The idea in plain language.
A part should send evidence to the whole whose predicted properties agree with it.
How it works
A capsule groups neurons into a vector representation of an entity. Lower-level capsules predict the properties of higher-level entities. The routing procedure increases the connection strength when those predictions agree, allowing a collection of parts to support a common interpretation. The paper tests this approach on digit-recognition tasks, including overlapping digits.
What to keep in mind
The reported digit-recognition experiments do not establish that capsule networks replace conventional visual models across large, varied datasets.
Source: Dynamic Routing Between Capsules. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.