Projects

[projects-header-1]

Multimodal Sentiment Analysis

This project delves into different methodologies for Multimodal Sentiment Analysis, with a focus on fusion techniques that integrate information from various modalities to improve sentiment analysis performance, as well as the architectures employing them. We explore Early Fusion, Late Fusion, Tensor Fusion, and their variants, alongside approaches like Multimodal Factorization Model (MFM), Multimodal Cyclic Translation Network (MCTN), and Multimodal Transformer (MulT). Each approach presents unique advantages and challenges, offering a range of tools to address the intricacies of sentiment analysis in multimodal data. Through a thorough review and analysis of these methodologies, this project aims to illuminate the current state-of-the-art in MSA. [Checkout the code here…]

Continual Learning for Autonomous Vehicles

This project focuses on predicting steering angles by leveraging image data captured in real-time. Through rigorous experimentation, the framework explores several continual learning strategies including Elastic Weight Consolidation, Experience Replay, and Temporal Consistency Regularization. By continuously learning from new experiences while retaining past knowledge, this framework aims to enable autonomous vehicles to navigate diverse environments with improved accuracy and robustness, thus advancing the capabilities of next-generation transportation systems. [Checkout the code here…]

[projects-header-2]

Vector Processor Simulator

The Python-based Vector Processor (VMIPS) functional and performance simulator offers a comprehensive platform for simulating the functionality and evaluation of the performace of a vector processor. The simulator is rigorously tested on assembly code for machine learning calculations - dot product, matrix multiplications and convolution. [Checkout the code here…]

ResNets

This project delves into the exploration of ResNet models trained for image classification on the CIFAR-10 dataset. Through meticulous fine-tuning of hyperparameters, experimentation with various ResNet depths, and implementation of strategies such as dropout layers and learning rate schedulers, the project achieved significant test accuracy of 95.37%. [Checkout the code here…]

[projects-header-3]

Semantic Image Segmentation for Autonomous Vehicles

This project focuses on semantic image segmentation tailored for autonomous vehicles, with a primary objective of detecting and segmenting various elements crucial for safe navigation, including drivable surfaces, vehicles, curbs, traffic signs, and sky. Leveraging a U-Net model, the project implements advanced deep learning techniques to accurately classify and segment objects within captured images. Specifically trained on the CARLA dataset, the U-Net model achieves an impressive accuracy rate of 96%, signifying its robustness and effectiveness in accurately identifying and delineating essential elements for autonomous vehicle perception. [Checkout the code here…]

Predictive Analysis of Diabetes

This project, conducted as part of the NYU Data Science bootcamp, focuses on predictive analysis within the PIMA Indian population, specifically aiming to predict the occurrence of diabetes. Employing a K-Nearest Neighbors (KNN) model as a foundational approach, the project explores avenues for improved modeling and feature engineering to better predict diabetes occurrences within this demographic. [Checkout the code here…]

Rugved Mhatre