Facebook Unveils SEER, a Self-Teaching AI Model for Image Recognition
Facebook introduces SEER, a self-supervised AI that learns to recognize images without human-labeled data. The model achieves state-of-the-art accuracy by training on a billion random Instagram photos.
Facebook researchers introduce SEER, a new artificial intelligence model that teaches itself to recognize images without relying on human-labeled data. The system, whose name stands for SElf-supERvised, processes raw, unlabeled images from Instagram and learns to identify objects entirely on its own. Facebook describes this breakthrough as a potential revolution in the field of computer vision.
Self-supervised learning eliminates the need for time-consuming human annotation by allowing the AI to learn directly from unstructured information. This approach enables SEER to train much faster and work with significantly larger, more diverse datasets than traditional models. By analyzing a massive feed of one billion public Instagram photos, the AI builds its own understanding of visual concepts.
SEER achieves an impressive 84.2% accuracy rate on the industry-standard ImageNet object recognition test, outperforming existing advanced models. The system relies on three main components: the SwAV algorithm for rapid image clustering, the scalable RegNets architecture, and a specialized library called VISSL. Together, these technologies allow SEER to learn six times faster than previous state-of-the-art systems.