Top-Down Regulation of Plasticity in the Birdsong System: “Premotor” Activity in the Nucleus HVC Predicts Song Variability Better Than It Predicts Song Features

Abstract

We studied real-time changes in brain activity during active vocal learning in the zebra finch songbird. The song nucleus HVC is required for the production of learned song. To quantify the relationship of HVC activity and behavior, HVC population activity during repeated vocal sequences (motifs) was recorded and temporally aligned relative to the motif, millisecond by millisecond. Somewhat surprisingly, HVC activity did not reliably predict any vocal feature except amplitude and, to a lesser extent, entropy and pitch goodness (sound periodicity). Variance in “premotor” HVC activity did not reliably predict variance in behavior. In contrast, HVC activity inversely predicted the variance of amplitude, entropy, frequency, pitch, and FM. We reasoned that, if HVC was involved in song learning, the relationship of HVC activity to learned features would be developmentally regulated. To test this hypothesis, we compared the HVC song feature relationships in adults and juveniles in the sensorimotor “babbling” period. We found that the relationship of HVC activity to variance in FM was developmentally regulated, with the greatest difference at an HVC vocalization lag of 50 ms. Collectively, these data show that, millisecond by millisecond, bursts in HVC activity predict song stability on-line during singing, whereas decrements in HVC activity predict plasticity. These relationships between neural activity and plasticity may play a role in vocal learning in songbirds by enabling the selective stabilization of parts of the song that match a learned tutor model.

INTRODUCTION

Experience permanently alters the brain through cellular and circuit changes during critical or sensitive periods of sensory development (Hensch 2005; Knudsen 2004; Wiesel and Hubel 1963). We do not yet understand the mechanisms that define sensitive periods in the development of motor circuits that control behavior. Vocal learning in songbirds provides an ideal model of a sensitive period in the development of a complex, sequential behavior. Zebra finch songbirds, like humans, learn their vocalizations (Doupe and Kuhl 1999; Marler 1970). During the sensorimotor period, finch vocalizations progress over an ∼55-day period from simple prototypes to a mature song motif, which is an ∼1-s series of complex sounds emitted in a stereotyped pattern (Bolhuis and Gahr 2006). Sounds separated by silence are termed “syllables.” Syllables are concatenated to form motifs, and motifs are concatenated to form songs. Emitted sounds are progressively matched to a learned tutor song using auditory feedback (Konishi 1965). The process of learning can be observed motif-by-motif, syllable-by-syllable, as the bird slowly sculpts his song toward the mature form (Deregnaucourt et al. 2005).

The neural song system is a series of anatomically distinct clusters of neurons (nuclei) in the thalamus, basal ganglia, and pallium (cortex) that are dedicated to the production and plasticity of song (Bottjer et al. 1984; Nottebohm et al. 1976). The nucleus HVC (Reiner et al. 2004) is a pallial brain region that controls song behavior (Nottebohm et al. 1976; Vu et al. 1994). During anesthesia (Volman 1993) and sleep (Nick and Konishi 2005a), HVC always responds preferentially to the bird's own song, even very early in song development. HVC activity during singing changes as song matures, most dramatically demonstrated by prolonged afterdischarges or bursts in juveniles that are rarely observed in adults (Crandall et al. 2007b). During the sensitive period for vocal learning in the wake state, a neural signal that is selective for the tutor song has been recorded in HVC (Nick and Konishi 2005b), which suggests that instructive auditory signals have access to the juvenile HVC during waking and, perhaps, during singing. In addition, HVC activity during sleep is positively correlated with overnight song stability in juveniles (Crandall et al. 2007a), which suggests that HVC sleep activity may have a role in song learning. Collectively, these data indicate that HVC activity is dynamically regulated during song learning and suggest that it is a locus of vocal plasticity. The simplest test of the role of HVC in song plasticity would be to lesion it. However, lesioning HVC prevents production of learned song (Nottebohm et al. 1976). Instead, male finches with HVC lesions produce primitive vocalizations that are song-like (Aronov et al. 2008; Simpson and Vicario 1990). Thus direct neural–behavioral comparisons during song learning have been used to ascertain the role of HVC in plasticity (Crandall et al. 2007a,b).

