How sleep affects the developmental learning of bird song

How sleep affects the developmental learning of

bird songS. Degregnaucourt et al. Nature 2005

introduced by K. Sasahara

12010年10月4日月曜日

Introduction• Sleep facilitates memory consolidation.

• ReactivationHippocampal areas activated during spacial learning were reactivated during sleep in humans.

• ReplaySome neurons of premotor nuclei show spontaneous bursting during sleep, similar to the pattern of activity in awake singing birds.

• Bridging between sleep and developmental learning

• A model systemSongbirds learn a correspondence between vocal-motor output and auditory feedback during development.


Materials and Methods

• Male zebra finches

• Continuous recording with a model tutoring

• Continuous measurements

• Syllable featurese.g., duration, pitch, AM, FM, entropy ...

• Similarity to the target

Dynamics of the VocalImitation Process: How a Zebra

Finch Learns Its SongOfer Tchernichovski,1*† Partha P. Mitra,2* Thierry Lints,1*

Fernando Nottebohm1*

Song imitation in birds provides good material for studying the basic biologyof vocal learning. Techniques were developed for inducing the rapid onset ofsong imitation in young zebra finches and for tracking trajectories of vocalchange over a 7-week period until a match to a model song was achieved.Exposure to amodel song induced the prompt generation of repeated structuredsounds (prototypes) followed by a slow transition from repetitive to serialdelivery of syllables. Tracking this transition revealed two phenomena: (i)Imitations of dissimilar sounds can emerge from successive renditions of thesame prototype, and (ii) developmental trajectories for some sounds followedpaths of increasing acoustic mismatch until an abrupt correction occurred byperiod doubling. These dynamics are likely to reflect underlying neural andarticulatory constraints on the production and imitation of sounds.

Vocal imitation is guided by auditory infor-mation, requires intact hearing, and is verysensitive to the age or reproductive conditionof the individual (1, 2). The brain circuits thatgovern this skill in songbirds have been de-scribed (3). We here report on conditions thatbring vocal learning under fine experimentalcontrol and provide a detailed acoustic anal-ysis of the sound transformations that under-lie the learning process.

Zebra finch (Taeniopygia guttata) malesdevelop their song between 35 and 90 daysafter hatching, a time known as the sensitiveperiod for vocal learning (4). This song con-sists of complex sounds (“syllables”) separat-ed by silent intervals (5). A song motif iscomposed of dissimilar syllables repeated ina fixed order (5). When a young male zebrafinch is reared singly in the company of anadult male, it develops a song that is a closecopy of the sounds and temporal order of thatmale’s song (4, 6). Acquisition of the audi-tory memory of the model song can start asearly as 25 days after hatching, but this onsetcan be delayed by withholding exposure tothe model (7, 8). Once acquired, a storedrepresentation of the model song can be con-verted to a motor imitation. This conversionhas been modeled by assuming simple Heb-bian and reinforcement learning rules (9).Nevertheless, past technical limitations en-countered when studying early song develop-ment have left much of the fine-grained struc-ture of the imitation process unexplored.

In many songbirds, as in humans, firstacquisition of auditory memories of learnedsounds occurs before the onset of vocal learn-

ing (10). Under such conditions, it can bedifficult to distinguish between the learnedand the innate component of the developingsounds. In the zebra finch, however, the sen-sory phase of model acquisition overlaps withthe period of motor development of learnedvocalization (5). We took advantage of thisoverlap to delay model acquisition so as toobtain a baseline of “untutored” song duringthe early subsong stage, and we then exam-ined the effect of exposure to a model songduring the remainder of the sensitive periodfor vocal learning. Untutored subsong wasrecorded, and then birds were trained, startingon day 43 after hatching, to peck at a key totrigger a short song playback from a smallspeaker housed within a plastic male model(11). Training persisted until the end of theexperiment. To enhance the identification oftime-frequency structure in subsong, whichtypically consists of poorly structuredsounds, we used multitaper spectral analysistechniques and estimated spectral derivativesthat act like “edge detectors” in the time-frequency plane of the spectrogram (12) (Fig.1). Figure 1B presents spectral derivatives ofthe emerging song of a bird just before train-

1Field Research Center, The Rockefeller University,Millbrook, NY 12545, USA. 2Bell Laboratories, LucentTechnologies, Murray Hill, NJ 07974, USA.

