A recent study from MIT has shown that computational models that mimic the structure and function of the human auditory system could significantly aid research into more sophisticated hearing aids, cochlear implants, and brain-machine interfaces. Modern computational models that use machine learning have already made progress in this area.
The MIT team carried out the most extensive study to date on deep neural networks trained to carry out auditory tasks, finding that the majority of these models produce internal representations sharing key properties with those in the human brain when listening to audio input.
The research also suggests that including background noise in the auditory input used for training these models brings them even closer to accurately mimicking activity in the human auditory cortex. The study’s senior author, Josh McDermott, suggests this represents a significant step forward in developing more accurate models of the human brain’s auditory system.
Deep neural networks consist of layers of information-processing units that can be trained on large datasets to perform specific tasks. They are widely used in different applications, and neuroscientists have begun exploring their potential use in modeling the human brain’s functionality. According to Greta Tuckute, an MIT graduate student, machine learning models can perform actions on a previously impossible scale.
By comparing a neural network’s processing units’ activation patterns in response to various auditory inputs like spoken words or sounds to fMRI brain scan results, it is possible to see similarities between these models and the human brain. This comparison function was first noted by McDermott and Alexander Kell in 2018, and since then, more researchers have used these models.
For this investigation, the researchers examined nine deep neural network models publicly available and developed 14 models of their own, mainly designed to undertake a single task such as recognizing words, identifying speakers, or detecting music genres. When these models were exposed to sounds used in human fMRI tests, they behaved similarly to the human brain.
Furthermore, the study supports the theory of hierarchical processing in the human auditory cortex, showing distinct computational functions at different processing stages. The researchers found early-stage model representations most closely resembled the primary auditory cortex, whereas those in the later stages aligned more with brain regions beyond the primary cortex. The findings will be used to develop even more accurate models of the human auditory system, assisting the development of more advanced hearing aids, cochlear implants, and brain-machine interfaces. This research was funded by an array of institutions, including the National Institutes of Health and the American Association of University Women.
