Capsule Networks for Attention Under Occlusion

Abstract

Capsule Neural Networks (CapsNet) serve as an attempt to model the neural organization in biological neural networks. Through the routing-by-agreement algorithm, the attention mechanism is implemented as individual capsules that focus on specific upstream capsules while ignoring the rest. By using the routing algorithm, CapsNets are able to attend overlapping digits from the MNIST dataset. In this work, we evaluate the attention capabilities of Capsule Networks using the routing-by-agreement with occluded shape stimuli as presented in neurophysiology. We do so by implementing a more compact type of capsule network. Our results in classifying the non-occluded as well as the occluded shapes show that indeed, CapsNets are able to differentiate occlusions from near-occlusion situations as in real biological neurons. In our experiments, performing the reconstruction of the occluded stimuli also shows promising results.