Speaker: Bernard Cathelain (Member of the Executive Board of Société des Grands Projets)

• Introduction: Learn about the Grand Paris Express, the massive transit project that will improve access and connectivity throughout the Ille de France region and Paris.  We will hear from Bernard Cathelain, who oversees the design and construction program for the Grand Paris Express transport project, as well as the industry, purchasing and environmental strategy functions, and Pierre-Emmanuel Becherand, Head of Design, Arts and Urban Development for the project. The Grand Paris Express comprises four new lines for the Paris Métro, plus extensions of the existing Lines 11 and 14.  A total of 200 kilometers (120 miles) of new tracks and 68 new stations are to be added, serving a projected 2 million passengers a day.
• Time: 11:00 AM – 12:30 PM | Thursday | April 24, 2025
• Meeting room: DUSP City Arena, MIT Building 9
• Bio: Bernard Cathelain, an engineer by training, is a graduate of the École polytechnique (class of 1980) and the École nationale des ponts et chaussées (1985). He began his career in 1985 as a project manager in the Ports and Waterways division of the engineering company BCEOM. The following year, he joined the Val d’Oise Departmental Equipment Department, where he was successively in charge of the functional district, in charge of the major works department, and then in charge of transport, traffic and IT at the Île-de-France Regional Council. In 1993, he joined Société des Autoroutes du Nord et de l’Est de la France (SANEF), where he was appointed Director of Construction, then Director of Engineering, Development and Environment in 1998. In 2001, he joined Aéroports de Paris (ADP) as head of the project management department. He became director of major works in 2003, before being appointed deputy managing director in charge of planning and development in 2008, a position he held until 2014. During this period, he was also a member of ADP’s Executive Committee and Chairman of the Board of Hub Télécom (now Hub One) from 2012 to 2014. On March 31, 2015, Bernard Cathelain was appointed to the Executive Board of Société des Grands Projets (formerly Société du Grand Paris). In this role, he oversees the design and construction program for the Grand Paris Express transport project, as well as the industry, purchasing and environmental strategy functions.

Speaker: Prof. Jakob Runge (Professor of Data Science (W3) at TU Dresden)

• Topic (TBD): Causal Inference from Time Series in Spatiotemporal Systems

April 24, 2025

Speaker: Zhe Fu (Ph.D. candidate at the University of California, Berkeley)

• Topic: Physics-Informed Learning and Control for Intelligent Transportation: Theory, Algorithms, and Experimental Validations
• Abstract: The rapid growth of data science is reshaping how we model and control physical infrastructure systems. Traditional PDE-based methods provide structured interpretability, whereas purely data-driven neural networks offer flexibility but often lack adherence to physical principles. In this talk, I will present a physics-informed learning and control framework that combines PDE-based modeling with neural networks, enabling improved understanding and predictive accuracy of transportation system dynamics. Specifically, I will introduce a Neural Finite Volume Method (NFVM) that preserves crucial physical properties, effectively bridging physics and data-driven approaches. Motivated by the potential of leveraging this improved understanding to influence real transportation systems, I developed control strategies using a small number of “leader” vehicles to guide traffic flow toward greater efficiency and lower energy consumption, with minimal system-wide intervention. These approaches include a kernel-based control method and an imitation learning strategy, with variations validated in a large-scale operational field experiment involving 100 autonomous vehicles. I will conclude by highlighting ongoing comparative studies to quantify how incorporating physics-informed modeling further enhances control performance in terms of efficiency, safety, and robustness.
• Bio: Zhe Fu is a Ph.D. candidate in Transportation Engineering and an M.S. candidate in Electrical Engineering and Computer Sciences (EECS) at the University of California, Berkeley. Her research lies at the intersection of transportation systems, control theory, and machine learning, with the goal of enabling intelligent and energy-efficient mobility in mixed autonomy environments. She has been recognized as a 2025 Eno Fellow and has received several national honors, including the Rising Stars in NSF CPS Award (2025), Rising Stars in Mechanical Engineering Award (2024), First Place Winner in the INFORMS Best Poster Competition (2023) and Runner-up Winner in Berkeley Grad Slam (2025). Her leadership, mentorship, and teaching efforts have been recognized by UC Berkeley and external organizations such as ITS/CTF, EDGE in Tech, H2H8 and AAa/e.

