Meta, Facebook's parent company, recently presented JEPA (Joint Embedding Predictive Architecture), a new artificial intelligence (AI) model developed by Yann LeCun, the group's AI chief scientist. JEPA is designed to go beyond current AI models by being able to understand the underlying world, learn, reason and plan tasks in an advanced way.
Unlike traditional generative AI models, JEPA is based on abstract representations of images or sounds, mimicking the way the human brain works, analyzing vast amounts of information daily as it observes its environment. This approach enables models to learn autonomously, without having to guess at missing information, thus reducing text or image generation errors.
JEPA also has the advantage of being faster and requiring less computing power than existing AI models. Yann LeCun believes that broad language models, such as ChatGPT, will gradually be replaced by models like JEPA over the coming years. Although this evolution will take time, he predicts that the first versions of AI based on JEPA could reach a level of intelligence similar to that of dogs.
One of the first use cases envisaged for JEPA is image recognition. Meta, which already uses AI to filter terrorist or pornographic content on its platforms, will be able to exploit JEPA in this field. In the longer term, this AI could be used to develop marketing strategies, draft legal arguments and many other applications.
JEPA open source
Meta is also keen to demonstrate, in the face of OpenAI, creator of ChatGPT, that it remains a powerful player in the AI field. Naila Murray, EMEA Research Director of FAIR (Facebook AI Research), says the idea that Meta is behind the curve in AI is false, pointing out that the company has been using AI in-house for many years, even if its past projects were perhaps less high-profile.
Although Meta has made JEPA open source, allowing the developer and research community to contribute to its improvement, the group remains cautious about the future. With the speed at which the AI field is evolving, it's possible that an even different model will be needed in the future. Joëlle Pineau, Vice President of AI Research at Meta, acknowledges that there are still thirty years of research to come in this field.
It's clear that AI continues to evolve rapidly, with advances such as JEPA seeking to push back the boundaries of artificial intelligence and approach or even surpass human intelligence. The future promises exciting developments in this constantly evolving field.