AIX (Academia-Industry X) hosts AI-related events and workshops connecting academia and industry, fostering collaboration to drive innovation, facilitate communication, and advance AI and emerging technologies.
Abstract: We consider the problem of learning to exploit learning algorithms through repeated interactions in games. Specifically, we focus on the case of repeated two player, finite-action games, in which an optimizer aims to steer a no-regret learner to a Stackelberg equilibrium without knowledge of its payoffs. We first show that this is impossible if the optimizer only knows that the learner is using an algorithm from the general class of no-regret algorithms. This suggests that the optimizer requires more information about the learner's objectives or algorithm to successfully exploit them. Building on this intuition, we reduce the problem for the optimizer to that of recovering the learner's payoff structure. We demonstrate the effectiveness of this approach if the learner's algorithm is drawn from a smaller class by analyzing two examples: one where the learner uses an ascent algorithm, and another where the learner uses stochastic mirror ascent with known regularizer and step sizes.
Bio: Yizhou Zhang is a PhD student at CMS dept in Caltech.
Title: Breaking into Academic Startups and Entrepreneur Ventures
Abstract: Translating research into real-world solutions has been the core theme of my work at the intersection of biomedical engineering, microfluidics, AI, and point-of-care diagnostics. In this talk, I will share my journey developing automated biomedical devices, from lab prototypes to market-oriented technologies. I will also provide an overview of common fundraising pathways for academic entrepreneurs, including angel investors, friends and family rounds, venture capital, and strategic collaborations with industry partners. To ground the discussion, I will present my latest startup effort: developing a Miniature AI Robot Scientist, designed to automate the full experimental workflow on a single benchtop. This compact, intelligent system aims to streamline R&D processes for small labs and early-stage biotech teams. Finally, I will reflect on the unique advantages and inherent challenges of launching a startup from an academic setting—balancing innovation, resource constraints, and the steep learning curve of entrepreneurship.
Bio: Shun Ye is the Founder of eBot Bio, which aims to cut R&D costs by 95%. He is also a PhD Candidate in the Bioengineering and Electrical Engineering departments and California NanoSystems Institute (CNSI) at UCLA. Before joining UCLA, he earned a Master's degree in Biomedical Engineering from Penn State University and completed a research fellowship in Analytical Microfluidics at the Chinese Academy of Sciences. Shun has six years of experience in biomedical instrumentation, automation, and AI. He has eight journal publications and four issued patents. He developed the OsciDrop technology, which was featured as a cover article in Analytical Chemistry and has been commercialized as the world's first all-in-one, automatic, chip-free digital PCR machine (D600, Maccura Biotech Co., Ltd.), with six fluorescent channels, and has received $20M in Venture Capital. He is currently co-supervised by pioneering medical device scholars Prof. Dino Di Carlo and Aydogan Ozcan. His doctoral research focuses on developing and translating novel biomedical devices and micro/nanotechnologies for AI-powered lab automation, point-of-care diagnostics, and other industry and clinical applications.
Yaozhong Shi
Title: Stochastic Process Learning via Operator Flow Matching
Abstract: This talk begins by demonstrating how traditional generative models, specifically diffusion and flow-matching methods, can be lifted from finite-dimensional settings to infinite-dimensional function spaces. Building on flow matching formulated in Hilbert space, it then shows how Gaussian process regression can be generalised to full stochastic-process regression. The talk concludes with an overview of the latest neural-operator architectures designed for efficient learning on irregular grids. Relevant papers are Stochastic Process Learning via Operator Flow Matching and Mesh-Informed Neural Operator : A Transformer Generative Approach.
Bio: Yaozhong Shi is a PhD student at Mechanical Eningeering (ME) in Caltech.
Title: Scalable, Learnable, and Interactive Autonomous Driving Systems
Abstract: This talk presents a series of learning-based frameworks for interactive and scalable decision-making in autonomous driving. We begin with Differentiable Integrated Prediction and Planning (DIPP), which couples prediction and planning in a differentiable architecture to enable planning-aware predictions. Building upon this, GameFormer introduces a level-k game-theoretic modeling approach using Transformers to capture mutual interactions among agents. DTPP (Differentiable Conditional Prediction and Cost Evaluation for Tree Policy Planning) leverages a structured tree-based planner with jointly learnable prediction and cost modules, allowing more efficient and human-aligned decisions. Finally, GenDrive applies diffusion models for generative driving policy learning, further enhancing flexibility and robustness. These frameworks demonstrate how principled integration of deep learning, game theory, and generative modeling can advance decision-making capabilities in complex autonomous driving environments.
Bio: Dr. Zhiyu Huang is currently a Postdoctoral Scholar at the UCLA Mobility Lab. He received his Ph.D. from Nanyang Technological University (NTU), working in the AutoMan Research Lab. His previous experience includes a research internship at NVIDIA Research and a visiting researcher position at UC Berkeley Mechanical Systems Control Lab. His research focuses on bridging robotics, AI, and autonomous mobility, with interests in deep learning, reinforcement learning, generative models, and vision-language models. He has authored 30+ papers in top-tier venues and received the IEEE ITS Best Dissertation Award Finalist and NTU MAE Best PhD Thesis Award.
