Face Detection: Am I beautiful? Or just pretty? – FAZ – Frankfurter Allgemeine Zeitung

Not everyone would take the matter so sporty: “I am twenty and pretty ugly, exactly right”, let a certain FischMaster3001 the world know via Twitter. The occasion was not late adolescent identity crisis, but the site “howhot”, which was developed at ETH Zurich and the FischMaster3001 had tried. Your software calculates on the basis of photos uploaded age and sex of the person portrayed. They also rated the attractiveness on a six-point scale, the. Of “Hmm , . “To” Godlike “enough. This aspect especially the side owes its enormous popularity. Since it was released on 5 January, they visited nearly five million users and twenty million photos uploaded to vote.

But behind the app is more than just a nice gimmick. Already in the 1940s, tried programmers to recreate the complexity and networking of the neurons in the human brain to breathe a kind of synthetic intelligence their calculators. Today, Apple, Google and others use the artificial neurons to filter spam and handwriting and speech recognition. Also on howhot works in the background, a neural network while the user is waiting for the verdict of the machine.

What is beauty, the machine has to find out for yourself

But is not it hard to estimate attractiveness a highly subjective matter, the last of all you want to leave a computer program? In fact, the howhot developer Rasmus Rothe says were formerly “features, make it appear beautiful or ugly a face, predefined by the people. The computers were rules taught in the form of an algorithm, with which he should recognize predefined patterns in the images “- for example the symmetry of a face

Rothes was howhot contrast downright antiautoritär educated -. No rules, no guidelines , The neural network has to learn for themselves what beauty means. Even with the age and sex determination is made howhot on their own. No one has taught him that old people have wrinkles and babies a round face.

Howhot sees itself simply to what see as many people as beautiful or ugly, and forms it into a self-learning standards for what is beautiful and ugly. As training material, based on which learned the software, the database was a Swiss Dating App. There assess user profile pictures of other users, by sliding it on the screen right or left. From the average rating of each of these pictures are each revealed an attractive value, the Rothe – together with the associated photo -. Einfütterte in his first still completely ignorant neural network

Elderly prepare still difficulties

Such a Photo initially reached the first layer of neurons. While this only recognizes simple shapes and scans looking for this the whole picture, process neurons further levels then the information of previous layers. In this way, they filter more complex patterns, such as whole faces out. It is crucial that the information received will be passed with different weights – is approximately realized that the size of the nose of people relevant to the attractiveness rating, so the neural network learns to weigh this information strongly

.