Hands-On Machine Learning for Algorithmic Trading

How capsule networks capture pose

CNNs have a key weakness due to their invariance to spatial translations of features. They do not learn relative positions of the various features they detect to recognize or locate an object. As a result, the image of a face with key features such as eyes, ears, and nose swapped would still be recognized as a human face, or a particular person.

Geoff Hinton, one of the leading researchers in the field, recently proposed a new architecture called capsule nets, which may yield significant improvements relative to CNN. The paper is called Dynamic Routing Between Capsules, and the new model stands out through its ability to encode the pose of an object. As a result, it is better in detecting objects that are ...