A study suggests that the human brain and artificial general-purpose neural networks process language sounds in similar ways
Researchers uncover striking parallels in the ways that humans and machine learning models acquire language skills. — How do brains learn?
Context & Ripple Effects
This study lands on a question Quanta's coverage has been circling since 2022, when [[a:1157802|computational neuroscientists argued that self-supervised learning models track brain function more closely than supervised ones]]. The new finding sharpens that claim from general behavior to a specific mechanism: humans and general-purpose neural networks appear to process language sounds alike, suggesting the overlap is not incidental but structural.
It also connects two threads that had run separately — the behavioral parallels between child and machine language learning, including efforts to train models on headcam footage from infants and toddlers, and the measurement side, where a GPT model paired with fMRI was already used to decode continuous language from human subjects non-invasively. If brains and networks share sound-processing strategies, both threads get a common explanatory frame.
First-order effects
- Computational neuroscientists gain a tractable model system for speech perception: instead of only measuring brains, they can probe a network that appears to solve the same sound-processing problem the same way.
Second-order effects
- Research programs built around developmental parallels — like the infant-headcam training effort — gain a mechanistic rationale, since similar acquisition outcomes now have a candidate shared cause in how sounds are encoded.
Third-order effects
- If the pattern holds across later results such as OpenAI's o1 inferring the phonological rules of made-up languages at near-human-expert level, neuroscience and AI research converge on one discipline: models become theories of the brain, and brain measurements become benchmarks for machines.
The trend: Self-supervised neural networks are converging with human language processing at the level of specific mechanisms, turning AI models into working hypotheses about how brains learn language.