Research Interests

Academic Research 

I have worked on the following areas in my academic life.

Broadly speaking, this is related to algorithms to tackle high-dimensional likelihood-based statistical problems. This involves developing and theoretically studying novel statistical methods (Bayesian and/or frequentist) for analyzing high-dimensional and complex data.  In the Bayesian context, this also involves designing new classes of posterior approximation algorithms, based on sampling and other approaches, which can produce low-error and targeted approximations rapidly even when the data are enormous.  While there have been substantial advances in the development and understanding of methodologies to target such problems, these methodologies fall short in many applied contexts and there is a need for both better algorithms and better understanding of robustness, scalability, and statistical and algorithmic efficiency. 

More specifically, this includes Monte Carlo methods for Bayesian inference.  A particular theme in my research has been the study and development of scalable Monte Carlo algorithms for Bayesian inference.  Accurate and fast algorithms are critical in the era of big data, where traditional Markov chain Monte Carlo (MCMC) algorithms are often too slow. These scalable algorithms can broadly classified into two categories: sub-sampling-based algorithms which use minibatches of data to construct stochastic estimates of quantities such as likelihoods and gradients, and divide-and-conquer algorithms. Both kinds of approaches typically lead to additional errors as compared to standard MCMC, and it is often critical to design algorithms that minimise these errors. 


Industry work 

I am interested in causal inference and machine learning techniques for causal inference. This is of importance in industry with many application areas such as understanding the causal impact of showing advertisements on customer behaviour, the causal impact of changing customer experience on a website on customer behaviour, etc. It is critical to understand causation rather than correlation in this context as the ultimate objective is to impact customer behaviour in desired ways rather than just predicting it.