New Pruning Method Boosts Echo State Network Efficiency

A new pruning technique, Dynamical Mode Pruning (DMP), has been developed for Echo State Networks (ESNs), which are widely used for temporal prediction. Unlike traditional methods that rely on static network properties, DMP evaluates neurons based on their dynamic influence on the network's state transitions. By identifying and removing neurons with low impact on these dominant modes, the method effectively reduces the over-parameterization and dynamic redundancy often found in ESN reservoirs. The DMP process involves ranking neurons by their contribution to key transition modes, derived from a trajectory-averaged Jacobian Gramian. After pruning, only the readout layer of the ESN is retrained. Experiments on various time-series benchmarks, including chaotic and real-world data, demonstrate that DMP can either enhance or preserve the forecasting accuracy of ESNs while making them more efficient. This suggests that considering the dynamical role of neurons offers a superior criterion for refining reservoir components compared to static structural importance alone.