Facebook AI Breakthrough: SEER Model Classifies a Billion Images Without Human Labels
Facebook unveils SEER, a self-supervised computer vision model that accurately classifies one billion random Instagram images without any human annotations. This achievement marks a major breakthrough in applying self-supervised learning to visual data.
Facebook introduces SEER, a self-supervised computer vision model that processes billions of unlabelled images and classifies them completely on its own. After analyzing one billion random public images from Instagram without any human annotations or captions, the system achieves an impressive object detection and classification accuracy of 82.4%, setting a new world record for self-supervised AI models in the visual domain.
The system relies on three primary components to achieve this remarkable feat: the SwAV algorithm, the RegNet convolutional neural network, and the VISSL library. SwAV uses online clustering to group similar images together six times faster than previous algorithms, while RegNet filters images using trillions of parameters to adapt to various memory requirements. Additionally, Facebook open-sources the PyTorch-based VISSL library to help researchers worldwide build their own self-supervised optical vision models.
This development represents a significant breakthrough because self-supervised learning historically thrives in natural language processing but struggles with the vast diversity of visual pixels. By successfully applying self-supervised learning to computer vision at a massive scale, SEER opens exciting new avenues for artificial intelligence. The technology provides a highly beneficial tool for studying vast, diverse data sets without the time-consuming and expensive need for manual human labelling.