Editorial: Deep neural network architectures and reservoir computing
Sou Nobukawa, Arya K. Bhattacharya, Akira Hirose · Frontiers in Artificial Intelligence · 2025
Over the past decade, deep learning (DL) techniques such as convolutional neural networks (CNNs) and long short-term memory (LSTM) networks have played a pivotal role in advancing the field of computational intelligence (Bengio et al. (2021)). Recent developments in deep neural network (DNN) architectures and computational infrastructure (particularly parallel computing), have further accelerated the progress by supporting the computational demands of optimizing large numbers of network parameters. These advancements have expanded the applicability of DL to a broad range of tasks in computational intelligence (Sharifani and Amini (2023)).Simultaneously, reservoir computing (RC) has attracted increasing attention (Tanaka et al. (2019)).Typically, RC consists of a fixed recurrent neural network (the reservoir) and a trainable readout layer. It exploits the nonlinear spatiotemporal dynamics of the reservoir to transform inputs, while learning is applied only to the output layer. This structure dramatically reduces the number of trainable parameters, resulting in high learning efficiency. However, conventional RC, which typically involves a single reservoir layer, has generally not matched the performance of deeper neural architectures used in mainstream DL.