Neural Network Models of Hearing Clarify Factors Limiting Cochlear Implant Outcomes

Authors: Annesya Banerjee1, Mark Saddler1, Josh McDermott3

1Harvard University
2Technical University of Denmark
3Massachusetts Institute of Technology

Introduction: Current cochlear implants (CI) fail to restore fully normal hearing. These shortcomings could arise from suboptimal stimulation strategies, degeneration in the auditory system, or limits on the brain’s ability to adapt to CI input. Computational models that predict auditory behavior from CI input may clarify how these different factors shape CI outcomes. Here, we developed artificial neural network models of hearing with CIs and investigated the effects of these different factors on speech recognition and sound localization performance.

Methods: We modeled normal hearing by training a feedforward convolutional neural network to recognize speech or localize sounds in noise given simulated auditory nerve input from an intact cochlea. We modeled CI hearing by testing this trained network on simulated nerve input from CI stimulation. To simulate learning to hear through a CI, we re-optimized the network for CI input. To simulate the best case (complete plasticity), all the network weights were re-optimized. To simulate potential limits on plasticity, only late-stage network weights were re-optimized.

Results: Models trained with CIs exhibited impaired speech recognition and sound localization relative to the normal hearing model. For speech recognition, re-optimizing the full model for CI input eliminated much of the deficit, resulting in better performance than is typical of CI users. Performance on par with CI users was achieved when re-optimization was limited to late stages of the network, consistent with plasticity constraints limiting speech recognition in CI users. For sound localization, large deficits remained even after reoptimizing the full model, suggesting device-related factors limiting performance.

Conclusion: Our results help clarify the roles of impoverished peripheral information and incomplete central plasticity in limiting CI users’ performance of realistic auditory tasks.