Sensitivity to temporal sequence
is a striking and nearly universal aspect of brain function –
not only at a sensory level, but also at a motor and cognitive level.
The ability of the brain to step rapidly through a learned sequence
of states underlies not only the performance of complex motor tasks
such as speech, but perhaps our ability to think and plan as well.
Despite the fundamental significance of temporal ordering in animal
behavior, little is known about the biophysical and circuit mechanisms
underlying the generation, learning and detection of complex sequences
in the brain.
Animal vocalizations provide
a marvelous example of these phenomena, and we are using the
songbird as an experimental system to explore detailed models
of neural sequence generation. Most songbirds, such as the zebra
finch, produce a stereotyped pattern of acoustic signals with
structure and modulation over a wide range of time-scales, from milliseconds
to several seconds (Figure 1). Another remarkable aspect of this
behavior is that the specific acoustic pattern produced by a songbird
is learned, rather than being innately controlled: Vocalizations are
learned from the parents through a series of well-defined stages. Moreover,
avian brain areas involved in song learning are closely homologous
to mammalian brain areas involved in motor learning. Thus, the
song control system may have a great deal to teach us about general
principles of sequence generation and learning in the vertebrate
Figure 1 - Spectrogram of zebra finch song.
Figure 2 - Anatomy of brain areas involved in vocal motor production
(blue) and vocal learning (green).
Recent Projects in the Songbird
For the past six years, our laboratory
has focused on studying the cellular, circuit, and mechanical underpinnings
of songbird vocalizations. In one current project, we are studying
nucleus RA, an area that projects directly to motor neurons of the
vocal organ (Figure 2). During song, RA neurons each generate
a distinctive and reproducible sequence of brief bursts of spikes.
Using a new miniature motorized microdrive developed in this lab (see
below), we have been able to record from large populations of RA neurons
(~50) in the singing bird to understand how premotor activity maps to
vocal output (Figure 3).
Figure 3 - Spike raster plots of four RA neurons, time
aligned to the spectrogram of the generated song motif.
We are also investigating how the complex
sequence of bursts in RA is generated. The primary motor-related
input to RA comes from nucleus HVC; previous studies in singing
birds found that HVC neurons generate relatively stochastic, dense
patterns of spikes. Furthermore, the relationship of these firing
patterns to the song suggested that the temporal control of song
is organized hierarchically, such that HVC codes for motor patterns
on the timescale of song syllables (50-100 ms), and RA codes for
motor and acoustic patterns on a much briefer timescale (~10 ms). Thus,
the prevailing view in the songbird field has been that burst patterns
in RA are largely generated by circuitry intrinsic to RA.
One difficulty with these earlier
experiments is that HVC is known to contain at least three
classes of neurons: neurons that project to RA, neurons that
project to area X (a brain area involved in vocal learning), and local
interneurons. We were the first to characterize the firing patterns
of antidromically identified HVC neurons in the singing bird, and have
shown that RA-projecting HVC neurons are extremely sparsely active
- each generating at most a single brief burst of spikes precisely at
one time in the song motif. We also found that these neurons burst
sequentially, with each neuron bursting at a different time in the
song (Figure 5).
What is the causal relationship
between the sparse bursts of HVC(RA)neurons and the bursts in
downstream nucleus RA? Do bursts in HVC(RA) neurons drive bursts
in RA, and if so, is every burst in RA driven from HVC? We are examining
these questions in a new head-fixed sleeping bird preparation that
permits sophisticated electrophysiological experiments that would
be difficult or impossible to do in the freely behaving bird. Recent
observations have shown that during sleep, RA neurons generate spontaneous
high-frequency bursts, and that the sequence of sleep burst produced
by an individual RA neuron can closely resemble the burst sequence of
that neuron during singing. We have shown that HVC neurons similarly
replay brief snippets of highly sparse song-like sequences during
sleep. Furthermore, we have shown that during sleep, nearly every RA
burst is driven directly by a small population of HVC(RA). The similarity
of song- and sleep- related burst patterns suggest that song-related
bursts in RA are likewise directly driven from HVC.
Figure 4 - Extracellular recording of a single RA-projecting
neuron in nucleus HVC in the singing zebra finch.
Figure 5 - Spike raster plot of eight sequentially recorded
RA-projecting HVC neurons in one zebra finch, during singing.
Model of Sequence Generation in the Songbird
We have proposed a new model for premotor
temporal control in the song system. The ensemble of RA neurons
active at any moment is driven by a small population of RA-projecting
HVC neurons that is active at only one time in the RA sequence.
As a result, these HVC neurons code for some transient event
or ‘state’ in the sequence, rather than longer timescale structure
such as song syllables as in the temporal hierarchical model. We
suggest that HVC(RA) neurons code for temporal position within
the RA sequence. Put in terms of songbird behavior, the motor actions
underlying song unfold when the ‘time code’ in HVC acts upon the
‘muscle map’ in RA. The pattern of activation in RA depends on the
synaptic connections from HVC(RA) neurons onto RA neurons, suggesting
that it is the pattern of HVC(RA)–RA synapses that contains the
‘score’ of the bird’s song (Figure 6). In collaboration with
Sebastian Seung's group
, we are examining with theoretical techniques the implications of this
sparse coding for song learning and for premotor neuronal representations
in nucleus RA.
Figure 6 - Model of sequence generation in the songbird.