Ph.D. student, Data Science · New York University
New York, NY
Hi, I'm Shiyang, and I also go by Lilian. I build generative models that find the quiet, low-dimensional structure underneath neural and behavioral recordings, and try to make that structure something we can actually interpret.
I'm a Ph.D. student at NYU's Center for Data Science, advised by John Pearson. I started my Ph.D. in Electrical and Computer Engineering at Duke in 2024 and transferred to NYU with my advisor in 2026. Right now I'm studying how mouse motor cortex controls reaching. Before that I worked on online models for streaming calcium imaging and on flow-based models of latent dynamics.
Before all this, I studied mathematics at Xi'an Jiaotong-Liverpool University and the University of Liverpool.
Things I think about most:
A running record of what I'm working on with the Pearson Lab, and what came before.
Now
Latent variable models of how mouse motor cortex controls reaching. I'm building a generative model that explains trial-to-trial variability through latent control signals, pairing a GP-VAE over the controls with flow matching to capture nonlinear, multimodal reach dynamics.
Earlier in the project, I used GLMs and Poisson linear dynamical systems to model how motor cortical output integrates distorted proprioceptive feedback.
A state-space model, Glow-SPLAT, that finds low-dimensional population dynamics in streaming, large-scale calcium imaging data. It's fitted with a particle-filter EM strategy and written in JAX, and it keeps pace with thousands of neurons faster than the microscope records them.
Earlier
A dual flow-matching model that learns identifiable, low-dimensional latent dynamics from neural and behavioral data. One flow field models how the system evolves over time; the other maps the data into a low-dimensional latent space and is trained to leave those dynamics undistorted, with a nested-dropout constraint keeping the latent space small. The result is more interpretable dynamics and better reconstructions than competing models, even on noisy datasets where conventional methods fail. Published at ICML 2026.
A Python system that classifies neurons from streaming calcium imaging while an experiment is running, with an interface for watching the results come in.
Education
Teaching
Industry
Recognition
Always glad to hear from people curious about the same things.