Figure 5.

Three DNN training paradigms used to study adaptation following cochlear degradation. This training and testing framework is run for each of the 12 cochlear states that we modeled. Both training and testing data sets contain neurograms from all 21 SNRs. Normal-hearing (NH) neurograms are used to train a baseline model, which simulates NH speech-in-noise word recognition. This baseline network is utilized by three networks during a second phase of training. During phase 2, the NH control is trained on NH neurograms and the unconstrained and constrained adaptation paradigms are trained on degraded neurograms. All three of the resulting models are then tested on a held out set of degraded neurograms. The end result is digit recognition accuracy as a function of SNR, which simulates a psychometric function. The NH-control paradigm (yellow) simulates performance after a sudden hearing loss, the unconstrained-adaptation paradigm simulates unlimited training after hearing loss (red-orange), and the constrained-adaptation paradigm (green) simulates limited training following hearing loss.