A Neural Ensemble Correlation Code for Sound Category Identification

Abstract

Humans and other animals effortlessly identify sounds and categorize them into behaviorally relevant categories. Yet, the acoustic features and neural transformations that enable the formation of perceptual categories are largely unknown. Here we demonstrate that correlation statistics between frequency-organized cochlear sound channels are reflected in the neural ensemble activity of the auditory midbrain and that such activity, in turn, can contribute to discrimination of perceptual categories. Using multi-channel neural recordings in the auditory midbrain of unanesthetized rabbits, we first demonstrate that neuron ensemble correlations are highly structured in both time and frequency and can be decoded to distinguish sounds. Next, we develop a probabilistic framework for measuring the nonstationary spectro-temporal correlation statistics between frequency organized channels in an auditory model. In a 13-category sound identification task, classification accuracy is consistently high (>80%), improving with sound duration and plateauing at ~ 1-3 seconds, mirroring human performance trends. Nonstationary short-term correlation statistics are more informative about the sound category than the time-average correlation statistics (84% vs. 73% accuracy). When tested independently, the spectral and temporal correlations between the model outputs achieved a similar level of performance and appear to contribute equally. These results outline a plausible neural code in which correlation statistics between neuron ensembles of different frequencies can be read-out to identify and distinguish acoustic categories.

A temporal integration mechanism enhances frequency selectivity of broadband inputs to inferior colliculus

“…Such nonlinear transformation is plausible, since ICNs can show multiple selectivities, including nonlinear or non–phase-locked response components, which don’t show up in the STRF and often differ from the linear components [20, 51, 52]. Functionally, differences between broadband and narrowband PBIs may be relevant for detecting transient sound elements and for detecting correlations between frequency channels, which are both relevant perceptually and for neural coding in the IC [20, 53, 54].…”

Magnetic Stimulation Allows Focal Activation of the Mouse Cochlea

“…Thus, data analysis was based on MUAs, which were quantified by an analog representation of multi-unit activity (aMUA). aMUA reflects the voltage signal power within the frequency range occupied by action potentials (56)(57)(58)(59)(60). This approach is advantageous over the more traditional measure of MUA based on thresholding and spikedetection since aMUAs are not biased by free parameters (e.g., threshold levels), and provide a high signal to noise ratio (59).…”

Neural Decoding of Inferior Colliculus Multiunit Activity for Sound Category identification with temporal correlation and deep learning

“…Areas of tonotopically organised frequency responses, varying from low to high frequencies with the depth of neuronal recording, are present in the main nucleus of the IC. This is called the tonotopic organisation of the recording area ([8]-Sadeghi et al, 2019). In the inferior colliculus (IC), sounds are broken down into higher order acoustic features such as temporal and spectral modulations.…”

“…Cathegories of stationary natural sounds with high-level acoustic structure, such as random sounds from bird chorus, crackling fire, running water, outdoor crowd, and rattling snake sounds were used in this study ([10]-Zhai et al, 2020). Collectively, data we want to use, suggest that stimulus-driven correlations between sets of neurons in the CI can provide an imprint for downstream neurons to recognise and classify natural sounds into categories ([8]-Sadeghi et al, 2019).…”

“…On the other hand, for a sound to be recognised by the brain, it must be heard for a certain amount of time. This duration is about 1 to 2 seconds ([8]-Sadeghi et al, 2019). The temporal summation of the messages transmitted by the different neurons will allow to code the intensity and duration of the stimulus ([6]-Richard et al, 2019).…”