The paper highlights the importance of feedback connections, recursive inference, lateral connections, and coming to the best explanation of a scene to solve recognition and segmentation problems. It argues that current deep learning systems take a lot of training data, and the model presented in the paper trains quickly with very little data.