Langfristige Speicherung von Gedächtnisinhalten in rekurrenten neuronalen Netzwerken durch das Zusammenspiel von synaptischer und struktureller Plastizität
Zusammenfassung der Projektergebnisse
Long-term memories are thought to be stored by cortical synapses. Yet, recent experiments show that an alarmingly high fraction of these synapses is removed and replaced on a daily basis. Hence, there must be mechanisms that maintain memories while their synaptic substrate is exchanged. In this project, we hypothesized that memories stored as Hebbian cell assemblies, i.e. strongly interconnected clusters of cells, naturally implement such mechanisms. We test this in a mathematical model comprising the two major activity-dependent processes that influence connectivity: synaptic plasticity, which changes the transmission efficacy or weight of the synapses, and structural plasticity, which creates and removes synapses. In this model, we explore two different possibilities to maintain cell assemblies: (1) Maintenance by constant slightly elevated activities within the assemblies and (2) maintenance by transient self-reactivations of the assemblies during phases when the network receives no external input. In the first case, we use the fact that synaptic plasticity induces a positive feedback between connectivity and activity, resulting in larger activities and weights in the strongly connected assemblies than in a weakly connected control (non-memory) population. As stronger synapses are typically less likely removed by structural plasticity, this leads to stabilisation of more synapses in the assemblies and decay of synapses between the control neurons. Such a bistable connectivity can be used to retain memory despite continous rewiring. Additionally, external input can be used to increase the activity in a population of cells, such that synapses within the population grow and become stable. This leads to the formation of an assembly in the strong connectivity state. Inhibitory input, in turn, destabilizes synapses and destroys the assemblies. Thus, assembly creation, removal and maintenance are realized by the same plasticity rules in an activity-dependent manner. During maintenance, synapse removal also prevents the slow outgrowth of assemblies observed in the absence of structural plasticity. Only at elevated pre- and post-synaptic activities the synapses stabilize, which maintains the connectivity within the assembly. These activities however also occur for synapses between two assemblies, such that morphological connectivity constraints are necessary to prevent this mechanism from merging memories together. Moreover, the assemblies maintained at low weights exhibit only weak pattern completion and show no significant difference in activation when presenting the learning stimulus again rendering them unsuitable as readily retrievable long-term memories. However, the strengthening of the connectivity upon a representation of the learning stimulus occurs much faster and is accomplished by building much less synapses. Thus, relearning is faster than learning, which is reminiscent of Ebbinghaus’ Savings effect. To obtain better retrievable memories, we investigated a second mechanism of assembly maintenance based on the phenomenon that strongly self-connected cell assemblies can spontaneously self-reactivate. Such self-reactivations induce high correlated activities within the assemblies, which are picked up by Hebbian synaptic plasticity. This leads to a maintentenance and sometimes even to a strengthening of the assembly connectivity during the convergence toward a long-term stable state. Also for this mechanism, there is no outgrowth or shrinkage of the assemblies and, even more importantly, no merging of individual assemblies within the network. Moreover, the maintenantance and strengthening of the assembly connectivity is robust against surpressing the self-reactivations for many hours, e.g., while processing incoming sensory information. We demonstrate that the assemblies maintained by this second mechanism can be reactivated by corrupted stimuli showing their potential for pattern completion. Hereby, the strengthening of connectivity also increases the robustness against corruption and hence the associative properties of the memory. Such strengthening of memories without further training can be related to the psychological phenomena of offline gains and sleep consolidation, which corresponds well to their occurence during resting phases of the network. In summary, we presented two possible mechanisms by which cell assemblies may be maintained despite synaptic turnover. Both can produce stable assemblies, which do not shrink or grow out and replace their synapses over time, and whose dynamics can be linked to different memory-related phenomena from psychology.
Projektbezogene Publikationen (Auswahl)
- ”Self-orgaized reactivations stable maintain and reinforce Hebbian cell assemblies despite synamptic turnover”, under review in eLife
Michael Fauth and Mark van Rossum
(Siehe online unter https://doi.org/10.7554/eLife.43717)
