A new study from MIT suggests that modern computational models powered by machine learning could potentially aid the design of better hearing aids, cochlear implants, and brain-machine interfaces. These models, specifically deep neural networks, are starting to encompass functions that replicate the structure of the human auditory system.
The study further illuminates how to best train these computational models. The researchers found that those trained on auditory input which includes background noise more closely replicate the activation patterns of the human auditory cortex. The internal representations generated by these models show similarities to the representations seen in the human brain when processing the same sounds.
Neural networks are composed of layers of information-processing units that can be trained to perform specific tasks. The internal representations generated by such networks during task performance can be evaluated in relation to the brain responses seen in fMRI scans. In this study, the researchers analyzed nine publicly available deep neural network models and created 14 of their own, based on two distinct architectures. The tasks that the models were coded to perform included word recognition, speaker identification, environmental sound recognition, and musical genre identification.
Models trained with natural sounds showed resemblance with brain activation patterns observed in humans. Results revealed that models trained to perform more than one task and conditioned with auditory input that included background noise exhibited representations most similar to those observed in human brain activity. Models specifically trained on different tasks showed efficiency in replicating different aspects of auditory perception. Consequently, such distinctions in task-specific conditioning can reflect specific tuning properties of the brain.
The study bolsters the concept of hierarchical processing within the human auditory cortex—processing which is divided into stages that support specific computational functions. Future work carried out by this research group will aim to develop computational models that can emulate human brain responses more effectively. It is hoped that successful implementation of such models could lead to advances in hearing aid, cochlear implant, and brain-machine interface technology.