My Publications

Image compression by storing the gray-scale image and some seeds of color pixels. Semi-supervised regression is used to recover the remaining colors. Active learning algorithm that minimize the recover error is proposed to select optimal color seeds. The use of encoded difference-image further improve the colorization quality drastically.

We use parallel vector field on the manifold to regularize (the gradient field of) a function so that it would be as "linear" as possible on the manifold.

Feature selection by minimizing the variance of the learned model parameters (similar objective function to optimal experimental design). A Laplacian regularizer is added to the model to let it adapt to the underlay geometry of the data manifold.

We proposed a simple unsupervised feature selection algorithm for multi-cluster data. The features are selected to optimally preserve the multi-cluster structure of the data. The optimization problem is solved via a sparse eigen-problem and a $\ell_1$-regularized least square problem.