HVC seems to code song behavior at the motif level, as opposed to the level of syllables or songs (Hahnloser et al. 2002; Vu et al. 1994; Yu and Margoliash 1996). Analysis of song development indicates that shaping of song occurs at the level of the motif, because syllables mature in the temporal context of the motif, with no transpositions (Tchernichovski et al. 2001). For these reasons, our laboratory has focused on neural–behavioral comparisons during motifs to study the role(s) of HVC in song plasticity (Crandall et al. 2007b). Bursts of population activity during motifs decrease in duration and increase in rate during vocal development (Crandall et al. 2007b). How might these bursts relate to behavior?

A basal ganglia-thalamic-cortical loop, the anterior forebrain pathway (AFP), is necessary for song plasticity (Bottjer et al. 1984; Brainard and Doupe 2000). Recent studies have clarified the role of the AFP in plasticity (Kao and Brainard 2006; Kao et al. 2005; Olveczky et al. 2005; Thompson et al. 2007): the AFP seems to induce variability in song behavior that is necessary for vocal learning through trial-and-error (Fig. 1). Furthermore, two of these studies indicate that gross anatomical lesioning or inactivation of one of the nuclei of the AFP [the lateral magnocellular nucleus of the anterior nidopallium (LMAN)] correlates with an immediate decrement in moment-by-moment song variability (Kao and Brainard 2006; Olveczky et al. 2005). Collectively, these data suggest that the role of the AFP role in song plasticity is to induce variability on short time scales and thus enable trial-and-error. What selects the sounds that are stabilized? What restricts motor plasticity during development? If the AFP alone controls song variability, why does activity in the afferent nucleus HVC change with development (Crandall et al. 2007a,b; Nick and Konishi 2005a; Volman 1993) and why is auditory information routed through HVC? Because a memory of the tutor song guides song learning through auditory feedback (Konishi 1965), a neural representation of the tutor song and auditory signals must play a role in song stabilization. If tutor song–selective neural auditory signals stabilize the bird's own song (BOS), where is the memory of the BOS stored? HVC responds selectively to both the BOS and the tutor song (Margoliash 1983; Margoliash and Konishi 1985; Nick and Konishi 2005b; Volman 1993) and projects directly to Area X, which is part of the AFP. In addition, the HVC neurons that project to Area X do not distinguish between hearing and production of the BOS in the mature adult (Prather et al. 2008), which suggests that the HVC neural signal driven by BOS auditory playback may be sensorimotor, as opposed to purely sensory, in nature. Consistent with a role for HVC in sensorimotor integration, HVC activity during singing (Crandall et al. 2007b) and in response to auditory playback of the BOS (Nick and Konishi 2005a; Volman 1993) and the tutor song (Nick and Konishi 2005b) changes with developmental song learning. Does HVC have a role in the auditory-guided restriction of plasticity during vocal learning?

FIG. 1.

Recent data have suggested that the neural circuitry that generates vocal fluctuations is located outside and parallel to the song motor pathway in the anterior forebrain pathway (AFP). A: the AFP seems to generate song variability, which is necessary for trial-and-error learning (Kao and Brainard 2006; Kao et al. 2005; Olveczky et al. 2005). Stimulation of lateral magnocellular nucleus of the anterior nidopallium (LMAN) induces behavioral variability, possibly by inducing variability in the motor pathway through synapses in robustus arcopallialis (RA) (Kao et al. 2005). B: lesioning or gross inactivation of LMAN decreases song variability (Kao and Brainard 2006; Kao et al. 2005; Olveczky et al. 2005). How is song stabilized during development and shaped by auditory signals? Auditory signals selective for the learned song are passed from HVC to both the motor pathway and the AFP. Accumulating data indicate that activity in song nucleus HVC changes during the sensorimotor sensitive period of vocal learning (Crandall et al. 2007a,b; Nick and Konishi 2005a,b; Volman 1993), suggesting that HVC is involved in auditory-guided plasticity. This study examines how HVC activity relates to behavior during singing on multiple time scales. DLM, dorsolateral part of the medial thalamus (inhibited by Area X projections); LMAN, lateral magnocellular nucleus; RA, robust nucleus of the arcopallium.

