Orchestrated Excitatory and Inhibitory Learning Rules Lead to the Unsupervised Emergence of Up-states and Balanced Network Dynamics

Abstract

Self-sustaining neural activity maintained through local recurrent connections is of fundamental importance to cortical function. We show that Up-states--an example of self-sustained, inhibition-stabilized network dynamics--emerge in cortical circuits across three weeks of ex vivo development, establishing the presence of unsupervised learning rules capable of generating self-sustained dynamics. Previous computational models have established that four sets of weights (WE[<-]E, WE[<-]I, WI[<-]E, WI[<-]I) must interact in an orchestrated manner to produce Up-states, but have not addressed how a family of learning rules can operate in parallel at all four weight classes to generate self-sustained inhibition-stabilized dynamics. Using numerical and analytical methods we show that, in part due to the paradoxical effect, standard homeostatic rules are only stable in a narrow parameter regime. In contrast, we show that a family of biologically plausible learning rules based on "cross-homeostatic" plasticity robustly lead to the emergence of self-sustained, inhibition-stabilized dynamics.