Encoding, reconstruction and representation of natural vocal communication signals

Abstract

How complex, natural communication signals are represented in the activity (spiking) patterns of neural populations is not well understood. I will describe data from a series of experiments that examine the spatiotemporal pattern of song-evoked spiking in large populations of simultaneously recorded neurons in the secondary auditory cortices of European starlings (Sturnus vulgaris), a species of songbird. Single neurons in these regions display composite receptive fields that incorporate large numbers (a dozen or more) orthogonal spectro-temporal features matched to the acoustics of species typical songs. This observation implies that the independent spiking response of any single neuron at any point in time is somewhat ambiguous with respect to the stimulus. I will discuss how this ambiguity can be resolved at the population level by considering an alternative representation of underlying latent stimulus space and the coincident pattern of firing among small groups of neurons. I show how this approach allows for high-fidelity reconstruction of spectrographic representations of novel birdsong from spiking responses. I then show how these coincident patterns of spiking can be measured directly in arbitrarily large neural populations, without reference to receptive fields, and conclude by discussing the implications of this combinatorial population encoding for the nature of perceptual representation.