Researchers publish debate on whether AI brain models explain brain activity
Reported by The Transmitter, udaipurkiran.com and Tech Times3 sources ↓
A November 11 Journal of Neuroscience paper by Bing Brunton and Patrick Mineault examines whether AI models that predict brain activity reveal how the brain works. Brunton argues that similar predictive accuracy from models based on different theories cannot identify the correct mechanism.
Key points
- Brunton says equally accurate models based on different theories cannot establish how the brain works
- Mineault argues brain-activity models can be studied and tested against real brains
- The authors planned to discuss the debate at a Society for Neuroscience session on November 15
Sources
Details
The Dual Perspectives article sets out contrasting views on whether predictive accuracy can lead to an explanation of brain function. Brunton argues that different models can fit the same neural recordings, so prediction alone does not show that a model captures the brain’s underlying processes. He also questions whether models based on limited measurements can represent cognition involving the body and action.
Mineault argues that models trained directly on brain recordings may help explain activity because researchers can examine them and test their predictions in living brains. Stanford’s Andreas Tolias points to an example from his lab: a model predicted a link between mouse pupil dilation and visual neurons’ sensitivity to ultraviolet light, which the team then tested in experiments.
The article describes mixed evidence on how these models improve. OmniMouse, trained on recordings from more than 2.3 million neurons across 73 mice, improved with more data, while adding model parameters helped only to a point. Other studies cited in the article found that more recording time per person improved some brain-decoding results, while adding participants had little effect; three foundation models also performed worse than a simple connectivity model on cognitive-score prediction.
Written from coverage as of
Images
- The Transmitter
- udaipurkiran.com
- Tech Times