Full lecture: http://bit.ly/PCA-alg
PCA is sometimes used as a pre-processing step to reduce the dimensionality of the data before applying a supervised learning algorithm. Such pre-processing can sometimes be detrimental, because PCA is unaware of class labels and can pick unfortunate dimensions to project to. Linear Discriminant Analysis (LDA) attempts to rectify the problem by projecting to the dimensions that maximize the separation between the classes.