Page 75 - Undergraduate Research Opportunities Program Proceedings 2017-18

School of Engineering
Department of Computer Science & Engineering

Supervisor: KIM Sung Hun / CSE
Student: GAO Tong / COMP Course: UROP1100, Fall
Fang Haoyang / MATH-PMA UROP1100, Fall

Neural dialog system has been an active area of research, as a dialog system makes it possible for human to interact
with computer using everyday language, lowering the threshold to fully utilize devices’ functionalities. One basic dialog
model is called sequence-to-sequence, which encodes the question into a context vector, and use a decoder to
reconstruct the corresponding answer given that context vector. However, this model does not consider the topic
clusters of answers thus producing monotonous responses. In our project, we present a modified version of sequence to
sequence model, aiming at improving the quality of replying questions with certain latent topics. This model is combined
from ideas of VAE and our supervisor. We implemented two improved models to answer questions with specific latent
topics. In this report, we first introduce the intuition and mathematic theory behind two models, then present our idea
with important code snippets. Finally, we propose some possible ways for further improvement on this model.

Supervisor: KIM Sung Hun / CSE
Student: KIM Dohoon / ELEC Course: UROP1100, Fall

The following report is a paper study of Unsupervised Neural Machine Translation, a novel method of NMT only using
monolingual corpuses of both languages. This is achieved by creating a cross-lingual embedding using the unsupervised
bilingual encoding method. The architecture of the RNN uses a shared encoder, with two separate decoders for the
translation of each language. In order to train the model unsupervised, new techniques such as denoising and back
translation is proposed to create pseudo parallel corpus for the RNN to learn on. The implementation of this method
was done using Tensor flow, up to the part of denoising of the cross-lingual embedded inputs.

Supervisor: KIM Sung Hun / CSE
Student: KOO Tin Lok / ELEC Course: UROP1100, Fall

The project topic I have chosen to study is “End-to-end representation learning for Correlation Filter based tracking”1. It
uses Correlation Filter to discriminate an image object from the surrounding images, thus allowing object tracking. One
key point in this project is to deep integrate Correlation Filter (CN) and Convolution Neutral Network (CNN), thus deep
features can be learnt.

Supervisor: KIM Sung Hun / CSE
Student: NG Chung Ki / ELEC Course: UROP1100, Fall

With the significant improvement of the computation power, it open up a lot of possibility of using computer to solve
problems in our lives. Nowadays conventional programs in the world, they are built with clear logical path for specific
task, and in order to build a sophisticated program, a thorough study of a problem is needed to develop a suitable model
for such task. As the amount of data in the world is growing, it is increasingly difficult for human to discover pattern and
creating the model. Deep-learning has offer a more sophisticated machine learning capability that could be more
efficient to learn about what the data representation. This report will briefly discuss about various concept of deep
learning.

65