Justin DomkeEmail : (my last name) at cs.umd.edu Office : A.V. Wiliams 4470 Phone : 301.405.1762 update: I've moved on from Maryland. If you would like to contact me, please do so using the information at my new web page. I am a grad student at the University of Maryland. My current research focus is on practical techniques for learning and inference in complex, large-scale probabilistic graphical models. (i.e. graphical models with high treewidth) My advisor is Yiannis Aloimonos |
|
||
 CV: pdf text | Publications | Miscellaneous
Thesis
Justin Domke. Tractable Learning and Inference in High-Treewidth Graphical Models, Doctoral Dissertation, University of Maryland, 2008
Selected Projects
|
|||
|
Learning Convex Inference of Marginals
Algorithms for inference of marginals can be seen as minimizing a "free energy" function, subject to constraints. In general graphs, however, both the function and constraints must be approximated, greatly complicating learning. The goal here is to circumvent these issues by "fitting the inference procedure": In learning, seek the parameters that give maximally accurate predictions, taking into account approximations that must be made for the sake of tractability in inference. |
UAI 2008 pdf |
||
|
Crossover Random Fields
One practical technique for dealing with an intractable graphical model is to simply delete edges of the graph until a tree remains. The obvious downside is that much of the dependence between variables is lost. The idea of this project is, rather than choosing a single tractable subgraph, to choose a series, with the results of inference on one influencing the next. This has given encouraging results for imaging problems, where the subgraphs can consist of models over either rows or columns. |
|
||
|
Image Transformations and Blurring
Consider two images of a surface, taken from different positions. Since cameras blur incoming light before measurement, these two images contain different information about that surface. This project tries to understand this through contrasting between the "ideal image" (the unblurred incoming light) and the "real image" (the actual measured signal, post blurring). This ultimately leads to the application of "multiple view image reconstruction"-- combining the information in several images into one. Notably, this can be done despite not knowing (or estimating) the blurring kernel. |
 
|
||
"If we could first know where we are, and whither we are tending, we could better judge what to do, and how to do it." Abraham Lincoln |
|||