April 17, 2025

Speaker: Dr. Yilei Shao (Professor, Dean of Shanghai AI-Finance School, East China Normal University; Director, United Nations University – ECNU Hub in AI-Finance)

• Topic: China’s AI Trajectory: From Labs to the Silicon Economy
• Abstract: China is emerging as a global leader in artificial intelligence, propelled by innovative academic research, dynamic industrial applications, and a rising Silicon-based Economy driven by massive data resources and advanced infrastructure building. Framed within China’s techno-industrial policy and digital governance, the lecture highlights how AI and silicon innovation are reshaping China’s economic strategy, technological sovereignty, and global influence.We will also examine the importance of global collaborative efforts bridging academia and industry, while also addressing challenges such as ethical considerations, talent cultivation, and international collaboration.
• Bio: Professor Yilei Shao, Founding Dean of Shanghai AI-Finance School (SAIFS) at ECNU, Director of United Nations University-ECNU Hub in AI-Finance. She holds a Ph.D. degree in Computer Science from Princeton University. She was the Secretary General of the International Joint Conference on Artificial Intelligence (IJCAI) China Office from 2021 to 2023. In her previous experience, she has worked at Goldman Sachs’s headquarters in New York. As a cross-disciplinary pioneer, she not only focuses on the application of AI in finance, but also actively explores its innovation integration in the fields of art and education, as well as other AI for social science issues, such as ethics and governance of AI technology.

April 10, 2025

Speaker: Dr. Xuan Jiang (PhD at University of California, Berkeley)

• Topic: Sparse Matrix Tuning: Efficient Fine-Tuning of Large Language Models
• Abstract: The recent surge in large language models (LLMs) has revolutionized natural language processing by delivering remarkable generalization capabilities. However, adapting these models to domain-specific tasks through fine-tuning remains a computational bottleneck, primarily due to the high memory and compute demands of updating all model parameters. In this talk, Dr. Xuan Jiang introduces Sparse Matrix Tuning (SMT), a novel parameter-efficient fine-tuning strategy that dramatically reduces both computational and memory overhead. SMT identifies and selectively updates the most critical sparse sub-matrices within pre-trained weight matrices—specifically targeting the attention mechanisms in LLMs—while freezing less essential layers. Drawing inspiration from gradient-based parameter selection techniques (informed by Fisher Information) as well as activation-based strategies, SMT overcomes the performance plateau observed in existing low-rank adaptation methods (e.g., LoRA and DoRA). Experimental evaluations on benchmarks such as Commonsense Reasoning and Math10K demonstrate that SMT not only bridges the accuracy gap with full fine-tuning but also achieves substantial speedups (up to 14.6×) and memory savings. This work opens a new pathway for scalable fine-tuning of LLMs, making high-performance adaptation accessible even on resource-constrained hardware.
• Bio: Dr. Xuan Jiang is a Ph.D. at the University of California, Berkeley, specializing in Transportation Engineering with a minor in high performance computing and AI. His interdisciplinary research spans high-performance parallel computing, deep learning, and scalable optimization techniques for large-scale systems. Dr. Jiang has contributed to advancing simulation-based optimization and efficient model fine-tuning, with a strong focus on bridging theoretical innovations and practical implementations. His work has been recognized in high-impact journals and conferences, and he is actively involved in both academic research and industry collaborations.

March 27, 2025

Speaker: Dr. Shenhao Wang (Assistant Professor in Artificial Intelligence at the University of Florida)

