The main goal of sentence simplification is rewriting text such that it becomes easier to read, while still retaining its original information and meaning. Current deep learning models are black-box models which focus on simplifying text while maintaining fluency. However, they do this at the expense of meaning preservation. Also, they are black-box models which means we do not know clearly what properties of the text need to be modified to do better simplifications. We are looking at models which can be controlled, use as less training data as possible and give insight into the simplification process.
Dhruv
Recently text generation models have been used to generate poetry, movie scripts, etc. We look at the task of generating lyrics for songs of different genres in a controllable manner. We use mel-spectrograms to represent songs and use it to condition the generation of lyrics.
Gaurav, Dhruv
The increasing amount of personal data is raising serious issues in the context of privacy, security, and data ownership. Entities whose data are being collected can benefit from mechanisms to manage the parties that can access their data and to audit who has accessed their data. Consent management systems address these issues. Consentio is a scalable consent management system based on the Hyperledger Fabric permissioned blockchain. The data management challenge we address is to ensure high throughput and low latency of endorsing data access requests and granting or revoking consent. Experimental results show that our system can handle as many as 6,000 access requests/s, allowing it to scale to very large deployments.
Rishav, Dhruv
Using Natural Language Processing (NLP) to measure toxicity, aggressiveness and incivility in online social networks. Accepted as a workshop paper (WASSA) at NAACL-HLT, 2019.
Dhruv Kumar
Used deep learning methods including multiple attention models for text classification implemented in Pytorch. The models are inspired not just by the state-of-the-art in classification but also by different tasks in NLP such as text entailment, Question Answering, etc.
Dhruv Kumar
Implemented the Bi-Directional Attention Flow(BiDAF) model from scratch in Tensorflow over the SQuAD1.1 dataset as part of the course project for the Deep Learning class. This was done as a course project for the course STAT-946 on Deep Learning.
Amir Pasha, Dhruv, Gaurav, Kashif
Trained a ResNet50 and VGG16 models using Transfer Learning on Cifar100 to obtain an accuracy of 78.53%. This was done for the data challenge which was conducted as part of the course STAT-946 on Deep Learning.
Dhruv Kumar
Applied my knowledge of Q-learning to learn the Cartpole environment of OpenAI’s Gym library.
Dhruv Kumar
Built a tic-tac-toe learning bot using Reinforcement Learning. This acted as a starting point for me into Reinforcement Learning.
Dhruv Kumar
Implemented a deep Recurrent Neural Network from scratch to learn about sequence networks, doing a character level generation of text. Ran the model on my Bachelor’s Thesis to learn my writing style. The model was quickly able to learn the grammar rules and important technical terminologies from my thesis.
Dhruv Kumar
It predicts a google stock opening price given the data for the last 20 days. Trained the model on multiple variables in the data from the last 5 years using a LSTM. Will improve this by adding sentiment knowledge to it, by using headlines of the stock for each day. It should not be much of an extension, since after getting the sentiment score from the headline about the stock, we can basically just add it as a new column to our matrix and our model learns it as a new feature. Although I haven’t fully optimized the hyper-parameters for now.
Dhruv Kumar
Built a face recognition web application on Flask and openCV for pmOne Analytics GmbH. The project’s goal was to identify faces from a video, given a person’s image. Since, we partnered Microsoft in this project, we used their Project Oxford API to train the model. The project was done as part of a tender bid towards a project by the German Government, aimed at identifying the people behind the Sexual Assault incident in Cologne on the new year’s eve of 2015/2016. Also used celery with RabbitMQ to make the application asynchronous.
Artus, Boris, Dhruv
This was my Bachelor Thesis, where I continued my previous project on Recommender Systems. I looked at ways to incorporate knowledge from metadata to improve the performance of Factorization Machines. Initially using the MovieLens 1M, I extracted the genre information from the summaries of movies and cast information by parsing the IMDB. Then, I designed a personalized model to utilize this information. We were able to more than double the performance achieved with tradition matrix factorization approaches like Collective Matrix Factorization, which I used as a baseline. We have been able to validate the approach with bigger datasets and will soon publish it.
Artus, Dhruv
This project was aimed at developing a recommender system, to incorporate semantic data as a side information for getting more accurate results. Proposed a modified Joint Matrix Factorization approach to incorporate information related to items and implicit user ‘s item preference to enhance the accuracy of the system. The model is based on time variant Recommender System, thus also utilises the time-related information of items. Experimental results show that this method gets approximately 0.60% more accurate recommendation results with faster converging speed than other existing approaches based on matrix factorization.
OP Vyas, Nidhi, Dhruv, Ronish, Shubham
Built this as part of my summer internship at Citi after my 3rd year. It was majorly a web application used for data visualization and a CRUD. Used MEAN stack and worked on d3.js for visualization. Build customization d3 modules wrapped in Angular directives. Was given a full time job offer after my internship.
Dhruv
It is based on Natural Language Processing (NLP). The goal of the project, is to develop a prototype that can feel the pulse of the E-Commerce website users with regard to the reviews they provide on the products they buy. What we generally find in the reviews section of these websites is rather deceptive. The star rating of the user is not always concordant with what he/she has written. Also there is no feature wise review available for products. We aim to bridge this gap with our project.
R. Sanyal, Dhruv, Shubham, Ronish
The application aims to provide all the information regarding Effervescence MMXIV the Cultural Fest of our College to the participants as well as to anyone who attends the fest. The application has day wise as well as genre wise details of the events. One can set reminders for all any events one is interested in as well contact the members of the management team through a single click via call, message and email facility. It also has Google Maps linked to it, which helps one find exactly the location where any particular event is happening. One could also mark his favorite events for a better navigation through the app.
Dhruv, Anmol, Ronish, Shubham
This project aims to develop an anonymous faculty feedback form through which all the students would login using their college login ID and password and give their opinion about a particular teacher teaching a particular subject. The results for each teacher would be compiled and displayed when the teacher logs in using a bar graph. The teachers will also have the privilege to compare their results over the years.
Dhruv, Ronish, Shubham
Feed of never-ending featured articles from Wikipedia. Ideal for killing time. Developed it in 2015. Inspired by Wikifeedia (search Github for the project). The difference though is that Wikifeedia gives random featured articles starting from the same alphabet(which is random too) whereas this gives random articles with no constraints whatsoever. Some of the code is borrowed from it. I had to change the underlying API and the way it was being utilized. Will soon update more details about the implementation.
Dhruv
A News Reading Android application built in 2013(before the new apps were cool). The aim was to build a better foundation in Android Development. Hits a Zee News API, crawls and fetches data in XML and displays the news sorted by time. Also has feature to search news through specific tags.
Dhruv