I am a PhD student in the Computer Science department at Virginia Tech's College of Engineering, working with Prof. Naren Ramakrishnan.
My research falls under the umbrella of Natural Language Processing, Deep Learning and Data Mining. I'm interested in developing AI systems for natural language understanding and generation. I also love developing Django webapps.
My research is graciously supported by DARPA and NSF.
Event Detection Using Hierarchical Multi-Aspect Attention
Event encoding can be viewed as a hierarchical task where the coarser level task is event detection and the fine-grained task is one of event encoding. In this work we present a novel attention mechanism and show its effectivenes on the event detection task when plugged into hierarchical attention models.
Sneha Mehta, M. Raihanul, Huzefa Rangwala, Naren Ramakrishnan
The Web Conference 2019
(Acceptance Rate: 20%)
PhotoSleuth: Combining Human Expertise and Face Recognition to Identify Historical Portraits
We created Photo Sleuth, a web-based platform that combines crowdsourced human expertise
and automated face recognition to support Civil War portrait identification.
Vikram Mohanty, David Thames, Sneha Mehta, Kurt Luther
An Exploratory Study of Human Performance in Image Geolocation
We perform an exploratory study of image geolocation tasks performed by novice and expert humans on a diverse image dataset we developed.
Our findings include a model of sensemaking strategies, a taxonomy of image clues, and key challenges and design ideas for image sensemaking and crowdsourcing.
We study the performance of a siamese network based deep neural architecture on the task of binary(Yes/No) visual question answering. Comparing the performance of a siamese network based VQA model to a non-siamese VQA model we find that having a pairwise loss helps perform better than a loss from a non-siamese VQA network.
We develop the topic analysis component of a robust Information Retrieval system for search and retrieval of large-scale tweets and webpages built on top of Solr, a general purpose open-source search engine. Our contribution enables semantic search and retrieval of tweets and webpages based on topics.
We employ a random forest approach for bird species identification. First we train 25 SVMs to predict 25 bird features for our dataset images. Then we train a random-forest to identify the specific bird species. We used the CUB-200-2011 bird dataset for this task.
Seam-carving is a dynamic programming based algorithm for content-aware image resizing developed by Ariel Shamir. In the above example, the algorithm reduces the width of the image by removing vertical seams of pixels between the girl and the cliff edges. > Code
Creating image panorama by image warping based on homographies. In the first example (left), the image on the bottom is the panorama created by warping (by computing a homography matrix) and stitching together the three images above.
In the second example(right), a photo of mine (a really embarrasing one now I realize) was warped and merged onto a billboard in a Times Square image. Find code and more examples > Code
Presenting at GroupSight (Workshop on Human Computation for Image and Video Analysis), HCOMP 2016 at Austin, Texas.