To assess the relationship of HVC activity to song plasticity and stability, we examined HVC population activity millisecond by millisecond during singing in juvenile and adult zebra finches. Instead of profoundly altering the song system using gross inactivation and lesioning, we used developmental song learning to naturally change behavior over thousands of motifs and show underlying neural mechanisms. These experiments are only possible in longitudinal recordings. Using established methods, we extracted motifs and corresponding neural activity across entire days and weeks (Crandall et al. 2007b). We used multiunit recordings that are very stable and more informative (better at predicting behavior) than any other intracortical signal, including single unit activity and local field potentials (Stark and Abeles 2007). We found that peaks in HVC activity predicted song stability, whereas activity troughs predicted plasticity. These data suggest that, in addition to its well-know role in song production (Nottebohm et al. 1976; Vu et al. 1994), HVC may also regulate song variability and, consequently, trial-and-error learning.

Aspects of this study have previously appeared in abstract form (Nick et al. 2007).

METHODS

Subjects

Forty-one juvenile (age 61–90 days; in the late sensorimotor stage of song development) and 22 adult (>90 days) male zebra finches (Taeniopygia guttata) were surgically implanted with chronic population recording electrodes. Of the juveniles, 17 had high-quality neural recordings (premotor RMS signal:noise > 2), but only 10 sang ≥100 motifs in at least 1 day and were used for this study. Of the adults, five had high quality neural recordings and sang ≥100 motifs. Three of the finches implanted as juveniles (Blue-70, Blue-15, and Blue-81) were also recorded after day 90 (≥100 motifs) and thus included in the adult group for those days. All song recordings were made in the absence of a female (undirected). Undirected song is thought to reflect song practice (Jarvis et al. 1998). All juvenile and three of the adult finches were reared in our facility on a 12:12 light cycle. Two adult birds were obtained from a commercial supplier. Although experience and not age is probably the strongest predictor of song system maturity, age is correlated with experience and presents a more quantifiable parameter. None of the finches used in these experiments were ever exposed to auditory playback. The finches were allowed to hear their own vocalizations. All procedures were approved by the University of Minnesota Institutional Animal Care and Use Committee.

Chronic physiological recording

Basic methods have been previously described (Crandall et al. 2007b). For population recordings, finches were implanted with a set of recording electrodes: one or two 50-μm nichrome-formvar electrodes in HVC (not plated, 1.1–1.8 MΩ, AM Systems, Carlsborg, WA), a 50-μm nichrome-formvar electrode adjacent to HVC for use as a reference electrode, and 75-μm silver ground and EEG electrodes. The headstage and recording environment were previously described (Nick and Konishi 2001; Schmidt and Konishi 1998).

All data were acquired with custom-written (Datafleet, Minneapolis, MN) LabView software (National Instruments, Austin, TX) at a sampling frequency of 44.1 kHz. During recording, a lightweight operational amplifier was attached to the bird and connected to a mercury commutator via a flexible cable. HVC neural activity was amplified 1,000 times and filtered 300–10,000 Hz. Song was monitored with a microphone (Earthworks, Milford, NH), high-pass in-line filtered at 100-Hz (Shure, Chicago, IL), and recorded. Localization of electrodes to HVC was confirmed with premotor activity in all cases and cresyl-violet histology in 7 of 16 finches.

Song behavior analysis

All data were analyzed with custom-written Matlab functions. Initial song analysis consisted of the sorting of sound data and exclusion of movement noise. Sound data were further sorted according to temporal properties. Preliminary songs were defined as sounds lasting ≥500 ms with time gaps of no more than 20 ms.

To specifically study activity during learned song behavior, a canonical motif was identified by a skilled observer from the oldest day available from each finch and used to extract motifs and corresponding neural activity throughout the recording period. Multiple motif forms could have been selected depending on the developmental stage. However, we were most interested in how HVC activity changed once the overall rhythm had been established and thus selected the most mature motif available. Because perfusion and histology were performed as soon as the electrode recording declined, the final mature motif may not have been achieved by some subjects.

We focused our analysis on behaviorally relevant vocalizations (instead of, for example, movements) by band-pass filtering our vocal data (1–8 kHz). Amplitude envelopes of the canonical motif and all preliminary songs were constructed by filtering the rectified sound recording with a Savitzky-Golay smoothing filter (4th-order polynomial fit; 20-ms frame size). Amplitude envelopes of the canonical motif and preliminary songs were cross-correlated (Crandall et al. 2007b). Sharp peaks in the cross-correlation indicated the onset of a motif that matched the canonical motif. Behavioral song motifs and corresponding neural activity were excised and saved for further analysis.