• Topic: Artificial Intelligence for Travel Demand Modeling and Generative Urban Planning
• Abstract: This talk introduces two early-stage research projects in the Urban AI lab at the University of Florida. The first project seeks to synergize graph neural networks and classical discrete choice models, thus developing graph-based deep choice models that improve the computational efficiency, interpretability, and application potentials in travel behavioral analysis. It transforms classical nested logic models through an innovative concept of alternative graph, which enables flexible nest specifications and connects individual analysis to complexity science. The second project focuses on the interaction between imagery and language, particularly with applications to design and planning process. It automatically generates urban imagery to describe urban landscape and street views, and further grounds urban development into planning documents with vision-language models. Besides the two concrete projects, the talk will also present visionary thinking into how AI can transform travel demand modeling and urban planning practices through the frameworks of deep choice modeling and vision-language integration.
• Bio: Dr. Shenhao Wang is an assistant professor and the director of the Urban AI Laboratory at the University of Florida. Dr. Wang completed his interdisciplinary Ph.D. in Computer and Urban Science at Massachusetts Institute of Technology in 2020. He received a Bachelor of Economics from Peking University (2014) and Bachelor of Architecture from Tsinghua University (2012), as well as a Master of Science in Transportation, and Master of City Planning from MIT (2017). As an urban and computer scientist, he develops novel AI approaches to focus on three research themes: (1) resilient and equitable urban systems, (2) travel behavior, and (3) design and planning automation with generative AI. Dr. Wang’s research centers on pioneering interdisciplinary work at the intersection of urban science, network dynamics, and artificial intelligence, driving theoretical innovation, educational advance, and practical social impact. As the Founding Director of Urban AI Lab, Dr. Wang leads to create a more sustainable, intelligent, and equitable urban future with artificial intelligence. His multi-million-dollar research grant has been funded by U.S. Department of Energy (DOE), Singapore-MIT Alliance for Research and Technology (SMART), University of Florida, as well as industrial and institutional partners. His publications has been published on Nature Cities, Nature Communications, Transportation Research Part A, Part B & Part C, IEEE Transactions on Intelligent Transportation Systems, Proceedings of ACM Knowledge Discovery and Data Mining, and top-tier journals and conferences.

March 20, 2025

Speaker: Dr. Frederike Dümbgen (Researcher at Inria Paris)

• Topic: Globally Optimal Solvers for Scalable Physical Intelligence
• Abstract: The past decades have witnessed significant advances in physical intelligence, enabling quadrupeds to navigate challenging terrains, humanoids to show remarkable feats, and robotic hands to display increasing dexterity. As we aim to deploy such systems to tackle real-world problems at scale, critical questions about the efficiency and robustness of the algorithms used to create these systems emerge. Approaches based on physical models demand substantial expertise for model selection, calibration, and real-time solvers. Purely data-driven methods are easier to set up and deploy but necessitate expensive data collection and face well-known safety concerns.
  In this talk, I argue that leveraging more capable global optimization techniques can address both approaches’ challenges. Recent breakthroughs have demonstrated that many non-convex problems encountered in robotics can be solved to global optimality in polynomial time using convex relaxations. Despite their promise, these approaches remain underutilized in robotics, partly due to their cumbersome mathematical formulation, difficulty meeting real-time constraints, and limitation to relatively simple models. I will present our recent efforts to overcome these barriers, including an automatic tool for problem formulation, efficient solvers that exploit sparsity, and integrating global solvers in end-to-end learned pipelines to improve learning. My vision is that by embracing global optimality for robotics, we can develop solutions that are not only capable but also safe and efficient, enabling us to address the critical challenges of our time.
• Bio: Frederike Dümbgen is a researcher in the WILLOW team of Inria Paris, where she has been working on optimization for robotics since May 2024. Before joining Inria, she spent two years as a postdoctoral fellow at the Robotics Institute of the University of Toronto, collaborating with Prof. Timothy D. Barfoot on certifiable optimization. Frederike holds a Ph.D. in Computer and Communication Sciences from École Polytechnique Fédérale de Lausanne (EPFL), Switzerland, where her research focused on developing systems for non-visual spatial perception using diverse sensor modalities. She earned her B.Sc. and M.Sc. in Mechanical Engineering from EPFL in 2013 and 2016, respectively, with a minor in Computational Science and Engineering. She performed her Master’s thesis at the Autonomous Systems Lab of ETH Zürich and gained experience in industry as an intern of Disney Research and ABB, among others. She was recognized as a R:SS Pioneer in 2024, as Google’s Women Techmarker in 2020, and has been co-chair of the RAS technical committee on model-based optimization since 2024.

