Ph.D. Candidate, Computational Intelligence Lab, Ajou UniversityAdvised by Prof. Jongbin Ryu
I make large neural networks efficient enough to run under tight compute limits, on edge devices and inside robot control loops.
Two papers accepted to EMNLP 2026: Train Overcomplete, Deploy Compact, and Layer-wise Curriculum Learning.
Paper accepted to BMVC 2026: Re-calibrated Contrastive Loss for Transformation-Aware Prompt Conditioning.
Selected for the Doctoral Research Grant, a two-year award from NRF Korea.
First place, IEEE Low-Power Computer Vision Challenge Track 1 at the CVPR 2025 workshop, deployed on Snapdragon via Qualcomm AI Hub.
Paper accepted to ACCV 2024: Neural Substitution for Branch-Level Re-parameterization.
Filed a PCT patent on a neural-network CT triage classifier.
Paper on the origins of CNN and ViT architectures published in Scientific Reports.
Paper on CT-based small bowel obstruction triage published in the International Journal of Surgery.
I want VLA policies to hold up inside real robot control loops, which means three things at once: training that transfers across embodiments, inference cheap enough to run every control step, and failures caught early enough for the policy to recover.
Structured pruning leaves a model that has to relearn what was removed. I work on giving that recovery enough capacity, and on scheduling compression so the error that accumulates in deeper layers is paid down first (EMNLP 2026).
Making vision models cheap without giving up accuracy: re-parameterization at branch level (ACCV 2024), transfer that keeps a vision-language model general (BMVC 2026), and adaptation at test time under distribution shift. The same drive won Track 1 of the 2025 IEEE Low-Power Computer Vision Challenge.
After pruning, the recovery module is too small to relearn what was removed. Training it overcomplete, then merging it back, gains up to 8.4 points at no deployment cost.
Empirical Methods in Natural Language Processing, 2026papercode
Fine-tuning a vision-language model compromises its generalization. Transformation-aware prompt conditioning, under a loss re-calibrated for multiple positives, retains both.
A neural-network classifier that reads abdominal CT to predict high-risk small bowel obstruction, filed as a PCT patent from the clinical study below.
PCT Patent Applicationfiled 2024
Judging who needs surgery for small bowel obstruction is hard from CT alone. A dual-branch 3D CNN flags high-risk patients at 0.90 AUROC on a multi-center cohort.
Real-time detection transformer for autonomous driving
Source-free domain adaptation for on-device face recognition
Masked-face recognition with domain adaptation
Lymphoma recognition and survival analysis from 3D CT
NRF KoreaSep 2025 - Aug 2027
Progressive Linear Activation for Depth-wise Reparameterization of Large-scale Models
NIA KoreaOct 2022 - Dec 2023
Track 1 of 45 teams, at the CVPR workshop2025
Layer Imitation: Efficient Vision-Language Model on the Mobile NPU
Google2023
Smiler: An Expression-Training App for Children with ASD
ACK, KOSIA2021
Recommender System for Web Search Based on NLP to Improve User Search Environment
Korea Digital Contents Society2021
Lightweight Video Search Using Video-to-Text Generation and Retrieval
Changing what the policy is trained to predict, so skills carry across robots instead of binding to one.
Cutting the compute spent on every control step while keeping task success.
Noticing that an action is going wrong early enough for the policy to recover.
Adapting under distribution shift at inference time, cheaply and without collapse.
Letting regularization respond to the data instead of applying one fixed strength.
I built and operated our lab's GPU cluster, which consists of about 20 nodes with eight RTX 3090s each. This work included containerized provisioning, per-node CUDA isolation, and training jobs that ran unattended for days. Outside of research, I taught AI tutorials for undergraduate research assistants and helped launch the Google Developer Student Club chapter at Ajou University.