Research Scientist, Facebook (Meta) AI Research

Abstract: In this talk, I will talk about self-supervised (or unsupervised) learning which tries to learn models without human supervision. I’ll present a series of works on self-supervised learning in computer vision. These self-supervised models automatically learn features that can recognize objects, identify images, and surpass traditional machine learning methods that use human labeling. Self-supervised models can easily use billions of unlabeled images, and inspecting the models shows us that they can automatically discover objects. I will further show how these ideas can be applied for learning general purpose models of the world that can reason from multiple inputs such as images, videos, 3D. Overall, machine learning models can learn powerful, fair and robust models of the world without human supervision.