Abstract
Autoencoders are regarded as one of the key functional components of deep learning architectures. In this study, we augment the well-known architectures of autoencoders by incorporating a concept of information granularity, which gives rise to so-called granular autoencoders. It is demonstrated that information granularity can be sought as an essential design asset whose optimal allocation produces the autoencoder with better representation capabilities. Several protocols of allocation of information granularity are presented and assessed with regard to their abilities to represent the data. Selected examples including those dealing with clustering time series are included.