Abstract: Understanding training and generalization dynamics remains a central challenge in machine learning, particularly as neural networks, datasets, and learning tasks become increasingly complex in practice. However, existing theoretical characterizations remain largely restricted to simplified models or limiting regimes. In this talk, I will present a general convergence framework for gradient descent on a broad family of neural networks and data beyond the neural tangent kernel regime, together with a characterization of their generalization dynamics. This framework also helps explain the role of residual connections, differences in layerwise generalization behavior, and the grokking phenomenon.
Bio: Yuqing Wang is a Data Science Postdoctoral Fellow in the Department of Applied Mathematics and Statistics at Johns Hopkins University. Before that, she was a Research Fellow at the Simons Institute at the University of California, Berkeley. She received her Ph.D. in Mathematics from the Georgia Institute of Technology in 2024. Her research focuses on the theoretical foundations of deep learning theory, with an emphasis on characterizing the training dynamics. She is the recipient of multiple rising star awards and the Sigma Xi Best PhD thesis award.
]]>
Abstract: Recent work has identified incremental learning in shallow networks trained on single-index and multi-index models. However, existing analyses often rely on simplifying settings, such as small initialization, correlation loss, or layer-wise training. These choices reduce neuron interactions and leave some feature learning dynamics under standard initialization unexplored. We study training dynamics for polynomial-width two-layer networks learning orthogonal multi-index targets under standard initialization using polynomially many samples. We first prove that incremental learning still occurs: the loss decreases sequentially according to the Hermite expansion of the target, with lower-order components learned before higher-order components recover the individual target directions. In this standard initialization regime, training also shows a competitive reallocation of parameter mass: after the total mass fits the target mean and stabilizes, mass shifts into the target subspace and then concentrates on aligned neurons. Our theoretical analysis uses slightly modified gradient flow, while vanilla gradient descent empirically exhibits the same qualitative dynamics. Technically, we introduce a symmetry-based finite-width approximation via symmetrized networks, rather than comparing directly with an infinite-width limit. This yields better control of approximation errors and may be of independent interest.
Bio: Mo Zhou is currently a postdoctoral researcher at the Paul G. Allen School of Computer Science & Engineering at the University of Washington, working with Simon Du and Maryam Fazel. He received his Ph.D. from Duke University, where he was advised by Rong Ge. His research focuses on the mathematical foundations of machine learning, particularly the training dynamics and optimization of neural networks.
]]>Register Online to attend via Microsoft Teams
]]>From Virtual to Operational: XR Applications in Aviation
Delta Innovation
XR Bytes is a graduate-student-led initiative dedicated to showcasing the trajectory of Extended Reality (XR) research across the Institute. Our goal is to highlight the versatile ways XR technologies are applied across various disciplines and majors, fostering an interdisciplinary community of innovation at Georgia Tech.
We actively encourage cross-departmental participation and welcome engagement from both internal academic units and external industry partners. Our goal is to foster a robust community of practice around XR at Georgia Tech.
Lunch will be served for those who RSVP.
To join our mailing list and RSVP for these seminar series, please sign up here: RSVP HERE.
]]>How Does Active Exploration in Virtual Reality Shape Spatial Memory?
Ph.D. student in Adult Development and Aging, College of Sciences
XR Bytes is a graduate-student-led initiative dedicated to showcasing the trajectory of Extended Reality (XR) research across the Institute. Our goal is to highlight the versatile ways XR technologies are applied across various disciplines and majors, fostering an interdisciplinary community of innovation at Georgia Tech.
We actively encourage cross-departmental participation and welcome engagement from both internal academic units and external industry partners. Our goal is to foster a robust community of practice around XR at Georgia Tech.
Lunch will be served for those who RSVP.
To join our mailing list and RSVP for these seminar series, please sign up here: RSVP HERE.
]]>
Abstract: Monte carlo sampling algorithms have seen a resurgence in interest as a technique to efficiently generate good candidate solutions to non-linear optimization problems.
