
Illustration: University of Oxford
According to Nature , scientists have developed a new technique called "mind captioning" that can generate accurate descriptions of what a person is seeing or imagining in their head, based on brain activity.
The method uses non-invasive functional magnetic resonance imaging (fMRI) and could open up new avenues in understanding how the human brain interprets the world .
The study, published in the journal Science Advances, showed that the system could “read” images or scenes that participants were looking at and translate them into fully semantic sentences. Experts say the technology could help people who have lost the ability to speak due to strokes or neurological conditions communicate more effectively.
In the experiment, the team of scientist Tomoyasu Horikawa (NTT Communications Science Laboratory, Japan) used a deep language model to analyze more than 2,000 video clips, turning each description into a digital "meaning signature".
They then trained another model to recognize the brain activity patterns corresponding to each signature as participants watched the videos.
Once fully trained, the model can read fMRI data from a person watching a new video and predict what they're seeing. Another text-generating model will look for a sentence that best matches this "semantic signature."
In one example, participants watched a scene of a person jumping off a waterfall. After several iterations, the AI gradually refined the description from “a rushing stream of water” to “a person jumping over a deep waterfall on a mountainside.”
The experiment also showed that when participants simply recalled videos they had seen, the system could still generate accurate descriptions, demonstrating that the brain stores and recreates images using similar mechanisms between seeing and recalling.
Researchers say the technique could improve communication for people using brain-computer interfaces, translating thoughts directly into language without the need for implanted devices.
However, experts also warn of the risk of privacy invasion if this technology is abused. Currently, the method only works with the consent of the participant and cannot read "private thoughts".
Source: https://tuoitre.vn/dot-pha-ai-doc-suy-nghi-ghi-phu-de-cho-nao-nguoi-20251107071355395.htm






Comment (0)