Early detection of Alzhiemer's

Decreasing language complexity can be indicative of Alzheimer's. I worked with my colleague under the guidance of Dr. Karl Stratos to develop an NLP based tool that is potentially useful for the early detection of Alzheimer's based on patients' speech.

  • Along with my colleagues, I worked on using transcripts generated by the speech-recognition team to quantify the complexity of speech.
  • I worked on preprocessing the speech transcripts and generating punctuated and capitalized text using BERT based transformer model.
  • I also worked on using LAL-Parser to perform constituency parsing of generated text that is useful to calculate syntactic complexity measures such as Yngve Score and Frazier's score
  • Through this project I gained experience with: Transformer architecture, text preprocessing, BERT, constituency parsing, and cognitive impairments caused by Alzheimer's.

Abstract Visual Learning using Vision Transformer

This project investigates whether Vision Transformer-based Deep networks perform better on Abstract Visual Learning tasks than CNN-based approaches.

  • I developed Vision Transformer-based Wide Relational Network to solve RAVEN's progressive matrices(RPM).
  • I worked on extracting image features from Vision Transformer and used them for the relational module that predicted correct answers to RPM questions.
  • I later trained and evaluated this neural network on the I-RAVEN dataset and compared its results with other approaches.

Recall prediction in Free Recall task using Transformer

I worked on predicting the next word in a free recall task. In a free recall task, participants see a list of words, and later they are asked to recall these words as they remember them. Given a presented list and partial recall, we aim to predict the next recalled word.

  • I developed a vanilla transformer based supervised model trained on the PEERS dataset that predicts the next recalled word.
  • These predictions were overall 40% accurate and captured most of the general recall trends observed in the data.

Gamified Career Test

Question-answer based career tests are unimaginative and inaccurate. To develop a better career test, using paradigms of gamification and AI, we developed a storyline game that determines the personality traits of users and predicts their career interests.

  • I worked on integrating the OCEAN model to find personality traits in a fantasy storyline that my project members and I wrote.
  • I also worked on implementing this storyline as a 2D game in Godot game e engine using gdscript.
  • We developed a decision tree based classifier model that mapped personality traits to potential careers for users.
  • In the end, we presented our work at an IEEE conference.

Satellite Image classification

For a hackathon coding challenge by ISRO, my colleagues and I worked on developing a classifier model to classify forest areas and farming areas from a LISS-IV satellite image.

  • I worked on using unsupervised clustering methods such as k-means and gaussian mixture models to cluster parts of images based on their similarity.

Organized Gathering for Social Distancing

As an early response to Covid-19, social distancing at public places such as grocery stores was essential. To effectively manage the flow of the crowd in such highly active sites, we developed an online website that suggested users time to visit these sites to maintain social distancing. We presented our idea at a hackathon.