Artificial Neural Network Models of Normal and Impaired Hearing
Mark Saddler, McDermott Lab
Recent work has shown that artificial neural networks optimized to perform auditory recognition tasks from simulated cochlear input replicate aspects of human auditory behavior. We extended this approach to investigate whether changes in peripheral auditory processing believed to accompany hearing impairment can account for the deficits experienced by hearing-impaired listeners. Psychophysical experiments were simulated on networks trained to perform auditory tasks using either healthy or impaired (with broader frequency tuning and reduced nonlinear amplification) cochlear representations. Surprisingly, networks trained and tested with the impaired cochlea performed very similarly to their unimpaired counterparts on pitch and speech-in-noise psychophysical experiments. By contrast, networks trained with the healthy cochlea and tested with the impaired cochlea replicated behavioral deficits of hearing impaired humans. Our results raise the possibility that some of the challenges experienced by hearing impaired listeners may not result from a less informative peripheral representation, and could instead reflect a lack of plasticity in the auditory system following alterations to the cochlea.
Segregation from Noise as Outlier Detection
Jarrod Hicks, McDermott Lab
Because sound events often occur amid the clutter of background noise, the auditory system must segregate foreground events from noise in order to make sense of the everyday acoustic environment. We explored whether listeners might identify foreground sound events by estimating distributions over environmental background sounds and registering outliers of these distributions as new events. To test this, we assessed listeners’ ability to detect brief (0.5 s) foreground sounds embedded in real-world background noise (3 s excerpts of sound textures). Critically, learning the background distribution in an online manner requires accumulating enough data samples to adequately estimate the distribution parameters. Thus, we predicted that foreground detection performance should increase with exposure to the background and level off once the background has been accurately estimated. Our results support this hypothesis, with foreground detection performance increasing over the first one-second exposure to the background. In addition, the peak foreground detection performance occurred later when background sounds were less homogenous (as measured by variability in their statistics over time), suggesting listeners collect a larger sample when more data is required to accurately estimate parameters of the background. These results are consistent with the idea that listeners estimate the distribution of ongoing background noise and segregate sound events that are outliers from this distribution.