A new study from the Massachusetts Institute of Technology (MIT) has found that modern computational models based on machine learning and structured similarly to the human auditory system could assist researchers in developing better hearing aids, cochlear implants, and brain-machine interfaces. The largest study of its kind on deep neural networks trained for auditory tasks shows that most models generate internal representations akin to those seen in the human brain when the same sounds are heard.
The participating researchers also gained insights on how to best train these types of models. Among other things, they highlighted that models trained on auditory input including background noise more accurately mimic the activation patterns of the human auditory cortex.
How the brain performs certain tasks could hypothetically be described by deep neural networks, a type of computational model comprised of many layers of information-processing units. These units could be used to decipher large volumes of data to execute specific tasks. The models’ processing units generate activation patterns in response to audio input, such as a certain word or sound. These model representations can then be compared to activation patterns seen in fMRI brain scans when humans listen to the same input.
Josh McDermott’s research group at MIT evaluated a larger set of models to assess whether a wider variety of models can approximate neural representations seen in the human brain. During the study, they analyzed nine publicly available deep neural network models trained for auditory tasks and created 14 unique models based on two different architectures. Most of these models were trained to perform a single task, while a couple of them were trained to perform multiple tasks.
The researchers found that the internal representations generated by the models exhibited a lot of similarities with those generated by the human brain. The most similar models were those trained on multiple tasks and with auditory input that included background noise. The study also lent credence to the theory that the human auditory cortex has hierarchical organization, where processing is divided into stages each supporting distinct computational functions.
In particular, earlier stages of the model were most similar to those observed in the primary auditory cortex, while representations generated in later stages tended to resemble those in brain regions beyond the primary cortex.
The team intends to use their findings to develop models that have higher success rates in reproducing human brain responses. Such models could aid the development of better hearing aids, cochlear implants, and brain-machine interfaces while broadening understanding of brain organization. The research received funding from several sources including the National Institutes of Health and an Amazon Fellowship from the Science Hub.