Projects

Developed a custom CUDA kernel for efficient 3D matrix multiplication, and integrate it into Python using Pybind11, aiming to enhance BERT model inference performance.

Trained Graph Convolution Networks as teacher models on MUTAG, PTC & PROTEINS datasets. 

Distilled knowledge to GraphSAGE & Graph Attention Networks achieving > 50% parameter reduction.

Implemented Stable Diffusion & GANs for text-to-image generation for scene creation with user prompts. Generated custom embeddings for scene-specific generation of images using CLIP Transformer.

Implemented Fully Convolutional Networks (Naive - FCN32) to perform semantic segmentation on CMP Facade database.

Improved performance by using skip connections with FCN-8 architecture to recover high level features to achieve a final IoU(intersection over union) of 0.46 & average pixel-wise test accuracy of 72%. 

Various collaborative filtering techniques from similarity-based, latent factors, Factorization Machines to Neural Collaborative Filtering are used to build a Movie Recommendation System using the Amazon Movie Reviews Dataset.

Implemented Variational Autoencoders and Conditional Variational Autoencoders (CVAE) for building a latent space for hand written digit generation.

Using image processing, convolutional neural networks and OpenCV, the task of hand gesture recognition has been successfully implemented along with a user interface built in Python (Tkinter).

Developed custom GANs for creating a suitable latent space for generating images with different artistic styles with FID score 0.029 and Inception score 2.57.

This project aims to analyze the situation of COVID by studying and analyzing the COVID cases along with the vaccination rates for different manufacturers using time series forecasting techniques like - Regression, ARIMA, LSTMs etc.