*All authors contributed equally to this work.†To whom correspondence should be addressed. E-mail: [email protected]

Fig. 1. An example oftraining. (A) Acclima-tion to the trainingapparatus from days30 to 42 after hatch-ing, in the presence ofa plastic model of anadult male (on middleperch). (B) Untutoredsubsong was recordedon day 43. Spectralderivatives provide arepresentation of songthat is similar but su-perior to the tradi-tional sound spectro-gram. Instead of pow-er spectrum versustime, we present di-rectional derivatives(changes of power) ona gray scale so thatthe detection of fre-quency contours is lo-cally optimized. Thiswas particularly usefulfor the analysis of ju-venile song. (C) Thekeys were then uncov-ered. The bird learnedto peck on either oneof the keys to induce ashort song playbackfrom the plastic mod-el. (D) Song playbackwas composed of tworenditions of the songmotif (the “model”)depicted. The overalldaily exposure waslimited to 28 s. Asshown, the bird’s songhad changed by (E) the second and (F) the third day of training.

R E S E A R C H A R T I C L E

30 MARCH 2001 VOL 291 SCIENCE www.sciencemag.org2564

on

May

26,

201

0 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

Target

2 days of training

3 days of training

Model tutor Pupil

O.Tchernichovski et al. Science 2001


Tracing Vocal Changes

green curve and bars). The difference between those samples is anestimate of the measurement error. We then examined vocalchanges from one day to the next, comparing a random sampleof 100 songs produced during one day with a random sampleproduced during the next day. The day-to-day vocal changes werehigher than baseline (Wilcoxon, P , 0.01) and decreased duringsong development.

We then examined vocal changes that occurred during night-sleep by comparing the last 100 songs produced during the eveningof one day with the first 100 songs produced in the morning of thenext day (red curve). Strikingly, the evening-to-next-morningchanges were higher than the day-to-day changes (P , 0.01).Similar results were obtained in a bird trained by a live tutor(Fig. 2b, striped bars). In adult birds (1 year old) we still observedthat the day-to-day variations were larger than the within-daybaseline, confirming retention of some plasticity32,33. However,transitory changes in the morning were not observed (see below).

Because vocal changes after 12 h of night-sleep were larger thanthe overall changes that occurred during 24 h, vocal changesmust benon-monotonic; that is, they must oscillate during a daily cycle.Indeed, we observed strong daily oscillations in syllable features,most pronouncedly in variance features that capture the richness ofacoustic structure within a syllable (Supplementary Table 1).Figure 3 shows how values of Wiener entropy variance (EV)increased from day to day during development (Fig. 3a, b). Strongdaily oscillations were observed shortly after training started(Fig. 3c, d), and decreased thereafter (Fig. 3e, f). Note that thedecrease in EVafter night-sleep was opposite to the overall trend of

increase. Pooling all the variance features showed that during earlysyllable development (Fig. 3h, red curve) the diversity in syllablefeatures was low in the morning, indicating a decrease in intra-syllabic structure after the night (repeated-measures analysis ofvariance (ANOVA), P , 0.001). Syllable structure recovered with2–3 h of intense morning singing. Suspending training with songplaybacks in the morning did not affect the recovery of songstructure (n ! 5,Wilcoxon, not significant (n.s.)). Daily oscillationsin song structure were much weaker during late syllable develop-ment (blue curve), and undetectable in 1-year-old birds (greencurve). Birds trained when adult, starting from day 90 (n ! 8),exhibited substantial vocal changes but showed very weak post-sleep changes. We compared post-sleep EV changes across agegroups, selecting syllables that showed similarmagnitudes of change(^s.e.m.) from one day to the next (5.1 ^ 0.3% versus 5.3 ^ 0.5%EV increase per day in birds trained from day 43 and in birds trainedfrom day 90, respectively). We found strong oscillations, with apost-sleep decrease in EV (216.7 ^ 1.5%) in birds trained fromday43, and a very weak decrease (22.4 ^ 2%) in birds trained fromday 90 (Mann–Whitney, P , 0.01; Supplementary Fig. 1 andSupplementary Data).The decrease in syllable structure in themorningmay suggest that

vocal changes during the day progress with the overall develop-mental trend, whereas vocal changes after night-sleep oppose thedevelopmental trend. To test this hypothesis we assigned a sign tovocal changes during night-sleep, by reference to the overalldevelopmental trend: positive if in the same direction, negative ifopposite. For example, if syllable mean pitch decreased during

Figure 1 Tracing vocal changes. a, Spectral derivatives29 of adult song motif with threesyllables. b, Smoothed histogram of syllable durations46 in an adult bird (^95%

confidence interval). Each peak corresponds to a song-syllable type of distinct duration.

c, To trace the development of syllable types we plot developmental maps in which eachrow represents the duration histogram of syllables during one day. Syllable types (ridges)

emerged shortly after training. d, Plotting two-dimensional distribution (duration versusWiener EV) shows syllable types as clusters (unclustered syllables are not shown).

