I am a Caltech computer scientist with a focus on applied machine learning.
My research experience is in artificial intelligence, computer vision, and robot learning, specifically deep reinforcement learning and learning from visual demonstrations. I work with Decisions, Optimization, and Learning @ Caltech (DOLCIT) and the Stanford Vision and Learning Lab (SVL).
My industry experience is in building algorithms and software for intelligent systems and large scale data analysis, specifically related to smart environments and Internet of Things. I have worked with GE Digital, the OpenFog Consortium, The Hive, and several startups.
CV, Github, LinkedIn
Papers | Talks | Posters
Multi Agent Option Learning for Team Planning
This ongoing research is advised by Professor Yisong Yue at Caltech and in collaboration with Stephan Zheng and Anshul Ramachandran. We aim to use reinforcement learning to learn the team coordination and planning that we see basketball and soccer teams exhibit. Specifically, we employ Multi-Agent Deep Deterministic Policy Gradients with unsupervised option learning in the RoboCup Half Field Offense Simulation environment. We hope to submit this to either ICML 2018 or NIPS 2018.
Grounded Neural Program Generation
This ongoing research is advised by Professor Silvio Savarese and Professor Fei-Fei Li at Stanford and in collaboration with De-An Huang, Animesh Garg, and Danfei Xu. We aim to tackle one shot visual imitation learning, but doing so while enforcing an internal task representation. More generally, we want to be able to go from a single demonstration of an unseen task to some representation of the task, then build a policy conditioned on the representation to complete the task. We hope to submit this to RSS 2018 or ECCV 2018.
Neural Task Programming
This project was advised by Professor Silvio Savarese and Professor Fei-Fei Li at Stanford and in collaboration with Danfei Xu, Animesh Garg, Yuke Zhu, and Julian Gao. We aim to tackle generalization in imitation learning, where given some domain knowledge in a type of task, one can learn a policy to execute a new task with just one video demonstration. Specifically, we use a policy instantiated as a Neural Program, conditioned on a video demonstration. This work is under review at ICRA 2018 and appeared as a talk and poster at CoRL 2017. For more details see here
Machine Learning: Applying Neural Networks to IoT Use Cases
I moderated a panel at the 2017 Internet of Things Solutions World Congress on machine learning applied to IoT. My fellow panelists included Dr Richard Soley, chairman of the IIC, Dr. Stan Schneider, CEO of RTI, and Dr. Joe Paradiso, Professor at the MIT Media Lab. Please find the full session here
Annotated Reconstruction of 3D Spaces Using Drones
This project began as a class project in Caltech CS/ME 132B and CS/EE 148, taught by Professors Joel Burdick and Pietro Perona. The goal of this project was the create an end to end perception, mapping, and expolration framework for drones. This involved (1) creating an optimized version of Faster RCNN to run on drone hardware, (2) developing a method to convert 2D images with bounding boxes and drone position and orientation to a 3D grid with annotations, and (3) designing a planning algorithm for exploring the space. My co-authors Anshul Ramachandran and Peter Kundzicz and I extended this work and submitted it to the 2017 IEEE MIT Undergraduate Research in Technology conference, where it won the Best Paper Presentation Award
Improving the Earthquake Early Response System Using Machine Learning
The goal of the project was to use machine learning methods to improve the accuracy and efficiency of the California Earthquake Early Response System. This system is implemented throughout California, and uses real time waveform data from several hundred seismological stations to raise an alarm of an earthquake. This system is critical, as it allows for a warning that despite only being ~10 seconds in advance is capable of halting processes at airports and metros, saving lives. In a team of 5 undergraduates, advised by Professor Yisong Yue, Professor Steven Low, and Postdoc Men-Andrin Meier, we developed a method that was both efficient and accurate, halving the existing number of false positives. We aim to publish the work soon and in the long term hope to incorporate our algorithm into the production system. This work also appeared as a poster at the 2017 Caltech Meeting of the Minds.
I can be reached by email at email@example.com