(a)

(b)

(c)

Figure 3: Abs. training set size versus test set accuracy for pixel-based features on MNIST, USPS, and DIGITS (IBk variants).