Arthur Asuncion

--------------

Research Interests

My area is artificial intelligence and machine learning. I'm part of the DataLab group with Prof. Smyth. I'm currently working on developing efficient and scalable statistical inference techniques for tasks such as topic modeling, collaborative filtering, and social network analysis. Computational efficiency is of increasing interest due to the increasing complexity of models (which may be nonparametric and hierarchical) and the sheer amount of data available to us.

Specifically, I'm interested in using parallel/distributed computing to speed up machine learning algorithms and allow them to handle massive data sets. I'm also interested in accelerating the approximate inference techniques themselves, by using improved MCMC samplers, faster variational techniques, particle filtering, improved contrastive divergence techniques, and other tricks. By combining these techniques with parallel computing, one can potentially achieve compounded speedups. For instance, one can learn a topic model on a corpus with thousands of documents in a single second by running a fast variational algorithm (CVB0) on an 8-core machine.

Other topics that interest me are computer music and signal processing. As an undergrad, I once made a wavelet-based signal processor that produced some interesting sounds.

Another area that I would like to pursue in my spare time is what I have termed "software biomimetics". Biomimetics is a discipline which attempts to mimic useful designs found in nature and apply these designs to technology. Visit Georgia Tech's Center for Biologically Inspired Design for some examples of biomimetics. What I would like to do is extend this biological mimicry to the software realm. Is it possible to extract software design principles from the software of life?