Oscar Leong
I am a Tenure-Track Associate Professor at the Institute of Natural Sciences and the School of Mathematical Sciences at Shanghai Jiao Tong University. From 2024 to 2026, I was a Tenure-Track Assistant Professor in the Statistics and Data Science department at UCLA. From 2021 to 2024, I was a von Kármán Instructor at Caltech in the Computing + Mathematical Sciences department, hosted by Venkat Chandrasekaran, where I also worked closely with Katie Bouman and the Computational Cameras group. I completed my PhD from Rice University in Computational and Applied Mathematics under the supervision of Paul Hand and was an NSF Graduate Research Fellow. I received my undergraduate degree in Mathematics from Swarthmore College.
My research interests lie in the mathematics of data science, inverse problems, machine learning, and optimization. Much of my work concerns solving signal recovery problems with approaches inspired by deep learning and uses tools from high dimensional probability, convex and star geometry, random matrix theory, and optimization to develop provable recovery guarantees. Recently, I have also become interested in the intersection of human and machine intelligence.
Family: my wife, Wani, is a developmental psychologist and on the faculty at Fudan University. We welcomed our first daughter, Gemma, in May of 2023.
Score Distillation Beyond Acceleration: Generative Modeling from Corrupted Data (with Yasi Zhang, Tianyu Chen, Zhendong Wang, Ying Nian Wu, and Mingyuan Zhou)
ICLR 2026. [OpenReview] (this paper was the result of merging DSD and RSD)
Riemannian AmbientFlow: Towards Simultaneous Manifold Learning and Generative Modeling from Corrupted Data (with Willem Diepeveen)
Submitted. [arXiv]
Discovering Structure from Corruption for Unsupervised Image Reconstruction (with Angela Gao, He Sun, and Katie Bouman)
IEEE Transactions on Computational Imaging 2023. [arXiv][ICASSP][project page][5-min video]
Optimal Regularization Under Uncertainty: Distributional Robustness and Convexity Constraints (with Eliza O'Reilly and Yong Sheng Soh)
Submitted. [arXiv]
Learning Difference-of-Convex Regularizers for Inverse Problems: A Flexible Framework with Theoretical Guarantees (with Yasi Zhang)
Submitted. [arXiv]
The Star Geometry of Critic-Based Regularizer Learning (with Eliza O'Reilly and Yong Sheng Soh)
NeurIPS 2024. [arXiv] [poster]
Optimal Regularization for a Data Source (with Eliza O'Reilly, Yong Sheng Soh, and Venkat Chandrasekaran)
Foundations of Computational Mathematics 2025. [arXiv]
A Recovery Theory for Diffusion Priors: Deterministic Analysis of the Implicit Prior Algorithm (with Yann Traonmilin)
AISTATS 2026. [arXiv]
Flow Priors for Linear Inverse Problems via Iterative Corrupted Trajectory Matching (with Yasi Zhang, Peiyu Yu, Yaxuan Zhu, Yingshan Chang, Feng Gao, and Ying Nian Wu)
NeurIPS 2024. [arXiv]
Compressive Phase Retrieval: Optimal Sample Complexity with Deep Generative Priors (with Paul Hand and Vladislav Voroninski)
Communications on Pure and Applied Mathematics 2024. [arXiv][journal]
Phase Retrieval Under a Generative Prior (with Paul Hand and Vladislav Voroninski)
NeurIPS 2018. Oral Presentation (0.6% of submissions). [arXiv]
What Makes a Good Example? Modeling Exemplar Selection with Neural Network Representations (with Fanxiao Wani Qiu and Alexander Latourrette)
Proceedings of the 48th Annual Meeting of the Cognitive Science Society (CogSci). [arXiv]
Feature, Alignment, and Supervision in Category Learning: A Comparative Approach with Children and Neural Networks (with Fanxiao Wani Qiu)
Proceedings of the 48th Annual Meeting of the Cognitive Science Society (CogSci). [arXiv]
August 2026: Congratulations to Yasi on successfully defending their PhD thesis! Yasi has done incredible work at the interface of generative AI, multimodality, and inverse problems and will be joining Google as a Research Scientist.
June 2026: New preprint out on MT-EditFlow, which uses reinforcement learning for better image editing with flow matching models.
May 2026: One paper accepted to ICML 2026! Great work from Yasi and Tianyu on integrating reinforcement learning for LLM-as-a-Judge in regression tasks.
April 2026: Two papers were accepted to CogSci 2026! Very excited that Wani and I's first interdisciplinary collaborations (paper 1 and paper 2) on the intersection of human and machine intelligence were accepted.
February 2026: Two new preprints out! These are extra special as they represent the first collaboration between my wife, Wani, and I on human and machine intelligence:
Paper 1: We rigorously analyzed the semi-supervised category learning ability of children (ages 5-7) and neural networks in a controlled setting, and find that their performance is influenced differently by an interaction of factors (including feature structure, difficulty of the task, and level of supervision).
Paper 2: How do we choose examples from a category to teach others? We show how we can use neural networks with dataset distillation techniques to computationally model how humans address this question! This was also a part of a wonderful collaboration with Sandy at USC.
January 2026: New preprint out with Willem on combining data-driven, pullback Riemannian geometry with AmbientFlow to simultaneously learn a manifold and generative model from corrupted data!
