Predictive coding for natural vocal signals in the songbird auditory forebrain

Predictive coding posits that sensory signals are compared to internal models, with resulting prediction-error carried in the spiking responses of single neurons. Despite its proposal as a general cortical mechanism, including for speech processing, whether or how predictive coding functions in single-neuron responses to vocal communication signals is unknown. As a proxy internal model, we developed a neural network that uses current sensory context to predict future spectrotemporal features of a vocal communication signal, birdsong. We then represent birdsong as either weighted sets of latent predictive features evolving in time, or as time-varying prediction-errors that reflect the difference between ongoing network-predicted and actual song. Using these spectrotemporal, predictive, and prediction-error song representations, we fit linear/non-linear receptive fields to single neuron responses recorded from caudomedial nidopallium (NCM), caudal mesopallium (CMM) and Field L, analogs of mammalian auditory cortices, in anesthetized European starlings, Sturnus vulgaris, listening to conspecific songs. In all three regions, the predictive features of song yield the single best model of song-evoked spiking responses, but unique information about all three representations (signal, prediction, and error) is carried in the spiking responses to song. The relative weighting of this information varies across regions, but in contrast to many computational predictive coding models neither predictive nor error responses are segregated in separate neurons. The continuous interplay between prediction and prediction-error is consistent with the relevance of predictive coding for cortical processing of temporally patterned vocal communication signals, but new models for how prediction and error are integrated in single neurons are required.