Effect of Synthetic Pretraining on Offline Reinforcement Learning, September 2023.

We Proposed a simple synthetic pretraining scheme that surpasses language pretraining on offline reinforcement learning. We aim to understand why simple synthetic data can be beneficial in offline RL.
[paper]

SUffix REtrieval-Augmented Language Modeling (SUREALM) June 2023.

We enhanced autoregressive language modeling with bidirectional knowledge through future context retrieval.
[paper][code]

Simultaneous Machine Translation with Deep Reinforcement Learning, Jan 2023.

We derived a learning algorithm for latency control in simultaneous machine translation using reinforcement learning framework.
[report]

AI for Scientific Research: Bio-treatment, May 2022.

We predicted vaccination response of patients using machine learning algorithms. A comprehensive pipeline including data pre-processing, feature engineering, cross validation, training, and evaluation is designed and implemented to provide the collaborator working in Bio-medicine with highly flexible and accessible tools with simple interfaces.
[website]

Object Detection with Self-Supervised Pre-training, May 2022.

Course project for CSCI-GA - 2572 Deep Learning. We used popular SSL methods for pretraining of ResNet-50 as well as Transformer architectures (DETR) and then finetuned detection heads on labeled dataset achieving reasonable performance (ranked top 3 in our class!).
[report][code]

Q&A Dialog System for Information Inquiry based on Similarity between Sentence Representations, March 2022.

Exploratory project fine-tuning sentence transformers on augmented data from limited templates. We used sentence transformer in a simple retrieval system to achieve inqury answering on faculty information of NYU Shanghai.
[code]