Capsule Routing is a mechanism in Capsule Networks that determines how signals from capsules in one layer are passed on to capsules in the next layer.
Unlike classical neural networks, which use fixed connections, Capsule Routing employs dynamic methods where the forwarding of information is based on agreements between predictions.
A well-known method is dynamic routing by agreement, where capsules weight their output signals depending on how strongly they agree with the predictions of higher-level capsules.
This allows the network to better capture hierarchical relationships and spatial contexts and to generate robust feature representations.
Capsule Routing is central to the ability of Capsule Networks to recognize objects despite translation, rotation, or occlusion.