March 13, 2025

Speaker: Riccardo Fiorista (PhD Student at the MIT JTL Urban Mobility Lab)

Speaker: Dingyi Zhuang (PhD Student at the MIT JTL Urban Mobility Lab)

March 6, 2025

Speaker: Han Wang (PhD Candidate at UC Berkeley)

• Topic: From Connected to Coordinated – Distributed Intelligence and Centralized Coordination for Connected Autonomous Vehicles
• Abstract: As connected and autonomous vehicles (CAVs) become increasingly integrated into modern transportation systems, the transition from mere connectivity to full coordination is essential for achieving efficiency, safety, and resilience. This seminar explores how distributed intelligence and centralized coordination can optimize traffic flow in mixed autonomy conditions, leveraging insights from the CIRCLES project—the world’s largest open-track field experiment deploying 100 CAVs on the I-24 MOTION testbed. We will discuss how reinforcement learning and real-time control mechanisms enable AVs to mitigate congestion with minimal penetration rates, achieving an 8% improvement in overall traffic efficiency. The session will also highlight key technical challenges, such as multi-agent decision-making, human-vehicle interactions, and large-scale deployment strategies.
• Bio: Han Wang is a Ph.D. candidate at UC Berkeley, affiliated with Berkeley Artificial Intelligence Research (BAIR) and Berkeley Deep Drive (BDD). In the CIRCLES project, he lead the developing real-time speed optimization for 100 CAVs and the full-backend development in the largest open-road field test to improve traffic flow. At BDD, He led the Predictive Occupancy Risk Assessment (PORA) project, creating probabilistic risk metrics and simulation-based safety evaluation for AVs in mixed traffic. Beyond his doctoral research, he explores AR/VR and LLM applications in simulation and game development, with awarded publication in Human-Computer Interaction(HCI) conferences.

Speaker: Nick Zolman (PhD Candidate at the University of Washington)

• Topic: SINDy-RL: Interpretable and Efficient Model-Based Reinforcement Learning
• Abstract: Deep Reinforcement Learning (DRL) has shown significant promise for uncovering sophisticated control policies that interact in environments with complicated dynamics, such as stabilizing the magnetohydrodynamics of a tokamak reactor and minimizing the drag force exerted on an object in a fluid flow. However, these algorithms require many training examples and can become prohibitively expensive for many applications. In addition, the reliance on deep neural networks results in an uninterpretable, black-box policy that may be too computationally challenging to use with certain embedded systems. Recent advances in sparse dictionary learning, such as the Sparse Identification of Nonlinear Dynamics (SINDy), have shown to be a promising method for creating efficient and interpretable data-driven models in the low-data regime. In this work, we extend ideas from the SINDy literature to introduce a unifying framework for combining sparse dictionary learning and DRL to create efficient, interpretable, and trustworthy representations of the dynamics model, reward function, and control policy. We demonstrate the effectiveness of our approaches on benchmark control environments and challenging fluids problems, achieving comparable performance to state-of-the-art DRL algorithms using significantly fewer interactions in the environment and an interpretable control policy orders of magnitude smaller than a deep neural network policy.
• Bio: Nick Zolman is a current PhD student in the Mechanical Engineering Department at the University of Washington, co-advised by Steve Brunton and Nathan Kutz, and is affiliated with the NSF AI Institute in Dynamic Systems. His research interests lie at the intersection of machine learning, dynamics, and control—with a particular emphasis on reduced order modeling, interpretability, and operations in the low-data limit. After graduating with a BS in Mathematics from the California Institute of Technology in 2016, he worked as a data scientist in the aerospace industry for over 5 years; examples of his industry work include building lithium-ion battery digital twin models to estimate battery health on-orbit and developing methods to extract low-dimensional representations of dynamics from high-dimensional scenes in the absence of ground truth state-measurements.
• Related paper: SINDy-RL: Interpretable and Efficient Model-Based Reinforcement Learning
• Material: GitHub | YouTube
• More sources: Data-Driven Dynamics and Controls [Brunton Lab] [Kutz Lab]

