External Links

Visual Question Answering from Image Sets Permalink

Research Project, Carnegie Mellon University, LTI, 2020

  • Adapted a Transformer based VQA model (LXMERT) for the task of Image-Set VQA and developed an Adversarial Regularization method to reduce dependance on language biases & improve performance on out-of-domain data
  • Introduced a new pre-training objective which utilized object bounding boxes extracted from an RCNN to improve model performance on object description questions by 4%

Unsupervised Scene change identification Permalink

Course Project, Carnegie Mellon University, LTI, 2020

  • Introduced a generative approach that uses a Beta-VAE to identify scenes changes in videos by measuring KL divergence between images and Eliminated manual effort in annotating data for downstream tasks (Super-Slomo)

YouTube Trending Analytics Website Permalink

Course Project, Carnegie Mellon University, LTI, 2019

  • Investigated factors that govern the YouTube trending page and visualized the presence of user and platform bias
  • Hypothesized the reasons for existence of bias and demonstrated their variability across different countries
  • Constructed a Machine Learning pipeline with XGBoost classifier to predict the likelihood of a video to trend
  • Video Demo: https://www.youtube.com/watch?v=UlanYkq_VII

Stance Detection to Identify Fake News Permalink

Course Project, PES University, Computer Science Dept., 2019

  • Developed a Bi-LSTM model which utilized Contextualized word Embeddings (ELMo), to detect discrepancies between claim present in a news article and other authoritative news sources to identify potential fake news
  • Demonstrated the superiority of the approach over existing online APIs for stance detection
  • Multidimensional Analysis of fake news including identifying fake accounts on social media, fake websties detection and analysis of spliced images to spread fake news.

Unconstrained Face Recognition Permalink

Course Project, PES University, Computer Science Dept., 2018

  • Introduced a novel pipeline architecture for face recognition which used the highly optimized CloudForest algorithm to achieve 10-15x training time improvement over other ensemble classifiers such as Random Forest