top of page
Abstract Cubes Structure

Research

My research focuses on an intersection of math, biomedicine, artificial intelligence, statistics and mathematical explainability of AI models. My work has been conducted both independently and at nationally renowned universities, in association with top academics from the country, proposing novel methodologies, architectures and ideas along the way. 

Click on the "Paper and Abstract" Icon to view a pdf of all papers.

EENN

Proposed a novel architecture called EENN which utilises an ensemble of neural networks. Biologically analysed genetic information obtained via SHAP: mapped CpG probes to gene symbols and subsequently to biomarkers, conducted protein-protein interaction analysis and survival analysis using Kaplan-Meier plots; wrote PyTorch implementation for the architecture. Wrote PyTorch implementation for the architecture with application on lung cancer classification. 

Publication: Under Review

Paper @ Research Square

Conducted at:

Paper & Abstract

Git Repository

GitHub

University of Delhi

ViT

Understood the mathematical intuitions of this state-of-the-art architecture, implementing a custom ViT in PyTorch. Proposed changes in the vision transformer after analysing the mathematical constructs of the model, experimenting with mathematical formulations in the self attention architecture after understanding the logical intuitions behind it. 

In Progress. 

Conducted at:

University of Delhi

Leaf disease detection via a Swin Transformer

Utilised a custom Swin Transformer architecture followed by a shallow feed-forward network to classify 38 classes of plant disease, and provided mathematical interpretation through a LIME module which extracted 4 most contributing super-pixels from single images. 

In Progress. Draft Available. 

Mentored at:

Paper & Abstract

University of Delhi

Real-time waste classification

Developed an artificial intelligence model to classify waste objects into biodegradable and non-biodegradable, recommending disposal procedures. Developed a novel front-end so as to not require server space or image storage, providing classification in seconds using AV classes in Swift.

Publication: Journal of Student Research

Top 15 Finalist @ IIT-Techfest

Associated with:

(See "Projects")

Paper & Abstract

Multiomic Lung cancer classification

Developed a novel methodology based on residual deep learning; implemented using Python, for different types bio-molecular omic data, developing a ResNet-inspired deep residual learning architecture, which are then fused using soft voting, and further analysed through a SHAP to identify relevant biomarkers. 

Conference Presentation: ASCS'23 (1 of 4 selected papers)

Proceedings: CERN's Zenodo

Publication: Under Review

Conducted at:

University of Delhi

Abstract

Conference Proc.

(Page 21)

Feature Extraction

Explored mathematic intuitions behind various models for novel and standard image transforms and feature extraction methods, applying on a dataset of plant disease images, further fed into a sequential model for comparison. Wrote custom implementations for standard and novel mathematical feature extraction models; applied to the problem of leaf disease classification

Publication: Journal of Mathematical Techniques and Computational Mathematics.

Conducted at:

Paper & Abstract

University of Delhi

Integrating human vision perception in ViT

Explored a novel method to modify datasets based on the way humans perceive vision. Explained the logical and mathematical intuition behind this idea, and applied it to solve the problem of waste segregation through vision transformers.

Presented @ IIT-Bombay

Top 15 Finalist @ IIT-B's Techfest

Presentation @ ICAIA'24

Publication : Springer

Conducted at:

Paper & Abstract

image.png

IIT-Delhi

Heart disease detection

Performed a comparative analysis of various machine learning classifiers on the task of heart disease detection, explaining the mathematics of hyper parameter tuning, and implementation of models in the front-end. 

Publication: International Journal of Innovative Science and Research Technology

Associated with:

(See "Projects")

Paper & Abstract

bottom of page