We instantly identify a person without delay as attractive to us or not. But it has not yet been possible to find out what factors and individual features of appearance determine this spontaneous unconscious sympathy. Sometimes attractiveness is associated with personal and cultural characteristics. However, Finnish scientists have found a rational approach to this absolutely irrational
So, is there a perfect match? Based on EEG data, scientists have taught generative adversarial neural networks(GAN) predict and recreate faces that will potentially appear attractive to us. Just imagine, the final prediction accuracy was> 80%. I wonder what will happen if the network can potentially influence the pairing of Tinder and similar apps? But let's figure it out in order.
Background
Psychologists around the world have been researching for a long time what is the metric of attractiveness. They have no concrete answer to this question. Although they managed to identify certain patterns. So, among the important features, geometry, proportions and symmetry / asymmetry of the face are distinguished. In addition to external visual stimuli, psychologists believe that our hormones, the level of self-esteem and personal attractiveness, social experience, etc.
, influence the perception of the likelihood of another person. However, before all these data were not enough to artificially create a person or a couple with 100% coincidence. Finnish scientists have made significant progress in the study of personal preferences. And the experiment helped them in this.
How was the experiment
The study involved 30 employees and students from the University of Helsinki. They used 30,000 celebrity photographs to train GAN. So the network was taught to create synthetic portraits. A total of 240 of them were modeled.
At first, the participants in the experiment looked at 32 images of 8 series. With each of them, they chose the least attractive faces. After that, they began to measure the reactions and response of the brain.
Using electroencephalography (EEG), scientists recorded a reactionbrain on artificial portraits. The volunteers were shown images and monitored reactions in real time, recording all observations. The advantage of the EEG in providing data on the feedback to triggers: sensations, events, cognitive or motor event.
What then?
Thanks to the brain-computer interface, the data was transferred to the GAN. And so it was possible to train the network to create attractive faces for specific volunteers. However, to what extent they would be attractive they had yet to check.
After 2 months, the scientists gathered the participants again. They put in their image bins new attractive ones as well as other neutral and / or unattractive ones. The volunteers received a matrix of 24 pictures. Attractiveness was assessed on a scale from 1 to 5. By pressing keys, participants rated the images.
As a result, it turned out that 86.7% of the images created by the GAN were identified by the experiment participants as attractive. Interestingly, another 20% of the images that were potentially created as unattractive turned out to be attractive to the volunteers. That is, the result of the network operation was false-negative.
Perfect couple
From the arguments in favor of the system working, most of the attractively created images received a score of more than 1 point compared to the created ones, as neutral. The scientists concluded that the GAN really learned to distinguish between attractive and unattractive brain reactions, and it does this with an accuracy of 83.3%.
At the end of the experiment, the scientists talked to the subjects. They were all pleased with the experiment. And many wondered how the neural network managed to recreate the perfect beauty. They asked for copies of the photographs for themselves. And some have pointed to the image's similarity to their current partner.
Shazam vs EEG?
In addition to images, scientists have learned to recreate the music they listen to based on brain activity.
Music has various characteristics: rhythm, timbre, melody, harmony. In addition, songs represent a specific sequence of repeated data. All these musical features are perceived in a certain way by our brain. When we receive a stimulus, our senses react differently.
Researchers from India and the Netherlands were able to learn how to recreate specific songs from EEG brain activity. The melody detection accuracy was 85%.
The network was trained on 20 volunteers who listened to 12 tunes. When the network worked on the data of one particular subject, the accuracy of identifying the melody was almost 85%. When the recognition process was launched without being tied to a person, the accuracy dropped by almost 77%.
