1) Multipie.mat and Mnist.dat contain the following:

  a) Xtr: A num_dimensions by num_examples matrix of training data
  b) Ytr: A num_examples   by 1 vector of training labels

  c) Xte: A 10 by 1 cell array, where each cell is a num_dimensions by num_examples matrix
  d) Yte: A 10 by 1 cell array, where each cell is a num_examples by 1 vector

  -The test data is represented in 10 random batches, which allows us to compute error plots.
  -Both the training and testing features have been projected into a low-dimensional space (with PCA)
   so as to capture 95% of the empirical data variance.
  -MNIST features:    vectorized pixel-images of deskewed pixels
  -MultiPIE features: normalized pixel-images of Viola-Jones face detections

2) Caltech.mat is similar, but the training data is also specified in a 10 by 1 cell array. We performed experiments with 30 training images per class and 1 test image per class. We needed to recompute the set of 30 training images for each of the 10 random batches of test images.

  -We use Lazebnik's et al implemention of spatial pyramid features, projected down to capture 95% of the their variance.