<![CDATA[PhD Defense | Structured Knowledge as a Lens for Trustworthy AI: Privacy, Unlearning, and Safety]]>36518 Title: Structured Knowledge as a Lens for Trustworthy AI: Privacy, Unlearning, and Safety
Date: Wednesday, November 18, 2026
Time: 10:00 AM – 12:00 PM (ET)
Location: Virtual via Microsoft Teams (link below). An in-person viewing room is also available at Coda C1215 Midtown (12th floor, Coda Building).
1. Dr. Pan Li (Advisor), Electrical and Computer Engineering, Georgia Institute of Technology
2. Dr. Justin Romberg, Electrical and Computer Engineering, Georgia Institute of Technology
3. Dr. Juba Ziani, Industrial and Systems Engineering, Georgia Institute of Technology
4. Dr. Duen Horng (Polo) Chau, Computational Science and Engineering, Georgia Institute of Technology
5. Dr. Pin-Yu Chen, IBM Research
Abstract
Privacy, unlearning, and safety are often evaluated at the level of individual records, facts, or requests. This dissertation uses structured knowledge as a lens to examine how relationships among these units shape what AI systems reveal, retain, and enable. For privacy, it develops graph diffusion mechanisms and distribution-dependent analyses that account for data structure. For unlearning, it reveals overlooked privacy risks for minority data and shows how seemingly forgotten facts remain recoverable through related knowledge, motivating evaluations and methods that address these dependencies. For safety, it demonstrates how restricted knowledge can be assembled across seemingly benign queries and develops a defender that learns when to intervene as information accumulates. Together, these contributions show how accounting for relationships among data, knowledge, and interactions leads to more faithful evaluations and more effective methods for trustworthy AI.
]]> shatcher8117942290462026-11-09 12:57:2617942290942026-11-09 12:58:1400eventStructured Knowledge as a Lens for Trustworthy AI: Privacy, Unlearning, and Safety
]]>2026-11-18T10:00:00-05:002026-11-18T12:00:00-05:002026-11-18T12:00:00-05:002026-11-18 15:00:002026-11-18 17:00:002026-11-18 17:00:002026-11-18T10:00:00-05:002026-11-18T12:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-18 10:00:002026-11-18 12:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]><![CDATA[]]><![CDATA[PhD Defense | Adaptive Subspace Tracking and Inverse Design for High-Dimensional Electromagnetics]]>36518 Title: Adaptive Subspace Tracking and Inverse Design for High-Dimensional Electromagnetics
Dr. Justin Romberg (Advisor), School of Electrical and Computer Engineering, Georgia Institute of Technology
Dr. Mark Davenport, School of Electrical and Computer Engineering, Georgia Institute of Technology
Dr. Saibal Mukhopadhyay, School of Electrical and Computer Engineering, Georgia Institute of Technology
Dr. Nima Ghalichechian, School of Electrical and Computer Engineering, Georgia Institute of Technology
Dr. Ryan Westafer, Georgia Tech Research Institute
Abstract
This dissertation develops theory, algorithms, and systems for high-dimensional optimization problems in electromagnetics. The first part studies subspace tracking and principal component analysis under hardware constraints such as limited computation, scarce memory, or acquisition restricted to a few linear measurements per sample. For Gaussian data in dimension d, it shows that an Oja-type algorithm using 2k adaptive linear measurements per sample estimates the top-k principal subspace at the minimax error rate of a two-level spiked model, up to a constant factor, when the sample size is known in advance, while measurement designs fixed in advance have a lower bound larger by a factor of order d/k at large sample sizes when the (k+1)-th eigenvalue is comparable to the k-th. It also develops regularized least squares algorithms on the Grassmann manifold for tracking time-varying subspaces. On the systems side, it describes AdaAFE-CIM, which uses subspace tracking for adaptive data compression in array processing hardware and reduces baseband power by 95.4% relative to full digitization in power-model simulations, together with an open-source end-to-end simulator for array processing under realistic hardware and channel effects. Finally, it presents inverse design methods for electromagnetic structures: a periodic level set function for gradient-free topology optimization, validated by fabrication and measurement for metasurfaces, and a neural adjoint method for reconfigurable antennas with exponentially large switch configuration spaces.
