Capsule networks as recurrent models of grouping and segmentation
Fig 4
Other network architectures do not explain uncrowding.
To verify that the ability of CapsNets to explain uncrowding is due to their architecture and not merely to the way they are trained, we replaced the recurrent routing by agreement between capsules by three different alternative architectures: a feedforward fully connected layer (yielding a classic ffCNN, a), a fully connected layer with lateral recurrent connections (b) and a fully connected layer with top-down recurrent connections to the layer below (c). The x-axes represent the (un)crowding configurations used during testing (shown on the right). The y-axes show the vernier discrimination performance as %correct for the whole configuration minus %correct for the central flanker alone as in Fig 3A. None of the ffCNN or recurrent CNN architectures we tested can produce uncrowding (compare with the CapsNet results in Fig 3A).