Biography
Hi! I am Xinchi Shi, a Year 3 undergraduate student majoring in Computer Science at University of Liverpool in the UK. My research interest lies in General Natural Language Processing (NLP), Machine Learning and Deep Learning.
Interests
- NLP, such as dialogue system and text generation
- Machine Learning
- Deep Learning
Education
-
BSc in Computer Science
University of Liverpool, 2022.09 - 2024.06
-
BSc in Information and Computing Science
Xi'an Jiaotong-Liverpool University, 2020.09 - 2022.06
Projects
*
Proposed a light-weight neural network that supports polyphonic output for music transcription
Developed an AI-powered music generation tool, allowing users without professional musical backgrounds to input a
cappella and receive a newly minted piece of music enriched with synthesized audio and a piano accompaniment
Designed a novel framework for fake news detection that overcame the low accuracy of conventional machine learning methods.
Fine-tuned BERT to create a model that improved the accuracy of Covid-19 fake news classification from 85.37% to 93.68%
Combined fine-tuned BERT (for frozen and unfrozen parameters) with CNN and bidirectional LSTM layers.
Achieved SOTA results on the fake news detection task and an improvement of 8.9% over baseline
Implemented a full-stack online social forum for University of Liverpool students, promoting enhanced communication and elevating
their university experiences, with the ability to serve over 200 concurrent users
Adopted agile methodologies (specifically the Scrum process) and spearheaded a team, crafting over 50 test cases
covering four primary user scenarios via the Pytest framework. Seamlessly integrated features such as permission
management, file/image uploading, categorization of posts, and email functionalities
Serendipity: Flask-Based Student Online Social Forums
Research Experience
Multi-Scale Document
Dewarping
via an End-to-End Deep Neural Networks Approach
Advised by Prof. Qiufeng Wang
Projects include:
- Implemented a deep learning-based approach for multi-scale document dewarping to tackle the challenges of small-target document warping.
- Integrated a ViT-based network for initial correction and subsequently applied a curvature correction method through sparse control point prediction. Collected and released a dataset comprising document images showcasing various scales and distortion levels.
- Leveraged the MS-SSIM and Local Distortion image quality assessment metrics, achieving performance gains of 1% and 7.5% respectively over the prevailing SOTA models.
See the Poster Here
OpenCV-Based Eutectic
Gallium-Indium (EGaln) Image Processing
Advised by Dr. Chenguang Liu
Projects include:
- Applied edge detection and morphological transformations to remove noise from images. Developed a measurement algorithm based on mouse events in OpenCV, allowing the precise measurement of syringe tip diameters.
- Adopted vectorized computation to build a versatile system for integration into the Ttinker GUI, enabling both manual and automatic measurements. Achieved a 15-fold speed increase.
See the Poster Here
Certificates
Neural Networks and Deep
Learning
Machine Learning
Awards and Honors
- XJTLU Academic Excellence Scholarship (ranked 4/277 in CS major)
- University of Liverpool 2+2 Excellence Scholarship (ranked 1/295 in CS major)
- XJTLU Summer Undergraduate Research Fellowship (twice)