Xinyu Lu
Senior Undergraduate • CS • UMich

Hey, this is Xinyu Lu. Greetings🥳 I am a senior undergraduate student majoring in Computer Science at the University of Michigan, and also expecting to get a dual degree in Electrical and Computer Engineering at Shanghai Jiao Tong University. My research interest lies in Machine Learning and Natural Language Processing, especially in embodied AI, situated human-robot dialogue, neural collapse, and knowledge graph construction from literature.

• Resume

Education

Research

Voice Instruction for the Visually Impaired in Situated Dialogue

• Advised by Professor Joyce Chai
• Goals: Develop a voice assistant to help the visually impaired people with situated description and task-oriented dialogue [Situated Dialogue, Grounded Language Processing, Robotics]
• Reviewing human-computer interaction and psychology studies of accessibility for people with disability

Neural Collapse in Transfer Learning

• Advised by Professor Qing Qu
• Goals: Study the relationship between neural collapse and model transferability, and design new training objectives and data augmentation methods to prevent the effect of neural collaspe on transferability [Machine Learning Theory, Neural Collapse, Transfer Learning]
• Reviewed literature that studied the effect of training objectives on transfer learning, and delivered presentation
• Implemented source code of training and validation of transferring learning under fully fune-tuning and fixed features settings
• Developed code of calculating neural collapse metric and angles between hidden layer features and classifiers

Knowledge Graph from Biomedical Literature

• Advised by Professor Jie Liu
• Goals: Construct a knowledge graph from biomedical literature by extracting the entities and relations, and create a website for interactive graph visualization [Machine Learning, Natural Language Processing, Biomedical Engineering]
• Designed and implemented a data preprocessing pipeline of abbreviation and coreference resolution and sentence simplification
• Implemented code of extracting entities and relations with OpenIE of Stanford CoreNLP
• Improved efficiency of programs dealing with the gigantic dataset by developing a multiprocess program

Simulation of DoS Attack in Networked Control System

[paper]

• Advised by Professor Jing Wu
• Goals: Design new components in the networked control system, apply neural network architecture and simulate the system based on TrueTime [Networked Control Systems, Machine Learning]
• Implemented and incorporated a neural network trained on National Grid data to the traditional plant component
• Improved the stability of the system by carefully adjusting the parameters in the controller component

Course Projects

VoicEmail: An Intelligent Email Voice Assistant

[code]

• EECS 498-006: Conversational Artificial Intelligence, advised by Professor Jason Mars
• Goal: Build an email AI assistant that supports email summarization and management based on conversation
• Developed a chat box as the user interface on the web with TypeScript and React
• Supported the conversion between text and speech in the web with the SpeechRecognition and SpeechSynthesis interfaces
• Executed email processing commands in the backend with EZGmail, a Python interface to the Gmail API

Generating Concise Content: Text Summarization with BERT

[code] [report]

• EECS 595: Natural Language Processing, advised by Professor Joyce Chai
• Goal: Use two approaches to improve the quality of text summarization: apply coreference resolution on the original text and generate summarization with BERT, and incorporate BERT model with the pointer generator
• Run coreference resolution on the CNN/DailyMail dataset
• Implemented the code of the fusion of BERT and pointer generator
• Fine-tuned the pre-trained BERT model with the processed CNN/DailyMail text

From Game Theory to Machine Learning: Reducing Overfitting

• EECS 498-005: Principles of Machine Learning, advised by Professor Qing Qu
• Goal: Design a new loss function to reduce the overfitting in image classification tasks
• Defined a new loss function by incorporating the formula of interactions defined by Harsanyi in the game theory
• Trained VGG16 and AlexNet on the CIFAR-10 dataset under the original and new loss functions

Contact Me

Welcome to send a message! By the way, I'm a fan of hiking, always looking for friends to hike together :-)