Software Engineer Intern at Waymo LLC.   June - September 2019

Project: Road Agent Interaction Understanding.

• Built an end-to-end model for road agent interaction prediction, which utilizes state-of-the-art deep learning framework to extract rich spatial temporal information from raw sensor input.
• As a use case, applied interaction modeling in trajectory prediction, which reduces the ℓ₂-loss by 10%.
• This project results in a working paper for CVPR submission and a patent application in preparation.

Machine Learning Intern at Facebook Inc.   June - September 2018

Project: Post Content Classification with User Engagement Signal.

• Extracted post embeddings from user-post engagement history with a sparse neural network model.
• Added the embedding as an additional feature into existing post subtopic classification model, which improved test accuracy by 1.5%.
• Augmented the training data with label propagation (propagated the ground-truth labels of posts to their nearest neighbors with unknown labels in the embedding space), and trained the subtopic classification model with larger training data, which improved the test accuracy by 1%.

Data Science Intern at Cardinal Operations   February - May 2015

Project: Demand Prediction at First-On-Sale Products at VIP.com (NYSE: VIPS).

• Applied clustering analysis to estimate the lost sales for each product based on its sold-out time.
• Built a demand prediction model using bagged regression tree, which improved test-set accuracy from 0.05 to 0.50.

Research Intern at Fields Institute   June - Aug 2013

Project: Dynamical Systems Models in Macroeconomics.

• Designed a new algorithm to estimate the parameters in the Goodwin model by utilizing an initial integral of the system, which improved the regression accuracy from 0.3 to 0.99.
• Collaborated with economists to extend the model by incorporating the influence of banks.
• Presented our results to a panel of roughly 100 researchers.

Action Recognition and Localization in Videos  [Poster] [Report]

• Built a spatio-temporal CNN model for action recognition, which improves the top-1 accuracy by 1% comparing with frame-wise spatial models.
• Localized the activation feature (temporal pixels) in classification, and visualization shows high accuracy.

Aiding Sentiment Evaluations with Social Network Structure  [Poster] [Report]

• Proposed a new CNN model to solve the language variation issue in sentiment analysis, which captured the interaction of word embedding and author network embedding.

Playing Chinese Checkers with Reinforcement Learning  [Poster] [Report]

• Designed a game-playing agent using minimax tree and function approximation.
• Proposed a new value iteration algorithm for adversarial zero-sum game setting.