Researchers at MIT may have found a breakthrough in understanding human hearing by studying deep neural networks. In a recent study, these computational models, trained to tackle auditory challenges, demonstrated an internal structure strikingly similar to the human auditory cortex when processing the same sounds. This finding could revolutionize the development of hearing aids, cochlear implants, and even brain-machine interfaces.
Exploring a total of 23 deep neural network models, some of which were custom designed by the MIT team, the study observed impressive alignment with human brain activation patterns. Working with gargantuan datasets, these models were subjected to a plethora of auditory tasks, ranging from word recognition to musical analysis. Notwithstanding the hum of background interruptions, the models that stood out were those trained under these noisy conditions, closely mirroring the human brain's ability to make sense of sounds amidst chaos. "If you train models in noise, they give better brain predictions than if you don’t," Jenelle Feather, the co-author of the study, remarked in a statement obtained by MIT News.
Dr. Josh McDermott, associate professor of brain and cognitive sciences at MIT and senior author of the paper, highlighted the comprehensive nature of the study, emphasizing the insight it provides into model training approaches. McDermott told MIT News, "The study suggests that models that are derived from machine learning are a step in the right direction, and it gives us some clues as to what tends to make them better models of the brain."
The study also suggested that human auditory processing may be hierarchical. Wondering about the nature of the brain's complexity, the researchers found that various stages of the deep neural networks mimicked distinct areas of the brain, specializing in different aspects of sound processing. Task-specific training particularly enhanced this effect, with models optimized for a speech-related task closely resembling speech-selective brain regions. Graduate student Greta Tuckute, who led the study, observed the specificity of the models: "Even though the model has seen the exact same training data and the architecture is the same, when you optimize for one particular task, you can see that it selectively explains specific tuning properties in the brain."
The potential applications of this research extend far beyond academia. Models replicating human responses might soon enhance the capability of hearing devices, providing a more natural hearing experience for users. The lab is intent on refining these models even further, leading to an improved understanding of brain function and the creation of new technologies. "A goal of our field is to end up with a computer model that can predict brain responses and behavior," McDermott said, hoping to unlock new frontiers in auditory research and device development.