Host: Dr. Xinyu Chen (Postdoc at the MIT JTL Urban Mobility Lab)

Commentator: Nina Cao (PhD Candidate in Mechanical Engineering at MIT)

February 27, 2025

Speaker: Dr. Ryan Qi Wang (Associate Professor in the Department of Civil and Environmental Engineering at Northeastern University)

• Topic: Human Mobility and Urban Inequality: Unraveling the Links Between Movement, Air Quality, and Justice
• Abstract: Cities today face pressing challenges, from large-scale crises and climate change to aging populations. These issues deeply impact human mobility, reshaping how people move, interact, and access essential services. This talk explores the intersection of urban mobility, environmental justice, and social inequality, drawing insights from research on mobility data, air quality exposure, and accessibility disparities. Using case studies such as the BostonWalks study, we analyze how mobility patterns shape disparities in pollution exposure, access to services, and resilience to shocks like COVID-19 and extreme weather events. By integrating mobility data with environmental and demographic insights, we uncover systemic inequalities in urban spaces, showing how marginalized groups often bear a disproportionate burden. This talk will highlight key findings on mobility rhythms, disparities in pollution exposure, and service accessibility for aging populations, offering a data-driven perspective on urban inequality.
• Bio: Ryan Qi Wang is an Associate Professor and Vice Chair for Research in the Department of Civil and Environmental Engineering at Northeastern University. Before joining Northeastern, Wang was a postdoc fellow at the Department of Sociology, Harvard University. He received his Ph.D. degree from the Department of Civil and Environmental Engineering at Virginia Tech. His research focuses on two interrelated areas: human movement perturbation under the influence of natural and manmade disasters, and mobility equality in big cities. His research has been published in Nature Human Behavior, Proceedings of National Academy of Sciences (PNAS), etc. His research group has received funding support from NSF, NIST, IARPA, MacArthur Foundation, USDOT, and other foundations and local government agencies.

February 13, 2025

Speaker: Prof. Mi Diao (Professor in the College of Architecture and Urban Planning at Tongji University)

• Topic: Electric Vehicle Adoption in China: Implications for Road Congestion and Carbon Emission Reductions
• Abstract: Electric vehicles (EVs) have been widely considered a pivotal strategy for reducing transport emissions and combating global climate change. However, while EVs emit less carbon per kilometer, their lower operating costs may incentivize increased vehicle usage, potentially exacerbating traffic congestion and offsetting some of their environmental benefits. This study examines the implications of rapid EV adoption on road congestion and carbon emission reductions in China, the world’s largest EV market and producer. Using a panel dataset from 49 major Chinese cities, we assess the causal impact of EV adoption on traffic congestion. Our results show that a 1% increase in the total number of EVs results in a 0.11% rise in traffic congestion. We identify a significant offset effect: Although EVs reduce operational carbon emissions, these gains are partially counteracted by additional emissions from internal combustion engine (ICE) vehicles due to worsened congestion. Our findings highlight the critical need for well-designed policies that not only promote EV adoption but also address the unintended consequences of increased vehicle usage, contributing to the goal of decarbonizing the transportation sector.
• Bio: Mi Diao is a professor in the College of Architecture and Urban Planning at Tongji University, China. Prior to this position, he was an assistant professor in the Department of Real Estate and a senior research fellow in the Institute of Real Estate and Urban Studies at National University of Singapore. Mi Diao received his PhD in Urban and Regional Planning and Master in City Planning from Massachusetts Institute of Technology (MIT), USA, and Bachelor’s and Master’s degrees in Architecture from Tsinghua University, China. At the nexus of urban planning, urban economics, and urban technology, Dr. Diao applies urban economics theories, emerging big data, and new analytics in tackling urban challenges. His research has appeared in leading academic journals such as Nature Sustainability, Journal of Urban Economics, Regional Science and Urban Economics, Urban Studies, Journal of Planning Education and Research, Transportation Research Part A, C, D, and Environment and Planning A, B.

