Artificial Neural Networks, Capsule Neural Networks

Convolutional Neural Networks have achieved much success in the recent years by demonstrating great results. On the other hand, there are inherent limitations, because these architectures do not take into an account pose of an entity, meaning that an object with different viewing angles is not being recognized as the same entity. Also, the relative positions of features are not being utilized. Capsule Neural Networks, recently made by Geoffrey Hinton and his collaborators are especially designed with this problem in mind. By grouping neurons in layers into capsules, we can create vectors containing information about poses. By using information about relations between capsules, we can create higher-level features which are invariant to rotation. This results in better generalization to unseen poses and decreases the required size of training datasets. In our work, we focus on applying this technique to standard image datasets and comparing the results with other approaches. We also test the benefits and shortcomings of Capsule Neural Networks.