Artificial intelligence recreates images by reading the brain – 06/04/2023 – Tech

From magnetic resonance imaging of brain activity, researchers managed to reconstitute with artificial intelligence the image of a train that had been shown to a person. The result was the silhouette of a locomotive surrounded by smoke. The colors were changed, but the shapes are close to the original ones.

Technology developed at Osaka University in Japan could help scientists map the code behind how the brain works. Exaggerating the possibilities to the limit, it would be possible with this new technique to reproduce images of dreams or thoughts on computers.

The researchers use computational models known as Stable Diffusion, similar to those employed by Dall-E —ChatGPT’s sister platform that generates images from texts. What the study at the Japanese university did was generate them from computerized resonance scans, translating brain stimuli into images compatible with the originals, according to the authors’ criteria.

For this, the study analyzed two parts of the brain: the occipital lobe, to capture shape and perspective, and the temporal lobe, to add meaning to the image.

Study author Yu Takagi, a professor at Osaka University, tells the report that the difference was this division of the MRI reading process.

The model was created using data from four people — each of whom looked at 10,000 portraits while undergoing an MRI scan.

Previous research has tried to translate brain stimuli into images, without analyzing the part of the brain that assigns meaning to objects and people. Even with larger amounts of data, the result was worse, according to the article.

The Osaka University study, however, has validity restricted to the four volunteers analyzed.

According to Professor Daniel Takahashi, from the Brain Institute at UFRN (Federal University of Rio Grande do Norte), the way the brain associates meaning with what the eyes see depends on individual experience and context. “Each person has a brain code.”

“Very likely, even if they get more volunteers, the result would not be as good, due to this characteristic of the brain”, adds Takahashi.

He also assesses the logistical difficulty of increasing the scope of the experiment. Each volunteer underwent 30 to 40 MRI sessions to view the 10,000 images. “Not everyone is up for it.”

Specialist in the application of artificial intelligence in health, professor from UFRGS (Federal University of Rio Grande do Sul) Luís Lamb says that before gaining practical applications, the results of the study need to be statistically validated, like a vaccine.

ETHICAL CONCERNS AND TECHNICAL LIMITS

Professor Dráulio Araújo, from the Instituto do Cérebro, says he is concerned about the possibilities of this technology in the future being used to find out, without permission, what people think or dream about.

This situation is still far away, according to Araújo. Models designed to tell what people see are more accurate than models designed to tell what people have in mind. “Visual stimuli generate a clearer pattern in brain activity.”

Today, artificial intelligences capable of describing in text what the person sees or thinks are more common. The accuracy of accuracy for technologies that say what the person sees is around 90%, and for what the person imagines or remembers, at 60%, according to the professor at the Instituto do Cérebro.

Araújo, who studies dreams and the effects of psychoactive substances on the human mind, says that he has already worked by labeling what people experience during sleep. “We’d ask a person to sleep inside the MRI, wake them up, ask them what they were dreaming about, and write them down.”

The process was repeated 200 to 300 times to generate enough data to feed an artificial intelligence.

The resource developed by the Japanese researchers is more complex than these, as it reconstitutes the image, with satisfactory resolution, after associating the vision stimuli with some meaning. “They inverted the traditional research process, which began 15 years ago”, says Araújo.

Despite this reflection, Araújo says that the study is important in helping to map which parts of the brain are stimulated by each mental activity. “We’re still in the basic science when it comes to understanding the cognitive system.”

Takahashi, also from the Brain Institute, says that current science doesn’t even know how the human brain works. Therefore, it is difficult to know even if the artificial intelligence of the Japanese study delivers accurate answers. Scientists are still working on hypotheses in this area.

“ChatGPT, for example, looks like people when it generates reasonable texts. But, as the technology does not have neurons, it is also possible to say, by this analysis, that its functioning has nothing to do with the human brain”, says Takahashi.

The same goes for the difference between the human brain and that of animals, according to the researcher. “It may be that the Japanese model works well with humans and animals and points to similarities or that it does not provide evidence for this discussion.”