Capsule Networks (CapsNets) are a neural network architecture paradigm introduced by Sabour et al. (2017) to overcome the limitations of CNNs in modeling spatial relationships between features.
A capsule is a group of neurons whose output signals represent a vector encoding both the probability of the existence of an object and its instantiation parameters (e.g., pose, deformation).
The core mechanism is dynamic routing-by-agreement, where lower-level capsules forward their outputs to higher-level capsules when their predictions agree, thereby modeling a hierarchical and part-whole structure.
This architecture enables better generalization under transformations and viewpoint changes compared to classical CNNs.
Definition:
“Capsule Networks are neural networks that use groups of neurons (capsules) with vector outputs to encode both the existence and instantiation parameters of objects, and utilize dynamic routing to model hierarchical relationships.”
Source:
Sabour, S., Frosst, N., & Hinton, G. E. (2017). Dynamic Routing Between Capsules. Advances in Neural Information Processing Systems (NeurIPS).