Robust Input Disentanglement Through Dendritic Calcium-Mediated Action Potentials

Abstract

In daily life, living beings encounter a continuous stream of mixed information, which has to be disentangled by the brain to form proper representations. Using computational modeling, we demonstrate that the interplay between dendritic calcium-mediated action potentials (dCaAPs) with synaptic plasticity and rewiring can enable single neurons to successfully perform this complex task. Compared to other types of dendritic spikes, dCaAPs exhibit a high triggering threshold, large, but graded spike amplitude, with lower amplitudes for stronger synaptic inputs. We show that these properties enable neurons to successfully learn to represent discrete items from a continuous input stream by facilitating the clustering of synapses with temporally correlated presynaptic activities onto the same dendritic branch. In comparison to NMDA spikes, dendrites generating dCaAPs can form representations of individual items more efficiently, independent of the temporal order of their presentation during learning -- whether randomly, sequentially, as part of a random stream of simultaneously shown input items, or even as items with shared properties. Thus, our results provide further evidence about the critical role of dCaAPs for the computational capabilities of single neurons.