4/15/2004 12:53 PM

1

Neural mechanisms of vocal sequence generation in the songbird

Michale S. Fee, Alexay A. Kozhevnikov, Richard H. R. Hahnloser*

McGovern Institute for Brain Research, Department of Brain and Cognitive Sciences,

Massachusetts Institute of Technology, Cambridge, MA 02139

* Present affiliation: Institute for Neuroinformatics, UNIZH/ETHZ, Zurich, Switzerland

Address for correspondence: Michale S. Fee McGovern Institute for Brain Research Massachusetts Institute of Technology Cambridge, MA 02139 fee@mit.edu

In press, Annals of the New York Academy of Sciences, special issue on birdsong

4/15/2004 12:53 PM

2

Abstract:

Little is known about the biophysical and neuronal circuit mechanisms underlying the

generation and learning of behavioral sequences. Songbirds provide a marvelous animal

model in which to study these phenomena. Using a motorized microdrive to record the

activity of single neurons in the singing bird, we are beginning to understand the circuits

that generate complex vocal sequences. We describe recent experiments elucidating the

role of premotor song-control nucleus HVC in the production of song. We find that HVC

neurons projecting to premotor nucleus RA each generate a single burst of spikes at a

particular time in the song, and may form a sparse representation of temporal order. We

incorporate this observation into a working hypothesis for the generation of vocal

sequences in the songbird, and examine some implications for song learning.

Temporal sequences in the organization of birdsong

Temporal structure plays a fundamental role in all aspects of brain function – not only

in sensory systems 1-4 – but at a motor and cognitive level as well 5,6. For example, when

we learn the alphabet, we do not learn twenty-six uncorrelated symbols, we really learn a

cognitive and motor sequence. Most of us can say the letters of the alphabet in just a few

seconds, but only if we say them in one particular order – the order in which we learned

them. This ability of the brain to step rapidly through a learned sequence of states underlies

not only our ability to say the alphabet, but also our ability to speak, to perform music and

athletic feats, and perhaps our ability to think and plan as well. Despite the fundamental

importance of temporal order to brain function, little is known about the biophysical and

circuit mechanisms that underlie the generation and learning of temporal sequences.

The songbird is a remarkable model system in which to study learned sequence

generation. Songbirds, such as the zebra finch, generate complex learned vocal sequences

to convey identity and attract mates [see Williams, this volume]. The zebra finch sings a

repeated song element called a song motif. The motif, typically of 0.5 to 1sec duration, is

composed of 3-7 smaller vocal gestures called song syllables 7,8. One useful property of

zebra finch song, particularly for the study of a motor system, is that the motifs are highly

4/15/2004 12:53 PM

3

stereotyped. That is, each time the zebra finch sings its song motif, the pattern of acoustic

signals may be nearly identical.

The generation of these vocal sequences is under the control of a small number of

premotor brain areas. Muscles of the vocal organ, or syrinx, are innervated by

motorneurons of the hypoglossal nucleus (nXIIts) 9,10. The motor neurons receive synaptic

inputs from a forebrain nucleus RA, which in turn receives a predominant premotor input

from another forebrain nucleus HVC 11-13 [see chapters by Suthers and Wild, this volume].

In this chapter, we describe some recent results on the mechanism by which HVC and RA

generate the complex sequences of motor commands that produce learned vocal behavior

in the songbird.

During singing, RA neurons generate a complex sequence of high-frequency bursts of

spikes, the pattern of which is precisely reproduced each time the bird sings a song motif 14.

By recording large numbers of RA neurons in the singing bird, we have recently found that

each RA neuron, on average, produces roughly 12 bursts of ~10ms duration. In addition,

each RA neuron has a fairly unique pattern of bursts. Thus, as a population, RA neurons are

active throughout the song vocalization, and at each time in the song roughly 12% of the

RA neuron population is active (Fig. 1A).

How are the complex burst sequences of RA neurons generated? Several recent

models of vocal sequence generation have suggested that short-timescale temporal

patterning is produced by circuitry within the premotor nucleus RA itself 14-17 (Fig. 1B). In

this view, each burst in an RA neuron is perhaps driven by bursts in a temporally preceding

group of RA neurons. Alternatively, the bursts patterns of RA neurons may be directly

driven by inputs from a brain area afferent to RA, for example from HVC (Fig. 1C). In this

view, each burst in an RA neuron is driven by activity in some population of HVC neurons.

Can we distinguish between these two views of sequence generation in RA? This is a

central issue we wish to address in this chapter.

How do we discriminate between a model in which the rapidly evolving ~10ms-long

burst patterns of RA neurons result from circuitry within RA, and a model in which this

activity is directly driven from HVC? A reasonable approach is to record from HVC

4/15/2004 12:53 PM

4

neurons that project to RA (HVC(RA) neurons). If RA bursts are driven directly from HVC,

one might expect to observe bursts of spikes in HVC(RA) neurons precisely time-locked to

bursts in RA and to the song vocalization. In contrast, if HVC(RA) neurons are not involved

in driving RA bursts, one might not expect to find precisely time-locked bursting activity in

HVC. In fact, this latter view was supported by recordings of HVC neurons in the singing

bird that generate rather noisy trains of spikes at a high rate throughout the song

vocalization, in striking contrast to the sparse bursting spike patterns of RA neurons 14.

Models of sequence generation: Experimental analysis.

To test these alternative hypotheses, we set out to record specifically from the

population of HVC neurons that project to RA. HVC contains at least three distinct classes

of neurons 18-20 [see also Mooney, this volume]: neurons projecting to area-X (HVC(X)),

neurons projecting to RA (HVC(RA)), and at least one class of local interneurons (HVC(I)).

