Transferring Reservoir Computing: Formulation and Application to Fluid Physics

Abstract

We propose a transfer learning for reservoir computing, and verify the effectivity of the proposed methods for the standard inference task of the Lorenz system. Applying the proposed methods to an inference task of fluid physics, we show the inference accuracy is drastically improved compared with the conventional reservoir computing method if available training data size is highly limited.