Mirah Shi

Email: mirahshi at seas.upenn.edu

Hello! I'm a third-year PhD student in computer science at the University of Pennsylvania, where I'm fortunate to be advised by Michael Kearns and Aaron Roth. I work at the intersection of machine learning theory, economics, and society, with a focus on online learning in the presence of human decision-makers. My research is supported by AWS AI. Before coming to Penn, I received my bachelors in math from Barnard College.

I pronounce my name my-ra shee.

Publications [google scholar]

• Collaborative Prediction: Tractable Information Aggregation via Agreement
  Natalie Collina, Ira Globus-Harris, Surbhi Goel, Varun Gupta, Aaron Roth, Mirah Shi
  [arXiv] Preprint
  Imagine there are multiple agents (e.g. a human and neural network) who hold different features and would like to collaboratively learn on their joint feature space --- without ever directly sharing their features. We show how, by simply exchanging predictions generated using their individual data, they can converge to a set of predictions that are accurate on their collective data.


• Sample Efficient Omniprediction and Downstream Swap Regret for Non-Linear Losses
  Jiuyao Lu, Aaron Roth, Mirah Shi
  [arXiv] Preprint
  Recent literatures in omniprediction and downstream (swap) regret study how to design predictions to optimize many losses at once. This paper introduces a unified framework: we show how to minimize "decision swap regret," which generalizes both omniprediction and downstream swap regret. Our results extend/improve upon both literatures.


• Algorithmic Aspects of Strategic Trading
  Michael Kearns, Mirah Shi
  [arXiv] Preprint
  What happens when traders---who each want to acquire some position in a stock---interact strategically? This paper investigates equilibria computation in a trading game.


• An Elementary Predictor Obtaining \(2\sqrt{T}+1\) Distance to Calibration
  Eshwar Ram Arunachaleswaran, Natalie Collina, Aaron Roth, Mirah Shi
  [arXiv] [slides] SODA 2025 (Also accepted to the NeurIPS 2024 Optimization for Machine Learning Workshop)
  We give an extremely simple, efficient, deterministic online algorithm that achieves low distance to calibration, answering an open question of Qiao and Zheng (2024).


• Forecasting for Swap Regret for All Downstream Agents
  Aaron Roth, Mirah Shi
  [arXiv] [slides] [poster] EC 2024 (Also accepted to the ESIF Economics and AI+ML Meeting 2024)
  How can we make forecasts that are simultaneously valuable (guarantee low swap regret) to any downstream agent? Calibration is one answer, but it suffers from poor convergence rates. Our techniques circumvent calibration to achieve low downstream swap regret at drastically improved rates.


• Center-Embedding and Constituency in the Brain and a New Characterization of Context-Free Languages
  Daniel Mitropolsky, Adiba Ejaz, Mirah Shi, Christos Papadimitriou, Mihalis Yannakakis
  [arXiv] Natural Logic Meets Machine Learning Workshop (NALOMA) 2022
  Proposes a biologically plausible implementation of a language parser that can handle recursion (i.e. embedded sentences) and generate constituency representations.

Teaching

I've been a teaching assistant for the following courses at Penn:

• NETS 4120 Algorithmic Game Theory (Spring 2024)
• CIS 6250 Theory of Machine Learning (Fall 2022)

• COMS 3261 Computer Science Theory (Spring 2021)

Other

• I co-organize the Theory Seminar at Penn. Feel free to reach out if you'd like to give a talk!
• I spent the summer of 2021 doing research at Pacific Northwest National Laboratory, hosted by Sinan Aksoy.