A team of researchers say they have used machine-learning to recreate images in our brains, from pictures subjects were looking at to things they remember seeing.
The research, which has not yet been peer-reviewed, was conducted by scientists from Kyoto University in Japan and led by Yukiyasu Kamitani. Using functional magnetic resonance imaging (fMRI), the team said they were able to reconstruct images seen by our brains.
In their paper, available on bioRxiv, a number of images were presented that were recreated by the artificial intelligence, known as a deep neural network (DNN). Each image was recreated pixel by pixel by the DNN, generating images that resembled the initial image.
“The results suggest that hierarchical visual information in the brain can be effectively combined to reconstruct perceptual and subjective images,” the team wrote in their paper.
The research builds on earlier work by the same team that found that brain activity patterns could be decoded into signal patterns. Other researchers have reported similar work in this field.
“This is a significant improvement on their earlier work,” Professor Geraint Rees, a neuroimaging expert from University College London, told The Times.
In this latest paper, the researchers used three subjects (two males aged 33 and 23, and one female aged 23). They were shown images of things like a post box and a lion, as well as geometric shapes and alphabetical letters.
The images were projected onto a screen in an fMRI scanner, with the heads of the subjects secured in place with a bar for them to bite down on. Each then took part in multiple scanning sessions, each lasting a maximum of 2 hours, over 10 months.
The participants stared at each image for a number of seconds before having a rest in the first experiment. In the next experiment, they had to simply remember one of the images they had seen previously and picture it in their mind.
Some more of the images processed by the DNN. Kamitani et al
Using the DNN, the researchers then attempted to decode the signals recorded by the fMRI scanner in order to produce a computer-generated reconstructed image of what the participants saw.
The results were rather remarkable, with the DNN able to reproduce images of a DVD player, feet with socks on, a fly, and more. However, it wasn’t too hot on other images, like a person with a cowboy hat or a snowmobile, but the results are impressive nonetheless.
“Our approach could provide a unique window into our internal world by translating brain activity into images,” the team noted.