Yeachan Park

Postdoctoral research fellow

Summary

I am a postdoctoral fellow at KIAS. I received my Ph.D. from Seoul National University (SNU) under the supervision of Prof. Myungjoo Kang.

I am interested in theoretical machine learning, with a focus on the expressive power of neural networks, the geometry of loss surfaces, and training dynamics. My research interests also extend to practical applications, including low-level vision models, privacy-preserving models, and math-solving models.

Research Interests

Machine Learning Theory

Expressive power of neural networks (universal approximation)
Geometry of Loss surface
Training Dynamics (Convergence, Grokking, Integrability, etc. )
Computer Arithmetic (Floating-point arithmetic, Fixed-point arithmetic)
Knowledge Distillation

Low-level Vision

Single Image Deraining

Priviate AI

Membership Inference Attacks
Differential Privacy

AI for Math

Autoformalization
Automated theorem proving
Mathematical Language Model

Differential Algebra

Differential Galois Theory
Morales–Ramis Theory (Integrability theory of dynamical systems)

Skills

Programming Languages: Python, C++, Matlab, Lean
Machine Learning Frameworks: Pytorch, Tensorflow, Chainer
Web Frameworks: FastAPI, React

Publications (* Equal contribution)

Expressive Power of ReLU and Step Networks under Floating-Point Operations, Neural Networks, Jul. 2024. Paper

- Yeachan Park, Geonho Hwang, Wonyeol Lee, Sejun Park

Self-knowledge distillation via dropout, Computer Vision and Image Understanding, 2023. Paper

- Hyoje Lee, Yeachan Park, Hyun Seo*, Myungjoo Kang

MCW-Net: Single image deraining with multi-level connections and wide regional non-local blocks, Signal Processing: Image Communication, 2022. Paper, Code

- Yeachan Park, Myeongho Jeon, Junho Lee*, Myungjoo Kang

Preprint

On Expressive Power of Quantized Neural Networks under Fixed-Point Arithmetic, submitted

- Geonho Hwang, Yeachan Park , Sejun Park

Acceleration of Grokking in Learning Arithmetic Operations via Kolmogorov-Arnold Representation, Paper, submitted

- Yeachan Park* , Minseok Kim*, Yeoneung Kim

Absence of Closed-Form Descriptions for Gradient Flow in Two-Layer Narrow Networks, Paper, submitted

- Yeachan Park

Membership inference attacks against object detection models, Paper

- Yeachan Park, Myungjoo Kang

Job Experience

Korea Institute for Advanced Study (KIAS)

- AI Research Fellow | Sep. 2022 - present

Projects

- Semiconductor Micro-defect Pattern Detection (SK Hynix) | Mar. 2017 – Dec. 2017.

- Semiconductor Micro-defect Type Pattern Classification (SK Hynix) | Jan. 2018 – Dec. 2018.

- Multi-lingual Scene Text Detection and Recognition (NAVER) | Feb. 2019 – Sep. 2019.

- Construction Site Disaster Prevention System (KCC Engineering & Construction) | Mar. 2019 – Feb. 2020.

- AI-based Emergency Medical System Development (Yonsei University Health System) | Mar. 2019 – Dec. 2021.

- Wafer Yield Prediction and Privately Safe Model Research (Samsung Electronics) | Mar. 2019 – Aug. 2022.

- CCTV Visitor Analysis Model (AI Nation) | Sep. 2021 – Aug. 2022.

Experience

- Data Analysis Test Questions Creation (LOTTE HRD Academy) | Mar. 2021

- HANA Card AI Hackerton Questions Creation (HANA Card) | Jun. 2021

- Practice Coach for AI Lecture (Elice) | Oct. 2021

- Play Carnival Startup Mentoring (Campus Town, Play Carnival) | Sep. 2021- Dec. 2021