- Palo Alto CA, USA 94304
- Email: [email protected]
Let's train machine to answer questions!
Natural Language Processing with Deep Learning class, teamed with: Daisy Ding. (01/2018-03/2018)
This project applies deep learning with the Bi-Directional Attention Flow
(BiDAF) network to train a model for the machine comprehension task on the
Stanford Question Answering Dataset (SQuAD).We implement the BiDAF model
that represents the context with character-level, word-level, and contextual-level
embeddings and utilizes the bi-directional attention flow to capture the interactions
between context and query. We experimented with exponential moving average
and conducted hyperparameter tuning. During the evaluation stage, we made
the answer-span prediction by searching the pair of start and end positions with
the highest joint probability. Our single model achieves competitive results of
75.594% F1 score and 65.299% EM on the test set.
Let's train machine trading!
Artificial Intelligence: Principles and Techniques class, teamed with: Tianchang He, Yunpo Li. (09/2017-12/2017)
In this project, we consider the problem of Portfolio Management
where a limited resource must be allocated among a set of assets to
maximize the expected return with transaction cost. We choose eigenportfolio
as baseline and greedy oracle. Our goal is to use deep reinforcement learning to reduce
the gap from oracle, and integrate enough information to characterize market trend.
The unforgettable four intelligence-intensive days with Xiaowei Xie and Chenyang Zhong! Thanks to the heated debate, and we successfully
formulated the so not well-defined problem finally!
Finalist Award for Mathematical Contest in Modeling 2016
Nowadays, space debris has become a rather urgent issue.
To address such a problem, especially in the Low Earth Orbit(LEO) where the
density of the junk reaches its peak, we develop a comprehensive model to deal with debris
of different sizes. Risks and benefits are carefully compared, contributing to the formulation
of final economic model to quantify cost and revenue.
I'm lucky to have my excellent team members: Xintian Han, Jiashuo Jiang and Molei Liu. Besides, the field
study to the company in Nanjing was a lot of fun! (05/2015-03/2016)
A data-driven idea in liquid crystal production has drawn increasing attention in manufacturing.
A precise beforehand judgement contributes much to the production efficiency.
Numerical simulations select promising candidates for target products with desirable properties,
thus largely shortening the procedure for trial and error, and reducing the production cycle.
In this project, I led the whole team to develop our own systematic two-stage model.
The first step was to predict liquid crystal properties from the given formula,
and we used boosting idea combining statistics and machine learning methods
such as support vector regression, neural network, regression tree, etc. The second step was the reverse,
to provide possible formulas for desired properties using conjugate gradient method.
We have so far achieved a precision within 5 times of the industrial standard.
It was a lot of fun with my dearest team members: BerBer,
Ruby, Phoebe, Syl, William, Kyle and Simon
in Qingdao, Hongkong and Beijing. Can't wait to meet you soon! Thanks for Dr. Lee Shui's generous support! (07/2014-08/2014)
Waste management includes the collection, categorization,
removal, continuing education regarding the different types of human waste produced.
An inquiry into the models and current state of technologies in use in facilitating
waste management in Hong Kong, Qingdao, and Beijing could be hopefully used as a roadmap to construct
a more comprehensive waste management plan for Mainland China and Hong Kong.
近期评论