Uno studio esplorativo sulle capsule networks e su come renderle più profonde

Capsule networks, a novel deep learning architecture proposed by Hinton, introduce the idea of encoding the instantiation parameters of the entities into groups of neurons, called capsules. These capsules are particularly suitable for handling different types of visual stimuli and coding features like pose (position, size, orientation), deformation, speed, etc. It is assumed that the brain has a mechanism for ¿routing¿ low-level visual information to the capsules best suited to manag- ing that data. Capsule networks aim to overcome the limitations of convolutional networks in object recognition. The present work aims to understand these limitations and to test the robustness of capsule networks to affine transformations and to novel viewpoints. This is achieved by implementing a PyTorch framework used to build deeper capsule networks and to conduct more experiments on the task of image classification on datasets more complex than MNIST. Despite their promising results, capsule networks require a large computational cost in terms of speed computation and memory usage. Further research in their optimization is probably needed for them to be a good replacement of convolutional networks for industrial applications.