Speaker: Chengyuan Zhang (PhD Candidate at McGill University)

• Topic: Stochastic Modeling and Simulations of Car-Following Behaviors
• Abstract: This talk presents a stochastic modeling and simulation framework designed to enhance the accuracy and realism of car-following models. By integrating time-series techniques with traditional car-following models, the framework addresses key limitations of conventional approaches, which often overlook historical driving behavior and assume independent errors. Within this framework, three models are introduced, leveraging the Intelligent Driver Model (IDM) or deep neural networks to capture behavioral uncertainty, while Gaussian processes (GPs) or autoregressive (AR) processes model temporal correlations. By incorporating past driving actions, this approach enables more realistic and reliable simulations of human driving behavior. Experiments conducted on large-scale naturalistic driving datasets validate the framework’s effectiveness, demonstrating its ability to generate enhanced probabilistic predictions and more realistic traffic flow simulations, ultimately offering deeper insights into driving dynamics.
• Bio: Chengyuan Zhang is a Ph.D. candidate at McGill University, supervised by Prof. Lijun Sun. He was a visiting researcher at the Robotics Institute, Carnegie Mellon University in 2023 and at the Department of Mechanical Engineering, UC Berkeley from 2019 to 2020. His research focuses on Bayesian learning, spatiotemporal modeling, traffic flow theory, and multi-agent interaction modeling within intelligent transportation systems. He aims to bridge the gap between theoretical modeling and practical traffic simulation using advanced statistical techniques. Driven by a passion for understanding human driving behavior, his work seeks to enhance microscopic traffic simulations, ultimately contributing to safer and more efficient transportation systems.
• Material: Slides or see below.

February 6, 2025

---

Past Events (2024)

Time: 8:30 AM-9:30 AM (Boston time) every Wednesday from September 18, 2024 to December 11, 2024.

December 11, 2024

Speaker: Dr. Zhanhong Cheng (Postdoc at the University of Florida)

• Topic: Residential Location Choice with Graph Neural Networks
• Bio: Zhanhong Cheng is now a Postdoctoral Associate at Urban AI Lab, University of Florida, USA. Prior to joining UF, he received PhD in Civil Engineering (Transportation) from McGill University, Canada. He is interested in understanding, planning, and optimizing urban mobility systems, along with the associated infrastructure and human behavior using expertise in artificial intelligence (AI), machine learning (ML), and transportation engineering. Currently, his research centers three interrelated areas: public transit (e.g., destination and OD matrix inference, travel time and demand forecasting), multimodal travel behavior (e.g., travel patterns in metro, bike-sharing, and E-taxi), and spatiotemporal data modeling (e.g., forecasting and imputation). Through his research, he aims to contribute to creating transportation systems that are more sustainable, efficient, and accessible.
• Material: Slides

November 27, 2024

Speaker: Dr. Eren Inci (Professor of Economics at Sabanci University)

• Topic: Price of Parking and Everything Else in Cities
• Abstract: In urban centers, parking emerges not merely as a commodity, but as a vital conduit, facilitating economic exchanges by connecting people to city offerings. The need for parking consumes significant portions of land, transforming these areas into guardians of mobility. As an important intermediate good and a voracious consumer of land, parking stands as a linchpin in the city’s economic machinery. Failure to accurately gauge its value risks distorting pricing across all sectors, casting a shadow of inefficiency over the entire urban economy. This essay explores illustrative examples from the author’s “Shoupista” work, examining how parking mispricing echoes through the urban landscape, distorting transportation costs, shaping retail price dynamics, and influencing the delicate price balance of housing markets.
• Bio: Eren Inci is a professor of economics at Sabanci University. He completed his undergraduate education in Management Engineering at Istanbul Technical University in 2002. He received his MA in Economics from Boston College in 2004 and his PhD in Economics in 2007. His research spans the fields of Urban Economics and Transportation Economics, aiming to enhance the livability and sustainability of cities. His work, particularly focused on the economic analysis of parking spaces, has contributed to the interdisciplinary study of traffic congestion and parking in cities. He has authored influential works on the development and analysis of urban parking policies. In 2014, he was awarded the Outstanding Young Scientist Award (TUBA-GEBIP) by the Turkish Academy of Sciences and the Young Scientist Award (BAGEP) by the Science Academy. In 2018, he received the TUBITAK Incentive Award. He served as the Vice Dean of Faculty of Arts and Social Sciences at Sabanci University between 2017 and 2024. He was an executive committee member of the International Transportation Economics Association from 2016 to 2024. He has held visiting roles at the Center for European Economic Research (ZEW), University of California, Riverside, University of British Columbia, and Massachusetts Institute of Technology.

