I am currently a first year Computer Science grad
student in UCLA. Before that I was a research assisant at National Taiwan University in Robin Lab. I finished my Bachelor's degree in Computer Science in National Tsing Hua University, under the supervision of Prof. Chun-Yi Lee.
My research lies primarily in the area of Natural Language Processing and Machine Learning; this includes but is not limited to Natural Language Processing, Machine Translation, Generalization in Planning, and Reinforcement Learning. I am particularly interested in how a robot can efficiently act as human.
H.-W. Yu, P.-Y. Wu, C.-A. Tsao, Y.-A. Shen, S.-H. Lin, Z.-W. Hong, Y.-H. Chang, and C.- Y. Lee,
Task-Agnostic Reinforcement Learning (TARL) Workshop at International Conference on Learning Representations (ICLR)., May 2019
We proposed a modular architecture to separate the instruction-to-action mapping problem to two stages, including Goal Exploration Process and Cross-entropy method.
[pdf]
H.-W. Yu, C.-H. Huang, and K.-Y. Su,
Conf. Technologies and Applications of Artificial Intelligence (TAAI), pp. , Dec. 2017
Treated a Yes/No question as an entailment decision with a Decomposable Attention model, including evidence extraction, data pre-processing, and decomposable attention model.
Links:
[pdf] |
[code]
Our app, PupGo, allows dogs to establish connection with nearby pet dogs. We aim at facilitating socialization for dogs based on events in reality, which is quite intuitive and simple to schedule with our app. PupGo users need to explore nearby events and choose to join it.
[video]
Allears aims to help music lover to find their partners and exchange taste of music. Based on music genre, our users could
upload their emoji of a day, checkin with their daily-pick
song, and match with the desired ones. I implemented the
recommendation system of this dating app.
We proposed a modular architecture to separate the instruction-to-action mapping problem to two stages via distance function. Our method are more adaptive to novel instruction and action space.
[pdf]
We combined NLP instruction and semantic image segmentation. Then, integrated Gated-Attention Model in Python, visual inputs, and Kobuki robot hardware system via ROS.
Links:
[video]
We improved the accuracy of the multi-nli dataset and the alignment of Natural Language Inference via integrating domain knowledge and neural network.
This is the solution project for Meichu 2017 hackathon competition, where we won 2nd place. In this project, we leveraged in knn as a spam email detecter and natural language inference model as response system.
We also developed a spam email filter website for demostration.
[website]
We developed and designed an online anonymous chatroom
(similar to Wootalk). The frontend is designed by JavaScript
which contains a count down timer.
[website]
Digital Logic Design & Hardware Design and Lab.
  National Tsing Hua University, Hsinchu, Taiwan