Investigating the motivation behind the desire for robot automation of different social groups. Worked on statistical modelling & analysis and finding correlations.

[Semester Project] Developed & deployed an audio-based trash classification model on a Franka Panda robot for automated waste segregation, using ROS for system integration and Deoxys for real-time control. Gained experience in Mujoco and Robosuite.(link)

Developed a method to control text-to-image generation system using diffusion models & guided adversarial prompts for effective training data generation. Achieved data-efficient generations, outperforming others in low-data regimes across various tasks & distribution shifts.

Working on part-based generation of multi-category objects using a recursive neural network based autoencoder utilizing the PASCAL-Part Dataset. Worked on tree-based dataset creation and data preprocessing.

Worked on mathematical equation recognition in terms of first year calculus problems to aid instructors in the assessment of students’ assignments and exams. Utilised CROHME 2019 dataset and AIDA Calculus dataset. Built custom dataset comprising of ~1500 images (99 claases). Implemented YOLOR to localise mathematical symbols within an image and further working on classification model

Worked on Squeeze-and-Excitation ResNet 101 and feature genuinization encoder-decoder model for fake/spoofed audio detection using conventional features. Used ASVSpoof 2019 dataset for evaluation of model. Proposed network outperforms shows higher accuracy over other classification models using Mel Spectrogram for all sound events.

Worked on Deep Belief Networks and Variational auto-encoder model for collaborative filter with explicit feedback. Model was trained and evalutated using MovieLens 1M and Netflix(1M subset) datasets.The proposed model outperformed state-of-the-art recommendation approaches(2020) in standard evaluation metrics.

Worked on the ongoing project ”Requirement of real-time compensation for turbulence and detecting moving objects” for removing heat turbulence from video footage. Evaluated existing models based on CNN, autoencoder, GAN and GNN for literature review.

Worked on standardizing MLOps to deliver an efficient workflow for managing the life-cycle of machine learning models from selection, training to inference, minimizing any duplication of previous tasks, all adhering to versioning of neural network models and training their datasets.