Revising recurrent neural networks from a granular perspective

In this paper, we formally deduce a new computational model, with a recurrent structure, by means of data granulation. The proposed scheme can be regarded as an Echo State Network (ESN), with an additional granular layer. ESNs have been recently revisited in the context of deep learning. In view of such a state-of-the-art, and coherently with the concept of data granulation, the aim herein is to propose a more efficient and transparent structure. The stability of the proposed scheme is formally discussed. The performance is shown by means of several benchmarks against the state-of-the-art methods. The proposed architecture exhibits a lower computational cost and a higher accuracy.