A new MIT study analyzed how modern deep neural networks compare to the human auditory system when processing sounds.
The researchers looked at 23 different computational models that had been trained on auditory tasks like speech and sound recognition.
Key findings:
- By comparing neural networks to fMRI scans, the MIT team confirmed the models developed intricate representations of audio processing comparable to those formed in our auditory cortex.
- Models trained on multiple auditory tasks at once and exposed to noises during training performed the best at mimicking human brain representations.
- Models trained on speech versus music mimicked the responses in the brain's speech- and music-processing regions, further confirmation that models can specialize based on the tasks they are optimized for.
“If you train models in noise, they give better brain predictions than if you don’t, which is intuitively reasonable because a lot of real-world hearing involves hearing in noise, and that’s plausibly something the auditory system is adapted to,” Jenelle Feather, Ph.D., co-lead researcher
Assumptions, assumptions: Until recently, we could all safely assume software could never approach the sophistication of human senses.
- But the pattern-recognition powers of deep learning algorithms can mirror complex biological functions when given enough processing power and data.
Why it matters
MIT’s research demonstrates that artificial intelligence can replicate hearing and, with further refinement, could help optimize hearing aids and cochlear implants.
The takeaway
Though still early, increasingly brain-like algorithms offer researchers a powerful new tool.
- Refining auditory models with machine learning is a promising path toward personalized hearing solutions that feel more natural and intuitive to users.
- Smarter hearing technology could drastically expand functionality for those with challenging hearing-loss conditions.
