Capsule networks with routing annealing

Abstract

Capsule Networks overcome some shortcomings of convolutional neural networks organizing neurons into groups of capsules. Capsule layers are dynamically connected by means of an iterative routing mechanism, which models the connection strengths between capsules from different layers. However, whether routing improves the network performance is still object of debate. This work tackles this issue via Routing Annealing (RA), where the number of routing iterations is annealed at training time. This proposal gives some insights on the effectiveness of the routing for Capsule Networks. Our experiments on different datasets and architectures show that RA yields better performance over a reference setup where the number of routing iterations is fixed (even in the limit case with no routing), especially for architectures with fewer parameters.