In addition to the intro that I liked a lot, I find interesting, or even important, the usage of an unsupervised method to generate the visual symbolic dictiona

In addition to the intro that I liked a lot, I find interesting, or even important, the usage of an unsupervised method to generate the visual symbolic dictionary, a convolutional autoencoder here. In arxiv:1412.6856 unsupervised object detectors emerge as byproducts in a CNN for another kind of task, for more interesting, photorealistic objects.