My current research interests include but is not limited to,
AI, machine learning, deep learning, NLP/NLU, graphical models, optimization, social and communication networks.
Previously I have also worked on,
Localization in wireless networks; Physical layer algorithms for LTE/WiMax networks, MIMO cooperative communication, space-time codes, synchronization in OFDM systems, and Mobile Ad-Hoc networks.
2016 - 2021, in Artificial Intelligence Center Lab
At Samsung Research, I developed new deep learning based algorithms for personalized natural language understanding, and continuous learning in digital AI assistants (e.g. Bixby, Alexa, Siri, Cortana). Our NLU models were commercialized and used by Samsung's Bixby assistant across all Samsung products (mobile, wearables, home appliances). We also developed a hybrid rule + ML algorithm/tool that can be used by ordinary software developers to rapidly develop NLU skills for AI assistants from cold-start (zero labeled data). (algorithms were implemented using Python/Tensorflow/Torch)
Aug 2010 – Oct 2016, Ph.D. dissertation advised by Prof. Sujay Sanghavi and Prof. Sanjay Shakkottai.
In my Ph.D. dissertation work I developed fast and scalable machine learning algorithms (unsupervised and semi-supervised) with applications in learning network structure (graphical models, overlapping communities, latent interest groups), and latent variable models (GMM, LDA topic models, subspace clustering, mixed regression). The algorithms were implemented (in Matlab/Python/C++) and tested on real and synthetic datasets to demonstrate their practical usefulness. We also derive theoretical performance guarantees for both sample complexity and runtime for our algorithms using mathematical tools from probability theory, tensor factorization, convex optimization, information theory, and percolation theory.
June 2008 – July 209, ME dissertation advised by Prof. P. Vijay Kumar
First we investigate the algebraic structure of an easily decodable Space-Time Block Code (STBC) called the Silver Code. We proved that the code is a right ideal in a certain Cyclic Division Algebra (CDA), and hence we prove the optimality of the code among a class of such easily decodable codes from CDAs. Next we construct Approximately Universal (AU) codes and protocols for a few two-hop networks (e.g. diamond network). Such codes achieve the optimal Diversity-Multiplexing Gain Tradeoff (DMT) of the network irrespective of the fading statistics of the individual links.
June 2006 – May 2007, BE Project under guidance of Prof. M.K. Naskar.
We construct distributed algorithms for maintaining the topology of a Mobile Adhoc Network. This enables the various nodes in a network follow a fixed routing scheme helping them to eliminate the routing overhead. Such a network can be very effective as an emergency military or civil network deployed in an area of crisis.