Researchers applied the mathematical theory of synchronization to clarify how recurrent neural networks (RNNs) generate predictions, revealing a certain map, based on the generalized synchronization, that yields correct target values. They showed that conventional reservoir computing (RC), a type of RNN, can be viewed as a linear approximation, and introduced a 'generalized readout' incorporating further order approximations. Using a chaotic time-series forecasting task, they demonstrated that this approach dramatically enhances both prediction accuracy and robustness.
↧