Hi guys, I'm a graduate student in Computer Science and Technology with a passion for machine learning, linear algebra, and linguistics. I also love nature, sports, movies, American TV shows and Chinese-style music!
2023-09 - 2026-03
Working on Master of Science in Computer Science and Technology.
7 courses and 17 credits as of June 2023
2019-09 - 2023-06
Majored in Naval Architecture and Ocean Engineering as a freshman, ranked 1/210, transferred to Computer Science and Technology as a sophomore, also minored in Translation and Interpretation.
A total of 82 courses and 204 credits
Hangzhou, ZheJiang, China
2023.07 -
- Training code LLM
Tai'an, Shandong, China
2022.11 - 2022.12
- Advise State Grid tech personnel on automatic document processing, lead Shangdong team in the final on behalf of Alibaba Cloud Computing Co. Ltd.
- Won the national championship.
SJTU, Shanghai, China
2022.10 - 2023.09
- Develop project website, guide students in natural language inference.
School of Foreign Languages, SJTU
2022.10 - 2023.1
- Guide senior linguistic students in NLP, assist the weekly teaching, design the coursework.
Deep Learning Natural Language Processing Spoken Language Understanding
Chinese SLU with @Reginald. We formulated SLU as a sequence tagging task, and applied both discriminative token labeling models and seq2seq models to it. We also introduced a dual-branch decoder architecture to utilize manually corrected ASR transcriptions for denoising.
Paper CodeDeep Learning Natural Language Processing Sentiment Analysis
Twitter Sentiment Analysis with @Reginald. We verified that RoBERTa's embedding is better for SVM than Word2vec, and that RoBERTa itself is better than LSTM 🤪. We also investigated the strengths and weaknesses of fine-tuning and prompot-tuning with limited training data.
Paper CodeParallel Computation CUDA Computer Vision
In this project I assembled a ResNet from scratch using CUDA. Particularly, I used Tensor Cores (proposed in Volta architecture) to implement convolution (by im2col) and linear layers. Forward pass only, no gradients!
CodeDeep Learning Computer Vision Sketch Recognition
Image classification? ResNet Is All You Need!
Or... is it? For free-hand sketches, RNN can be useful as well!
Even better, we applied a dual-branch model to it, and fused the decision values of RNN and CNN.
Reinforcement Learning Variational Auto-Encoder Mahjong
Playing mahjong with reinforcement learning algorithms including DQN, double DQN, dueling DQN, and parametric noise under the Bayesian learning framework of Variational Oracle Guiding.
Paper