Expectation-modulated temporal dynamics in a sensory neural population during behavior.

Real-world perception is thought to arise from the coordinated activity of neuronal populations reflecting a combination of sensory inputs, existing knowledge, and contextually modulated expectations. We hypothesize that this integration is realized in the temporal dynamics of perceptual population responses. Here, we use latent-space modeling and dynamical systems analysis to characterize activity across populations of simultaneously recorded neurons in a secondary auditory forebrain region of songbirds (European starlings), engaged in an expectation-cued song-syllable categorization task. We show that during perception, population spiking responses are organized into well-structured trajectories through low dimensional sub-spaces, characteristic of attractor dynamics and carrying information about stimulus identity and behavioral response. Compared to passive listening, population trajectories during active task engagement occupy a significantly greater portion of the underlying subspace, leading to enhanced decodability of categorization behavior across different contexts. On trials where expectations for the song syllable are not violated, the population trajectory is maximally informative: variance relative to these trajectories predicts response accuracy in other contexts, and other trajectories following uninformative cues eventually converge onto validly cued trajectories. Expectation overall increases the speed with which the population trajectory can be accurately decoded. Collectively, the results reveal a strong link between the temporal dynamics of sensory populations, perception, and behavior.