Assistant Professor · CSIE, National Taiwan University
Shih-Hsin Wang 王士欣
I develop machine learning methods that are both theoretically grounded and practical, spanning geometric deep learning, generative models, and AI for Science. I care about turning intuition into principled, reliable, and efficient tools — and applying them to problems in molecules, biology, and other sciences.
Recent
Experience →Jul 2026
I was named a Yushan Young Fellow by Taiwan's Ministry of Education (Yushan Fellow Program).
Jul 2026
I am organizing two minisymposia on generative models, transport, and inverse problems at SIAM IS26 & MDS26 (Salt Lake City, Nov 2026). [Details]
Jul 2026
I am actively recruiting undergraduate project students and MS/PhD students for my new lab at NTU CSIE. See Join the Lab for openings and how to apply.
Jul 2026
I will join the Department of Computer Science and Information Engineering at National Taiwan University as an Assistant Professor in August 2026.
Notes
All notes →Research Areas
Flow-Based Generative Models
From Noise to Data
Overview: Data, Distributions, and Particles 中文 Not in English yet — read it in 中文. What Generative Models Actually Learn 中文 Not in English yet — read it in 中文. Samples as Particles: The Dynamical View 中文 Not in English yet — read it in 中文. Vector Fields: Where Should Each Point Go? 中文 Not in English yet — read it in 中文. Probability Paths: A Road Between Distributions 中文 Not in English yet — read it in 中文. Denoising: Generation as Local Repair 中文 Not in English yet — read it in 中文. The Score Function: A Local Sense of Direction 中文 Not in English yet — read it in 中文. Velocity as a Regression Target: What Flow Matching Learns 中文 Not in English yet — read it in 中文. Sampling as Solving an ODE: Discretization, Steps, and NFE 中文 Not in English yet — read it in 中文. Three Languages: Noise, Score, Velocity 中文 Not in English yet — read it in 中文. Review Quiz: Ten Questions to Walk the Whole Map Again 中文 Not in English yet — read it in 中文.
Geometric Deep Learning
Invariance and Equivariance
The Same Taipei Map, Seen Through Two Phone Views 中文 Not in English yet — read it in 中文. Why CNNs Naturally Fit Maps and Images 中文 Not in English yet — read it in 中文. What If a Map Is a Set of Points? 中文 Not in English yet — read it in 中文. A Map Is Also a Graph 中文 Not in English yet — read it in 中文. When the Phone Rotates, Should the Model Rotate Too? 中文 Not in English yet — read it in 中文. Point Clouds Need Two Kinds of Symmetry 中文 Not in English yet — read it in 中文. Symmetry Is Not a Free Lunch 中文 Not in English yet — read it in 中文.
Academic Skills
Paper Writing
Research
- Geometric Deep Learning Equivariant graph neural networks and expressive geometric graph representations.
- Generative Models Diffusion models, flow matching, single-step models, and test-time guidance.
- AI for Science Modeling complex structures like molecules and proteins.
Selected publications
All publications →ICML 2026 GenBio [Spotlight]
ICML 2026