This paper considers a recently proposed method for unsupervised learning and dimensionality reduction, locally linear embedding (LLE). LLE computes a compact representation of high-dimensional data combining the major advantages of linear methods (computational efficiency, global optimality, and flexible asymptotic convergence guarantees) with the advantages of non-linear approaches (flexibility to learn a broad of class on non-linear manifolds). We assess the performance of the LLE algorithm on a real-world data (face images in different poses) and compare the results with those obtained with two different approaches (PCA and SOM). Extensions to the original LLE algorithm are proposed and applied to the problem of pose estimation.