Abstract
The formation of associations between memory items, that enables recall of one memory component by activating another, is a fundamental operation of higher brain function. Recent neural recordings provide insight into the way how such associations are encoded on the level of neurons in the human medial temporal lobe (MTL). We show that important features of these experimental data can be reproduced by a generic neural circuit model consisting of excitatory and inhibitory spiking neurons with data-based short- and long-term synaptic plasticity. A key result of the experimental data and the model is that the association process causes the emergence of overlaps between the assemblies of neurons that encode the memory components. These overlaps appear in the experiments and the model at the same time when the association becomes computationally functional. Hence our model elucidates computational and plasticity processes that are likely to shape memory systems in the brain.
Significance statement One commonly assumes that memory items are encoded by sparsely distributed groups of neurons, often referred to as assemblies, that fire whenever a memory item is activated. An important question is how combinations of several memory items are encoded. Recent experimental data suggest that the assemblies for memory items expand during an association process, so that overlaps of the assemblies emerge. This result is surprising from the perspective of neural network models, where one commonly assumed that assemblies for memory items remain largely invariant. We show that a simple model for recurrent neural circuits with data-based forms of synaptic plasticity reproduces the new experimental data, and thereby provides the basis for more data-based neural network models for memory associations.
