Current research
My current research is focused on nonparametric Bayesian learning. Frank Wood and I have developed a novel nonparametric Bayesian model called the structured infinite hidden semi-Markov model. It is a flexible generalization of the non-parametric variant of the Bayesian HMM, the HDP-HMM. Our model is the first nonparametric variant of the HMM to explicitly parameterize the state dwell duration (as opposed to the implicit geometric distribution of the HDP-HMM). It is also capable of having a wide variety of state transition topologies — for example, only allowing transitions to nearby states or to “later” states. It recovers the HDP-HMM and Bayesian HMM as special cases.
I am collaborating with Liam Paninski and Ari Pakman on a model for inferring the locations and strengths of synaptic connections based on sparse, noisy voltage measurements of dendritic trees. We adapted Efron et al.’s LARS-lasso algorithm. Our algorithm takes advantage of special properties of the problem so that it can run very efficiently — much faster than “LARS” or general quadratic programming techniques
Past research
Liam Paninski and I developed an algorithm and series of heuristic approximations for efficient experimental design of dendritic voltage sampling schemes. Part of this work involved extending an algorithm that Professor Paninski had previously developed for approximating the forward Kalman filter using a low rank matrix technique to full the forward-backward Kalman smoother case. This approximation allows the smoother to run in linear time (and space) rather than cubic time (and quadratic space) in the size of the system’s state space.
From 2007-2009 I was a technical co-op at the MITRE Corp., where I worked on natural language processing problems such as co-reference finding, action recognition in free text, and named entity recognition. Most of this work made use of conditional random fields. Before MITRE I spent a summer working on the robot Trisk at the MIT Media Lab, in Professor Deb Roy’s Cognitive Machines Group.
© 2011 Jonathan Huggins | Generated by webgen | Design by Andreas Viklund.