Song was quantified based on six features: amplitude, Weiner entropy, frequency, pitch, FM, and pitch goodness (a measure of sound periodicity). These features were calculated from the raw motif in 9.3-ms bins with sliding 1-ms steps with a subprogram of Sound Analysis Pro (Tchernichovski et al. 2000), ported to Matlab by S. Saar.

Multiunit physiology analysis and comparisons to song features

To examine patterns of population spiking activity during the motif, amplitude envelopes of neural activity during motifs were created by band-pass filtering 300–6,000 Hz, rectifying, and smoothing with a Savitzky-Golay smoothing filter (4th-order polynomial fit; 50-ms frame size).

For each bird, neural data for all available motifs were temporally aligned, and the mean and variance for each millisecond was taken across all motifs for each day. Likewise, data for each song feature were aligned, and the mean and variance for each millisecond was taken across all motifs. For the three birds that spanned the 90-day age, data preceding and on/following day 90 were analyzed separately and classed as “Juvenile” or “Adult,” respectively.

Statistical analysis and data presentation

Vectors representing the mean or variance for neural activity or vocal features were compared using multiple linear regression. Neural activity preceding song features by 30–70 ms was used for comparisons. Pearson's correlation coefficients (Rs) were compared between adults and juveniles using unpaired Student's t-test (2-tailed). For linear regression comparisons of juveniles and adults, ideally each bird would contribute only one Pearson R value for each feature and time lag, even though multiple days were recorded. Thus for each bird, single vectors for each feature and for neural activity were calculated by taking the mean of each parameter for each millisecond across multiple days. All bars represent means and all error bars represent SE. Significance was defined as α = 0.05. A Bonferroni correction for multiple comparisons was applied to the five time-point analyses (30–70 ms neural preceding song) of the same vocal feature. The criterion for significance in these cases was P < α/5 = 0.01.

RESULTS

Lulls in HVC activity occur during plastic song syllables

Qualitative comparison of song behavior and smoothed HVC population activity showed that lulls in HVC bursting were correlated with increased plasticity in the corresponding song syllable. The example shown in Fig. 2 shows data from a relatively mature finch (99 days) that had a partially crystallized song. There was strong HVC population burst rhythm except during a syllable that was still quite variable with regard to duration (Fig. 2; other examples in Supplementary Figs. 1 and 2).1 Note the decrease in HVC activity during the variable syllable (solid arrowheads and yellow boxes). These exemplar data suggest that song plasticity increases when HVC activity decreases. The remainder of the study quantifies and elaborates this idea.

FIG. 2.

HVC activity oscillates with lower power during song syllables that are plastic. Data from a relatively mature animal show differences in activity between relatively stable and more plastic syllables. The yellow boxes and black arrowheads indicate a syllable that varied in duration within single songs. Note the dampening of HVC oscillations during the plastic syllable relative to other activity within the same motif. HVC activity was rectified and smoothed. Scale bar: 200 ms. Representative data are from bird Blue-81, age 99 days.

Temporal alignment of song motifs shows reliable patterns in HVC activity and song features

To quantitatively test the hypothesis that HVC oscillations relate to song plasticity, it is important to identify temporal song segments or syllables that are plastic and compare them over multiple renditions. However, if the song segment is plastic, it is difficult to identify it based on its own song features because they are inherently unstable. Therefore we used the temporal context of the motif to align song elements, both stable and plastic.

We measured HVC activity during singing across multiple days. Zebra finches were maintained in the same recording chamber for many days before and during recording. Finches were only disturbed for food and water replenishment. The electrodes were never experimentally adjusted. These efforts resulted in stable recordings over many days and, in a few cases, for several weeks. For all data shown, the finch had recovered from surgery for ≥6 days, been in the recording chamber for ≥3 days, and sang in the chamber at least 1 day. All HVC activity in this manuscript occurred during undirected song motifs. “Juveniles” were 61–90 days of age (in the late sensorimotor stage of song development) and “adults” were >90 days of age (Supplementary Table 1).

