Can a machine learn from a few labeled pixels to predict every pixel in a new image?