Do you want your thoughts read from a distance? Think it’s a stupid question. However, this is not the case. Scientists do not have much time left for such a fantastic phenomenon to become a reality.
Scientists at the University of Texas at Austin have published an article in the journal Nature Neuroscience on the working methods of their new invention – a “semantic decoder.” This is a similar ChatGPT neural network that uses a language model to analyze MRI images and “decipher” patients’ thoughts. Unlike analogues, it does not require data on neuronal activity, that is, subjects do not need to implant electrodes in the brain.
Functional magnetic resonance imaging (fMRI) allows you to determine the activity of brain neurons to increase blood flow, which is necessary to supply them with oxygen. The tomograph receives high-resolution images with a frequency of about ten seconds: during this time, a person manages to pronounce about 20 words. That is why scientists needed a large-scale language model-transformer to decrypt the data.
The authors of the work trained the neural network on the fMRI results of three volunteers whose brain activity was monitored while reading audiobooks. The model learned to identify neural patterns activated in response to different word sequences rather than each word individually, that is, reconstructed the overall meaning of thoughts. The accuracy of the “semantic decoder” was surprisingly high, for example, the phrase “I have not yet received a driver’s license” sounding in my head turned into “She has not yet begun to learn to drive.”
The trained system also successfully coped with a more complex decryption: for example, if a person heard two texts at the same time, but mentally concentrated on only one of them or spoke phrases to himself. Even when showing videos (no sound), the neuronal activity recorded by fMRI allowed the neural network to reconstruct the sequence of events on the record. But still, the “semantic decoder” is not able to read the data of any patient at the moment.
First, the system is separately trained on the tomography data of each person individually. Secondly, the subject should focus on the relevant thoughts, actively helping the neural network – it will not cope on its own. In cases where the subjects were asked to be deliberately distracted, the AI demonstrated extremely low fMRI decoding accuracy. However, in medicine – for example, to help paralyzed patients who remain conscious – a decoder can prove exceptionally useful.
In the near future, mobile gadgets will be a thing of the past. Their place will be taken by devices for transmitting thought over a distance.
Add to favorites