Group Equivariant Capsule Networks

We present group equivariant capsule networks, a framework to introduce guaranteed equivariance and invariance properties to the capsule network idea. We restrict pose vectors and learned transformations to be elements of a group, which allows us to prove equivariance of pose vectors and invariance of activations under application of the group law. Requirements are a modified spatial aggregation method for capsules and a generic routing by agreement algorithm with abstract rules, which we both present in this work. Further, we connect our equivariant capsule networks with work from the field of group convolutional networks, which consist of convolutions that are equivariant under applications of the group law. Through this connection, we are able to provide intuitions of how both methods relate and are able to combine both approaches in one deep neural network architecture, combining the strengths from both fields. The resulting framework allows sparse evaluation of feature maps defined over groups, provides control over specific equivariance and invariance properties and uses routing by agreement instead of pooling operations. It provides interpretable and equivariant representation vectors as output capsules, which disentangle evidence of object existence from its pose.