Because we were unable to distinguish these classes of neurons on the basis of their spike

waveform, we used antidromic stimulation 21,22 to identify these neurons. By placing

bipolar electrodes in RA and in X, we can stimulate the axon terminals of an HVC neuron

that projects to either of these brain areas 23. We found that stimulation in X by a single

biphasic pulse (0.2ms duration) activates HVC(X) neurons and HVC(I) neurons, but not

HVC(RA) neurons. Likewise, stimulation in RA activates HVC(RA) neurons and HVC(I)

neurons, but not HVC(X) neurons. The projection neurons (of either type) were easily

distinguished from interneurons by examining the variability in latency from stimulus to

spike response. Projection neurons produced the first spike with a timing jitter less than

50µs, whereas putative interneurons produced the first spike with a timing jitter more

than500µs. 1

1 Antidromic responses in HVC were characterized in two stages: First, we characterized the distribution

of latencies and latency variability of each HVC neuron type to stimulation in RA and X in awake and sleeping birds. The responses were found to fall into four clear categories based on 1) site of stimulation (RA or X), and 2) latency variability (<50µs or > 500µs). Second, in a separate set of experiments we used the spike collision test to verify the identity of a subset of neurons in each category (Fuller and Schlag, 1976). These experiments were done in sleeping birds, in which all HVC neuron types exhibit significant spontaneous rates. By this test, all neurons responding to RA stimulation with a latency variability < 50us exhibited spike collision; all neurons responding to X stimulation with a latency variability < 50µs exhibited

4/15/2004 12:53 PM

5

Recording in singing birds poses several technical challenges. First, singing is a

natural social behavior and zebra finches will sing only while relatively unrestrained,

necessitating the use of a small device, or microdrive, mounted on the head to hold and

position electrodes in the brain. The microdrive allows an extracellular recording electrode

to be advanced and retracted to isolate the electrical signal from a single neuron; electrode

positioning may be accomplished by holding the animal and manually turning a small

screw on the microdrive. Unfortunately, handling a zebra finch to operate the microdrive

greatly reduces the likelihood that the bird will sing once a neuron is isolated and the bird is

released. We have solved this problem using a motorized microdrive. This microdrive

permits us to remotely and independently position up to three electrodes, and has resulted

in a factor of 50 increase in the per-animal yield of neurons recorded during singing 24, as

compared to an earlier non-motorized version of the microdrive 25. Once an HVC neuron is

isolated using the microdrive and identified using antidromic stimulation, the bird is

induced to sing by presenting a female zebra finch. After the bird sings several tens of song

motifs, or after the neuron signal is lost, the electrode is advanced until another neuron is

isolated and identified. The process can be repeated many times, building up a dataset of

firing patterns in a single bird.

In carrying out our experiments in HVC of singing birds, we found that antidromic

activation is essential not so much to distinguish different HVC neuron types, as to avoid

the tremendous selection bias associated with differences in spontaneous firing rates in the

different classes of HVC neurons. HVC(I) neurons have relatively high spontaneous rates in

the awake non-singing bird (2-40Hz). In contrast, HVC projection neurons have very low

rates of spontaneous activity: HVC(X) neurons generate spontaneous spikes at <1Hz, and

we have not yet observed a spontaneous spike in an HVC(RA) neuron in an awake adult

zebra finch. As a result, by simply searching for neurons in HVC with an extracellular

spike collision; finally, no putative interneurons (those neurons responding to RA or X stimulation with latency variability >500 µs) exhibited spike collision. We concluded that identification on the basis of stimulation site (RA or X) and latency variability is unambiguous. All neurons recorded in the singing bird were identified solely on the basis of stimulation site and latency variability. Spike collision tests were not carried out in these recordings because of the extremely low rates of spontaneous activity of projection neurons in the awake adult bird.

4/15/2004 12:53 PM

6

recording electrode, one is most likely to find an interneuron. For the same reason,

HVC(RA) neurons cannot be found by searching for spontaneous activity in an awake non-

singing bird (the birds do not sing often enough to search for these neurons during singing).

We have avoided this source of selection bias by searching for antidromically-activated

spikes during ongoing stimulation (1Hz) in RA and/or X. Even with this approach,

HVC(RA) neurons are still difficult to isolate because of their small size; we find them at a

rate far lower than expected from their relative abundance (50-80% of the HVC neuronal

population 26).

Using antidromic identification, we were able to record from RA-projecting neurons

during singing and test our hypotheses as to the origin of the complex burst sequences in

RA. Figure 2A shows the extracellular signal from an isolated HVC(RA) neuron during three

sequential song motifs. Most HVC(RA) neurons generated a single burst of 3-4 spikes during

each song motif 23. The firing rate during bursts of song-related HVC(RA) neurons was

680±138 Hz (not including one outlier that generated a 50 Hz burst, see Fig. 2B). The

bursts generated by a single neuron were highly stereotyped across song motifs, as

determined from the alignment of burst onset to vocal output (0.66ms rms jitter), the

number of spikes per burst (±1), and inter-spike intervals within a burst (0.13 ms rms

jitter).

Most HVC(RA) neurons were active during vocalizations — either songs or calls.

Roughly 16% (3/18) of vocal-related HVC(RA) neurons burst during calls, but not during

song. Another 11% (2/18) of vocal-related neurons burst during introductory notes, but not

during the song motif. An additional 7 units were identified by antidromic stimulation as

HVC(RA) neurons, but these were not spontaneously active, nor were they active during

singing nor during any calls that could be elicited. Because spikes were only observed from

these neurons during antidromic stimulation, there is less certainty about the classification

of these units. However, if these are assumed to be HVC(RA) neurons, then up to half of

antidromically-activated HVC(RA) neurons in our data set were not active during song

motifs, consistent with previous observations that only half of HVC(RA) neurons showed

elevated ZENK expression during singing 27.

