A neural network adds descriptions to YouTube videos

DeepMind AI adds descriptions to short YouTube videos
(illustration: CC0 Public Domain)

Flamingo, a visual language model created by the Google DeepMind team, writes descriptions for short YouTube Shorts videos. They are published quickly and often lack descriptions and meaningful titles, making them hard to find. Flamingo aims to solve this problem.

The neural network writes the descriptions by analyzing the opening frames of the video and offers explanations for what is happening, for example: “a dog is holding a pile of biscuits on its head”. AI-generated text descriptions are stored in a metadata format that will help “better categorize videos and match search results to user queries.”

AI solves a problem specific to the YouTube Shorts section: creators often ignore video metadata, and the videos themselves are mostly viewed in the general feed and, as a result, cannot be found through search.

The descriptions generated by Flamingo are not shown to the viewers or even the creators of the video, explained Todd Sherman, director of product management for YouTube Shorts, quoted by The Verge. The text of these descriptions complies with the ethical standards of Google products, although it is unlikely that the AI ​​will try to present the video in a bad light.

Flamingo already works on YouTube and adds its descriptions to all new videos in the Shorts section – a significant part of already published materials, and especially the most popular ones, have also gone through the procedure.

The administrators of the platform do not rule out the possibility that the AI ​​model will also start working with full-length long videos on YouTube, although there is no particular need for this: people spend hours, days, weeks and even months shooting and editing these materials, so the addition of metadata to them is only a small part of the video creation process.

And since viewers choose long-form videos based on the thumbnail and title, the creators of these posts have a natural incentive to fill out the metadata correctly.

Leave a Reply

Your email address will not be published. Required fields are marked *