During my research internship at Google, I worked on improving end-to-end neural network models that enable a robot to navigate following natural language route instructions in a simulated home environment. Given a route intsruction in a home environment, the model learned a policy to map egocentric visual observations to low level navigation actions via reinforcement learning. Models were trained and tested on an internal dataset.

During my research internship at Facebook Applied Machine Learning, I worked on evaluating supervised dialog models for a task-oriented dialog system. The system had to identify which of the possible tasks a user was interested in completing, and engage in a dialog to fill slots required to perform the task. The focus of my internship was to compare the performance of various models that used dialog text, as well as extracted dialog state features to predict the best system response dialog act at each stage.

The broad goal of this internship was to personalize ad messages for different target audiences. Given a skeleton message, we used Twitter data of users to identify segment-specific word usage patterns that could be exploited to enrich a given skeleton ad message. One of the most interesting ideas we used in this work was identifying the head of a noun phrase to choose appropriate adjectives for it. Our work won the best paper award at CiCLing 2015 and was also filed as a patent whose details are as follows -
Rishiraj Saha Roy, Guna Prasaad Jeganathan, Aishwarya Padmakumar, and Ponnurangam Kumaraguru, “Automated Linguistic Personalization of Ad Messages for Targeted Campaigns”, Patent Application Number P4592-US, United States Patent Office, filed on 10 December 2014 (Patent pending).

Google Trader was a service that allowed people to buy and sell products and services, later replaced by Google+ Business. Our work during the internship involved adding a feature to Trader that allowed people to save their searches and get updated results on them. This internship was my first experience of working with big data and large-scale software and where I learnt to use map-reduce and version control systems, and was exposed to dependency injections and remote procedure calls.

SAP was where I learnt about what are and how to create web services, specifically REST web services. We were contributing to the revamping of the website of a social endeavour of SAP called Charitra. Charitra is an application that connects individuals and organizations that need something with individuals and organizations that are willing to give. During this time, I also learnt about HANA-DB - an in-memory, column store database developed by SAP.