Recent work from Gurnee et al. show that the activations of neural language models have high correlation with the spatial and temporal properties of their inputs. However, the work didn't estabilish a causal link between two. Our experiments, filling a part of this blank, show that the spatial representation has causal influence on the model's behavior. Editing LLM's spatial representations can improve the model's performance on a simple spatial task.


Paper Preprint

Does the 2D image generative model has an internal model of 3D geometry? Can a 2D neural network see beyond the X-Y dimension of a matrix of pixels? Our project found controllable representations of 3D geometry inside diffusion model.


Paper

Video Presentation (54K Views)

Wonder how attention flows inside your Vision Transformer? What visual patterns are recognize by machine's attention? Does machine's attention resemble human's visual cognition? Collaboration with Catherine Yeh. My main contribution is the visualization of learnt attention in vision transformer models.

We aim to further improve the accuracy and clinical applicability of echocardiography video segmentation by extending the analysis from half-heartbeat (End-diastolic to End-systolic phases) to multi-heartbeat video. Uisng supervised deep learning, we proposed a fully automated neural network-based segmentation method for sparsely annotated multi-heartbeat echocardiogram. Our proposed neural network exploited the spatiotemporal convolutions in extracting video features. Our paper about this project has been accepted for an oral presentation at SPIE Medical Imaging 2022: Image Processing Conference, webpage .

We trained a 3D-UNet to segment and predict the motion of cardiac structures in the Echocardiography Video. The convolutional neural network is implemented in PyTorch and trained with dual Nvidia Titan RTX graphics cards. The experimental results show that the motion tracking task enhances the learning of video segmentation. Video segmentation, compared to the traditional frame segmentation, produces temporally coherent segmentation on cardiac structures, such as left ventricles. The temporally coherent segmentation further improves the estimations on patients' left-ventricular ejection fractions using segmented echocardiography video.

KALMUS is a Python package for the computational analysis of colors in films. It provides quantitative tools to study and compare the use of film color. This package serves two purposes: (1) various ways to measure, calculate and compare a film’s colors and (2) various ways to visualize a film’s color.

Course Project of Math 343: Numerical Analysis
This project implements and extends a stochastic approximation (SA) method proposed by Song et al. [1]. This SA method aims to find an optimal set of buying and selling thresholds for trading a stock. The method assumes the stock has a mean-reversion property so the stock price will oscillate around a long-term average.

[1] Q. S. Song, G. Yin, and Q. Zhang, “Numerical methods for buying-low-and-selling-high stock policy,” 2008 American Control Conference, 2008, https://doi.org/10.1109/ACC.2008.4586627

Portfolio Website for Math 230: Data Visualization
This website hosts the visualization and generative artworks that I created during the course.

The project is implemented in R, and the portfolio website is built with Distill.

Linear probe found representations of scene attributes in a text-to-image diffusion model

Yida Chen's course project for MATH 230: Data Visualization instructed by Prof. Abby Flynt at Bucknell University.

The implementation of CLAS-FV described in "Fully automated multi-heartbeat echocardiography video segmentation and motion tracking".

Kaggle COVID-19 Radiography Classification Project

Ruby Inspect: online study notes of the programming language Ruby

KALMUS Color Toolkit for color analysis and visualization in film studies. Archived JOSS release of the software, v1.3.9. doi: 10.21105/joss.03156.