4/15/2004 12:53 PM

7

To display the behavior of the population of HVC(RA) neurons, we show the firing

patterns of all song-related HVC(RA) neurons aligned to the song. Because the song is

highly stereotyped, we choose one song motif as a template and line up all motifs (sung by

the same bird) to the template motif: Alignment can be done with sub-millisecond timing

precision. Because the song and the neural signals were recorded simultaneously, we can

display the firing patterns of all neurons recorded in the same bird, aligned to the template

motif (Fig. 2B). All HVC(RA) neurons recorded in this bird burst at different times in the

song, with no apparent relation to syllable onsets or offsets.

It is clear that HVC(RA) neurons generate sparse patterns of bursts precisely time-

locked to the song vocalization. How does this observation relate to our two initial

hypotheses about the generation in burst sequences in RA? As we have argued above, the

presence of precisely song-locked bursts in HVC(RA) neurons is consistent with the view

that bursts in RA neurons are directly driven by inputs from HVC. In fact, our observations

suggest an interesting possible model for the generation of RA burst sequences. We

hypothesize that at each moment in the song a small ensemble of HVC(RA) is active. We

further hypothesize that the effect of a burst of activity in a population of HVC(RA) neurons

is to drive a single burst of activity in an ensemble of RA neurons with a short latency

(~4.5ms) and a short duration (~10 ms). Thus, at each moment in the song, the ensemble of

active RA neurons is driven by a small subpopulation of HVC(RA) neurons. Our working

hypothesis for the generation of RA burst sequences is summarized in Figure 3A.

In the framework of this model, we can estimate the number of HVC(RA) neurons

coactive at each time in the song. A typical motif duration is (roughly speaking) 600ms and

the average burst duration of HVC(RA) neurons is 6ms. Each song-related HVC(RA) neuron

is therefore active for ~1% of the duration of the motif. Assuming that song-related

HVC(RA) neurons are active at random times, uniformly and independently distributed

across the motif, then on average 1% of song-related HVC(RA) neurons should be co-active

at any given time. Thus, of the total ~40,000 HVC(RA) neurons in each hemisphere 26, of

which ~20,000 may be song-related, we estimate that ~200 HVC(RA) neurons are co-active

at each moment in the song motif.

4/15/2004 12:53 PM

8

The hypothesis that a subpopulation of HVC(RA) neurons is active at each time in the

song motif is based on our previous finding that, at any time in the song, roughly 10% of

RA neurons are active (Leonardo and Fee, submitted). However, based solely on our

recordings of HVC(RA) neurons in the singing bird, we cannot exclude the possibility that

the population of HVC(RA) neurons is active at only a few discrete times in the motif, each

time triggering some sequence-generating circuitry intrinsic to RA (Fig. 3B). We were not

able to record from enough HVC(RA) neurons in a single bird to demonstrate that there is at

least one of these neurons active at each moment in the song. (Given the 6ms burst width of

HVC(RA) neurons, this would likely require many HVC(RA) neurons to be recorded in the

same animal.) How then is it possible to discriminate between the two different models

shown in Figure 3?

The question is essentially one of causality: Is every burst (in a co-active ensemble of

neurons) in RA driven by an immediately preceding burst in a population of HVC(RA)

neurons? To address the issue of causality between HVC and RA requires recording

simultaneously in these brain areas, and manipulating these circuits to tease apart the role

of both the intrinsic circuitry in RA and the input to RA. Unfortunately these are very

difficult experiments, particularly in an unrestrained singing bird. Fortunately, there is an

alternative approach – the sleeping bird.

It has recently been shown that when a zebra finch sleeps, neurons in nucleus RA

generate patterns of bursts 28. The sleep-related patterns of an individual RA neuron can be

compared to the burst patterns generated by the same neuron during singing. Remarkably,

the burst patterns observed during sleep can be nearly identical to those observed during

singing, suggesting that during sleep the motor pathway can ‘replay’ the patterns it

generates for singing (Fig. 4A). Such ‘replay’ of neural activity patterns is reminiscent of

that seen in rat hippocampus during sleep after the animal explores a novel environment,

and may be involved in learning - specifically in the transfer of short term memories into

long-term memories 29. The significance of sleep replay in song learning is still unresolved,

but we have used the phenomenon of sleep replay as a sort of ‘fictive singing’(i.e. motor

4/15/2004 12:53 PM

9

output without movement, 30) to study the circuitry that underlies the generation of burst

sequences.

To address the generative mechanism of RA burst sequences we have performed

paired recordings from HVC(RA) neurons and from RA neurons in the sleeping bird. The

HVC(RA) neurons burst only rarely during sleep, roughly once every 18 seconds. (Bursts

were defined as events where the instantaneous firing rate continuously exceeds 100Hz).

In contrast, RA neurons exhibited ongoing dense patterns of bursting, on average one burst

every 1.2 seconds. Interestingly, the ratio of burst densities in RA neurons and HVC(RA)

neurons during sleep was roughly the same as during singing. If RA sequences are driven

by HVC, then by aligning the RA neuron activity to the bursts of the simultaneously

recorded HVC(RA) neuron, we should expect to see a coherent pattern. In fact, for most RA-

HVC(RA) neuron pairs, the RA neurons exhibited a statistically significant pattern of bursts

aligned to the HVC(RA) neuron burst (Fig. 4B). The RA burst sequences often extended

over a range of several hundred milliseconds before and after an HVC(RA) burst.

Do HVC(RA) neurons also burst sequentially during sleep? Simultaneous recordings of

pairs of HVC(RA) neurons in the sleeping bird showed this to be the case in about a third of

recorded pairs (Fig. 4C). Thus during sleep, just as during singing, RA neurons generate

complex sequences of bursts that are temporally locked to sparse sequences in HVC.

