Data-driven modelling of rodent auditory processing of pure tones as low-order polynomial encodings of spectral features*

Hearing is a key biological function to perceive and interact with the environment. While we understand the basic structure of biological auditory systems —responsible for processing sound—, the computations these substructures perform remain considerably less understood.

At the level of a single auditory cell, mapping its response to sounds of varying frequencies and intensities produces a receptive field known as the frequency response area (FRA) of the cell. This common encoding description is known to show characteristic nonlinear shapes, which are often relatively simple for responses to simple sounds like pure tones, even in higher-order neurons receiving inputs several steps removed from the physical sound waves. However, these FRAs has not yet been modelled computationally.

Our research here is based on spike-sorted tetrode recordings of mouse primary auditory cortex (A1), as well as guinea pig A1 and inferior colliculus presented with pure tone stimuli. We develop a novel, data-driven, and biologically-plausible model using the neural engineering framework, where a single-layer spiking neural network encodes a pure tone as a polynomial on the tone's frequency and intensity at the synaptic connections to an end-layer of spiking neurons, which produce the observed FRA shapes. Notably, low polynomial orders are able to approximate the responses of these neurons located downstream in the auditory hierarchy, which receive information that has already passed through and been processed at multiple points, and higher polynomial orders lead to overfitting. We also argue that it is not possible to reproduce the nonlinearities observed in the FRAs with a linear encoding model, regardless of the nonlinearity implemented by the final neuronal integration.

 

Our work here might inform future modelling, simulations, and algorithms of sound processing, as well as auditory brain-computer interfacing.