Meta has unveiled a computer vision model that learns more like humans.

 

Meta has unveiled a computer vision model that learns more like humans.  I-JEPA is capable of generating human-like images by completing incomplete sketches. Backed by Meta's top AI scientist Yann LeCun, I-JEPA learns by creating an internal model of the world and comparing abstract image representations. I-JEPA uses the concept of self-supervised learning, which allows AI systems to learn from raw observations without labeled training data. The model predicts high-level abstractions instead of pixel-level details. Its approach mimics human learning by capturing common sense background knowledge. Meta says the model performs well on computer vision tasks, with potential applications in video understanding, robotics, and self-driving cars. It's also more computationally efficient than other models. Meta is presenting the paper on I-JEPA  at CVPR 2023 and has open-sourced the training code and model checkpoints. I-JEPA brings Meta closer to LeCun's vision of machines learning internal models of the world.