My current research is focused on bringing in structured outside knowledge into end-to-end vision systems. My accepted CVPR 2017 paper investigates the use of knowledge graphs in image classification. Future work will investigate the use of graphs in other vision problems and mining knowledge from the web.
Some past and current on generative models for computer vision. In just the past few years researchers have started creating models that can create novel images approaching the complexity and resolution of natural images. I am particularly excited by the implications of this work on representation learning and semi-supervised learning.
One of my ongoing projects. The basic idea is to get a real-time system that can accurately estimate the pose of a human subject using a camera. The project is with Professor Yaser Sheikh at Carnegie Mellon. My main responsibility is to improve the the speed (real time, remember) using GPU acceleration.
Another of my ongoing projects. This work is the implementation of the idea introduced in a recent ICML Paper -- A Physics-Based Model Prior for Object-Oriented MDPs. My contribution was the implementation of the overhead camera tracking system.
This project was a natural extension of a paper demonstrating basis function changes for Temporal Difference to LSPI. I worked on this with Professor Charles Isbell at Georgia Tech.
An old idea I had about applying deep belief network learning for Temporal Difference. The game of choice was Go because of the large state space of the game.
RoboJackets is the Georgia Tech student robotics organization. I was involved in the Intelligent Ground Vehicle Competition (IGVC) Team first as a programmer and later as project manager.
A project I worked on very briefly during my semester abroad with Professor Cédric Pradalier. The webpage for the project can be found here.
Manifold is an open source infrastructure for the modeling and simulation of many-core processors. My work with Professor George Riley was to test the behavior of the system on pthreads programs.
An independent project I started my freshman year. The basic idea is to train a neural network to accurately predict when some of its inputs are missing or invalid.
Here is just a collection of the various tidbits of code and other deliverables. A fun mix of assembly programs, vhdl projects and silly java apps.