Computational models of memory

Computational models of memory are constrained from a number of directions. A physically accurate model of memory must simultaneously describe the details of behavioral performance as well as the mechanistic neural processes that support such performance. Such a model should simultaneously explain both changes in behavior due to manipulations of the neural substrate, such as lesions or pharmacological manipulations, as well as describing the detailed neurophysiology of the underlying structures, as measured by fMRI, EEG/ERP and the activity of single neurons. We are far from such a comprehensive view of the neural basis of the types of memory typically studied by cognitive psychologists. However, some progress has been made towards this ideal in several domains. This work can be divided into content areas, distinguished by the types of memory and brain region supporting them, or level of approach. This article describes the levels of approach that are possible in constructing computational models of memory and reviews progress in one content area—computational models of working memory.