Data & Applied Scientist 2

Contributing to a new product, PRISM Feed. It will be the default experience on Bing Image Search

• Owner of module for diversifying and re-ranking recommended documents for a new product called PRISM
• Developed scalable, modular code from the scratch. Shipped to en-* market
• Working on Query Classifier for identifying relevant user signals for personalization

Tools: C#, PowerBI, Internal Compute and Development Tools

Associate Researcher

Developing machine learning models trained over images and videos and then deploying them on resource-constraint devices.

• Prototyped and deployed face detection and object detection models on Huawei NPUs
• Mentored four co-op intern students working on different ML inference projects
• Contributor to drowsiness detection system for cars, featured as a major product by Huawei CBG CEO
• Consulted on multiple projects for architectural design and embedded systems related optimization
• Tools: C++, Python, Tensorflow, Android JNI

Applied Research Intern

The internship aimed at analyzing the microscopy images of multi-cell cultures and extract intensity, shape, and texture-based features using machine learning techniques to quantify the cell-cell interactions.

• Developed and deployed a robust and scalable segmentation and feature extraction pipeline for microscopic images.
• Classified and analysed different cells identified through segmentation using supervised and semi-supervised techniques.
• Quantified and validated the contact-dependent effects between lung cancer cells and fibroblasts.

Tools: Python, AWS services, CellProfiler, ImageJ (Fiji)
Links: Poster

Research Engineer

Sensor-based In-house Tracking System

• Devised an offline tracking method for residents based on sensors in the house
• Published in a Springer conference
• Developed a robust generic Scala script to preprocess raw sensor-based data. Preprocessing time reduced from two weeks to three days

Tools: Apache Spark, Scala, Python, NumPy
Award: SPOT Award for the preprocessing script

Voice Activity Detection

• Literature review on the current state-of-the-art human presence detection methods
• Implementation of the feature extraction to optimize in memory and time
• Used TensorFlow Lite to validate the output of the human presence model on a Tizen-based smart speaker

Tools: C++, Python, TensorFlow, TensorFlow Lite
Award: SPOT Award for the reduction of the inference time from 7 sec to 2.9 sec

k-space Imputation and MRI Reconstruction

Explored denoising autoencoder based U-Net and perceptual GAN to impute k-space to improve the process of MRI reconstruction. Proposed the DAE-UNet method and trained using fastMRI dataset.

• Idea is to impute k-space first and then reconstruct image.
• Observed that as the area of interest in a MRI image is much less as compared to the size of the entire image, it might be possible that a model learns to reconstruct the background and still get a low MSE

Team: Vaibhav Saxena, Pulkit Mathur, Duc Truong

• Links:
• Report

The Tempest

Theatrical presentation of the play, The Tempest, on a 180-degree curved screen leveraging the concept of proprioception and 3-D drawing for the character Ariel

• Inclusion of proprioception to trigger the events based on zones selected instead of explicit display of menu items
• Performed comparative analysis between pre-defined and user-defined zones to understand the user’s comfort level

Team: Andrejs Ru, Sahil Narula, Sneha Desai
Tools: Unity, C#

• Links:
• Report
• Slides
• Code

Text Readability Analysis using Language Models

• Designed a new fine-grained, computational measure of readability. Main conjecture was that the predictability of a text, as determined by standard language models, is a viable metric of its readability
• Evaluated our metric on various datasets with known readability scores, such as simple and standard Wikipedia articles, labeled datasets from prior work on automatic readability scoring systems, and even created a new dataset by parsing freely available school textbooks
• Obtained good results for all the three datasets, unlike earlier approaches

Tools: Python

• Links:
• Report
• Slides

Using Spatial Transformer Networks for Egocentric Images

• Implemented spatial transformer networks for object recognition from egocentric images
• Evaluated it on GTEA (Georgia Tech Egocentric Activities) dataset. The model showed better results than a traditional CNN model

Tools: Python, TensorFlow

• Links:
• Report
• Slides

Assistive Device for Hand Amputees

ADHA is a device that is designed to support people who have had both their arms amputated above or below the elbow. It provides two major types of support:

• Enable the user to connect to the digital world by forming a convenient interface to a computer's keyboard and mouse.
• Enable the user to transport heavy objects by having a transport robot that they can control.

Tools: Arduino, Python

• Links:
• Documentation
• Chassis Design
• Code