Representational drift from a population view of memory consolidation

In the classical views of systems memory consolidation, memories are transferred from the hippocampus to cortex and between cortical areas for long-term storage [1]. But with the advance of neuronal recording techniques, it has become apparent that most variables of the outside world are learned and represented in multiple brain regions simultaneously. In some brain regions, these representations change dynamically while the behavior of the animal remains stable [2]. Here, we propose a population perspective on memory consolidation that reproduces many features of this “representational drift”. We develop an analytically tractable model that is inspired by the structure of insect mushroom bodies and an earlier model of working memory [3], in which memories are not transferred between brain regions but rather redistributed across or within regions. Memories remain stable during consolidation by confining synaptic plasticity to the null space of a memory readout. The resulting changes in population activity are in agreement with the systematic changes observed in vivo. Neuronal tuning curves show a diversity of changes, including stability, gradual drifts in the preferred stimulus, abrupt remapping, and a loss or gain of tuning. Decoders trained on a single session show limited generalization over time, while multi-day decoders reveal invariant subspaces in population activity. Finally, the model shows that a varying degree of plasticity across neurons can lead to a largely linear increase in the squared deviation of population activity from a reference day, although the changes in tuning are deterministic and not driven by diffusion. In summary, we present a model of systems memory consolidation which combines the classical view of memory consolidation as a memory transfer mechanism with the modern view of highly distributed memory representations.
[1] Roxin & Fusi 2013 [2] Driscoll et al. 2017 [3] Goldman 2009