Facebook AI Develops SEER, a Self-Supervised Vision Model That Classifies Billions of Unlabeled Images
Facebook AI introduces SEER, a groundbreaking computer vision model that accurately classifies one billion random Instagram images without relying on human labels. This self-supervised system achieves an 84.4% accuracy rate by utilizing a new algorithmic pipeline.
Facebook AI unveils SEER, a self-supervised computer vision model that processes and classifies billions of images without any human annotations or labels. By feeding one billion random, unlabeled public images from Instagram into the system, the model achieves an impressive 84.4% classification accuracy. This milestone represents the highest performance ever recorded for a self-supervised AI model in the computer vision domain.
The system relies on three primary components to accomplish this massive task. The SwAV algorithm groups similar images together through online clustering, reducing training time by six times compared to previous methods. Additionally, a highly adaptable Convolutional Neural Network called RegNet handles trillions of parameters simultaneously, while the PyTorch-based VISSL library provides the scalable infrastructure necessary for self-supervised training at such an enormous scale.
Experts consider SEER a major breakthrough because self-supervised learning previously finds much more success in natural language processing than in visual recognition due to the vast diversity of pixels. By successfully applying this technique to optical vision, Facebook opens new avenues for analyzing highly diverse, unlabeled datasets. The company also open-sources the VISSL library, giving researchers worldwide the tools to build their own advanced self-supervised vision models.