Representational drift as a correlate of memory consolidation
Denis Alevi*+1,2, Felix Lundt+1, Simone Ciceri1, Kristine Heiney1, Henning Sprekeler1,2,3
1Modelling of Cognitive Processes, Technische Universität Berlin, Berlin, Germany 2Bernstein Center for Computational Neuroscience Berlin, Berlin, Germany 3Science of Intelligence, Research Cluster of Excellence, Berlin, Germany +Equal contribution *Email: denis.alevi@tu-berlin.de
Introduction
Neural representations – and their population geometry – often change over time despite stable behavior, a phenomenon termed representational drift [1-4]. It is debated if drift is driven by a random process or if it has a directed component, and if it serves a computational function [5]. Systems memory consolidation is a promising candidate [6], because it predicts a temporal reorganization of neural memory engrams. However, it remains unclear how classical theories of consolidation relate to the population-level view of drift and how apparently unstructured drift could be driven by a directed consolidation process. Methods We present a computational model for engram dynamics under memory consolidation and explore the resulting representational drift. Assuming that engram changes are driven by reactivations, the model displays recurrent neural network (RNN)-like dynamics, but evolves on long time scales of memory consolidation. This allows us to reinterpret common dynamical phenomena in RNNs in light of memory consolidation and relate them to experimentally observed drift. In simulation, we study how single cell tuning curves and the geometry of neural representations change over time, when not all neurons are observed and develop analytical results for the effect of subsampling, based on Green’s functions and random matrix theory. Results Our model redistributes memory engrams across neural populations while maintaining stable memory recall through null-space dynamics [2]. The model can display power-law forgetting without requiring a diversity of learning rates [7]. Low-rank dynamics induce selective consolidation and semantization. In line with experimental findings on representational drift, individual neurons exhibit diverse tuning changes: stability, gradual drift, and abrupt changes of preferred stimulus. Multi-day decoders [2] reveal invariant subspaces on the full population, but degrade quickly under subsampling. A theoretical analysis shows that the dynamics of subsampled populations can be predominantly driven by the unrecorded population, which generates seemingly noise-driven dynamics.
Discussion Our phenomenological model of engram dynamics bridges the gap between the area-centered perspective of systems consolidation and the population-level perspective of representational drift. Our results show that despite systematic population dynamics, a recorded subset of the neural population can appear to have unstructured dynamics [2]. Recent evidence for stable geometric structure during representational drift in CA1 [7] is consistent with our model of RNN-like engram dynamics, and we hypothesize that unstable population geometry [3] could also be explained by subsampling. Overall, our model offers a functional interpretation of drift as a means to redistribute engrams for improved memory retention.
Acknowledgements Kristine Heiney is funded by a Postdoctoral Research Fellowship from the Alexander von Humboldt Foundation. References [1]https://doi.org/10.1016/j.cell.2017.07.021 [2]https://doi.org/10.7554/eLife.51121 [3]https://doi.org/10.1038/s41586-021-03628-7 [4]https://doi.org/10.1007/s00422-021-00916-3 [5]https://doi.org/10.1016/j.conb.2022.102609 [6]https://doi.org/10.1371/journal.pcbi.1003146 [7]https://doi.org/10.1101/2025.02.04.636428
