A great animal to study these questions in is the zebrafinch: Finches learn their songs from their fathers. The ethology of song-learning is richly documented, and now there is beautiful neural data from premotor song nuclei, taken as the bird sings.

Problems ripe for study include the learning of sparse, robust neural sequences; motor learning and refinement of premotor maps; and understanding why the peculiar premotor codes found in HVC and RA may be helpful for the representation and learning of song.

Song learning in a network of spiking neurons


The birdsong motor circuit is a hierarchical structure: nucleus HVC projects to premotor nucleus RA, which in turn drives motor neurons. Recent experiments show that RA-projecting HVC neurons have temporally sparse neural sequences that drive activity in RA. The role of RA appears to be the conversion of abstract neural sequences in HVC into motor activity.

Our work aims to show how an appropriate map of HVC to motor activity could be learned via plastic connections between HVC and RA. Such learning is commonly thought to be driven by reinforcement (Doya & Sejnowski 1995, Troyer & Doupe 2000), with a reward signal generated by comparing the bird's vocal output with an internally stored copy (template) of its tutor song.

We have constructed a reinforcement model with spiking neurons that learns HVC-to-RA connections in a feedforward network of HVC, RA, and a motor layer. We assume that HVC provides a sparse sequence, and learning is governed by a synaptic plasticity rule that exploits correlations between fluctuations in the motor output due to noisy neural inputs, and a positive scalar global reward that depends on the match between network output and the stored template. We explore motor fluctuations arising from the inherent stochasticity of HVC-to-RA synapses, or from (possibly LMAN-generated) noise injected into RA. The learning rule performs stochastic gradient ascent on the reward, and is robust over a wide range of parameters.

Temporal sparseness of the premotor drive and learning speed


Recent experiments (Hahnloser et al, 2002) reveal that HVC neurons display temporally sparse sequential activity throughout singing, where individual RA-projecting HVC neurons fire just once per song motif. What is the role of such exteremely sparse temporal coding in HVC? By simulating the song learning process in a feedforward network of HVC, RA, and motor layers, and through a linear analysis of the network, we show that the time required to learn the correct motor outputs, given a sequential HVC drive, depends strongly on the temporal sparseness of the premotor (HVC) inputs. Sparse HVC inputs result in fast learning by decoupling the premotor drive for different parts of the song. We find that the minimum time required to learn a song is proportional to the number of times each HVC neuron is active during a song motif, and would double if each HVC neuron were to burst twice per motif. This work points to the potential importance of sparse coding for speed of learning in general motor-related learning tasks.