Projects
Big Data Mining and Management Course Projects
We completed three major projects throughout the course: Project 1 uses BERT for sentiment analysis, Project 2 uses Deepwalk to train graph embeddings, and Project 3 uses Wide & Deep learning model to estimate business ratings for the users. [CODE] Search Engine
Project for the Search Engine for Web and Enterprise Data Course at HKUST. This project contains the scraper and indexer for the HKUST CS Department's official website, which is written in Java. It also contains a frontend interface of the search engine built using Apache Tomcat, as well as the database for storage built using RocksDB. [CODE] Fake News Classification
Research project for the course, Machine Learning, at HKUST. In this research, we conduct the Performance Analysis of Different Word Embeddings and Transformers on Fake News Detection. We compare between Word2Vec, GloVe, and Elmo; and also between BERT, ALBERT, and DistilBERT. We adopt three phases, each focusing on a different analyses of fake news detection: Phase 1 emphasized on comparing the performances of different embedding layers and transformers in a general setting, Phase 2 attempted to observe the capabilities of representative models under low-resource settings, and Phase 3 explored the possibility of transfer learning through pretraining of these representative models. [PAPER] [SLIDES] [CODE] 2048 Game
Project for the Computer Organization Course at HKUST. This is a rebuild of the classic 2048 game that is usually played on portable devices. This project is built entirely using MIPS. [CODE] Human-Computer Interaction Course Projects
A personal portfolio displaying the list of projects I did for my Human-Computer Interaction course at HKUST. We completed three projects throughout a few months: sustainable ad campaign, augmenting paper as an information career, and human-robot interaction. [WEBSITE] Tower Defense Game
Project for Software Engineering Course at HKUST. This project features a tower defense game with the player building different towers to destroy the monsters that will be generated once a while. This game is built using Java and JavaFX framework. [CODE] 3D Visualization of Chinese Word Embeddings
This website shows a 3D visualization of word embeddings of Chinese characters and words. A form is also provided to select k-nearest neighbor words given a word based on some filter options. The corresponding embeddings to words are experiment results from the research at our lab, and the website is used as demo at ACL 2019. [CODE] Tensorflow.js Real-time Object Recognition
This project is created while I was attending the Sydney | Hong Kong Ideation Exchange Program partnered between HKUST and the Macquarie University. The topic is about how to make the transportation, specifically the metro system, more sustainable in Sydney and Hong Kong according to the UN Sustainable Development Goals. Our idea was to provide a platform for the metro company to incorporate camera vision technology to their surveillance cameras already existing in the stations, to more effectively synchronize the trains between stations to minimize the waiting times for passengers. [WEBSITE] [SLIDES] [CODE] 
Top 2 Pitch Teams
Generative Adversial Network with Object Localization for Fashion Trend Prediction & Style Inspiration
Research project for the course, Deep Learning in Computer Vision, at HKUST. In this work we use Generative Adversarial Networks (GAN) with Faster R-CNN to create predictive fashion samples from the data collected from multiple sources of recent high-fashion catwalks. With sufficient amount of training, we found that the DCGAN model is able to output a set of images resembling a model walking on a catwalk with newly proposed patterns as the clothing design as our final product. [PAPER] [SLIDES] [CODE] Deep Learning for NLP Course Assignments
During my first year at HKUST, I studied Deep Learning for Natural Language Processing, which made me interested in this field. We learned about the modern technologies applied on text, such as word embeddings, CNN, RNN, seq2seq, self-attention, and even transformer models. The assignments of this course are as follows: Twitter sentiment analysis using Logistic Regression, Twitter sentiment analysis using CBOW embeddings and FeedForward Neural Network, Amazon product review rating classifier using word-based CNN, and Sherlock Holmes novel generation using RNN. [CODE] Comparing the Efficiency of Random Search and Tree-Structured Parzen Estimator Algorithms to Optimize Convolutional Neural Networks for Image Recognition
This research work dives into comparsion of two algorithms in optimizing hyperparameters for the most popular kind of neural networks, CNNs. I completed this research individually with the help of my school advisor as a requirement for the IB Extended Essay. The focus of the work is to see which algorithm performs better in what kind of scenarios since each set of hyperparameter configuration can potentially impact the performance of the image classification model. Majority of the research is implemented using Keras. [PAPER] Food-Recognition Health Mobile Application
A food-recognition Android app that tracks the calories the users consume and gives intelligent food suggestions based on their current diet and goal. The app performs automatic food-recognition through the smartphone camera, which is implemented using Tensorflow Lite and the Food 101 dataset of food images. This project is submitted as part of my high school coursework for IB Computer Science. [CODE] Club Management Website Platform
An online club exhibition and management platform designed for club system to promote the effects of club publicity and to simplify club management process for both club leaders and students. As a team of four, we built this website for my high school from scratch and this website proved to be successful as the number of active users frequently exceeds 1000. This platform is still being used by the next generation of students and teachers. [WEBSITE] [CODE] 