The journey from solving hand-written recognition to understanding more general natural images is a challenging one, requiring the creation of several invariances that can match the performance of the best neural networks with far fewer examples.