Prediction error dependent changes in brain connectivity during associative learning

One of the fundaments of associative learning theories is that surprising events drive learning by signalling the need to update one’s beliefs. It has long been suggested that plasticity of connection strengths between neurons underlies the learning of predictive associations: Neural units encoding associated entities change their connectivity to encode the learned associative strength. Surprisingly, previous imaging studies have focused on correlations between regional brain activity and variables of learning models, but neglected how these variables changes in interregional connectivity. Dynamic Causal Models (DCMs) of neuronal populations and their effective connectivity form a novel technique to investigate such learning dependent changes in connection strengths. In the work presented here, I embedded computational learning models into DCMs to investigate how computational processes are reflected by changes in connectivity. These novel models were then used to explain fMRI data from three associative learning studies. The first study integrated a Rescorla-Wagner model into a DCM using an incidental learning paradigm where auditory cues predicted the presence/absence of visual stimuli. Results showed that even for behaviourally irrelevant probabilistic associations, prediction errors drove the consolidation of connection strengths between the auditory and visual areas. In the second study I combined a Bayesian observer model and a nonlinear DCM, using an fMRI paradigm where auditory cues differentially predicted visual stimuli, to investigate how predictions about sensory stimuli influence motor responses. Here, the degree of striatal prediction error activity controlled the plasticity of visuo-motor connections. In a third study, I used a nonlinear DCM and data from a fear learning study to demonstrate that prediction error activity in the amygdala exerts a modulatory influence on visuo-striatal connections. Though postulated by many models and theories about learning, to our knowledge the work presented in this thesis constitutes the first direct report that prediction errors can modulate connection strength