"You have to dream before your dreams can come true."
About Me
I am a first year graduate student pursuing Masters in Computer Science at University of Massachusetts. I completed my Bachelor's from Indian Institute of Technology, Kanpur with major in Mathematics and Scientific Computing and minor in Artificial Intelligence.
I have a keen interest in the field of Computer Vision in combination with Natural Language Processing. My recent projects involved Image and Video Captioning under Professor Vinay Namboodiri. I am also interested in the application of Gaussian Processes to regression based problems.
Apart from academics, I enjoy to pen down my thoughts and opinions- Blog. I am also a basketball player having represented my college twice. As the Head events of the Entrepreneurship Cell at IITK, I have organised events like eSummit and TEDxIITKanpur.
You can find my resume Here.
Past Affiliations
 |
 |
 |
 |
 |
Research Interests & Projects
The buzzword "Machine Learning", finding inspiration from humans who learn from experiences, succesfully lured me into exploring the field. Machine Learning tools combined with image processing algorithms encouraged me to delve into deeper neural networks. With a strong undergraduate background and current Master's knowledge, I aspire to fuel my research interests and put it into application making a significant contribution in this advancing world.
Interests
- Computer Vision
- Deep Learning
- Gaussian Processes
Projects
-
Replicated the torch implementation results of Dense Captioning model to obtain region-specific captions.Refined the captions by putting filter on the overlap area and caption similarity to reduce redundancy. Categorized captions through topic based clustering and fed them into anencoder-decoder model trained to generate sentences from phrases obtaining successful results in producing paragraphs for images in Visual Genome Dataset.[Report] [Slides]
-
Implemented the State-of-the-art model of Sequence to Sequence-Video to Text which exploits the temporal information of videos. Incorporated audio features to improve the confidence in activity prediction which refined captions. Also Combined the Deep Compositional Captioning model to use a language model trained on DBpedia reducing dependency on annotated dataset.[Report] [Slides]
-
Refined the data obtained from surveillance camera videos of the campus and carried out background subtraction for object detection which were then tracked using box overlap in consecutive frames.Employed grey-scale, HOG, hierarchial HOG and SIFT features for clustering as well as classification.Experimented with SVM, Random Forest, Adaboost, Convolutional neural nets for classification of vehiclesand pedestrians obtaining the highest accuracy of 95.37%.[Report] [Slides]
-
Studied Gaussian Processes which involves inferring a distribution rather than giving a point estimate. Assumed a Gaussian prior with a zero mean and squared exponential covariance over the predictingfunction and used Bayes Theorem to obtain a posterior which is also a Gaussian Distribution for forecasting. Tried various covariance functions to predict forest fires using GPML libraries in MATLAB.[Slides]
Internships
-
Analysed data containing app usage history to obtain user specific pattern in the app preferences.Tried k-means, agglomerative Mean-Shift and Markov models to group users with similar features. Trained SVM, decision trees & Random Forest for predicting ads with maximum click through ratesto provide user targeted advertisements.
-
Build an Analytical Engine to get insight into Supply-Demand Matching for real time debugging. Tomcat based web application in Java backed by Elastic Search for efficient querying, aggregating data. Real time feedback using Kafka which provided data for analysis from serving systems having necessary details.
-
Developed a model to predict weekly and quarterly sales of medicines using past 2 years data. De-trended & deseasonalized the data to obtain stationary sequences for applying time-series models. Used Exponential Smoothing & ARIMA[Auto Regressive Integrated Moving Average] methods to forecast the sales of retail stores in the neighborhood.
