A capsule network is a neural architecture that groups units into capsules representing entities and models relationships among their properties. Geoffrey Hinton introduced the approach in 2017 as a possible alternative to convolutional neural networks, aiming to add reasoning about entities within images. section 7.6
An elegant idea meets the benchmark
Capsule networks reached state-of-the-art results on smallNORB and came close on MNIST, but their training and scaling difficulties prevented them from matching leading CNNs on CIFAR and ImageNet. They became a cautionary example of why an appealing architectural idea still needs competitive empirical results. section 7.6
