(A) Top: the early phase of song learning is characterized by high vocal variability (left panel). This is associated with relatively larger number of weaker inputs from HVC to RA neurons (middle panel). A simple model of this circuit organization (‘Materials and methods’), in which RA neurons integrate input from precise HVC time-keeper neurons and variable LMAN neurons, produces variable firing patterns in RA (right panel). Bottom: song learning is associated with a strengthening and pruning of HVC-RA connections and a concomitant decrease in song variability. Pruning the relative number of active HVC-RA inputs in our model while strengthening the remaining ones (middle panel) dramatically decreases the variability in song-related RA firing (right panel). The simulations of RA spike trains were done using model parameters consistent with our experimental data from age groups 2 (top, plastic song) and 3 (bottom, crystallized song), respectively. Note that the simulations for the two neurons are independent (i.e., the ‘older’ is not derived from strengthening and pruning connections of the ‘younger’). See ‘Materials and methods’ for further details. (B) The average cross-correlation (CC) between spike trains during different ‘song’ renditions of the simulated RA neuron increases with the degree of strengthening and pruning of HVC-RA synapses. The open circles correspond to the age groups 2 and 3 in our data set. Solid black line represents simulations using ‘standard’ model parameters chosen to conform to our experimental data (see ‘Results’, Figures 3–5, and ‘Materials and methods’). Dashed lines show simulations where inputs from HVC were either only strengthened (red) or pruned (blue), relative to the ‘standard’ model at age group 2. The x-axis (bottom labels) shows the fraction of active HVC inputs to RA (‘Materials and methods’), as well as the average strength of HVC inputs (top labels). (C–E) Effects on variability in RA firing from: (C) strengthening/weakening LMAN input by up to 50%, (D) changing NMDA:AMPA ratio at the LMAN-RA synapse, and (E) changing the gain (F-I relationship) of the RA neuron. All changes are relative to the ‘standard’ model (black lines) in ‘B’. (F–H) Effects on variability in RA firing stemming from changes to LMAN firing patterns. Changes are relative to the ‘standard’ model (black lines). (F) Different LMAN spike trains tested in our model (for 50 ‘song renditions’). Top: Poisson spike train. Middle: Poisson spike train with 30% of the spikes in the form of Poisson bursts (see ‘Materials and methods’). Bottom: time-varying Poisson firing. Blue curve shows the average instantaneous firing rate as a function of time in song with 50% modulation compared to the baseline rate. (G) Effect of altering the burstiness of LMAN neurons. Here, Poisson burst were added to the normal Poisson spike train. (H) Effect of increasing the song-locking of LMAN firing.