SCS Undergraduate Thesis Topics
|Jay Park||Aarti Singh||Active Sampling for Estimating Gaussian Graphical Models|
We consider the problem of learning the structure of high-dimensional Gauss Markov Random fields using L1-regularized linear regression. In many applications, such as in sensor networks and proteomics, it may be costly to obtain joint observations repeatedly from all the variables involved. To address this, we propose an active learning algorithm that directs the learning process by selectively focusing the sampling resources on more uncertain variables as it proceeds. We show theoretically that this results in significant savings in terms of the number of samples required per variable. Furthermore, we demonstrate experimental results that corroborate our theoretical findings.