A Computer Engineering undergraduate currently in my final year, it excites me to learn about the numerous opportunities the field of computer science offers. I have explored them thoroughly and plan to pursue the interdisciplinary area between Computer Vision and Natural Language Processing. A Master’s Degree at the University of York will be an extension of my undergraduate work in the Visual and Language Space and will enable me to delve deeper to carry forward my research into Visual Scene Understanding and Navigation.
During my undergraduate study I did a summer student program at the Indian Institute of Technology, Kanpur where I conducted research on Unsupervised Visual Grounding. At this internship, I developed the baseline and designed a novel approach to ground image captions. I read many relevant papers and in understanding their results, I realised that the language model in the encoder-decoder framework was lacking in terms of performance when compared with the vision model. Using this conclusion, I developed the framework for context-specific concept batches. This approach ensures better semantic and contextual clarity within the model. I am now in the process of training my model and it shows promising results. The concept batches are more specific. This means that more of the dataset is trained over, rather than a few recurring common concepts. Completeness over datasets helps ensure that even entities that are not as well as more common objects, can be trained over and hence semantically learned.
In the course of this internship, I faced a problem that many Grounding and consecutively Vision and Language systems encounter – Image Specificity or a lack thereof. Initially defined in CVPR 2015, the Image Specificity problem causes a drop in performance of vision and language models in the case of unspecific images. With real world datasets, quality or rather specificity of an image cannot be guaranteed. In order to better understand and combat this problem, I am conducting my undergraduate thesis on this topic. I use generation of diverse image captions as a proxy task to determine Image Specificity. While this topic is simple in theory, it has far reaching consequences and applications. Images could be assigned weights depending on how specific or unspecific they are. This would help Vision and Language models learn better from highly specific images and be less affected by noise during training.
While in my third year, it was my curiosity that drove me to become involved with research under one of my professors. We were a team of four and we were lucky to receive a research grant from the University of Mumbai for our proposal,namely ‘Optimisation of Rainwater Harvesting Systems, using GIS and Modelling in Arid Regions of Maharashtra’. We designed and developed a framework for finding the most optimal locations for reservoirs based on remote sensing. Using a Digital Elevation Model, the stream networks and pour points and water retention of different soil types in the area, we modelled the flow and calculated the amount of water that could be collected. Our framework showed some promising results in arid regions of Maharashtra, India. We identified three possible sites to collect naturally flowing stream water, in the extremely dry planes of the region.
To gain a ‘real world’ perspective and to accumulate new skills in the computer field, in the summer of 2017, I interned at i3Systems to classify images of identification documents. I learned various Image processing, feature extraction and analysis techniques and was introduced to the world of Deep Learning. Finally, my model showed a 20% increase in accuracy over the existing model. I am currently working as a Data Science Intern at Hashtag Loyalty and have completed three projects till date. The first project involved observing customer transactions at a business and analysing it to find patterns. I performed time series forecasting to predict footfall and reservations on a given day. The second project I worked on was generating customer segmentation and behaviour prediction, for which I used Kohonen’s Self Organising Maps. This system is now in the beginning stages of deployment and preliminary results will be obtained in a couple of months. Lastly, I am working on an Email subject line optimisation model, which takes a caption as input, and suggests similar meaning words with higher hits, that can be used in place of the current ones. It is now available as a service through an API endpoint using Flask.
Undergrad study is, for me, a time for exploration and my curiosity and self-motivation leads me to try many different fields of computer science. This augmented my education in unquantifiable measure. I participated in a multitude of hackathons which has helped me not only to conceptualise ideas but also to engage my creative talents in a problem solving manner. Notably, my team was in the Top 6 in the nation-wide Smart India Hackathon of the ISRO chapter for our ‘Personal Identification System using Keystroke Dynamics’. Through these Hackathons I have also developed my leadership and communication skills. In CodeShastra, a state-wide hackathon, I helped develop a Crop Recommendation System based on Remote Sensing. We used high resolution vector and raster data from satellites along with soil nutrient analyses and ideal conditions for crops. By intersecting these sources we had to find crops that are best suited for a given location. Here I coordinated between all the members of my team and integrated all the components of the pre-processing framework and am proud to say that this web-app earned us second place.
My desire to pursue a graduate course in Computer Engineering at your esteemed University will be my dream come true. The flexibility of the curriculum, the variety of areas of concentration and the added incentive of being tutored by your renowned faculty will help shape and augment my skills to better suit the needs and demands of the industry I aspire to excel in.The subjects offered by your Graduate program are in line with my interests and thus best suited to my academic endeavours. I am positive that my association with your university and the erudite international environment your campus offers will definitely result in mutual discovery to keep up with the changing needs of the industry the world over.
