Listening in noise is a core problem in everyday hearing. Sound sources of interest routinely occur amid irrelevant distractors, as when you talk with someone in a bustling coffee shop. This background ?noise? distorts the pattern of spikes in the auditory nerve, often to a profound degree. Thus, to recognize sources of interest, the auditory system must somehow separate or suppress the effects of the background. Typical human hearing is remarkably noise-robust, but listeners with age-related hearing loss or other forms of impaired hearing struggle in noisy environments ? and are not much helped by contemporary hearing aids. Previous work on the neural basis of noise robustness has typically employed simple, synthetic noise sources, which lack the structure present in real-world sounds, and this work has focused on subcortical regions or on primary auditory cortex. Reasoning that real-world conditions might necessitate more complicated solutions, in the applicant's doctoral work, he considered everyday sources of noise, and leveraged the large-scale coverage afforded by fMRI to examine noise robustness throughout human auditory cortex. Real-world ?background noise? was operationalized as a natural sound with statistical properties that are stable over time (i.e., are stationary), conveying little new information about the world (e.g., swamp insects, an air conditioner, rain on pavement). The applicant measured fMRI responses in human listeners to a broad set of natural sounds presented in quiet, as well as embedded in the real-world background noises. Primary auditory cortical responses were substantially altered by the background, but non-primary responses were substantially more robust. This effect was not seen for simple synthetic backgrounds as had been used in previous work, suggesting that becoming robust to real-world background noises require different mechanisms. The applicant's thesis work demonstrates where noise invariance arises, but understanding how will require data with finer spatial and temporal resolution, and thus the proposed postdoctoral research will consist of training in single-unit electrophysiology using marmosets.
Aim 1 A builds on previous work examining single- unit noise robustness in artificial conditions, extending such work to real-world noise.
Aim 1 B leverages texture models to probe what aspects of real-world backgrounds disrupt the encoding of foregrounds.
Aim 2 A deploys linear reconstruction techniques to probe population representations.
Aim 2 B involves optimizing deep neural networks for noise invariance tasks, and using them as an encoding model to predict single-unit responses. Furthermore, such networks will be deployed as nonlinear decoding algorithms, reconstructing stimuli from neuronal populations. Throughout all aims, the work will characterize neuronal responses in non-primary areas, and in particular in parabelt, which is understudied in primates. The proposed work may enable improvements in hearing aid algorithms or neural prosthetics. Lastly, this training will lay the groundwork for the applicant's long-term goal of developing a marmoset model for hearing loss.
Typical human hearing is remarkably robust to the presence of background noise, but listeners with age- related hearing loss or other forms of impaired hearing struggle in noisy environments ? and often do not benefit from contemporary hearing aids in these conditions. In my doctoral work, using fMRI in humans I showed that real-world background noise engages different mechanisms than the simpler synthetic noise typically employed in previous work. In my postdoctoral work I will be trained in marmoset electrophysiology to zoom in to the level of neural circuits to probe the mechanisms that generate auditory cortex's robustness to background noise.
