Learning to learn about uncertain feedback

Unexpected outcomes can reflect noise in the environment or a change in the current rules. We should ignore noise but shift strategy after rule changes. How we learn to do this is unclear, but one possibility is that it relies on learning to learn in uncertain environments. We propose that acquisition of latent task structure during learning to learn, even when not necessary, is crucial. We report results consistent with this hypothesis. Macaque monkeys acquired adaptive responses to feedback while learning to learn serial stimulus-response associations with probabilistic feedback. Monkeys learned well, decreasing their errors to criterion, but they also developed an apparently nonadaptive reactivity to unexpected stochastic feedback, even though that unexpected feedback never predicted problem switch. This surprising learning trajectory permitted the same monkeys, naïve to relearning about previously learned stimuli, to transfer to a task of stimulus-response remapping at immediately asymptotic levels. Our results suggest that learning new problems in a stochastic environment promotes the acquisition of performance rules from latent task structure, providing behavioral flexibility. Learning to learn in a probabilistic and volatile environment thus appears to induce latent learning that may be beneficial to flexible cognition.