Alumni Spotlight Talks

Title: nput-Adaptive Autoencoding: A Training-Data-Free Approach for Deep Generative Image Recovery

Abstract: Inverse imaging problems (IIPs)—including denoising, inpainting, non-linear deblurring, and medical imaging tasks such as MRI and CT reconstruction—are fundamental challenges in computational imaging. Recent advances in deep learning have led to a range of generative approaches for IIPs, from data-less methods like Deep Image Prior (DIP) to data-intensive diffusion models (DMs). In this talk, I will introduce Autoencoding Sequential DIP (aSeqDIP), a training-data-free approach that mitigates DIP’s overfitting via a sequential input-adaptive optimization with autoencoding regularization. Empirical results across linear and non-linear image recovery problems, show that aSeqDIP achieves competitive or superior performance in terms of reconstruction quality and/or run-time efficiency.

Title: Going between math, data science, plant biology, and education. Journeys into interdisciplinarity

Abstract: When I arrived CMSE (and the US) as a new international PhD student, I was mentally ready for the cultural shock from switching countries. I was not ready for the cultural shock from switching disciplines, from mathematics to biology. That was a far more difficult bridge to cross. However, CMSE was the best place to make that transition happen. With plenty of support from my advisors and CMSE, I thrived at the intersection of applied topology, data science, and plant biology. Moving back and forth between disciplines has taught me valuable lessons on how to foster precisely these interdisciplinary bridges. 

My interdisciplinary journey has taken me to the University of Missouri, where I've had the chance to use image processing to track vampire plants, applied topology to model mRNA localization in soybean cells, and develop curricula and materials to teach data science with a specific biology focus. As biology transitions into a data-driven era, a meaningful interpretation of large datasets is a limiting factor. The solutions to grand societal challenges we face lie in data science. And more interdisciplinary scientists are critical to address these challenges.

Title: Building data-driven behavior models for epidemiological agent-based models

Abstract: The COVID-19 pandemic highlighted the importance of human behavior in mitigating the spread of infectious diseases. Current approaches for studying behavior largely rely on surveys, which are often inadequate for integration into large-scale epidemiological models. While significant progress has been made in making synthetic populations, there is less focus on creating behavioral profiles from real-world data. We address these gaps by developing data-driven models of mitigation behavior adoption that account for demographical, geospatial, and temporal variations among individuals. We used open-source data from the U.S. COVID-19 Trends and Impact Survey, COVID States project, and census to build models that estimate the percent of individuals that wore a mask, went to work, used public transit, spent time with and avoided contact with others during the pandemic. Our approach used the machine learning-driven algorithm MissForest to impute missing data. The results were then combined with COVID States and census data and up sampled using iterative proportion fitting, ensuring the preservation of demographical relations across temporal and geospatial scales. Our estimates are validated against external datasets, demonstrating comparable accuracy and robustness. This innovative approach to human behavior modeling not only enhances the complexity and accuracy of epidemiological models but also offers a framework for complementing surveys and understanding emergent human behavior.

Title: Extended Magnetohydrodynamics Models for High Energy Density Physics Simulations

Abstract: This presentation will provide an overview of some recent advancements in several of Sandia’s plasma modeling codes. It will discuss improvements in extended magnetohydrodynamics modeling capabilities within productionline codes and research focused on developing non-quasi-neutral multifluid models suitable for the high energy density conditions found in pulsed-power experiments. Special emphasis will be placed on avenues for future modeling and algorithmic enhancements

Title: Ongoing Work at Amazon and the Role of Recent AI Developments

Abstract: This talk will outline some of the main challenges in e-commerce, web services and advertising technology, and describe how recent progress in AI is being applied in these areas at Amazon. The speaker will give a brief overview of current Amazon products related to these domains, as well as a few possible directions for future services. Two examples from published research projects—one on improving ranking performance with uncertainty information, and another on automated evaluation of code changes—will be discussed, if time permits, to illustrate how research is carried out in an industry setting.

Title: Consensus based optimization with adaptive Momentum Estimation and its application on control

Abstract: We propose a gradient-free optimization framework, Adaptive Momentum Consensus-Based Optimization (Adam-CBO), for solving high-dimensional stochastic optimal control problems. Traditional policy gradient methods in reinforcement learning suffer from high variance and scalability issues, particularly in nonconvex and continuous control settings. To address these challenges, we extend the Consensus-Based Optimization (CBO) paradigm by introducing momentum dynamics and adaptive second-moment estimation, akin to the Adam optimizer. This results in the Momentum CBO (M-CBO) and Adam-CBO algorithms, which optimize policy parameters without relying on explicit gradients or discretization of the state-action space. We establish theoretical convergence guarantees under mild assumptions and demonstrate the effectiveness of our methods through numerical experiments on linear-quadratic control, the Ginzburg–Landau model, and mean-field control problems. The results show that Adam-CBO achieves high accuracy and scalability, outperforming baseline methods, especially in high-dimensional and nonconvex environments.

Title: Evolution of Ads Conversion Optimization Models at Pinterest

Abstract: Accurate conversion rate (CVR) prediction is fundamental to modern advertising systems. At Pinterest, our modeling architecture has evolved from a foundational multitask learning (MTL) framework to progressively incorporate advanced feature interaction modules such as Deep & Cross Network v2 (DCNv2), MaskNet, and Transformers. This presentation details the next stage of this evolution: the successful adaptation of a Deep Hierarchical Ensemble Network (DHEN) for the sparse-signal CVR prediction task. Key contributions include: (1) the deployment of DHEN as a unified backbone within our MTL system; and (2) the integration of rich, sequential user features from both on-site and off-site activities. The resulting model achieves state-of-the-art performance, demonstrating significant improvements over preceding architectures and establishing a new benchmark for conversion optimization

Title: Multi-Stage Ranking at Meta Ads: Architecture, Challenges, and Innovations

Abstract: Meta is committed to building an industry-leading Ads Ranking System that embodies state-of-the-art machine learning excellence. Designed to serve billions of users and rank trillions of ads every day, the Meta Ads Ranking System is architected for scalability, efficiency, and precision. At its core, the system employs a multi-stage ranking pipeline that progressively refines candidate ads through a sequence of specialized models, each tailored to distinct optimization objectives.

This talk will introduce the design of Meta’s multi-stage ranking system and the core challenges we face in data augmentation, knowledge distillation, and foundational modeling. If time permits, I will also share my experience in translating the interdisciplinary engineering skills I developed in the field of bioinformatics at CMSE to the domain of ads ranking.

Title: Using data science and artificial intelligence to support physics education

Abstract: Data science and artificial intelligence are transforming educational research, enabling deeper insights into learning behaviors and institutional effectiveness. The large-scale collection of student learning and assessment data opens up novel avenues for evaluating student outcomes and institutional practices. In this talk, I will provide examples from across my career. Specifically, I will highlight three studies: using machine learning to investigate whether a department successfully changed their graduate admissions process, applying data science techniques to examine achievement gaps across majors, and exploring how generative artificial intelligence can be used as part of an instructor-in-the-loop system to support student learning. Together, these examples demonstrate the power of data science not just to observe, but to improve educational systems.