Meta has presented a new tool that allows the generation of images and texts with artificial intelligence (AI) using the prediction of certain parts of the content and imitating human reasoning.
The company has explained that this solution is the result of an idea by Meta’s chief AI scientist, Yann LeCun, who proposed “a new architecture aimed at overcoming the main limitations of the most advanced AI systems,” as stated in a statement. .
The result of their work is the Image Joint Embedding Predictive Architecture (I-JEPA), a tool that collects data from the outside world, creates an internal model of it, and compares abstract representations of images, instead of comparing the pixels themselves.
The company has recalled that humans “learn an enormous amount of prior knowledge about the world by passively observing it”, an aspect that it considers “key to enabling intelligent behavior”.
For this reason, the objective of this model is to predict the representation of a part of a content, such as an image or a text, based on the context offered by other parts of the composition.
Once I-JEPA collects all this information, it is in charge of predicting the missing pixels of an image or the words that do not appear in a certain text, to give it a natural and realistic meaning.
Meta has also commented that, unlike other generative AIs, theirs uses “abstract prediction goals” for which unnecessary detail is removed at the pixel level, allowing the model to learn additional semantic features.
The company has finally indicated that it continues to work on expanding the focus of this tool so that it learns “more general” models based on more specific modalities. For example, allowing spatial and temporal predictions about future events with a video from a simple context.
