The neural network “Vkontakte” was taught to identify suicidal content

The neural network “Vkontakte” was taught to identify suicidal content

Detects dangerous and fraudulent materials, spam and pornography.

Social network “Vkontakte” began to use “trained” a proprietary algorithm that assigns the images a special rating to determine if they contain illegal or dangerous content.

The algorithm recognizes the images of the objects — people, animals, plants. Identified as a “prohibited” content does not fall in the search results and “smart” news feed either completely blocked

The technology is also able to analyze the text overlays on the images.

NewsSocial networks have declared war on suicide

The company said that the neural network is able to determine the objects and entities depicted in the photo. She understands not only the “people”, “child”, “animal”, “plant”, but also can determine, for example, a piece of clothing, a symbol of the “groups of death”.

In General, the neural network has several thousand classes that are assigned to photos — from 0 to 1 according to several parameters. If the algorithm characterizes the photos as “dangerous”, it is sent to the moderators for verification. If a photo is recognized as “moderately hazardous”, it disappears from the search results and in news feed lowers its priority. The mechanism helps to stop the spread of spam, fraud, child pornography, illegal services, and other illegal activities.

A separate script is provided to fight with suicidal content. Records and pictures on this topic will be automatically deleted and their authors blocked.

If a user uploads a picture or mentions of hashtags previously used as a call for suicide, “Vkontakte” will block the page and offers to provide reasons for the publication

Based on the answers is determined when the page is unlocked, and further actions: the user may be offered the assistance of psychologists Fund “Your turf” or security Advisory page.

The algorithm also helps to create a “smart” news feed, which is introduced in “Vkontakte” in the past year, records from friends and communities are not displayed in chronological order and by priority, the most interesting to a specific user.

The assignment of classes need not only to remove the illegal images, but also for the formation of the ribbon user taking into account his interests. If a person “likes” the photos with cats, personalized ribbon in the priority he will be more likely to display records with cats. Today, according to “Vkontakte”, “smart” tape is used by 80% of the active users of the social network.

Head of Big Data and Machine Learning “Vkontakte”, Andrew Laws said that now the neural network was tested for the analysis of the text over the images. According to him, technology is required in order to make news user “the most useful and interesting.”

The neural network examines only the images uploaded for sharing, that is, send photos in private messages are not analyzed, — said Andrey Laws. Using this technology in the future we will be able to deal with spam and fraud. We have trained our neural network on the basis of complaints of users to different content.

The technology of neural networks known for many years, but it was made popular multi-player app that turns photos into works of art. The first was a Prisma from a native of Mail.Ru Group, later the same app called Vinci released “Vkontakte”, and then she Mail.Ru Group released app Artisto, which is already edited video.

Unlike the standard algorithms that are programmed to perform certain actions, a neural network works on the principle of the human brain

It taught us to analyze data and make decisions on their own. For example, Google is now developing a neural network in your service interpreter so that she could translate texts from one language to another, as it would have done.

The founder of the company “ARB didzhital” (working in the field of data analysis and machine learning, including the creation and application of artificial neural networks in applied business tasks) Vladislav Arbatov said that in any social network there is the problem of suicidal manifestations and illegal activity, but deal with it differently. According to him, already, there are neural networks, which are not only able to recognize objects in the photo, but also to make meaningful descriptions.

For example, just blocks Instagram posts under the hashtag “blue” and proposes to seek qualified help, but it is not very efficient. But the use of VK more suitable method with image recognition is a big plus and will be more effective, — said Vladislav Arbatov. — Text content can now be perform, to search, to targetirovat advertising based on it, but the image must first be recognized. Roughly speaking, this is a picture in a text form suitable for many applications. Instagram, Facebook is also using machine learning technology, how to identify content 18+ and more relevant to the user of the tape, and Facebook still uses an algorithm for people with disabilities, users with low vision can run “reader”, find out what the picture is.

Researcher of the University Innopolis Leonard Johan said that image classification — a major achievement in the field of neural networks.

Neural networks can cope with the simple task of object recognition is far better than human error in the detection of a uniform object, as a rule, is only a few percent — said Leonard Johan. — Of course, visual suicidal and erotic content is a subjective concept that depends largely on context, so such content is more difficult to recognize compared to ordinary objects. Analysis of image content — the object of attention in the scientific community with promising results. It is an interesting but difficult task. Of course, in the process there will be mistakes, but with increasing volume of data collected, the results will be better.

The expert added that opinions regarding the development trend, the classification of images and video differ

Some members of the community of specialists in deep learning claim that a neural network can develop a deeper perception of the world just by tracking very large volumes of images and videos. Others believe that the neural network will learn this world, since true understanding only occurs during complex interactions with him.