industry conversations.
new Our paper Courtroom Analogy is accepted at ICML 2026 🎉
Presented F‑EDL at NeurIPS 2025 in San Diego.
Gave an invited talk on F‑EDL at the INFORMS Annual Meeting.
Invited seminar on uncertainty quantification methods for deep learning at POSCO Labs.
Received Best Poster Award (3rd Prize) at Samsung AI Forum 2024.
Presented DAEDL at ICML 2024 in Vienna.
Full list on Google Scholar. * equal contribution · bold denotes me.
In International Conference on Machine Learning (ICML), 2026
A new framework for uncertainty-aware classification that reframes the prediction problem through a courtroom metaphor — separating evidence collection, deliberation, and verdict to produce more interpretable and well-calibrated predictions.
In Advances in Neural Information Processing Systems (NeurIPS), 2025
Extends evidential deep learning by predicting a flexible Dirichlet distribution over class probabilities, improving uncertainty generalization under classical, long-tailed, and noisy in-distribution settings.
In International Conference on Machine Learning (ICML), 2024
Improves evidential uncertainty estimation by incorporating feature-space density information at prediction time, alongside a new parameterization designed to improve classification and out-of-distribution detection.
To be submitted to NeurIPS 2026
Under review at KDD 2026
Major revision at IEEE Transactions on Automation Science and Engineering
Single-pass UQ methods, evidential deep learning, efficient Bayesian deep learning, conformal prediction, and calibration of modern neural networks.
Learning under distribution shift, class imbalance, noisy labels, semi-supervised learning, and out-of-distribution detection.
Multivariate time-series anomaly detection, graph structure learning, probabilistic state-space models, and industrial process monitoring.
NeurIPS 2025 (Poster) · KIIE Annual Conference · INFORMS Annual Meeting (Invited Session)
Invited Seminar — POSCO Labs
ICML 2024 (Poster) · Samsung AI Forum (Invited Poster) · Qualcomm AI Fellowship Korea (Finalist Session) · KIIE Spring Conference
A condensed version. Download full CV (PDF) →