| Résumé |
The mammalian auditory system extracts features from the acoustic environment based on the responses of spatially distributed sets of neurons in the subcortical and primary cortical auditory structures. The characteristic responses of these neurons (linearly approximated by their spectro-temporal receptive fields, or STRFs) suggest that auditory representations are formed on the basis of a time, frequency, rate (temporal modulations) and scale (spectral modulations) analysis of sound. However, how these four dimensions are integrated and processed in subsequent neural networks remains unclear. In this work, we present a new methodology to generate computational insights into the functional organization of such processes. We first propose a systematic framework to explore more than a hundred different computational strategies to process the output of a generic STRF model. We then evaluate these strategies on their ability to compute perceptual distances between pairs of environmental sounds. Finally, we conduct a meta-analysis of the dataset of all these algorithms’ accuracies to examine whether certain combinations of dimensions and certain ways to treat such dimensions are, on the whole, more computationally effective than others. We present an application of this methodology to a dataset of ten environmental sound categories, in which the analysis reveals that (1) models are most effective when they organize STRF data into frequency groupings - which is consistent with the known tonotopic organisation of receptive fields in A1 -, and that (2) models that treat STRF data as time series are no more effective than models that rely only on summary statistics along time - which corroborates recent experimental evidence on texture discrimination by summary statistics. |
