Zheyu Wen

Zheyu Wen

Research Assistant/Master student

University of Michigan, Ann Arbor

Biography

Zheyu Wen is research assistant and master student in University of Michigan, Ann Arbor. His research interests include Computer Vision, Spatial temporal dynamic modeling for resting fMRI data and image processing. He is going to pursue Ph.D. program in Fall 2021.

Interests

  • Computer Vision
  • Spatial temporal dynamic
  • brain image

Education

  • MS in Electrical and Computer Engineering, 2021

    University of Michigan Ann Arbor

  • BS in Communication Engineering, 2019

    University of Electrical and Scientific Technology of China

Research Experience

 
 
 
 
 

Research Assistant

University of Michigan, Ann Arbor

May 2020 – Present MI

Variational Autoencoder algorithm in resting state fMRI data.

  • Extracted code was used to do individual identification (as brain ID). K-means clustering is performed to analyze brain patterns reflected by latent space.
  • Extend the data from 2D fMRI data to 3D fMRI data. 3D data contains white matters. I analyzed latent code considering ‘channel bandwidth’ to see the non-informative latent variable and informative variable in VAE model.

Neural Ordinary Differential Equation for temporal dynamic model.

  • I model the temporal dynamic of resting state fmri data as ordinary differential equation, and analyze the saddle point for latent of latent code. This method brings a new view for spatial temporal modeling compared to existing method like AR model and HMM.
 
 
 
 
 

Graduate Researcher

University of Michigan, Ann Arbor

September 2019 – December 2019 MI

Reconstruction on SPECT data. demo.

  • Constructed Convolutional Neural Network by Keras in Python to compress and reconstructed SPECT image’s projection. Explored low rank property of SPECT projection. Analyzed several metrics of reconstruction like NRMSE, Contrastive Rate, Contrastive to Noise Ratio.
 
 
 
 
 

Undergraduate Researcher

UESTC

July 2018 – July 2019 China

Intelligent Communication.

  • Used Neural network to replace computation costing part of BiG-AMP which is used in image reconstruction.

Matrix Factorization code

  • Studied the Bilinear Generalized Approximate Message Passing algorithm and designed Variational Bayesian to improve the accuracy. Utilized Matlab to simulate the algorithm and compared the two algorithms in the same condition.

Selected Projects

*
All Computer Vision Artificial Intelligence

Google Football AI training

Discretize the whole football field into meshgrid. Used Monte Carlo Search Tree to learn the action strategy of football AI to deal with the state captured by environment. See demo for more information.

Speech to Face

Extract latent code from both Speech signal and face image signal. Forced them to distribute close to each other. In testing time, generate talking face from speech data. code

Video prediction from single starting image

Extract language features from VisualCOMET to predict the context information given single image. Combined with the video generation backbone (Ordinary DifferentialEquation and 3D convolutional VAE) and Noise Contrastive Estimation (NCE) contrastive loss, we generate sharper and temporally coherent videos. Colab notebook

GMusic

Took advantage of GAN algorithm to learn music style and generate its own music after training. demo

Contact