Research Interests:

Reinforcement Learning, Meta-Learning, Curriculum Learning, Multi-task Learning, Sequential Decision Making, Graphical Models, Interpretable ML. Applications in healthcare, biology, medicine, natural language and general AI.


1 Xinshi Chen, Yan Zhu, Haowen Xu, Muhan Zhang, Liang Xiong, Le Song.Learning Two-Time-Scale Representations For Large Scale Recommendations. Under Review.

2 Haowen Xu, Hao Zhang, Zhiting Hu, Xiaodan Liang, Ruslan Salakhutdinov, Eric Xing. AutoLoss: Learning Discrete Schedule for Alternate Optimization. ICLR2019

Work Experience, Beijing

Algorithm Engineer in Natural Language Processing (NLP) team , 09/2017 - 02/2018

  • Designed an attention based sequence labeling model, which achieves state-of-the-art results on NER taskon benchmart dataset, reducing the inference time of company‚Äôs online service product by 90%.
  • Proposed a deep architecture with paragraph reasoning modules for document classification, which takesadvantage of both symbolic reasoning and deep nerual nets.

Research Project

Protein sequence alignment via deep learning and reasoning

Georgia Tech, Advisor: Prof. Le Song, 02/2020 - present

  • Designed a deep architecture for protein sequence alignment problem to incorporate prior knowledge via adifferentiable reasoning layer, which achieves state-of-the-art performance on benchmark dataset.
  • The model is highly structured and performs well in case of small data, which is an important propertyfor solving biological problems.

Learning two-time-scale user model for recommendation system

Georgia Tech (joint project with Facebook AI, Personalization Team), Advisor: Prof. Le Song, 02/2020 - present

  • By modeling active and inactive users in different ways, the overall hybrid user model is simple yet effective,achieving at least 7% improvement on two largest benchmark datasets.
  • This hybrid model tackles the challenges of bothlong-rangesequence modeling for active users, and thecold-startproblem for inactive or new users jointly.

Meta Learning: learning of dynamic learning schedule

Carnegie Mellon University, Advisor: Prof. Eric Xing, 03/2018 - 07/2018

  • Proposed a meta-learning framework, which provides a generic way to learn the discrete optimizationschedule from metadata, allowing for a dynamic learning schedule in ML problems. Accepted by ICLR2019.

Auditory signal segmentation and classification via deep learning

Johns Hopkins University, Advisor: Prof. Xiaoqin Wang, 08/2016 - 05/2017

  • Designed a deep architecture for marmoset vocalization segmentation and classification working underhigh signal-to-noise ratio condition, which improves the detection F1-score by 50%.
  • The earliest work to apply deep learning to animal vocalization detection.

Single photon simulation of Intrinsic Imaging system

Johns Hopkins University, Advisor: Prof. Xiaoqin Wang, 08/2016 - 05/2017

  • Built a single photon simulation model in C++ and conducted simulation experiments to model thepropagation of polarized photon in brain tissue.
  • Experimental results showed that polarized photon can improve imaging depth, which guided the buildingof an intrinsic imaging system.


I am currently a Master student in CS&E in Georgia Tech, working with Prof. Le Song. I completed my bachelor degree in Tsinghua University in 2017. See more about my research in the Projects page.

Other materials:

Word Embedding Review