A passive computational marker for individual differences in non-reinforced learning

Abstract

AbstractAlthough research about preference formation and modification has classically focused on the role of external reinforcements, there is also increasing evidence for a key role of non-externally reinforced cognitive mechanisms such as attention and memory in preference modification. In a novel paradigm for behavioral change called the Cue-Approach training (CAT) task, preferences are modified via the mere association of images of stimuli with a neutral cue and a rapid motor response, without external reinforcements. The procedure’s efficacy has been replicated across dozens of studies, and the net behavioral change was linked with increased activity in a frontal value-based decision-making brain region during the post-training probe choice phase. However, the cognitive mechanisms during the training phase itself have not been elucidated. Based on the structure of the task alongside recent findings of the involvement of striatal and supplementary motor regions during training, we hypothesized that a motor-related learning process could be a prospective candidate. To test this hypothesis, we developed a computational model of the motor response pattern during training in a large corpus of data collected from 864 participants across 29 different CAT experiments. Using Bayesian modeling of the meta-analysis data, we developed a computational marker for individualized learning in the training task, which was found to be associated with the preference modification effect in the subsequent probe task, both at the participant-level as well as in the more granular individual-item level. Following the conclusions of the meta-analysis, in two additional experiments (a pilot study and a larger preregistered replication study) we aimed to affect learning efficacy by manipulating the training procedure difficulty. As hypothesized and preregistered, training difficulty was captured by the new computational marker identified on the previously collected samples. Manipulation of the training difficulty also resulted in a differential preference modification effect, suggesting a causal relationship between the motor learning captured by the computational model and the post-training behavioral change effect. Our work highlights a novel non-reinforced preference modification pathway, suggesting that attention and motor learning are linked to preference formation, and provides a computational framework to identify individualized training markers which could predict future behavioral change effects.