Current Projects
I am always on the lookout for motivated and hardworking students.
General prerequisites for working with me: You should be motivated to work on mathematical aspects of problems and have good analytical reasoning and communication skills.
If you want to work with me, you would require strong foundations in probability and linear algebra/matrices, mathematical aptitude, comfort with programming (python is preferred, but the language can be picked up later).
If you want to work with me on a specific project, please send me an email describing your interests and background, and a copy of your latest CV.
Distributed inference and learning with limited communication
Classical approaches to statistical inference and machine learning assume that all data is available at a central location.
However, with recent concerns about privacy has brought about new approaches like federated learning.
If we take the example of learning user preferences (rating movies, for example), it would be helpful if a central server could learn the model without accessing all user data.
In addition to privacy issues, there is also the cost of communicating data from each user to the server.
Here is a cartoon by Google with a non-technical introduction to federated learning.
Distributed estimation also has other applications. Take the example of a wireless sensor network where we have multiple low cost sensors measuring certain parameters, and a central server (sometimes called a fusion center) wants to infer something from these measurements (maybe the average, or some such function of all the measurements).
The link from the sensors to the server will in general be noisy and communication-limited. In such a scenario, how do we design a set-up to perform reliable inference at the server?
See this paper and this one for recent work on distributed mean estimation with limited communication. See this paper for the broad scope.
The problem is rooted deeply in estimation theory, and has applications in signal processing and machine learning.
Requirements
- Comfort with mathematics: The project will involve some amount of mathematical derivations (mainly probability) and simulations/implementations on python. The student is expected to be comfortable with mathematical reasoning, and will have to take courses on probability, statistical inference, optimization and machine learning at IIT Hyderabad.
- Comfort with programming in one of C/C++/python/matlab: The project involves a significant amount of programming in python, but the language can be picked up later.
- Most importantly, enthusiasm to work on these problems and learn new things.
- Prior exposure to statistical inference and machine learning is a plus, but not essential.
Weakly supervised building identification from satellite imagery
Object detection in images and segmentation are classical problems that are studied in computer vision and image processing. Detection of buildings and other objects in satellite imagery is helpful in digital addressing, finding land use statistics, and surveys. However, India has very diverse demographics, and obtaining high-quality annotated data is very expensive.
This project revolves around designing a framework for building detection and segmentation using limited annotations (such as point annotations). See this paper and this one on general object counting/detection using point supervision.
Requirements
- Comfort with programming in python and some exposure to deep learning
- Any background in computer vision is a plus
Other projects
I am also generally open to work on other problems. See this page for some of my recent research interests.