Sparse and Misaligned Data

In the learning and recognition methods we introduced till now, we have not made strong assumptions about the data: Nearest neighbor, SVM, distance metric learning, normalization and decision trees do not explicitly assuming distributional properties of the data; PCA and FLD are optimal solutions under certain data assumptions, but they work well in many other situations too; parametric probabilistic models assume certain functional form of the underlying data distribution, but GMM and nonparametric probabilistic methods have relaxed these assumptions.