Changes in the position of clusters reveal vocal changes. e, Colour-coding circadian timeshows vocal changes after night-sleep.

articles

NATURE |VOL 433 | 17 FEBRUARY 2005 | www.nature.com/nature 711©!!""#!Nature Publishing Group!

!

• Three clusters emerge and stabilize with development

• Variability decreases in the morning


Vocal Changes During Night-sleepFull bars: model tutorStriped bars: live tutor

night-sleep but increased during development, we assigned changeduring night-sleep as negative. Note that if vocal changes duringnight-sleep were due to random perturbations, the net changeacross many samples should approach zero. As shown in Fig. 2c,analysis across the 12 birds (including all features of all syllables)confirmed that the net effect of vocal changes after night-sleep wasnegative. Furthermore, deterioration was observed in both meanand variance feature values (Supplementary Table 1) and wasmaximal from day 50 to day 55. To separate the effect of trainingfrom that of chronological age, we repeated the experiment in birds

trained fromday 60 (Fig. 2d, n ! 6). As shown, delaying the onset oftraining delayed the effect (see Supplementary Data and Sup-plementary Fig. 2). Isolated birds (n ! 6) that were kept in thesame conditions but were not trained did not show a significanttrend (Fig. 2e, f). The magnitude of post-sleep deteriorationdiffered significantly across the groups: highest when trainingstarted early, lower when training started later, and lowest in theuntrained birds (Fig. 2f, median test, P , 0.05). Overall, song wasless structured andmore primitive after night-sleep but only duringdays of rapid learning.

Figure 2 Vocal changes during night-sleep. a, Vocal changes (absolute values;medians ^ s.e.m.) observed during song development in 12 birds trained from day 43.

Green, vocal changes across random samples obtained during the same day (baseline);

blue, vocal changes from one day to the next; red, vocal changes from one evening to the

next morning (during night-sleep). b, The same data, with medians across songdevelopment (full bars) and in a live tutored bird (striped bars). c–e, Vocal changes during

night-sleep with reference to the overall developmental trend, in birds trained from day 43

(c), in birds trained from day 60 (d) and in untrained (isolated) birds (e). f, The same dataas in c–e, showing medians across song development (median test, P , 0.05 across

groups and post hoc between birds trained from day 43 and isolates; results are medians;

error bars are s.e.m.).

articles

NATURE |VOL 433 | 17 FEBRUARY 2005 | www.nature.com/nature712©!!""#!Nature Publishing Group!

!

• Baseline < Day-to-day < Pre-post sleep

• Isolates < 60d < 43d

• Adults still retain the day-to-day variability, suggesting learning plasticity


Post-sleep Deterioration and Recovery

Post-sleep deterioration and learningWe now turn to measurements of vocal learning and examine theeventual similarity to the model song (measured on day 90) againstthe magnitude of night-sleep deterioration. Similarity analysis isbased on comparing detailed temporal structure across slidingwindows of two sounds, yielding similarities as a percentage

measure29. The overall magnitude of post-sleep deteriorationduring development is positively correlated with the eventualsimilarity to the model song (Fig. 4a, Spearman r ! 0.6,P ! 0.038; note, however, that correlation does not prove causal-ity—see further analysis in Supplementary Data). In contrast, thenet change in feature values during the entire learning period did

Figure 3 Recovery of syllable structure during the morning. a, Three examples ofincreasing syllable structure during development. b, Developmental change in thestructure of the same syllable as captured by Wiener EV. c, Changes in the structure ofthat syllable during night-sleep. d–f, Tracking EV values continuously shows a decrease inEV values after the night-sleep of day 46 (d) but not after the night-sleep of day 89 (e, f ).g, Tracing EV values continuously during development shows oscillations between days45 and 60. EV values have been smoothed with a running median (period ! 40 data

points). h, Collapsing together all variance features across birds shows the time course ofsyllable structure recovery during the day. Red curve, early syllable development (days

50–60); blue curve, late syllable development (days 85–90); green curve, 1-year-old

birds. Features units were transformed to median absolute deviation from the mean

(MAD). Mean daily MAD values are shifted to zero, to enhance the display of hourly

changes in syllable structure. Error bars are s.e.m.

articles


!

