Scientists at MIT successfully taught a machine learning model to perform an auditory process in a similar way to the human brain.
Some species in the animal kingdom can perceive a wide variety of sounds, including infrasounds and ultrasounds, which are way out of the human ear’s reach.
With a hearing range comprised of frequencies between 20Hz – 20,000Hz, there’s a whole auditory world to which we’re deaf.
Cats, for example, have super hearing abilities. They can perceive sounds between 40 and 46,000 hertz.
Although the human brain developed in terms of an auditory process range, it excels at processing, interpreting, isolating, prioritizing, and responding to auditory information.
Our hearing field is narrow but far more specialized than other species. We even start lending our ears to the world around us in the womb and we make great use of the sounds audible to us.
Deep Neural Model That Recognizes Words and Music
The human brain’s amazing ability to make sense of sounds has helped us survive, evolve, and develop advanced languages.
We listen to music and even augmented music to relieve stress and express ourselves. It is even by listening to the humming atoms in a super-cold Bose-Einstein Condensate that we gain a greater understanding of the creation of the Universe.
Apart from understanding more about ourselves, learning more about the inner workings of the auditory process of the brain may help us build AI that perceives and reacts to sensory information. In turn, the way deep neural networks go about processing auditory data can inform us about the brain’s neural organization.
But, who is conducting these experiments?
Neuroscientists at MIT recently reported what they describe as the first machine-learning model that can recognize speech and music in a way similar to the human brain.
To learn how to identify speech, MIT’s AI model analyzed thousands of two-second audio recordings with the task of identifying the word in the middle.
The deep neural networks did the same with music and had to identify the musical genre despite the presence of background noise.
“That’s been an exciting opportunity for neuroscience,” said MIT graduate student Alexander Kell, “in that we can actually create systems that can do some of the things people can do, and we can then interrogate the models and compare them to the brain.”
According to researchers, their model performed as any human listener would in both auditory tasks. Sometimes, they even made human-like mistakes.
“Brainy” Neural Networks Help Neuroscientists Model the Brain
MIT researchers not only created a deep neural network that mimics the brain’s auditory tasks, they also shed some light on the way the human auditory system works.
Neuroscientists already know about the hierarchical organization of the visual cortex. We know that a specific region in the brain interferes at a specific stage to process visual information.
However, they weren’t sure about the auditory system.
The MIT team found that an AI model was at its best performing these two tasks when its architecture allowed for staging them into two analysis steps.
Researchers then went a step further to model the brain and prove that it uses hierarchical architecture to decode auditory information.
It turns out that neural networks in the primary auditory cortex also rest on a hierarchical architecture to process sounds.
Researchers think their findings “offer convincing evidence that the early part of the auditory cortex performs generic sound processing while the higher auditory cortex performs more specialized tasks.”
Next, the team intends to explore other auditory tasks with their deep neural network model. They will mainly focus on aspects such as identifying the source of sounds and certifying the findings associated with the inner workings of the brain.