Srinitish Srinivasan's Portfolio

Explore my research, projects and more

About Me

Hello! I'm Srinitish Srinivasan, a passionate researcher and innovator in the field of Artificial Intelligence and Mathematics. My skills involve Linear Algebra, Statistics, Calculus, Machine and Deep Learning apart from being proficient in several programming languages and development frameworks such as Python, C, C++, Javascript, PyTorch, Tensorflow, Flask and Django. My current research interests lie within the domain of Graph Theory and Graph Neural Networks especially Dynamic Graphs. I love reading new research, connecting with different people and exchanging new ideas. If I am not working on some deep learning model, I am either listening to K-Pop or watching a K-Drama :))

Research Experience

University of West of England

Statement: Developing a toolkit for user acceptance of various languages through a qualitative and quantative approach towards language modelling.

  • Research on developing embeddings for Indian Languages such as Tamil by analysing spectrograms. Further, the study focuses on a computational model that builds towards developing a toolkit for user acceptance of language technologies.
  • Analyse an individual user's reaction to different languages and study their responses using Machine and Deep Learning methods.

University of Lincoln

Statement: Improving the domain accuracies of object detection models such as YOLOv5 on Out of Distribution(OOD) samples through Self Supervised Learning and Explainable AI Techniques

  • Conducted research on Self-Supervised Learning and pseudo-labeling using YOLOv5. This involved pre-training YOLOv5 and YOLOv3 backbones to generate pseudo labels for multiple objects within an image.
  • Worked on mitigating Out-of-Distribution (OOD) errors in object detection during inference time and implemented a method to transfer backbone weights from YOLOv5 to Ultralytics YOLOv5 following custom pre-training.

Vellore Institute of Technology, India

Statement: Side Effect Prediction through Adverse Drug Reaction Analysis by modelling Drug-Drug Interactions using Graph Neural Networks and Self-Supervised Learning

  • Designed a comprehensive pipeline to fetch SMILE strings and convert them into featurized Molecular Graphs. Developed Graph Variational Autoencoders and created a Dual Branched Graph Neural Network framework to model drug-drug interactions.
  • Implemented a contrastive training method to enhance GNN resilience against Adversarial Attacks.

Center of Cyber Physical Systems, Vellore Institute of Technology, India

Statement: Semantic Segmentation of Crop and Weed for Precision Spraying.

  • Developed a technique for semantically segmenting crops and weeds using hyperspectral images during early growth stages. Evaluated several semantic segmentation loss functions and constructed a modified U-Net architecture.

Projects

Molecular Property Prediction using Graph Isomorphism and Contrastive Self-Supervised Learning

  • Developed an adversarial pre-training approach based on model perturbation and information loss to robustly pre-train a Graph Convolution-based encoder.
  • Integrated the pre-trained model with a Graph Isomorphism framework to predict molecular properties, achieving state-of-the-art results on MoleculeNet benchmarks.
Molecular Property Prediction Code Adversarial Attack Defence Code

Real Time Surgical Smoke Detection using Graph Neural Networks and 3D CNNs

  • Designed a feature extraction system combining a Graph Neural Network and a 3D Convolutional Neural Network to detect smoke in surgical videos.
  • Used temporal frame positions to form the Graph Adjacency Matrix, exploring a broad range of features dynamically during training.
View Project

Differential Equations using Equilibrium Driven Neural Networks

  • Implemented a numerical solution using neural networks to solve ordinary and partial differential equations while considering mathematical constraints.
  • The implemenation takes consideration of boundary conditions such as Dirichilet and Neumann Boundary Conditions for Partial Differential Equations.
View Project

Self Supervised Psuedo-labelling using Autoencoders built on YOLOv5 Backbone

  • Pre-trained the YOLOv5 backbone to localize multiple objects within an image and developed a pipeline to transfer the pre-trained backbone weights to Ultralytics YOLOv5 for fine-tuning.
View Project

Mitigation of Spurious Correlations in YOLOv5

  • Identified patterns in misclassified instances of YOLOv5 and developed instance-based mitigation techniques, including Image Inpainting and Weighted Box Fusion (WBF).
View Project

Deep Learning Based Asymmetric Cryptographic Scheme using GANs

  • Created a deep learning-based technique using GANs to encrypt medical images. The network comprises 2 generators and 2 discriminators to maintain cycle consistencies.
View Project

Plastic Detection using Reflectance Images of Hyperspectral Bands

  • Performed a deep learning-based mapping from RGB space to reflectance images followed by reflectance-based thresholding for plastic detection.
  • The project involves an initial spectral analysis and dataset creation by mapping the multispectral images to its reflectance images using the reflectances of each band against both surface and plastic.
View Project

Tamil Speech to Text using Spectrogram Analysis

  • Developed during a 72-hour hackathon, involving speech processing techniques like silence removal, L1 Formant analysis, and Short Time Fourier Transforms to segment Tamil audio into phonetics,thus enabling us to win the hackathon.
View Project

Publications

Detecting Side Effects of Adverse Drug Reactions Through Drug-Drug Interactions

Made use of Dual Path Graph Neural Network Models with Self Supervised Learning using Graph Variational Autoencoders to model Adverse Drug Eeactions(ADRs) and detect the side effects it causes.

View Publication
View Source Code

Autoencoder based approach for the mitigation of spurious correlations

Developed a technique using Autoencoders and Image Painting to analyze and mitigate spurious correlations in multi-distribution samples. Further, we analysed the patterns occuring in the Out of Distribution(OOD) samples using autoencoders built on the YOLOv5 backbone through a self-supervised pseudolabelling technique as a method of model explainability.

View Publication
View Source Code

Color and Brush Stroke Pattern Recognition in Abstract Art using DCGANs

Incorporated DC-GANs to analyze brush stroke patterns in Abstract Art. The study further involved a random walk into the latent space to develop logical relationships between various colors used by artists.

View Publication
View Code

Seminars and Talks

Temporal Graph Neural Networks

Date:March 2024

Organisation:Vellore Institute of Technology

Gave a talk on temporal graph neural networks inclusive of representation of dynamic graphs, developing embeddings, mathematics of graph operations and PyTorch implementations.

Link to Code: GitHub
Link to Presentation: Slides

Google IO Extended: Generative AI

Date:June 2023

Organisation:Google Developers Student Club

Gave a presentation on recent advancements on Googles's recent advancements on Generative AI such as Palm API, MediaPipe and representation of words and sentences as embeddings.

Details: Event Details