In contrast to HVC(RA) neurons, HVC interneurons (HVC(I)) burst densely during

sleep, and were highly synchronized to each other. Simultaneous recordings of HVC(I)

neuron pairs showed that on average 74% of the bursts in one neuron were synchronized

with a burst in the other neuron (within a 10ms window). Similarly, paired recordings of

HVC(RA) and HVC(I) neurons showed that on average 61% of bursts in an HVC(RA) neuron

were synchronized with a burst in the HVC(I) neuron. Thus, a single HVC(I) neuron can

serve as a ‘read-out’ of activity in the population of HVC(RA) neurons.

Let us now return to the question of whether each step of the RA burst sequence is

driven directly from HVC. There is one feature that clearly differentiates the two models in

Figure 3. If activity in HVC drives every burst in RA (Fig. 3A), then every burst in RA

should be immediately preceded by a burst of activity in HVC. If, on the other hand, HVC

4/15/2004 12:53 PM

10

triggers an autonomous sequence in RA, then a large proportion of RA bursts will not be

immediately preceded by a burst in HVC (Fig. 3B). Is there a way to estimate the fraction

of RA sleep bursts that are preceded (by a small latency) by a burst in the population of

HVC(RA) neurons?

Given the sparseness of HVC(RA) bursts, it would seem that one would have to record

from many HVC(RA) neurons to estimate this fraction. However, because a single HVC(I)

neuron can serve as a ‘read-out’ of activity in the population of HVC(RA) neurons, we can

answer this question by examining the relation between RA bursts and HVC(I) bursts.

Simultaneous recordings of RA neurons and HVC(I) neurons showed that bursts in the RA

neuron were preceded with high probability by a burst in the HVC(I) neuron (Fig. 5A). The

probability that an HVC(I) burst falls within a window 0ms to 10ms preceding an RA

neuron burst is on average 0.58 for all RA-HVC(I) pairs recorded (35 pairs). This number is

highly significant because the probability that an HVC(I) burst fell in any random 10ms

window was only 0.05 (Fig 5C). Because on average 58% of RA bursts were preceded by

an HVC interneuron burst, and 61% of HVC(RA) bursts were synchronized with an HVC

interneuron burst, our results are consistent with the possibility that every RA burst is

driven by HVC(RA) neurons. In other words, if every RA burst is driven from HVC, it is

expected that an HVC(I) neuron will not be active before roughly 40% of bursts in an RA

neuron, because this HVC(I) neuron fails to ‘read-out’ ~40% of the HVC(RA) bursts.

Of course, the analysis given here is correlative and cannot completely exclude the

unlikely possibility that burst sequences in HVC(I) neurons and burst sequences in RA

neurons are each generated by independent, but precisely timed circuitry within HVC and

RA. In this way, the strong correlation between HVC(I) bursts and RA bursts would not

result from a causal link between HVC(RA) and RA neurons.. Strong synchrony between

HVC and RA could in principle also be produced by common drive from another brain

area. However, there are no known inputs common to HVC and RA likely to produce such

correlations.

Our observations suggest that, during sleep, the large majority of sleep-replay bursts

in RA are directly driven by immediately preceding bursts in a sub-population of HVC(RA)

4/15/2004 12:53 PM

11

neurons, as depicted in Figure 3A. Of course, we cannot naively assume that the same

mechanisms are operating in the singing bird. In particular, we cannot yet rule out the

possibility that RA burst sequences in the singing bird may be generated largely by

circuitry intrinsic to RA, and that this circuitry is simply being ‘tickled’ into action (i.e.

raised above some threshold level) by inputs from HVC. However, given the similarity of

sleep-related and singing-related burst sequences in both HVC and RA, and the strongly

bursting nature of HVC inputs to RA, the simplest explanation is that burst sequences in

RA, during sleep and singing, are driven by direct feed-forward input from HVC.

What is the role of the recurrent excitatory and inhibitory connections observed

within RA? Although no recurrent excitatory EPSPs have been observed between pairs of

RA projection neurons, strong inhibitory interactions within RA have been demonstrated 31.

While our observations suggest that, at least during sleep, RA burst sequences are driven

from HVC, recurrent connections in RA could still play an important role in shaping the

patterns of RA activity. For example, inhibition could temporally or spatially sharpen RA

activity patterns. The role of recurrent circuitry in RA could be addressed experimentally in

the sleeping bird by comparing intracellular recordings of sub-threshold patterns of

synaptic input to an RA neuron before and during the suppression of recurrent circuitry in

RA with GABA injection.

RA efference and the generation of motor sequences during singing

Let us turn now from the origin of burst patterns in RA, to the effect of RA bursts in

the downstream motor pathway. What is the basic unit of signaling from RA neurons to

brainstem vocal nuclei? During singing, RA neurons generate exquisitely stereotyped

bursts. That is, the intervals between spikes within a given burst can be reproduced with

less than 50us jitter each time the bird sings the song motif 14. The spike timing precision in

RA has led to the suggestion that RA forms a temporal code for song; that is, the vocal

output is specified by the precise timing of individual spikes within the bursts of one or

more RA neurons 32. We consider the simpler hypothesis that RA neurons converge onto

downstream targets, such as brainstem motor neurons, in which motor control signals are

4/15/2004 12:53 PM

12

derived from a weighted sum of synaptic currents from RA neurons. In this view, the

precise timing of RA spikes could play a role in reducing motif-to-motif variability in

motor output, but would not constitute a temporal code for motor output, in the usual sense

of the term. The precise spike timing in RA may be simply a consequence of the

mechanism by which RA bursts are generated, i.e., that they are directly driven by highly

stereotyped bursts in HVC(RA) neurons. We may speculate as to why HVC(RA) neurons

generate highly stereotyped bursts in the first place. One possibility is that precisely timed

spikes within HVC(RA) bursts reflect the biophysical mechanisms involved in the generation

of the precisely timed temporal sequences observed in HVC. In other words, the stereotypy

of bursts of RA and HVC(RA) neurons is related, not to a specialized neural code, but to the

demands of producing a stereotyped and precisely timed motor behavior.

