Feedforward processing of sounds in the primary auditory cortex

Neurons in the auditory system, up to the primary auditory cortex (A1), represent sounds by decomposing them into constituent frequencies. However, it is unclear what systematic computations are performed on these frequency-tuned receptive fields (RFs) to build more complex and selective RFs at higher stages of auditory processing. Therefore, we studied what computational principles may underlie feed-forward sound processing in the A1 of awake primates.;First, we characterized tone-responsive RFs of A1 neurons that may input to higher auditory processing stages. Contrary to previous data, we found that most A1 neurons were sharply and separably tuned to both frequency and sound level. Regardless of intensity, sounds were represented by neural populations with similar underlying tuning widths, response profiles and firing rates, resulting in a level-invariant representation of sounds in A1. These excitatory RFs were further shaped by broadly tuned suppression at side-band frequencies and persistent inhibition over a wide range of frequencies at loud levels. These neurons formed a highly selective filter, only responding to sparse, tonal and narrowband sounds.;Second, we investigated how these neural responses may be integrated to result in selectivity to more complex features. Many superficial layer neurons did not respond to pure tones, but responded robustly to composite features present in complex sounds such as marmoset vocalizations. These responses were explainable by the precise and nonlinear integration of two tone-tuned subunits with specific frequency and onset time differences between them. Over the population, such local, non-linear interactions could underlie a wide range of observed complex 'tuning' properties. These superficial layer neurons may form the subsequent layer of auditory processing.;Lastly, we asked how well such nonlinear receptive fields could perform in a biologically relevant task, such as classification of vocalization types, compared to template-match based methods. We found that such nonlinear RFs carried more information relevant to classification than pure-tone or noise based RFs. Importantly, these RFs approached the performance of maximally informative fragment-based RFs, suggesting that the observed nonlinear RFs are a viable and efficient way of representing sounds. These findings further our understanding of how sounds are processed by higher auditory cortical areas.