The amazing imitation capabilities of songbirds show that they can memorize sensory sequences and transform them into motor activities which in turn generate the original sound sequences. This suggests that the bird's brain can learn (1) to reliably reproduce spatio-temporal sensory representations and (2) to transform them into corresponding spatio-temporal motor activations by using an inverse mapping. Neither the synaptic mechanisms nor the network architecture enabling these two fundamental aspects of imitation learning are known. We propose an architecture of coupled neuronal modules that mimick areas in the song bird and show that a unique synaptic plasticity mechanism can serve to learn both, sensory sequences in a recurrent neuronal network, as well as an inverse model that transforms the sensory memories into the corresponding motor activations. The proposed membrane potential dependent learning rule together with the architecture that includes basic features of the bird's brain represents the first comprehensive account of bird imitation learning based on spiking neurons.
展开▼
