Efficient Memory Encoding Explains the Interactions Between Hippocampus Size, Individual Experience, and Clinical Outcomes: A Computational Model

Abstract

The relationship between hippocampal volume and memory function has produced mixed results in neuroscience research, which suggests an additional influencing mechanism. To explore the role of an experience-dependent encoding mechanism, we developed an autoencoder model of the cortex-hippocampus loop and examined how its memory representation is affected by a penalty that prioritizes sparseness. We trained our model with the Fashion MNIST database and a loss function to modify synapses via backpropagation of mean squared recall error. The model exhibited experience-dependent efficient encoding, representing frequently-repeated objects with fewer neurons and smaller loss penalties; representations of similar sizes were found for objects repeated equally. Our findings clarify perplexing results from neurodevelopmental studies by linking increased hippocampal size and memory impairments in autism spectrum disorder (ASD) to decreased sparseness, and explaining dementia symptoms of forgetting with varied neuronal integrity. Our findings propose a novel model that connects observed relationships between hippocampal size and memory to the environmental demands and the underlying biological constraints, contributing to the development of a larger theory on experience-dependent encoding and storage and its failure. Author SummaryThe hippocampus is the brain region where memories are initially formed. A larger hippocampus seems to be associated with better memory performance, but this is not a static relationship; studies suggest that its size might grows with demand (e.g., the need to memorize more information) and that, in certain conditions, larger or smaller volumes of the hippocampus are not associated with better or worse memory. We demonstrate that it is possible to make sense of these findings by assuming that the hippocampus balances the need to correctly remember with the need to minimize resources used to store it, and that the hippocampal size might change as a function of memory demands (the characteristics of the information to memorize) and the underlying biology (e.g., the presence of hippocampal cell damage, or a reduction in hippocampal GABA receptors). We test this hypothesis by building a neural network model; the model correctly predicts existing, puzzling findings in the literature, and can even capture subtle interactions between different phenomena, such as predicting that individuals with Autism Spectrum Disorder would be more susceptible to memory loss in dementia.