Abstract

Most existing neural networks for learning graphs address permutationinvariance by conceiving of the network as a message passing scheme, where eachnode sums the feature vectors coming from its neighbors. We argue that thisimposes a limitation on their representation power, and instead propose a newgeneral architecture for representing objects consisting of a hierarchy ofparts, which we call Covariant Compositional Networks (CCNs). Here, covariancemeans that the activation of each neuron must transform in a specific way underpermutations, similarly to steerability in CNNs. We achieve covariance bymaking each activation transform according to a tensor representation of thepermutation group, and derive the corresponding tensor aggregation rules thateach neuron must implement. Experiments show that CCNs can outperform competingmethods on standard graph learning benchmarks.

 

Introduction (beta)

None

 

Conclusion (beta)

None