Facebook AI Unveils SEER, a Self-Supervised Vision Model That Classifies a Billion Images
Facebook introduces SEER, a self-supervised computer vision model that accurately classifies one billion random Instagram images without any human labels. This breakthrough combines a novel clustering algorithm, a massive convolutional network, and an open-source PyTorch library.
Facebook recently unveils SEER, a self-supervised computer vision model that analyzes and classifies one billion random, unlabeled Instagram images with 82.4% accuracy. This impressive performance sets a new world record for self-supervised AI models in the visual domain, proving that machines can learn complex visual concepts without human annotations.
The system achieves this feat through a powerful combination of three main components: the SwAV online clustering algorithm, the highly scalable RegNet convolutional neural network, and the VISSL PyTorch library. Together, these technologies allow SEER to process trillions of pixels, group similar images together, and adapt to various runtime and memory requirements significantly faster than previous methods.
This development marks a major breakthrough for applying self-supervised learning to computer vision, a field where such techniques previously succeed mostly in natural language processing. To encourage further innovation, Facebook open-sources the VISSL library, enabling researchers worldwide to build their own diverse self-supervised models for optical vision tasks.