Projects

SightSync is a project that implements a Shazam-like functionality for video and audio clips. The system takes a database of videos, extracts frames and audio features, and stores them in Milvus vector database collections. It then allows users to search for a specific video or audio clip, finding the closest matches in both the visual and audio databases.

This project was done towards the coursework CSCI576: Multimedia Systems Design taught by Prof. Parag Havaldar in Fall 2023, at USC

In this project, we implement the Forward-Forward algorithm porposed by Prof. Geoffrey Hinton and compare it with the traditional backpropagation algorithm. We also study the architecturual differences in FF for various datasets and propose popular architectures such as CNN, Self-Attention with Forward-Forward. Finally, we log system metrics for both the algorithms and analyze them.

This project was done towards the coursework - CSCI566: Deep learning and its applications, taught by Prof. Jesse Thomason, Spring 2023, at USC

The proliferation of fake news and clickbait articles on social media platforms has become a major concern for individuals, organizations, and society as a whole. To combat this issue, there is a need for effective models that can accurately detect these types of content. In this paper, we present an in-depth analysis of existing models for detecting fake news and clickbait, and study how the two problems are correlated by studying the predictions from these models. Our models have been trained and tested on a variety of publicly available datasets our conclusions show that models that detect fake news are to a certain extent reliable to make predictions on clickbait data as well.

This project was done towards the coursework CSCI544: Applied Natural Language Processing, taught by Prof. Mohammad Rostami, and Prof. Xuezhe Ma in Spring 2023, at USC.

Presented the paper by Nikhil Mehta et. al (ACL 2022), where they propose inference operators to continually improve the social media context (represented as Relational Graph Convolutional Neural Network), to aid fake news detection.

This presentation was done towards the coursework CSCI544: Applied Natural Language Processing, taught by Prof. Mohammad Rostami, and Prof. Xuezhe Ma in Spring 2023, at USC.

Presented the paper by Levy & Goldberg (2014), where my role as a teacher I ensured my students understood the mathematical grounding of why neural embeddings can be thought of as implicit matrix factorization, and explained key insights from the paper. I finally concluded with an intriguing application of neural word embedding algorithm (skip-gram model) for recommendation systems.

This presentation was done towards the coursework - CSCI566: Deep learning and its applications, taught by Prof. Jesse Thomasson, Spring 2023, at USC

Time series forecasting is the practice of using historical data to predict future values. It can be challenging, due to the complexity of time-series data and the difficulty of accurately modeling patterns and trends. In this project, we tackled the ‘Store Sales- Time Series Forecasting’ challenge on Kaggle, which involved forecasting sales at a grocery retailer in Ecuador. We trained multiple models, starting with linear regression as our baseline. After careful analysis, we improved upon the baseline and achieved an error score of 0.39525, a 13% improvement over our linear regression model.

This project was done towards the coursework - CSCI567: Machine Learning, taught by Prof. Vatsal Sharan, Fall 2022, at USC

Implemented MNTD Networks proposed by Xu et al. (2020) for the trojan detection challenge hosted by USC’s AI safety group, and managed to beat the baseline performance using ensembling technique. Consequently won the IInd place in the competition