MIT researchers have found that computational models based on machine learning that simulate the human auditory system are drawing closer to potentially helping in the creation of improved hearing aids, cochlear implants and brain-machine interfaces. The study is the most comprehensive comparison so far made between these computer models and the human auditory system. Notably, the research showed that such models generate internal “representations” similar to those found in the human brain when hearing the same noises.
The MIT research also found that including background noise when training these models resulted in an outcome more closely aligned with the responses of the human auditory cortex. The study therefore helps to provide insights into the best way to train these models.
Deep neural networks, being computational models comprising many layers of information-processing units, have been widely used in a range of applications. This study analysed nine available models and created 14 originals from two different architectures. Nearly all of these were trained to perform one task, such as recognising words or identifying a speaker.
When listening to natural sounds used in human fMRI experiments, the models’ internal representations generally showed similarity to those generated by the human brain. Interestingly, those models were shown to most closely resemble human brain responses when trained on multiple tasks and on auditory input combined with background noise.
The researchers also suggested that the human auditory cortex displays elements of hierarchical organisation, with processing divided into stages carrying out different computational functions. It was further indicated that models trained on different tasks were better at mimicking different aspects of hearing. For instance, models trained on a speech-related task were shown to be more closely attuned to speech-specific areas.
Josh McDermott, the senior author of the study, says that the aim of the research field is to develop a computer model capable of predicting brain responses and behaviour, which could lead to significant advances. The research was funded by the National Institutes of Health, among other bodies.
