Japanese neural network to learn to "read the mind"

It seems that computers can already read our minds. The Google Autocomplete, possible friends in Facebook and targeted advertising that pops up in your browser, and sometimes make you wonder: how does it work? We are slowly but surely moving in the direction of the computer, reading our thoughts, and new research from Kyoto, Japan, was unambiguous step in this direction.

Japanese neural network to learn to

A team of scientists from Kyoto University used a deep neural network to read and interpret the thoughts of men. It sounds unbelievable? In fact, they are doing a similar not for the first time. The difference is that the previous methods - and results - have been easier, they deconstructed the image based on and pixels, and the main circuit. The new technology, called "deep image reconstruction" goes beyond the binary pixels and allows researchers to decode the image with multiple layers of color and texture.

"Our brain processes visual information by hierarchical extraction of features of different levels or components of varying complexity," says Yukiyasu Kamitani, one of the scientists who participated in the study. "These neural network model or AI can be used as a proxy for the hierarchical structure of the human brain."

The study took place 10 months. As part of the three-person study examined images of three different categories: natural phenomena (animals or humans), artificial geometric figures and letters of the alphabet. monitors brain activity was measured either during image playback, or after. To measure brain activity after viewing the images, people just asked to think about the images that are shown to them.

The recorded activity is then fed to the neural network, which is "decoded" data and use them to generate their own interpretations of the thoughts of people.

Humans (and, indeed, all mammals) visual cortex located in the back of the brain in the occipital lobe, which is the cerebellum. The activity in the visual cortex was measured using functional magnetic resonance imaging (fMRI), which is transformed into hierarchical neural network especially profound.

Japanese neural network to learn to

Since the random images, the network repeatedly optimizes the pixel values ​​of the image. Properties induction in neural network image are similar to the features of the decoded brain activity.

What is important, the scientists studied the model using only natural images (human or natural), but has learned to reconstruct the artificial forms. This means that the model is really a "generated" image, starting from the activity of the brain, and not compared this activity with existing examples.

Not surprisingly, the model with hard decode brain activity when people were asked to recall the image, and it is easier when they are directly view these pictures. Our brains can not remember all the details seen pictures so the memories will always be blurred. The reconstructed image from the study retain some resemblance to the original images that are viewed participants, but for the most part look as minimally detailed splash. However, the accuracy of the technology will only get better, but at the same time will expand and possible applications.

Imagine "instant art", where you can create a work of art, just imagine it in my head. Or if the AI ​​will record the activity of your brain while you sleep, and then recreate your dreams for analysis? Last year, completely paralyzed patients were able for the first time to communicate with their families by means of brain-computer interface.

There are lots of possible applications of the model used in the study of Kyoto. But the brain-computer interface can also recreate the eerie images, unless we learn to properly interpret brain activity. the cost of failure in the event of an incorrect reading of thoughts may be too high.

With all this, the Japanese company is not alone in its efforts to create a read minds AI. Elon Musk founded Neuralink to create a brain-computer interface that connects people and computers. Kernel working on a chip that can read and write the neural code. Mind reading slowly develops.