Hyunsoo Lee
I am an undergraduate student in Electrical and Computer Engineering, double majoring in Mathematical Sciences at Seoul National University (SNU). I am currently a research intern at the SNU Machine Perception and Reasoning Lab, advised by Prof. Jonghyun Choi. Before that, I was a visiting student and research intern at UC Irvine, where I worked with Prof. Stephan Mandt, and I also conducted research at SNU with Prof. Young Min Kim and Prof. Bohyung Han.
Research Vision
- I study how intelligent systems form, control, and adapt representations across structured inputs.
- My current research moves from controllable generation toward neuroscience-inspired learning mechanisms, especially memory, generalization, and interpretability.
- My long-term goal is to build interpretable intelligence that can support critical real-world problems.
My work has focused on generative visual computing, where images, point clouds, human motion, and 3D geometry are treated as structured visual signals that can be controlled and extended across tasks. This research has led to publications in leading computer vision and machine learning venues, including NeurIPS, CVPR, ECCV, ICML, and WACV.
My current direction moves from controlling generation toward understanding learning itself. I am developing a new research agenda in neuroscience-inspired AI, using ideas from memory operations and brain-inspired learning systems to study how artificial neural networks learn and generalize. My aim is to develop learning mechanisms that are not only stronger, but also more interpretable.
The long-term question I care about is how to build intelligence that can solve problems beyond current human capability. I see generative modeling as a lens for studying how systems form and manipulate representations, and I now aim to connect this perspective with neuroscience-motivated mechanisms and continual learning. Ultimately, I want to build AI systems whose mechanisms can be explained, scaled, and applied to meaningful real-world domains such as education, scientific and engineering design, and environmental modeling.
Publications
* denotes equal contribution. † denotes corresponding author.
-
Point2Pose: A Generative Framework for 3D Human Pose Estimation with Multi-View Point Cloud DatasetIn WACV, 2026Frames 3D human pose estimation from raw point clouds as conditional generation and builds a pose framework using diffusion and flow matching.
Education
- Compulsory military service included (Aug 2023 - Feb 2025).
- Early visual understanding projects formed the origin of my research in computer vision.
Research Experience
- Advisor: Jonghyun Choi
- Conducting research on neuroscience-motivated memory mechanisms for representation learning.
- Advisor: Stephan Mandt
- Conducted research on guidance methods for diffusion models in inverse problems and collaborative generation.
- Advisor: Young Min Kim
- Conducted research on geometry deformation and versatile content generation with diffusion models.
- Advisor: Bohyung Han
- Conducted research on training-free, diffusion-based image-to-image translation.