Authors: Annesya Banerjee* 1 ; Mark R Saddler 2 ; Josh McDermott 2
1 Harvard University
Cochlear implants (CI) are one of the great success stories in biomedical engineering, but nonetheless fail to restore fully normal hearing in individuals with sensorineural deafness. One plausible limitation on current CIs is suboptimal algorithms for converting sound into electrical stimulation. Models that can predict what a person hears with CI input could help develop better stimulation strategies for CIs. Here, we investigate models of CI-mediated hearing based on deep artificial neural networks, which have recently been shown to reproduce aspects of normal hearing behavior and replicate hierarchical organization in the auditory system.
We modeled normal hearing by training a deep neural network to recognize speech using simulated auditory nerve input from an intact cochlea. We modeled CI hearing by testing this same trained network on simulated auditory nerve input from a CI. To simulate possible consequences of learning to hear through a CI, we retrained this network on CI input. Further, to model the possibility that only part of the auditory system exhibits this plasticity, in some models we retrained only the late stages of the network.
When the entire network was reoptimized for CI input, the model exhibited near-normal speech intelligibility scores. Performance on par with CI users was achieved only when just the late stages of the models were reoptimized (keeping the weights of the early stages unmodified).
Our results are consistent with the possibility that limitations on CI-mediated speech perception relate to incomplete plasticity that prevents the rest of the auditory system from optimally decoding CI input. Overall, our work validates deep neural networks with altered peripheral input as a candidate model of auditory perception in CI users and suggests that the difficulties of CI users could partly reflect plasticity limitations in the human brain, rather than being entirely due to impoverished auditory nerve representations from CI stimulation.