Thresholding Based Methods for Robust Learning

Learning in presence of outliers is a critical problem that can heavily affect performance of the learning algorithms in practice. In this talk, we present a general approach for learning with outliers, where we iteratively estimate the model parameters with estimated inliers and threshold out point which seems unlikely to be generated from the model to obtain more refined set of inliers. We instantiate this general approach for the outlier efficient PCA problem and demonstrate that it leads to nearly optimal solution in O(PCA) computation time.