RUB » Bibliotheksportal

Bakr Al Beattie, Petro Feketa, Karlheinz Ochs, Hermann Kohlstedt

• Reservoir computing is an efficient and flexible framework for decision-making, control, and signal processing. It uses a network of interacting components varying from abstract nonlinear dynamical systems to physical substrates. Despite recent progress, the hardware implementation with inherent parameter variability and uncertainties, such as those mimicking the properties of living organisms' nervous systems, remains an active research area. To address these challenges, we propose a constructive approach using a network of FitzHugh-Nagumo oscillators, exhibiting criticality across a broad range of resistive coupling strengths and robustness without specific parameter tuning. Additionally, the network's activity demonstrates spatial invariance, offering freedom in choosing readout nodes. We introduce an alternative characterization of criticality by analyzing power dissipation, and demonstrate that criticality supports the robustness of the classification accuracy with respect to the readout shrinkage. Our results indicate criticality as a valuable property for classification problems, and provides design concepts for bio-inspired computational paradigms.