Burst-dependent plasticity and dendritic amplification support target-based learning and hierarchical imitation learning

Most biological models are composed of single-compartment neurons.  Despite recent findings on dendritic computational properties, these features have been only rarely exploited. We introduce a multi-compartment model for pyramidal neurons, in which bursts and dendritic input segregation allow for biologically plausible target-based learning. The solution of a problem is represented as a spatio-temporal pattern of bursts, suggested to apical tufts of neurons. A store-and-recall and a nontrivial navigation tasks are learned, the latter by introducing a hierarchical architecture of layers enabling the decomposition into simpler subtasks (hierarchical imitation learning).