Title: Solving Nash Equilibrium in Large-scale Multi-agent System for Mean Field Game
Abstract: This talk introduces the foundations of game theory and recent advances in solving Nash Equilibria (NE) using reinforcement learning (RL). We discuss how game types extend from two-player to mean field games (MFGs), where solving NE via traditional PDE/HJB approaches becomes intractable. To address this, we present an RL-based framework that minimizes exploitability—a measure of deviation from true NE. We review key methods, including fictitious play and online mirror descent, and conclude with our recent work on a master policy for MFGs that adapts to varying initial distributions and common noise.
Bio: Zida Wu is a Ph.D. candidate in Electrical and Computer Engineering at UCLA. His research lies at the intersection of robotics, machine learning, and game theory, with a focus on developing scalable and robust solutions for complex multi-agent systems. His previous work spans estimation, planning, localization, and control. He is deeply passionate about bringing advanced techniques into real-world robotic applications!
Title: Safer Intersections for Connected Autonomous Drving
Abstract: In recent years, we have witnessed the rapid advancement of autonomous driving. The development has progressed from modular systems to end-to-end autonomous driving architectures. With the rise of Transformers and large language models, the field is now entering a new phase where world models are playing increasingly important roles. This shift is driven by several factors. For instance, long-tail scenarios and complex urban intersections pose significant challenges due to high traffic density, intricate layouts, and diverse agent behaviors. Intersections, in particular, are critical components of mobility systems—they are central to both traffic modeling and decision-making processes. Beyond technical complexity, intersections also raise major safety concerns. These environments are safety-critical, involving not only vehicles but also pedestrians and vulnerable road users such as cyclists, motorcyclists, and scooter riders. Ensuring both stability and safety in such multifaceted scenarios is paramount. Our goal, therefore, is to provide additional perception redundancies and enhance scenario understanding, enabling autonomous systems to perform more reliably in these challenging conditions.
Bio: Please check out his personal website for more info: Zhaoliang Zheng
Lizhi Yang
Title: Learning to Walk - Recent Advances in Robot Locomotion
Abstract: This talk introduces the development of robot learning for locomotion. It is a brief survey of the continual research thread in the community spanning both model-based and model-free methods. Common tools are also introduced.
Bio: Lizhi Yang is a Ph.D. Candidate in MCE at Clatech. His research lies in robot learning and robot safety. His previous works spans reinforcement learning for robot locomotion and safety-critical control. He is passionate about bringing humanoid robots into the wild.
Title: Beyond Closure Models: Learning Chaotic-Systems vis Physics-Informed Neural Operators
Abstract: Accurately predicting the long-term behavior of chaotic systems is crucial for various applications such as climate modeling. However, achieving such predictions typically requires iterative computations over a dense spatiotemporal grid to account for the unstable nature of chaotic systems, which is expensive and impractical in many real-world situations. An alternative approach to such a full-resolved simulation is using a coarse grid and then correcting its errors through a \textit{closure model}, which approximates the overall information from fine scales not captured in the coarse-grid simulation. Recently, ML approaches have been used for closure modeling, but they typically require a large number of training samples from expensive fully-resolved simulations (FRS). In this work, we prove an even more fundamental limitation, i.e., the standard approach to learning closure models suffers from a large approximation error for generic problems, no matter how large the model is, and it stems from the non-uniqueness of the mapping. We propose an alternative end-to-end learning approach using a physics-informed neural operator (PINO) that overcomes this limitation by not using a closure model or a coarse-grid solver. We first train the PINO model on data from a coarse-grid solver and then fine-tune it with (a small amount of) FRS and physics-based losses on a fine grid. The discretization-free nature of neural operators means that they do not suffer from the restriction of a coarse grid that closure models face, and they can provably approximate the long-term statistics of chaotic systems. In our experiments, our PINO model achieves a 330x speedup compared to FRS with a relative error ~10%. In contrast, the closure model coupled with a coarse-grid solver is 60x slower than PINO while having a much higher error ~186% when the closure model is trained on the same FRS dataset.
Bio: Chuwei Wang is a PhD student at CMS dept in Caltech.
Title: Plug-and-Play Diffusion Model for Inverse Problems
Abstract: Inverse problems are fundamental in many domains of science and engineering, where the goal is to infer the unknown source from indirect and noisy observations. These problems are challenging due to their ill-posedness, complexity in the underlying physics, and unknown measurement noise. This talk introduces a general family of probabilistic methods—Plug-and-Play Diffusion Prior (PnPDP) methods—that leverages both data-driven diffusion priors and knowledge of the underlying physics to solve ill-posed inverse problems. We introduce InverseBench, a benchmarking framework designed to evaluate PnPDP approaches in a systematic and easily extensible manner. We discuss experimental observations based on InverseBench, sharing insights into the efficacy and limitations of different PnPDP methods.