Another question that naturally arises is that of the timescale on which RA bursts

affect vocal output. Evidence that RA activity has a brief transient effect on vocal output is

provided by experiments to measure the latency and duration of vocal perturbations

induced by brief electrical stimulation in RA [see also 33]. We carried out preliminary

measurements in two birds in which bipolar stimulating electrodes were previously

implanted in RA for antidromic identification of HVC(RA) neurons. The song of each of

these birds contained a long harmonic stack syllable (>50ms duration). A computer was

used to trigger a brief (0.2 ms) electrical stimulation in RA 10ms after the onset of the

harmonic stack syllable. Recordings of the song vocalization show that RA stimulation

induced a brief increase in the pitch of the vocalization, with a duration (full width at half

maximum) of ~15ms and a latency of ~15ms between the stimulation and the peak of the

perturbation (Fig. 6). These results suggest that RA activity is integrated on a short

timescale by downstream motor circuitry, and has a similarly brief effect on vocal output.

At present, little is known about the role of brainstem motor and respiratory circuits

downstream from RA. However, a reasonable starting point for thinking about motor

control in the songbird is that each RA neuron contributes transiently with some effective

weight to the activity of the syringeal muscles 17, as has also been proposed for cortical

control of arm movement in primates 34. The total force produced by a muscle, including

4/15/2004 12:53 PM

13

that required to overcome inertia, damping, and spring forces, can be expressed as a linear

weighted sum of instantaneous activities of higher premotor neurons 35. (Consistent with

this approach, recent recordings from syringeal motor neurons reveal a linear firing rate

change in response to somatic current injection 36.) Of course, syringeal muscle forces are

related in a complex manner to vocal output 37-39, but in further discussion here we will

neglect muscle dynamics (i.e. inertia and damping) and syringeal dynamics, and will adopt

the simplified view that vocal acoustic parameters such as pitch are given directly by the

summed weighted contributions of RA neuron activity.

Simple hypothesis for song generation and learning

We can now put our picture of sequence generation in RA into a working hypothesis

of vocal sequence generation and learning (Fig. 7): HVC(RA) neurons are each active at only

one particular moment in the song motif. Furthermore, each moment in the song motif is

associated with a unique population of coactive HVC(RA) neurons. These neurons drive a

brief (~10ms) burst of activity in a small fraction (~10%) of RA neurons. The ensemble of

neurons activated is determined by the pattern of synaptic connections from HVC to RA. In

this model, the effect of RA on vocal output is of short duration, and is produced by the

convergence of RA neurons, with a set of fixed synaptic weights, onto a single neuron

representing motor output 17. One might think of this motor output as controlling a

particular vocal parameter, such as pitch. The sequence of motor, or muscle, configurations

is thus driven at each time step by activity in HVC(RA) neurons, where both the activity of

HVC(RA) neurons and their effects downstream are highly localized in time.

How would song learning proceed in the model presented above? The convergence of

the anterior forebrain pathway, a circuit necessary for song learning and maintenance 40-42,

onto RA in the song motor-control pathway, suggests that song learning is mediated by

plasticity in RA, either in the recurrent synaptic connections within RA or in the input to

RA 43,44. In our model, the pattern of activity in RA, and thus the vocal output, is

determined by the pattern of HVC(RA) –to-RA synapses; in this view, vocal learning is

controlled by synaptic plasticity at these synapses. Song learning is then simply the process

4/15/2004 12:53 PM

14

of modulating the effective connectivity from sparsely active HVC neurons to downstream

motor neurons such that, at each moment, the correct motor output is produced: RA serves

as a switchboard on which HVC(RA) neurons are wired up to the correct motor output.

At present, it is unclear when in the ontogeny of the zebra finch sparse HVC(RA)

sequences arise. If sensorimotor learning is the process of mapping the sparse sequence in

HVC(RA) to motor output, it seems plausible that the sparse HVC(RA) sequence develops

before the onset of sensorimotor learning. Learning the mapping from HVC to RA would

be most efficiently carried out after the sparse HVC sequence is established. Note that, in

principle, HVC(RA) sequences are universal; i.e. for every zebra finch song of a given

length, HVC(RA) activity can be plotted as an diagonal band equivalent to that in Figure 7B,

regardless of the vocal content of the song. Thus, auditory/vocal experience may not be

necessary in the development of sparse sequences in HVC, consistent with the possibility

that HVC sequences form spontaneously before the developmental onset of sensorimotor

learning.

We have hypothesized that each unique 5-10ms interval in the zebra finch vocal

repertoire is associated, one-to-one, with a small ensemble of co-active HVC(RA) neurons.

Our hypothesis implies that the number of coactive HVC(RA) ensembles may scale with

total duration of unique song repertoire of the bird. In other words, if we can think each co-

active HVC(RA) ensemble as a piece of ‘tape’ in the bird’s ‘tape player,’ then birds with a

longer repertoire should have more ‘tape.’ If the number of HVC(RA) neurons in each co-

active ensemble (i.e. ~200 neurons in the zebra finch) is roughly constant across birds, and

the burst duration of each ensemble is roughly constant across birds, then we would expect

the number of HVC(RA) neurons to vary linearly with the duration of the unique vocal

repertoire. In fact, HVC volume, and thus possibly the number of HVC(RA) neurons,

correlates positively with repertoire size and phrase duration in zebra finches45. Likewise,

large differences in HVC volume between male and female songbirds are reflected in large

differences in song repertoire size, within individual songbird species 46. Additionally, it

has been found that across many species of songbirds, total song repertoire size is

correlated with HVC volume [see DeVoogd, this volume] 47. Note that repertoire size is

4/15/2004 12:53 PM

15

typically quantified as the number of different song types produced by an individual. It

would be interesting to directly compare across many species the number of HVC(RA)

neurons with the repertoire duration, as defined by the total duration of all unique vocal

elements produced by an individual. Deviations from a linear relation could indicate that, in

some birds, HVC(RA) neurons are shared across different parts of the vocal repertoire, or

could suggest inter-species variations in the HVC(RA) burst duration, or variations in the

number of neurons within co-active HVC(RA) ensembles.

