MIT researchers have found that computational models designed with machine learning techniques are becoming more accurate in mimicking the structure and function of the human auditory system. They believe these models could assist in the development of improved hearing aids, cochlear implants and brain-machine interfaces. In an extensive study of deep neural networks trained for auditory tasks, it was found that most models create internal representations that are similar to those generated in the human brain when listening to the same sounds.
According to the study, the best approach to training these models involves using auditory input that incorporates background noise, which typically activates the human auditory cortex.
Deep neural networks consist of multiple layers of information-processing units that are trained on vast quantities of data to accomplish designated tasks. Such models have seen wide application in numerous fields and are now seen as a potential way to explain how the human brain performs certain tasks.
The researchers in this study employed nine publicly available deep neural network models that had been trained for auditory tasks while building an additional 14 models. The models were mainly single-task ones such as recognising words, identifying speakers, recognising environmental sounds or identifying musical genres. There were two models which could perform multiple tasks.
The study supported the belief that the human auditory cortex operates in a hierarchical manner, processing distinct computational stages. The results also revealed that earlier model stages were more closely linked to the primary auditory cortex, while later model stages had more in common with brain regions beyond the primary cortex.
The research has led the MIT team to further their efforts to understand human brain responses and look into the prospect of improved hearing aids, cochlear implants, and brain-machine interfaces. The study’s now focusing on predicting brain responses using computational models improved with the help of machine learning.
