Clustered Echo State networks for signal denoising and frequency filtering

• Ano de defesa: 2020
• Autor(a) principal: Oliveira Junior, Laercio de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Artificial neural networks; Clustered networks; Complex networks; Echo state networks; Redes com clusters; Redes complexas; Redes neurais artificiais; Reservoir computing
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/59/59143/tde-28022021-205755/

This dissertation aims to study a type of Artificial Neural Networks (ANNs), known as Reservoir Computing, specifically, the Echo State Networks (ESNs). ESNs are Recurrent Neural Networks (RNNs), which make input-output mapping through a high dimensional nonlinear projection, called reservoir. In a classic ESN, the internal connection matrix of the reservoir usually is formed by an Erdös-Rényi random graph. Recent studies have also investigated Clustered ESNs (CESNs), which replaces the random network inside the reservoir by a clustered network. Both types of ESNs have been applied to time series prediction problems. In this work, an ESN with a clustered Barabási-Albert network (Barabási-Albert CESN), and a deep ESN with clustered reservoir layers (Deep CESNs) are designed. Moreover, we propose to apply ESNs in two new different tasks: the frequency filtering problem and the noise filtering problem of time series. We also compare the performance of the classical ESN and its various extensions in these two tasks. Numerical results show that the proposed ESNs (Barabási-Albert CESN and Deep CESNs) outperform the classical ESN, indicating that the organization of reservoirs in clustered or layered networks can improve the learning performance of ESNs.