The sparse representation of temporal order in HVC has interesting implications for

vocal learning. For example, if there is an error in the vocal output at one particular time in

the song, synaptic outputs need only be modified for the sub-population of the HVC(RA)

neurons that were active at that time (or preceding by synaptic and axonal latency of

~20ms). If HVC(RA) neurons were active multiple times in the song, then correcting an error

at one time would introduce errors at other times in the song. Because HVC(RA) neurons are

active very sparsely, the learning process at different times become uncoupled, thus

allowing learning to occur at a faster rate. Recent theoretical work confirms that the time

required to learn motor sequences in simple feed-forward network models of HVC and RA

is minimized when HVC(RA) neurons burst exactly once per song motif 48. Thus the sparse

code observed in HVC hints at possible evolutionary pressures for optimizing the premotor

neural machinery for the task of song learning.

Our observations also raise the question of how the sparse HVC(RA) sequences are

generated. Using the same logic we applied to the origin of RA burst sequences (Fig. 1B

and C), we can ask whether sparse sequences arise from circuitry intrinsic to HVC, or

whether they are directly driven from nuclei that project to HVC, such as nucleus Interface

(NIf) or nucleus Uvaeformis (Uva). We expect that a combination of the techniques

described, in both the singing and sleeping bird, will continue to be useful as we pursue an

understanding of the circuits that underlie the generation and learning of temporal

sequences in the brain.

4/15/2004 12:53 PM

16

References

1. Carr, C. E. Processing of temporal information in the brain. Annu Rev Neurosci 16, 223-43 (1993).

2. Casseday, J. H., Ehrlich, D. & Covey, E. Neural tuning for sound duration: role of inhibitory mechanisms in the inferior colliculus. Science 264, 847-850 (1994).

3. Buonomano, D. V. Decoding temporal information: A model based on short-term synaptic plasticity. J Neurosci 20, 1129-1141 (2000).

4. Tallal, P. in Temporal Coding in the Brain (eds. Buzsaki, G., Llinas, R., Singer, W., Berthoz, A. & Christen, Y.) 291-299 (Springer, Berlin, 1994).

5. Ivry, R. The representation of temporal information in perception and motor control. Curr Opin Neurobiol 6, 851-857 (1996).

6. Gallistel, C. R. The organization of action: A new synthesis. (Lawrenc e Erlbaum Associates, Hillsdale, NJ, 1980).

7. Immelmann, K. in Bird Vocalizations (ed. Hinde, R. A.) 61-74 (Cambridge Univ. Press, New York, 1969).

8. Price, P. H. Developmental determinants of structure in zebra finch song. J Comp Physiol Psychol 93, 268-277 (1979).

9. Vicario, D. S. & Nottebohm, F. Organization of the zebra finch song control system: I. Representation of syringeal muscles in the hypoglossal nucleus. J Comp Neurol 271, 346-354 (1988).

10. Wild, J. M. Descending Projections of the Songbird Nucleus Robustus Archistriatalis. J. Comp. Neurol. 338, 225-241 (1993).

11. Nottebohm, F., Kelley, D. B. & Paton, J. A. Connections of vocal control nuclei in the canary telencephalon. J Comp Neurol 207, 344-357 (1982).

12. Wild, J. M. Neural pathways for the control of birdsong production. J Neurobiol. 33, 653-70 (1997).

13. Vicario, D. S. Organization of the zebra finch song control system: II. Functional organization of outputs from nucleus Robustus archistriatalis. J Comp Neurol 309, 486-494 (1991).

14. Yu, A. C. & Margoliash, D. Temporal hierarchical control of singing in birds. Science 273, 1871-1875 (1996).

15. Laje, R. & Mindlin, G. B. Diversity within a Birdsong. Physical Review Letters 89, 288102 (2002).

16. Troyer, T. W. & Doupe, A. J. An associational model of birdsong sensorimotor learning. I. Efference copy and the learning of song syllables. J Neurophysiol 84, 1204-1223 (2000).

17. Doya, K. & Sejnowski, T. J. in Central Auditory Processing and Neural Modeling (eds. Poon, P. W. F. & Brugge, J. F.) 77-88 (Plenum, New York, 1998).

18. Kubota, M. & Taniguchi, I. Electrophysiological Characteristics of Classes of Neuron in the HVC of the Zebra Finch. J. Neurophysiol. 80, 914-923 (1998).

19. Dutar, P., Vu, H. M. & Perkel, D. J. Multiple cell types distinguished by physiological, pharmacological, and anatomic properties in nucleus HVc of the adult zebra finch. J Neurophysiol 80, 1828-1838 (1998).

4/15/2004 12:53 PM

17

20. Mooney, R. Different Subthreshold Mechanisms Underlie Song Selectivity in Identified HVc Neurons of the Zebra Finch. J. Neurosci 201, 5420-5436 (2000).

21. Swadlow, H. Neocortical efferent neurons with very slowly conducting axons: strategies for reliable antidromic identification. J Neurosci Meth 79, 131-141 (1998).

22. Fuller, J. H. & Schlag, J. D. Determination of antidromic activation by the collision test: problems of interpretation. Brain Res 112, 283-298 (1976).

