Computational models that mirror the structure and functioning of the human auditory system could lead to improvements in hearing aids, cochlear implants, and brain-machine interfaces, researchers at MIT say. The team has conducted the largest study yet of deep neural networks trained to perform auditory tasks, and found that most generate internal representations bearing similarities to those seen in the human brain when listening to the same sounds. The research also highlighted that models trained on auditory input which included background noise were more accurate in replicating the activation patterns of the human auditory cortex.
“We found that the models trained in noise give better brain predictions compared with those that are not. This makes sense because a lot of real-world hearing involves listening in noise, which is quite possibly something the auditory system is adapted to,” said Jenelle Feather, a PhD candidate and author of the study. The neural networks were given sounds previously used as stimuli in human fMRI experiments, and the researchers looked at the internal model representations that had most similarity to those created by the human brain.
Another important finding from the study was its support for the concept of a hierarchical organization in the human auditory cortex, where processing is split into stages supporting distinct computational functions. The study found that models trained on different tasks excelled at replicating different aspects of hearing. For instance, models trained on a speech-related task closely resembled areas specific to speech.
Professor Josh McDermott, the senior author of the study, says the research offers the most comprehensive comparison to date of these models with the auditory system, and provides clues as to what can lead to improved models of the brain. The study signals that models derived from machine learning are pointing in the right direction, he adds.
The research now aims to use the findings to create models more successful at reproducing human brain responses. Should such models become a reality, they could assist in the improvement of hearing aids and cochlear implants. “A goal of the field is to find a computer model that predicts brain responses and behavior. We think that this could open several doors,” McDermott concludes.
