Purushottam Kar

Purushottam Kar 

Purushottam Kar, Ph.D.
Consulting Researcher
Microsoft Research India, Bengaluru

Assistant Professor (on long leave)
Department of Computer Science and Engineering
The Indian Institute of Technology Kanpur

Welcome to my homepage. You will find my contact details and some articles and talks of mine on this site.
Please use the links on the left hand menu to navigate.

Leave Notice: Starting August 2020, I am on long leave from IIT Kanpur.


  • [02 August 2020] [New] Our work on Bayesian optimization techniques for designing effective lockdown and other non-pharmaceutical interventions to check the spread of pandemics such as CoViD-19 will appear in the Transactions of the INAE. Code for our method and a preprint are available [here].

  • [27 May 2020] [New] Our work on accelerated program repair with applications to AI-based e-tutors for introductory programming courses will appear at AIED 2020. Code for our method and a preprint are available [here].

  • [02 June 2019] Our work on accelerating extreme classification algorithms that are used for recommendation and labelling tasks with millions of labels, will appear at IJCAI 2019. Code for our method and a preprint are available [here].

  • [25 March 2019] Check out our recent works on using the IRLS heuristic for robust regression (to appear in AISTATS 2019), and our work on bandit algorithms that are resilient to data corruption (to appear in Machine Learning J.). Preprints are available [here].

  • [25 March 2019] Our monograph on non-convex optimization is a helpful guide to understanding the design and analysis of scalable algorithms for solving non-convex optimization problems in machine learning. Purchase the official copy [here] or get a free copy [here].

  • [5 September 2017] Our paper on consistent robust regression will appear at NIPS 2017. Our work develops the first polynomial time algorithm that can solve the age-old problem of linear regression in a statistically consistent manner even when a large number of training data samples are adversarially corrupted.

  • [15 May 2017] Our paper on context-dependent clustering of multi-armed bandits will appear at ICML 2017. Our work demonstrates how improved recommendation systems can be developed by combining explore-exploit-style bandit techniques with collaborative filtering techniques.