]]> shatcher8117939060152026-11-05 19:13:3517939060582026-11-05 19:14:1800eventAdaptive Subspace Tracking and Inverse Design for High-Dimensional Electromagnetics]]>2026-11-19T14:00:00-05:002026-11-19T16:00:00-05:002026-11-19T16:00:00-05:002026-11-19 19:00:002026-11-19 21:00:002026-11-19 21:00:002026-11-19T14:00:00-05:002026-11-19T16:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-19 02:00:002026-11-19 04:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]><![CDATA[]]><![CDATA[PhD Defense | Resource Pooling in Hyperconnected Logistics Networks]]>36518Title: Resource Pooling in Hyperconnected Logistics Networks
H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology
Committee
1 Dr. Benoit Montreuil (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
2 Dr. Walid Klibi, Centre of Excellence for Supply Chain Innovation and Transportation, Kedge Business School
3 Dr. Yao Xie, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
4 Dr. Valerie Thomas, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
5 Dr. Mathieu Dahan, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
Abstract
Hyperconnected logistics networks comprise interconnected hubs and organizations whose resources can be shared, so that capacity committed at one location can serve demand arising elsewhere in the network. Resource pooling therefore extends capacity planning under uncertainty beyond individual facilities and firms to include the coordination of shared resources and the discovery of outside capacity. This dissertation develops stochastic optimization, decomposition, and learning methods for resource-pooling decisions, from deploying known resources to expanding the pool beyond the network. We first formulate the dynamic deployment of pooled workers and autonomous mobile robots across urban parcel hubs as a rolling-horizon two-stage stochastic program that respects worker shift and location preferences, and solve large-scale instances with an accelerated Benders decomposition and clustering-based relax-and-fix heuristics. Then we design battery swapping and charging networks that support long-haul truck electrification by keeping swapping close to freight demand while allowing charging to shift to grid-suitable hubs. We model this design problem as a multi-period, two-stage stochastic location–inventory–fleet program and solve it with a Lovász-cut Benders decomposition that exploits the submodularity of the battery recourse function. Lastly, we introduce open capacity pooling, in which a disrupted network discovers capacity held by previously unengaged providers through a closed loop of coordination, search, and admission under a limited budget of requests for quotation. Each request targets a capacity gap remaining after incumbent firms coordinate through the alternating direction method of multipliers (ADMM) without sharing their operating models, and is directed to a provider ranked by a machine-learning layer built on directory records and large language model readings of provider self-descriptions. We evaluate the proposed methods in computational studies grounded in real-world data from industry partners and public sources. The findings show how resource pooling helps logistics networks respond to uncertain demand and disruptions by coordinating shared resources and incorporating capacity from outside providers.
]]> shatcher8117939058842026-11-05 19:11:2417939059622026-11-05 19:12:4200eventResource Pooling in Hyperconnected Logistics Networks]]>2026-11-16T14:00:00-05:002026-11-16T16:00:00-05:002026-11-16T16:00:00-05:002026-11-16 19:00:002026-11-16 21:00:002026-11-16 21:00:002026-11-16T14:00:00-05:002026-11-16T16:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-16 02:00:002026-11-16 04:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]><![CDATA[Machine Learning Seminar Series Fall 2026 | Measuring Cultural Variation: from Twitter to LLMs]]>27863Abstract: People's language use reflects their culture. As a result, LLMs should adapt to the cultures of their many users, but they struggle to do so. Today's LLMs are multilingual but not multicultural: they speak fluent Hindi or Mandarin yet reflect the (often stereotypical) American perspectives implicit in their training data. We first demonstrate how language can be used to quantify cultural variation, analyzing Twitter texts to quantify variation in individualism and collectivism across US counties. We then present a benchmark to assess how well LLMs understand cultural variation—a set of questions that should be answered differently across countries—and demonstrate that LLMs struggle to respond appropriately. Achieving truly global generative AI requires not just linguistic diversity, but LLMs that adapt to their multicultural users.
Bio: Lyle Ungar is a Professor of Computer and Information Science at the University of Pennsylvania, where he holds secondary appointments in Psychology, Bioengineering, Genomics and Computational Biology, and Operations, Information and Decisions. His group builds chatbots and language analysis tools for psychological and medical research. They study the drivers of physical and mental well-being and their cultural variation, and develop socially sensitive AI-based tutors and coaches.
For CODA guest access, please contact shatcher8@gatech.eduat least 2 business days prior to the event.
