Projects at Princeton
Constructing online smooth safety filters using differential dynamic programming - Link to full paper
Learning to control using a convex combination of controllers - Link to full paper
Learning PID controllers by automatic differentiation - Link to full paper
Safe learning using control barrier functions - Link to full paper
Planning for fibers of Subaru Prime Focus Spectrograph - Link to full paper
Summer internship project on reinforcement learning at Nokia Bell Labs - Link to arxiv draft
Adversarial training for robust reinforcement learning ( Will blog about it by 2026! )
Talk on graph neural network theory in ORFE deep learning theory seminar - Link to slides
Some important course projects include:
Regularized matrix factorization
Please find the report attached.
Sample complexity of model-based vs model-free reinforcement learning
As part of theoretical machine learning COS 511, we compared the sample complexity of state-of-the-art model-based and model-free reinforcement learning algorithms on the gridworld, cartpole and mountain car environments. Please find the report attached.
Comparison of optimization algorithms for Machine Learning
As part of the course on optimization for machine learning COS 598D, I compared five optimization algorithms, namely SGD, Momentum, Variance Reduction, Adam and second-order methods on four machine learning datasets. Please find the report attached.
Tennis Game Playing Agent: Modeling and Analysis
As part of the course on stochastic optimization ORF 544, I modeled a tennis game playing agent using tools from stochastic optimization and reinforcement learning using an innovative modeling framework invented by Prof. Warren. B Powell. Please find the report attached.
Projects before Ph.D
I have done a wide variety of projects right from my undergraduate days. Some of them have led to publications.
Efficient Deep Learning for Videos
I worked under Prof. Balaraman Ravindran and Prof. Anand Raghunathan (Purdue university) on achieving computation and energy savings when applying deep learning to videos. We primarily focused on object detection in videos. We leveraged the temporal similarity between frames in a video to achieve computational savings for video processing. This work is available in arxiv ( https://arxiv.org/abs/1809.01701 ) and accepted at the Proceedings of The ACM India Joint International Conference on Data Science and Management of Data (CODS-COMAD’19). This got featured in TechXplore. ( https://techxplore.com/news/2018-09-fast-videos-region-of-interest.html )
Frequency Assignment in Open environments
I worked at the Center of Excellence in Wireless Technology on reformulating the traditional frequency assignment problem in the literature to encompass the SNR and throughput requirements of the end users. We have patented this work and this work has been accepted at the National Conference on Communications 2018.
Direct position tracking in GPS using the Vector Correlator
For my MS thesis, I worked with Prof. Grace Gao (University of Illinois at Urbana-Champaign) on a novel positioning technique known as direct positioning. I proved the advantages offered by direct positioning, namely robustness to noise and multipath.
Industry project at Google - Centimeter accurate positioning of StreetView cars using carrier phase GPS
I did an internship at Google from May-August 2014. I was given the challenging task of obtaining centimeter accurate positioning of Street View cars using carrier phase GPS. We achieved the goal when the car was under cruise control. However, in urban environments, when the car was accelerating and decelerating often, it was difficult to obtain lock on the carrier phase. Due to frequent loss of lock, the integer ambiguity resolution algorithm did not have enough time to resolve the integers accurately. I recommended changes to be made to the current setup in order to successfully use carrier phase for Street View car positioning.
Privacy protection for location-based services
In this project with Prof. Grace Gao, we opened up the GPS black box and leveraged the range measurements for private proximity detection. This work was accepted in the proceedings of the 27th International Technical Meeting of the Satellite Division of the Institute of Navigation, ION GNSS 2014 (link). Further, this work also resulted in a journal publication at IEEE Aerospace and Electronic Systems (link) and an inside GNSS magazine article (pdf).