As in a previous study (Crandall et al. 2007b), cross-correlation of the amplitude envelopes of each bird's relatively mature motif and his entire song database enabled precise extraction of song motifs and corresponding neural activity. For analysis, HVC activity was rectified and smoothed. Song motifs were analyzed using six song features calculated by a subprogram of Sound Analysis Pro (Tchernichovski et al. 2000): amplitude, entropy, frequency, pitch, FM, and pitch goodness (sound periodicity). Temporal patterns relative to the motif (Fig. 3 A) were found in both the neural activity (Fig. 3B) and song features (Fig. 3, C and D). The reader may note the bars of red (high levels) and blue (low levels). Although the overall pattern of HVC activity was stable, there were small changes in some bars (e.g., they shifted in time relative to the motif and/or changed in duration). To quantify the pattern of HVC activity across each day, we took the mean and variance of the HVC activity for each millisecond. Similar to this millisecond level analysis of HVC activity, daily patterns of song features were quantified by taking the mean and variance across all motifs for each millisecond of the motif.

FIG. 3.

Example data from a single day show alignment of bursts of HVC neural activity, song amplitude, and song FM. A: a representative motif for comparison with cross-day data. B: HVC activity bursts at reliable times during the motif. The arrow indicates that the mean of neural activity was calculated for each millisecond across motifs. C: song amplitude varies across each motif but is relatively stable across motifs at a given millisecond. The arrow indicates that the mean and variance of song features were calculated for each millisecond across motifs. D: song FM is also relatively stable across motifs at a given millisecond. Example data are from bird Blue 70, age 90 days.

HVC activity correlates with the mean and variance of some song features

To directly assess potential effects of HVC activity on song features, we compared the mean and variance vectors computed for each millisecond of the motif. We used linear regression to compare the relationship of song features at each millisecond to the neural activity preceding that millisecond. Figure 4 shows the comparison of mean HVC activity (50 ms prior) to respective song features: the mean and variance of amplitude and FM. As would be expected for a premotor area, HVC activity was correlated with the mean amplitude of song (Fig. 4A). Mean HVC activity inversely predicted amplitude variance (Fig. 4B). In contrast to amplitude, HVC did not directly predict FM (Fig. 4C). More interestingly, HVC weakly predicted FM variance.

FIG. 4.

Neural activity correlates with some song features, but not others. A: mean HVC activity correlates with mean song amplitude. Top: the same representative sonogram is shown at the top of each feature plot for temporal orientation purposes only. Middle: mean amplitude (green) is plotted on the same axis as mean smoothed HVC activity to show temporal correlations. Note that when HVC activity is high, so typically is the mean amplitude. Bottom: linear regression shows a direct relationship between HVC activity and song amplitude. B: amplitude variance is negatively correlated with HVC activity. Middle: amplitude variance (red) seems to peak when HVC activity is low and vice versa. Bottom: HVC activity inversely correlates with amplitude variance, as shown by linear regression. C: mean FM does not seem to be correlated with HVC activity. D: in contrast, the variance of FM peaks when HVC activity is low, and vice versa (middle). Bottom: this leads to a small inverse correlation. Example data are from bird Red 136, age 65 days.

Although HVC activity predicts some song features, the predictive capacity is not developmentally modulated or temporally precise

To quantitatively assess the effects of HVC activity on song behavior across development, we compared the Pearson Rs for mean HVC activity versus the mean of each song feature using unpaired t-test (Fig. 5). The relationship of song features at each millisecond was compared with the neural activity preceding that millisecond by 30–70 ms. Increased HVC activity directly predicted increased amplitude at all time points examined (Fig. 5, row 1). The peak relationship was with HVC activity preceding the corresponding song millisecond by 40 ms, although the relationship in the time range 30–50 ms was fairly stable. This suggests that HVC's prediction of amplitude is not temporally precise. Mean HVC activity also weakly predicted decreased entropy and increased pitch goodness in juveniles (Fig. 5, rows 2 and 3). However, these effects were not temporally precise (compare 30–60 ms) or significantly different between juveniles and adults. The reader may note the near-significant comparison at 50 ms for pitch goodness (significance with a Bonferroni correction is P < 0.01). Collectively, these data indicate that HVC activity levels do not directly specify song features with temporal precision or developmental regulation.

FIG. 5.

HVC activity directly predicts mean amplitude, but little else. Bar plots of juvenile (black; n = 10 finches) and adult (gray; n = 8 finches) mean Pearson Rs show no developmental modulation of HVC song feature relationships. Each row contains the temporal data from a single song feature. Each column contains the feature for a given HVC song lag. With a Bonferroni correction for multiple comparisons, significance is defined as P < 0.01.

