1614874332 meet facebooks powerful new image recognition seer ai

Meet Facebook’s Powerful New Image Recognition SEER A.I.

If Facebook has an informal slogan, equivalent to Google’s “Don’t Be Evil” or Apple’s “Think Different”, it’s “Fast and Break Things Move.” This means, at least in theory, that one must iterate to try news things and not be afraid of the possibility of failure. Although, in 2021, on social media, currently being blamed for a heap of social ideologies, the phrase should, perhaps, be amended for: “move things faster and fine.”

One of the many areas of social media, not only Facebook, for this, some special images have been disseminated online. This is a challenging problem by any stretch of the imagination: some 4,000 photos are uploaded to Facebook every 4 seconds. This is equivalent to 14.58 million images per hour or 350 million photographs per day. To handle this task manually, each Facebook employee must work in a 12-hour shift to approve or veto the uploaded image every nine seconds.

Facebook hacked
Digital trend graphic

It is not likely to happen soon. This is why the task of classifying images is entrusted to artificial intelligence systems. A new piece of Facebook research, published today, describes a new, large-scale computer vision model called SEER (which is “self-supervised”) trained on more than 1 billion public images on Instagram , It can improve the most state-of-the-art self-monitoring image-recognition systems, even if the images are of low quality and are difficult to read.

This is a development that its creators can claim, “[pave] Way for a more flexible, accurate and adaptable computer vision model. “It can be used” to “keep harmful images or memes away from our platform.” It can be equally useful for visually impaired users to automatically create all-text-description images, marketplaces, or Better automated categorization of items sold on Facebook shops and a multitude of other applications that require better computer vision.

Welcome to the self-supervised revolution

According to Digital AI, Priya Goyal, a software engineer at Facebook AI Research (FAIR), said, “Using self-supervision, we can train on any random image. “[That] This means that, as harmful materials evolve, we can train a new model on rapidly evolving data and, as a result, react faster to situations. “

Self-supervision Goyal refers to a brand of machine learning that requires little in the way of human input. Semisupervised learning is an approach to machine learning that sits somewhere between supervised and unsupervised learning. In supervised learning, training data is fully labeled. In unsupervised learning, there is no label training data. In learning the semicircle … well, you get the idea. This is for machine learning, keeping half an eye on your child while they charge autonomously around the park. Self-supervisory learning has been used for everything from machine translation to question answering to transformational influences in the world of natural language processing. Now, it is also being implemented for image recognition.

Brain network on the depiction of nerves
Chris Degraw / , Getty Images

“Unsupervised learning is a very broad term that suggests that learning does not use supervision at all,” Goyal said. “Self-supervised learning is a subset – or more specific case – of untrained learning, because self-supervision derives supervisory signals automatically from training data.”

What self-supervised learning means for Facebook is that its engineers can train models on random pictures, and do so quickly while achieving good performance on many tasks.

Goyal said, “By being trained on any random Internet image, we can capture the visual diversity of the world. ” On the other hand, supervised learning requires data annotation, which limits visual understanding of the world because the model is trained to learn only very limited visual annotated concepts. In addition, creating annotated datasets limits the amount of data our system can be trained on, so supervised systems are likely to be more active. “

This means AI systems that can learn better from whatever information is given, without relying on curated and labeled datasets that teach them how to identify specific objects in a photo. In a world that moves as fast as online, it is equally important. This should mean that smart image recognition works more quickly.

Other possible applications

“We can use self-supervised models to solve problems in domains that have very limited data or no metadata,” Goyal said. “Being able to train high-quality, self-supervised models from just random, unlabeled and uneducated images, we can train the model on any Internet image, and it allows us to diversify the visual content Allows to capture and minimize biases otherwise presented. Data duration. Since we require no labels or data periods for training self-supervised models, we can quickly create and deploy new models to solve problems. “

As with all of FAIR’s work, right now it is firmly in the research stages, rather than technology that will roll out on your Facebook feed in the next few weeks. This means that it will not be deployed immediately to solve the problem of harmful images spreading online. At the same time, this means that conversations about the use of AI to identify fine details in uploaded images are premature.

Like it or not, though, image-classified AI tools are getting smarter. The big question is whether they use to break things further or start fixing them back again.

Editors recommendations

Similar Posts