Post-sleep deterioration and learningWe now turn to measurements of vocal learning and examine theeventual similarity to the model song (measured on day 90) againstthe magnitude of night-sleep deterioration. Similarity analysis isbased on comparing detailed temporal structure across slidingwindows of two sounds, yielding similarities as a percentage

measure29. The overall magnitude of post-sleep deteriorationduring development is positively correlated with the eventualsimilarity to the model song (Fig. 4a, Spearman r ! 0.6,P ! 0.038; note, however, that correlation does not prove causal-ity—see further analysis in Supplementary Data). In contrast, thenet change in feature values during the entire learning period did

Figure 3 Recovery of syllable structure during the morning. a, Three examples ofincreasing syllable structure during development. b, Developmental change in thestructure of the same syllable as captured by Wiener EV. c, Changes in the structure ofthat syllable during night-sleep. d–f, Tracking EV values continuously shows a decrease inEV values after the night-sleep of day 46 (d) but not after the night-sleep of day 89 (e, f ).g, Tracing EV values continuously during development shows oscillations between days45 and 60. EV values have been smoothed with a running median (period ! 40 data

points). h, Collapsing together all variance features across birds shows the time course ofsyllable structure recovery during the day. Red curve, early syllable development (days

50–60); blue curve, late syllable development (days 85–90); green curve, 1-year-old

birds. Features units were transformed to median absolute deviation from the mean

(MAD). Mean daily MAD values are shifted to zero, to enhance the display of hourly

changes in syllable structure. Error bars are s.e.m.

articles


!

50-60 day85-90 day

1year

• More structure with development

• Daily oscillation in song structure:50-60d > 85-90dN.S. in 1yr


Progression of Song Learning

not correlate with better imitation (Spearman r ! 0.25, P ! 0.443),further confirming that the initial acoustic distance from the songmodel does not predict song learning17. In sum, the magnitude ofdaily oscillations in song structure, but not the total developmentalchange, seems to be a predictor of accurate song learning. A similartrend was observed in birds trained from day 90 (SupplementaryFig. 1).We propose that the morning period of less structured (and

perhaps more plastic) sounds gives the bird an opportunity toexplore its vocal abilities and improve imitation. To examine thedynamic relations between daily changes in song structure and theprogression of vocal learning, we selected the most complex syllableof each bird’s song and computed similarity to the model syllablecontinuously (for each syllable produced) from day 55 to day 70after hatching. As expected, similarity values dropped during night-sleep (repeated-measures ANOVA, P , 0.001) and then increasedto a plateau (Fig. 4b). To measure a bird’s overall progress towardsimitating the tutor song, we defined a measure called ‘recordsimilarity’, the quality of a bird’s best success so far in copying thetutor song (Fig. 4b, c). For robustness, we computed the 95th centileof the similarities within a 1-h moving window. The recordsimilarity at a given time is defined as the best 95th centile achievedso far. Improvements in record similarity are computed withreference to developmental time, not to the previous hour, thusexcluding improvements that are due solely to recovery (Fig. 4b, c).Strong improvements in record similarity were achieved during

the third hour of the photophase (Fig. 4d), overlapping with the latephase (third hour) of the recovery of song structure (Fig. 3h, redcurve). Little improvement was achieved during the second half ofthe day. This pattern was not a consequence of different rates ofsinging across the day, because dividing the increases in recordsimilarity by the singing rate (Fig. 4e) gave similar results, with largeincreases in record similarity from 3 to 6 h and very low increasesduring the second half of the day. Once syllable structure reached anasymptote at midday, record similarity did not increase muchfurther, despite continued singing.

Causes of post-sleep deteriorationThe underlying cause of the observed effects might be sleep, butthere are alternative hypotheses. First, sleep inertia, a transitionalstate of lowered arousal34, and possibly altered states of the vocalapparatus (‘morning voice’), could explain the deterioration in songstructure. Alternatively, vocal changes might be caused by circadianchanges in hormonal state. To test both hypotheses, we preventedthe bird from singing for 2 h during one morning (n ! 6, range:50–57 days) by taking the cage from the sound box into thelaboratory35. The bird did not attempt to sing but exhibited usualactivities including eating and calling. When singing resumed (3–4 h after lights were turned on) we would have expected to see at

Figure 4 Progression of song learning. a, Relation between magnitude of post-sleepdeterioration and similarity to the song model. b, An illustration of similaritymeasurements during six consecutive days. Grey dots represent the raw similarity scores

and the blue curve shows the 95th centile. c, An illustration of record similaritymeasurements during one day. Red segments illustrate record similarity values and the

blue curve shows the 95th centile. d, Gain in record similarity to the song model during theday across birds (means ^ s.e.m.). Similarity scores were computed continuously from

day 55 to day 70 in 12 birds. e, As in d, but divided by the number of songs produced ineach 1-h time bin (means ^ s.e.m.).

