Neural networks, or, more properly, networks, artificial neural are structured in computer systems approach to computing based connections that mimic the human brain's processing engine.
Google is to teach their algorithms using neural networks to generate thumbnail images in the videos of its users on YouTube.
The video put on YouTube is complemented by a miniature showing, or attempts to display an image, which the subject, who is present or something you can identify what we'll see. This miniature is more or less attractive it allows people to see or not that video, because the image is itself the key to success of content available on YouTube.
While it is possible for people to put the image that best suits the content, the vast majority do not do that and leave the burden to YouTube to generate a thumbnail automatically. But Google wants to make it more scientific and objective task.
When we send a video to YouTube, Google automatically selects one frame at a second of the video as this is being loaded and applies to post multiple filters to choose well, have a good thumbnail, something that makes sense.
This process is flawed and does not always work in a useful way, in fact we can see examples in this image that Google itself provided:
Now the method will be others, and we can see how they intend to implement:
Unlike the task of identifying whether a video contains your favorite animal, judging by the visual quality of a video frame that can be very subjective - people often have different views and preferences in choosing frames as thumbnails of the video. One of the main challenges we faced was how to collect a large set of successful training examples to feed our neuronal network. Fortunately, on YouTube, in addition to thumbnail generated by algorithms, it also has many videos with carefully designed custom thumbnails and sent by the creators.
Basically, Google making use of custom thumbnails as training, allows his team can give concrete examples, serves as intensive training to teach a neural network for this "see" a little more as they think humans, mainly leading them to choose certain image at the time of forming a thumbnail that identifies the video content.
The way we present the algorithm already choose the images shows that there is an evolution, as we can see above in the example that Google has made available to your blog, but there is still a continuous learning because the contents are in such a comprehensive way that everything can challenge the "intelligence" of the artificial neural network.
Today a human neuronal mechanics, tomorrow a set of actions ... one day the brain to function artificially.
Post a Comment