Science Daily reports (here) on work by D. Durstewitz and others reported in Neuron, Abrupt Transitions between Prefrontal Neural Ensemble States Accompany Behavioral Transitions during Rule Learning.

While it is clear that new rules are often deduced through trial-and-error learning, the neural dynamics that underlie the change from a familiar to a novel rule are not well understood…”The ability of animals and humans to infer and apply new rules in order to maximize reward relies critically on the frontal lobes. In our study, we examined how groups of frontal cortex neurons in rat brains switch from encoding a familiar rule to a completely novel rule that could only be deduced through trial and error…they found that the same populations of neurons formed unique network states that corresponded to familiar and novel rules. Interestingly, although it took many trials for the animals to figure out the new rule, the recorded ensembles did not change gradually but instead exhibited a rather abrupt transition to a new pattern that corresponded directly to the shift in behavior, as if the network had experienced an “a-ha” moment…Taken together, these findings provide concrete support for sudden transitions between neural states rather than slow, gradual changes. “In the present problem solving context where the animal had to infer a new rule by accumulating evidence through trial and error, such sudden neural and behavioral transitions may correspond to moments of ‘sudden insight,'” concludes Dr. Durstewitz.

So the evidence is gathered and a new rule devised over time, but the new rule is put into use suddenly. I wonder how that is done: the rule is evaluated and modified elsewhere and then installed in the prefrontal network, or each neuron in that area of the prefrontal lobe is associated with other cells (not normally thought part of any network) and these cells accumulate the changes until the system reaches some sort of threshold and flips to the new configuration, or some other mechanism. My (uneducated) guess is that the results of trail-and-error are stored and computed elsewhere.