Projects
Evaluation of Kernel-based approach for Non-linear Causal Clustering
I am currently collaborating with the Neuroinformatics Lab of Universität Wien, headed by Prof. Moritz Grosse-Wentrup, to evaluate kernel-based approaches for non-linear causal clustering of samples. I am also performing ablation studies on the distance correlation kernel to better understand how well it is suited to find the clusters with unknown unconditional independence equivalence classes in the sample data.
Belief Elicitation on the Impact of Algorithmic Decision Making
I am currently collaborating with Prof. Ujwal Gadiraju of Technische Universiteit Delft, Netherlands to elicit human presumptions and behaviour in response to Algorithmic Decision Making in a modified Ultimatum Bargaining setting.
Deployed a DRF backend, Angular 12 frontend, PostgreSQL DB coupled with Redis + Celery task management application on a Heroku + GitHub deployment pipeline.
Feature Gating for Deeper Graph Convolution Networks
Worked under Prof. Shanmuganathan Raman of IIT Gandhinagar to find better explanations of single-instance predictions of Graph Neural Networks by identifying optimal subgraph structure and feature set.
To tackle this problem we introduced the idea of feature gating for assigning importance to nodes and node features in GNNs. To this end we showed that using an identity mapping and a modified form of residual connection along with feature gating we create deep GCN models which achieve state-of-the-art results for semi-supervised node classification.
[Work is currently under review in the IEEE Signal Processing Letters]
Single Exposure Coded Snapshot
Implemented a MATLAB solution for coded aperture compressive temporal imaging to recover a sequence of frames from a single coded-snapshot to achieve temporal gains in video acquisition.
This is an implementation of Video from a Single Exposure Coded Photograph. Do check this paper out! Its really interesting.
Compressed Sensing Over Graph Structures
Implemented the DICeNod algorithm for identifying nodes with the highest betweenness centrality in social networks using compressive sensing without full knowledge of the network topological structure. Using the limited information and general knowledge of social networks we approximated the global betweenness centralities by reconstructing the egocentric local centrality vector to find the top-k information flow hot spots.
X-Ray Anomaly Detection using CNNs
Led a team of 4 as a part of the Institute Technical Summer Project 2020, responsible for building and deploying a WebApp and an Open-API Endpoint for hosting a 5 model ensemble which accurately classifies and localizes 5 common thoracic diseases in a heatmap, from Chest X-Rays with an average AUC of 0.915 which is quite close to the SOTA of 0.94, to reduce the False Negative cases among radiologists.
Unplag - Plagiarism Checker
Worked in a team of 3 to build an effective web-based pairwise plagiarism checker for source code files. We applied TF-IDF metric on preprocessed Abstract Syntax Trees for C++ and Python using Clang and Radon. Applied the same metric on cleaned root-stemmed data for plain textual inputs. Finally we integrated the computational model with a multithreaded Django REST Framework backend, an Angular 9 frontend and a NodeJS CLI with a secure nginx file-server and stateless JWT authentication.
Gestures For 3D Space (Seasons of Code)
A Seasons of Code project under WnCC IITB for designing both the front and back-end of a real-time gesture detection module using Convolutional Neural Networks and my own dataset. The model which I have used has an AUC of 0.9959 and it has a pretty good reception when it comes to real world application.
How did I make my datasets ? Check it out here
Where are the model training codes ? Check it out here
Neural Networks and Deep Learning (Summer of Science)
A nice reading project, that I had taken up in my summer break of 2020 as a part of the Summer Of Science, MnP IIT Bombay.
This report covers some basic principles and techniques of Neural Networks and ML, ranging from solving simple Linear Algebra problems to creating your own Deep Learning library. A beginner's guide to AI.
Siamese Neural Network for One-Shot Learning of Hand Gestures using Minimal Datasets
The aim of this project build Siamese Neural networks for hand gesture detection and to study the effect of dataset volumes and network depth on the Neural Net's performance.
The Angular Django SOCKET.IO cross-platform Messenger
This is a light-weight cross-platform chatting application built on a Django REST API Framework as Backend (deployed on Heroku), Angular Frontend (deployed on GitHub Pages) and a SOCKET.IO WebSocket server (I'm glad Heroku supports sockets 😛).
I have achieved cross-platforming using a Cordova backend. This app was just for providing myself a full-stack development experience from ground-up, and I don't know if I will continue maintaining or developing it further.
Image Enhancer (Super Resolution using SOTA GANs)
An Open-Source, PyQt5 based Desktop App for up-scaling images using state-of-the-art GAN techniques.
This is something which I created purely out of my own interest as a hobby-dev to learn more about "Super Resolution" (as they are shown in detective movies 😛) and also about Generative Adversarial Networks !