- Keywords: capsule networks, routing, attention
- TL;DR: We present a new routing method for Capsule networks, and it performs at-par with ResNet-18 on CIFAR-10/ CIFAR-100.
- Abstract: We introduce a new routing algorithm for capsule networks, in which a child capsule is routed to a parent based only on agreement between the parent's state and the child's vote. Unlike previously proposed routing algorithms, the parent's ability to reconstruct the child is not explicitly taken into account to update the routing probabilities. This simplifies the routing procedure and improves performance on benchmark datasets such as CIFAR-10 and CIFAR-100. The new mechanism 1) designs routing via inverted dot-product attention; 2) imposes Layer Normalization as normalization; and 3) replaces sequential iterative routing with concurrent iterative routing. Besides outperforming existing capsule networks, our model performs at-par with a powerful CNN (ResNet-18), using less than 25% of the parameters. On a different task of recognizing digits from overlayed digit images, the proposed capsule model performs favorably against CNNs given the same number of layers and neurons per layer. We believe that our work raises the possibility of applying capsule networks to complex real-world tasks.