[ad_1]
This week, Facebook announced the deployment of a large-scale machine learning system called Rosetta, which it uses to automatically and proactively identify "inappropriate or harmful content" in images on the social network. In other words, Facebook has developed an AI that can tell if a meme is offensive.
The social media giant does not intend to use this information for memes, of course. In a blog post, Facebook describes how the algorithm can be used to detect text on a photo of a storefront, a street sign or a restaurant menu.
With recent appearances at Capitol Hill, the news that 26% of Americans have removed the application of their phones, and worries about the false news that have hit Facebook's heels since the election, will surely be a priority.
Rosetta's intelligence begins with a two-step process: detecting images that may contain text, and then recognizing the actual contents of the image. This model does not only concern the English language, because Facebook says that it supports different languages and encodings, including Arabic and Hindi, which means that the system can also read from left and right.
Rosetta is already used by Facebook and Instagram teams to improve the quality of photo searches, improve the accuracy of photos in the news feed and identify hate speech.
In the past, Facebook has struggled to correctly identify hate speech or misleading information and, in the documents consulted by Motherboard, Facebook's own training material incorrectly stated that an image of the 2010 earthquake in Jeiegu in China
The use of artificial intelligence to assess the severity of speech has already encountered difficulties. It has recently been discovered that Google's IA Perspective, which is used to detect detrimental comments, can be easily trapped with typos, spaces between words, and the addition of words unrelated to the original sentence.
In other words, with approximately 350 million photos uploaded each day on the social network, Rosetta and Facebook are fighting each other.
Source link
