Juyong Lee

I am a master student in the Cognitive Learning for Vision and Robotics (CLVR) Lab at KAIST, advised by Joseph J. Lim. I received B.S. degree by double majoring both Computer Science and Engineering & Mathematics at POSTECH, where I could fortunately work with Dongwoo Kim, Jaesik Park, and Wankyun Chung. The picture is me at the Grand Canyon during exchange student at Stanford!

CV  /  Google Scholar  /  Email

profile photo

I'm interested in understanding the fundamental principles of intelligence and building human-like learning machines. My current research interests lie in helping AI agents to embrace physical reasoning ability and to invent new tools.

Hyperbolic VAE via Latent Gaussian Distributions
Seunghyuk Cho, Juyong Lee, Dongwoo Kim
ICML 2023 workshop TAGML (Extended work of this)

Verification on the efficacy of employing Gaussian manifold on density estimation over images with implicit hierarchical structure. I, especially, focused on applying Gaussian manifold as the hyperbolic latent geometry for the world model in RL.

GM-VAE: Representation Learning with VAE on Gaussian Manifold
Seunghyuk Cho, Juyong Lee, Dongwoo Kim
KAIA 2022 (Award Of Excellence)

Proposing a new distribution (namely, PGM normal distribution) over the set of the diagonal Gaussian distributions with the Fisher information metric (namely, Gaussian manifold). We propose a variational auto-encoder whose latent space is the Gaussian manifold, showing higher numerical stability compared to the canonical hyperbolic distributions.

A Rotated Hyperbolic Wrapped Normal Distribution for Hierarchical Representation Learning
Seunghyuk Cho, Juyong Lee, Jaesik Park, Dongwoo Kim
NeurIPS 2022
paper / code

A simple yet effective alteration of a hyperbolic wrapped normal distribution (HWN), to better utilize the geometric properties of the diagonal HWN. We analyze the geometric properties of the diagonal HWN and alleviate the limitation by introducing the rotated hyperbolic wrapped normal distribution (RoWN).

Style-Agnostic Reinforcement Learning
Juyong Lee, Seokjun Ahn, Jaesik Park
ECCV 2022
paper / code

Style-agnostic reinforcement learning algorithm, with both neural style transfer and adversarial learning. The style, here, refers to task-irrelevant details in the image, such as the background color. Agents learn style-agnostic representation by generating adversarial perturbation in the latent space and to achieve generalized performances.

A 3D cell printed muscle construct with tissue-derived bioink for the treatment of volumetric muscle loss
Yeong-Jin Choi, Young-Joon Jun, DongYeon Kim, Hee-Gyeong Yi, Su-Hun Chae, Junsu Kang, Juyong Lee, Ge Gao, Jeong-Sik Kong, Jinah Jang, Wan Kyun Chung, Jong-Won Rhie, Dong-Woo Cho
Biomaterials 2019

A novel treatment for volumetric muscle loss with decellularized extraceullar matrix bioink using 3D cell printing technology. I, especially, participated in situ functional analysis of the 3D printed muscle cell, to electrically stimulate the peroneal nerve of mice legs.

The source code is from here