Predictive Coding Dynamics Improve Noise Robustness in A Deep Neural Network of the Human Auditory System
Ching Fang, Erica Shook, Justin Buck, Guillermo Horga, Columbia University, United States
Contributed Talks 4 Lecture
Grand Ballroom A-C
Sun, 28 Aug, 11:45 - 12:05 Pacific Time (UTC -8)
The human auditory system is robust to many types of corrupting noise. However, the neural mechanism that drives this robustness is unclear and has rarely been explored in large-scale auditory network models. There is converging evidence from the visual neuroscience and computer vision literature that feedback connections play an important role in making biological and artificial networks robust to noise. Furthermore, there is evidence of substantial feedback connections in the human auditory cortex. Here, we augment a feedforward deep neural network trained to identify speech in corrupting noise using a recently introduced predictive coding scheme. We find that the introduction of feedback connections improves speech identification across several types of corrupting noise. An analysis of network activity showed that feedback connections denoise representations which drives this performance improvement. We also find that the extent of this performance improvement depends on which layers are augmented with feedback connections. Overall, this work demonstrates that increasing the biological realism of a deep neural network of the auditory system improves robustness to corrupting noise, a key feature of human audition. Furthermore, our network model of auditory predictive coding provides a testbed for hypotheses regarding the dynamics of auditory representations.