Variance of HVC activity predicts variance in some song features, but the predictive capacity is not developmentally modulated

One might predict that variance in the activity of a premotor area would be reflected in the variance of behavior. Surprisingly, this is not what we generally found in the time range typically defined as “premotor” in HVC (40–60 ms; Fig. 6; Supplementary Fig. 3) (McCasland and Konishi 1981; Troyer and Doupe 2000). HVC activity variance 50 ms before the corresponding song feature is shown in Fig. 6. The variance of HVC activity weakly predicted variance in song amplitude but little else. These relationships were not significantly developmentally modulated. These data suggest that a great deal of HVC activity preceding song does not directly specify spectral aspects of vocalizations.

FIG. 6.

Variance in HVC activity directly predicts variance in amplitude but little else. Bar plots of juvenile (black; n = 10) and adult (gray; n = 8) mean Pearson Rs show no significant developmental modulation of variance of HVC variance of song feature relationships. Each row contains the temporal data from a single song feature, with the compare HVC activity preceding the song feature by 50 ms.

Mean HVC activity predicts variance in five song features, three with temporal precision and developmental regulation

Unlike the variance in HVC activity, which had little predictive value for song variation, mean HVC activity did predict variance in most song features. In juveniles, the variances of all features except pitch goodness were inversely predicted by HVC activity [note the nonzero R values for juveniles (black) in Fig. 7]. The variance of amplitude and Weiner entropy were equally well predicted by HVC activity in juveniles and adults (Fig. 7, rows 1 and 2). In contrast, FM was developmentally regulated (Fig. 7, row 6). In addition, the developmental regulation of frequency and pitch was near significant (Fig. 7, rows 4 and 5). Notably, the developmental modulation was significant for FM variance when it was compared with HVC activity 50 ms prior, but not other time points, indicating that the developing HVC–FM variability relationship was temporally precise (Fig. 7, row 6, column 3). Likewise, the relationship of the variance of frequency and pitch to mean HVC activity peaked at 50 ms in the juvenile. Thus the strongest relationships between neural activity and song variance were with neural activity leading behavior by 50 ms. These results confirm prior reports that have estimated an ∼50-ms neurobehavioral delay (McCasland and Konishi 1981; Troyer and Doupe 2000).

FIG. 7.

HVC inversely predicts song feature variance. In juveniles (black; n = 10), all features examined except for pitch goodness were inversely predicted by mean HVC activity. This relationship did not change with development for amplitude and entropy. In contrast, the relationship of neural activity and frequency, pitch, and FM trended to be stronger in juveniles than in adults (gray; n = 8). This difference was significant for FM. As in Fig. 5, each row contains the temporal data from a single song feature. Each column contains the feature for a given HVC song lag. With a Bonferroni correction for multiple comparisons, significance is defined as P < 0.01.

DISCUSSION

This study provides the first evidence that the song nucleus HVC may restrict plasticity during sensorimotor learning. Evidence indicates that HVC is involved in the production of learned song (Nottebohm et al. 1976; Simpson and Vicario 1990; Vu et al. 1994). This would suggest that HVC activity should be correlated with song. However, direct comparison of HVC activity and emitted sounds during singing showed that the mean HVC activity did not correlate well with the mean of most song features, nor did the variance of HVC activity correlate well with the variance of most song features. In contrast, mean HVC activity levels correlated with decreased song variance, millisecond by millisecond (Fig. 8). In addition, the relationship of mean HVC activity levels and song variance was temporally precise and, for frequency components, developmentally modulated. These data strongly suggest that HVC activity is actively and dynamically involved in real-time song learning. Furthermore, they showed an additional, novel mechanism through which premotor brain areas may shape vocal behavior through selective stabilization of specific time windows of sequentially modulated sounds.

FIG. 8.

In summary, this study shows that HVC activity peaks relate to stability in frequency components, amplitude, and entropy in juveniles, whereas activity troughs predict plasticity in these song features. In contrast, HVC activity in adults does not predict stability or plasticity in frequency components but does relate to more stable amplitude and entropy.

Does HVC activity affect song stability through the AFP?

