We propose a hierarchical neural network architecture for unsupervised learning of
equivariant part-whole decompositions of visual scenes. In contrast to the global …