Working memory shapes neural geometry in human EEG over learning

Abstract

Working memory has been traditionally studied as a passive storage for information. However, recent advances have suggested that working memory is prospective rather than retrospective, meaning that its content undergoes transformations that will support future behaviour. One perspective that underscores this notion conceptualises memory processes as a computational resource that can be used to reduce the complexity of computation at decision time. Here, we explore this perspective by examining whether the process of maintenance shapes neural geometry and leads to low-dimensional representations during storage and later decision time. We recorded EEG in 25 human participants who learnt to solve a XOR task. We hypothesised that separating task features by a working memory delay would result in participants temporally decomposing the XOR computation, by prospectively processing one of the task features early in trial time. In line with our predictions, participants transformed the first feature from a sensory to an abstract format and maintained this pre-processed information throughout the delay. This process was related to the low-dimensional representation required at decision time early in learning, a representation that has recently been shown to support later cross-generalisation. These results demonstrate that low-dimensional representations, elsewhere associated with slow learning, might also provide a mechanism for maintenance processes in working memory.