I am a first year Phd student in the Machine Learning Department at Carnegie Mellon University. My advisor is Carlos Guestrin. Currently I am working on parallel machine learning on graphical models. Other interests include Computer Vision, Operating Systems and Computer Architecture and Go.
While the CS algorithm community has been analyzing parallel algorithms for decades, the Machine Learning community is still far behind and has not caught up with recent changes in CPU architecture. One could think that parallelizing a program is mostly an engineering task and is therefore uninteresting; but what do you do with an inherently sequential algorithm? (such as most online learning methods). It might also be possible that the inherently parallel solution is vastly less efficient than the optimal sequential solution. More intriguing is the possibility that one may be able to "sacrifice" some accuracy to get greater parallelism.