An associative learning account of sensorimotor adaptation

ABSTRACTTwo influential paradigms, sensorimotor adaptation and eyeblink conditioning, have deepened our understanding of the theoretical and neural foundations of motor learning, and in particular, the role of the cerebellum. Although there has been some cross-pollination between these two lines of research, they typically operate within distinct theoretical frameworks, with the incremental updating of an internal forward model explaining adaptation, and associative learning processes explaining eyeblink conditioning. Here we ask if a unified framework might be parsimonious, directly linking sensorimotor adaptation to associative learning. Using a task that isolates implicit sensorimotor adaptation, we paired movement-related feedback with neutral auditory or visual cues that served as conditioning stimuli (CSs) to test two key signatures of associative learning-differential conditioning and compound conditioning. We observed clear Pavlovian effects in both cases: Implicit trial-by-trial changes in movement kinematics were reliably modulated by the CSs in the predicted directions. Moreover, after compound conditioning, we observed a robust negative correlation between the responses of individuals to the two elemental CSs of the compound, consistent with the additivity principle posited by the Rescorla-Wagner model of classical conditioning. Computational modelling demonstrates that these results cannot be captured by the conventional algorithm used to explain the operation of a forward model. We believe that associative learning effects in implicit sensorimotor adaptation provide a proof-of-concept for linking multiple motor learning paradigms within a similar theoretical framework.

Do rats learn conditional independence?

If acquired associations are to accurately represent real relevance relations, there is motivation for the hypothesis that learning will, in some circumstances, be more appropriately modelled, not as direct dependence, but as conditional independence. In a serial compound conditioning experiment, two groups of rats were presented with a conditioned stimulus (CS1) that imperfectly (50%) predicted food, and was itself imperfectly predicted by a CS2. Groups differed in the proportion of CS2 presentations that were ultimately followed by food (25% versus 75%). Thus, the information presented regarding the relevance of CS2 to food was ambiguous between direct dependence and conditional independence (given CS1). If rats learnt that food was conditionally independent of CS2, given CS1, subjects of both groups should thereafter respond similarly to CS2 alone. Contrary to the conditionality hypothesis, subjects attended to the direct food predictability of CS2, suggesting that rats treat even distal stimuli in a CS sequence as immediately relevant to food, not conditional on an intermediate stimulus. These results urge caution in representing indirect associations as conditional associations, accentuate the theoretical weight of the Markov condition in graphical models, and challenge theories to articulate the conditions under which animals are expected to learn conditional associations, if ever.