Pengyuan "Eric" Lu

Ph.D. Candidate, Computer and Information Science

University of Pennsylvania

eric.lu.py@gmail.com | curriculum vitae

Levine 614, 3330 Walnut St., Philadelphia, PA 19104

About Me

I am an upcoming fourth year Ph.D. student at Computer and Information Science Department, University of Pennsylvania, research fellow of PRECISE Center and advised by Prof. Insup Lee. My research interests focus on machine learning-based components in cyber-physical systems (CPS). Specifically, I aim to improve desirable properties such as safety, security, liveness and real-timeness to ML-based CPS in complex environments and defend the systems from adversaries. I became a Ph.D. candidate in April 2021.

Before joining University of Pennsylvania, I completed my undergraduate program as a Bachelor of Science in Computer Science, University of Virginia, while assisting Prof. Lu Feng in CPS research.

Keywords: life-critical cyber-physical systems, machine learning, deep learning, embedded systems

My Research

All figures are credited to the research teams as in the author lists.

I publish my works under my legal name "Pengyuan Lu".

Submitted Work

Scalable Semi-supervised Learning with Few Labeled Data via Data Programming

P. Lu, S. Lee, A. Watson, D. Kent, I. Lee, E. Eaton and J. Weimer

International Conference on Very Large Databases (VLDB'22)

We address the problem that current continual learning / lifelong machine learning requires lots of labeled training data, but existing semi-supervised continual learning (SSCL) frameworks are not scalable and costs computational resources. Therefore, we developed a tool, namely Mako, that leverages data programming to pseudo-label large unlabeled data to support low-cost SSCL. Mako maintains high per-task and overall accuracy, has low catastrophic forgetting, reduces computational overhead, passes scalability tests, and enables parallel implementation.


Keywords: machine learning, continual learning, semi-supervised learning, scalability

Published Work

Confidence Composition for Monitors of Verification Assumptions

I. Ruchkin, M. Cleaveland, R. Ivanov, P. Lu, T. Carpenter, O. Sokolsky and I. Lee

ACM/IEEE International Conference on Cyber-Physical Systems (ICCPS'22)

Real-time system assumptions can be violated during runtime in physical environments, and such discrepancies can be easily fixed (calibrated) for individual atomic assumptions but hard for combined ones such as safety guarantee. Therefore, we propose a set of methods to composite calibrated assumptions and prove that the combined assumption propositions have probabilistically upper-bounded discrepancy. [link]


Keywords: cyber-physical systems, formal methods, probabilistic computation

FRED: Fall Risk Evaluation Database Based on EHR Data

P. Lu, X. Li, S. Jang, A. Lee, S. Pugh, A. Watson, R. Bjarnadottir, R. Lucero, G. Demiris, A. Nenkova, J. Weimer and I. Lee

IEEE/ACM Conference on Connected Health Applications, Systems, and Engineering Technologies (CHASE '21)

This is a poster that describes our fall risk evaluation database that organize time-stamped patient information in EHR, so that each patient can be categorized into increasing, decreasing, constantly high and constantly low fall risks in different time periods, benefiting future machine learning-based analysis. [link]


This work has been selected for 2021 MidAtlantic Bioinformatics Conference. The project is also uploaded to PhysioNet.


Keywords: database systems, internet of medical things, time series, machine learning

RT-ACL: Identification of High-Risk Youth Patients and their Most Significant Risk Factors to Reduce Anterior Cruciate Ligament Reinjury Risk

A. Watson, P. Lu, E. Greenberg, J. Lawrence, T. Ganley, I. Lee and J. Weimer

IEEE/ACM Conference on Connected Health Applications, Systems, and Engineering Technologies (CHASE '21)

A machine learning-based framework that predicts ACL retear risks in youth patients, leveraging weak supervision with expert-generated weak labeing functions. [link]

This paper is nominated for Best Paper Award.

Keywords: machine learning, weak supervision, internet of medical things

Real-Time Attack-Recovery for Cyber-Physical Systems using Linear-Quadratic Regulator

L. Zhang, P. Lu, F. Kong, X. Chen, O. Sokolsky and I. Lee

International Conference on Embedded Software (EMSOFT '21)

A model-based recovery controller that guides an attacked CPS to a target set while maintaining it safe before a conservative deadline, with the system dynamics in linear ODE model but all sensors are assumed to be corrupted. [link]

Keywords: cyber-physical systems, control systems, safety guarantees

Attack-resilient Sensor Fusion for Cooperative Adaptive Cruise Control

P. Lu, L. Zhang, B. Park and L. Feng

IEEE International Conference on Intelligent Transportation Systems (ITSC '18)

A defense algorithm specifically designed to protect connected automous vehicle platoons from compromised sensors, utilizing the communication channels among them to adjust system states. [link]

Keywords: cyber-physical systems, control systems, distributed systems, safety guarantees

... and there are many other ongoing projects.

My Services

  • Reviewer of Conference and Workshop on Neural Information Processing Systems (NeurIPS 2022)

  • Reviewer of International Conference on Machine Learning (ICML 2022)

  • Reviewer of ACM/IEEE International Conference on Cyber-physical Systems (ICCPS 2022)

  • Reviewer of Informatics for Health and Social Care (IHSC 2021)

Cat Gallery