Drifting assemblies for persistent memory: Neuron transitions and unsupervised compensation

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, associative 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 synapses and neural representations. The assemblies drift freely as noisy autonomous network activity and spontaneous synaptic turnover 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 temporal evolution of fear memory representations and suggest that memory systems need to be understood in their completeness as individual parts may constantly change.

Cortical cell assemblies and their underlying connectivity: anin silicostudy

“…It is possible that wiring efficiency can be optimized further, based on experience, through structural and functional plasticity. Recent modelling studies investigated how starting from a random initial connectivity, plasticity rules and network activity lead to assembly formation, maintenance and competition for member neurons (Fauth and Van Rossum, 2019;Kossio et al, 2021;Gastaldi et al, 2021). Conversely, our network model has strongly non-random connectivity, constrained by neuronal morphology (Reimann et al, 2015(Reimann et al, , 2017a, and can thus be viewed as a circuit in a non-random plastic state, but unshaped by experience.…”

“…Early theoretical work in the field explored the potential link between memories and cells that fire and therefore wire together, concentrating on the storage and retrieval of memories in strongly recurrent networks, such as the CA3 area of the hippocampus (Hopfield, 1982). Theories evolved and improved, but modeling studies about cell assemblies still concentrate on plasticity rules underlying the learning, storage and recall of various patterns (Krotov and Hopfield, 2016; Fauth and Van Rossum, 2019; Kossio et al, 2021; Gastaldi et al, 2021). Thus, their focus lies on how function shapes structure, with little or no emphasis on the biologically accurate aspects of structural connectivity, such as low connection probabilities and an abundance of directed motifs (Song et al, 2005; Perin et al, 2011; Reimann et al, 2017b).…”

Purely STDP-based assembly dynamics: Stability, learning, overlaps, drift and aging

“…Models of assembly maintenance, however, found that fast homeostatic plasticity was needed in addition to Hebbian learning. This introduces competition between synapses and prevents pathological growth of assemblies and exploding activity [5,[7][8][9][11][12][13]. Homeostatic plasticity has been observed in experiments, but it is much slower than Hebbian plasticity and does therefore not suffice to prevent runaway potentiation [14][15][16][17][18] (see, however, [19] for a different view and [10,20] for a small timescale implementation of homeostasis via inhibitory STDP).…”

“…neurons that are part of both assemblies [21][22][23]; the size of these overlaps appears to correspond to the strength of the associations between the concepts encoded by the assemblies. Previous models of assemblies stabilized by recurrent synaptic plasticity and fast homeostatic normalization usually do not show prominent overlaps [5,[7][8][9]12]. An example of a network with weight plasticity, structural plasticity, multiple synapses per connection and short-term depression that can store two strongly overlapping assemblies was given in [24].…”

Dendritic modulation enables multitask representation learning in hierarchical sensory processing pathways

“…A puzzling observation, first discovered in high-level areas 63,64 and later also in early sensory regions 65–67 , is that the participation of a neuron in the representation of a sensory stimulus changes over time. This representational drift raises questions about the framework of classical representation learning, and about how stable perception can be achieved 68,69 . As we show, changes to the sensory representation could help in extracting high level information in further processing layers.…”