Degree: Ph.D. Candidate
Department: Department of Computing Science, New Jersey Institute of Technology
Location: New Jersey, United States
Email: hy9@njit.edu

I am a Ph.D. candidate in Computer Science at the New Jersey Institute of Technology (NJIT), supervised by Prof. Przemyslaw Musialski.
My research focuses on 3D reconstruction, generative AI, and neural surface representation.
In particular, I develop implicit neural networks for 3D shape reconstruction and representation, bridging geometry processing and machine learning.

Before joining NJIT, I earned dual Bachelor’s degrees from Jilin University —
a B.Sc. in Information and Computational Science (College of Mathematics)
and a B.Eng. in Computer Application (College of Computer Science).

I am currently seeking job opportunities — please feel free to reach out!

Publications

[SIGGRAPH Asia 2025]

A Finite Difference Approximation of Second-Order Regularization for Neural-SDFs Developed a lightweight optimization framework replacing computationally expensive second-order derivatives with finite-difference operators, improving model efficiency and numerical stability.
Enabled large-scale neural modeling using only first-order gradients — an approach generalizable to AIGC and reinforcement learning systems requiring interpretable and efficient optimization.

[Pacific Graphics 2025]

FlatCAD: Fast Curvature Regularization of Neural SDFs for CAD Models
Proposed an off-diagonal loss design that stabilizes model training and doubles convergence speed.
Demonstrated the framework’s adaptability across domains — from 3D surface learning to general deep network regularization and AI optimization under limited compute budgets.

[ISVC 2025]

Scheduling the Off-Diagonal Weingarten Loss of Neural SDFs for CAD Models
Designed an adaptive scheduling strategy that dynamically adjusts regularization weights during training to maintain convergence stability.
Improved overall model accuracy by 35%, providing a general mechanism for balanced loss design in large-scale AI and multi-agent learning scenarios.

[ICMLA 2024]

Shrinking: Reconstruction of Parameterized Surfaces from Signed Distance Fields
We propose a method for reconstructing explicit, parameterized surfaces from neural SDFs by iteratively contracting a parameterized sphere to match the target shape. Unlike Marching Cubes, our approach preserves continuity, differentiability, and surface parameterization, enabling downstream applications such as texture mapping, geometry processing, animation, and simulation.