(2021-22) CTRL-sim: Controllable medical text simplification
Low Health Literacy has several adverse effects including poor patient self-care, lack of timely communication of health issues, and even increased risk of hospitalization and mortality. CTRL-sim is a project that addresses this challenge by focusing on the accessibility of online medical content. I worked with Rosni Vasu, Michihiro Yasunaga and, Qian Yang to develop a novel medical text simplification dataset and baseline models. Our models leverage the latest advances in large pretrained language models and controllable text generation allowing users to selectively simplify the contents of a medical text.
(2021-22) Healthy Recommender System that understands human values affecting item choice
Reward-based approaches for recommending items have been found to be more generalizable than supervised learning, as we can directly optimize for long-term engagement and user satisfaction. Further, modeling intentions or values behind human choices can be advantageous when the purpose of the recommender system is to manipulate those choices. In this work, I we first developed a multi-label value classification model that can map human food choices to explicit human values based on review texts. Next, I am developing a a hybrid recommender system that uses the value and tag-based item representations to jointly optimize for long-term engagement and aggregate healthiness of the consumer food choice. This work is done in collaboration with Jian Vora and Erdem Biyik. We are investigating the interaction between active manipulation of weights in the value space and display of explicit aggregate healthiness information.
(2019) SynAV: Synthesizing adversarial traffic for training self-driving cars
There is a real paucity of edge-case scenarios in real-world driving data. This makes it hard to train autonomous vehicles for such situations are really encountered. I explored a traffic scenario data augmentation technique that first learns realistic driving behavior from real-world data and then tweaks this learned behavior gradually to synthesize anomalous and potentially dangerous traffic situations.
(2016-2019) Learning human preferences from rich guidance
Drawing from preference learning, hierarchical reward learning, rich queries and Bayesian models of human guidance we are designing interactive AI agents that can learn faster by asking the right questions. They use rich models of human guidance to ask queries autonomously.
(2015 - 2016) Do you want your car to drive like you?
In this project, we investigated value-misalignment between demonstrations and preferences in the domain of self-driving cars. We found that people prefer self-driving cars that drive very differently from themselves.
(2014 - 2015) Privacy preserving Bluetooth Localization
Bluetooth is increasingly used for indoor localization due to lower range and hence noise than WiFi signals. Such location information could be symbolic in terms of proximity relative to a device or a person with known location or a service seeking entity. It could also be the absolute coordinate position or relative coordinate position on the map of a building. I am implementing Bluetooth localization for efficient human arrival time estimation in a Human-Robot rendezvous problem. People do not like to be tracked. Our short survey at the Gates Hillman Center in Carnegie Mellon University says the same. So is it possible to accomplish the tracking without securing any personally identifiable information from the phone or potentially identifiable information like MAC address? I am developing a method for peer-to-peer localization where a network of phones can identify you by your device name. What if people are not happy about disclosing or changing their device names for localization purpose and the process latency is too high for reliable proximity detection? As a second alternative I am trying to exploit several features extracted from the inquiry scan packets of Bluetooth device recorded by a sniffer.
(2013 - 2014) Dictionary learning and sparse coding for occupancy estimation
Most modern temperature control systems are designed to consider maximum occupancy count afforded by each room. This works well for offices which have less than five occupants, but for large shared spaces like conference room or classrooms, without a designated owner, the actual occupancy can vary significantly. Rooms are scheduled for meetings and classes assuming attendance of all registered participants resulting in inefficient space scheduling and usage. Setting a wrong occupancy based temperature has a negative effect on either energy usage or occupant comfort. Recent research has shown that 42% of the annual energy of building can be saved with knowledge of fine-grained occupancy. In this project, conducted in collaboration with Professor Anind Dey from Carnegie Mellon University and Dr. Kamalika Das from NASA Ames Research Center, I developed a sparse non-negative matrix factorization (SNMF) based prediction algorithm for occupant count from single sensor carbon dioxide readings.
(2011 - 2013) Sensor-based predictive modeling for intelligent lighting in grid-integrated buildings
As part of Internet-of-Things and Smart Lighting project, we developed a low cost wireless sensor enabled intelligent office lighting system for future grid-integrated buildings. Closed-looped intelligent lighting control relies on dense sensing. Dense sensing can be avoided by optimal sensor deployment that takes advantage of the spatial correlation of light distribution. The spatial correlations are encoded in models which are piecewise linear predictions of indoor light discretized by clustering for sky conditions and sun positions. We call these models virtual sensors. For more information refer to our publication in IEEE Sensors Journal.
Some of my shorter projects geared towards human activity detection and modeling include Kinect-based posture correction with voice-feedback and motion and proxemics detection using thermal array sensor. In the latter project I used correlation between background subtracted 64 infra-red sensor pixels to infer motion near the doorway of a room.
Many of the codes developed for the above projects can be found in my git repository.