Figure 5 Vocal changes after song prevention and sleep manipulation. a, Recovery ofsong structure could be circadian (scenario 1) or it might require singing activity (scenario

2). b, Vocal changes (in reference to the developmental trend) were measured after night-sleep plus 2 h of song prevention (yellow bar, mean ^ s.e.m.). Results were similar to

those of post-sleep deterioration alone (grey bars, n.s., n ! 6 birds). c, If deterioration ofsong structure during sleep were due to lack of practice, 8 h of song prevention should

cause deterioration (scenario 1) or the bird might continue the morning recovery when

singing resumed (scenario 2). d, Vocal changes during 8 h of song prevention (yellow bar)

were similar to morning changes (from 08:00 to 10:00, blue bar, n.s.) and differed from

vocal changes during night-sleep (grey bars, P , 0.02, n ! 9 birds). e, Vocal changesafter melatonin-induced sleep (n ! 6 birds). Results in b–e are means and s.e.m.

articles


!

• Eventual similarity∝ Magnitude of daily oscillation× Total developmental change

• Morning intense singing guides strong improvements in similarity.

• Morning less structured (more plastic) sounds gives an opportunity to explore learnability.

ρ=0.6, P=0.038

Record similarity (the best 95th centile so far)

95th centile


Causes of Post-sleep Deterioration

not correlate with better imitation (Spearman r ! 0.25, P ! 0.443),further confirming that the initial acoustic distance from the songmodel does not predict song learning17. In sum, the magnitude ofdaily oscillations in song structure, but not the total developmentalchange, seems to be a predictor of accurate song learning. A similartrend was observed in birds trained from day 90 (SupplementaryFig. 1).We propose that the morning period of less structured (and

perhaps more plastic) sounds gives the bird an opportunity toexplore its vocal abilities and improve imitation. To examine thedynamic relations between daily changes in song structure and theprogression of vocal learning, we selected the most complex syllableof each bird’s song and computed similarity to the model syllablecontinuously (for each syllable produced) from day 55 to day 70after hatching. As expected, similarity values dropped during night-sleep (repeated-measures ANOVA, P , 0.001) and then increasedto a plateau (Fig. 4b). To measure a bird’s overall progress towardsimitating the tutor song, we defined a measure called ‘recordsimilarity’, the quality of a bird’s best success so far in copying thetutor song (Fig. 4b, c). For robustness, we computed the 95th centileof the similarities within a 1-h moving window. The recordsimilarity at a given time is defined as the best 95th centile achievedso far. Improvements in record similarity are computed withreference to developmental time, not to the previous hour, thusexcluding improvements that are due solely to recovery (Fig. 4b, c).Strong improvements in record similarity were achieved during

the third hour of the photophase (Fig. 4d), overlapping with the latephase (third hour) of the recovery of song structure (Fig. 3h, redcurve). Little improvement was achieved during the second half ofthe day. This pattern was not a consequence of different rates ofsinging across the day, because dividing the increases in recordsimilarity by the singing rate (Fig. 4e) gave similar results, with largeincreases in record similarity from 3 to 6 h and very low increasesduring the second half of the day. Once syllable structure reached anasymptote at midday, record similarity did not increase muchfurther, despite continued singing.

Causes of post-sleep deteriorationThe underlying cause of the observed effects might be sleep, butthere are alternative hypotheses. First, sleep inertia, a transitionalstate of lowered arousal34, and possibly altered states of the vocalapparatus (‘morning voice’), could explain the deterioration in songstructure. Alternatively, vocal changes might be caused by circadianchanges in hormonal state. To test both hypotheses, we preventedthe bird from singing for 2 h during one morning (n ! 6, range:50–57 days) by taking the cage from the sound box into thelaboratory35. The bird did not attempt to sing but exhibited usualactivities including eating and calling. When singing resumed (3–4 h after lights were turned on) we would have expected to see at

Figure 4 Progression of song learning. a, Relation between magnitude of post-sleepdeterioration and similarity to the song model. b, An illustration of similaritymeasurements during six consecutive days. Grey dots represent the raw similarity scores

and the blue curve shows the 95th centile. c, An illustration of record similaritymeasurements during one day. Red segments illustrate record similarity values and the

blue curve shows the 95th centile. d, Gain in record similarity to the song model during theday across birds (means ^ s.e.m.). Similarity scores were computed continuously from

day 55 to day 70 in 12 birds. e, As in d, but divided by the number of songs produced ineach 1-h time bin (means ^ s.e.m.).

