A homeostatic mechanism or statistics can maintain input-output relations of multilayer drifting assemblies

Associative memories might be represented by groups of strongly interconnected neurons, called neuronal assemblies. Recent experimental findings suggest that such assemblies are not static but composed of different neurons at different times [1]. A theoretical model reproducing these findings is given by drifting assemblies [2]: In a gradual process, assemblies exchange individual neurons between one another due to ever-present synaptic plasticity. So far, only single-layer networks of drifting assemblies have been considered. Biological neuronal assemblies are, however, thought to be distributed over several brain regions. How networks of drifting assemblies generalize to multiple layers is unknown.

Here we propose a model for multilayered networks of drifting assemblies. We introduce a novel form of homeostatic plasticity which we refer to as distributed homeostatic normalization. It promotes between-layer connectivity by separately normalizing intra- and interlayer weights. We show that distributed homeostatic normalization is capable of ensuring an even distribution of assemblies over two layers. In contrast to homeostatic plasticity mechanisms of previous models, distributed homeostatic normalization can act on biologically plausible timescales [3] and still have the desired impact on the network structure. For large assemblies we find that statistics on its own also leads to a sufficient distribution of the assemblies over multiple layers.

Our model demonstrates how continuous pathways from input neurons over multiple layers to output neurons can be established and maintained in the presence of representational drift. The faithfulness of such input-output relations is essential for the conservation of memory and behavior.