AI experiment shows ‘mind-reading’ capability
人工智能实验显示出 “读心术 “能力
After waking up, you may feel frustrated you cannot recall the dreams you had last night. Artificial intelligence may be able to help.
Previously, there have been AI models that can turn text into images. They do this by learning from a large amount of data from both texts and images. This time, researchers from Osaka University in Japan have trained an AI system called Stable Diffusion to re-create images based on people’s brain scans, Science magazine reported.
The researchers used an online data set provided by the University of Minnesota, which consisted of brain scans from four participants as they each viewed a set of 10,000 photos. The scans were recorded by functional magnetic resonance imaging.
The AI then learned about the brain activities by analyzing changes in blood flow shown by the fMRI data — when a part of the brain is activated, more blood will flow to it. It then matched the brain activity with the photos.
Through this method, AI learned how human brains would react when seeing different photos.
Finally, the researchers tested the AI on additional brain scans from the same participants when they viewed photos of a toy bear, airplane, clock and train.
If the person looked at an airplane, for example, the AI would use the brain scan data to create an image of a very blurry airplane. Then, it would turn on the previous “text-to-image” model and improve the quality of the image by feeding itself the keyword “airplane”.
The final images were “convincing” to about 80 percent accuracy, according to the researchers.
The new study created a novel approach that incorporates text and images to “decipher the brain”, Ariel Goldstein from Princeton University said.
In the future, scientists hope the technology can be used to record imagined thoughts and dreams or allow people to understand how differently other animals perceive reality.