Inhibitory-stabilization is sufficient for history-dependent computation in a randomly connected attractor network

For effective information processing, the response to a sensory stimulus should depend on both the incoming stimulus and the history of prior stimuli. Existing models of neural circuits based on multiple attractor states produced with strong self-excitation can exhibit these properties, but they do not stabilize at biologically realistic firing rates. We demonstrate how a randomly connected inhibition-stabilized attractor network can preserve the computational abilities of recurrent excitatory networks, while stabilizing at arbitrarily low firing rates. Not only does excitatory-inhibitory balance stabilize network activity, inhibitory-stabilization also plays a functional role in history-dependent computation: transient oscillations made possible by inhibitory feedback are sufficient for state-dependent responses to stimulation. Such networks may underlie many cognitive tasks, suggesting a functional role for inhibition-stabilized dynamics in cortical computation. Author summaryGeneral cognitive behavior requires the interpretation of incoming information within its recent context. For example, in sports, a single sensory stimulus: the referees whistle, can convey diverse messages: "begin," "foul," or "goal" dependent on the events immediately preceding the whistle. We will refer to such situations as "history-dependent," indicating that the appropriate behavioral response to a given stimulus depends on the prior history of stimulation. History-dependent behaviors include counting, oral communication, and sequence discrimination. In each of these instances, behaviorally relevant information depends less on the characteristics of a single stimulus, but rather on the entire set of stimuli, often including their order. Thus, to perform a wide range of cognitive tasks, the brain must possess a mechanism for short-term memory in which neural responses to a given stimulus depend both on the characteristics of that stimulus and on the recent history of stimulation. Here we study how the dynamics of networks of neurons could support history-dependent behaviors.