Revisiting the memory capacity in reservoir computing of directed acyclic network

Area of Science:

• Computational neuroscience
• Machine learning

Background:

• Reservoir computing (RC) offers hardware implementation potential and low training costs.
• Interpreting RC dynamics via memory capacity (MC) is an active research area.
• Directed acyclic networks (DANs) are used as reservoir architectures in RC.

Purpose of the Study:

• To theoretically bound the global memory capacity (MC) of RC with DAN reservoirs.
• To investigate the influence of hyperparameters on MC.
• To identify and exploit memory communities within DANs for improved RC performance.

Main Methods:

• Theoretical analysis of MC bounds in DANs.
• Hyperparameter dependency analysis of MC.
• Application of the struc2vec network embedding method to detect memory communities.
• Development and evaluation of a clustered RC model.

Main Results:

• Global MC of RC with DANs is theoretically bounded by the longest path length.
• Hyperparameters significantly influence MC.
• Memory communities within DANs correlate with MC and provide interpretability.
• The clustered RC model shows enhanced prediction accuracy and reduced training cost on chaotic time series.

Conclusions:

• The longest path in a DAN provides a theoretical upper bound for RC's MC.
• Network embedding methods can reveal interpretable community structures related to memory.
• Exploiting these memory communities leads to improved and more efficient clustered RC models.