Figure 5 Vocal changes after song prevention and sleep manipulation. a, Recovery ofsong structure could be circadian (scenario 1) or it might require singing activity (scenario

2). b, Vocal changes (in reference to the developmental trend) were measured after night-sleep plus 2 h of song prevention (yellow bar, mean ^ s.e.m.). Results were similar to

those of post-sleep deterioration alone (grey bars, n.s., n ! 6 birds). c, If deterioration ofsong structure during sleep were due to lack of practice, 8 h of song prevention should

cause deterioration (scenario 1) or the bird might continue the morning recovery when

singing resumed (scenario 2). d, Vocal changes during 8 h of song prevention (yellow bar)

were similar to morning changes (from 08:00 to 10:00, blue bar, n.s.) and differed from

vocal changes during night-sleep (grey bars, P , 0.02, n ! 9 birds). e, Vocal changesafter melatonin-induced sleep (n ! 6 birds). Results in b–e are means and s.e.m.

articles


!

Sleep inertia?Circadian changes?

Lack of practice?

Sleep-related activity?e.g., Neural song replay without auditory feedback

least some recovery of song structure if the cause had been sleepinertia or circadian hormonal state (Fig. 5a), but song deteriorationremained similar to that observed the day before (Fig. 5b,Wilcoxon,n.s.). Thus, post-sleep deterioration is not simply a circadianphenomenon and cannot be explained by sleep inertia.

Second, it could be that maintaining song structure during daysof rapid learning requires intense practice, and that the lack ofsinging practice during sleep causes deterioration in motor per-formance. To test this hypothesis, we raised birds under a lightregime of 16 h light and 8 h of darkness (16 h:8 h LD) to allowmatching of the 8-h night-sleep with an 8-h song preventioninterval during the photophase (n ! 9, range 50–57 days). Wemonitored song development on-line to detect the time of rapidlearning, and confirmed that post-sleep deterioration took placeafter 8 h of darkness. In the morning we allowed the bird to partlyrecover its song structure for 2 h. Then we prevented singing for 8 h;finally we allowed the bird to restart singing during the evening ofthe same day.When singing resumed, we would have expected to seedeterioration if lack of practice had an effect, and continuedimprovement otherwise (as would be expected when comparingthe second to the third hour of singing in the morning; Fig. 5c). Asshown, vocal changes continued in a positive direction after the 8 hof song prevention (Fig. 5d). The difference between the effects ofnight-sleep and those of song preventionwas statistically significant(Wilcoxon, P ! 0.02).

Finally, we tested for causal relations between sleep and deterio-ration of song structure by inducing sleep during the day.Melatonincan induce sleep in zebra finches, and electrophysiological record-ings suggest that song replay occurs readily in this induced sleepstate27. We allowed each bird (n ! 6, age 54–63 days) to recover itssong structure for 4 h after night-sleep and then injected 3mg ofmelatonin27. All birds fell asleep within 15min, and slept for 2–3 hduring 4 ^ 0.5 h of vocal rest. When singing resumed we observedsong deterioration in all birds, comparable to that observed afternight-sleep (Fig. 5e, Wilcoxon, n.s.).

Vocal changes during sleep could relate to synaptic21 or cellu-lar36,37 remodelling that might occur during sleep. Regardless of thespecific mechanism, our results indicate the involvement of anactive process, perhaps neural song-replay during sleep. How mightthe bird’s ‘internal singing’26,27 during sleep give rise to the observedplasticity? If the lack of auditory feedback during replay during sleephas a similar effect to that of abolishing32,33,38 or perturbing39

auditory feedback, we would expect to see drifts in song structureafter sleep replay. We analysed post-deafening deterioration in eightadult birds and found that the deterioration of intra-syllabictemporal structure was similar to that observed after night-sleepin young birds (Fig. 6). Therefore, song replay along with lack ofauditory feedback could by itself explain the decrease in song

structure after prolonged sleep. If feedback-guided singing neverresumes (that is, after deafening), this effect is detrimental, but inthe intact bird, transitory drifts in song structure could promotelearning.

DiscussionMany procedural and declarative learning skills are enhanced aftersleep1–11 with no apparent oscillations in performance. Oscillationsin performance have been reported in the behaviour of severalspecies, but possible relations to sleep were not investigated. Forexample, bumblebees can learn to approach and probe unfamiliarflowers faster as they gain experience during a foraging day, butthese skills deteriorate strongly overnight40. Another example is thedevelopment of voluntary movement in vertebrates. Movementpatterns appear gradually in the newborn infant, progressingfrom simple to complex, but after prolonged immobility, move-ments become simple and primitive41 (‘warm up’ phenomenon). Asimilar effect was observed in the development of exploratorybehaviour in the juvenile rat42.Our study bridges these past observations by demonstrating a

