An Effective Hit-or-Miss Layer Favoring Feature Interpretation as Learned Prototypes Deformations

Neural networks designed for the task of classification have become a\ncommodity in recent years. Many works target the development of more effective\nnetworks, which results in a complexification of their architectures with more\nlayers, multiple sub-networks, or even the combination of multiple classifiers,\nbut this often comes at the expense of producing uninterpretable black boxes.\nIn this paper, we redesign a simple capsule network to enable it to synthesize\nclass-representative samples, called prototypes, by replacing the last layer\nwith a novel Hit-or-Miss layer. This layer contains activated vectors, called\ncapsules, that we train to hit or miss a fixed target capsule by tailoring a\nspecific centripetal loss function. This possibility allows to develop a data\naugmentation step combining information from the data space and the feature\nspace, resulting in a hybrid data augmentation process. We show that our\nnetwork, named HitNet, is able to reach better performances than those\nreproduced with the initial CapsNet on several datasets, while allowing to\nvisualize the nature of the features extracted as deformations of the\nprototypes, which provides a direct insight into the feature representation\nlearned by the network .\n