Welcome! My name is David. I am currently a PhD student at Vanderbilt University working with Prof. Meiyi Ma on eXplainable AI in Healthcare and Deep Learning research.
I am passionate about advancing the fields of eXplainable AI (XAI), healthcare, and machine learning. My work focuses on leveraging cutting-edge machine learning techniques to improve physical therapy, nursing simulations, and formal verification in AI. Through interdisciplinary collaborations and innovative approaches, I strive to create impactful solutions that bridge the gap between research and real-world applications.
Jul 2024 – Present
Collaborators: Daniel Levin, Gautam Biswas, Alyssa White, advised by Meiyi Ma
- Developing explainable AI methods to analyze video segments from Tobii Glasses' first-person perspective during nursing simulation training sessions.
- Designed an unsupervised segmentation method optimized for processing long videos efficiently.
- Focused on interpretable models to link gaze dynamics with task performance and learning outcomes.
Feb 2024 – Present
Collaborators: Meiyi Ma
- Creating an IMU-guided method to enhance classification accuracy and efficiency in multimodal data.
- Improved Temporal Segment Networks by incorporating motion-based insights for better frame selection.
Jan 2024 – Oct 2024
Collaborators: Meiyi Ma
- Developed a Continual Multitask Learning framework, addressing challenges in continual multitask learning without requiring replay buffers.
- More details will be available soon.
Sep 2023 – May 2024
Collaborators: Meiyi Ma, Taylor Johnson, Diego Manzanas Lopez
- Led the development of a framework designed to evaluate and test the comprehensiveness of state-of-the-art attribution methods.
- Utilized Neural Network Verification (NNV) to analyze and validate the boundaries and trustworthiness of sampling-based attribution methods.
- Designed experiments to showcase the method's robustness in providing deterministic and targeted explainability for neural networks.
- More details will be available soon.
May 2023 – June 2024
Collaborator: Meiyi Ma
- Led the development of EXACT, a novel method for segmenting exercises within multivariate time series data using PyTorch.
- Designed and implemented a deep learning model leveraging a U-Net architecture with temporal positional encoding to accurately identify and categorize exercise phases from complex sensor data.
- Integrated dense labeling techniques with state-of-the-art neural network architectures, improving segmentation precision and reliability.
- Conducted extensive experiments across various datasets, demonstrating the method’s adaptability and superiority over traditional segmentation techniques.
- Developed a modular Python software framework to facilitate easy replication, experimentation with different architectures, and application to diverse time series segmentation tasks.
Aug 2022 – May 2023
Collaborators: Meiyi Ma, Pamela Wisniewski
- Led the development of MicroXercise, a system that integrates Siamese Neural Networks with saliency maps to provide detailed feedback on physical therapy exercises.
- Designed and implemented a Siamese Neural Network capable of determining input similarity and generating detailed attribution scores for interpretability.
- Incorporated saliency map techniques and layer-wise relevance propagation into the architecture for explainability across modalities such as images and signals.
- Built a comprehensive feedback system with visual and textual components to deliver actionable insights based on IMU sensory data.
- Conducted a mixed-methods study involving semi-structured interviews, surveys, and quantitative analysis to evaluate system effectiveness and user perception.
- Highlighted the system’s role in fostering self-awareness among users by serving as an observer during exercise sessions and encouraging the development of a mind-muscle connection.
- Utilized Python and PyTorch for model development, with Swift and Unity for additional application integration.
Aug 2021 – May 2022
Collaborator: Meiyi Ma
- Led the development of PhysiQ, a framework for continuous tracking and quantitative measurement of off-site exercise activity through passive sensory detection.
- Designed a multi-task spatiotemporal Siamese Neural Network to evaluate both absolute exercise quality through classification and relative quality based on individual progress through similarity comparison.
- Implemented PhysiQ to assess exercises based on three key metrics: range of motion, stability, and repetition.
- Conducted data collection and annotation for 31 participants with varying exercise quality levels, enabling robust evaluation of the framework.
- Demonstrated PhysiQ's effectiveness with a detection accuracy of 89.67% for exercise quality levels and an average R-squared correlation of 0.949 in similarity comparison.
- Highlighted the role of PhysiQ in addressing challenges posed by lack of supervision and self-correction during at-home physical therapy exercises.
- Published a detailed overview of this work on Medium.
-
Wang, Hanchen David, Bae, Siwoo, Sun, Xutong, Thatigotla, Yashvitha, and Ma, Meiyi.
EXACT: A Meta-Learning Framework for Precise Exercise Segmentation in Physical Therapy.
International Conference on Cyber-Physical Systems (ICCPS), Nov. 2024, Under Review. -
Wang, Hanchen David, Bae, Siwoo, Chen, Zirong, and Ma, Meiyi.
Learning with Preserving for Continual Multitask Learning.
International Conference on Learning Representations (ICLR), Oct. 2024, Under Review. -
Cohn, Clayton, Davalos, Eduardo, Vatral, Caleb, Fonteles, Joyce, Wang, Hanchen David, Ma, Meiyi, and Biswas, Gautam.
Multimodal Methods for Analyzing Learning and Training Environments: A Systematic Literature Review.
ACM Computing Surveys, Aug. 2024, Under Review.
More on Publications
Database Management System: (Fall 2020)
Database Management System: undergrad course project equivalent to graduate level course on end-to-end implementation of a database at UC Irvine.
Categories: C/Cpp, Git
Single Sign On: (Winter 2020)
Single Sign On: internship At St. Jude Research Hospital, worked as Information Services
Categories: AngularCLI, TypeScript, Node.js
Online Movies Store Web Application: (Jun. 2019 – Aug. 2019)
Online Movies Store Web Application Project: undergrad course project at the University of California, Irvine. It is a 10-week project class that heavily targets the implementation of an application.
Categories: Java, MySQL, HTML, CSS, Maven
Stay Together: (Fall 2018)
Stay Together Project: my second year first hackathon working related to healthcare on how we can treat opioid-addicted patient better.
Categories: Swift, Storyboard
Teaching Assistant (Aug. 2021 - May 2022)
@Vanderbilt University
- Graded homeworks weekly for the students bodies and hosted weekly office hours to help students with the concepts, homework instructions, and debuggings.
- Drew on the whiteboard to demonstrate many concepts including how operating system utilizes fork and the relationship between parent and child in processes.
- Hosted additionally meeting with students to confront coding challenges and brainstormed with them to resolve problems in C programming
Undergraduate Grader (Sep. 2019 – Dec. 2019, Mar. 2020 – June 2020)
@UC Irvine
- Designed and developed extensive test cases for command-line programs, Bash-like shell, dynamic memory allocator, and client/server network programming in C.
- Collaborated with the professor and team of 2 graders to design C programming assignments which could be programmatically tested with the auto-grader.
- Assisted students 1-on-1 with debugging programs and understanding test case failures.
Information Services Intern (Jun. 2019 – Aug. 2019)
@ St. Jude Research Hospital
- Developed Single Sign-On project using Agile & Waterfall development cycles and researched Active Directory Federation Service applications for authentication.
- Implemented idP-initiated SSO with Angular framework as frontend and Node.js as backend server to process SAML assertion from ADFS. Deployed application on Tomcat to test the Active Directory by using SecureAuth.
- Optimized the application with MIT-Licensed NPM plugin for SAML 2.0, samlify, to establish connectivity with Active Directory over the Node.js.