Yash Patel

Yash Patel

PhD student in Statistics at University of Michigan, interested in robust decision-making and neural operators. Previously computer vision/graphics engineer at Facebook. Math graduate from Princeton University '18 (certificates in CS and Statistics/ML).

Yash Patel

Bio

I am currently a 4th year PhD student at the University of Michigan studying Statistics advised by Ambuj Tewari. I graduated as a Math major from Princeton in 2018 with certificates in CS and Statistics/ML, where I worked with Matt Weinberg on adversarial behaviors on blockchain networks. From 2018 to 2021, I worked as a computer vision/graphics engineer at Meta, where I worked on realtime dense 3D reconstruction (point clouds/KinectFusion) and dynamic real-time disocclusion on meshes. I previously worked on the software for the Manifold camera, specifically adding support for rendering on cloud farms and improving the efficiency of the depth estimation algorithm. I also implemented, trained, and optimized (with layer fusion, SNPE, and QAT quantization) an end-to-end deep learning-based feature for Portal to run at realtime on the Snapdragon SoC. See my full resume here.

Research Interests

My research interests are in the use of robust machine learning predictions for making decisions downstream with probabilistic guarantees. Such ``predict-then-optimize’’ problems naturally arise in many settings, such as in vehicle routing and model-based control. With machine learning predictions being increasingly leveraged in such multi-stage decision-making pipelines, I am very excited in studying both new robustification schemes, such as those extending conformal prediction, and the accompanying optimization algorithms to efficiently solve the resulting robust decision-making problems.

My current research interests are shifting to the subclass of problems in this space where the upstream prediction itself is a function, owing to the recent progress on neural operator learning for accelerating PDE solvers. I am very excited to explore the implications on robust control and related domains from this work.

Earlier in my PhD, I explored alternative approaches for uncertainty quantification, focusing on variational inference, and its use in computational biology and astrophysics.

Conference Publications

Conformal Contextual Robust Optimization, Patel Y, Rayan S, Tewari A. International Conference on Artificial Intelligence and Statistics, 2024 Oral Presentation

Amortized Variational Inference with Coverage Guarantees, Patel Y, McNamara D, Loper J, Regier J, Tewari A. International Conference on Machine Learning, 2024

Manuscripts In Submission

Conformal Prediction for Robust Control, Patel Y, Rayan S, Tewari A.

Conformalized Late Fusion Multi-View Learning, EO Rivera*, Patel Y*, (* denotes equal contribution), Tewari A.

Workshop Proceedings

Non-Parameteric Conformal Distributionally Robust Optimization, Patel Y, Cao G, Tewari A. ICML 2024 Workshop on Structured Probabilistic Inference & Generative Modeling

Diffusion Models for Probabilistic Deconvolution of Galaxy Images, Li Y, Xue Z, Patel Y, Regier J. ICML Machine Learning for Astrophysics Workshop, 2023

RL Boltzmann Generators for Conformer Generation in Data-Sparse Environments, Patel Y, Tewari A. NeurIPS Machine Learning in Structural Biology (MLSB) Workshop, 2022

Scalable Bayesian Inference for Finding Strong Gravitational Lenses, Patel Y, Regier J. NeurIPS Machine Learning and the Physical Sciences Workshop, 2022

Patents

Holographic Calling for Artificial Reality, AP Pozo, J Virskus, G Venkatesh, K Li, SC Chen, A Kumar, R Ranjan, BK Cabral, SA Johnson, W Ye, MA Snower, Y Patel. US Patent App. 17/360,693

Awards

2x NSF GRFP Honorable Mention (2020, 2022)

Outstanding First-Year Ph.D. Student Award (2022)

Outstanding Graduate Student Instructor Team Award (2022)

Graduate Student Service Award Award (2024)

Mentoring

During my PhD, I have also had the opportunity to mentor the following fantastic undergraduate and master’s students.

Guyang (Kevin) Cao (Honors Thesis 2023-2024; Undergraduate Research Program in Statistics 2023): Non-parametric Conformal Distributionally Robust Optimization. Next step: Ph.D. in Computer Science at University of Wisconsin-Madison

Zhiwei Xue (Undergraduate Research Program in Statistics 2023): Diffusion Models for Probabilistic Deconvolution of Galaxy Images. Next step: Ph.D. in Computer Science at National University of Singapore

Yuhang Li (Undergraduate Research Program in Statistics 2023): Diffusion Models for Probabilistic Deconvolution of Galaxy Images. Next step: Master’s in Computer Science at University of Illinois, Urbana-Champaigna

Zhong Zheng (Undergraduate Research Program in Statistics 2023): Atomic Maps Reconstruction for Cryo-EM Data with Continuous Heterogeneity. Next step: Master’s in Computational Data Science at Carnegie Mellon University

Chengsong Zhang (2022): Molecular Conformer Generation. Next step: Master’s in Computer Science at University of Illinois, Urbana-Champaigna

Teaching

  • STATS 485 (Capstone Seminar), Fall 2023, Fall 2024.

  • STATS 504 (Practice and Communication in Applied Statistics), Winter 2023.

  • STATS 315 (Introduction to Deep Learning), Winter 2022, Fall 2022: Co-designed course for its 1st offering (F22)

  • STATS 250 (Introduction to Statistics), Fall 2021

Miscellaneous

Outside of research, I also really enjoy lifting and reading. In a previous life, I dabbled in other subfields I was interested in, like computer architecture and computer graphics; here are some artifacts of those interests: