Learning The Discriminative Power-Invariance Trade-Off