Exploring Neural Models for Predicting Dementia from Language Master thesis

automatic dementia prediction, multimodal embedding, neural coherence model

• Use neural network models for dementia prediction to avoid task-specific features.
• The model for combining multimodal features gives performances comparable to traditional model using feature selection.
• The proposed coherence score features do not show much difference between dementia patients and healthy controls.

A Neural Model for Predicting Dementia from Language MLHC 2019 Paper

automatic dementia prediction, Hierarchical Attention Networks

• Apply a neural framework, Hierarchical Attention Networks, for dementia prediction.
• The HAN-AGE model achieves state-of-the-art performance on a widely used dementia dataset.
• Visualization and statistical analysis reveal that the attention mechanism of the model manages to capture similar key concepts as the information unit features specified by human experts.

Chinese Character Generation Report

conditional convolutional VAE/GAN, PyTorch

• Preprocess a Chinese handwritten character dataset, obtaining the bitmap of each character and the corresponding GBK encoding.
• Use three different neural generative models (conditional convolutional GAN, conditional convolutional VAE and a hybrid model combining both cCVAE and cCGAN) to generate Chinese characters conditioned on their GBK encodings.

A Distribution Similarity Based Regularizer for Learning Bayesian Networks Report

Directed Graphical Model, KL Divergence Regularization, Bhattacharyya Distance Regularization

• Use parameter sharing and multi-task learning to encourage similar factors.
• Propose a novel regularization term by encouraging similar distributions over factors.
• Evaluate different models on the task of approximating the perturbations of wave propagation in inhomogeneous materials.

Semi-supervised Image Captioning via Reconstruction Report

LSTM, Gumbel Softmax, PyTorch

• Try to tackle the task of generating image descriptions without {image, ground-truth caption} pair.
• Build a mapping between image feature and the generated caption feature.
• Propose a reconstruction loss (L2 loss between the original image feature and the reconstructed one) to train the model in a semi-supervised way.

Multimodal Learning Exercises

CNN, RNN, PyTorch

• Image classification and image captioning on MS COCO dataset.
• Image retrieval and language retrieval by unifying visual-semantic embeddings.

Robocode Tank Learning System Report

Reinforcement Learning, Neural Network, Java

• Developed a robocode tank using neural net and Q learning.
• Firstly implemented the robocode tank learning system (Q learning) using a look up table. Then replaced the table with a two-layer neural network.
• Using experience replay to enhance learning performance.

Machine Learning Exercises

Machine Learning, Julia

• Implemented several machine learning algorithms.
• Applied some algorithms into practice, e.g. using K-Means to quantize images and filling in the missing bottom half of MNIST images with sigmoid belief network.

Healthcare Data Analysis Platform

Mahout Machine Learning

• Participated in building a healthcare data platform, providing a friendly data analysis environment for doctors.
• Implemented the data engine module based on Mahout machine learning library, including logistic regression, random forests, K means, Canopy and Fuzzy k means.

Pocket Finance

Computational Finance, iOS Development, Swift

• Developed an iOS app implementing simulations for three investment strategies.
• Multiple simulation results could be displayed on one graph for comparison.
• Implemented option pricing by computing Black-Scholes equation using a binomial tree.

Chinese Currency Serial Number Recognition

Neural Network, OCR

• Implemented an algorithm to select an optimal threshold from gray levels for picture so as to get rid of background noise as much as possible.
• Implemented a parallel algorithm for picture segmentation.
• Using segmented pictures as training set, built a BP neural network for serial number recognition.

Automatic Obstacle Avoidance Mini Vehicle

Embedded System(pcDuino), Bluetooth Communication, C

• Invented a mini vehicle working on two modes: automatic or under control by a mobile app.
• Associated with microcontroller, wireless communication, ultrasonic distance measurement, etc.

OpenGL Room

A Computer Graphics course project, using OpenGL to construct an indoor scene.

• First-person view (control movement by mouse/keyboard), including collision detection.
• User can construct, rotate and zoom basic geometries, and add texture to them.
• Support importing obj file, controlling the light, taking snapshot and playing embeded avi video, etc.

Instant Message Application

Android Development, Web Development Demo

• Developed a chat app on both web side and mobile (Android) client.
• Implemented features like sign up/sign in; find, add, delete and group contacts; send messages between web client and Android app, etc.
• Repo for Web Client:
  https://github.com/arankong/IMS
• Repo for Android App:
  https://github.com/arankong/SocialApp

Numerical Methods

Library Management System

HTML, PHP, JavaScript and MySQL

• Developed a system to manage library books.
• Supported features including add/delete books; add/delete card and borrow/return books.

Arduino Saxophone

Embedded System(Arduino) Demo