cv

Apple

Machine Learning Intern, RF Systems

• Tabular generative models for synthetic data generation to model rare scenarios.

Samsung Research America

AI Research Intern, Standards and Mobility Innovation Lab

• Sequence modeling for design of polar codes with Transformers using policy gradient methods.

Qualcomm Research India

Research Engineer, WiFi Systems Team

• Developed and deployed multiple transceiver algorithms for next-generation WiFi chipsets.

PhD

University of Texas at Austin

Electrical and Computer Engineering

Bachelors and Masters

Indian Institute of Technology, Madras

Electrical Engineering

LightCode: Light Analytical and Neural Codes for Channels with Feedback

IEEE Journal on Selected Areas in Communications (JSAC)

Introduces LightCode, a hybrid analytical and neural code for channels with feedback. Achieves performance gains over existing feedback codes.

Nested Construction of Polar Codes via Transformers

Proceedings of the 2024 IEEE International Symposium on Information Theory (ISIT)

Presents a method to construct nested polar codes using Transformer models. Demonstrates performance improvements over conventional polar codes.

Exploring Explainability in Video Action Recognition

3rd Explainable AI for Computer Vision (XAI4CV) Workshop at CVPR

Explores explainability techniques for video action recognition models. Proposes novel visualizations and explanations to understand model decisions.

Task-Aware Distributed Source Coding Under Dynamic Bandwidth

Advances in Neural Information Processing Systems (NeurIPS)

Introduces a task-aware approach for distributed source coding in dynamic bandwidth environments. Utilizes deep learning to optimize bandwidth usage and improve coding efficiency.

DeepPolar: Inventing Nonlinear Large-Kernel Polar Codes via Deep Learning

Proceedings of International Conference on Machine Learning (ICML)

Proposes DeepPolar, a method to invent nonlinear large-kernel polar codes using deep learning techniques. This approach outperforms traditional polar codes in terms of error-correcting performance.

Learning RL-Policies for Joint Beamforming Without Exploration: A Batch Constrained Off-Policy Approach

arXiv preprint

Develops a reinforcement learning approach for joint beamforming without exploration, using batch-constrained off-policy methods to optimize performance.

Interpreting Neural Min-Sum Decoders

Proceedings of the 2023 IEEE International Conference on Communications (ICC)

Provides insights into neural min-sum decoders by analyzing their decision-making process. Proposes a method to interpret these models for better transparency.

Compressed Error HARQ: Feedback Communication on Noise-Asymmetric Channels

Proceedings of the 2023 IEEE International Symposium on Information Theory (ISIT)

Proposes a compressed HARQ feedback mechanism for noise-asymmetric channels. Achieves better reliability and lower overhead than traditional HARQ.

TinyTurbo: Efficient Turbo Decoders on Edge

Proceedings of the 2022 IEEE International Symposium on Information Theory (ISIT)

Proposes TinyTurbo, an efficient turbo decoding method designed for edge devices. Demonstrates low latency and reduced resource usage.

Ultra Low-Rate Neural Image Compression

• Developed an ultra-low rate (<0.1 bpp) compression framework leveraging vision foundation models.
• Designed a novel cross-attention aggregation technique to improve the alignment between input image and reconstructed image using textual captions as side information.
• Improving realism-fidelity trade-off in reconstruction using RLHF guidance and preference datasets.

Synthetic Datasets using Tabular Generative Models

• Developed and trained a diffusion-based foundation model for RF calibration, focusing on synthetic data creation to model failures at the receiver.
• Reduced data collection needs by over 10× through the use of high-quality synthetic tabular datasets.
• Improved regression performance by training on synthetic data, achieving ∼22% reduction in MSE.

Improving In-Context Learning (ICL) in LLMs using Structured Noise

• Developed techniques to enhance ICL performance by improving the separation of demonstrations.
• Proposed a method to select the optimal separator by analyzing perplexity for each demonstration.
• Formulated an explanation for the empirical observations using Bayesian inference.

Data Augmentation using Generative Models

• Explored parameter-efficient fine-tuning methods for text-to-image models for data augmentation.
• Demonstrated gains of up to 3.4% in classification accuracy by augmenting the true datasets with synthetic images generated using low rank approximation (LoRA) with DreamBooth.

Construction of Polar Codes using Sequence Modeling

• Modeled Polar code construction as a sequential decision making problem and designed a nested construction technique using transformer models and policy gradient methods.
• Demonstrated significant gains (up to 0.8 dB) compared to patented Polar code in 5G-NR standards.

Task-Aware Variable Rate Compression of Distributed Sources

• Designed distributed representation learning algorithm to optimize compression for downstream tasks.
• Proposed a dimensionality reduction technique to encourage low-rank representations, allowing variable-rate compression using a single model.