Drifting Assemblies for Persistent Memory

Abstract

Change is ubiquitous in living beings. In particular, the connectome and neural representations can change. Nevertheless behaviors and memories often persist over long times. In a standard model, memories are represented by assemblies of strongly interconnected neurons. For faithful storage these assemblies are assumed to consist of the same neurons over time. Here we propose a contrasting memory model with complete temporal remodeling of assemblies, based on experimentally observed changes of connections and neural representations. The assemblies drift freely as spontaneous synaptic turnover or random activity induce neuron exchange. The gradual exchange allows activity dependent and homeostatic plasticity to conserve the representational structure and keep inputs, outputs and assemblies consistent. This leads to persistent memory. Our findings explain recent experimental results on the temporal evolution of fear memory representations and suggest that memory systems need to be understood in their completeness as individual parts may constantly change.

Representational drift in the mouse visual cortex

“…To generate consistent perception, the visual system must cope with changes in the coding of visual information. 76,77 It has been suggested that a system that carries a high-dimensional distributed code may maintain its functionality under representational drift by either confining the drift to the null space of the code, or via a compensatory plasticity of the downstream reader. 29,32 In both cases, the similarities across representations of different stimuli are expected to be somewhat conserved over time, even under a significant change in the representations themselves.…”

Memory out of context: Spacing effects and decontextualization in a computational model of the medial temporal lobe

“…One common puzzle arising within this literature on drift is how long-term memories survive amidst drift and what function drift might serve. It has been claimed (and shown via modeling) that long-term representations can continue to survive even with substantial drift (Kalle Kossio, Goedeke, Klos, & Memmesheimer, 2020; Mau et al, 2020). Using a recurrent neural network model, Clopath, Bonhoeffer, Hübener, and Rose (2017) showed that long-term stability can be maintained with a “backbone” of stable neurons and recurrent activity.…”

“…Using a recurrent neural network model, Clopath, Bonhoeffer, Hübener, and Rose (2017) showed that long-term stability can be maintained with a “backbone” of stable neurons and recurrent activity. Relatedly, while there may be drift from the perspective of some external input (e.g., the environment), internal representations between neural representations and their downstream readers may remain relatively coherent, which could help to compensate for drift in earlier representations (Kalle Kossio et al, 2020; Mau et al, 2020; Rule et al, 2019).…”

Computational Roles of Intrinsic Synaptic Dynamics

“…The cause and role of representational drift largely remain unknown. Some modeling works have proposed the framework in which extrinsic synaptic dynamics change neural codes to be consistent with the learning criterion and intrinsic dynamics add isotropic random changes, thereby the representation drifts in the "neutral" dimensions of the learning criterion [71,107]. For example, the network model of [71] was driven by both intrinsic dynamics and the gradient of expected reward.…”