1. Mortality prediction on MIMIC-III using a latent topic model

    Introduced the idea of treating different types of clinical notes from EHR (physician notes, nurses notes, etc) as different modalities in learning latent topics from MIMIC-III data and using the learned topics for making clinically relevant predictions.

  2. Group Equivariant Deep Reinforcement-Learning

    Used an equivariant CNN for approximating Q-values using Deep Q-Network algorithm for symmetric RL environments like Snake and Pacman. Currently under review in an ML conferencce.

  3. Method Summarization from Code

    This work was carried out while I was an intern at IBM Research AI Lab, Bangalore from May’18 to Aug’18. I worked with Rahul AR on using sequence-to-sequence models models for solving problems prevalent in Software Engineering – the automatic generation of method names from method bodies.

  4. Automatic ROI detection from Histopathological slides

    This work uses an RCNN architecture to automatically detect suspicious (of being carcinogenic) regions in images of histopathological slides. It works for images at various zoom levels. The main challenges faced during this project was dealing with the large sizes of images (nearly 22000x37000) which we circumvented by dividing into multiple grids and stitching them back together for visualizing results. This project was carried out under the guidance of Prof. T K Srikanth. The boxes denote suspicious regions.

  5. Scaling up Simhash

    This is a dimensionality reduction algorithm that maintains an estimate of the cosine similarity between the original high-dimensional data points. This work was done in collaboration with Dr. R Pratap and is currently under review at a top tier AI conference.

  6. Detection of star clusters using Pattern Analysis

    This work was done with Prof. Sarita Vig while I was an intern at the Indian Institute of Space Sciene and Technology from May’17 to Aug’17. Worked on a direct application of pattern analysis techniques for a highly relevant problem statement in astrophysics. Used the K-Nearest Neighbor algorithm for density estimation and detection of star clusters. The images below show contour plots of a locus of simulated points with equal probability density. Highly dense regions correspond to star clusters.

© Made with love, code and coffee

Powered by Hydejack v8.1.1