A Bayesian model for a Pavlovian-instrumental transfer hypothesis

A Pavlovian conditioned stimulus (CS) associated with a reward can enhance an instrumental response directed to the same or other rewards. This effect is called Pavlovian-instrumental transfer (PIT). In recent years, lesion studies using rats have gained insight into its neural substrates dissociating between specific PIT (where CS and instrumental response share the same reward) and general PIT (where they do not) (Corbit and Balleine, 2005, 2011). Despite these advances, the functional differences between specific and general PIT and how Pavlovian cues interact with instrumental response are still not clear. Here we try to explain Pavlovian-instrumental transfer effects by using a latent causes Bayesian model. Previous work in the Pavlovian conditioning literature (Courville et al., 2005) suggests that during Pavlovian conditioning rats do not simply learn associations between two events (CS and reward); instead, they actually try to figure out the real hidden causes behind them by constructing a latent cause model. We expanded that view to include instrumental actions and so explain the interactions between Pavlovian conditioning and instrumental conditioning. Our model correctly reproduces both the presence of specific and general PIT and the absence of general PIT when the CS is associated to the reward of another instrumental action. By framing the PIT effects explanation in Bayesian terms, our model offers a new integrated view on their functional mechanisms and new testable predictions.