MIT researchers have conducted a study that suggests machine learning models might be used to better design hearing aids, cochlear implants, and brain-machine interfaces. In the largest research project to date involving deep neural networks trained to carry out auditory tasks, the researchers showed that most of these models mimic the human brain’s behaviour when listening to sounds. Models that were trained using auditory input inclusive of background noise also demonstrated similar activation patterns to those seen in the human auditory cortex.
The multi-author study represents the most extensive comparison of machine learning models to the human auditory system so far. Neural networks use layers of units, which process information to perform tasks. These models produce activation patterns whenever they perform a task, like recognising a word or sound, and these activation patterns can be compared with those seen in fMRI scans of the human brain when listening to the same input.
The research studied nine deep neural network models that had been trained for auditory tasks and also created 14 models based on two different architectures. Most of these models were trained using a single task, such as recognising words, identifying the speaker, recognising environmental sounds, and identifying musical genre.
The models were presented with natural sounds used in human fMRI experiments and found that internal model representations were similar to those produced by the human brain. Models that were trained with more than one task and on auditory input with background noise had a higher similarity to human brain activity.
The study supports the idea that the human auditory cortex hierarchy might play a significant role in processing sounds. The study’s next step will be to use the new knowledge to develop models that can even more accurately replicate human brain responses. These results could provide further insight into how the human brain specifically processes sound and could be used to produce more effective hearing devices and brain-machine interfaces.
This research was funded by several sources, including the National Institutes of Health and various fellowships from MIT and external organisations. Its findings are publicly available in PLOS Biology.
