Gast ::
| Hauptseite > Publikationsdatenbank > Connectivity patterns that enable working memory in recurrent networks with dendritic subunits |
| Hauptseite > Publikationsdatenbank > Connectivity patterns that enable working memory in recurrent networks with dendritic subunits |
| Conference Presentation (After Call) | FZJ-2023-00976 |
; ; ;
2022
Abstract: Maintain a working memory of observed features is a crucial neuronal computation. In spiking networks, this computation depends upon the existence of signals within neurons that persist after spike generation. Commonly, these signals are thought to be plasticity-related. Here, we hypothesize that the dendritic subunit forms a complementary substrate for working memory. Dendritic subunits support the generation of N-Methyl-D-Aspartate driven depolarizations (NMDA-spikes), which have a time-scale that outlasts the membrane time-scale by a factor of five. Moreover, voltage dynamics within dendritic subunits are unperturbed by somatic spikes, thus allowing temporal information to persist within neurons.To investigate the capacity of dendritic subunits to retain stimulus information, we reduce a layer 5 pyramidal cell to a prototypical model with dendritic subunits, representing distal basal compartments, and recurrently connect these excitatory (E) neurons to each other's dendrites. Inhibitory (I) interneurons provide global inhibition to balance the network. As NMDA-spikes are generated by coincident inputs to a dendritic subunit, we find that introducing modularity in the E-to-E connections, and increasing the degree of intraneural clustering -- such that neurons from the same cluster preferentially target the same dendritic subunit -- elicits NMDA-spikes in the network. Interestingly, NMDA-spikes also emerge at a fixed modularity when intra-neural clustering is increased, showing that this quantity -- often ignored in connectivity studies -- is crucial in determining the emergent network state. Finally, we quantify memory retention in the network by computing the capacity of linear readout units to reconstruct past inputs based on the current network state. We find that network states with NMDA-spikes can retain past inputs for up to $\sim$200 ms longer than network states without them. Thus, our results show that spiking networks can maintain a working memory even without synaptic plasticity, through the activity of dendritic subunits.
|
The record appears in these collections: |