Gioele Zerini, Andrea Ceni, Andrea Cossu, Davide Bacciu, and Claudio Gallicchio, "Non-Dissipative Random Oscillators Networks through Negative Feedback Coupling", in Proc. World Congress on Computational Intelligence, 2026.
Abstract: The careful design of neural architectures is a prominent research area in machine learning. Recurrent neural networks are notoriously tricky to train, making their architectural design often a determining factor for performance. We adopt the reservoir computing approach, where the recurrent component of the network is left untrained, but its architectural design guarantees effective sequential data processing abilities. Inspired by the Random Oscillators Network (RON), we propose the Non-Dissipative RON (ND-RON) and benchmark it against popular time series datasets, showcasing its effectiveness against RON and other popular reservoir computing approaches. We provide mathematical proof of the increased ability of ND-RON to propagate information without dissipation over long time spans, surpassing other reservoir computing approaches and remaining competitive with fully trainable recurrent models while requiring one order of magnitude fewer parameters.
More Information
