Predictive dynamics improve noise robustness in a deep network model of the human auditory system

The human auditory system is robust to many types of corrupting noise. However, the neural mechanism that drives this robustness is unclear. Experimental evidence suggests that top-down predictions play an important role in processing stimuli in noisy contexts. Feedback connections are abundant in the auditory cortex, and have been hypothesized to carry predictive signals but are often omitted in large scale models. We evaluate the potential role of predictive dynamics in auditory robustness by introducing these dynamics into a large-scale model of the human auditory system. Specifically, we augment a feedforward deep neural network trained to identify speech in corrupting noise using a recently introduced predictive feedback scheme. We find that predictive dynamics improve network performance on a challenging speech recognition task. These performance gains were associated with denoising of network representations and alterations in layer dimensionality. Finally, we find that the model captures brain data outside of the speech domain. Overall, this work demonstrates that predictive dynamics are a candidate mechanism for denoising representations in the human auditory system. Furthermore, our network model provides a testbed for hypotheses regarding the dynamics of auditory robustness.