We analyzed the HVC song nucleus in a behavioral context and found that, in addition to its well known role in song production (Nottebohm et al. 1976; Simpson and Vicario 1990; Vu et al. 1994), HVC also seems to decrease song variability on moment-to-moment time scales. Potentially, increased HVC activity could more strongly drive the HVC projection to the nucleus robustus arcopallialis (RA; song “motor cortex”) and thereby reduce behavioral variability. However, the relationship between HVC activity levels and the variance of song was differentially regulated during development based on song feature, suggesting a more interesting model. Increased HVC activity could entrain the activity of the AFP via HVC's only other efferent connection, to Area X (song basal ganglia). Previous correlative lesioning and inactivation studies have argued that the AFP, which is not in the direct motor pathway, drives song variability (Fig. 1; Kao and Brainard 2006; Olveczky et al. 2005). Lesioning the output nucleus of the AFP (LMAN) correlates with decreased song variability and impaired vocal learning (Bottjer et al. 1984; Kao et al. 2005; Olveczky et al. 2005). If the HVC-to-RA motor pathway alone could induce song variability and enable trial-and-error learning, why would lesioning the AFP block song learning (Bottjer et al. 1984)? It is possible that the memory of the learned tutor song is stored or compared with ongoing behavioral feedback in the AFP. However, studies from many laboratories using a variety of techniques indicate that the locus of the tutor memory and the site of memory-feedback comparison are found upstream of HVC (Bolhuis and Gahr 2006; Jarvis and Nottebohm 1997; Mello and Clayton 1994; Nick and Konishi 2005b; Phan et al. 2006). Although some neural processing loops could place the AFP upstream of HVC, synaptic delays and coding efficiency support the more parsimonious explanation that auditory areas and/or the nuclei that project to HVC (uva, nucleus interfacialis, and the medial magnocellular nucleus) contain the tutor song memory and compare it to auditory feedback. In light of the data presented here, we propose that modulation of HVC activity level controls song variability and trial-and-error learning by controlling the AFP. If the effect of HVC on song variability is orchestrated through the AFP, it cannot be considered direct motor control but higher order planning or prediction. If the effects of HVC on song variability are direct through RA, the proposed role of the AFP in inducing song variability and thereby enabling vocal experimentation (Olveczky et al. 2005) should be reassessed.

Cellular model of motor control in the song system

How might HVC population activity affect song variability? Population activity is relatively uniform across the nucleus HVC (Nick and Konishi 2005a; Schmidt 2003). Inhibitory interneurons maintain spatially broad oscillations (Buzsaki and Chrobak 1995) and promote synchronous activity in cortex/pallium (Galarreta and Hestrin 2001; Hasenstaub et al. 2005). The most active neurons in HVC are fast-spiking neurons, which dominate multielectrode single-unit recordings (Crandall et al. 2007b) and which may become more synchronous with development (N. Aoki, D. Q. Nykamp, and T. A. Nick, unpublished observations). Collectively, these data led us to propose that faster oscillatory bursting activity characteristic of the adult HVC (Crandall et al. 2007b) is caused by the synchronous firing of interneurons. Interneurons shape the activity of projection neurons (Cobb et al. 1995; Klausberberger et al. 2003). We propose that synchronous activity in HVC interneurons entrains the activity of Area X–projecting HVC neurons and thus entrains the activity of the AFP. Control of the AFP would control variability in the behavior, because the output nucleus of the AFP (LMAN) seems to induce variability in vocal output (Kao et al. 2005; Olveczky et al. 2005). We hypothesize that as HVC interneuron bursts become stronger in development, the activity of Area X–projecting neurons becomes more entrained, the AFP becomes more entrained, the song becomes less variable, and plasticity becomes more difficult to induce. A companion hypothesis predicts that HVC interneurons entrain RA-projecting neurons as well to produce more stable behavior. It is also possible that the effects on song variability that we observe result at least partially from the introduction of synaptic “noise” directly onto the HVC neurons that project to RA, particularly when HVC activity is dominated by lower-frequency activity in juveniles (Crandall et al. 2007b). However, the fact that lesions of the AFP block song plasticity (Bottjer et al. 1984; Brainard and Doupe 2000; Williams and Mehta 1999) suggests that control of song variability and plasticity involves the AFP (see above). We have examined correlations between neural activity and behavior in finches that have been relatively unperturbed by our experiments but strongly perturbed by the developmental processes that naturally occurred during our long-term recordings. Our findings were only possible because we allowed the system to develop naturally without experimental manipulation. The next set of experiments will test the cellular hypotheses that arose from these longitudinal neurobehavioral studies using technically easier, shorter-term recording and acute perturbation with pharmacological and electrical methods.

