What Makes a Brain Smart? Reservoir Computing as an Approach for General Intelligence

Recurrent connectivity, balanced between excitation and inhibition, is a general principle of cortical connectivity. We propose that balanced recurrence can be achieved by tuning networks near their critical branching (CB) points when spike propagation is formalized as a branching process. We consider critical branching networks as foundations for artificial general intelligence when they are analyzed as reservoir computing models. Our reservoir models are based on principles of metastability and criticality that were developed in statistical mechanics in order to account for long-range correlations in activities exhibited by many types of complex systems. We discuss reservoir models and their computational properties, and we demonstrate their versatility by reviewing a number of applications.