Sheetal Shalini

Implemented Hierarchical and Coarse-to-fine attention using a sparsemax approximator. Proposed two novel attention mechanisms using Convolutional Neural Networks (CNNs) and Recurrent networks like LSTMs to store the context of previous attention maps.

Solved the tasks of Natural Language Inference (NLI) and Recognizing Question Entailment (RQE) using an ensemble of BERT, MT-DNN, and Infersent models. We then used the NLI and RQE modules as feature extractors to filter and re-rank answers in the medical domain using multi-task learning.
Our model achieved the highest Spearman's Rho and Mean Reciprocal Rank of 0.338 and 0.9622 respectively, on the ACL-BioNLP workshop MediQA Question Answering shared-task 2019.

Implemented and compared the performance of Bottom-up top-down attention mechanism and Memory Attention Composition (MAC) cells for the task of Visual Question Answering.

Developed a Movie Recommender System based on user-movie similarity, Pearson Correlation Coefficient (PCC) and Probabilistic Matrix Factorization (PMF). Achieved Root Mean Squared Error (RMSE) of 0.9025.

Developed Global, Query-based and Personalized Topic-Sensitive Page Rank combined with search-relevance scores to rank documents. Achieved Mean Average Precision (MAP) of 26%.

1. Implemented Canny Edge and Harris Corner Detection, Image Stitching with Perspective Transform and Object Detection with Naive Bayes, Decision Trees and SVM classifier.
2. Achieved an accuracy of 93.24%.

1. Performed Image Segmentation using U-Net and achieved a validation accuracy of 92.94%.
2. Performed Classification of Skin Cancer images using ResNeXt and achieved a validation accuracy of 79.25%.

1. This project aimed at enriching the existing crime ontology by incorporating images from news and social media, and deriving meaning out of them.
2. Crawled through the multitude of multilingual information available on the internet using python libraries like Goose and JusText and developed an ontology of entities and events connecting these entities.
3. Performed news summarization of online newspaper crime reports, tokenization and POS tagging using NLTK and Shallow Parser, Named… Show more

1. This project aimed at enriching the existing crime ontology by incorporating images from news and social media, and deriving meaning out of them.
2. Crawled through the multitude of multilingual information available on the internet using python libraries like Goose and JusText and developed an ontology of entities and events connecting these entities.
3. Performed news summarization of online newspaper crime reports, tokenization and POS tagging using NLTK and Shallow Parser, Named Entity Recognition (NER), Word Sense Disambiguation (WSD), relationship extraction and machine translation.
4. Performed image tagging, retrieval, correlation and information extraction from social networking sites like Facebook, Twitter and Flickr.
5. Obtained the Combined Ontology by taking the Newspaper ontology as the base and enhancing it with relevant data from the Social Media Ontology. Show less

1. Implemented Dijkstras algorithm using Binary, Binomial and Fibonacci heaps.
2. Compared their running times for large number of nodes.

1. Used graph mining, machine learning and OddBall algorithm to detect link-based static anomalies in weighted graphs of social networks.
2. OddBall uses the density, weights and eigen values of the egonet (1-step neighbourhood) of each node to calculate the outlierness score. It's based on the 3 power laws, EDPL (Egonet density power law), EWPL (Egonet weight power law) and ELWPL (Egonet Lambda weight power law).
3. Perfomed the algorithm on facebook dataset from SNAP (Stanford Network… Show more

1. Used graph mining, machine learning and OddBall algorithm to detect link-based static anomalies in weighted graphs of social networks.
2. OddBall uses the density, weights and eigen values of the egonet (1-step neighbourhood) of each node to calculate the outlierness score. It's based on the 3 power laws, EDPL (Egonet density power law), EWPL (Egonet weight power law) and ELWPL (Egonet Lambda weight power law).
3. Perfomed the algorithm on facebook dataset from SNAP (Stanford Network Analysis Project). Assigned edge weights to the unweighted facebook graph based on the similarity of features of the end nodes of each edge.

Technologies used were R and python. Show less

1. Implemented a trilateral filter to denoise brain MRI with Rician Noise.
2. Used image processing techniques like Rough Set Theory (RST) to obtain the Rough Entropy Threshold (RET), Rough Edge Map (REM) and Rough Class Labels (RCL) to produce the third component of a bilateral filter.

1. Developed a system to digitalize the procedures of attendance and professor feedback.
2. Worked on technologies like Video and Image Processing using OpenCV in python, Local Binary Pattern Histograms (LBPH) and Eigen Face recognizers, machine learning algorithms like feature extraction using SURF (Speeded up robust features), k-means clustering, Naive Bayes classification and k-fold cross validation.

1. Developed a bus reservation system using Django, MySQL and JS.
2. Integrated it with Google Maps API, to automatically select the nearest pickup and drop points of the passenger based on his source and destination addresses.

Designed the lexical, syntax, semantic and intermediate code generation phases of a C compiler using Lex and Yacc.

1. Developed a distributed and dynamic load balancing system using OpenMPI.
2. Implemented truthful payment mechanism to yield minimum overall expected response time as proposed in the paper 'Algorithmic Mechanism Design for Load Balancing in Distributed Systems'.

Developed a library management system using MySQL, PHP, HTML and LAMP framework. This system facilitates students and professors to view all the listed books and choose them beforehand as well as provides special access to admins.

Developed a django application to dynamically track the locations of the users and employ machine learning techniques like Naive Bayes classification, genetic algorithm and fuzzy c-means clustering to predict the amount of risk involved based on their age, gender and expertise of driving and notify them of the same.

An 8 week-long android workshop organized by Google, in which we developed several android applications like Anagrams using HashSets, Scarne's dice with asynchronous programming, word stack, Ghost game with binary search trees and tries, puzzle with heaps, priority queues and A* algorithm, word ladder with graphs, Black Hole using the Monte Carlo method as well as Continental Divide with Dynamic Programming.

Developed a chat application called BizzChat which involves a client-server architecture, and includes general Settings as well as Profile features like Backup and File drag and drop provisions.

1. Developed an android application for the visually impaired.
2. The application includes storage and retrieval of documents, emergency calls and messages, and location assistance using automatic speech generation and recognition.
3. Worked on technologies like Android Studio, Microsoft Azure, speech to text conversion, Google Maps integration and SMS API.

Developed a parking lot automation software called 'ParkWhizz'. It facilitates online booking of parking slots in commercial places.