Dynamical Mode Pruning: A New Method to Optimize Echo State Networks
A recent study introduces Dynamical Mode Pruning (DMP), a technique designed to prune reservoirs in Echo State Networks (ESNs) by evaluating neurons according to their influence on key transition modes obtained from a trajectory-averaged Jacobian Gramian. This method differs from traditional pruning techniques that depend on static connectivity or activation metrics, which might miss neurons essential for input-driven state changes. DMP eliminates low-impact neurons and only retrains the readout layer, aiming to streamline reservoir components while preserving or enhancing forecasting accuracy. Tests on chaotic and real-world time-series data indicate that DMP either maintains or boosts forecasting accuracy, highlighting the importance of dynamical influence over mere static structural significance. The research is published on arXiv with the identifier 2608.04593 in the Computer Science > Machine Learning category.
Key facts
- The paper proposes Dynamical Mode Pruning (DMP) for Echo State Networks (ESNs).
- DMP ranks neurons by their contribution to dominant transition modes from a trajectory-averaged Jacobian Gramian.
- Existing pruning methods rely on static connectivity or activation statistics.
- DMP removes low-impact units and retrains only the readout.
- Experiments on chaotic and real-world time-series benchmarks show improved or preserved forecasting accuracy.
- The paper is available on arXiv with ID 2608.04593.
- The paper is categorized under Computer Science > Machine Learning.
- DMP suggests dynamical influence is a useful criterion for reservoir refinement.
Entities
Institutions
- arXiv