About MeI am now at Google X. I was previously a Staff Software Enginner in the SmartPricing team at Walmart Labs, where I worked on
using machine learning algorithms for dynamic pricing on walmart.com. Two of the important projects that I have worked on at Walmart are: Thompson Sampling for dynamic pricing of items, dynamic pricing of overstock items. In tandem with product managers and my colleagues, I have worked on end-to-end implementation of dynamic pricing systems at walmart. This means designing, implementation and monitoring of the algorithms. Previously, I was a Postdoctoral Research Associate in the Optimization group at the Wisconsin Institutes for Discovery, UW-Madison. My mentors there were Rebecca Willett and Rob Nowak. At Madison I worked on two distinct research threads. In the first thread of work we worked on creating flexible, shallow, high-dimensional models, called single-index models, and designed machine learning algorithms to learn such models from data. Our algorithms are flexible enough to be applied to various kinds of high-dimensional statistical tasks, with structures of choice (sparsity, group sparsity, low rank, <insert any new structure you want>). If you want an algorithm that is more flexible than logistic regression, but at the same time less computationally intense than deep models then you should try out our algorithms (look at the AAAI 2017, NIPS 2015 papers in the publication list). We also applied our algorithms to the problem of matrix completion a.k.a. recommender systems, and obtained better results than the state-of-the-art matrix completion solvers. In the second thread of my research at UW-Madison, I worked on multi-armed bandit algorithms. We looked at large scale bandit problems where the reward structure of the arms can be arranged as a positive semidefinite matrix. We designed and implemented algorithms that exploit the PSD structure, and return the best arm without even sampling all the arms. The final result was a cool algorithm that combines elements of Nystrom completion from matrix algebra, successive elimination from multi-armed bandits. If you ever were looking for an alternative to Nystrom algorithm, then our algorithms can be a good choice (look at our AISTATS 2017 paper). Prior to my stint at UW-Madison, I got my PhD from Georgia Tech under the supervision of Alex Gray working on a thesis in active learning. We designed active learning algorithms and established bridges between the disparate fields of active learning and multi-armed bandits. As you must have guessed by now, I work in disparate areas of machine learning (multi-armed bandits, bandit optimization, active learning, matrix completion, non-parametric methods, high-dimensional learning). I do not like to pigeonhole myself into any one class of methods. In recent times I have been studying graphical models, approximate inference and of course deep learning. My goal is to teach myself enough in these areas and see how they intersect with the sequential/bandit world. Bandits on deep generative models anyone? Here is my short CV and here is a longer one. |
Publications |
|