The Role of Moderate Memory Activation During Learning and Post-learning Sleep in Driving Representational Change

Abstract

When attempting to retrieve a target memory, related memories often come to mind. This can be beneficial for when we want to append new information to a set of related memories (i.e. integration), but could also be disadvantageous when related, non-target memories make it hard to access a target memory (i.e. interference). One way the brain can decrease interference at retrieval is to decrease neural overlap between related memory representations, a process known as differentiation. The Non-Monotonic Plasticity Hypothesis (NMPH) predicts that differentiation is achieved when one memory is strongly activated and an overlapping memory is moderately activated, leading to a weakened connection between these two memories. Rapid eye movement (REM) sleep, in particular, has also been shown to help resolve interference. We conducted an fMRI experiment (designed by Dr. E McDevitt and Dr. K Norman; primary data collection by Dr. E McDevitt) to test the hypothesis that REM sleep drives learning-related representational change–specifically, neural differentiation–in the hippocampus. We used a statistical learning task previously shown to induce differentiation when a predicted item is moderately activated due to its failure to appear in perception (violation events). We additionally manipulated whether participants had sleep or no sleep, and if their sleep contained REM or non-REM only, between the learning experience and the final measurement of representational change. In addition to examining how these experimental conditions (task-related violation/non-violation conditions and offline sleep condition) impact brain activity (the “traditional” analysis approach), we analyzed the data in the opposite direction using a “decoding” approach in which we aimed to decode experimental conditions from brain activity using machine learning classifiers. The results of our traditional analyses (originally done by Dr. E McDevitt) revealed the REM group to be the only group with significant violation-related differentiation in the right CA2/3/DG subfield of the hippocampus. Moreover, the decoding analyses revealed that our learning and sleep manipulations indeed led to discriminable changes in brain activity in various brain regions (hippocampus, medial prefrontal cortex, occipital-temporal cortex), which allowed classifiers to successfully decode sleep and task conditions from patterns of post-learning and post-sleep brain activity.