Capsule networks with non-iterative cluster routing

Capsule networks use routing algorithms to flow information between consecutive layers. In the existing procedures, capsules produce predictions (termed votes) for of next layer. a nutshell, next-layer capsule’s input is weighted sum over all votes it receives. this paper, we propose non-iterative cluster capsule networks. proposed routing, vote clusters instead individual capsules, and each sends its centroid capsule. Generally speaking, The that comes from with smaller variance assigned larger weight in process. Compared state-of-the-art networks, achieve best accuracy on Fashion-MNIST SVHN datasets fewer parameters, smallNORB CIFAR-10 moderate number parameters. also disentangled representation generalize well images captured at novel viewpoints. preserve 2D spatial an image channels: if channels are rotated, object reconstructed these will be rotated by same transformation.