Peinuan Qin

• Updating ...

Tech stack: Python / Pytorch

• Collected 359405 commentaries and processed to build a 173M dataset. Took GPT-2 as backbone and combined title, summary of movies for movie commentary generation.
• Proposed a comprehensive end-to-end trained Chinese corpus punctuation addition model. Collected and processed 2G+ Chinese texts in over 15 major categories and trained the model.
• The generated narration fits the main content of the film and flows logically. Built a preliminary punctuation addition model for multi-category texts (still updating). Thesis is in preparation.

Tech stack: Python / Pytorch

• Collected patients’ motion data with Kinect, followed by data smooth and filtering.
• Proposed a gait annotation method. Led the annotation, data cleaning & visualization tools development.
• Proposed a sequential-based gait quantitive assessment algorithm with rich spatial-temporal features and gait semantic segmentation.
• Proposed Key Semantic Positioning Accuracy to evaluate model ability more comprehensively instead of relying only on precision, recall, and f1-score.
• Attained precise gait parameters and decreased the error rate of parameter calculation to 3.17%. Produced Deep Learning Assisted Gait Parameter Assessment of Neurodegenerative Diseases: Model Development and Validation and submitted to JMIR. ReadyGo have received Chinese national medical device registration and is used for commercial purposes.

Tech stack: Python / Pytorch / Keras

• Investigated the influencing factors of knowledge distillation. Conducted many experiments on distillation structure and mode. Proposed to model and fuse features from different levels and distilled the aggregation information and generated a research report MFFD: Multi-scale feature fusion distillation.
• Proposed a lightweight shared-mask attention mechanism and a multi-level attention structure and produced a journal paper MLAN: Multi-Level Attention Network.

Tech stack: Python

• Studied and practiced traditional machine learning algorithms and evaluation methods.
• Peformed domain analysis (with FFT) and denoising for brainwave signals.
• Extracted brainwave features and reduced features by PCA.
• Visualized signal and features and analysed confusion matrix to measure the classification model effect.
• Improved the classification result by 5% on given dataset compared to the baseline.

Tech stack: Python / Pytorch

Built a motion assessment platform based on deep learning models and the patient’s 3D motion data collected by Kinect. The device is already in use in major hospitals in China, helping middle-aged and elderly people with neurodegenerative diseases with contactless early diagnosis and rehabilitation.

Tech stack: HTML5 / CSS / JavaScript / MongoDB / Python / Vue / Django / Nodejs

Built and deployed a disease data website as an agile team scrum master. In this program, a Nodejs server is taken for normal requests and a Django server for chart generation. The front end was based on element-UI & Vue. Agile concept is used to drive projects forward, manage risk and build usable products iteratively.