Projects

 Research Projects

1. Learning to Gather Information With Large Language Models (March '25 - Ongoing)

Guide: Prof. Sanmi Koyejo, Computer Science, Stanford University

2. Treatment Effect Estimation in Causal Inference (July '23 - June '24)

Guide: Prof. Sunita Sarawagi, Computer Science, IIT Bombay

I worked with Professor Sarawagi and Lokesh on developing a novel pair loss for treatment effect estimation, which uses approximate counterfactuals from the training data. We extended the CATENets code and compared our loss against baselines using JAX.  This technique attained around 15% reduction in PEHE error over cutting-edge methods, and the gain was statistically significant across several different model architectures and datasets. This work was accepted at ICML '24 at Vienna, Austria.

I then worked on utilizing counterfactual simulators to extract pre-treatment representations without access to true counterfactuals. Moreover, I am developed a theoretical framework to understand how imperfect counterfactual simulators can be used to estimate treatment effect from post-treatment covariates. This work was published at the NeurIPS Causal Representation Learning Workshop '24 and its extension was published at TMLR.

3. Client Selection for Communication Efficient Federated Learning (May '23 - Feb '24)

Guides: Profs. Petar Popovski & Shashi Raj Pandey, Aalborg University, Denmark 

During my summer break in 2023, I interned at the Connectivity Section at Aalborg University. I worked on improving upon existing Shapley-Value based client selection algorithms for federated learning to increase communication efficiency. By using an efficient truncated Monte-Carlo sampling approximation I was able to make the algorithm computationally tractable. Our method achieved faster convergence and higher accuracies nearing centralized training for varying levels of privacy, data, and systems heterogeneity under timing constraints compared to several state-of-the-art methods. The algorithm also shows much more stable convergence with a lower standard deviation in test accuracy. This work resulted in a publication at IEEE Networking Letters.

4. Joint Probability Estimation Using Low-Rank Tensor Decomposition (May '22 - March '23)

Guides: Prof. Ajit Rajwade, Computer Science, IIT Bombay & Karthik Gurumoorthy, Walmart

The goal was to extend the existing literature on density estimation in the low sample regime using low-rank tensor factorisation. We proposed a novel scheme for computing radon projections of continuous 2-dimensional marginals that reduced the problem to 1-dimension, achieving lower sample complexity over previous estimators. We obtained lower Jensen-Shannon divergence compared to Gaussian Mixture Models and other methods on several different density families including mixtures of Laplacians, Gaussians and discrete variables. This work resulted in a publication at EUSIPCO '23 in Finland.

Ongoing Course Projects

Courses starting from September

Completed Course Projects

1. Automatic Speech Recognition (Spring '24)

Presented the paper Voicebox by Meta research

2. Image Synthesis (Spring '24)

Presented a paper on neural implicit surface queries for graphics

3. Organization of Web Information (Spring '24)

Implemented a fact retrieval system to verify claims using Wikipedia data 

4. AI, Data, and Policy (Fall '23)

Literature survey on policy for bias in medical AI 

5. Estimation and Identification (Fall '23)

 Implementing error-state Kalman filter using IMU data in MATLAB

6. Advanced Machine Learning (Spring '23)

Extended score-based diffusion models with Hamiltonian Monte Carlo and wavelet-based sampling schemes

7. Optimization in Machine Learning (Spring '23)

Combined submodular optimization for efficient Neural Architecture Search with information-theoretic measures (PRISM) for dynamic subset selection; improved model accuracy over existing methods for NAS

8. Error Correcting Codes (Fall '22)

Experimentally verified performance guarantees of Raptor codes in Python

9. Wavelets (Fall '22)

Analysed performance of wavelet pooling in denoising autoencoders using Keras

10. Neuromorphic Engineering (Fall '22)

Designed a Speech Recognition Liquid State Machine in MATLAB

11. Advanced Image Processing (Spring '22)

Implemented paper on video denoising using singular value thresholding for low-rank matrix completion. Extended technique to various Schatten-p norms; applied to image inpainting and recommender systems

12. Microprocessors (Spring '22) 

Designed a pipelined RISC microprocessor in VHDL using Intel Quartus

13. Programming for Data Science (Fall '21)

Performed exploratory data analysis on the utilization of energy sources around the world

Other Projects

1. Time-Frequency Methods for Microdoppler Signal Separation (Spring '22)

Research under Prof. Vikram Gadre, EE, IIT Bombay; analysed Fourier-Bessel Transform and Doppler-focusing techniques for signal separation; Inverse Radon Transform for frequency estimation from noisy spectrograms

2. Brain-Computer Interface: Institute Technical Summer Project (Summer '21)

Simulated BCI with EEG data for motor movement using a CNN; chosen in top 6 projects across 43 teams

3. Reinforcement Learning: Snake Game (Summer '21)

Developed snake game in PyGame; implemented RL strategies like SARSA and Q-Learning to train the snake

4. Psychology X Reinforcement Learning (Reading Project, Summer '21)

Studied connections between psychology and RL, existing mathematical models of learning