Karan Chadha

Karan Chadha

PhD Student in Machine Learning

Stanford University

Biography

I am a fourth year Electrical Engineering PhD student at Stanford University, advised by Prof. John Duchi. In my research, I use techniques from statistics and optimization to study different Machine Learning paradigms and give principled solutions to the problems that arise. I am currently working on problems in differential privacy and federated learning. I am also interested in studying the nuances in learning problems when the data sources are social in nature, and in modeling the impact deployed ML systems have on shaping individual choices.

Before joining Stanford, I did my bachelor’s and master’s at IIT Bombay. During my time there, I worked on Game Theory, Restless Bandit Algorithms and MCMC on Graphs with Prof. Vivek Borkar, Prof. Ankur Kulkarni and Prof. Jayakrishnan Nair.

Outside work I enjoy biking, soccer, cooking and anything outdoors.

Download my resumé .

Interests
  • Differential Privacy
  • Machine Learning
  • User heterogeneity
  • Social data in ML
Education

  • B.Tech + M.Tech in Electrical Engineering (Communications and Signal Processing), 2019

    Indian Institute of Technology, Bombay

Recent Publications

Check Google Scholar for the most recent list.
Karan Chadha, Hilal Asi, Gary Cheng, John Duchi (2022). Private optimization in the interpolation regime: faster rates and hardness results. In ICML 22.

PDF Arxiv

Karan Chadha, Gary Cheng, John Duchi (2022). Accelerated, Optimal, and Parallel: Some Results on Model-Based Stochastic Optimization. In ICML 22.

PDF Arxiv

Karan Chadha, Gary Cheng, John Duchi (2021). Federated Asymptotics: a model to compare federated learning algorithms. (In Submission).

PDF Arxiv

Karan Chadha, Hilal Asi, Gary Cheng, John Duchi (2020). Minibatch stochastic approximate proximal point methods. In NeurIPS 20.

PDF Arxiv

Karan Chadha, Ankur A. Kulkarni (2019). Aggregate play and welfare in strategic interactions on networks. Journal of Mathematical Economics.

PDF Arxiv

See all publications

Work in Progress

Private confidence sets using aggregation of subsampled bootstraps
We give differentially private algorithms for estimating coverage probabilities and computing valid confidence sets, and prove upper bounds on the error of our estimates and the length of our confidence sets. Our methods outperform normal approximation based confidence sets whenever bootstrapping outperforms normal approximation in the non-private case.

Contact