sleep-associated oscillation in performance that has a role indevelopmental learning. A potential reason why such a phenom-enon has not been reported before is that the effect is specific to (orstronger in) developmental learning. A second reason for this mightbe the requirement for exhaustive, long-term behavioural measure-ments spanning many days to establish the effect.We advance some speculations about the utility of these sleep-

related oscillations in performance. Our speculations are based onthe observation that a larger oscillation is associated with better finalimitation. The oscillations could reflect a compromise betweenplasticity and consolidation of structure: periodic increases inplasticity might allow improvement of similarity to the songmodel through the correction of inappropriately consolidatedstructure.Finally, it is worth noting that certain optimization algorithms

such as simulated tempering43 use non-monotonic trajectories ofparameters. The ideas of simulated annealing44 and tempering43

come from metallurgical processes such as the reduction of brittle-ness of steel through tempering, which involves a cyclical process ofreheating and cooling. A

MethodsAnimal careAll experiments were performed in accordancewith guidelines of the National Institutes ofHealth and have been reviewed and approved by the Institutional Animal Care and UseCommittee of the City College of New York and the Wesleyan University.

Experimental designWe used 50 zebra finches (Taenyopygia guttata) from the City College of New Yorkbreeding colony. Colony management and experimental design have been describedpreviously17. All birds were kept in isolation from days 30 to 90 after hatching.

Recordings from deafened birds were obtained from eight birds from the breedingcolony of the Wesleyan University. These birds were kept in a large aviary with theirparents and siblings of both sexes until days 70–90, when they were removed and housedtogether in groups of two to five.

Experimental groupsTraining from day 43 to day 90: 12 birds from 8 families were trained with songplaybacks17, starting from day 43 after hatching. We used three different songs (four birdsper song model). Birds were raised from hatching under an artificial photoperiod of12 h:12 h LD.

Training fromday 60 to day 90: six birds from three different families were trained fromday 60 with the three song models used in the first group (12 h:12 h LD).

Training from day 90 to day 120: eight birds from seven different families were trainedfrom day 90 with the three song models used in the first group (12 h:12 h LD). Birds stayedin acoustic isolation until day 120.

No training: six birds from four different families were kept in similar sound boxesfrom day 30 to day 90 but were not trained (12 h:12 h LD).

Singing prevention: in 15 birds from 10 families, photoperiod was shifted from12 h:12 h LD (days 0–30) to 16 h:8 h LD thereafter. Birds were trained from day 43 today 90 (five birds per song model). We prevented singing once in each bird by taking thebird (with its cage) out of the training box and placing it in the laboratory under

Figure 6 Comparison of post-sleep deterioration and post-deafening deterioration.a, Post-sleep changes in birds trained from day 43. b, Post-deafening changes (2–4weeks after deafening). Each slice indicates the mean relative contribution of each feature

to the overall effect. The sign indicates whether the feature value increased or decreased.

Var, variance.

articles


!

Post-sleep Post-deafening


Conclusions and Discussions

• Sleep-related oscillation in developmental learning

• Post-sleep deterioration vs. morning intense practice

• Oscillation in behavioral performance

• Bumblebees, juvenile rats, newborn babies

• Not observed in many other procedural and declarative learning skills

• Optimization algorithm using non-monotonic parameter trajectories e.g., Simulated tempering: Heating vs. reheating


Further Reading

LETTERS

Sleep and sensorimotor integration during early vocallearning in a songbirdSylvan S. Shank1 & Daniel Margoliash1,2

Behavioural studies widely implicate sleep in memory consolida-tion in the learning of a broad range of behaviours1–4. Duringsleep, brain regions are reactivated5,6, and specific patterns ofneural activity are replayed7–10, consistent with patterns observedin previous waking behaviour. Birdsong learning is a paradigmaticmodel system for skill learning11–14. Song development in juvenilezebra finches (Taeniopygia guttata) is characterized by sleep-dependent circadian fluctuations in singing behaviour, withimmediate post-sleep deterioration in song structure followedby recovery later in the day15. In sleeping adult birds, spontaneousbursting activity of forebrain premotor neurons in the robustnucleus of the arcopallium (RA) carries information about day-time singing16. Here we show that, in juvenile zebra finches, play-back during the day of an adult ‘tutor’ song induced profound andtutor-song-specific changes in bursting activity of RA neuronsduring the following night of sleep. The night-time neuronalchanges preceded tutor-song-induced changes in singing, firstobserved the following day. Interruption of auditory feedbackgreatly reduced sleep bursting and prevented the tutor-song-specific neuronal remodelling. Thus, night-time neuronal activityis shaped by the interaction of the song model (sensory template)and auditory feedback, with changes in night-time activity preced-ing the onset of practice associated with vocal learning. We hypo-thesize that night-time bursting induces adaptive changes inpremotor networks during sleep as part of vocal learning. By thishypothesis, adaptive changes driven by replay of sensory informa-tion at night and by evaluation of sensory feedback during the dayinteract to produce the complex circadian patterns seen early invocal development.

