Learning without gradients
Derivative-free and forward-only optimization for models that are black boxes or too costly to backpropagate through.
Researcher · Optimization for Efficiency & Robustness
Seonghwan Park 박성환
I'm a researcher at the Korea Electronics Technology Institute (KETI), working with Dr. Min-Gyu Park on 3D vision and physical AI. I received my M.S. in Artificial Intelligence from POSTECH under Prof. Namhoon Lee, where I studied optimization for deep learning.
I'm broadly interested in efficiency and robustness in vision and machine learning, which I approach through the lens of optimization. I now carry this lens into 3D vision and physical AI.
01 News
Our paper on recovering black-box 3D models from pose shifts was accepted to WACV 2027 (Round 1).
Our papers on efficient OOD detection and monocular avatar reconstruction were accepted to ECCV 2026.
Joined the Korea Electronics Technology Institute (KETI) as a researcher, working on 3D vision and physical AI under Dr. Min-Gyu Park.
Our paper on robust concept bottleneck models was accepted to NeurIPS 2025.
Completed my M.S. in AI at POSTECH, advised by Prof. Namhoon Lee.
Our paper on black-box prompt tuning was accepted to ICLR 2025.
Started a project with Samsung Advanced Institute of Technology (SAIT) on visual prompting for MMFM-based VQA.
Our paper on backprop-free federated learning was presented at the ICML 2023 Federated Learning Workshop.
Started my M.S. in AI at POSTECH, joining the Computational Optimization Lab under Prof. Namhoon Lee.
02 Research
Flatness & robustness
Loss-landscape flatness as a lever for robustness to noisy labels and for efficient out-of-distribution detection.
3D vision & physical AI
Reconstructing humans and their interactions with objects from monocular input, toward robotics grounded in vision priors.
03 Publications
Year
Topic
-
Recovering Black-Box Point Cloud Models from Pose Shifts via Derivative-Free Optimization
WACV 2027Accepted as a Round 1 paper
-
-
-
-
-
FedFwd: Federated Learning without Backpropagation
ICML 2023 WorkshopFederated Learning and Analytics in Practice
-
On the Effectiveness of Sharpness-Aware Minimization with Large Mini-batches
ICML 2023 WorkshopHigh-dimensional Learning Dynamics
No publications match this filter.
04 Experience
Jan 2026 — present Now
Researcher · KETI
3D vision & physical AI.
Supervisor: Dr. Min-Gyu Park.
Jan 2023 — Dec 2025
Research Assistant · Computational Optimization Lab, POSTECH
Model efficiency & robustness through optimization.
Supervisor: Prof. Namhoon Lee.
Jun 2025 — Nov 2025
Samsung Research Project
Multi-modal foundation models for anomaly detection.
Jun 2024 — May 2025
Samsung Advanced Institute of Technology (SAIT) Project
Visual prompting for MMFM-based visual question answering.
Jan 2022 — Dec 2022
Undergraduate Research Intern · ML Lab, Hanyang Univ.
Prompt tuning for LLM-based VQA.
Supervisor: Prof. Eun-Sol Kim.
05 Misc
Awards
- 2025 · NeurIPS Travel Grant
- 2025 · CKAIA Best Paper Finalist
- 2022 · HYU Academic Excellence Award · top 3%
Services
- Reviewer · ICLR 2026, ECCV 2026, NeurIPS 2026, WACV 2027, TMLR
- Military Service · Republic of Korea Air Force (Oct 2018 — Aug 2020)