Metastable attractors in secondary motor cortex underlie self-initiated behavior

The decision to choose specific self-initiated actions and execute them at certain times depends on several deterministic factors including the animal's goals and internal state, and external stimuli. Variability in self-initiated actions may have a stochastic origin as well, as suggested by the large variability in action timing. Here, we investigated the neural mechanisms underlying behavioral variability in a self-initiated waiting task. We recorded neural ensemble activity in secondary motor cortex (M2), which was previously shown to encode timing variability in the task. We found that M2 ensemble activity unfolded through sequences of metastable attractors. Attractor onset times showed large variability across-trials, yet reliably predicted upcoming self-initiated actions, allowing us to establish a dictionary between attractors and actions. To explain the mechanism generating reliable sequences of metastable attractors, we proposed a mesoscale circuit model, where M2 is reciprocally coupled to a subcortical area. Transitions between attractors in the model are generated by low-dimensional correlated variability, which we evinced in the empirical data as low-dimensional noise correlations. Our work establishes a framework for investigating the circuit origin of self-initiated behavior, based on mesoscale recurrent networks supporting attractor dynamics. Our theory suggests a
robust mechanism underlying self-initiated actions based on correlated neural variability.