November 20, 2024

Speaker: Dr. Mehrdad Ghadiri (Postdoc at MIT)

• Topic: Approximately Optimal Core Shapes for Tensor Decompositions
• Abstract: This work studies the combinatorial optimization problem of finding an optimal core tensor shape, also called multilinear rank, for a size-constrained Tucker decomposition. We give an algorithm with provable approximation guarantees for its reconstruction error via connections to higher-order singular values. Specifically, we introduce a novel Tucker packing problem, which we prove is NP-hard, and give a polynomial-time approximation scheme based on a reduction to the 2-dimensional knapsack problem with a matroid constraint. We also generalize our techniques to tree tensor network decompositions. We implement our algorithm using an integer programming solver, and show that its solution quality is competitive with (and sometimes better than) the greedy algorithm that uses the true Tucker decomposition loss at each step, while also running up to 1000x faster.
• Related paper: Approximately Optimal Core Shapes for Tensor Decompositions [Slides]

November 13, 2024

Speaker: Dr. Filipe Rodrigues (Associate Professor at the Technical University of Denmark)

• Topic: Reinforcement Learning for Network Optimization in Transportation
• Abstract: Optimization problems over dynamic networks have been extensively studied and widely used in the past decades to formulate numerous real-world transportation problems. However, traditional optimization-based approaches do not scale to large networks, and the design of good heuristics or approximation algorithms often requires significant manual trial-and-error. In this talk, I argue that data-driven strategies can automate this process and learn efficient algorithms without compromising optimality. To do so, I will introduce network control problems through the lens of reinforcement learning and propose a bi-level graph-network-based framework, where we (1) specify a desired next state via RL, and (2) solve a convex program to best achieve it, leading to drastically improved scalability and performance. The benefits of the proposed approach will be demonstrated for Autonomous-Mobility-on-Demand rebalancing and dynamic pricing, and inventory management applications. I will then introduce a framework for offline RL of these bi-level/hierarchical from datasets generated by arbitrary behavior policies (historical data). The learned policies are shown to significantly outperform end-to-end offline RL in terms of performance and robustness.
• Bio: Filipe Rodrigues is associate professor at the Technical University of Denmark (DTU) in the Machine Learning for Smart Mobility (MLSM) lab, where his research is primarily focused the use of machine learning for understanding and optimizing urban mobility and human behaviour. Previously, he was a H.C. Ørsted / Marie-Skłodowska Curie Actions (COFUND) postdoctoral fellow, also at DTU, working on spatio-temporal models of mobility demand with emphasis on modelling uncertainty and the effect of special events. He has published more than 60 articles in leading conferences and journals in both transportation and machine learning. His research interests span machine learning, reinforcement learning, intelligent transportation systems, and urban mobility.

Speaker: Dr. T. Konstantin Rusch (Postdoc at the MIT CSAIL Lab)

• Topic: Message-Passing Monte Carlo: Generating low-discrepancy point sets via Graph Neural Networks
• Abstract: Discrepancy is a well-known measure for the irregularity of the distribution of a point set. Point sets with small discrepancy are called low discrepancy and are known to efficiently fill the space in a uniform manner. Low-discrepancy points play a central role in many problems in science and engineering, including numerical integration, computer vision, machine perception, computer graphics, machine learning, and simulation. In this work, we present a machine learning approach to generate a new class of low-discrepancy point sets named Message-Passing Monte Carlo (MPMC) points. Motivated by the geometric nature of generating low-discrepancy point sets, we leverage tools from Geometric Deep Learning and base our model on graph neural networks. We further provide an extension of our framework to higher dimensions, which flexibly allows the generation of custom-made points that emphasize the uniformity in specific dimensions that are primarily important for the particular problem at hand. Finally, we demonstrate that our proposed model achieves state-of-the-art performance superior to previous methods by a significant margin. In fact, MPMC points are empirically shown to be either optimal or near-optimal with respect to the discrepancy for low dimension and small number of points, i.e., for which the optimal discrepancy can be determined.
• Related paper: Message-Passing Monte Carlo: Generating low-discrepancy point sets via graph neural networks

