Fully Distributed, Flexible Compositional Visual Representations via Soft Tensor Products