Research

This page describes my projects briefly. Please follow the hyperlink for an in detail description. The links to code or the paper are available wherever possible.

  • Visual Similarity Measure for Recommendations

    UMass Amherst (Oct 2017 - Dec 2017)
    Motivation for this project is in observing that visual preference of humans affect their decisions. In this work, we implement various models to incorporate this visual information adn show that the results are comparable to the state of the art.


  • Dependency Parsing of Biomedical Articles

    IESL, UMass Amherst (June 2017 - August 2017)
    This project was done during the summer of 2017 as part of IESL, UMass Amherst. The aim of the project is focused at developing an efficient dependency parser to parse Biomedical articles from PubMed articles. The challenge was to build this parser without any labeled data. We use domain adaptation and techniques like self training to deal with this problem.


  • Open Domain Dialogue Generation

    UMass Amherst (Mar 2017 - May 2017)
    Development of Conversational agents is an important area to advance the area of General Intelligence. In this project we explored the state of the art Seq2Seq Models. Seq2Seq architecture has proven to produce great results in traanslation. We argue that this model lacks memory which isn't needed for translation but is crucial in dialogue agents. We prove this by showing that the training loss converges faster and the results are better.


  • Generating harder CAPTCHAs using GAN

    UMass Amherst (Nov 2016 - Dec 2016)
    With the advances in deep learning, breaking CAPTCHAs has become trivial. Instead of using this advances to break the CAPTCHAs, we wanted to use it to improve the quality and make them stronger. The approach we used was to create a generative model using GANs. This was particularly chosen because we could provide feedback and improve the generative model accordingly.


  • Live 3D Scene Modeling using Smartphones

    IIT Mandi (Aug 2015 - May 2016)
    The aim of this project is to create an app that would take video of an object or a scene as input and produce a corresponding 3D model. There are many approaches to scene reconstruction but the challenge here was to develop such algorithm, which would be lightweight and very time efficient. We observe that contrary to other approaches, we have continuous video feed and hence capitalize on this fact to construct our implementation.


More information will be updated soon.