(a)

(b)

(c)

Figure 6: Abs. training set size versus test set accuracy for gradient features on MNIST, USPS, and DIGITS (SVM variants).