Song learning in the context of top-down motor control

All areas of the cerebral cortex are subject to top-down control, through which complex information at higher processing stages shapes lower-level processes (Gilbert and Sigman 2007). However, the developmental assembly of top-down control is completely uncharted. Furthermore, potential roles of top-down control in the development of complex behaviors are not understood. Top-down control in the song system may represent an expectation of the bird's own song based on prior experience. Such feed-forward predictions alter the influence of bottom-up sensory inputs (Driver and Frith 2000). As a consequence, strong top-down control may be anathema to sensory-guided plasticity (Engel et al. 2001). If so, initially weak top-down influences may be crucial for sensorimotor learning (Buschman and Miller 2007; Engel et al. 2001). During song development, we speculate that release of top-down control may be necessary for bottom-up influences to affect network activity and shape the vocalization. After song maturation, increased top-down control of song through increased oscillatory activity may explain the lack of immediate effects of perturbation of sensory feedback (Zevin et al. 2004).

The oscillatory pattern in HVC may represent an expectation or prediction of the learned song that is driven by auditory signals. In the mature adult, the prediction could be driven by both the motor program, contained in the pattern of firing of HVC RA-projecting neurons and/or extrinsic inputs, and by auditory stimuli that match the product of the motor program, the bird's own song. In the juvenile, the sensorimotor expectation should be weaker, according to this hypothesis. Previous work has shown that the auditory response to playback of the bird's own song increases with development (Nick and Konishi 2005a). A tutor song matching signal (Nick and Konishi 2005b) could strengthen the bird's own song prediction based on the degree of the match between the song produced and the tutor song. Because the prediction would not directly drive behavior, Hebbian processes would be released from the constraints of synaptic motor and auditory feedback delays (Troyer and Doupe 2000). Prolonged bursting in juvenile, but not adult, HVC neurons (Crandall et al. 2007b) provides a potential mechanism through which the delay between premotor activity that predicts sound A and auditory feedback signals from sound A could be bridged during a sensitive period in development and thereby increase interconnections of effective functional networks.

The correlations that we report are relatively weak. This may be due to the fact that HVC is inherently sensorimotor, not just motor, and thus sensory feedback signals during singing may cloud the purely premotor–behavioral relationship. The weak correlations may be necessary for song learning, which proceeds over an ∼55-day period. Stronger effects of HVC activity on song variability may not allow enough plasticity or stability for optimal song learning. Whatever the case, no correlation between song system activity and song variance has ever been reported, and thus the effects of differing degrees of correlation have no context for evaluation.

Neuroethological implications

Why would a premotor area selectively modulate plasticity in frequency components during a vocal sensitive period? Song is an indicator trait that conveys the fitness of a male zebra finch to prospective female mates. Such a trait can only be reliable if it is costly (Zahavi 1975). There seem to be trade-offs between song frequency bandwidth and syllable rate, with an upper performance limit defined by the highest bandwidth and rate (Ballentine et al. 2004; Podos 1997). Female swamp sparrows and canaries prefer males performing nearest this upper performance limit (Ballentine et al. 2004; Draganoiu et al. 2002). HVC seems to have a role in song timing and thus rate (Hahnloser et al. 2002; Solis and Perkel 2005). We have provided the first evidence that HVC also may have a role in regulating song FM. Furthermore, the relationship of HVC to variance of frequency components changes with development. This finding is consistent with the “developmental stress” hypothesis that proposes that, for song indicator traits, the costs occur during development rather than production (Nowicki and Searcy 2005). Because the relationship of HVC to FM variance is developmentally regulated and the burst rate of HVC activity during singing increases during song learning (Crandall et al. 2007b), it is possible that HVC is the site of bandwidth rate optimization during developmental song learning.

GRANTS

This work was supported by The John Merck Scholars Program, National Institute on Deafness and Other Communication Disorders Grants R01-DC-007384 and K02-DC-008521, the Minnesota Medical Foundation, The Grant-in-Aid of Research, Artistry, and Scholarship from the University of Minnesota Graduate School to T. A. Nick, National Institute of General Medical Sciences Grant NI5T32-GM-008471-15 to N. F. Day, and the University of Minnesota Undergraduate Research Opportunities Program to A. K. Kinnischtzke and M. Adam.