CRCNS: Identifying principles of auditory cortical organization with machine learning

The human auditory system transforms incoming acoustic information into distinct auditory “objects” that can be interpreted, localized, and integrated with information from the other senses. A human listener can resolve, for example, a monophonic musical recording into piano and guitar, or parse speech audio into a sequence of words. We do not currently understand how this is achieved, nor how transformations in sound processing across cortical regions contribute, especially beyond primary auditory cortex. On the other hand, we now have available another, more easily investigated system that—as of this last decade—solves such problems at human performance levels: the artificial neural network (ANN). Although crude as biophysical models, ANNs strongly resemble biological neural networks in terms of computation and representation. We propose to compare single-unit electrophysiology in macaque auditory cortex with state-of-the-art ANNs trained to solve ecologically relevant auditory tasks. This combination allows us to track how transformations in perceptual representations are distributed and instantiated in the brain; to experiment with multiple ANN architectures and tasks to test hypotheses about why the representations take the forms we observe; and to refine iteratively our stimulus protocols and models. Our objectives are (1) to evaluate these ANNs as encoding models for neurons in macaque auditory cortex, recorded during auditory discrimination tasks; (2) to use the internal structure of ANNs to generate and test novel hypotheses about the topographical and hierarchical organization of non-primary auditory cortex; and (3) to demonstrate “stimulus- based control” of neurons throughout nonprimary auditory cortex: optimizing stimuli via the ANN and then playing to the animals in closed-loop neurophysiology experiments.