Meta's New AI Model I-JEPA Is 'Smarter' Than Other Computer Vision Models

Meta says I-JEPA delivers strong performance on multiple computer vision tasks, and it’s much more computationally efficient than other widely used computer vision models.

I-JEPA delivers strong performance on multiple computer vision tasks. Representative image.

Source : Getty

Facebook parent Meta on Tuesday announced it is introducing the first artificial intelligence (AI) model based on a key component of its Chief AI Scientist Yann LeCun's vision. Meta is calling this AI model the Image Joint Embedding Predictive Architecture (I-JEPA), which learns by creating an internal model of the outside world and compares abstract representations of images (rather than comparing the pixels themselves).

According to Meta, I-JEPA delivers strong performance on multiple computer vision tasks, and it’s much more computationally efficient than other widely used computer vision models.

"We're open-sourcing another AI tool: Image Joint Embedding Predictive Architecture, a new way for machine learning to predict high-level concepts rather than focus on pixel-level details. This approach more closely reflects how people understand the world. To build AI assistants, we need models that perceive the world and make predictions. This research is another step in that direction," Mark Zuckerberg, CEO, Meta, announced on his Facebook wall.

The representations learned by I-JEPA can also be used for many different applications without needing extensive fine-tuning. For example, the social networking giant trains a 632M parameter visual transformer model using 16 A100 GPUs in under 72 hours, and it achieves "state-of-the-art" performance for low-shot classification on ImageNet, with only 12 labelled examples per class. Other methods typically take two to 10 times more GPU-hours and achieve worse error rates when trained with the same amount of data.

The company added that at a high level, the JEPA aims to predict the representation of part of an input (such as an image or piece of text) from the representation of other parts of the same input. Because it does not involve collapsing representations from multiple views/augmentations of an image to a single point, the hope is for the JEPA to avoid the biases and issues associated with another widely used method called invariance-based pretraining.

To recall, in a bid to take on AI chatbot rivals, the Facebook parent unveiled its own artificial intelligence (AI) chatbot named Metamate earlier this week. The social networking giant recently gave a sneak peek into its "incredible breakthroughs" in AI and gave a hint at unveiling a productivity assistant for its staff. Metamate is currently being introduced internally to a small group.