By Lina Sorg
Motor patterns are shaped at many different spatial and temporal scales, each of which contains its own ongoing source of variability. A changing neural substrate—the part of the central nervous system that underlies a specific behavior, cognitive process, or psychological state—is responsible for stable learned motor behaviors, though such behaviors can also result from macroscopic changes in the environment or alterations to physical structure (i.e., injury or aging). “Maintaining stable behaviors requires robustness to variations at both very small and very large scales,” Alison Duffy of the University of Washington said. During a minisymposium at the 2022 SIAM Conference on the Life Sciences, which is currently taking place in Pittsburgh, Pa., in conjunction with the 2022 SIAM Annual Meeting, Duffy studied the learning and maintenance of song behavior in adult zebra finches given neural perturbation.
Duffy seeks to understand the way in which motor sequences are learned and maintained over time. “We learn many motor sequences during development and even later, then maintain them over long periods of time despite physical or environmental change,” she said. “This might seem simple on the surface because you are repeating the same thing over and over. But when you think about how this might be built and maintained, it opens up some interesting challenges.”
The zebra finch mating song is an example of a learned motor behavior that spans numerous spatial and temporal scales; it even serves as a canonical example of trial-and-error learning in neurobiology. “Studying this behavior also gives us an opportunity to connect computational learning algorithms and theories of learning with a biological system,” Duffy said.
Duffy began with an overview of the infant zebra finch’s learning process. Immediately after hatching, baby birds listen to and memorize an adult male bird’s song. A series of nuclei in the primary motor pathway produce this song, and spiking in the pathway is highly correlated with song activity. Shortly after birth, the juveniles partake in a practice period and produce highly variable babbling vocalizations — much like those of human infants. Over time, the vocalizations increasingly begin to resemble the adult teacher’s song. Exposure to the teacher’s song—as well as auditory feedback from the baby birds’ own practice sessions—are essential to the learning process.
Figure 1. Process of Alison Duffy's experimental collaborators to silence 50 percent of the excitatory projection neurons in adult zebra finches.
Local mechanisms of motor sequence stability occur in HVC, a designated region in the primary motor pathway of the songbird brain. “The HVC nucleus acts like the clock of the song,” Duffy said. The song is strong on the precise sparse timing sequence of spiking activity that emerges from HVC, and each burst serves as a unique time stamp. Even during neurogenesis—the process by which new neurons form in the brain—most adult finches’ songs remain completely unchanged.
Armed with this knowledge, Duffy posed the following question: How does HVC generate robust representations of sequence timing despite large-scale perturbations? In other words, how do neural circuits generate stable behavior with fluctuating components? To explore these queries, Duffy’s experimental collaborators performed a cell-specific perturbation on adult finches. They injected a virus into HVC that silenced 50 percent of the excitatory projection neurons in the circuit, thus triggering a degradation in both the timing and acoustic structure over the course of the perturbation (see Figure 1).
Prior to perturbation, certain song components are highly repeatable across all renditions. Precise portions of the song completely disappear upon perturbation, only to gradually reform and tighten with time. Different portions of the song return at different times. By the 32nd day after perturbation, the restored song trajectory closely resembles the original form, with several minor alterations (see Figure 2).
Figure 2. Visualization of the zebra finches' recovered song trajectory on the first day of perturbation and 32 days thereafter. By day 32, the restored trajectory closely resembles the original form.
After quantifying the distortion of discrete syllables and parametrizing their population in a high-dimensional acoustic feature space, Duffy presented a trajectory over days of the closest syllable distances to the original song for a single bird. Syllables are nearly fully restored after roughly 15 days. “This is a remarkedly rapid recovery in the behavior, given that approximately 50 percent of the network responsible for generating the timing of the song has been removed,” Duffy said.
Over the course of the recovery that follows song collapse, excitatory synaptic inputs onto the silenced projection cells within HVC triple; the cells’ lack of participation clearly elicits a strong response within the network. In addition, the synaptic connections onto the unperturbed projection cells double. “This is a pretty remarkable increase onto the connections of the unperturbed portion of the circuit, considering that originally the perturbation removed half of the preexisting connections,” Duffy said.
Naturally, she then wondered about the mechanisms that contribute to song recovery and circuit reorganization. To investigate whether recovery depends on practice (much like the initial trial-and-error learning period), Duffy’s collaborators repeated the first experiment but prevented the birds from singing for 10 days after the perturbation. “We found that significant song recovery takes place, even with minimal practice,” Duffy said. She next employed computational modeling to investigate local plasticity mechanisms’ ability to restore dynamics within a model of HVC, paying particular attention to the way in which the mechanisms reconfigure network architecture after massive cell loss.
After introducing local plasticity mechanisms to interact with the model’s dynamics, Duffy considered the impact of synaptic scaling via cell-autonomous firing-rate homeostasis. “This form of homeostasis preserves overall levels of activity without altering the pattern of synaptic connectivity into a cell,” she said. “It has been shown in other theoretical models to act as a counterbalance to spike-timing-dependent plasticity.” She found that silencing 50 percent of the projection neurons destroys activity in the network, but cell-autonomous synaptic scaling can rebuild the sequential dynamics after many renditions. When looking at the synaptic weights on individual cells, the network reorganization returns to baseline (unlike the doubling of synaptic strength that occurs in the experimental results). Finally, Duffy postulated a modulation of the single-cell scaling rule to achieve a stable population level firing rate, ultimately allowing the output to fully recover.
In summation, the adult zebra finch song recovers rapidly after massive perturbation to HVC, even with minimal practice. As evidenced in the computational model, local plasticity mechanisms can restore dynamics in HVC after 50 percent projection cell loss. “Our modeling work finds a more complete recovery and aligns with experiment by employing a population-level firing rate homeostasis rule in HVC,” Duffy said. “This rule makes use of silent cells and can shift dynamics onto new subgroups of neurons, which is another form of resilience that might exist in the true system.”
Acknowledgments: Carlos Lois, Zsofia Torok, and Bo Wang conducted the experimental work in the lab of Carlos Lois at the California Institute of Technology. Alison Duffy, David Bell, and Adrienne Fairhall completed the modeling and analysis work in the lab of Adrienne Fairhall at the University of Washington.
|
Lina Sorg is the managing editor of SIAM News. |