Dissociation of task acquisition from expression during learning reveals latent knowledge
I will present a new behavioral framework to interpret neural recordings during learning. We trained mice, rats and ferrets on tasks such as discriminating between two tones with a lever press or a lick; animals were rewarded for correct actions and punished for incorrect ones. During learning, we interleaved reinforced trials with “probe” trials in which we omitted reinforcement. Across different types of tasks, all animals performed remarkably better in probe trials during learning. Reinforcement feedback, although critical for learning, paradoxically, may hide the expression of underlying knowledge. Probing behavioral performance during learning in the absence of reinforcement is therefore a powerful behavioral tool that now allows us to distinguish between what an animal knows (“acquisition”, probe trials) and how it demonstrates that knowledge (“expression”, reinforced trials). We constructed a computational model in which learning of action values takes place only during reinforced trials, while the changes between contexts (reinforced and probe trials) do not change the learned values of different options, but modulate only the read-out parameters to take into account factors such as impulsivity or exploration. This mechanism successfully led to a difference at the level of behavioral expression between contexts, without any change of the underlying action values which represent task acquisition. Importantly, selectively scaling the gain of feed-forward inhibition provided the best fit of behavioral data with a small number of adjusted parameters for mice, rats, and ferrets. Preliminary neural data acquired using large-scale two-photon imaging and optogenetics suggests that auditory cortical networks and cholinergic projections may be involved in both task acquisition and expression.