Distributed Machine Learning on Data Streams
Royal Institute of Technology (KTH), Sweden
Prof. Christer Norström / Dr. Mohammed El-Beltagy / Prof. Šarūnas Girdzijauskas
Pose estimation/Motion modelling using deep learning, Real time feedback on poses, Reinforcement learning, Distributed Machine learning
After completing my bachelor in engineering (Computer Science) from Indian Institute of Engineering Science and Technology, I worked as a consultant for PricewaterhouseCoopers for three years. Thereafter, I pursued my dual master degree (with EIT Digital) from University of Rennes1 (France) and TU-Berlin (Germany) in Cloud computing with specialization in Cloud and data analytics. My master thesis dealt with detecting redundancy in textual relational dataset using NLP and deep learning. On completion of my master in 2017, I started working in a start-up as a data scientist. I will start working as a Marie Curie PhD fellow, in the RAIS project with Racefox and KTH. This position complements my skills and my research interests perfectly.
My current research will focus on distributed machine learning on data streams. I would like to explore the scope of distributed machine learning in resource constrained such as edge computing infrastructure. In addition to that I am also quite interested to research on reinforcement learning. I would very much like to apply my research work in the paradigm of outdoor sports, such as soccer, skiing, and others.
My present research work focusses on motion modelling using artificial intellgience. Presently I am exploring the domain of human pose estimation in a minimalistic approach. Primarily, we are focussing on estimating human pose only from one IMU sensor providing accelerometer readings. This approach would enable a low cost pose estimation technique that can easily be integrated with any IOT ecosystem revolving around accelerometer readings (e.g. mobile phone) and will greatly contribute towards personalized coaching feedback for athletes (for whom a good pose equivalents to good technique. Furthermore, it can be extended to multiple other use cases of health and well being. We will build a system leveraging machine learning techniques that would take in accelerometer signals from running users of Racefox and estimate poses for those users in a fast and effective manner. Furthermore, we plan to extend our research to explore the possibility of training our deep learning/machine learning models over a distributed infrastructure and experiment with different edge settings for deployment of our model. Additionally, we also plan to explore the scope of reinforcement learning for personalization and recommendation.
1. Master thesis at TU Berlin, Big data management group. “Detection of textual inconsistencies in relational dataset”
2. Van Kempen, A., Crivat, T., Trubert, B., Roy, D., & Pierre, G. (2017, April). MEC-ConPaaS: An experimental single-board based mobile edge cloud. In 2017 5th IEEE International Conference on Mobile Cloud Computing, Services, and Engineering (MobileCloud) (pp. 17-24). IEEE.