Minhyuk An

I'm currently a master's student in the department of Artificial Intelligence at Yonsei University. I am working at the Soft Computing Laboratory (SCLab) and am fortunate to be advised by Sung-Bae Cho.
Previously, I graduated from Inha University with B.S. in the department of Information and Communication Engineering and was advised by Sungeun Hong and Hakil Kim.

Email: als7928@yonsei.ac.kr
LinkedIn / Github

Education

Yonsei University, Seoul, South Korea (2023.09 — 2025.08)
Master's degree in Artificial Intelligence
GPA: 4.12 / 4.50
Coursework: Advanced Video Processing / Lightweight Networks / Multicore Programming / Multimodal Deep Learning / Information Theory / Advanced Data Mining / Graph and Network Analysis

Inha University, Incheon, South Korea (2018.03 — 2023.08)
Bachelor of Science in Information and Communication Engineering
GPA: 3.89 / 4.50
Coursework: Computer Vision / Digital Image Processing / Digital Signal Processing / Linear Algebra / Probability and Statistics / Object-Oriented Programming / Data Structures / Algorithms / System Programming / Operating Systems / Database Systems / Computer Networks

Technical Skills

Programming Languages: Python / MATLAB / C / C++

Libraries and Tools: Pytorch / NumPy / OpenCV / Albumentations / Pillow / PyG / NetworkX / Transformers / SciPy / Scikit-learn / Pandas / cuDF / Matplotlib / Seaborn / WandB / Tensorboard / TensorRT / Docker / Conda / Git / Overleaf

Operating Systems: Linux

Research

I have a keen interest in AI/ML as a whole with data of any kind—images, text, audio, graphs, and more! Particularly, I am deeply fascinated by gaining a deeper understanding of data for real-world applications.

Prompt Optimization for Large Language Models with Generalized Feedback via Semantic Centrality
Minhyuk An (Leader), Tae-Hoon Kang, Haerin Byeon, Sung-Bae Cho
Preprint, ACL 2025 (under review)
Paper / Code

Applied prompt optimization using semantic centrality-based feedback

Achieved significant improvements of LLM responses by 1.88%p over baselines using my proposed algorithm

AI Agents / Large Language Models / Prompt Optimization / Black-box Optimization / Dialogue & Interactive Systems / Monte Carlo Tree Search

Local-Global Blending Graph Neural ODE Network for Graph Classification
Minhyuk An, Sung-Bae Cho
Preprint, IJCAI 2025 (under review)
Paper / Code

Designed a Graph Neural ODE that combines local-global representations dynamically

Achieved significant improvements up to 3.8%p

Graph Neural Network / Representation Learning / Graph Mining / Neural Ordinary Differential Equations

Undergraduate Projects

	Samsung AI Challenge : Camera-Invariant Domain Adaptation Minhyuk An (Leader), Gihun Son Competition (ranked 30/211), (2023.08.21 — 2023.10.02) Code Background: Prediction performance catastrophically drops on highly distorted images from different domains Applied a semi-supervised image segmentation method utilizing data augmentation and pseudo-labeling Improved model performance with limited labeled data Ranked 30/211 Domain Adaptation / Domain Adaptive Semantic Segmentation / Semi-Supervised Learning / Pseudo-Labeling / Adaptive Cutmix
	Crowd Counting Using Diffusion-Based Latent Space Gihun Son, Minhyuk An (PM), Siwon Lee Capstone Project (Advisor: Sungeun Hong), (2023.03 — 2023.08) Paper / Poster / Code Background: Crowd counting often requires expensive data like depth or thermal images Developed a diffusion-based generative model for RGB-to-crowd density map translation Achieved comparable results with a single accessible modality by leveraging generative models Crowd Counting / Contiditional Image Generation / Diffusion Models
	Applying Bag of Tricks for Improving Accuracy Minhyuk An Personal Toy Project, (2022.09 — 2022.12) Report / Code Background: CNN performance improvements often require complex changes or large-scale data Applied simple techniques from the Bag of Tricks paper like label smoothing, mixup, and learning rate scheduling to boost ResNet accuracy Learned how to improve accuracy using simple techniques Knowledge Distillation / Learning Rate Scheduling / Label Smoothing / Mixup Augmentation / ResNet
	Study on Mitigating Class Imbalance Minhyuk An Personal Toy Project, (2022.09 — 2022.12) Report1 / Report2 Background: Class imbalance frequently degrades model performance in deep learning tasks Built an automated data synthesis system using easily available background images and foregrounds from minority classes to generate a balanced dataset Effectively trained deep learning models, proving valuable as training data using synthetic datasets Class Imbalance / Data Synthesis / Data Augmentation / Semantic Segmentation / Data Sampling
	2D/3D Object Detection for Autonomous Vehicles Minhyuk An and 13 members Competition, (2021.09 — 2022.11) Demo / Report / Note1 / Note2 Built a camera–LiDAR fusion object detection system using deep learning for real-world deployment Improved model inference speed by approximately 2x using TensorRT, enabling operation in a real-time environment Collected custom datasets and developed a detection system for traffic lights, cones, and signs 2D/3D Object Detection / Custom Datasets / YOLOv4 / Frustum PointNets / TensorRT / Sensor Fusion / Calibration / ROS
	Speech Classification for Gender and Speaker Identification Minhyuk An (Leader) and 2 members Group Project, (2021.11 — 2021.12) Report Built a classification system using fundamental frequency and K-means clustering This project sparked my interest in AI/ML Audio Analysis / Digital Signal Processing / K-means Clustering