23. Hahnloser, R. H. R., Kozhevnikov, A. A. & Fee, M. S. An ultra-sparse code underlies the generation of neural sequences in a songbird. Nature 419, 65-70 (2002).

24. Fee, M. S. & Leonardo, A. Miniature motorized microdrive and commutator system for chronic neural recordings in small animals. J. Neurosci. Meth. 112, 83-94 (2001).

25. Venkatachalam, S., Fee, M. S. & Kleinfeld, D. Ultra-miniature headstage with 6-channel drive and vacuum-assisted microwire implantation for chronic recording from the neocortex. J Neurosci Meth 90, 37-46 (1999).

26. Wang, N., Hurley, P., Pytte, C. & Kirn, J. R. Vocal Control Neuron Incorporation Decreases with Age in the Adult Zebra Finch. Journal of Neuroscience 22, 10864-10870 (2002).

27. Jarvis, E. D., Scharff, C., Grossman, M. R., Ramos, J. A. & Nottebohm, F. For Whom the Bird Sings: Context-Dependent Gene Expression. Neuron 21, 775-788 (1998).

28. Dave, A. S. & Margoliash, D. Song replay during sleep and computational rules for sensorimotor vocal learning. Science 290, 812-816 (2000).

29. Louie, K. & Wilson, M. A. Temporally structured replay of awake hippocampal ensemble activity during rapid eye movement sleep. Neuron 29, 145-56 (2001).

30. Wallen, P., Grillner, S., Feldman, J. L. & Bergelt, S. Dorsal and Ventral Myotome Motoneurons and Their Input During Fictive Locomotion in Lamprey. Journal of Neuroscience 5, 654-661 (1985).

31. Spiro, J. E., Dalva, M. B. & Mooney, R. Long-range inhibition within the zebra finch song nucleus RA can coordinate the firing of multiple projection neurons. J Neurophysiol 81, 3007-3020 (1999).

32. Chi, Z. & Margoliash, D. Temporal Precision and Temporal Drift in Brain and Behavior of Zebra Finch Song. Neuron 32, 899-910 (2001).

33. Vu, E. T., Mazurek, M. E. & Kuo, Y. Identification of a forebrain motor programming network for the learned song of zebra finches. J Neurosci 14, 6924-6934 (1994).

34. Fetz, E. E. & Cheney, P. D. Postspike facilitation of forelimb muscle activity by primate corticomotoneuronal cells. Journal of Neurophysiology 44, 751-772 (1980).

35. Todorov, E. Direct cortical control of muscle activation in voluntary arm movements: a model. Nature Neuroscience 3, 391-398 (2000).

36. Sturdy, C. B., Wild, J. M. & Mooney, R. Respiratory and Telencephalic Modulation of Vocal Motor Neurons in the Zebra Finch. Journal of Neuroscience 23, 1072-1086 (2003).

4/15/2004 12:53 PM

18

37. Suthers, R. A., Goller, F. & Pytte, C. The neuromuscular control of birdsong. Philos Trans R Soc Lond B Biol Sci 354, 927-39 (1999).

38. Fee, M. S., Shraiman, B., Pesaran, B. & Mitra, P. P. The role of nonlinear dynamics of the syrinx in the vocalizations of a songbird. Nature 395, 67-71 (1998).

39. Gardner, T., Cecchi, G., Magnasco, M., Laje, R. & Mindlin, G. B. Simple gestures for birdsongs. Phys. Rev. Letts 87, 208101 (2001).

40. Bottjer, S. W., Miesner, E. A. & Arnold, A. P. Forebrain lesions disrupt development but not maintenance of song in passerine birds. Science 224, 901-903 (1984).

41. Bottjer, S. W., Halsema, K. A., Brown, S. A. & Miesner, E. A. Axonal connections of a forebrain nucleus involved with vocal learning in zebra finches. J. Comp. Neurol. 279, 312-326 (1989).

42. Brainard, M. S. & Doupe, A. J. Interruption of a basal ganglia-forebrain circuit prevents plasticity of learned vocalizations. Nature 404, 762-766 (2000).

43. Stark, L. L. & Perkel, D. J. Two-stage, input-specific synaptic maturation in a nucleus essential for vocal production in the zebra finch. Journal of Neuroscience 19, 9107-16 (1999).

44. Mooney, R. Synaptic basis for developmental plasticity in a birdsong nucleus. J Neurosci 12, 2464--2477 (1992).

45. Airey, D. C. & DeVoogd, T. J. Greater song complexity is associated with augmented song system snatomy in zebra finches. Neuroreport 11, 2339-2344 (2000).

46. Brenowitz, E. A. Comparative Approaches to the Avian Song System. Journal of Neurobiology 33, 517-531 (1997).

47. DeVoogd, T. J., Krebs, J. R., Healy, S. D. & Purvis, A. Relations between song repertoire size and the volume of brain nuclei related to song: comparative evolutionary analyses amongst oscine birds. Proc. R. Soc. Lond. B 254, 75-82 (1993).

48. Fiete, I. R., Hahnloser, R. H. R., Fee, M. S. & Seung, H. S. Temporal sparseness of the premotor drive is important for rapid learning in a neural network model of birdsong. J. Neurophysiol. In press (2004).

Acknowledgements

Recordings of RA neurons in the singing bird were carried out in collaboration with

A. Leonardo. The work described here was supported in part by the National Science

Foundation.

4/15/2004 12:53 PM

19

Figure Captions

Figure 1: Burst sequences in RA of the singing zebra finch, and two models of the

generation of these sequences. A) Instantaneous firing rate of six RA neurons recorded

sequentially in one zebra finch, during singing. Each RA neuron generates a complex

sequence of roughly ten brief bursts (~10ms average duration). The bursts are distributed

throughout the song such that on average 12% of RA neurons are active at each time. At

top is shown the spectrogram of one song motif, to which neural activities have been

aligned. B) A cartoon model of burst sequence generation in RA in which circuitry within

nucleus RA generates the short (~10ms) timescale structure in RA firing patterns. C) A

cartoon model of RA sequence generation in which short timescale structure in RA is

driven directly from premotor nucleus HVC.

