Researchers from Osaka University have developed an artificial intelligence system that can create images based on brain scans. The technology uses the Stable Diffusion text-to-image generator to generate pictures of objects, such as teddy bears, clock towers and airplanes. Once refined, the technology could be used to help people communicate, understand dreams, and interpret the perception of the world by different species. The model uses thousands of parameters, which is much simpler than other similar technologies that require millions of parameters.
The team used functional magnetic resonance imaging (fMRI) data to teach an A.I. model to link brain activity with images. The fMRI scans of four participants were used, who had viewed 10,000 different images as part of a previous study. The A.I. system can recreate the perspective and layout of the object the participant viewed but generates cloudy and nonspecific figures. However, by recognizing the object from the text descriptions of the training images, the technology can insert the object into the image, achieving an 80% accuracy rate.
The recreated images are similar to the original images, but with some noticeable differences. The technology has some limitations as it can only recreate images of objects that were included in the training material, and broadening the technology to include more people would require training the model on each individual’s brain scan. As such, the technology is not practical for daily use.
While the technology shows promise, broader concerns exist around A.I. technologies, such as copyright laws, biased policing, misinformation and invasion of privacy. Demis Hassabis, CEO of DeepMind, emphasized the need for caution with powerful technologies such as A.I.