Zebra finch song is driven by a forebrain clock. In nucleus HVC, each neuron that projects to RA fires a single brief burst at one precise moment of the song motif and is silent the rest of the time. This page rebuilds that ultra-sparse code from the numbers and equations in Hahnloser, Kozhevnikov & Fee (2002).
What is from the paper and what is assumed. Burst statistics, timing jitter, sleep firing rates, correlation values and every equation come from the paper’s text and Methods, and are checked against it below. The song is synthetic (no recording could be used here). Interneuron rates during song, the HVC→RA weights and the synaptic kernel are illustrative; sections that use them carry an illustration tag.
Each row is one rendition of the motif and each block is one neuron. HVC(RA) neurons fire a brief burst at a single moment, and do so on every rendition with sub-millisecond jitter. Interneurons fire throughout the song. Press play to hear the motif and follow the cursor.
The generator is set from the paper’s means and s.d.; the bursts are then detected with the paper’s R(t) > 100 Hz rule and measured. With only ten neurons, sample means vary from draw to draw; press New neurons to see the spread.
Different HVC(RA) neurons burst at different times, “with no obvious timing relation to the onset or offset of song syllables”. Ordered by burst time, the population tiles the motif. At every moment a different small group is active: a code for time in the song, not for syllables.
The paper proposes that the canonical sequence in HVC “is translated into a complex sequence in RA by a set of synaptic weights”. Here each RA neuron sums delayed, weighted HVC(RA) bursts. Click the RA trace to correct an “error” at that moment by weakening the synapses active just before it. With the sparse code only that moment changes. With a dense code, where every HVC neuron bursts many times, the same correction spreads across the song.
In the head-fixed sleeping bird, HVC(RA) neurons still burst sparsely and RA neurons fire bursts locked to them, before and after. The paper quantifies this with a conditional correlation K(τ) and a shuffle test, both implemented here exactly as in its Methods. Lidocaine in HVC abolished RA sleep bursts, leaving RA’s 10–20 Hz regular spiking.
“The motor actions underlying song unfold when the ‘time code’ in HVC acts on the ‘muscle map’ in RA.” Here are all the pieces running together: the song, the HVC clock, and RA reading it out.
Each neuron bursts once per motif (6.1 ms, 4.5 spikes, 0.66 ms jitter). Together they mark every moment of the song.
See the clock ↑Which RA neurons fire at each moment depends on the synapses from the HVC neurons active then.
See the readout ↑RA neurons burst about 13 times as often as HVC(RA) neurons and project to vocal and respiratory motor neurons.
See RA ↑RA bursts follow interneuron bursts by about 4.5 ms and vanish when HVC is silenced. About 70% of RA sleep bursts are HVC-driven.
See sleep ↑Θ is the Heaviside step. τ > 0 means the RA neuron follows the HVC(RA) neuron. c̄ is the mean of c(τ) over shuffled trains.
| Quantity | Value | State |
|---|---|---|
| HVC(RA) bursts per motif | ≤ 1 | singing |
| HVC(RA) burst duration | 6.1 ± 2 ms | singing |
| Spikes per burst | 4.5 ± 2 | singing |
| Intraburst rate | 613 ± 210 s⁻¹ | singing |
| Burst jitter (r.m.s.) | 0.66 ± 0.14 ms | singing |
| HVC(X) bursts per motif | 0–5 | singing |
| Interneuron spontaneous rate | 11 ± 7 s⁻¹ | awake, not singing |
| HVC(RA) burst rate | 0.06 ± 0.05 s⁻¹ | sleep |
| RA : HVC(RA) burst ratio | 13 ± 3 | sleep |
| HVC(RA) sleep burst | 6.5 ± 1.8 ms · 3.2 ± 0.8 spikes · 347 ± 81 s⁻¹ | sleep |
| RA sleep burst duration | 11.5 ± 3.5 ms | sleep |
| RA regular spiking | 10–20 Hz | sleep |
| Motif length | about 1 s | (synthetic: 0.9 s) |