Capsule Networks as alternative to DCNN

Capsule Networks is an architecture of Deep Neural Networks specifically developed for image recognition. The main component of this architecture is the capsule, a group of artificial neurons that outputs a vector representing the instantiation parameters of an entity present inside the image. Information flows from one layer of capsules to the next through the Routing by Agreement algorithm, specifically developed for CapsNets. The aim of this thesis is to test Capsule Networks on different tasks where Convolutional Neural Networks are normally used. These tests include Image Classification on different datasets and resolutions, Image Segmentation and Generative Adversarial Networks. The results obtained reveal a promising but immature architecture; main problems reside in the difficulty of the Routing by Agreement algorithm in scaling to big and deep architectures. Images generated through a GAN using a CapsNet discriminator result in more accurate representations of objects depicted in various viewpoints, compared to a CNN discriminator. Further research should be focused on improving the CapsNet architecture for bigger and deeper networks, maintaining the concept of capsules as a key element.