Meta presents five innovations to bring artificial intelligence closer to human cognitive abilities

In a world where artificial intelligence is advancing at breakneck speed, Meta asserts itself as a leading player by unveiling five groundbreaking innovations. These advances aim to bring the cognitive abilities of machines closer to those of humans. Through bold projects, the company seeks to equip its AI with an understanding of the world, the ability to interact, and a depth of analysis, all with the aim of transforming our relationship with technology and making machines profoundly more human. Meta, one of the tech giants, has reached a remarkable milestone in the field of artificial intelligence (AI) with the presentation of five groundbreaking innovations. These advances aim to bring AI closer to human cognitive abilities, by equipping machines with heightened senses, intuition, and a better understanding of our world. With these tools, Meta seeks to transform the way machines interact with us and evolve. Perception Encoder: A Digital Super-RetinaAt the heart of this series of innovations is the Perception Encoder, a model that acts as a digital super-retina. This system allows AIs to decode the world around them with incredible precision. Imagine a robot capable of spotting a stingray camouflaged in a sandy landscape or a small bird hiding in the background. With this technology, machines can not only interpret images and videos but also handle complex tasks such as reading illustrated documents. They have moved far from the Stone Age, where they simply imitated human speech; they can now grasp complexity and nuance. The Perceptual Language Model: Multidimensional Understanding Next comes the Perceptual Language Model (PLM), which connects perception to comprehension. This innovative model was powered by massive volumes of synthetic data, including an impressive video dataset of 2.5 million well-labeled examples. Thanks to this database, PLM is able to associate images with verbal concepts, thus optimizing its effectiveness in understanding complex scenes. Better still, this innovation is open source, allowing the research community to explore and evaluate it using different methodologies. Locate 3D: Interacting with the Physical World When it comes to interacting with our environment,

Meta Locate 3D

is emerging as a major breakthrough. The model allows a machine to understand simple instructions such as « find the vase near the TV. » To do this, it scans the space in 3D, interprets the request, and identifies the requested object, even when several similar items are present. This system is based on an efficient architecture, combining scene encoding and language understanding, making it a giant leap towards truly interactive intelligence.

Dynamic Byte Latent Transformer: Linguistic Robustness **In an effort to improve the reliability of AI, Meta introduced the Dynamic Byte Latent Transformer. Unlike traditional language models that are word-based, this one operates at the byte level. This makes it less susceptible to errors, typos, or even invented terms. By ensuring greater stability and speed in responses, this model offers a greater sense of accuracy than ever before, enabling smoother exchanges between humans and machines.**Collaborative Reasoner: A Cooperative AI

Finally, the latest innovation presented is the Collaborative Reasoner

This model was designed to facilitate true collaboration between machines and humans. Imagine two bots discussing a problem together, supporting each other to correct their errors and find the most relevant solution. Using a simulation engine called Matrix, Meta was able to train these AIs by creating large-scale dialogues, which led to performance improvements of up to 29% on collaborative reasoning tasks.

Notez cet article