RESEARCH
Duke University
My research mainly focussed on the applications of information theory to real world problems. I worked on the intersection between information theory, machine learning, engineering, statistical physics and statistics.
FUNDAMENTAL LIMITS FOR COMMUNITY DETECTION
The problem of detecting the community structure of networks as well as closely related problems involving low-rank matrix factorization arise in applications throughout science and engineering. In recent years there has been a significant amount of work focused on characterizing asymptotically exact limits of the mutual information and reconstruction error. The theoretical formulas provide insight into the ability to recover the community structure in the network. Recent work focuses on network models with a high degree of symmetry or a small number of communities.
My work focused on the the fundamental limits of detection and recovery associated with a broad class of probabilistic network models, that includes the stochastic block model with labeled-edges. The formulas are described in terms of low-dimensional estimation problems in additive Gaussian noise that are numerically tractable. The approach used to obtain the formulas builds upon a number of recent theoretical advances at the interface of information theory, random matrix theory, and statistical physics.
BRAIN-COMPUTER INTERFACE AS A COMMUNICATION CHANNEL
The P300 Speller is a technology that enables people suffering from neuromuscular disorders to communicate with others using electroencephalography (EEG) measurements. Specifically it uses the P300 event-related potential which is elicited in response to specific stimulus events. Unfortunately the performance of the BCI is affected by the refractory effects which induces a temporal dependence on the EEG responses.
We proposed a model for the P300 speller as a communication channel with memory. By studying the maximum information rate on this channel, we gain insight into the fundamental constraints imposed by refractory effects. We then propose modifications that could improve the performance and reduce the impact of the refractory effects.
Publications
V. Mayya and G. Reeves, “Mutual information in community detection with covariate information and correlated networks”, 57th Annual Allerton Conference on Communication, Control, and Computing, 2019
H. Mathews, V. Mayya, A. Volfovsky, and G. Reeves, “Gaussian mixture models for stochastic block models with non-vanishing noise”, 2019 IEEE International Workshop on ComputationalAdvances in Multi-Sensor Adaptive Processing (CAMSAP), IEEE, 2019
G. Reeves, V. Mayya and A. Volfovsky. “The Geometry of Community Detection via the MMSE Matrix”, IEEE International Symposium on Information Theory, 2019.
V. Mayya, B. Mainsah, and G. Reeves. “Information-Theoretic Analysis of Refractory Effects in the P300 Speller,” Asilomar Conference on Signals, Systems, and Computers, 2016.
V. Mayya, B. Mainsah, and G. Reeves.“Modeling the P300-Based Brain-Computer Interface as a Channel with Memory,” Allerton Conference on Communication, Control, and Computing, 2016.
T. Schaller, J. Wheaton, D. Reilly, L. Pegorsch, V. Mayya, C. Bales, Z. Xu, and S. Barnwal ,“Review and Analysis of the Costs and Benefits of Trees in the City of Durham”, Issue 5, Duke Science Review.
PAST PROJECTS
BLUE DEVIL OCEAN ENGINEERING
Less than 5% of the ocean floor on our planet. The goal of this project was to develop a method to obtain a map of the ocean floor using a low cost method. We were a finalist in the Shell Ocean Discovery XPRIZE, where teams were required to launch from shore or air and make a high resolution bathymetric map of a 500 km2 area at 4 km depth.
Our approached involved using a hybrid drone to carry a pod filled with sensors to a particular location in the ocean. The pod descends into the ocean, using sonar pings to collect underwater data. once it reaches the ocean floor, it is winched back to the surface and picked up by the drone. The pod is then dropped at another location in the ocean to repeat the process.
This was large project involving multiple teams of students over several years. I managed several multidisciplinary teams during the development process, including a Data+ team in May-July 2017.
PATHOS
Breast cancer patients undergoing BCS have a one in five chance of requiring a reoperation, because surgeons cannot determine if the entire tumor has been removed before ending the surgery. Quite simply, this is because a standard pathological procudure called frozen section is unavailable for breast cancer surgeries, due to the inherent mechanical properties of breast tissue.
My teammates and I developed a medical device and method to prepare frozen sections of breast tissue, which was previously impossible to prepare. We collaborated with the Biomedical Engineering Department at the Johns Hopkins University and the Pathology Department at the Johns Hopkins Hospital.
I was involved in multiple stages of the development of this project: from conducting multiple stakeholder interviews with clinicians to developing a business plan and fund raising. I developed the UV resin used for stabilizing the tissue and ran experiments on human tissue to test efficacy of the solution.
HEMOGLOBE
It is estimated that nearly 41% of pregnant women are anemic. Maternal anemia can significantly increase the risk of an adverse outcome for the mother and child.
HemoGlobe is a medical device aimed at tackling widespread maternal anemia in the developing world through a noninvasive hemoglobin measuring device. It utilizes the health worker's phone to measure hemoglobin levels using photoplethysmography and sends the anemia level to a public health database in real time, which would be available to public health officials.
I managed a team of students working on product development. I worked on the development of machine learning algorithm to estimate hemoglobin levels from plethysmographic data. I also planned and conducted data collection studies in different hospitals in India and Nepal, coordinating with our local partners in these locations.