Bio: Hongkai Zheng is a PhD candidate at Caltech in Computing + Mathematical Sciences advised by Yisong Yue. His research interests lie in the realm of deep generative modeling and inverse problems. He develops scalable and efficient generative models and designs algorithms to solve ill-posed problems in a probabilistic framework.
Sijie Ji
Title: AI in Practice: Creating Accessible Healthcare via Ubiquitous Computing
Abstract: The integration of Artificial Intelligence (AI) into healthcare has become a transformative force, reshaping the way medical services are delivered and experienced. A particularly exciting frontier in this domain is the concept of ubiquitous computing, which refers to the seamless integration of advanced computational technologies into everyday environments. By blending the power of AI with the pervasive presence of computing devices, ubiquitous computing offers unprecedented opportunities for making healthcare more accessible, efficient, and personalized.
Bio: Dr. Sijie Ji is a Schmidt Science Fellow working at the Division of Engineering and Applied Science, California Institute of Technology, with a joint position at UCLA. She obtained her Ph.D. from Nanyang Technological University, Singapore. Sijie’s research interests span the broad area of sensing, deep learning, ubiquitous computing and their applications in digital healthcare systems, cyber-physical systems (CPS), and next-generation AIoT systems. She was an organizer and PC member for multiple conferences such as SenSys, MobiSys, BuildSys, MobiCom, ICDCS, VTC etc. She is a recipient of N2Women Young Research Fellowship. She was named CPS Rising Star, Singapore 100 Women in Tech.
Sze Chai (Mickey) Leung
Title: Data driven methods for flow control, sensing and turbulence modeling
Abstract: Optimal sensor placement (OSP) is critical for efficient, accurate monitoring, control, and inference in complex real-world systems. We propose a machine learning-based feature attribution framework to identify OSP for target predictions. Feature attribution quantifies input contributions to a model’s output and shows promise for OSP. Thus, we develop an attribution-based framework for identifying optimal sensor locations and compared its performance with other existing methods. (Relevant document: Integrated Gradients for Optimal Surface Pressure Sensor Placement for Lift Prediction of an Airfoil Subject to Gust.) Subgrid-scale (SGS) models enable large-eddy simulations to perform fluid-dynamic simulations with low computational costs by resolving only the large scales and modeling the SGS effects. We develop a novel sparse nonlinear model for predicting both the SGS stress and the numerical errors produced during the simulations which are often neglected in the other models. Our model shows promising a-posteriori test results, as evidenced by the stabilization of the simulation. (Relevant document: Consistent data-driven subgrid-scale modeldevelopment for large-eddy simulation.)
Bio: Sze Chai (Mickey) Leung is a PhD Student at Mechanical and Civil Engineering (MCE) in Caltech.
Abstract: Over the past century, microscopy has evolved significantly through advances in hardware design. However, pushing the boundaries of imaging performance using purely optical and mechanical innovations has become increasingly challenging. Meanwhile, the rapid growth in computational power has transformed the way we process and analyze imaging data, enabling a paradigm shift in microscopy. By offloading complexity from hardware to algorithms, computational microscopy—a rising interdisciplinary field—offers a powerful approach to simplify system design, correct aberrations, and extract more information from data. In this talk, I will explore how computational techniques can enhance modern imaging systems, with case studies in biological and pathological applications that highlight the potential of this synergistic approach.
Bio: Haowen Zhou is a Ph.D. candidate at Department of Electrical Engineering in California Institute of Technology, advised by professor Changhuei Yang. His research focuses on computational microscopy for 2D and 3D imaging, with broad applications in biological and clinical sciences. Haowen received his B.S. degree in Optical Engineering from Huazhong University of Science and Technology, China in 2019, an M.S. degree in Electro-Optics from University of Dayton in 2021, and an M.S. degree in Electrical Engineering from California Institute of Technology in 2024. He is a Gupta Sensing to Intelligence fellow, a 2024 SPIE Optics and Photonics Education Fellow, and a Schmidt Graduate Research Fellow.
Abstract: As machine learning shows its capabilities in scientific research, grounding models in physics is essential for achieving robustness and reliable generalization. This talk explores the integration of fundamental physical principles into machine learning frameworks. Physics-Informed Neural Networks (PINNs) embed partial differential equations (PDEs) directly into the neural network's loss function. While promising, PINNs present significant training challenges, particularly in balancing the different loss components, which can prevent convergence. There are novel methods designed to address these challenges. We introduce MultiAdam, a dynamic loss reweighting algorithm that balances different loss terms based on easy-to-access metrics, and CondPINN, a technique that improves the error landscape for more stable convergence by preconditioning. Furthermore, we present PINNacle, a comprehensive benchmark consisting of the evaluation of various PINN methods across a range of physical problems. We hope the research in physics-inspired machine learning can pave the way for AI models capable of simulating diverse physical phenomena, generating novel designs, and advancing scientific discovery.
Bio: Jiachen Yao is a PhD student at Caltech in Computing + Mathematical Sciences advised by Anima Anandkumar. His research focuses on deep generative modeling and scientific problems, learning strong priors from data and physics to model complex systems.
Contact
Please feel free to forward this message to those who might be interested. If you have questions, please contact us at Jiajing Mao.
