Minimum Complexity Deep Echo State Networks

In the context of Recurrent Neural Networks (RNN), suitable for the processing of temporal sequences, the Reservoir Computing approach and specifically the Echo State Networks (ESN) are gaining particular interest. Their peculiarity is the simplicity of model training without paying in terms of predictive performance. A point of these networks on which it is possible to make improvements is the initialization of some parts of the network, reducing the overall number of hyper-parameters and increasing the ease of use. Innovatively, this thesis presents a new ESN model, called Minimum Complexity Deep Echo State Network (MCDeepESN). The main objective of the model described is, on the one hand, to simplify the overall architecture of the multilayer ESN model (DeepESN), modifying some parts of the network, on the other hand, to maintain and possibly improve the predictive performance of the model. Several architectural alternatives of the MCDeepESN model have been implemented and numerous experiments have been executed during the work that brought to light this thesis. To test the performance of the proposed architecture, several datasets were used, both synthetic and from the real-world, including medical data for automatic diagnostics. From the experiments executed it was possible to see that the performance achieved by the MCDeepESN is equal or superior to other previous models, with the additional advantage of the decreased complexity of the model