Capsule Networks

An alternative to pooling in convolutional neural networks.

Problems with Pooling

Sometimes when outputting a smaller representation of an image, spatial information is lost. We can use capsule networks to avoid this.

What are Capsule Networks?

They detect parts in an object within the context of the spatial information.

Capsule networks are made of parent and child nodes that build up a complete picture of an object.

Capsule Tree

What are Capsules?

Capsules are a collection of nodes. Each node contains information about a specific part (width, orientation, colour), and outputs a vector with:

• Magnitude ($m$) = the probability that a part exists; a value between 0 and 1.

• Orientation ($\theta$) = the state of the part properties


Pytorch Implementation

Here's a Github repo of a PyTorch implementation.