Sunday, September 16, 2018

The neural network has learned to conduct their own public in " Vkontakte»

Abstract humor
20-year-old administrator of the public "Abstract humor" Artem Chirkov taught the neural network to create and publish memes. The algorithm combines several images from popular communities into one and publishes them on the project page. The author of the system told TJournal about her work, the future of neural networks and training as a programmer.

In 2015, Chirkov wrote the first version of the bot in Python. He downloaded pictures from popular communities, cut off the top and bottom parts, and then combined them into one image. Together with friends Chirkov taught the algorithm to send pictures to the proposed publication.

Abstract humor

In 2016, the student made four scripts from the system. One was looking for pictures and cut the text on them, the second helped to avoid too frequent access to" Vkontakte " through the API, the third saved images from the public, and the fourth contained the main code of the bot. The system did not take into account the context, so you had to manually select memes for publication. In addition, the bot was looking for pictures with black and white font Impact, and they began to go out of circulation.

Autonomous operation
Chirkov abandoned the public until may 2018, when they found a way to put the neural network to recognize text and mathematical analysis for filtering results. Once a week, the system collects information about new memes from the public "MDK", "Leprosy", IGM and "Eaglet", analyzes them and stores information about the areas with the text. She then runs a daily script that glues the zones together and schedules eight publications. The developer taught the neural network to check the result by manually selecting a thousand" suitable and unsuitable " images. As a result, the project began to work completely independently.

Abstract humor

Under the hood
To create the structure of the system, Chirkov used a ready-made neural network that determines and evaluates the content of the text. Without it, half of the pictures were with unreadable text, now there are about 5 %. The machine is alien to the notion of humor, and memes only make sense when two pictures add up as a joke. Bot has its own understanding of context: he analyzes the secondary factors such as position, size, color, and how to separate the setups from punchline.

Automation of humor
Chirkov believes that his public is a good example of future automation. According to him, if you can automate humor, it can be done with all other areas. Over time, algorithms will learn to understand the context and will be able to create funny pictures, the administrator of "Abstract humor"believes.

Artem Chirkov
The author of the algorithm said that he was born in Perm and has long been fond of programming, and the last two years studying to be a programmer in Canada. Until adulthood, he always wanted to become famous and become a "boy genius", but since then maximalism "lost". Programming was a way for him to combine what he liked: computers, games, the Internet and earnings.

Chirkov recommended the Python programming language to all novice programmers, as it was there that he was able to "finally do something useful." For new developers it is advised to find a field where I can apply their skills and to learn programming in simple projects.

Similar system
In June 2018, scientists at Stanford University introduced a new machine learning model that creates memes. For training, they used a set of 400,000 images on various topics from the site ahhh! The researchers opted for memes in the style of "advising animals".

No comments:

Post a Comment