prediction and inference across scales of granularity

Abstract

The utility of our distributed memory system is two-fold. First, it is to provide robust inferences as to the identity and affordances of stimuli in our environment. These inform our predictions as to what may happen next so we can determine what we ought to do about it. Here, I cover three studies that probe the structure of neural representations and the computations performed on them that give rise to these functions. First, I study how inference processes shape the retroactive effects of arousal on memory for related items, such that the specificity with which participants learn a contingency structure determines the specificity of the item relations they reactivate and later retain. As the neural gradients by which representational specificity varies is hotly contested, I then apply cutting-edge machine learning techniques to a large resting-state fMRI sample to reveal novel distinctions of representational specificity just within the anterior hippocampus, reconciling disparate findings and laying the groundwork for future study. Lastly, I show how dopaminergic signals reverse the effect of cortical replay on memory reconsolidation, such that labile traces that are replayed under conditions of low VTA connectivity are better stabilized and left with fewer errors, while traces that are replayed under conditions of high VTA connectivity are more strongly updated and left with more errors. Future work applying these concepts to the generalization of contingency structures across sensory modalities is discussed.