Abstract
We study the information processing abilities of active matter in thereservoir computing (RC) paradigm, using a model that is externally driven toinfer the future state of a chaotic signal. The simulated system closelyfollows a previously reported model. We uncover an exceptional dynamical regimeof agent dynamics that has been overlooked heretofore. It appears robustlyoptimal across varying physical parameters and inference tasks, thus providingvaluable insights into computation and inference with physical systems moregenerally. The ability to form effective mechanisms for information processingare primarily determined by the system's own intrinsic relaxation abilities.These are identifiable when probing the system without a specific inferencegoal and manifest when testing minimalistic single-particle reservoirs. Theregime that achieves optimal computation is situated just below the criticaldamping threshold, involving a microscopic dynamical relaxation with multiplestages. The optimal system is adaptable under chaotic external driving, due toa diversity in response mechanisms that emerge like rapid alternations betweenquasi-stationary and highly nonlinear dynamical states. Both coherent andincoherent dynamics contribute to their operation, partly at dissimilar scalesof space and delay time. Correlations on agent dynamics can indicate thebest-performing regimes and onsets of tight relationships between theresponding system and the fluctuating driver. As this model of computation isinterpretable in physical terms, it facilitates re-framing inquiries regardinglearning and unconventional computing with a fresh rationale for many-bodyphysics out of equilibrium.