Figure 2: Spike activity of identified HVC neurons in the singing zebra finch. A)

Extracellular recording of one antidromically identified RA-projecting HVC neuron during

three renditions of the song motif. B) Instantaneous firing rate of 8 RA-projecting HVC

neurons recorded in the same bird. These neurons generated a single burst of high-

frequency firing, each at a different time in the song.

Figure 3: Two possible models for RA burst sequence generation consistent with the firing

patterns of RA-projecting HVC neurons. A) At each time in the sequence, an ensemble of

HVC neurons activates an ensemble of RA neurons to which it is synaptically connected.

Conversely, each burst in RA is driven by a synaptically-connected ensemble of HVC

neurons. B) Bursts in HVC trigger short self-propagating sequences in the RA circuitry.

Figure 4: Song replay during sleep. A) Comparison of burst patterns of an RA neuron

during singing (top trace) with a burst pattern during sleep (bottom trace). Adapted from

Dave and Margoliash (2000) with permission. Burst patterns during sleep can closely

resemble burst patterns during singing. B) Simultaneous extracellular recording from an

4/15/2004 12:53 PM

20

HVC(RA) neuron (filled trace) and an RA neuron (empty trace) during sleep. Plotted is the

instantaneous firing rate of the RA neuron aligned to the onset of the sparse bursts of the

HVC(RA) neuron. C) Simultaneous recording of two sequentially-bursting HVC(RA) neurons,

during sleep.

Figure 5: During sleep, most bursts in RA are preceded by bursts in HVC interneurons. A)

Simultaneous recording from an RA neuron and an HVC interneuron during sleep. B)

Probability as a function of time that the HVC interneuron generated a spike relative to the

time at which the RA neuron spiked, calculated for the neuron pair in panel A. C)

Probability that the HVC interneuron generated a spike in a window 0ms to 10ms

preceding the RA spike, P(Int|RA) (x-axis) versus the probability that the HVC interneuron

generated a spike in any random time window, P(Int), for all HVC(I) –RA pairs recorded.

Each point represents a different pair. C) Distribution of probability ratios P(Int|RA)/P(Int)

over all the neuron pairs. The ratio represents how much more likely the HVC interneuron

was to spike, given the presence of a following RA spike.

Figure 6: Effect of brief electrical stimulation in RA during singing. A) Zebra finch song

syllable containing a harmonic stack. B) A single 0.2 ms pulse (~300 uA) in RA produces a

brief transient perturbation in the vocal output. C) Pitch of stimulated song syllable

(dashed) and unstimulated song syllable (solid). Note the transient pitch change with 15ms

latency to the peak and 15 ms duration (full width at half-maximum deviation).

Figure 7: Working hypothesis of vocal sequence generation in the zebra finch song control

system. A) HVC(RA) neurons burst at a single time in the song motif. At each time in the

song, a coactive population of ~200 HVC(RA) neurons activates ~10% of RA neurons,

which in turn converge (through motor neurons) to produce a motor control signal (e.g.

muscle activation). B) Representation of firing rate of each population of neurons during a

brief sequence (~70 ms duration, corresponding roughly to a short song syllable). Note:

4/15/2004 12:53 PM

21

The alignment of burst onsets and offsets is only for artistic clarity; at present we have no

conclusive evidence for such alignment.

RA

(RA)HVC

10ms

RA

(RA)HVC

10ms

Intrinsic dynamics in RA Feedforward activation from HVC

A

B C

Figure 1

100 ms

1 kHz

RA

neuro

ns

Fre

quency

(kH

z)

0

1

2

3

4

5

6

4

8

Song motif

Fee et. al.

100 ms1 kHz

0.5 s

1

2

3

4

HV

Cn

eu

ron

s

B

A

Fre

qu

en

cy

[kH

z]

0

8

Song motif

Motif # 1 2 3

5

6

7

8

Figure 2Fee et. al.

4 ms

Syllable: a

Sound

cb

(RA

)

HVC(RA)

RA

(RA

)H

VC

1

2

10ms

RA

RA

RA

(RA)

(RA)

HVC

HVC

(RA

)H

VC

1

1

1

2

2

33

5

5

44

10ms 10ms

10ms

Intrinsic dynamics in RA triggered from HVC

Feedforward activation from HVCA

B

Figure 3Fee et. al.

�

�

�

�

��

��

��

��

�

�

��

�

�

�

A

B100 ms

100 ms

1 kHz

singing

sleeping

C

Figure 4Fee et. al.

RA

12

HVC(RA)

HVC(RA)HVC(RA)

P(I

nt|R

A)

Time of RA spike

P(Int|RA)

P(I

nt)

#P

airs

A

B

D

C

Figure 5Fee et. al.

HVC Interneuron

RA neuron

Time [ms]

0

0.0

0.0

0

5

10

1 10 100

0.0

0.1

0.2

250ms

0.2

0.5

0.4

1.0

0.6

200 400-200-400

P(Int|RA) P(Int)

100 ms Stim

Stim

A B

C

Figure 6

0 20

0

4

8

6

2

40 60 ms

Fra

ctional

Pitch

Change

Fre

quency

[kH

z]

10%

-10%

Fee et. al.

RAR

AM

oto

r

Motor

(RA)HVC

(RA

)H

VC

1 2 3 54

1

2

3

5

6

7

4

10ms 10ms

10ms

76

�

�

�

�

�

�

�

�

�

�

�

�

�

�

Figure 7Fee et. al.

A

B