November 6, 2024

Speaker: Dr. Qiusheng Wu (Associate Professor in the Department of Geography & Sustainability at the University of Tennessee, Knoxville)

• Topic: Open-Source Python Package and Software Development in Geospatial Data Science
• Abstract: The rapid growth of geospatial data and the increasing demand for open-source solutions have significantly reshaped the field of geospatial data science. This presentation explores the development and practical application of open-source Python packages for geospatial analysis. We will highlight key tools such as Geemap, Leafmap, and SAMGeo, which streamline workflows for data processing, visualization, and analysis, enabling researchers, developers, and organizations to effectively leverage spatial data. Furthermore, we will discuss best practices for open-source software development in the geospatial domain, covering essential topics such as package templates, documentation tools, integration testing, and strategies for fostering community engagement. Attendees will gain valuable insights into enhancing their geospatial projects through the effective use of open-source resources.
• Bio: Qiusheng Wu is an Associate Professor and the Director of Graduate Studies in the Department of Geography & Sustainability at the University of Tennessee, Knoxville. He also serves as an Amazon Visiting Academic. Dr. Wu specializes in geospatial data science and open-source software development, with a research focus on utilizing big geospatial data and cloud computing to study environmental changes, particularly in surface water and wetland inundation dynamics. He is the creator of several widely used open-source Python packages, including geemap, leafmap, and segment-geospatial, which are designed for advanced geospatial analysis and visualization. Explore his open-source contributions on GitHub at https://github.com/opengeos.

October 16, 2024

Speaker: Dr. Vassilis Digalakis Jr (Assistant Professor of Operations Management, HEC Paris)

• Topic: Slowly Varying Regression under Sparsity
• Abstract: Stability in machine learning (ML) is important for ensuring consistent performance and enhancing interpretability. For example, in healthcare, patient mortality risk prediction models may change upon retraining with new batches of data, but are likely to do so in a “smooth” way; in sustainability, the same holds true for energy consumption prediction over consecutive hours of the day or air quality prediction over adjacent spatial areas. Abrupt changes in a model’s structure or the resulting analytical insights can lead to hesitation for adoption. In this talk, I will present my research on stable ML through the lens of industry collaborations in healthcare and sustainability. First, I will introduce the framework of slowly varying regression under sparsity, allowing sparse regression models to exhibit controlled variations under some temporal, spatial, or general graph-based structure. Assessing stability in decision trees presents new challenges compared to regression models: to address this, I will propose a novel distance metric for decision trees and use it to determine a tree’s level of stability. Finally, I will present a model-agnostic framework to stabilize interpretable models’ structures or black-box models’ insights upon retraining with new data. We have tested the proposed methodologies on numerous real-world case studies and have shown that a controlled—and often negligible—decrease in predictive power can significantly improve the models’ stability and interpretability.
• Bio: Vassilis Digalakis Jr. is an Assistant Professor of Operations Management at HEC Paris. He completed his PhD in operations research at MIT in 2023. His research addresses the gap in analytics and ML adoption in high-stakes applications like healthcare and sustainability, where the lack of transparency in decision-making is a barrier. Leveraging tools from optimization, he develops “trustworthy” analytics and ML methodologies—ensuring that models exhibit characteristics such as interpretability, stability and robustness, fairness, and privacy. He has collaborated, among others, with OCP, the world’s largest producer of phosphate products, to develop a robust optimization framework guiding their $2Bn investment in renewable energy, and with FEMA to help them fairly decide on locations for COVID-19 mass vaccination centers. His research has been published, among others, in Operations Research and M&SOM, and has earned awards, including the 2023 Harold Kuhn Award, finalist in the 2023 M&SOM Practice-Based Research Competition, and the 2021 INFORMS Pierskalla Award.
• Related paper: Slowly Varying Regression Under Sparsity