Figure 1.
Diagram of the neural control signals underlying birdsong generation. Note that this figure is a composite; HVC, RA, and the vocal muscles have not yet been studied simultaneously in a single individual bird.
- Top row: HVC encodes the sequence of sound elements in the song. Each row of the raster shows the activity profile of a single HVC neuron during 500 ms of song production (different colors represent different neurons; adapted from Hahnloser et al. 2002).
- Second row: Unique patterns of neural activity across the RA population map the sparse HVC code into continuous muscle control signals. Each row of the raster shows the activity profile of a single RA neuron during song production (different colors represent different neurons; adapted from Leonardo & Fee 2004).
- Third row: Vocal activity corresponds directly to the temporal structure of the song (adapted from Wild et al. 1998).
- Bottom row: Spectrogram of the bird's song. The y axis shows the frequencies of sound produced at each moment in the song.
[Click for enlargement] |
By Anthony Leonardo, Harvard University
We have all heard the songs of birds, heralding the approach of spring. As
beautiful as these songs are, there is an even greater beauty to be appreciated,
literally beneath the surface, in the elegant mechanisms used by the bird's
brain to generate the song. Just as the cover of a watch can be removed to
reveal a dazzling array of interlocking gears whose combined activity encodes
time, in this note I will attempt to sketch out the broad strokes of our
current understanding of song control, and reveal some of the beauty of how
birds sing.
It has been known for many years that songbirds have a specialized set of
brain nuclei dedicated to song learning and song production (Nottebohm et al.
1976). Although these two processes are intertwined, our understanding of song
production is most detailed through three stages of motor control (Figure 1).
The high-level control nucleus, HVC, encodes the unique sequence of sounds
that make up the bird's song (Hahnloser et al. 2002; Yu & Margoliash 1996).
Each HVC neuron is activated for 5–10 ms at a single time during the song
(Figure 1, top row; Hahnloser et al. 2002). So, for example, HVC neuron 1 is
associated with sound 1, HVC neuron 2 is associated with sound 2, etc. However,
HVC does not generate sound itself, nor does it even project to the motor
neurons that drive the sound-producing vocal muscles; in this sense, HVC really
only encodes the sequence, rather than the content, of the sound elements in
the song.
HVC lies at the top of the song generation hierarchy; at the bottom lies the
vocal muscles and the syrinx (the vocal organ). Contractions of these muscles
puts the syrinx into different dynamical states so that different sounds are
produced when the abdominal muscles force air out of the air sacs and through
the syrinx (Fee et al. 1998; Goller & Suthers 1996; Wild et al. 1998) (Figure
1, third row). The song contains some sound elements that change very rapidly,
in a few milliseconds, and other sound elements that change very slowly, over
hundreds of milliseconds (Figure 1, bottom row). Sounds that change quickly are
generally produced by fast changes in vocal muscles activity, and sounds that
change slowly are produced by fixed configurations of the vocal muscles. There
is thus a relatively direct correspondence between the time course of changes
in these sounds, and the time course of activation in the vocal muscles (Figure
1, third row).
The temporal structure in the song and the vocal muscle signals spans a wide
range of time scales, yet the HVC neurons that encode the song sequence are
active only briefly. This raises an interesting question: how is the sparse HVC
activity transformed into continuous patterns of vocal muscle activity?
Sequence-encoding neurons in HVC project to premotor neurons in nucleus RA;
these RA neurons connect directly to the motor neurons in the brainstem that
drive the vocal muscles (Vicario 1991). The central role of RA is to map the
temporal sequence onto particular acoustic patterns (Leonardo & Fee 2005). Like
HVC neurons, RA neurons are also active only briefly, for ~10 ms. Unlike HVC
neurons, however, RA neurons are active at multiple times in the song and
therefore provide little information about the temporal position in the song
sequence. Individual RA neurons thus encode neither song temporal structure nor
song sequence (Figure 1, second row).
It is at the level of RA population activity, rather than individual
neurons, that we begin to understand how neural activity is transformed into
sound. There is a large convergence between RA and the vocal muscles: roughly
1000 RA neurons (Gurney 1981) connect, via motor neurons in the brainstem, to
each vocal muscle (Greenewalt 1968). This is somewhat like a series of 1000
input pipes (RA neurons) connecting to a single output pipe (a vocal muscle).
Even if we turn the water of the individual input pipes on and off very quickly,
as long as one of them is always on at any instant in time, there will always
be a constant flow of water from the output pipe. Similarly, we find that
although individual RA neurons turn on and off very quickly, the population
activity, if combined appropriately, can be transformed into a continuous
muscle control signal (Leonardo & Fee 2005). The water pipe analogy highlights
another unusual aspect of the RA code. If we were to choose 100 input pipes at
random, and poured 1 cup of water into each simultaneously, we would get 100
cups of water from the output pipe—regardless of which input pipes we choose.
This property is called degeneracy: there is a many-to-one mapping from input
pipe combinations to the same output. The convergence from RA to the vocal
muscles creates a similar degeneracy, allowing many different patterns of RA
activity to generate the same sound (Leonardo & Fee 2005). We suspect this
degeneracy in the neural code may allow the bird to have more efficient song
learning—but this is a hypothesis that has yet to be verified.
Zebra finch song is generated by the concerted activity of many brain
areas, and by carefully studying the system as a whole one begins to appreciate
the beauty of the underlying mechanics. Just as the hands on a watch can seem
to move continuously, revealing no more than a whispered tick-tick-tick of the
gears within, birdsong, with its wide range of fast and slow temporal
structure, is generated by neural circuitry that runs on a single fast clock.
Each tick of this clock, encoded in an HVC-RA activity pattern, describes the
song at one moment in time. Although song itself reveals little of these
neural steps, they lie just below the surface waiting to be heard.
References
Fee M., Shraiman B, Pesaran B, Mitra P. 1998 The role of nonlinear dynamics
of the syrinx in birdsong production. Nature 395: 67–71. [Link]
Goller F, Suthers R. 1996. Role of syringeal muscles in controlling the
phonology of bird song. Journal of Neurophysiology 76: 287–300.
[Link]
Greenewalt C. 1968. Bird Song: Acoustics and Physiology.
Washington, D.C.: Smithonian Institution Press.
Gurney ME. 1981. Hormonal control of cell form and number in the zebra
finch song system. Journal of Neuroscience 1: 658–673.
[Link]
Hahnloser RHR, Kozhevnikov AA, Fee MS. 2002. An ultra-sparse code
underlies the generation of neural sequences in a songbird. Nature
419: 65–70.
[Link]
Leonardo A, Fee M. 2005. Ensemble coding of vocal control in a songbird.
Journal of Neuroscience: in press.
Nottebohm F, Stokes T, Leonard C. 1976. Central control of song in the
canary. Journal of Comparative Neurology 165: 457–486.
Vicario DS. 1991. Organization of the zebra finch song control system: II.
Functional organization of outputs from nucleus Robustus archistriatalis.
Journal of Comparative Neurology 309: 486–494.
Wild J, Goller F, Suthers R. 1998. Inspiratory muscle activity during bird
song. Journal of Neurobiology 36: 441–453.
[Link]
Yu AC, Margoliash D. 1996. Temporal hierarchical control of singing in
birds. Science 273: 1871–1875.
|
|