The Hamiltonian Markov chain Monte Carlo (HMC) algorithm is a good fit for this task because it enjoys mixing rates that are by its design robust to factors that can confound other MCMC methods.
A key paradigm to view quantum computing for optimization also treats it as a sampler of good candidate solutions. Existing quantum methodology shows that classical Markov chain methods (on discrete state spaces) enjoy a speed up in their mixing times when quantized. We construct a quantization scheme for HMC that extends this result while preserving its classical robustness. We sketch early results on applying this technique to the convex body volume estimation problem.
Bio: Soumyadip Ghosh is a member of IBM Research in the Algorithms and Applications group at Yorktown Heights. His research focuses on stochastic optimization and decision making under uncertainty, with a recent focus on constructing efficient hybrid quantum classical algorithms with applications to optimization formulations that arise in finance, insurance and healthcare. He has worked previously on smarter grid and energy analytics, cloud infrastructure, and supply chain management. He is an associate editor of Stochastic Models and has served as a committee member for INFORMS / WSC.
]]>Evaluating Lightweight Contact Feedback in Dual-Operator Collaborative Humanoid Teleoperation
Ph.D. student in Robotics, George W. Woodruff School of Mechanical Engineering, College of Engineering
XR Bytes is a graduate-student-led initiative dedicated to showcasing the trajectory of Extended Reality (XR) research across the Institute. Our goal is to highlight the versatile ways XR technologies are applied across various disciplines and majors, fostering an interdisciplinary community of innovation at Georgia Tech.
We actively encourage cross-departmental participation and welcome engagement from both internal academic units and external industry partners. Our goal is to foster a robust community of practice around XR at Georgia Tech.
Lunch will be served for those who RSVP.
To join our mailing list and RSVP for these seminar series, please sign up here: RSVP HERE.
]]>
Abstract: Abstract: Decision-making under uncertainties has been a subject of study for many years. In particular, the decision making (optimization) for complex systems such as engineering and economic systems involves human-decision phase. This triggers the occurrences of uncertainties in the decision making phase in addition to the widely-existed inherent uncertainties in the actual systems. In this talk, such challenges will be addressed by summarizing the year’s efforts of the presenter when he was with the University of Manchester before 2016. At first, a generic framework on the decision making for uncertainty systems has been formulated into a “closed loop control design problem”, where the uncertainties in both decision making phase and the actual systems have been represented by the uncertainty injections to the “controller” and to the “plant to be controlled”. The uncertainties in human decision making phase is characterized by their probability density functions (PDF) of the brain-cell’s signal collected via brain-computer-interface (BCI). Using such a generic framework, decision making under uncertainties are formulated as a PDF control problem of objective functions and constraints- leading to a total solution for the stochastic optimization for complex systems. Two case studies will be presented together with the rationale and claims that most of the existing stochastic optimizations are special cases of the presented theory. Note that the presentation is based upon the published materials by the author.
Bio: Professor Hong Wang (Member of European Academy of Sciences and Arts, FIEEE, FIET, FInstMC, SMIEEE) received the BSc form Anhui University of Science and Technology in 1982, the master’s and Ph.D. degrees from the Huazhong University of Science and Technology, Wuhan, China, in 1984 and 1987, respectively. He was a Research Fellow with Salford University, Salford, U.K., Brunel University, Uxbridge, U.K., and Southampton University, Southampton, U.K., before joining the University of Manchester Institute of Science and Technology (UMIST), Manchester, U.K., in 1992. He was a Chair Professor in process control of complex industrial systems with the University of Manchester (UoM), U.K., from 2002 to 2016, where he originated probability density function control theory in 1996 and was the Deputy Head of the Paper Science Department, the Director of the UMIST Control Systems Centre from 2004 to 2007, which is the birthplace of Modern Control Theory established in 1966. He was a University Senate member and a member of general assembly during his time in Manchester. Whilst he has been an Emeritus Professor of UoM, from 2016 to 2018, he was with the Pacific Northwest National Laboratory (PNNL), Richland, WA, USA, as a Laboratory Fellow and Chief Scientist,
and was the Co-Leader and the Chief Scientist for the Control of Complex Systems Initiative. He joined the Oak Ridge National Laboratory in January 2019.
