Neural network modeling of memory gradient in Alzheimer's disease

Abstract

Several studies have documented a temporal gradient in the memory of persons with Alzheimer's disease: patients are better able to recall more distant memories. The significance of this gradient is unclear: does the disease selectively interfere with the recall of recent memories, or does it prevent the memories from being adequately recorded? To address this question, neural networks were used to simulate learning over time. Once trained with a group of patterns, the networks were damaged to simulate the lesions associated with Alzheimer's disease. By altering the number of times a network was trained with a given pattern before additional patterns were added, and by varying the number of patterns in the training set, the direction of the temporal gradient was changed. The factors that determine the direction of the gradient are in place before the network is damaged. This suggests that the gradient associated with Alzheimer's disease is not a direct result of brain lesions that are hallmarks of the disease, but instead develops from an alteration of the learning process that begins long before dementia develops.