Karan Singh
email:

Karan Singh is a postdoctoral researcher at Microsoft Research in the Reinforcement Learning group. In November 2021, he completed his PhD in Computer Science at Princeton University, advised by Elad Hazan. While at Princeton, Karan was awarded the Porter Ogden Jacobus Fellowship (press, more), Princeton University's highest graduate student honor. Before that, he completed his bachelors at Indian Institute of Technology (IIT) Kanpur, where he received the President's Gold Medal (press) for the best academic performance in the graduating class.

Karan's research addresses statistical and computational challenges in feedback-driven interactive learning, spanning both prediction and control. His results draw from the algorithmic toolkits of optimization and online learning, together with techniques from dynamical systems and control theory.

His PhD dissertation work on Nonstochastic Control proposes an algorithmic (vs. traditionally analytic) foundation for control theory, and outlines provably efficient instance-optimal control algorithms (1, 2, 3, 4, 5) that go beyond both average-case notions of optimal control and worst-case notions in robust control.

His prior works delineate principled approaches (6, 7, 8) for learning and prediction in dynamical systems that do not "forget" (exhibit long-term correlations). Recently, he has also been investigating a plausible systems-level theory of machine learning (e.g. 9, but broader), where guarantees on the aggregate could be synthesized from functional (rather than behavioral) characteristics of individual subsystems.

Email  |  CV  |  Google Scholar

News
Peer-reviewed Publications
All publications list authors in the alphabetical order, except those indicated with †.
Boosting for Online Convex Optimization
with Elad Hazan
International Conference on Machine Learning (ICML), 2021
proceedings | arXiv
A Regret Minimization Approach to Iterative Learning Control
with Naman Agarwal, Elad Hazan, Anirudha Majumdar
International Conference on Machine Learning (ICML), 2021
proceedings | arXiv
Improper Learning for Nonstochastic Control†
with Max Simchowitz, Elad Hazan
Conference on Learning Theory (COLT), 2020
proceedings | arXiv
No-Regret Prediction in Marginally Stable Systems
with Udaya Ghai, Holden Lee, Cyril Zhang, Yi Zhang
Conference on Learning Theory (COLT), 2020
proceedings | arXiv
The Nonstochastic Control Problem
with Elad Hazan, Sham Kakade
Algorithmic Learning Theory (ALT), 2020
proceedings | arXiv
Logarithmic Regret for Online Control
with Naman Agarwal, Elad Hazan
Neural Information Processing Systems (NeurIPS), 2019 Oral Presentation (<0.5% of submissions)
Also, Best Paper Award at the OptRL workshop at NeurIPS 2019
proceedings | arXiv
Online Control with Adversarial Disturbances
with Naman Agarwal, Brian Bullins, Elad Hazan, Sham Kakade
International Conference on Machine Learning (ICML), 2019
proceedings | arXiv
Provably Efficient Maximum Entropy Exploration
with Elad Hazan, Sham Kakade, Abby Van Soest
International Conference on Machine Learning (ICML), 2019
proceedings | arXiv
Efficient Full-Matrix Adaptive Regularization
with Naman Agarwal, Brian Bullins, Xinyi Chen, Elad Hazan, Cyril Zhang, Yi Zhang
International Conference on Machine Learning (ICML), 2019
proceedings | arXiv
Spectral Filtering for General Linear Dynamical Systems
with Elad Hazan, Holden Lee, Cyril Zhang, Yi Zhang
Neural Information Processing Systems (NeurIPS), 2018 Oral Presentation (<0.5% of submissions)
proceedings | arXiv
Learning Linear Dynamical Systems via Spectral Filtering
with Elad Hazan, Cyril Zhang
Neural Information Processing Systems (NeurIPS), 2017 Spotlight (<5% of submissions)
Also, Spotlight Prize at New York Academy of Sciences' ML Symposium, 2018
proceedings | arXiv
The Price of Differential Privacy for Online Learning
with Naman Agarwal
International Conference on Machine Learning (ICML), 2017
proceedings | arXiv
Efficient Regret Minimization in Non-Convex Games
with Elad Hazan, Cyril Zhang
International Conference on Machine Learning (ICML), 2017
proceedings | arXiv
Preprints and Technical Reports
A Boosting Approach to Reinforcement Learning
with Nataly Brukhim, Elad Hazan
Prelim version at ICML Workshop on RL Theory, 2021
Dynamic Learning System
with Elad Hazan, Cyril Zhang
US Patent 11,138,513 B2, approved Oct 2021
Machine Learning for Mechanical Ventilation Control†
with Daniel Suo, Cyril Zhang, Paula Gradu, Udaya Ghai, Xinyi Chen, Edgar Minasyan, Naman Agarwal, Julienne LaChance, Tom Zajdel, Manuel Schottdorf, Daniel Cohen, Elad Hazan
Machine Learning for Health (ML4H), 2021 Workshop Track
Featured in Princeton Engineering news.
Deluca -- A Differentiable Control Library: Environments, Methods, and Benchmarking†
with Paula Gradu, John Hallman, Daniel Suo, Alex Yu, Naman Agarwal, Udaya Ghai, Cyril Zhang, Anirudha Majumdar, Elad Hazan
NeurIPS Workshop on Differentiable Computer Vision & Physics, 2020 Oral Presentation
Towards Provable Control for Unknown Linear Dynamical Systems
with Sanjeev Arora, Elad Hazan, Holden Lee, Cyril Zhang, Yi Zhang
International Conference on Learning Representatios (ICLR), 2018 Workshop Track
Dynamic Task Allocation for Crowdsourcing†
with Irineo Cabreros, Angela Zhou
ICML Workshop on Data Efficient Machine Learning, 2016