A study by Massachusetts Institute of Technology (MIT) shows that machine learning-based computational models are making strides towards mimicking the human auditory system, potentially improving the design of devices like hearing aids and cochlear implants. The research indicated that these models’ internal data structures have similarities to those seen in the human brain in response to auditory signals. The team suggested that equipping these models with the ability to process background noise could help them more accurately reflect the human auditory cortex activity.
The computational models in this study, commonly referred to as deep neural networks, are known for their exceptional performance on large datasets in recently developed applications. The scientific community has shown an interest in these frameworks due to the possibility of using them to understand and depict how the human brain performs specific tasks. When an auditory input is given to a neural network, the processing units generate activation patterns that can then be compared with those from functional magnetic resonance imaging (fMRI) brain scans of humans experiencing the same input.
MIT had previously reported similarities between patterns created by neural networks and those observed in fMRI scans of the human brain. Due to the widespread use of these computational models, the researchers sought to assess more of these networks to see if the apparent similarity to brain activity was a standard feature. They analysed nine publicly available deep neural network models trained for auditory tasks while creating 14 models based on two different structures.
These models were trained to perform various tasks, including recognizing words, identifying speakers, recognizing environmental sounds and identifying musical genres. These models were then exposed to natural sounds previously used in human fMRI studies. The internal patterns these models exhibited were comparable to those created by the human brain. Models that were trained for multiple tasks and on input with background noise, in particular, were found to resemble human brain structures more intimately.
The study also reinforced that the human auditory cortex likely has a hierarchical organization, where processing is divided into stages supporting distinct computational functions. The models were found to reflect this: early stages of the model corresponded more closely to primary auditory cortex activation, while later stages resembled brain activity beyond the primary cortex.
The models excelled at mimicking certain aspects of human auditory processing depending on the task they were trained for; for instance, models trained on speech-related tasks corresponded more closely with speech-specific brain regions. Leveraging these findings, the researchers aim to create models that more accurately reproduce human brain responses, which will help augment auditory devices and improve our understanding of brain organization.
