Deep learning for pereception and prediction|  Sept 2022 - Present
Plus, Cupertino

Working on design, implementation and deployment of deep learning models for tacking problems in perception and prediction stack of autonomous driving

Object detection and tracking| Internship |  May - Aug 2021
Plus, Cupertino

Worked on state of the art image based anchor-free object detection and tracking implementation in PyTorch.

Machine learning for Physical Layer Wireless Communication | Aug 2018 - Present
UC Berkeley | Collaborators: Josh Sanz, Caryn Tran, Kailas Vodrahalli, Prof. Anant Sahai

Designed a blind interactive learning protocol for modulation schemes in the multi-agent setting without codesign. Experimentally verified the universality and robustness of the protocol and showed that it achieves bit error rates similar to the optimal baseline. Working on integrating other parts of the communication pipeline including equalization and error correcting codes to enable end-to-end learning of communication schemes. 

Learning Stabilizing Control under Multiplicative Noise | July 2019 - January 2020
UC Berkeley| Collaborators: Moses Won, Prof. Gireeja Ranade

Exploring use of neural networks to discover control strategies for stabilizing a system under multiplicative noise. Proposed an architecture and training procedure tailored for the control problem that enables the network to generalize and output controls for rollouts longer than the training horizon. Showed that the neural network based control strategybeats current best known strategies including optimal linear strategies. 

Classification versus Regression for Minimum Norm Interpolating Solutions | August 2019 - August 2021
UC Berkeley | Collaborators: Vidya Muthukumar, Adhyyan Narang, Prof Anant Sahai; UC Berkeley, Prof. Daniel Hsu; Columbia University, Prof. Mikhail Belkin; Ohio State University

Analyzed the classification and regression loss of minimum norm interpolating solutions in the overparameterized setting. Related the classification error to statistical signal processing concepts of shrinkage and false-discovery and computed sharp upper and lower bounds for these quantities. Showed the existence of a regime where asymptotically classification performs well but regression does not.

Generalization for multiclass classfication| August 2021 - August 2022
UC Berkeley | Collaborators: Rahul Arya, Prof Anant Sahai

Analyzed the multi-class classification loss of minimum-norm interpolating solutions in an asymptotic overparameterized setting where both the number of underlying features and the number of classes scale with the number of training points.Proved that the the multiclass problem is ``harder'' than the binary one due to the relatively fewer training examples per class in the multiclass setting

Neural Network based Control for Witsenhausen problem | Jan-Aug 2018
UC Berkeley | Guide: Prof. Anant Sahai

Revisited the classical decentralized stochastic control problem of Witsenhausen using neural networks. Concluded that biasing architectures towards favorable solutions is required to escape the local minimas of the non-convex problem. Numerically showed that in higher dimensions it is possible to outperform the best known one dimensional strategy. 

Compressed Sensing in Radio Astronomy | Master's Thesis | July 2014 - May 2015
Indian Institute of Technology Bombay | Advisors: Prof. Sibi Raj Pillai & Prof. V. Rajbabu 

Thesis PDF : here

Scheduling Algorithms for Wireless Communication | Research Internship |  May - July 2013
Bell Labs Alcatel Lucent, Bangalore | Advisor: Anand Muralidhar

Analyzed the performance of scheduling algorithms for device to device communication in cellular networks. Implemented and tested a novel coloring algorithm in Python and concluded that it increases total network throughput by up to 70% as compared to existing schemes like CSMA/CA. Theoretically proved approximate optimality of the greedy coloring algorithm by obtaining bounds on its performance.