VVTNS seventh Season Opening Lecture

The Algorithm(s) of Birdsong Learning

How do songbirds acquire and refine their complex vocal repertoires? Learning a song requires both evaluating sensory evidence and determining which changes in behavior improve performance. These challenges have traditionally been studied in separate experimental paradigms: auditory-feedback adaptation, reinforcement learning, and imitation. In this talk, I will examine the behavioral and neural algorithms underlying each and ask how they might fit within a common learning architecture. First, when auditory feedback is shifted in pitch, birds change their vocal pitch in the opposite direction. As in humans, compensation is typically incomplete and becomes proportionally smaller for larger shifts. This pattern is consistent with Bayesian causal inference: birds combine internal predictions of pitch with auditory evidence while discounting feedback that is unlikely to have been self-generated. Second, I will show why classical policy-gradient reinforcement learning falls short in explaining the nonlinear pitch trajectories observed in conditioning experiments. A normative extension accounts for these trajectories and aligns with the modular organization of the brain’s song-control system. In this architecture, a learning component generates exploratory variation and is updated by reward. Other components contribute to motor output but are not updated by reward on every trial. Third, I will turn to imitation learning. Songbirds simplify this computational challenge by acquiring vocal units—syllables—independently of their sequential order. The intrinsic reward guiding the acquisition of this vocal inventory can be characterized by a simple sum–max rule: each target syllable contributes a partial reward determined by its best pitch match among the bird’s developing vocalizations, and these rewards are summed across targets. A light-tailed pitch-similarity function ensures that large mismatches contribute little reward. Together, these findings identify complementary mechanisms for interpreting sensory feedback, assigning credit to motor actions, and matching vocal output to remembered targets. The next challenge is to explain how these mechanisms interact within a unified theory of birdsong learning. I will conclude with initial attempts to use AI to develop this synthesis, with the aim of turning a collection of task-specific algorithms into a coherent framework with testable predictions and potential relevance beyond the avian brain.