Contact me:

---

Heat Diffusion Simulation

This project is based on an ACM Transactions on Graphics article by Crane et al., Geodesics in Heat: A New Approach to Computing Distance Based on Heat Flow. Keenan Crane introduced the heat method for computing the geodesic distance efficiently. In practice, this method updates the distance an order of magnitude faster than the previously existing methods. This project is a reproduction of the algorithm proposed in the paper.

---

NeurIPS Reproducibility Challenge 2019

This project is submitted to Reproducibility Challenge @ NeurIPS 2019. Reproducibility Challenge @ NeurIPS 2019 is the most recent edition of the reproducibility challenge for members of the machine learning community to dive deep into cutting-edge research by aiming to re-implement parts of a paper. This project is based on the paper Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss with the focus of algorithm re-implementation and baseline improvement. A general scheme for baseline improvement with learning rate step decay and triangular policy is proposed in this project.

---

Animated Characters Creation Tool

This project deploys a hierarchy of transformations to draw characters. The characters are a collection of rigid objects connected by various parametric joints, such as single axis rotary joints, spherical joints, etc. Unique characters can be made using the implemented elements. The transformation hierarchy is developed to pose the characters and ultimately create a short key frame animation. Each key frame is then interpolated to produce a continuous animated clip.

---

Ray Tracer

This project applies physics based ray tracing that aims to simulate the way light bounces off objects, in turn creating more realistic shadows, reflections, and lighting effects. This is a program that can render ray-traced images of complex scenes. Some functionalities include anti-aliasing, accelerated ray-surface intersection, depth of field, soft shadows, glossy reflection, texture mapping, various shading techniques, etc.

---

Investigation of ResNet in Image Classification Problems

ResNet, short for Residual Network, is one of the recent year breakthroughs in machine learning community. It has now become a classic neural network structure used as a back bone for many computer vision tasks. This model allows extremely deep neural networks to be trained, by solving the problem of vanishing gradients in training very deep neural network prior to its existence. In this work, we investigated the application of ResNet on the modified MNIST classification task. A task-specific ResNet model was implemented and experimented with several factors such as the depth of network, different choices of optimizers, and the batch sizes.

---

Depth of Field Dolly Zoom Camera

This program is a simulated 35mm autofocus dolly zoom camera. The camera provides the ability to select focus point, set sperture size, automatically adjust focal length according to the focus point, and create bizarre dolly zoom cinematic effect. The scene is rendered by averaging multiple renders with the accumulation buffer, and the users can set how many renders to be used.

---

Reddit Comment Classification

Supervised text classification is an automated process of classification of text into predefined categories. The goal of this project to to classify a comment from reddit.com as coming from one of 20 selected subreddits (subform communities). We have explored this with several different text cleaning and feature extraction methods. Different classic machine learning algorithms are also applied and compared in terms of their performance in the context of data distribution: Multinomial Naive Bayesian classification (NB), Logistic Regression (LR), Random Forest classification (RF), Support Vector Machines (SVM) and ensumble methods.

---

© Copyright 2020 Doreen He. All rights reserved.