[2K6-ES-2] Machine learning: Modeling

[2K6-ES-2-01] Selecting Neurons of Deep Echo State Networks as a Readout Layer via Generalized-Ensemble Algorithms

〇Keizaburo Nishikino¹ (1. SoftBank Corp.)

Keywords:Reservoir computing, Echo state network, Deep echo state network, Generalized-ensemble algorithm

A deep echo state network (DeepESN), a type of recurrent neural network with an adaptable readout layer and fixed layers, effectively extracts hidden states from temporal sequences. However, it is computationally complex to identify appropriate neurons and layer depths of DeepESNs as the readout layer from many candidates. To address this problem, we propose applying generalized-ensemble algorithms for selecting neurons of DeepESNs based on the information criteria. We show that our approach can identify suitable neurons and layer depths of DeepESNs efficiently. We also show that the forecast accuracy is better than other neural network architectures.