]]> Christa Ernst117861127422026-08-07 14:25:4217937342272026-11-03 19:30:2700eventAll Seminars Held 12pm - 1pm in CODA]]>2026-11-11T12:00:00-05:002026-11-11T13:00:00-05:002026-11-11T13:00:00-05:002026-11-11 17:00:002026-11-11 18:00:002026-11-11 18:00:002026-11-11T12:00:00-05:002026-11-11T13:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-11 12:00:002026-11-11 01:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]>For CODA guest access, please contact shatcher8@gatech.edu at least 2 business days prior to the event.]]><![CDATA[PhD Defense | Optimization Driven Machine Learning for Healthcare]]>36518Title: Optimization Driven Machine Learning for Healthcare Date: Monday, November 9, 2026 Time: 4 PM (ET) Location: George Tower 1502
Himadri S. Pandey Machine Learning PhD Candidate H. Milton Stewart School of Industrial and Systems Engineering Georgia Institute of Technology
Committee
Dr. Gian-Gabriel P. Garcia (Advisor), Industrial and Systems Engineering, University of Washington
Dr. Jing Li, Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Lauren N. Steimle, Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Kai Wang, Computational Science and Engineering, Georgia Institute of Technology
Dr. Landon Lempke, School of Medicine, Virginia Commonwealth University
Abstract
Machine learning has shown great promise for improving healthcare through more accurate diagnosis, risk prediction, and personalized care. However, accurate predictions alone do not guarantee better decisions. In practice, healthcare providers must make decisions with limited time, staffing, and resources. Yet these constraints are often overlooked during model development. This dissertation investigates how optimization can be integrated into machine learning to develop decision-making tools that explicitly account for these operational constraints. It addresses three interconnected questions: When is individualized information worth collecting? What information should we prioritize when resources are limited? And who should receive scarce interventions, and when?
The first part of this dissertation examines the value of information in clinical decision-making, developing a framework to quantify how individualized baseline testing changes diagnostic decisions in sports-related concussion assessment. The second part develops an operationally constrained machine learning framework that jointly selects diagnostic assessments and trains interpretable prediction models while accounting for uncertain administration times, clinical knowledge, and operational requirements. The third part introduces a mixed-integer optimization framework for ranked feature selection that optimizes feature rankings simultaneously across multiple resource budgets, accounting for feature acquisition costs and the predictive value of each successive feature set. The final part addresses the dynamic allocation of scarce healthcare resources among pediatric patients whose clinical conditions evolve over time. It proposes a multi-action restless multi-armed bandit framework and a decision-focused learning approach that combines neural networks and long-run resource planning to coordinate clinical interventions under resource constraints.
Together, these contributions advance a shift in healthcare AI from “What can we predict?” to “What actions should we take and for whom, given the resources available?”
]]> shatcher8117932809022026-10-29 13:35:0217932809602026-10-29 13:36:0000eventOptimization Driven Machine Learning for Healthcare]]>2026-11-09T16:00:00-05:002026-11-09T18:00:00-05:002026-11-09T18:00:00-05:002026-11-09 21:00:002026-11-09 23:00:002026-11-09 23:00:002026-11-09T16:00:00-05:002026-11-09T18:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-09 04:00:002026-11-09 06:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]><![CDATA[PhD Defense | Decentralized Learning of Interdependencies for Multi-Component Dynamic Systems]]>36518Title: Decentralized Learning of Interdependencies for Multi-Component Dynamic Systems
H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology
Committee
1 Prof. Nagi Gebraeel (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
2 Prof. Pascal Van Hentenryck, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
3 Prof. Yao Xie, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
4 Prof. Aaron Stebner, George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology
5 Prof. Jeff Shamma, The Department of Industrial and Enterprise Systems Engineering, University of Illinois Urbana-Champaign
Abstract
Modern industrial systems comprise multiple interconnected subsystems whose temporal dynamics are often interdependent, making it essential to understand how changes in one subsystem influence others. Learning these interdependencies is particularly challenging in decentralized environments where raw data cannot be shared, and individual subsystems operate proprietary models that cannot be modified. This dissertation develops decentralized learning frameworks for characterizing cross-client temporal interdependencies in both linear and nonlinear dynamical systems while preserving the autonomy of individual client models. For linear systems, we establish a federated Granger causality framework that enables the discovery of unknown cross-client interdependencies. We further use this federated Granger causality framework to investigate the propagation of estimation uncertainty and its implications for interdependency learning. For nonlinear systems, the dissertation introduces graph attention-based learning with a predefined candidate graph structure, establishing theoretical relationships between learned attention coefficients and Jacobian-based interdependencies. It further develops a complementary framework for learning nonlinear interdependencies without requiring a predefined graph skeleton, where the downstream task of root cause analysis is explored. The proposed methods are evaluated on synthetic and industrial datasets, demonstrating their utility for decentralized prediction and interdependency characterization.
