PhD Graduates Dissertation Research Spotlights

ICME PhD Graduate - Ishani Karmarkar
Advisors: Prof. Aaron Sidford, Prof. Ellen Vitercik
Title: Optimization over Graphs, Policies, and Games
Abstract: This thesis develops provably efficient algorithms for structured optimization problems arising in data science and machine learning applications. We focus on three broad classes of such problems: optimization over graphs, policies, and games. Across a range of settings within these classes, we identify structure that admits compact models of the underlying optimization instances, and we co-design classical optimization techniques such as sketching, variance reduction, and proximal methods to leverage either static or dynamic versions of these models. The resulting algorithms achieve improved runtime, query, or sample complexity for several fundamental problems in these domains.

ICME PhD Graduate: Cynthia Xinran Li
Advisor: Prof. Mirabela Rusu
Title: Deep Learning for Prostate Cancer on MRI: Detection, Model Aggregation, and Active Surveillance
Abstract: Magnetic resonance imaging (MRI) is increasingly used to detect prostate cancer, yet its interpretation remains challenging due to subtle differences between benign and malignant tissue. In this thesis, we develop deep learning methods that improve automated prostate cancer detection by incorporating clinical knowledge and leveraging complementary models, and demonstrate their clinical utility for predicting cancer progression.

ICME PhD Graduate: Hui Lan
Advisor: Prof. Vasilis Syrgkanis
Title: Robust Machine Learning Methods for Heterogeneous Treatment Effect
Abstract: Treatment effects often vary substantially across individuals, populations, and treatment environments, making the estimation of heterogeneous treatment effects (HTE) central to personalized decision making, policy evaluation, and modern causal inference. At the same time, modern observational datasets frequently involve unobserved confounding, complex treatment assignment mechanisms, and high-dimensional covariates, which can create challenges in statistical estimation. This dissertation studies the nonparametric estimation and selection of heterogeneous treatment effects in these settings, with an emphasis on developing statistically robust and computationally flexible machine learning methods for causal inference, and provides insights into how to perform model selection for HTE in the absence of ground truth counterfactuals. A central theme throughout the dissertation is the integration of ideas from semi-parametric efficiency theory and statistical learning theory to enable reliable heterogeneous effect estimation under complex data settings.

ICME PhD Graduate: James Burgess
Advisor: Prof. Serena Yeung-Levy
Title: Vision and language foundation models in scientific reasoning
Abstract: Foundation models, which are large neural networks trained on large and diverse datasets, have shown strong general-purpose capability across language, vision, and multimodal tasks. This dissertation asks how to apply such models to scientific work, focusing on vision-and-language foundation models for scientific reasoning. We begin with two earlier projects on visual representation learning: an orientation-invariant embedding method for microscopy images, and an approach for controlling 3D perspective in image generation using foundation-model representations. The thesis then turns to two methodological questions. The first question is how to define and evaluate the tasks we hope foundation models can perform. We construct a benchmark of expert-level scientific problems posed by working microscopy researchers, and diagnose where current models fail; in the process, we develop new methods for removing shortcuts from language model benchmarks. The second question is how to build systems on top of these models. We briefly discuss an effort to perform large-scale vision-language training on biomedical images, before detailing a study of language agents that retrieve and reason over the scientific literature to answer expert questions.

ICME PhD graduate: Ian Madden
Advisor: Prof. Jenny Suckale
Title: Numerical modeling for understanding the glacially-driven development of sedimentary landforms
Abstract: Gravitational stresses push grounded glaciers of ice sheets towards the ocean. As glaciers slide over the topographies at their bases, their motion and overburden impose large shear stresses on subglacial till, an often water-saturated, deformable sediment that forms the mechanical foundation beneath many ice streams. Tills reciprocate a dissipative shear stress to slow glacial motion, but failure in tills can instead enhance glacial motion. The tendency of till to deform depends on its bulk strength, effective stress, porewater pressure, and grain-scale structure. Under the right conditions, deformation of subglacial tills can organize sediment into a vast array of sedimentary landforms. 

