A team of four Japanese scientists based in Kyoto University, have just used artificial intelligence to decode and recreate images from the human mind. Though machine learning has been used to generate visuals from MRI (brain scan) data of someone thinking of basic shapes and colours, Guohua Shen, Tomoyasu Horikawa, Kei Majima and Yukiyasu Kamitani have made a significant leap, using deep neural networks (A.I.) to decode far more sophisticated images, like pictures of animals and vehicles.
The neural network of general purpose A.I. is so detailed and complex that it can be used as “a proxy for the hierarchical structure of the human brain”, to quote Yukiyasu Kamitani. Previously, using machine learning, the method was to “assume that an image consists of pixels or simple shapes”, he continues, “but it’s known that our brain processes visual information hierarchically, extracting different levels of features or components of different complexities”.
Three subjects, over ten months, were shown photographs of animals, objects and letters of the alphabet. In some instances, a subject’s brain activity would be measured while they were looking at one of the images, and in other instances brain activity was measured when a subject was asked to think about what they’d just seen. The A.I. would then reverse-engineer (decode) their measurements to recreate the images the subjects saw. Below is a flow chart of the decoding process.
Below are images resulting from reconstructions made by the A.I. while the subjects were looking at the original pictures.
And finally, images reconstructed from the subjects memory of the original pictures. Though recalling an image in exacting detail is impossible for humans, note the striking similarities across the rows between what the subjects are remembering and what the A.I. has developed from their memories.
The creators of Google’s Deep Mind [general purpose] A.I., specifically Demis Hassabis, stated that their artificial neural network was developed to help them better understand the nature of intelligence. That neural network went on to teach itself to walk, having never been exposed to the action in any way, and became the best ‘Go’ (an ancient Chinese abstract strategy board game) player to have ever existed in all 2,500 years of the game being played. We appear to be on the verge of a technological advancement that will take the human race to places we couldn’t possibly conceive of… but with the creation of intelligence potentially greater than our own, what becomes of us?
Original Article: Here