Meta introduced an AI that it says can fill in images more accurately than other tools on the market.
Meta presented an AI on Tuesday with the ability to complete images using “common sense”. The company explained that its model doesn’t compare pixels like other available tools do.but can understand abstract representations from “prior knowledge about the world». In this way, the company said, it can fill in raw images more accurately than other tools on the market.
He baptized him as Image Joint Embedding Predictive Architecture: Predictive Image Joint-Embedding Architecture (I-JEPA). The model is based on the vision of Meta’s chief AI scientist, Yann LeCun. His idea “is to create machines that can learn internal models of how the world works”published the company on its blog.
Systems like ChatGPT are trained under what is known as the “supervised” learning method. That is, from a large set of labeled data. I-JEPA, instead of labeled data, directly analyzed images or sounds, explained Meta, the parent company of Facebook, Instagram and WhatsApp. They call this other method “self-supervised” learning.
If we show some pictures of cows to young children, they will eventually be able to recognize any cow they see. In the same way, I-JEPA can identify representations through comparisons.
Meta posted an example of how its AI was able to fill in images of various animals and a landscape. The model was able to “semantically” recognize which parts were missing thanks to the context: the head of the dog or the leg of the bird, for example.
“Human and non-human animals appear capable of learning enormous amounts of prior knowledge about how the world works through observation and through an incomprehensibly small number of interactions in an unsupervised and task-independent manner,” LeCun explained in a journal publication. February 2022. It is worth proposing the hypothesis, he said then, that this accumulated knowledge can “form the basis of what is usually called common sense».
This “common sense” is what would guide AI models to know what is probable, what is possible and what is impossible. For this reason, Meta says, I-JEPA would not make errors that are common in images generated by other AIs, such as hands with more than five fingers.
AI systems based on labeled datasets (such as ChatGPT) are often very good at the specific tasks they were trained to do. “But it is impossible to label everything in the world”, Meta explained in another report of his 2021 investigations.
There are also some tasks for which there simply isn’t enough labeled data. If AI systems can gain a deeper understanding of reality beyond their training, “They will be more useful and ultimately bring AI closer to human-level intelligence.”
Achieving “common sense” would be like reaching the dark matter of AI, had explained Meta in 2021. The company believes that this type of AI can learn much faster, plan how to perform complex tasks, and easily adapt to unfamiliar situations.
After insisting for a long time with the Metaverse, now Meta has begun to draw more attention to its AI developments. In May it launched the AI Sandbox, a “testing ground” for early versions of AI-powered advertising tools. For now, the tests are focused on the Text writing, background generation and image overlay.
Those from Menlo Park have also presented LLaMa, their great language model, and SAM, an AI capable of recognizing elements and meanings within an image. In addition, Mark Zuckerberg, CEO of the company, said that they plan to develop a virtual assistant focused on improving the social life of its users.
I-JEPA, like the other developments announced by Meta, is currently designed to be tested by the scientific community and not by the general public. This has been the great hallmark of Meta compared to the competition.
Technological and scientific news in 2 minutes
Receive our newsletter every morning in your email. A guide to understand in two minutes the keys to what is really important in relation to technology, science and digital culture.