His research focuses on stochastic distribution control, fault diagnosis and tolerant control, and intelligent controls with applications to complex system area. He has published 6 books and 200+ journal papers in these areas together with 10+ awards/prizes at international conferences.
He is an associate editor for IEEE Transactions on Neural Networks and Learning Systems and was an Associate Editor of IEEE Transactions on Automatic Control (2002 – 2004), IEEE Transactions on Control Systems Technology (2013 – 2018), and the IEEE Transactions on Automation Science and Engineering. He is also a member for three IFAC Committees for the three areas of his research.
Abstract: We study a class of nonconvex–nonconcave minimax problems in which the inner maximization problem satisfies a local Kurdyka–Łojasiewicz (KL) condition that may vary with the outer minimization variable. In contrast to the global KL or PL conditions commonly assumed in the literature—which are significantly stronger and often too restrictive in practice—this local KL condition accommodates a broader range of practical scenarios. However, it also introduces new analytical challenges. In particular, as an optimization algorithm approaches a stationary point, the region over which the KL condition holds may shrink, leading to a more intricate and potentially ill-conditioned landscape. To address this challenge, we show that the associated maximal function is locally generalized Hölder smooth. Leveraging this key property, we develop an inexact proximal gradient method for solving the minimax problem, where the inexact gradient of the maximal function is computed by applying a proximal gradient or sequential convex programming method to a KL-structured subproblem. Under mild assumptions, we establish complexity guarantees for computing an approximate stationary point of the minimax problem.
This is joint work with Xiangyuan Wang (University of Minnesota).
Bio: Zhaosong Lu is a Full Professor in the Department of Industrial and Systems Engineering at the University of Minnesota. He received his Ph.D. in Operations Research from Georgia Tech. His research focuses on the theory and algorithms of continuous optimization, with applications in data science and machine learning. Dr. Lu has published extensively in leading journals, and his research has been supported by major funding agencies, including AFOSR, NSF, and ONR. He has served on several prize committees, such as the INFORMS George Nicholson Prize Committee and the ICCOPT Best Paper Award Committee. He has also served as an Associate Editor for leading journals, including Mathematics of Operations Research, SIAM Journal on Optimization, Computational Optimization and Applications, and Journal of Global Optimization.
]]>Essentials of Negotiations and Stakeholder Influence level-sets the participants' understanding of negotiation influence and strengthens preparation, planning and execution activities involved with both simple and complex negotiations. The program includes industry techniques and tools for traditional supplier negotiations, as well as tips for internal cross-functional leadership. Participants walk away with a standard industry and customized individual experience which includes their personal Negotiation Style “DNA” to help them embrace their own natural tendencies and strengths. The program includes mock negotiations to reinforce techniques and tactics immediately in a “no judgement zone” environment.
This course is ideal for sourcing initiative leaders, project leaders, business unit leaders, operations managers, sales leaders and procurement & supply management-related professionals who are involved with supplier selection, contract development and supplier performance management.
Contracting and Legal Oversight provides participants with a holistic and integrated understanding of contract law, contract types, key industry standard contract terms, and contract structure to improve their confidence when creating or modifying contract documents. The program is geared to reinforce standards of excellence for professionals who are responsible for delivering contractual agreements and mitigating financial risk for their organization.
The online version of the course is comprised of (3) instructor-led LIVE group webinars, homework, and pre-work (e.g. installing and testing software on your computer, testing connectivity with Canvas LMS and BlueJeans meeting software, etc.) to be completed before the first day of the course.