January 2026: Two papers were accepted to ICLR 2026! One on distillation for learning from corrupted data and the other on multi-turn editing, both led by Yasi and Tianyu. Congrats to them on their hard work, and thanks to Mingyuan for the great collaboration.
January 2026: Our paper on signal recovery theory for diffusion priors was accepted to AISTATS! Thanks to Yann for the great collaboration.
October 2025: Presented our work on optimal regularization at University of Minnesota's Data Science Seminar at the Institute for Mathematics and its Applications. Thanks to Gilad for the invitation!
October 2025: Two new preprints out:
Paper 1: Collaboration with Eliza and Yong Sheng on leveraging distributionally robust optimization to develop an uncertainty-aware framework for optimal regularization.
Paper 2: Big effort led by Johannes, Hok Shing, Matthias, and Sebastian on a unified comparison between the various learned regularization frameworks for imaging inverse problems.
September 2025: Two new preprints out:
Paper 1: Collaboration with Yann on developing signal recovery theory for diffusion priors. We show how to interpret score-based priors as time-varying, approximate projections onto a model set, and derive recovery guarantees in linear inverse problems for a variety of data distributions, such as low-rank GMMs.
Paper 2: Led by Tianyu and Yasi, EdiVal-Agent introduces an automated, object-centric framework and new benchmark for scalable, fine-grained evaluation of multi-turn editing, integrating vision-language models with expert tools to better align with human judgments.
September 2025: I gave a talk at the Allerton Conference in the Diffusion Models for Inverse Problems session on our distillation work. Thanks to Jane, Minh, and Yue for the invitation!
May 2025: New preprint out extending our DSD work on learning clean diffusion models from linearly corrupted data via score distillation!
March 2025: New preprint out on how distillation can be used as a surprisingly effective technique to learn clean diffusion models from noisy data! This project is a part of a wonderful, ongoing collaboration with Mingyuan Zhou's group at UT Austin and was jointly led by Tianyu and Yasi.
February 2025: New preprint out with Yasi on learning a class of nonconvex regularizers given by the difference-of-convex functions!
October 2024: Our paper on optimal regularization has been accepted to Foundations of Computational Mathematics!
September 2024: Two papers were accepted to NeurIPS 2024! The first on the star geometry of critic-based regularizer learning, joint with Eliza and Yong Sheng, and the second on flow matching priors for linear inverse problems, led by my student Yasi.
September 2024: I gave a talk at the Allerton Conference in the Optimization session on optimal regularization using star geometry. Thanks to Richard and Salar for the invitation!
September 2024: I gave a talk in Edward Castillo's DMIC Lab AI Medical Imaging Seminar on generative networks for inverse problems without ground-truth data. Thanks to Jorge for the invitation!
September 2024: New preprint out with Eliza and Yong Sheng on using star geometry to better understand neural network-based adversarial regularizers!
November 2023: Honored to have been elected as a member of the IEEE Computational Imaging Technical Committee (CI TC).
November 2023: I gave a talk in the Math Machine Learning seminar, joint with MPI MiS and UCLA, on generative networks for inverse problems.
September 2023: Our paper on solving inverse problems without explicit priors by exploiting common structure has been accepted to IEEE Transactions on Computational Imaging!
February 2023: Our paper on establishing optimal sample complexity in phase retrieval with deep generative priors has been accepted to Communications on Pure and Applied Mathematics!
February 2023: Two papers accepted to ICASSP 2023, one on Rohun's SURF project and the other with Angela, He, and Katie on image reconstruction without explicit priors!
December 2022: New preprint out with Eliza, Yong Sheng, and Venkat on characterizing optimal regularizers for a data distribution!
December 2022: Honored to have been awarded an MGB-SIAM Early Career Fellowship.
November 2022: Rohun's SURF project on denoising priors for phase retrieval has been uploaded to arXiv as a preprint!
October 2022: Honored to have been selected as a Rising Star in Data Science by UChicago.
July 2022: I gave a talk at ICCOPT on new work with Yong Sheng and Venkat on our variational analysis of learning convex regularizers. A preprint will be posted soon!
May 2022: I presented during the CMX student/postdoc seminar on some new, exciting work with Angela, He, and Katie on learning image models directly from noisy data! Please see our webpage for more info.
February 2022: Mateo Díaz, Yong Sheng Soh, and I are organizing a pair of sessions on "Convex and nonconvex methods for matrix factorization problems" at ICCOPT 2022!
December 2021: I will be serving as a local arrangement chair for ICCP 2022!
October 2021: Will be presenting in-person (first time in awhile!) during Caltech's CMI seminar on using deep generative models in inverse problems.
September 2021: I'll be virtually presenting our generative prior work at the "Generative Regularization Approaches for Inverse Problems" Minisymposium at the IFIP TC7 Conference.
May 2021: Our paper on analyzing subgradient descent for phase retrieval using non-Lipschitz matrix concentration accepted to Communications in Mathematical Sciences.
April 2021: I successfully defended my dissertation! Excited to announce I'll be joining Caltech in the Department of Computing + Mathematical Sciences as a von Kármán Instructor in the Fall.
Here is a video of my oral presentation at NeurIPS 2018 on our Deep Phase Retrieval paper along with a link to a 3 minute summary of our work:
