Decoding of images from the human brain

Today in the journal Science published scientific work ( pdf ) describing the new method of automatic reconstruction of images from the human brain. The authors of the scientific work believe that their method is more accurate than previous work in this area: the characters of the alphabet and simple contrasting figures are well recognized.

So far, the decoder recognizes only the pictures that a person actually sees at the moment, but in the future it should work on imaginary images. In this case, it will be possible, for example, to type the text with eyes closed.


')
Image recognition is carried out by a probabilistic method using a Bayes classifier. In this case, the image decoder does not need training.

To remove the signal, the method of functional magnetic resonance imaging (fMRI) is used - a type of magnetic resonance imaging, which is used to measure changes in blood flow. There is a slight delay between neural activity and changes in blood flow, as can be clearly seen in the demo video.



Anyone can test the image recognition program: it is laid out on Github .

Magnetic resonance imaging data and a converter can be found here .

The result of the decoder is shown in the screenshots.





Science article