Combined previous work with the newest trend. Inspired by scale laws paper, started to work on comparing cost of the following three methods when achieving same accuracy:1) in-context learning 2) few-shot finetuning 3) synthetic data generation with large models. Developed skills: Slurm, GitHub (maintain a project repo), CUDA, conda environment.

Worked on QuALITY (a question and answering dataset) with synthetic data augmentation. Shifted research direction after GPT-4’s emergence. Developed skills: python, PyCharm, GitHub, conda environment.

Helped with experiments complement through Python coding and method details checking. Organize the GitHub repo in the final stage.

Analyzed VLN agents by adding noise to the pretraining data. Proposed a unigram + landmark method to lower the cost of data creation and augmentation. Designed a nonsense-data augmentation method for effective VLN pretraining.

My co-2nd-auther work has been sent to the EMNLP 2023 conference.

link will be public soon

Communication system

Piano audio recognition based on MATLAB

Communication system simulation based on MATLAB platform

Audio recognition project based on sound dynamics signals (team leader)