ICME PhD graduate: Rajat Dwaraknath
Advisors: Prof. Mert Pilanci, Prof. Lexing Ying
Title: Structured Randomization and Optimization for Scalable Learning and Inference on Modern Systems
Abstract: Randomized numerical linear algebra often treats sparsity as a path to efficiency, but on GPUs this intuition can fail. The very randomness that underpins the theoretical strength of sparse sketches, such as the sparse Johnson-Lindenstrauss transform (SJLT), also creates irregular memory access and contention patterns that are poorly matched to modern computing hardware. The main focus of the dissertation is FlashSketch, a GPU-efficient sketching system built around BlockPerm-SJLT, a new sparse sketch designed to address this tension through hardware-informed structured randomness. BlockPerm-SJLT is a block-sparse sketch where each block is an SJLT, and the blocks are coupled through a union of edge-disjoint randomized permutations. This construction preserves the robustness of sparse sketching while imposing the regularity needed for efficient GPU execution. The resulting CUDA implementation advances the Pareto frontier between sketching quality and speed across randomized numerical linear algebra and end-to-end machine learning tasks.

ICME PhD graduate: Yinuo Ren
Title: Generative Modeling via Stochastic Processes: Mathematical Foundations and Scalable Algorithms
Advisors: Prof. Lexing Ying, Prof. Grant M. Rotskoff 
Abstract: Generative modeling has become a central task in modern machine learning, with the goal of learning, representing, and inferring complex high-dimensional distributions. In recent years, a particularly powerful class of methods has emerged through the language of stochastic processes, especially flow- and diffusion-based models in continuous spaces, as well as discrete diffusion models and diffusion language models for structured categorical data. These approaches have achieved state-of-the-art performance across a wide range of domains, including computer vision, language and sequence modeling, and the modeling of complex physical and biological systems. This dissertation is motivated by the observation that, despite their diversity and the distinct technical lenses through which they have been developed, these models share a common mathematical logic: generation as the transport of probability distributions through time. From this perspective, the dissertation develops a unified stochastic-process framework for generative modeling that clarifies the mathematical structure of the field while guiding the design of scalable algorithms and principled methods for real-world applications.

ICME PhD graduate: Elliot Epstein
Advisor: Prof. Kay Giesecke
Title: Structured and Efficient Sequence Modeling for High-Dimensional and Long-Horizon Time Series
Abstract: This dissertation studies how to exploit structure to make sequence models effective on time-series problems that are challenging due to large numbers of interacting units and long temporal horizons.

ICME PhD graduate: Shaun Datta
Advisor: Prof. Adam Bouland
Title: Experimental Quantum Complexity Theory
Abstract: Quantum computing is premised on the expectation that quantum computers solve certain problems exponentially faster than classical computers can. Realizing this promise in practice has been a central challenge for the field. In recent years, this challenge has been reshaped by the advent of prototype quantum devices capable of surpassing classical simulation even on the world's best supercomputers---a prospect that establishes a new frontier of computation. At this frontier, there are significant gaps between theory and experiment that invite a renewed dialogue between the two disciplines.

ICME PhD graduate: Junyi Zou
Advisor: Prof. Lu Tian
Title: Reliable and Interpretable Methods for Scientific Machine Learning: Causality, Sparsity, and Robust Feature Selection
Abstract: Machine learning has demonstrated remarkable predictive capabilities, yet its application to scientific discovery and critical healthcare domains is often hampered by a fundamental ``black-box" dilemma: standard deep learning models lack reliability in low-signal regimes, struggle to adhere to known physical mechanisms, and suffer from over-parameterization that obscures scientific insight. This thesis proposes a unified methodological framework to bridge the gap between data-driven flexibility and domain-driven rigor, organizing these contributions into three synergistic advancements. 
View dissertation 

ICME PhD graduate: Tiffany Fan
Advisor: Prof. Eric Darve
Title: Structure-Informed Machine Learning for Scientific Computing
Abstract: This dissertation advances structure-informed machine learning for scientific engineering, with a focus on interpretable representation learning and physics-aware surrogate modeling. Traditional multi-scale, multi-physics simulations remain computationally prohibitive for tasks such as optimization, control, and uncertainty quantification. To address these challenges, we develop data-driven frameworks that integrate physical context and geometric structure into learning algorithms.
View dissertation 

ICME PhD graduate: Leah Collis
Advisor: Prof. Lexing Ying
Title: Graph-Based Representation Learning for Protein Dynamics and Tensor Network Methods for the High-Dimensional Kolmogorov Backward Equation
Abstract: This thesis develops two distinct approaches for making high-dimensional stochastic dynamics more tractable. Part I introduces a data-driven method for learning a low-dimensional representation of protein dynamics. Part II develops a tensor-network approach for approximating Markov operators in high dimensions.
View dissertation