Research Projects

지속 가능한 협업형 멀티 모달 평생 학습 프레임워크 개발 (2023.12 — 2024.12)
(Sustainable Collaborative Multi-Modal Lifelong Learning Framework Development)
Funded by Institute for Information & communications Technology Promotion (IITP) (정보통신기획평가원 정보통신방송연구개발사업)

그래프 데이터로부터 생활패턴 발견기술 연구 (2023.09 — 2023.11)
(Information Retrieval Based on GNN for Understanding the Living Patterns)
Report
Funded by Electronics and Telecomunications Research Institute (ETRI) (한국전자통신연구원 정보통신방송연구개발사업)

상용 자율주행차 주행 데이터 (2022.05)
(Training Data Construction Project: Commercial Autonomous Vehicle Driving Data)
Funded by National Information society Agency (NIA) (한국지능정보사회진흥원 인공지능학습용데이터구축사업)
— Developed a system that utilizes instance segmentation to understand the background scene and automatically constructs synthetic datasets by augmenting objects

Vertically Integrated Projects (다학년연구프로젝트) (2021.09 — 2022.12)
(Development of Precision Localization for Autonomous Vehicles and Collaborative Service Technology with Autonomous Drones)
Funded by Inha University
— Acquired expertise, insights into research, problem-solving and collaboration skills.

Patent

교사 모델을 이용한 지식 증류 방법 및 장치 (10-2024-0141487) (출원, 2024.10)
(Method and Apparatus for Knowledge Distillation Using Teacher Models)
Application

Achievements & Experience

Completion of Fundamentals of Accelerated Data Science (2025.03.26)
Deep Learning Institute, NVIDIA
Certificate
— Learned the fundamental concepts of data science and GPU-accelerated data preparation & feature extraction

Teaching Assistant (2024.09 — 2024.12)
CSI3103: Programming Language Structures (Sung-Bae Cho)
Yonsei University
— Taught functional programming and common LISP to 45 undergraduate students

Accreditation for Engineering Education (2018.03 — 2023.08)
공학교육인증 이수 (ABEEK)
Inha University

Completion of Perception Technologies for Autonomous Driving (2023.01.02 — 2023.01.27)
H-Mobility Class, Hyundai NGV
Certificate

Academic Excellence Scholarship (Fall 2022)
Inha University

Research Excellence Award (2022.12)
Inha University

Future Mobility Urban Challenge Award (2022.11)
Seoul National University
— Developed object detection systems (2D, 3D) for traffic signals, signs and cones

International Student Automotive Innovation Competition (2022.10)
Korea Auto-Vehicle Safty Association (한국자동차안전학회)
— Developed object detection systems (2D, 3D) for traffic signals, signs and cones

Research Excellence Award (2022.01)
Inha University

Undergraduate Research Student (2021.12 — 2022.05)
Computer Vision Laboratory, Inha University (advisor: Hakil Kim)
Note
— Developed a program using instance segmentation to analyze background scenes
— Automated synthetic dataset generation by augmenting foreground objects
— Developed an interactive foreground extraction tool based on the GrabCut algorithm
— Designed a user-friendly GUI with simple path setup and configuration management
— Recognized domain differences in synthetic data leading to performance drop; emphasized importance of domain adaptation

The Republic of Korea Army (2020.02 — 2021.08)
Discharged after mandatory military service (Sergeant)

Academic Excellence Scholarship (Spring 2018)
Inha University

The finest template from Jon Barron