Electrical and Systems Engineering
Monday, November 10th Raisler Lounge (Towne 225)
Temporal Knockoffs: Variable selection for time-varying systems with e-processes
Bio: I am a second-year Ph.D. student co-advised by Prof. Pratik Chaudhari and Prof. Christos Davatzikos. My research focuses on developing methods for causal inference, with applications to longitudinal and cross-sectional studies of neurodegenerative diseases. Prior to my Ph.D., I completed my undergraduate degree at IIT Bombay, where I worked on topological data analysis under the guidance of Prof. Debasish Chatterjee and Prof. Neeta Kanekar.
Abstract: One of the primary goals of ‘explainable AI’ is the identification of a small subset of explanatory variables in an attempt to understand interesting phenomena. The Markov blanket constitutes one such subset, essential for tasks involving causal interpretation, prediction, and robustness. In medical imaging, identifying such variables is particularly important for achieving generalization across sites and mitigating domain shifts induced by scanner or population biases. Existing approaches based on the model-X knockoffs framework (Barber & Candès, 2015) provide finite-sample control of the false discovery rate (FDR) under the IID assumption. However, longitudinal data violate this assumption and exhibit temporal dependencies, non-stationarity, making the standard knockoff constructions invalid. In this work, we explore a principled extension of knockoff-based variable selection to time-varying systems by leveraging ideas from betting games and e-processes in sequential hypothesis testing. We explore its applicability to both synthetic datasets as well as test it on real-world longitudinal neuro-imaging data from ADNI.
Monday, November 24th Raisler Lounge (Towne 225)
Nonconvex Linear System Identification
Bio: Uday Kiran is a third-year Ph.D. student under the supervision of Prof. René Vidal. He earned his bachelor’s degree in Electrical Engineering from the Indian Institute of Technology, Hyderabad, in 2023. His research interests include optimization techniques and establishing statistical guarantees for control systems, machine learning, and signal processing applications.
Abstract: The goal of system identification (SysID) is to learn a mathematical model from a corpus of temporal observations of a system’s inputs and outputs. SysID is a fundamental problem in engineering, with applications ranging from circuit design to robot control. Classical approaches to linear SysID rely on convex relaxations that offer strong theoretical guarantees. However, these methods often suffer from scalability issues and are not well-suited for large-scale systems. In this talk, I will present a nonconvex optimization approach to linear SysID that overcomes these limitations. It is well known that converting convex problems to nonconvex ones can lead to significant computational advantages, but they typically lack theoretical guarantees. I will show how we managed to achieve the best of both worlds: a faster and theoretically sound algorithm.
Monday, March 24th Raisler Lounge (Towne 225)
Real-Time Spatiotemporal Motion Planning for Autonomous Robots
Bio: Yuwei Wu is a third-year Ph.D. student in the GRASP Laboratory under the supervision of Prof. Vijay Kumar. Her research focuses on motion planning and trajectory optimization for mobile robots, particularly in dynamic, uncertain, and complex real-world environments. She received her B.Eng. degree in Transportation Engineering from Beijing Jiaotong University, China, in 2019, and her M.S.E. degree in Systems Engineering from the University of Pennsylvania in 2022. Before pursuing her Ph.D., she worked with Prof. Fei Gao at Zhejiang University on trajectory optimization for quadrotors.
Abstract: The representation of geometric environments and moving objects plays a critical role in obstacle avoidance. However, not all representations are well-suited for real-time planning that requires efficient representation, querying, and updating. A key challenge in this domain is computing a good approximation of the environment for trajectory optimization. In this talk, I will introduce the concept of Safe Flight Corridors (SFC) and how to locally optimize them along a predefined path with waypoints initialized by a geometric or kinematic planner. I will then explore the role of SFCs in environmental representation for learning-based trajectory optimization, particularly for quadrotors. Finally, I will present how temporal constraints can be integrated to generate temporal corridors for dynamic obstacle avoidance and discuss potential extensions incorporating different sensor inputs.
Monday, March 3rd Raisler Lounge (Towne 225)
Understanding Deep Learning via Simple Models
Bio: Behrad is a Ph.D. student in the Electrical and Systems Engineering department at the University of Pennsylvania, advised by Prof. Hamed Hassani. His research focuses on deep learning theory, high-dimensional statistics, and information theory. He obtained his M.A. in Statistics from the Wharton School of the University of Pennsylvania in 2024, and his B.Sc. in Electrical Engineering from the Sharif University of Technology in 2020, with highest distinctions.
Abstract: In this talk, I begin by providing a subjective overview of the theoretical and practical trends in modern deep learning, which motivates a deep learning theory based on simple models. As a case study, I demonstrate that the random features (RF) model—a minimalist framework—can reveal new and previously unknown behaviors of deep models under distribution shift. Recognizing the RF model’s limitation in that it does not learn features, I propose an extension that incorporates feature learning while remaining fully analyzable. I then illustrate how this new model can demystify and theoretically motivate layer-wise preconditioning in deep learning optimization. Additionally, I show that the extended model exhibits intriguing phenomena such as grokking and emergence.