To explore the role of sleep in the early phases of song learning, wecharacterized the properties of single RA neurons in head-fixed,sleeping juvenile zebra finches during nights before and after theonset of tutor song exposure. Sleep was defined on the basis of beha-vioural and electroencephalographic criteria (see SupplementaryInformation)17. During sleep, RA neurons tended to discharge irre-gularly or ‘burst’, as seen in distributions of inter-spike intervals(ISIs; Fig. 1a, black curve)16. Starting on the night after the first dayof exposure to the tutor song, there was a sharp increase in theamount of high-frequency spiking activity (Fig. 1a, red curve).Across all birds, we quantified the effect of tutor song exposure as anormalized change in the percentage of ISIs#40ms, which showed asignificant increase starting on the night after the first day of tutorsong exposure and persisted thereafter (Fig. 1b). A significant resultwas also obtained considering firing rates that were normalized bylinearly scaling the ISIs for each cell (P5 0.02, repeated measuresanalysis of variance (ANOVA), a5 0.05; Fig. 1b, open circles). Thisverified that the increase in high-frequency activity was not depen-dent on changes in mean spike rates after tutoring, but was the resultof a specific increase in high-frequency activity including bursting.

Within each bird there was some variation in the amount of high-frequency activity of RA cells on nights after the onset of song learn-ing, but the tendency towards shorter ISIs was apparent in most cells(Fig. 1c).

Emerging RA bursting activity, furthermore, was shaped by thespecific tutor song that a bird heard. Nightly mean ISI distributionswere calculated for all RA neurons recorded for each bird after tutorsong exposure (which showed little difference from night to night;Fig. 2c), and nightly mean distributions were averaged together togenerate one mean curve per bird. For the resulting post-exposurecurves, within the high-frequency range (ISIs#40ms), the shapes—as assessed using Pearson correlation coefficients—were more sim-ilar in birds hearing the same tutor song than in birds hearing dif-ferent tutor songs (see Supplementary Information). This groupingof ISI distribution shapes by tutor song can be visualized by compar-ing the average ISI distributions for individual birds on nights beforetutor song exposure to those on nights after tutor song exposure(Fig. 2a, black and coloured lines, respectively). The differencesbetween groups can be visualized by comparing global average ISIdistributions—one for each group of birds hearing a given tutor song(Fig. 2b).

Once a bird was exposed to a tutor song, a prototypical post-exposure ISI distribution shape was quickly obtained and thenmain-tained. To quantify this, we compared (using Pearson correlations)the nightly ISI distributions (#40ms) for each bird before and aftersong exposure to the corresponding global mean curve (Fig. 2b),excluding data from the bird being analysed from the global meandistributions. Before tutor song exposure, both the within- andbetween-group comparisons (Fig. 2d, black and grey dots, respect-ively) had large variability and were not significantly different fromeach other on any night (P5 0.25 to P5 0.73). By the first night aftertutor song exposure, ISI distributions had already assumed theirpost-exposure shape, showing increased Pearson correlations withmuch lower variability (Fig. 2d, red dots), whereas the across-groupcorrelations did not increase (Fig. 2d, pink dots). These differenceswere statistically significant starting from the first night (P5 0.0082,two-tailed t-test, a5 0.05), and continued to show a significant dif-ference throughout the post-exposure period (P, 0.05 for 8 out of11 post-exposure nights; see also Supplementary Fig. 2), emphasizingthat stable changes in ISI distributions were rapidly achieved andthen maintained in the days after tutor song exposure.

Zebra finches begin singing as early as 25 days of age18, so our,40-day-old birds had extensive sensorimotor experience before neuralrecordings were performed. To explore the influence of singing andauditory feedback on the structure of RA neuronal bursting, weperformed two additional experiments. We prevented singing bysurgically muting two birds (a third bird, M2, sang in spite of thesurgery), and raised a second group of four birds in continuous100 dB white noise environments to suppress auditory feedback.

1Department of Psychology, 2Department of Organismal Biology and Anatomy, University of Chicago, Chicago, Illinois 60637, USA.

Vol 458 |5 March 2009 |doi:10.1038/nature07615

73 Macmillan Publishers Limited. All rights reserved©2009


Documents

How sleep affects the developmental learning of bird song