This course is ideal for contract managers, procurement professionals, sourcing initiative leaders, project managers and all procurement & supply management-related professionals involved with bid contract development, contract execution or supplier performance management.
Category Management and Sourcing Leadership is designed to deepen participants' knowledge base of core activities in the procurement & supply management function. The program covers the sourcing process, specifications gathering, common bid package alternatives, cross-functional collaboration and supplier evaluation & selection. Participants will walk away ready to develop bid packages more thoroughly to help drive sourcing decisions for their organizations. This "hands on" delivery focuses on the professional serving as the main liaison between the buying organization and the selling organization in the company sourcing process.
This course is ideal for sourcing initiative leaders, procurement professionals, project managers, finance analyst, contract managers and all procurement & supply management-related professionals involved with bid package development, bid package analysis, negotiations preparation, contracting and supplier selection activity.
Abstract: Modern manufacturing generates increasingly rich in-situ sensing data, creating new opportunities for AI to enable automated and intelligent quality control decisions. However, conventional quality control using supervised learning relies on abundant labeled data and assumes that future product defects or process faults resemble those known during training. In practice, new defect types emerge, while abnormal conditions may be rarely observed or completely unknown. These challenges are particularly important for in-situ quality control, where defects must be detected or correctly classified for real-time decision-making, including newly emerging defects with limited or unavailable labels. Moreover, for latent defects that cannot be directly inspected online, defects must instead be predicted from indirect process-sensing signals. This requires mapping process-signal changes to possible defects despite scarce or unavailable defect training samples. This talk explores how advances in AI can address these challenges and enable more adaptive and intelligent in-situ quality control and decision-making for smart manufacturing.
Bio: Dr. Judy Jin is the A. Galip Ulsoy Collegiate Professor of Engineering and Professor of Industrial and Operations Engineering at the University of Michigan. Her research lies at the intersection of data science and quality engineering, with a focus on synergistically integrating engineering models, AI, and advanced quality control methods to improve system design and operational performance. She has served as PI/Co-PI on more than $20 million in federally and industry-funded research. Her work has received numerous honors, including 18 Best Paper Awards, the S.M. Wu Research Implementation Award from SME, the Forging Achievement Award from FIERF, the NSF CAREER Award, and the NSF PECASE Award.
Dr. Jin currently serves as Editor-in-Chief of IISE Transactions. She has also served as Vice President of INFORMS, Chair of the INFORMS Quality, Statistics and Reliability Section, and President of the IISE Quality Control and Reliability Engineering Division. She is a Fellow of ASME, IISE, and INFORMS.
This course provides a deep dive into the ways in which artificial intelligence (AI) optimizes supply chain efficiency. Participants will explore generative AI fundamentals, prompt engineering, and practical applications such as automated inventory systems, predictive maintenance, and route optimization. The course also covers ethical AI use, good and bad use of generative AI (GenAI), and rapidly emerging use cases. By the end, professionals will be skilled in applying AI to enhance supply chain processes and drive success in their organizations.
This course targets supply chain managers, data analysts, logistics professionals, procurement specialists, and business leaders aiming to harness GenAI for enhanced supply chain operations. It is ideal for those interested in GenAI-driven efficiency, strategic insights, and navigation of GenAI's role in transforming supply chain processes.
This course provides a deep dive into the ways in which artificial intelligence (AI) optimizes supply chain efficiency. Participants will explore generative AI fundamentals, prompt engineering, and practical applications such as automated inventory systems, predictive maintenance, and route optimization. The course also covers ethical AI use, good and bad use of generative AI (GenAI), and rapidly emerging use cases. By the end, professionals will be skilled in applying AI to enhance supply chain processes and drive success in their organizations.
This course targets supply chain managers, data analysts, logistics professionals, procurement specialists, and business leaders aiming to harness GenAI for enhanced supply chain operations. It is ideal for those interested in GenAI-driven efficiency, strategic insights, and navigation of GenAI's role in transforming supply chain processes.