ETD

• Artificial Intelligence
• Deep Learning
• Echo State Network
• Modularity
• Neural Network
• Reservoir Computing

In this work, we present two novel architectures, Modular Cycle Echo State Network (MC-ESN) and Modular Antisymmetric Echo State Networks (MA-ESN), in the field of Reservoir Computing (RC). By changing the usual deep hierarchical setting to a structured one that relies on information flow across networks. Subsequently, we extend these structured modularities to a promising RC model: Recurrent Oscillator Networks, which offers a more expressive internal representation than simple ESNs. Both novel architectures are inspired by recent advances in the RC field and use deep-layer construction with a structured modular topology. We evaluate the proposed models on several benchmark tests for RC, showing that they are able to outperform on memory capacity and long-term prediction tasks by comparing state-of-the-art shallow and deep RC architectures. We briefly analyse their linearized behaviour in order to provide insights into the role of modularity and depth in RC.