Optimal hearing aid design through restoration of the neural code

Abstract

Hearing loss introduces complex distortions in the neural coding of sound that current hearing aids fail to address. Here, we combine electrophysiology and deep learning to identify novel sound processing strategies to correct these distortions. We use large-scale intracranial recordings from the gerbil inferior colliculus to train deep neural network models of neural coding (ICNets) to serve as in silico surrogates for brains with normal and impaired hearing. We then use the ICNets to train another network (AidNet) to act as an optimal hearing aid, providing the individualized sound processing required to elicit normal neural activity in impaired brains. We find that AidNet outperforms state-of-the-art hearing aid processing by a wide margin in correcting distortions in neural coding and deficits in simulated phoneme recognition. Much of this advantage was retained when AidNets parameters were swapped across animals with similar hearing thresholds, suggesting that AidNet can compensate for complex effects of hearing loss without direct access to neural recordings. These results demonstrate that closed-loop optimization can identify novel and generalizable strategies for neural restoration, providing a foundation for the development of more effective hearing aids as well as other sensory devices and neurotechnologies.