Capsule Network
A neural network architecture proposed by Geoffrey Hinton that uses groups of neurons (capsules) to better capture spatial hierarchies and part-whole relationships in data.
Motivation
Standard CNNs lose spatial relationships through pooling. Capsule networks preserve the pose (position, orientation, scale) of detected features, enabling better understanding of how parts compose into wholes.
Status
Despite theoretical elegance, capsule networks haven't achieved the scalability or performance of standard CNNs or vision transformers in practice, and remain primarily a research interest.