MIT researchers have conducted the largest study to date of deep neural networks trained for auditory tasks. These computational models, which mimic the structure and function of the human auditory system, have the potential to improve hearing aids, cochlear implants, and brain-machine interfaces. The study shows that the majority of the models generate representations which are similar to those occurring in the human brain when hearing the same sounds. Models trained with background noise more closely imitate the processes in the human auditory cortex.
The MIT study represents the most thorough comparison to date of these models against the auditory system, offering insights into what makes these machine learning derived models more effective mimics of the brain. Deep neural networks can handle large volumes of data for specific tasks and are increasingly being used to describe the way the human brain completes certain tasks. Their processing units generate activation patterns in response to each audio input, which can be compared to the activation patterns seen in humans listening to the same input.
In 2018, it was discovered that neural networks trained to complete auditory tasks showed familiarities with human processes. In the years since, these models have become wide-spread, causing researchers to examine a larger set of models to see if they could approximate the neural representations seen in the human brain.
The MIT research group analyzed nine publicly available deep neural network models and created 14 of their own. Most were trained for a single task while two were trained for multiple tasks. The study found that the internal model representations tended to resemble those from the human brain, particularly when trained on multiple tasks and background noise. McDermott’s lab now plans to use these findings to develop models that are even more successful at reproducing human brain responses, potentially leading to advancements in hearing aid, cochlear implant, and brain-machine interface technology.
The study also lends support to the idea of hierarchical processing in the human auditory cortex, where processing is divided into stages that support different computational functions. Furthermore, the study found models trained for specific tasks were better at copying certain aspects of audition. For instance, those trained for a speech-related task more closely resembled speech-specific areas.
With the overall aim of creating a computer model capable of predicting brain responses and behavior, these findings represent a step forward in reaching this goal, anticipated to significantly advance various technologies among other things.
