Predictive, interactive multiple memory systems.

Most lesion studies in animals, and neuropsychological and functional neuroimaging studies in humans, have focused on finding dissociations between the functions of different brain regions, for example in relation to different types of memory. While some of these dissociations can be questioned, particularly in the case of neuroimaging data, we start by assuming a "modal model" in which at least three different memory systems are distinguished: an episodic system (which stores associations between items and spatial/temporal contexts, and which is supported primarily by the hippocampus); a semantic system (which extracts combinations of perceptual features that define items, and which is supported primarily by anterior temporal cortex); and modality-specific perceptual systems (which represent the sensory features extracted from a stimulus, and which are supported by higher sensory cortices). In most situations however, behavior is determined by interactions between these systems. These interactions reflect the flow of information in both "forward" and "backward" directions between memory systems, where backward connections transmit predictions about the current item/features based on the current context/item. Importantly, it is the resulting "prediction error"--the difference between these predictions and the forward transmission of sensory evidence--that drives memory encoding and retrieval. We describe how this "predictive interactive multiple memory systems" (PIMMS) framework can be applied to human neuroimaging data acquired during encoding or retrieval phases of the recognition memory paradigm. Our novel emphasis is thus on associations rather than dissociations between activity measured in key brain regions; in particular, we propose that measuring the functional coupling between brain regions will help understand how these memory systems interact to guide behavior.