Exploration and skill in songbird motor learning

PhD work with Michale Fee at MIT

Most of what people and animals do well is learned, from speech to a tennis serve. Learning a motor skill begins with variable, exploratory behavior and ends with precise, stereotyped performance. Zebra finches learn their song the same way: a juvenile babbles, and an adult sings one song with millisecond precision. Because distinct brain regions underlie the two modes, the bird lets each be studied on its own. For my PhD, I recorded the premotor activity of neural ensembles at both ends of song learning.

HVC bursts tile song continuously, with a syllable-locked rhythm

In mammalian motor cortex, a long-running debate asks whether neurons encode features of movements or carry intrinsic dynamics that pattern them. The same debate arose for HVC, the songbird premotor nucleus whose projection neurons each burst sparsely at one moment in the song. One model holds that the bursts form a continuous sequence, like a clock; another holds that they cluster around vocal gestures.

In a large dataset of HVC neurons recorded in singing birds, projection neuron bursts were spread continuously and nearly uniformly through song, as the clock model predicts. Firing rates also carried a 10 Hz rhythm locked to syllables, peaking before syllable onsets and suppressed before offsets (Lynch, Okubo et al. 2016, Neuron).

Testing how LMAN generates variability took new tools

Exploration drives reinforcement learning, and in songbirds the nucleus LMAN generates song variability and injects it into the motor pathway. Two models of how it does so, a balanced excitatory–inhibitory network and an excitable medium with propagating waves, predict opposite correlations between nearby neurons. Telling them apart required recording several neurons at once in a singing bird. For that, I built a multi-electrode microdrive with remotely positionable electrodes, which together with its headstage weighs under 1.6 g, and co-developed a lightweight head-mounted microscope for calcium imaging (Scherrer, Lynch et al. 2023, Nature Methods).