Neural Mechanisms of Sequence Generation and Learning
Introduction
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 brain.
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. |
|