]]> shatcher8117931284272026-10-27 19:13:4717931285042026-10-27 19:15:0400eventDecentralized Learning of Interdependencies for Multi-Component Dynamic Systems]]>2026-11-10T10:00:00-05:002026-11-10T12:00:00-05:002026-11-10T12:00:00-05:002026-11-10 15:00:002026-11-10 17:00:002026-11-10 17:00:002026-11-10T10:00:00-05:002026-11-10T12:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-10 10:00:002026-11-10 12:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]><![CDATA[IRIM Fall 2026 Seminar | Restoring Dexterity in Telerobotics: An Engineer’s Journey Into Neuroscience]]>27863Abstract: The human body can perform dexterous manipulation across a variety of environments. However, some environments present challenges due to limitations in function, distance, and scale. In many of these cases, access can be restored through the use of a telerobot. Dexterous manipulation via a telerobot is feasible only if it can effectively extend the user’s inherent sensorimotor control scheme to the distal environment. In this talk, I will discuss research from our lab that explores the intricate relationship between touch perception, sensorimotor control, and telerobotics in both healthy and impaired individuals. Specifically, I will focus on our efforts to 1) restore “neurally efficient” control in upper-limb prosthetics, 2) enhance skill learning and skill transfer in robotic minimally invasive surgical training, and 3) assess and augment touch perception in healthy and impaired sensorimotor systems. Throughout, I will emphasize the need for increased collaboration between engineers and neuroscientists to deepen our understanding of these long-standing societal challenges and move closer to long-term solutions.
Bio: Jeremy D. Brown is the John C. Malone Associate Professor in the Department of Mechanical Engineering at Johns Hopkins University and the Solomon H. Snyder Department of Neuroscience at the Johns Hopkins School of Medicine. He is also a member of the Laboratory for Computational Sensing and Robotics (LCSR) and the Malone Center for Engineering in Healthcare. Prior to joining Hopkins, Jeremy was a Postdoctoral Research Fellow at the University of Pennsylvania in the Haptics Research Group, which is part of Penn’s General Robotics, Automation, Sensing, and Perception (GRASP) Laboratory. He received his M.S. and Ph.D. degrees in Mechanical Engineering at the University of Michigan. He also holds B.S. degrees in Applied Physics and Mechanical Engineering from Morehouse College and the University of Michigan, respectively, as a graduate of the Atlanta University Center’s Dual Degree Engineering Program. Jeremy directs the Haptics and Medical Robotics (HAMR) Laboratory, whose research focuses on human-robot interaction and haptic feedback, with applications in upper-limb prosthetics, robot-assisted surgery, and rehabilitation. By integrating human perception, motor control, and biomechanics, his work advances the design of intuitive and effective human-machine interfaces. He has been honored to receive several awards, including the NIH R21 Trailblazer Award, the IEEE RAS Technical Committee on Haptics Early CAREER Award, the NSF CRII Award, the NSF CAREER Award, and the Sloan Foundation Fellowship. He was also named a scholar to the NIH funded Interdisciplinary Rehabilitation Engineering Career Development Program (IREK-12). He is a member of the Institute of Electrical and Electronics Engineers’ (IEEE) Robotics and Automation Society, the American Society of Mechanical Engineers (ASME), and the National Society of Black Engineers (NSBE). Brown’s work has appeared in several peer-reviewed journals and news outlets, including Nature Scientific Reports, the Journal of Neuroengineering and Rehabilitation, IEEE Transactions on Haptics, IEEE Transactions on Biomedical Engineering, IEEE Transactions on Neural Systems and Rehabilitation Engineering, the IEEE Transactions on Medical Robotics and Bionics, Forbes, and The New York Times.
]]> Christa Ernst117861244542026-08-07 17:40:5417930400962026-10-26 18:41:3600eventAll Seminars Held on Wednesdays 12:15 - 1:15pm]]>2026-11-18T12:15:00-05:002026-11-18T13:15:00-05:002026-11-18T13:15:00-05:002026-11-18 17:15:002026-11-18 18:15:002026-11-18 18:15:002026-11-18T12:15:00-05:002026-11-18T13:15:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-11-18 12:15:002026-11-18 01:15:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]>christa.ernst@research.gatech.edu
]]><![CDATA[PI Website]]><![CDATA[Machine Learning Seminar Series Fall 2026 | Session VII]]>27863Abstract: TBA
Bio: Enric Boix-Adsera is an assistant professor of statistics and data science. His parents are originally from Barcelona, and he grew up mostly in Princeton, New Jersey. Before coming to Wharton, he earned a doctorate in electrical engineering and computer science from MIT.
Boix-Adsera studies the learning processes of artificial intelligence models, such as ChatGPT, with the goal of improving their efficiency and reliability.
For CODA guest access, please contact shatcher8@gatech.edu at least 2 business days prior to the event.
]]>2026-12-02T12:00:00-05:002026-12-02T13:00:00-05:002026-12-02T13:00:00-05:002026-12-02 17:00:002026-12-02 18:00:002026-12-02 18:00:002026-12-02T12:00:00-05:002026-12-02T13:00:00-05:00America/New_YorkAmerica/New_Yorkdatetime2026-12-02 12:00:002026-12-02 01:00:00America/New_YorkAmerica/New_Yorkdatetime<![CDATA[]]>For CODA guest access, please contact shatcher8@gatech.edu at least 2 business days prior to the event.]]><![CDATA[PI Website]]>