
AI can pluck images directly from a person’s brain. Given an fMRI scan of someone looking at a picture, an algorithm can reconstruct the original picture from the scan. Though the results aren’t yet perfect, they are still often recognisable and hint at what may be possible in the future.
Guohua Shen at Japan’s Advanced Telecommunications Research Institute and his colleagues experimented with three types of images: “natural” pictures of things like bats, snowmobiles and stained glass, artificial shapes like squares and plus signs, and alphabetical letters.
The shapes and letters were identifiable, but the reconstructions of the natural images tended to be blurry and difficult to parse. Once improved, AIs like this might allow computers to know what we’re thinking about.
Advertisement
“These decoding methods could be used for human-computer interaction in the future,” says at Purdue University in Indiana. “It is exciting that you could be able to know what a person is dreaming or thinking just by analyzing the brain signals.”

Most neural networks of this sort have two steps: first they decode the data from your brain scan into a few specific features that the algorithm can understand, then they either reconstruct or identify the picture that those features represent. In order to do this, the network is trained on a pre-assembled set of images, sometimes using over a million pictures.
However, the image features do not always map directly to patterns of brain activity, so Shen’s system does the same job with only one step. It translates brain activity measured by the fMRI machine directly into a reconstruction of the image at which the subject is looking.
This algorithm is trained on a sample from a previous study consisting of 6,000 pictures and the fMRI readouts from subjects who were looking at those pictures. In comparison Google’s image recognition systems are trained on 300 million images. Getting hold of that many fMRI scans will be tricky and expensive, but could lead to huge improvements.
Read more: Mind-reading AI uses brain scans to guess what you’re looking at
Reference: bioRxiv, DOI: